This is a plain text version of the ORIGINAL FILE. Please see that file instead.




Models, Resources, Applications


Models, Resources, Applications

Igor A. Bolshakov and Alexander Gelbukh


All rights reserved. No part of this publication may be reproduced,

stored in a retrieval system or transmitted, in any form or by any means,

electronic, mechanical, recording, photocopying, or otherwise, without

the prior permission of the publisher.


Dirección de Publicaciones

Tresguerras 27, 06040, DF


Torre de Rectoría, 9° Piso, Ciudad Universitaria, 045100, México DF


Carretera Picacho-Ajusco 227, 14200, México DF

ISBN: 970-36-0147- 2

Impreso en México / Printed in Mexico

Serie Ciencia de la Computación

The growth of the amount of available written information originated in the Renaissance with the invention of printing press and increased nowadays to unimaginable extent has obliged the man to acquire a new type of literacy related to the new forms of media besides writing. One of such forms is the computer—an object of the modern world that increases the degree of freedom of human action and knowledge, where the fantasy becomes reality, and the new common alphabet penetrates the presence marked by such a phenomenon as computing.

However, even though this phenomenon has become a part of our everyday life, the printed text has not been substituted by the electronic text; on the contrary, they have become into symbiotic elements that constitute fundamental means for accelerating the transition to the advance society and economy restructured towards the science, technology, and promotion and dissemination of knowledge. Only through such spread of knowledge is it possible to create a scientific culture founded on the permanent quest for the truth, informed criticism, and the systematic, rigorous, and intelligent way of human actions.

In this context, the Computer Science Series published by the Center for Computing Research (CIC) of the National Polytechnic Institute in collaboration with the National Autonomous University of Mexico and the Economic Culture Fund editorial house (Fondo de Cultura Económica) presents the works by outstanding Mexican and foreign specialists—outstanding both in their research and educational achievements—in the areas of tutoring systems, system modeling and simulation, numerical analysis, information systems, software engineering, geoprocessing, digital systems, electronics, automatic control, pattern recognition and image processing, natural language processing and artificial intelligence.

In this way, the publishing effort of the CIC—which includes the journal Computación y Sistemas, the Research on Computing Science series, the technical reports, conference proceedings, catalogs of solutions, and this book series—reaffirms its adherence to the high standards of research, teaching, industrial collaboration, guidance, knowledge dissemination, and development of highly skilled human resources.

This series is oriented to specialists in the field of computer science, with the idea to help them to extend and keep up to date their information in this dynamic area of knowledge. It is also intended to be a source of reference in their everyday research and teaching work. In this way one can develop himself or herself basing on the fundamental works of the scientific community—which promotion and dissemination of science is.

We believe that each and every book of this series is a must-have part of the library of any professional in computer science and allied areas who consider learning and keeping one’s knowledge up to date essential for personal progress and the progress of our country. Helpful support for this can be found in this book series characterized first and foremost by its originality and excellent quality.

Dr. Juan LuisDíaz De León Santiago

Center ForComputing Research



Preface. 5

I.    Introduction.. 15

II.   A Historical Outline. 33

III.  Products of Computational Linguistics: Present and Prospective  53

IV.  Language as a Meaning Û Text Transformer.. 83

V.   Linguistic Models. 129

Exercises. 153

Literature. 167

Appendices. 173

Detailed contents

Preface. 5

A new book on computational linguistics. 5

Objectives and intended readers of the book.. 9

Coordination with computer science. 10

Coordination with artificial intelligence. 11

Selection of topics. 12

Web resources for this book.. 13

Acknowledgments. 13

I.    Introduction.. 15

The role of natural language processing.. 15

Linguistics and its structure. 17

What we mean by computational linguistics. 25

Word, what is it?. 26

The important role of the fundamental science. 28

Current state of applied research on Spanish.. 30

Conclusions. 31

II.   A Historical Outline. 33

The structuralist approach.. 34

Initial contribution of Chomsky.. 34

A simple context-free grammar.. 35

Transformational grammars. 37

The linguistic research after Chomsky: Valencies and interpretation   39

Linguistic research after Chomsky: Constraints. 42

Head-driven Phrase Structure Grammar.. 44

The idea of unification.. 45

The Meaning Û Text Theory: Multistage transformer and government patterns  47

The Meaning Û Text Theory: Dependency trees. 49

The Meaning Û Text Theory: Semantic links. 50

Conclusions. 52

III.  Products of Computational Linguistics: Present and Prospective  53

Classification of applied linguistic systems. 53

Automatic hyphenation.. 54

Spell checking.. 55

Grammar checking.. 58

Style checking.. 60

References to words and word combinations. 61

Information retrieval. 63

Topical summarization.. 66

Automatic translation.. 70

Natural language interface. 73

Extraction of factual data from texts. 75

Text generation.. 76

Systems of language understanding.. 77

Related systems. 78

Conclusions. 81

IV.  Language as a Meaning Û Text Transformer.. 83

Possible points of view on natural language. 83

Language as a bi-directional transformer.. 85

Text, what is it?. 90

Meaning, what is it?. 94

Two ways to represent Meaning.. 96

Decomposition and atomization of Meaning.. 99

Not-uniqueness of Meaning Þ Text mapping: Synonymy.. 102

Not-uniqueness of Text Þ Meaning mapping: homonymy.. 103

More on homonymy.. 106

Multistage character of the Meaning Û Text transformer.. 110

Translation as a multistage transformation.. 113

Two sides of a sign.. 116

Linguistic sign.. 116

Linguistic sign in the mmt.. 117

Linguistic sign in hpsg.. 118

Are signifiers given by nature or by convention?. 119

Generative, mtt, and constraint ideas in comparison.. 120

Conclusions. 127

V.   Linguistic Models. 129

What is modeling in general?. 129

Neurolinguistic models. 130

Psycholinguistic models. 131

Functional models of language. 133

Research linguistic models. 134

Common features of modern models of language. 134

Specific features of the Meaning Û Text Model. 137

Reduced models. 141

Do we really need linguistic models?. 143

Analogy in natural languages. 145

Empirical versus rationalist approaches. 147

Limited scope of the modern linguistic theories. 149

Conclusions. 152

Exercises. 153

Review questions. 153

Problems recommended for exams. 157

Literature. 167

Recommended literature. 167

Additional literature. 168

General grammars and dictionaries. 169

References. 170

Appendices. 173

Some Spanish-oriented groups and resources. 173

English-Spanish dictionary of terminology.. 177

Index of illustrations. 180

Index of authors, systems, and terminology.. 182


Why did we decide to propose a new book on computational linguistics? What are the main objectives, intended readers, the main features, and the relationships of this book to various branches of computer science? In this Pre­face, we will try to answer these questions.

A new book on computationallinguistics

The success of modern software for natural language processing impressesour imagination. Programs for orthography and grammar correction, informationretrieval from document databases, and translation from one natural languageinto another, among others, are sold worldwide in millions of copies nowadays.

However, we have to admit that such programs still lack real intelligence.The ambitious goal of creating software for deep language understanding and production,which would provide tools powerful enough for fully adequate automatictranslation and man-machine communication in unrestricted natural language, hasnot yet been achieved, though attempts to solve this problem already have a historyof nearly 50 years.

This suggests that in order to solve the problem, developers of newsoftware will need to use the methods and results of a fundamental science, inthis case linguistics, rather than the tactics of ad hoc solutions. Neither increasing the speed of computers, nor refinement of programming tools, nor further development of numerous toy systems for language “understanding” in tiny domains, will suffice to solve one of the most challenging problems of modern science—automatic text understanding.

We believe that this problem, yet unsolved in the past century, will besolved in the beginning of this century by those who are sitting now on studentbenches. This book on computational linguistics models and their applicationsis targeted at these students, namely, at those students of the Latin Americanuniversities studying computer science and technology who are interested in thedevelopment of natural language processing software.

Thus, we expect the students to have already some background in computerscience, though no special training in the humanities and in linguistics inparticular.

On the modern book market, there are many texts on Natural LanguageProcessing (nlp), e.g. [1, 2, 7, 9]. They are quite appropriate as the further step in education for the students in computer science interested in the selected field. However, for the novices in linguistics (and the students in computer science are among them) the available books still leave some space for additional manuals because of the following shortages:

· Many of them are English-oriented. Meanwhile, English, in spite of all its vocabulary similarities with Spanish, is a language with quite a different grammatical structure. Unlike Spanish, English has a very strict word order and its morphology is very simple, so that the direct transfer of the methods of morphologic and syntactic analysis from English to Spanish is dubious.

· Only few of these manuals have as their objective to give a united and comparative exposition of various coexisting theories of text proces­sing. Moreover, even those few ones are not very successful since the methodologies to be observed and compared are too diverse in their approaches. Sometimes they even contradict each other in their definitions and notations.

· The majority of these manuals are oriented only to the formalisms of syntax, so that some of them seemingly reduce computational linguistics to a science about English syntax. Nevertheless, linguistics in general investigates various linguistic levels, namely, phonology, morphology, syntax, and semantics. For each of these levels, the amount of purely linguistic knowledge relevant for computational linguistics seems now much greater than that represented in well-known manuals on the subject.

Reckoning with all complications and controversies of the quicklydeveloping discipline, the main methodological features of this book on computationallinguistics are the following:

· Nearly all included examples are taken from Spanish. The other languages are considered mainly for comparison or in the cases when the features to be illustrated are not characteristic to Spanish.

· A wide variety of the facts from the fundamental theory—general linguistics—that can be relevant for the processing of natural languages, right now or in the future, are touched upon in one way or another, rather than only the popular elements of English-centered manuals.

· Our educational line combines various approaches coexisting in computational linguistics and coordinates them wherever possible.

· Our exposition of the matter is rather slow and measured, in order to be understandable for the readers who do not have any background in linguistics.

In fact, we feel inappropriate to simply gather disjoint approaches undera single cover. We also have rejected the idea to make our manual a referencebook, and we do not have the intention to give always well-weighted reviews and numerous references through our texts. Instead, we consider the coherence, consistency, and self-containment of exposition to be much more important.

The two approaches that most influenced the contents of this book are thefollowing:

· The Meaning Û Text Theory (mtt), developed by Igor Mel’čuk, Alexander Žolkovsky, and Yuri Apresian since the mid-sixties, facilitates describing the elements, levels, and structures of natural languages. This theory is quite appropriate for any language, but especially suits for languages with free word order, including Spanish. Additionally, the mtt gives an opportunity to validate and extend the traditional terminology and methodology of linguistics.

· The Head-driven Phrase Structure Grammar (hpsg), developed by Carl Pollard and Ivan Sag in the last decade, is probably the most advanced practical formalism in natural language description and processing within the modern tradition of generative grammars originated by Noam Chomsky. Like the mtt, hpsg takes all known facts for description of natural languages and tries to involve new ones. As most of the existing formalisms, this theory was mainly tested on English. In recent years, however, hpsg has acquired numerous followers among researchers of various languages, including Spanish. Since the main part of the research in nlp has been fulfilled till now in the Chomskian paradigm, it is very important for a specialist in computational linguistics to have a deeper knowledge of the generative grammar approach.

The choice of our material is based on our practical experience and onour observations on the sources of the problems which we ourselves and our colleaguesencountered while starting our careers in computational linguistics and whichstill trouble many programmers working in this field due to the lack offundamental linguistic knowledge.

After coping with this book, our reader would be more confident to begin studying such branches of computational linguistics as

· Mathematical Tools and Structures of Computational Linguistics,

· Phonology,

· Morphology,

· Syntax of both  surface and deep levels, and

· Semantics.

The contents of the book are based on the course on computationallinguistics that has been delivered by the authors since 1997 at the Center forComputing Research, National Polytechnic Institute, Mexico City. This coursewas focused on the basic set of ideas and facts from the fundamental sciencenecessary for the creation of intelligent language processing tools, withoutgoing deeply into the details of specific algorithms or toy systems. Thepractical study of algorithms, architectures, and maintenance of real-worldapplied linguistic systems may be the topics of other courses.

Since most of the literature on this matter is published in English regardlessof the country where the research was performed, it will be useful for thestudents to read an introduction to the field in English. However, Spanishterminological equivalents are also given in the Appendix (see page 173).

The book is also supplied with 54 review questions, 58 test questionsrecommended for the exam, with 4 variants of answer for each one, 30 illustrations,58 bibliographic references, and 37 references to the most relevant Internetsites.

The authors can be contacted at the following e-mail addresses: igor*, gelbukh* (gelbukh*; see also ( The webpage for this book is

Objectives and intended readers of the book

The main objectives of this book are to provide the students with fewfundamentals of general linguistics, to describe the modern models of hownatural languages function, and to explain how to compile the data—linguistictables and machine dictionaries—necessary for the natural language processingsystems, out of informally described facts of a natural language. Therefore, wewant to teach the reader how to prepare all the necessary tools for thedevelopment of programs and systems oriented to automatic natural languageprocessing. In order to repeat, we assume that our readers are mainly studentsin computer sciences, i.e., in software development, database management,information retrieval, artificial intelligence or computer science in general.

Throughout this book, special emphasis is made on applications to the Spa­nishlanguage. However, this course is not a mere manual of Spanish. A broader basisfor understanding the main principles is to be elucidated through some examplesfrom English, French, Portuguese, and Russian. Many literature sources providethe reader with interesting examples for these languages. In our books, weprovide analogous examples for Spanish wherever possible.

Significant difficulties were connected with the fact that Latin Americanstudents of technical institutes have almost no knowledge in linguistics beyondsome basics of Spanish grammar they learned in their primary schooling, atleast seven years earlier. Therefore, we have tried to make these books understandablefor students without any background in even rather elementary grammar.

Neither it should be forgotten that the native language is studied in theschool prescriptively, i.e., how it is preferable or not recommendable to speak and write, rather than descriptively, i.e., how the language is really structured and used.

However, only complete scientific description can separate correct oradmissible language constructions from those not correct and not belonging tothe language under investigation. Meantime, without a complete and correctdescription, computer makes errors quite unusual and illogical from a humanpoint of view, so that the problem of text processing cannot be successfullysolved.

Coordination with computer science

The emphasis on theoretical issues of language in this book should not beunderstood as a lack of coordination between computational linguistics andcomputer science in general. Computer science and practical programming is apowerful tool in all fields of information processing. Basic knowledge ofcomputer science and programming is expected from the reader.

The objective of the book is to help the students in developing appliedsoftware systems and in choosing the proper models and data structures forthese systems. We only reject the idea that the computer science’s tools ofrecent decades are sufficient for computational linguistics in theoreticalaspects. Neither proper structuring of linguistic programs, nor object-orientedtechnology, nor specialized languages of artificial intelligence like Lisp orProlog solve by themselves the problems of computational linguistics. All thesetechniques are just tools.

As it is argued in this book, the ultimate task of many applied linguisticsystems is the transformation of an unprepared, unformatted natural languagetext into some kind of representation of its meaning, and, vice versa, thetransformation of the representation of meaning to a text. It is the main taskof any applied system.

However, the significant part of the effort in the practical developingof an npl system is not connected directly with creating the software for this ultimate task. Instead, more numerous, tedious, and inevitable programming tasks are connected with extraction of data for the grammar tables and machine dictionaries from various texts or human-oriented dictionaries. Such texts can be originally completely unformatted, partially formatted, or formalized for some other purposes. For example, if we have a typical human-oriented dictionary, in the form of a text file or database, our task can be to parse each dictionary entry and to extract all of the data necessary for the ultimate task formulated above.

This book contains the material to learn how to routinely solve suchtasks. Thus, again, we consider programming to be the everyday practical toolof the reader and the ultimate goal of our studies.

Coordination with artificial intelligence

The links between computational linguistics and artificial intelligence (ai) are rather complicated. Those ai systems that contain subsystems for natural language processing rely directly on the ideas and methods of computational linguistics. At the same time, some methods usually considered belonging only to ai, such as, for example, algorithms of search for decision in trees, matrices, and other complex structures, sometimes with backtracking, are applicable also to linguistic software systems.

Because of this, many specialists in ai consider computational linguistics a part of ai [2, 40, 54]. Though such an expansion hampers nothing, it is not well grounded, in our opinion, since computational linguistics has its own theoretical background, scientific neighbors, methods of knowledge representation, and decision search. The sample systems of natural language processing, which wander from one manual on ai to another, seem rather toy and obsolete.

Though the two fields of research are different, those familiar with bothfields can be more productive. Indeed, they do not need to invent things alreadywell known in the adjacent field, just taking from the neighborhood what theyprefer for their own purposes. Thus, we encourage our readers to deeplyfamiliarize themselves with the area of ai. We also believe that our books could be useful for students specializing in ai.

Selection of topics

Since the mtt described below in detail improves and enriches rather than rejects the previous tradition in general linguistics, we will mainly follow the mtt in description and explanation of facts and features of natural languages, giving specific emphasis on Spanish.

To avoid a scientific jumble, we usually do not mark what issues arecharacteristic to the mtt, but are absent or are described in a different way in other theories. We give in parallel the point of view of the hpsg-like formalisms on the same issue only in the places where it is necessary for the reader to be able to understand the corresponding books and articles on computational linguistics written in the tradition originated by Chomsky.

We should assert here that the mtt is only a tool for language description. It does not bind researchers to specific algorithms or to specific formats of representation of linguistic data. We can find the same claims about a purely linguistic orientation in the recent books on the hpsg as well, though the latter approach seems more computer-oriented. In prospect, we hope, those two approaches will complement and enrich each other.

Web resources for this book

The webpage of this book is You can find thereadditional materials on the topics of this book, links to relevant Internetresources, and errata. Many publications by the authors of this book can befound at


We are profoundly grateful to Prof. Igor Mel’čuk for placing in our disposal his recent works and for reading through a nearly ready version of the book, with a lot of bitter criticism. We are grateful also to Prof. Yuri Apresian and his research team in Moscow, especially to Prof. Leonid Tsinman and Prof. Igor Boguslavsky, for providing us with the materials on the applications of the Meaning Û Text Theory and for some discussions. Dr. Patrick Cassidy of MICRA, Inc., has read parts of the manuscript and provided us with helpful advice.

We express our most cordial thanks to our doctoral student Sofía N. GaliciaHaro, who performed a lot of administrative chores and paperwork, freeing ourtime for the work on the book. She also was our main advisor in everything concerning Spanish language.

Finally, we thank our families for their great patience and constant support.

I.  Introduction

Is it necessary to automatically process natural language texts, if we can just read them? Moreover, is it difficult anyway, when every child easily learns how to read in primary school? Well, if it is difficult, is it then possible at all? What do we need to know in order to develop a computer program that would do it? What parts of linguistics are most important for this task?

In this introductory chapter, we will answer these questions, and themain answers will be: yes, it is necessary; yes, it is very difficult; and yes, it is possible.

The role of natural language processing

We live in the age of information. It pours upon us from the pages ofnewspapers and magazines, radio loudspeakers, tv and computer screens. The main part of this information has the form of natural language texts. Even in the area of computers, a larger part of the information they manipulate nowadays has the form of a text. It looks as if a personal computer has mainly turned into a tool to create, proofread, store, manage, and search for text documents.

Our ancestors invented natural language many thousands of years ago forthe needs of a developing human society. Modern natural languages aredeveloping according to their own laws, in each epoch being an adequate toolfor human communication, for expressing human feelings, thoughts, and actions.The structure and use of a natural language is based on the assumption that theparticipants of the conversation share a very similar experience and knowledge,as well as a manner of feeling, reasoning, and acting. The great challenge ofthe problem of intelligent automatic text processing is to use unrestrictednatural language to exchange information with a creature of a totally differentnature: the computer.

For the last two centuries, humanity has successfully coped with theautomation of many tasks using mechanical and electrical devices, and thesedevices faithfully serve people in their everyday life. In the second half of the twentieth century, human attention has turned to the automation of naturallanguage processing. People now want assistance not only in mechanical, butalso in intellectual efforts. They would like the machine to read an unpreparedtext, to test it for correctness, to execute the instructions contained in the text, or even to comprehend it well enough to produce a reasonable responsebased on its meaning. Human beings want to keep for themselves only the final decisions.

The necessity for intelligent automatic text processing arises mainlyfrom the following two circumstances, both being connected with the quantity ofthe texts produced and used nowadays in the world:

· Millions and millions of persons dealing with texts throughout the world do not have enough knowledge and education, or just time and a wish, to meet the modern standards of document processing. For example, a secretary in an office cannot take into consideration each time the hundreds of various rules necessary to write down a good business letter to another company, especially when he or she is not writing in his or her native language. It is just cheaper to teach the machine once to do this work, rather than repeatedly teach every new generation of computer users to do it by themselves.

· In many cases, to make a well-informed decision or to find information, one needs to read, understand, and take into consideration a quantity of texts thousands times larger than one person is physically able to read in a lifetime. For example, to find information in the Internet on, let us say, the expected demand for a specific product in the next month, a lot of secretaries would have to read texts for a hundred years without eating and sleeping, looking through all the documents where this information might appear. In such cases, using a computer is the only possible way to accomplish the task.

Thus, the processing of natural language has become one of the main problemsin information exchange. The rapid development of computers in the last twodecades has made possible the implementation of many ideas to solve theproblems that one could not even imagine being solved automatically, say, 45years ago, when the first computers appeared.

Intelligent natural language processing is based on the science calledcomputational linguistics. Computational linguistics is closely connected with applied linguistics and linguistics in general. Therefore, we shall first outlineshortly linguistics as a science belonging to the humanities.

                          Linguistics and its structure                         

Linguistics is a science about natural languages. To be more precise, itcovers a whole set of different related sciences (see Figure I.1).

General linguistics is a nucleus [18, 36]. It studies the general structure of various natural languages and discovers the universal laws of functioning of natural languages. Many concepts from general linguistics prove to be necessary for any researcher who deals with natural languages. General linguistics is a fundamental science that was developed by many researchers during the last two centuries, and it is largely based on the methods and results of grammarians of older times, beginning from the classical antiquity.

As far as general linguistics is concerned, its most important parts are the following:

· Phonology deals with sounds composing speech, with all their similarities and differences permitting to form and distinguish words.

· Morphology deals with inner structure of individual words and the laws concerning the formation of new words from pieces¾morphs.

· Syntax considers structures of sentences and the ways individual words are connected within them.

Figure I.1. Structure of linguistic science.

· Semantics and pragmatics are closely related. Semantics deals with the meaning of individual words and entire texts, and pragmatics studies the motivations of people to produce specific sentences or texts in a specific situation.

There are many other, more specialized, components of linguistics as awhole (see Figure I.1).

Historical, or comparative, linguistics studies history of languages by their mutual comparison, i.e., investigating the history of their similarities and differences. The second name is explained by the fact that comparison is the main method in this branch of linguistics. Comparative linguistics is even older than general linguistics, taking its origin from the eighteenth century.

Many useful notions of general linguistics were adopted directly fromcomparative linguistics.

From the times of Ferdinand de Saussure, the history of language has been called diachrony of language, as opposed to the synchrony of language dealing with phenomena of modern natural languages only. Therefore, diachrony describes changes of a language along the time axis.

Historical linguistics discovered, for example, that all Romance languages (Spanish, Italian, French, Portuguese, Romanian, and several others) are descendants of Latin language. All languages of the Germanic family (German, Dutch, English, Swedish, and several others) have their origins in a common language that was spoken when German tribes did not yet have any written history. A similar history was discovered for another large European family of languages, namely, for Slavonic languages (Russian, Polish, Czech, Croatian, Bulgarian, among others).

Comparative study reveals many common words and constructions within eachof the mentioned families—Romance, Germanic, and Slavonic—taken separately.

At the same time, it has noticed a number of similar words among thesefamilies. This finding has led to the conclusion that the mentioned familiesform a broader community of languages, which was called Indo-Europeanlanguages. Several thousand years ago, the ancestors of the people now speakingRomance, Germanic, and Slavonic languages in Europe probably formed a commontribe or related tribes.

At the same time, historic studies permits to explain why English has somany words in common with the Romance family, or why Romanian language has somany Slavonic words (these are referred to as loan words).

Comparative linguistics allows us to predict the elements of one languagebased on our knowledge of another related language. For example, it is easy toguess the unknown word in the following table of analogy:









Based on more complicated phonologic laws, it is possible even to predictthe pronunciation of the French word for the Spanish agua (namely [o], eau in the written form), though at the first glance these two words are quite different (actually, both were derived from the Latin word aqua).

As to computational linguistics, it can appeal to diachrony, but usuallyonly for motivation of purely synchronic models. History sometimes gives goodsuggestions for description of the current state of language, helping the researcherto understand its structure.

Contrastive linguistics, or linguistic typology, classifies a variety of languages according to the similarity of their features, notwithstanding the origin of languages. The following are examples of classification of languages not connected with their origin.

Some languages use articles (like a and the in English) as an auxiliary part of speech to express definite/indefinite use of nouns. (Part of speech is defined as a large group of words having some identical morphologic and syntactic properties.) Romance and Germanic languages use articles, as well as Bulgarian within the Slavonic family. Meantime, many other languages do not have articles (nearly all Slavonic family and Lithuanian, among others). The availability of articles influences some other features of languages.

Some languages have the so-called grammatical cases for several parts of speech (nearly all Slavonic languages, German, etc.), whereas many others do not have them (Romance languages, English—from the Germanic family, Bulgarian—from the Slavonic family, and so on).

Latin had nominative (direct) case and five oblique cases: genitive,dative, accusative, ablative, and vocative. Russian has also six cases, andsome of them are rather similar in their functions to those of Latin. Inflectedparts of speech, i.e., nouns, adjectives, participles, and pronouns, havedifferent word endings for each case.

In English, there is only one oblique case, and it is applicable only tosome personal pronouns: me, us, him, her, them.

In Spanish, two oblique cases can be observed for personal pronouns,i.e., dative and accusative: le, les, me, te, nos, las, etc. Grammatical cases give additional mean for exhibiting syntactic dependencies between words in a sentence. Thus, the inflectional languages have common syntactic features.

In a vast family of languages, the main type of sentences contains a syntactic subject (usually it is the agent of an action), a syntactic predicate (usually it denotes the very action), and a syntactic object (usually it is the target or patient of the action). The subject is in a standard form (i.e., in direct, or nominative, case), whereas the object is usually in an oblique case or enters in a prepositional group. This is referred to as non-ergative construction.

Meantime, a multiplicity of languages related to various other families,not being cognate to each other, are classified as ergative languages. In a sentence of an ergative language, the agent of the action is in a special oblique (called ergative) case, whereas the object is in a standard form. In some approximation, a construction similar to an ergative one can be found in the Spanish sentence Me simpatizan los vecinos, where the real agent (feeler) yo ‘I’ is used in oblique case me, whereas the object of feeling, vecinos, stays in the standard form. All ergative languages are considered typologically similar to each other, though they might not have any common word. The similarity of syntactical structures unites them in a common typological group.

Sociolinguistics describes variations of a language along the social scale. It is well known that various social strata often use different sublanguages within the same common language, wherever the same person uses different sublanguages in different situations. It suffices to compare the words and their combinations you use in your own formal documents and in conversations with your friends.

Dialectology compares and describes various dialects, or sublanguages, of a common language, which are used in different areas of the territory where the same language is officially used. It can be said that dialectology describes variations of a language throughout the space axis (while diachrony goes along the time axis). For example, in different Spanish-speaking countries, many words, word combinations, or even grammatical forms are used differently, not to mention significant differences in pronunciation. Gabriel García Márquez, the world-famous Colombian writer, when describing his activity as a professor at the International Workshop of cinematographers in Cuba, said that it was rather difficult to use only the words common to the entire Spanish-speaking world, to be equally understandable to all his pupils from various countries of Latin America. A study of Mexican Spanish, among other variants of Spanish language is a good example of a task in the area of dialectology.

Lexicography studies the lexicon, or the set of all words, of a specific language, with their meanings, grammatical features, pronunciation, etc., as well as the methods of compilation of various dictionaries based on this knowledge. The results of lexicography are very important for many tasks in computational linguistics, since any text consists of words. Any automatic processing of a text starts with retrieving the information on each word from a computer dictionary compiled beforehand.

Psycholinguistics studies the language behavior of human beings by the means of a series of experiments of a psychological type. Among areas of its special interest, psycholinguists studies teaching language to children, links between the language ability in general and the art of speech, as well as other human psychological features connected with natural language and expressed through it. In many theories of natural language processing, data of psycholinguistics are used to justify the introduction of the suggested methods, algorithms, or structures by claiming that humans process language “just in this way.”

Mathematical linguistics. There are two different views on mathematical linguistics. In the narrower view, the term mathematical linguistics is used for the theory of formal grammars of a specific type referred to as generative grammars. This is one of the first purely mathematical theories devoted to natural language. Alternatively, in the broader view, mathematical linguistics is the intersection between linguistics and mathematics, i.e., the part of mathematics that takes linguistic phenomena and the relationships between them as the objects of its possible applications and interpretations.

Since the theory of generative grammars is nowadays not unique amonglinguistic applications of mathematics, we will follow the second, broader view on mathematical linguistics.

One of the branches of mathematical linguistics is quantitativelinguistic. It studies language by means of determining the frequencies of various words, word combinations, and constructions in texts. Currently, quantitative linguistics mainly means statistical linguistics. It provides the methods of making decisions in text processing on the base of previously gathered statistics.

One type of such decisions is resolution of ambiguity in text fragments to be analyzed. Another application of statistical methods is in the deciphering of texts in forgotten languages or unknown writing systems. As an example, deciphering of Mayan glyphs was fulfilled in the 1950’s by Yuri Knorozov [39] taking into account statistics of different glyphs (see Figure I.2).

Applied linguistics develops the methods of using the ideas and notions of general linguistics in broad human practice. Until the middle of the twentieth century, applications of linguistics were limited to developing and improving grammars and dictionaries in a printed form oriented to their broader use by non-specialists, as well as to the rational methods of teaching natural languages, their orthography and stylistics. This was the only purely practical product of linguistics.

Figure I.2. The ancient Mayan writing system was deciphered with statistical methods.

In the latter half of the twentieth century, a new branch of appliedlinguistics arose, namely the computational, or engineering, linguistics. Actually, this is the main topic of this book, and it is discussed in some detail in the next section.

What we mean by computational linguistics

Computational linguistics might be considered as a synonym of automatic processing of natural language, since the main task of computational linguistics is just the construction of computer programs to process words and texts in natural language.

The processing of natural language should be considered here in a verybroad sense that will be discussed later.

Actually, this course is slightly “more linguistic than computational,”for the following reasons:

· We are mainly interested in the formal description of language relevant to automatic language processing, rather than in purely algorithmic issues. The algorithms, the corresponding programs, and the programming technologies can vary, while the basic linguistic principles and methods of their description are much more stable.

· In addition to some purely computational issues, we also touch upon the issues related to computer science only in an indirect manner. A broader set of notions and models of general linguistics and mathematical linguistics are described below.

For the purposes of this course, it is also useful to draw a line between the issues in text processing we consider linguistic—and thus will discuss below—and the ones we will not. Inour opinion, for a computer system or its part to be considered linguistic, itshould use some data or procedures that are:

· language-dependent, i.e., change from one natural language to another,

· large, i.e., require a significant amount of work for compilation.

Thus, not every program dealing with natural language texts is related tolinguistics. Though such word processors as Windows’ Notebook do deal with theprocessing of texts in natural language, we do not consider them linguisticsoftware, since they are not sufficiently language-dependent: they can be usedequally for processing of Spanish, English, or Russian texts, after somealphabetic adjustments.

Let us put another example: some word processors can hyphenate wordsaccording to the information about the vowels and consonants in a specific alphabet and about syllable formation in a specific language. Thus, they are language-dependent. However, they do not rely on large enough linguistic resources. Therefore, simple hyphenation programs only border upon the software that can be considered linguistic proper. As to spell checkers that use a large word list and complicated morphologic tables, they are just linguistic programs.

Word, what is it?

As it could be noticed, the term word was used in the previous sections very loosely. Its meaning seems obvious: any language operates with words and any text or utterance consists of them. This notion seems so simple that, at the first glance, it does not require any strict definition or further explanation: one can think that a word is just a substring of the text as a letter string, from the first delimiter (usually, a space) to the next one (usually, a space or a punctuation mark). Nevertheless, the situation is not so simple.

Let us consider the Spanish sentence Yo devuelvo los libros el próximo mes, pero tú me devuelves el libroahora. How many words does it contain? One can say 14 and will be right, since there are just 14 letter substrings from one delimiter to another in this sentence. One can also notice that the article el is repeated twice, so that the number of different words (substrings) is 13. For these observations, no linguistic knowledge is necessary.

However, one can also notice that devuelvo and devuelves are forms of the same verb devolver, and libros and libro are forms of the same noun libro, so that the number of different words is only 11. Indeed, these pairs of wordforms denote the same action or thing. If one additionally notices that the article los is essentially equivalent to the article el whereas the difference in grammatical number is ignored, then there are only 10 different words in this sentence. In all these cases, the “equivalent” strings are to some degree similar in their appearance, i.e., they have some letters in common.

At last, one can consider me the same as yo, but given in oblique grammatical case, even though there are no letters in common in these substrings. For such an approach, the total number of different words is nine.

We can conclude from the example that the term word is too ambiguous to be used in a science with the objective to give a precise description of a natural language. To introduce a more consistent terminology, let us call an individual substring used in a specific place of a text (without taking into account its possible direct repetitions or similarities to other substrings) a word occurrence. Now we can say that the sentence above consisted of 14 word occurrences.

Some of the substrings (usually similar in the appearance) have the samecore meaning. We intuitively consider them as different forms of some commonentity. A set of such forms is called lexeme. For example, in Spanish {libro, libros}, {alto, alta, altos, altas}, and {devolver, devuelvo, devuelves, devuelve, devolvemos...} are lexemes. Indeed, in each set there is a commonality between the strings in the letters they consist of (the commonality being expressed as patterns libro‑, alt‑, and‑), and their meanings are equivalent (namely, ‘book’, ‘high’, and ‘to bring back’, correspondingly). Each entry of such a set—a letter string without regard to its position in the text—is called wordform. Each word occurrence represents a wordform, while wordforms (but not word occurrences) can repeat in the text. Now we can say that the sentence in the example above contains 14 word occurrences, 13 different wordforms, or nine different lexemes. The considerations that gave other figures in the example above are linguistically inconsistent.

A lexeme is identified by a name. Usually, one of its wordforms, i.e., aspecific member of the wordform set, is selected for this purpose. In the previousexamples, LIBRO, ALTO, and DEVOLVER are taken as names of the corresponding lexemes. Just these names are used as titles of the corresponding entries in dictionaries mentioned above. The dictionaries cover available information about lexemes of a given language, sometimes including morphologic information, i.e., the information on how wordforms of these lexemes are constructed. Various dictionaries compiled for the needs of lexicography, dialectology, and sociolinguistics have just lexemes as their entries rather than wordforms.

Therefore, the term word, as well as its counterparts in other languages, such as Spanish palabra, is too ambiguous to be used in a linguistic book. Instead, we should generally use the terms word occurrence for a specific string in a specific place in the text, wordform for a string regardless to its specific place in any text, and lexeme for a theoretical construction uniting several wordforms corresponding to a common meaning in the manner discussed above.

However, sometimes we will retain the habitual word word when it is obvious which of these more specific terms is actually meant.

The important role of the fundamental science

In the past few decades, many attempts to build language processing orlanguage understanding systems have been undertaken by people withoutsufficient knowledge in theoretical linguistics. They hoped that they wouldsucceed thanks to clever mathematical algorithms, fine programming in Assemblerlanguage, or just the speed of their computers. To our knowledge, all suchattempts have failed. Even now it is still worthwhile to explain the necessityto have fundamental knowledge for those who would develop such systems, andthus to clarify why we decided to start a course in computational linguisticsfrom notions of general linguistics.

General linguistics is a fundamental science belonging to the humanities.An analogy with another fundamental science—physics—is appropriate here.

A specialist with deep knowledge of physics would easily understand thestructure of any new electrical device.

Moreover, since the fundamentals of physics are changing very slowly,such a specialist would be able to understand those new or revolutionaryengineering principles that did not even exist at the time when he or she studiedin a university. Indeed, the underlying fundamentals of physics have remainedthe same.

On the contrary, somebody with narrow specialization only in laserdevices might be perplexed with any new principle that does not relate to hisor her deep but narrow knowledge.

The same situation can be observed with linguistics. Even experiencedprogrammers who have taken solid courses on software systems for naturallanguage processing might become helpless in cases where a deeper penetration into linguistic notions and laws is needed.

What is more, they will hardly be able to compile a formal grammar,grammar tables, or a computer dictionary of a natural language, whereas theprogram heavily relies on such tables and dictionaries as on its integralparts.

They will not even be able to understand the suggestions of professionallinguists (who regrettably in many cases prove to know little aboutprogramming) and will not be able to explain to them the meaning of the datastructures the program needs to use.

On the contrary, a good background in linguistics will allow the newspecialists in computational linguistics to work productively in an interdisciplinaryteam, or just to compile all the necessary tables and dictionaries on theirown.

It might even guide them to some quite new ideas and approaches. Thesespecialists will better understand the literature on the subject, which is veryimportant in our rapidly developing world.

Current state of applied research on Spanish

In our books, the stress on Spanish language is made intentionally andpurposefully. For historical reasons, the majority of the literature on naturallanguages processing is not only written in English, but also takes English asthe target language for these studies. In our opinion, this iscounter-productive and thus it has become one of the causes of a lag in appliedresearch on natural language processing in many countries, compared to the United States. The Spanish-speaking world is not an exception to this rule.

The number of Spanish-speaking people in the world has exceeded now 400million, and Spanish is one of the official languages of the United Nations. Asto the human-oriented way of teaching, Spanish is well described, and the RoyalAcademy of Madrid constantly supports orthographic [33] and grammaticalresearch [30] and standardization. There are also several good academic-typedictionaries of Spanish, one the best of which being [28].

However, the lexicographic research reflected in these dictionaries istoo human-oriented. Along with some historical information, these dictionariesprovide semantic explanations, but without a formal description of the main linguistic properties of lexemes, even in morphologic and syntactic aspects.

Formal description and algorithmization of a language is the objective ofresearch teams in computational linguistics. Several teams in this field orientedto Spanish work now in Barcelona and Madrid, Spain. However, even this israther little for a country of the European Community, where unilingual andmultilingual efforts are well supported by the governmental and internationalagencies. Some research on Spanish is conducted in the United States, for example, at New Mexico State University.

In Mexico—the world’s largest Spanish-speaking country—the activity incomputational linguistics has been rather low in the past decades. Now, theteam headed by Prof. L.A. Pineda Cortés at National Autonomous University of Mexico is working on a very difficult task of creation of a program that will be able to perform a dialogue in Spanish with a human. A very useful dictionary of modern Mexican Spanish, developed by the team headed by Prof. L.F. Lara Ramos [26] (see also [47]), is oriented to human users, giving semantic interpretations and suggestions on good usage of words.

Some additional information on Spanish-oriented groups can be found inthe Appendix on the page 173.

As to the books by Helena Beristáin [11], Irene Gartz [15], and J.L. Fuentes [14] on Spanish grammar, they are just well structured[1] manuals of language oriented to native speakers, and thus cannot be used directly as a source of grammatical information for a computer program.

One of the most powerful corporations in the world, Microsoft, hasannounced the development of a natural language processing system for Spanishbased on the idea of multistage processing. As usually with commercialdevelopments, the details of the project are still rather scarce. We can onlyguess that a rather slow progress of the grammar checker of Word text processorfor Windows is related somehow with these developments.

Thus, one who needs to compile all facts of Spanish relevant for itsautomatic processing faces with a small set of rather old monographs andmanuals oriented to human learners, mainly written and first published in Spainand then sometimes reprinted elsewhere in Latin America.

Meantime, a development of natural language processing tools is quitenecessary for any country to be competitive in the twenty-first century. We hopethat our books will contribute to such developments in Mexico.


The twenty-first century will be the century of the total informationrevolution. The development of the tools for the automatic processing of thenatural language spoken in a country or a whole group of countries is extremelyimportant for the country to be competitive both in science and technology.

To develop such applications, specialists in computer science need tohave adequate tools to investigate language with a view to its automaticprocessing. One of such tools is a deep knowledge of both computationallinguistics and general linguistic science.

II.  A Historical Outline

A course on linguistics usually follows one of the general models, or theories, of natural language, as well as the corresponding methods of interpretation of the linguistic phenomena.

A comparison with physics is appropriate here once more. For a long time,the Newtonian theory had excluded all other methods of interpretation of phenomenain mechanics. Later, Einstein’s theory of relativity incorporated the Newtoniantheory as an extreme case, and in its turn for a long time excluded othermethods of interpretation of a rather vast class of phenomena. Such exclusivitycan be explained by the great power of purely mathematical description ofnatural phenomena in physics, where theories describe well-known facts and predictwith good accuracy the other facts that have not yet been observed.

In general linguistics, the phenomena under investigation are much morecomplicated and variable from one object (i.e., language) to another than inphysics. Therefore, the criteria for accuracy of description and prediction ofnew facts are not so clear-cut in this field, allowing different approaches tocoexist, affect each other, compete, or merge. Because of this, linguistics hasa rich history with many different approaches that formed the basis for thecurrent linguistic theories.

Let us give now a short retrospective of the development of generallinguistics in the twentieth century. The reader should not be worried if he orshe does not know many terms in this review not yet introduced in this book.There will be another place for strict definitions in this book.

The structuralist approach

At the beginning of the twentieth century, Ferdinand de Saussure had developed a new theory of language. He considered natural language as a structure of mutually linked elements, similar or opposed to each other. Later, several directions arose in general linguistics and many of them adhered to the same basic ideas about language. This common method was called structuralism, and the corresponding scientific research (not only in linguistics, but also in other sciences belonging to the humanities) was called structuralist.

Between the 1920’s and the 1950’s, several structuralist schools wereworking in parallel. Most of them worked in Europe, and European structuralismkept a significant affinity to the research of the previous periods in its terminologyand approaches.

Meantime, American structuralists, Leonard Bloomfield among them, made claims for a fully “objective” description of natural languages, with special attention to superficially observable facts. The order of words in a sentence was considered the main tool to become aware of word grouping and sentence structures. At this period, almost every feature of English seemed to confirm this postulate. The sentences under investigation were split into the so-called immediate constituents, or phrases, then these constituents were in their turn split into subconstituents, etc., down to single words. Such a method of syntactic structuring was called the phrase structure, or constituency, approach.

Initial contribution of Chomsky

In the 1950’s, when the computer era began, the eminent American linguistNoam Chomsky developed some new formal tools aimed at a better description of facts in various languages [12].

Among the formal tools developed by Chomsky and his followers, the twomost important components can be distinguished:

· A purely mathematical nucleus, which includes generative grammars arranged in a hierarchy of grammars of diverse complexity. The generative grammars produce strings of symbols, and sets of these strings are called formal languages, whereas in general linguistics they could be called texts. Chomskian hierarchy is taught to specialists in computer science, usually in a course on languages and automata. This redeems us from necessity to go into any details. The context-free grammars constitute one level of this hierarchy.

· Attempts to describe a number of artificial and natural languages in the framework of generative grammars just mentioned. The phrase structures were formalized as context-free grammars (cfg) and became the basic tool for description of natural languages, in the first place, of English. Just examples of these first attempts are extensively elaborated in the manuals on artificial intelligence. It is a good approach unless a student becomes convinced that it is the only possible.

A simple context-free grammar

Let us consider an example of a context-free grammar for generating verysimple English sentences. It uses the initial symbol S of a sentence to be generated and several other non-terminal symbols: the noun phrase symbol NP, verb phrase symbol VP, noun symbol N, verb symbol V, and determinant symbol D. All these non-terminal symbols are interpreted as grammatical categories.

Several production rules for replacement of a non-terminal symbol with a string of several other non-terminal symbols are used as the nucleus of any generative grammar. In our simple case, let the set of the rules be the following:

S     ® NP VP

VP   ® V NP

NP ® D N

NP ® N

Each symbol at the right side of a rule is considered a constituent of the entity symbolized at the left side. Using these rules in any possible order,we can transform S to the strings D N V D N, or D N V N, or N V D N, or N V N, etc.

An additional set of rules is taken to convert all these non-terminalsymbols to the terminal symbols corresponding to the given grammatical categories. The terminals are usual words of Spanish, English, or any other language admitting the same categories and the same word order. We use the symbol “|” as a metasymbol of an alternative (i.e. for logical OR). Let the rules be the following:

N    ® estudiante | niña | María | canción | edificio...

V     ® ve | canta | pregunta...

D    ® el | la | una | mi | nuestro...

Applying these rules to the constituents of the non-terminal stringsobtained earlier, we can construct a lot of fully grammatical and meaningfulSpanish sentences like María ve el edificio (from N V D N) or la estu­diante canta una canción (from D N V D N). Some meaningless and/or ungrammatical sentences like canción ve el María can be generated too. With more complicate rules, some types of ungrammaticality can be eliminated. However, to fully get rid of potentially meaningless sentences is very difficult, since from the very beginning the initial symbol does not contain any specific meaning at all. It merely presents an abstract category of a sentence of a very vast class, and the resulting meaning (or nonsense) is accumulated systematically, with the development of each constituent.

On the initial stage of the elaboration of the generative approach, the idea of independent syntax arose and the problem of natural language processing was seen as determining the syntactic structure of each sentence composing a text. Syntactic structure of a sentence was identified with the so-called constituency tree. In other words, this is a nested structure subdividing the sentence into parts, then these parts into smaller parts, and so on. This decomposition corresponds to the sequence of the grammar rules applications that generate the given sentence. For example, the Spanish sentence la estudiante canta una canción has the constituency tree represented graphically in Figure II.1. It also can be represented in the form of the following nested structure marked with square brackets:

Figure II.1. Example of constituency tree.

This structure shows the sentence S consisting of a noun phrase NP and a verb phrase VP, that in its turn consists of a verb V followed by a noun phrase NP, that in its turn consists of a determiner D (an article or pronoun) followed by a noun N that is the word canción, in this case.

Transformational grammars

Further research revealed great generality, mathematical elegance, andwide applicability of generative grammars. They became used not only fordescription of natural languages, but also for specification of formal languages,such as those used in mathematical logic, pattern recognition, and programming languages. A new branch of science called mathematical linguistics  (in its narrow meaning) arose from these studies.

During the next three decades after the rise of mathematical linguistics,much effort was devoted to improve its tools for it to better correspond tofacts of natural languages. At the beginning, this research stemmed from thebasic ideas of Chomsky and was very close to them.

However, it soon became evident that the direct application of simple context-free grammars to the description of natural languages encounters great difficulties.Under the pressure of purely linguistic facts and with the aim to betteraccommodate the formal tools to natural languages, Chomsky proposed the so-called transformational grammars. They were mainly English-oriented and explained how to construct an interrogative or negative sentence from the corresponding affirmative one, how to transform the sentence in active voice to its passive voice equivalent, etc.

For example, an interrogative sentence such as Does John see Mary? does not allow a nested representation as the one shown on page 37 since the two words does and see obviously form a single entity to which the word John does not belong. Chomsky’s proposal for the description of its structure consisted in

(a) description of the structure of some “normal” sentence that does permit the nested representation plus

(b) description of a process of obtaining the sentence in question from such a “normal” sentence by its transformation.

Namely, to construct the interrogative sentence from a “normal” sentence “John sees Mary.”, it is necessary

(1) to replace the period with the question mark (*John sees Mary?),

(2) to transform the personal verb form see into a word combination does see (*John does see Mary?), and finally

(3) to move the word does to the beginning of the sentence (Does John see Mary?), the latter operation leading to the “violation” of the nested structure.

This is shown in the following figure:





does see






Not nested:









A transformational grammar is a set of rules for such insertions, permutations, movements, and corresponding grammatical changes. Such a set of transformational rules functions like a program. It takes as its input a string constructed according to some context-free grammar and produces a transformed string.

The application of transformational grammars to various phenomena of naturallanguages proved to be rather complicated. The theory has lost its mathematicalelegance, though it did not acquire much of additional explanatory capacity.

The linguistic research after Chomsky:

Valencies and interpretation

After the introduction of the Chomskian transformations, many conceptions of language well known in general linguistics still stayed unclear. In the 1980’s,several grammatical theories different from Chomsky’s one were developed withinthe same phrase-structure mainstream. Nearly all of them were based on the cfgs, but used different methods for description of some linguistic facts.

One very important linguistic idea had been suggested already by Chomskyand adopted by the newer theories. It is the subcategorization of verbsaccording to their ability to accept specific sets of complements. Thesecomplements are also called actants, or valency fillers, which we will use interchangeably. We also will informally use the term valency for valency filler, though valency is a link, whereas a valency filler is a linked word or a word group.

The term valency is also used in chemistry, and this is not by accident. The point is that each specific verb has its own standard set of actants (usually nouns). Within a sentence, the actants are usually located close to the predicative verb and are related with it semantically. For example, the Spanish verb dar has three actants reflecting (1) donator (who gives?), (2) donation (what is given?) and (3) receptor (to whom is given?).

In texts, the valencies are given in specific ways depending on the verb,e.g., with a specific word order and/or a specific preposition for eachvalency. All three actants of the verb dar can be seen in the sentence Juan (1) dio muchas flores (2) a Elena (3). The last two actants and their position in a sentence after the verb dar can be expressed with the pattern:

dar <donation> a <receptor>

The description given above reflects linguistic phenomena including bothsyntactic and semantic aspects of valencies. In particular, the names <donator>, <donation>, and <receptor> reflect valencies in their semantic aspect. As to the generative grammar approach, it operates only with constituents and related grammar categories. Under this approach, the pattern called subcategorization frame has the form:

dar N1 a N2,

where N1 and N2 stand for noun phrases, without exposure of their semantic roles. Thus, these phrases are not distinguishable semantically, they only can be given different syntactic interpretations: N1 is a direct complement and N2 is an indirect complement.

We have induced only one of the possible subcategorization frames for theverb dar. To reflect the structure of the sentence Juan (1) dio a Elena (3) muchas flores (2), with the same semantic valencies given in a different order, we are compelled to introduce another pattern:

dar a <receptor> <donation>

with the corresponding subcategorization frame

dar a N1 N2.

Direct and indirect complements swap over, while their semantic rolesstay the same, bit it is not clear in such a subcategorization frame.

Categorization and subcategorization are kinds of classification. Any consistent classification implies separation of entities to several non-intersecting groups.However, in the case under our study, the verb dar should be considered belonging to two different subcategories. Or else two verbs dar should be introduced, with equal meanings and equal semantic valency sets, but with different subcategorization frames. In fact, the situation in languages with the free word order is even more complicated. Indeed, the verb dar can have their donation and receptor actants staying before the subject, like in the sentence A Elena (3) le dio Juan (1) muchas flores (2). Such an order requires even more subcategorization frames obligatorily including the subject of the sentence, i.e. the first valency or the verb, with the role of donator.

The actants are used with verbs more frequently than the so-called circonstants. The circonstants are expressed by adverbs or, similarly to actants, by prepositional groups, i.e., through a combination of a preposition and (usually) a noun. However, the way they are expressed in text does not usually depend on a specific verb. Thus, in the first approximation, the difference between actants and circonstants can be roughly summarized as follows.

· In the syntactic aspect, actants are expressed peculiarly depending on the specific verb, whereas circonstants do not depend on the specific verb in their form, and

· In the semantic aspect, actants are obligatory participants of the situation described by the verb, while circonstants are not.

Only one, obligatory and usually the most important, participant of thesituation is expressed by many languages in a quite standard form, namely thesubject of the sentence. In Spanish, English, and several other languages (butnot in all of them!), it usually precedes the syntactic predicate of the sentence and is represented with a noun without any preposition. Since the subject is expressed in the same manner with all verbs, it is not specified explicitly in the subcategorization frames. However, it is efficient only for languages with strict word order.

As we could see above, the semantic interpretation of different phrases within a sentence cannot be given in the frame of the purely generative approach. It can only distinguish which noun phrase within a sentence is subject, or direct complement, or indirect complement, etc. In deep semantic interpretation (“understanding”), additional theoretical means were needed, and they were first found out of the generative approach. In late 60s, Charles Fillmore [13] has introduced semantic valencies under the name of semantic cases. Each verb has its own set of semantic cases, and the whole set of verbs in any language supposedly has a finite and rather limited inventory of all possible semantic cases. Just among them, we can see semantic cases of donator, donation, and receptor sufficient for interpretation of the verb dar. To “understand” any verb deeper, some rules connecting subcategorization frames and semantic cases had been introduced.

Linguistic research after Chomsky: Constraints

Another very valuable idea originated within the generative approach wasthat of using special features assigned to the constituents, and specifying constraints to characterize agreement or coordination of their grammatical properties. For example, the rule NP ® D N in Spanish and some other languages with morphologic agreement of determinants and the corresponding nouns incorrectly admits constituents like *unas libro. To filter out such incorrect combinations, this rule can be specified in a form similar to an equation:

NP(Gen, Num) ® D(Gen, Num) N(Gen, Num),

where Gen and Num are variables standing for any specific gender and any specific number, correspondingly. The gender Gen and the number Num of a determinant should be the same as (i.e., agree with) the gender Gen and the number Num of the consequent noun (compare Spanish la luna and el sol, but not *unas libro). Such a notation is shorthand of the following more complete and more frequently used notation:

NP  ® D  N .

The following variant is also used, where the same value is not repeated, but is instead assigned a number and then referred to by this number. Thus,  stands for the value Gen and  for the value Num as specified where these numbers first occur:

NP  ® D  N .

The same constraint can be expressed in an even shorter equivalent form, where  stands for the whole combination of the two features as specified when the symbol  first occurs:

NP  ® D N.

Each feature of a constituent can be expressed with a pair: its name andthen its value, e.g., “gender: Gen”. The rule above determines which values areto be equal. With each constituent, any number of features can be expressed,while different constituents within the same rule possibly can have differentsets of specified features.

Another example of the correspondence between features of constituents is the following rule of agreement in person and number between the subject and thepredicate of a Spanish sentence (the syntactic predicate is a verb in finite, i.e., in personal form):

S ® NP(Pers, Num) VP(Pers, Num).

It means that, in the Spanish phrase like yo quiero or ellas quieren, there is a correspondence between the values of grammatical persons Pers and numbers Num: in the first phrase, Pers = 1st, Num = singular, while in the second phrase, Pers = 3rd, Num = plural on the both sides.

The constraint-based approach using the features was intensively used bythe Generalized Phrase Structure Grammar (gpsg), and now is generally accepted. The featured notation permits to diminish the total number of rules. A generalized rule covers several simple rules at once. Such approach paves way to the method of unification to be exposed later.

Head-driven Phrase Structure Grammar

One of the direct followers of the gpsg was called Head-Driven Phrase Structure Grammar (hpsg). In addition to the advanced traits of the gpsg, it has introduced and intensively used the notion of head. In most of the constituents, one of the sub-constituents (called daughters in hpsg) is considered as the principal, or its head (called also head daughter). For example, in the rule:

S ® NP(Pers, Num) hVP(Pers, Num),

the VP constituent (i.e., the syntactic predicate of a sentence with all connected words) is marked as the head of the whole sentence, which is indicated by the symbol h. Another example: in the rule:

NP (Gen, Num) ® D (Gen, Num) hN (Gen, Num),

the noun is marked as the head of the whole noun phrase NP.

According to one of the special principles introduced in hpsg, namely the head principle, the main features of the head are inherited in some way by the mother (enclosing) constituent (the left-side part of the rule).

In the previous examples, the features of the predicate determinefeatures of the whole sentence, and the features of the noun determine thecorresponding features of the whole noun phrase. Such formalism permits toeasier specify the syntactic structure of sentences and thus facilitatessyntactic analysis (parsing).

As it was already said, the interpretation in early generative grammarswas always of syntactic nature. For semantic interpretation (“understanding”), additional theoretical means were introduced, which were somewhat alien to the earlier generative structure mainstream. By contrast, each word in the hpsg dictionary is supplied with semantic information that permits to combine meanings of separate words into a joint coherent semantic structure. The novel rules of the word combining gave a more adequate method of construing the semantic networks. Meantime, Chomskian idea of transformations was definitely abandoned by this approach.

                                The idea of unification                               

Having in essence the same initial idea of phrase structures and their context-free combining, the hpsg and several other new approaches within Chomskian mainstream select the general and very powerful mathematical conception of unification. The purpose of unification is to make easier the syntactic analysis ofnatural languages.

The unification algorithms are not linguistic proper. Rather they detectsimilarities between parts of mathematical structures (strings, trees, graphs,logical formulas) labeled with feature sets. A priori, it is known that somefeatures are interrelated, i.e., they can be equal, or one of them covers the other.Thus, some feature combinations are considered compatible while met inanalysis, whereas the rest are not. Two sets of features can be unified, ifthey are compatible. Then the information at an object admitting unification(i.e., at a constituent within a sentence to be parsed) combines the informationbrought by both sets of features.

Unification allows filtering out inappropriate feature options, while theunified feature combination characterizes the syntactic structure underanalysis more precisely, leading to the true interpretation of the sentence.

As the first example of unification operations, let us compare featuresets of two Spanish words, el and muchacho, staying in a text side by side. Both words have the feature set [gender = masculine, number = singular], so that they are equivalent with respect to gender and number. Hence, the condition of unification is satisfied, and this pair of words can form a unifying constituent in syntactic analysis.

Another example is the adjacent Spanish words las estudiantes. The article las has the feature set [gender = feminine, number = plural]. As to the string estudiantes, this word can refer to both ‘he-student’ of masculine gender and ‘she-student’ of feminine gender, so that this word is not specified (is underspecified) with respect to gender. Thus, the word occurrence estudiantes taken separately has a broader feature set, namely, [number = plural], without any explicit indication of gender. Since these two feature sets are not contradictory, they are compatible and their unification [gender = feminine, number = plural] gives the unifying constraint set assigned to both words. Hence, this pair can form a unifying mother constituent las estudiantes, which inherits the feature set from the head daughter estudiantes. The gender of the particular word occurrence estudiantes is feminine, i.e., ‘she-students,’ and consequently the inherited gender of the noun phrase las estudiantes is also feminine.

As the third example, let us consider the words niño and quisiera in the Spanish sentence El niño quisiera pasar de año. The noun niño is labeled with the 3rd person value: [person = 3], whereas the verb quisiera exists in two variants labeled with the feature set [person = 1 or person = 3], correspondingly. Only the latter variant of the verb can be unified with the word niño. Therefore this particular word occurrence of quisiera is of the third person. The whole sentence inherits this value, since the verb is its head daughter.

The Meaning Û Text Theory:

Multistage transformer and government patterns

The European linguists went their own way, sometimes pointing out someoversimplifications and inadequacies of the early Chomskian linguistics.

In late 1960´s, a new theory, the Meaning Û Text model of natural languages, was suggested in Russia. For more than 30 years, this theory has been developed by the scientific teams headed by I. Mel’čuk, in Russia and then in Canada, and by the team headed by Yu. Apresian in Russia, as well as by other researchers in various countries. In the framework of the Meaning Û Text Theory (mtt), deep and consistent descriptions of several languages of different families, Russian, French, English and German among them, were constructed and introduced to computational practice.

One very important feature of the mtt is considering the language as multistage, or multilevel, transformer of meaning to text and vice versa. The transformations are comprehended in a different way from the theory by Chomsky. Some inner representation corresponds to each level, and each representation is equivalent to representations of other levels. Namely, surface morphologic, deep morphologic, surface syntactic, deep syntactic, and semantic levels, as well as the corresponding representations, were introduced into the model.

The description of valencies for words of any part of speech and of correspondence between the semantic and syntactic valencies have found their adequate solution in this theory, in terms of the so-called government patterns.

The government patterns were introduced not only for verbs, but also forother parts of speech. For a verb, gp has the shape of a table of all its possible valency representations. The table is preceded by the formula of the semantic interpretation of the situation reflected by the verb with all its valencies. The table is succeeded by information of word order of the verb and its actants.

If to ignore complexities implied by Spanish pronominal clitics like me, te, se, nos, etc., the government pattern of the Spanish verb dar can be represented as

Person X gives thing Y to person Z

X = 1

Y = 2

Z = 3

1.1 N

2.1 N

3.1 a N

The symbols X, Y, and Z designate semantic valencies, while 1, 2, and 3 designate the syntactic valencies of the verb. Meaning ‘give’ in the semantic formula is considered just corresponding to the Spanish verb dar, since dar supposedly cannot be represented by the more simple semantic elements.

The upper line of the table settles correspondences between semantic andsyntactic valencies. For this verb, the correspondence is quite simple, but itis not so in general case.

The lower part of the table enumerates all possible options of representationfor each syntactic valency at the syntactic level. The options operate with part-of-speechlabels (N for a noun, Vinf for a verb in infinitive, etc.) and prepositions connecting the verb with given valency fillers. In our simplistic example, only nouns can fill all three valencies, only the preposition a is used, and each valency have the unique representation. However, such options can be multiple for other verbs in Spanish and various other languages. For example, the English verb give has two possible syntactic options for the third valency: without preposition (John gives him a book) vs. with the preposition to (John gives a book to him).

The word order is generally not related with the numeration of the syntacticvalencies in a government pattern. If all permutations of valencies of aspecific verb are permitted in the language, then no information about wordorder is needed in this GP. Elsewhere, information about forbidden or permittedcombinations is given explicitly, to make easier the syntactic analysis. Forexample, the English verb give permits only two word orders mentioned above.

Thus, government patterns are all-sufficient for language description andsignificantly differ from subcategorization frames introduced in the generative grammar mainstream.

The Meaning Û Text Theory: Dependency trees

Another important feature of the mtt is the use of its dependency trees, for description of syntactic links between words in a sentence. Just the set of these links forms the representation of a sentence at the syntactic level within this approach.

Figure II.2. Example of a dependency tree.

For example, the Spanish sentence La estudiante mexicana canta unacanción can be represented by the dependency tree shown in Figure II.2. One can see that the dependency tree significantly differs from the constituency tree for the same sentence (cf. Figure II.1).

Up to the present, the proper description of the word order and wordagreement in many languages including Spanish can be accomplished easier bymeans of the mtt. Moreover, it was shown that in many languages there exist disrupt and non-projective constructions, which cannot be represented through constituency trees or nested structures, but dependency trees can represent them easily.

In fact, dependency trees appeared as an object of linguistic research inthe works of Lucien Tesnière, in 1950’s. Even earlier, dependencies between words were informally used in descriptions of various languages, including Spanish. However, just the mtt has given strict definition to dependency trees. The dependency links were classified for surface and deep syntactic levels separately. They were also theoretically isolated from links of morphologic inter-word agreement so important for Spanish.

With dependency trees, descriptions of the relationships between thewords constituting a sentence and of the order of these words in the sentencewere separated from each other. Thus, the links between words and the order inwhich they appear in a sentence were proposed to be investigated apart, andrelevant problems of both analysis and synthesis are solved now separately.

Hence, the mtt in its syntactic aspect can be called dependency approach, as contrasted to the constituency approach overviewed above. In the dependency approach, there is no problem for representing the structure of English interrogative sentences (cf. page 39). Thus, there is no necessity in the transformations of Chomskian type.

To barely characterize the mtt as a kind of dependency approach is to extremely simplify the whole picture. Nevertheless, this book presents the information permitting to conceive other aspects of the mtt.

           The Meaning Û Text Theory:Semantic links           

The dependency approach is not exclusively syntactic. The links betweenwordforms at the surface syntactic level determine links between correspondinglabeled nodes at the deep syntactic level, and after some deletions,insertions, and inversions imply links in the semantic representation of the same sentence or a set of sentences. Hence, this approach facilitates the transfer from syntactic representations to a semantic one and vice versa.

According to the mtt, the correlation between syntactic and semantic links is not always straightforward. For example, some auxiliary words in a sentence (e.g., auxiliary verbs and some prepositions) are treated as surface elements and disappear at the deep syntactic level. For example, the auxiliary Spanish verb HABER in the word combination han pedido disappears from the semantic representation after having been used to determine the verb tense and mode. At the same time, some elements absent in the surface representation are deduced, or restored, from the context and thus appear explicitly at the deep syntactic level. For example, given the surface syntactic dependency tree fragment:

su ¬ hijo ® Juan,

the semantically conditioned element NAME is inserted at the deep syntacticlevel, directly ruling the personal name:

su ¬ hijo ® NAME ® Juan

Special rules of inter-level correspondence facilitate the transition tothe correct semantic representation of the same fragment.

The mtt provides also the rules of transformation of some words and word combinations to other words and combinations, with the full preservation of the meaning. For example, the Spanish sentence Juan me prestó ayuda can be formally transformed to Juan me ayudó and vice versa at the deep syntactic level. Such transformations are independent of those possible on the semantic level, where mathematical logic additionally gives quite other rules of meaning-preserving operations.

We should clarify that some terms, e.g., deep structure or transformation, are by accident used in both the generative and the mtt tradition, but in completely different meanings. Later we will return to this source of confusion.

All these features will be explained in detail later. Now it is importantfor us only to claim that the mtt has been able to describe any natural language and any linguistic level in it.


In the twentieth century, syntax was in the center of the linguistic research, and the approach to syntactic issues determined the structure of any linguistic theory. There are two major approaches to syntax: the constituency, or phrase-structure, approach, and the dependency approach.The constituency tradition was originated by N. Chomsky with the introduction of the context-free grammars, and the most recent development in this tradition is Head-driven Phrase Structure Grammar theory. The dependency approach is used in the Meaning Û Text Theory by Igor Mel’čuk. Both approaches are applicable for describing linguistic phenomena in many languages.

III.  Products of Computational Linguistics:

Present and Prospective

For what purposes do we need to develop computational linguistics? What practical results does it provide for society? Before we start discus-sing the methods and techniques of computational linguistics, it is worthwhile giving a review of some existing practical results, i.e., applications, or products, of this discipline. We consider such applications in a very broad sense, including in this category all known tasks of word processing, as well as those of text processing, text generation, dialogue in a natural language, and language understanding.

Some of these applications already provide the user with satisfactorysolutions for their tasks, especially for English, while other tasks and languageshave been under continuous research in recent decades.

Of course, some extrapolations of the current trends could give completely new types of systems and new solutions to the current problems, but this is out of scope of this book.

Classification of applied linguistic systems

Applied linguistic systems are now widely used in business and scientificdomains for many purposes. Some of the most important ones among them are thefollowing:

· Text preparation, or text editing, in a broad sense, particularly including the tasks listed below:

– Automatic hyphenation of words in natural language texts,

– Spell checking, i.e., detection and correction of typographic and spelling errors,

– Grammar checking, i. e., detection and correction of grammatical errors,

– Style checking, i. e. detection and correction of stylistic errors,

– Referencing specific words, word combinations, and semantic links between them;

· Information retrieval in scientific, technical, and business document databases;

· Automatic translation from one natural language to another;

· Natural language interfaces to databases and other systems;

· Extraction of factual data from business or scientific texts;

· Text generation from pictures and formal specifications;

· Natural language understanding;

· Optical character recognition, speech recognition, etc.

For the purposes of this book, we will give here only a short sketch ofeach application. Later, some of these topics, with more deep explanations, canbe touched upon once more.

Automatic hyphenation

Hyphenation is intended for the proper splitting of words in natural languagetexts. When a word occurring at the end of a line is too long to fit on thatline within the accepted margins, a part of it is moved to the next line. Theword is thus wrapped, i.e., split and partially transferred to the next line.

The wrapping can be done only at specific positions within words, whichgenerally, though not always, are syllable boundaries. For example, in Spanishone can split re-ci-bo, re-u-nir-se, dia-blo, ca-rre-te-ra, mu-cha-chas, but not in the following positions: *recib-o, *di-ablo, *car-retera, *muc-hac-has.

In this way, hyphenation improves the outer appearance of computer-producedtexts through adjusting their right margins. It saves paper and at the sametime preserves impression of smooth reading, just as without any hyphenation.

The majority of the well-known text editors are supplied now withhyphenation tools. For example, Microsoft Word has the menu item Hyphenation.[2]

Usually, the linguistic information taken for such programs is ratherlimited. It should be known which letters are vowels (a, e, i, o, u in Spanish) or consonants (b, c, d, f, g, etc.), and what letter combinations are inseparable (such as consonants pairs ll, rr, ch or diphthongs io, ue, ai in Spanish).

However, the best quality of hyphenation could require more detailedinformation about each word. The hyphenation can depend on the so-calledmorphemic structure of the word, for example: sub-ur-ba-no, but su-bir, or even on the origin of the word, for example: Pe-llicer, but Shil-ler. Only a dictionary-based program can take into account all such considerations. For English, just dictionary-based programs really give perfect results, while for Spanish rather simple programs are usually sufficient, if to neglect potentially error-prone foreign words like Shiller.

Spell checking

The objective of spell checking is the detection and correction oftypographic and orthographic errors in the text at the level of word occurrence considered out of its context.

Nobody can write without any errors. Even people well acquainted with therules of language can, just by accident, press a wrong key on the keyboard(maybe adjacent to the correct one) or miss out a letter. Additionally, whentyping, one sometimes does not synchronize properly the movements of the handsand fingers. All such errors are called typos, or typographic errors. On the other hand, some people do not know the correct spelling of some words, especially in a foreign language. Such errors are called spelling errors.

First, a spell checker merely detects the strings that are not correctwords in a given natural language. It is supposed that most of the orthographicor typographic errors lead to strings that are impossible as separate words in this language. Detecting the errors that convert by accident one word into another existing word, such as English then ® ?than or Spanish cazar ® ?casar, supposes a task which requires much more powerful tools.

After such impossible string has been detected and highlighted by the program,the user can correct this string in any preferable way—manually or with thehelp of the program. For example, if we try to insert into any English text thestrings[3] *groop, *greit, or *misanderstand, the spell checker will detect the error and stop at this string, highlighting it for the user. Analogous examples in Spanish can be *caió, *systema, *nesecitar.

The functions of a spell checker can be more versatile. The program canalso propose a set of existing words, which are similar enough (in some sense) to the given corrupted word, and the user can then choose one of them as the correct version of the word, without re-typing it in the line. In the previous examples, Microsoft Word’s spell checker gives, as possible candidates for replacement of the string caió, the existing Spanish words shown in Figure III.1.

In most cases, especially for long strings, a spell checker offers onlyone or two candidates (or none). For example, for the string *systema it offers only the correct Spanish word sistema.

The programs that perform operations of both kinds are called orthographiccorrectors, while in English they are usually called spell checkers. Ineveryday practice, spell checkers are considered very helpful and are used bymillions of users throughout the world. The majority of modern text editors aresupplied now with integrated spell checkers. For example, Microsoft Word usesmany spell checkers, a specific one for each natural language used in the text.

The amount of linguistic information necessary for spell checkers is muchgreater than for hyphenation. A simple but very resource-consuming approachoperates with a list, or a dictionary, of all valid words in a specific language.It is necessary to have also a criterion of similarity of words, and somepresuppositions about the most common typographic and spelling errors. A deeperpenetration into the correction problems requires a detailed knowledge ofmorphology, since it facilitates the creation of a more compact dictionary that has a manageable size.

Figure III.1. Alternatives for the word *caió.

Spell checkers have been available for more than 20 years, but some quiteevident tasks of correction of words, even taken separately, have not been yetsolved. To put a specific example, let us consider the ungrammatical string *teached in an English text. None of the spell checkers we have tried suggested the correct form taught. In an analogous way, if a foreigner inserts into a Spanish text such strings as *muestrar or *disponido, the Spanish spell checkers we have tried did not give the forms mostrar and dispuesto as possible corrections.

Grammar checking

Detection and correction of grammatical errors by taking into account adjacentwords in the sentence or even the whole sentence are much more difficult tasksfor computational linguists and software developers than just checking orthography.

Grammar errors are those violating, for example, the syntactic laws orthe laws related to the structure of a sentence. In Spanish, one of these lawsis the agreement between a noun and an adjective in gender and grammaticalnumber. For example, in the combination *mujer viejos each word by itself does exist in Spanish, but together they form a syntactically ill-formed combination. Another example of a syntactic agreement is the agreement between the noun in the role of subject and the main verb, in number and person (*tú tiene).

The words that must agree can be located in quite different parts of thesentence. For example, it is rather difficult for a program to find the errorin the following sentence: *Las mesas de madera son muy largos.

Other types of grammatical errors include incorrect usage of prepositions,like in the phrases *debajo la puerta, or *¡basta con verla!, or *casarse a María. Some types of syntactic errors may be not so evident even for a native speaker.

It became clear long ago that only a complete syntactic analysis (parsing)of a text could provide an acceptable solution of this task. Because of thedifficulty of such parsing, commercial grammar checkers are still ratherprimitive and rarely give the user useful assistance in the preparation of atext. The Windows Sources, one of the well-known computer journals, noted, in May 1995, that the grammar checker Grammatik in the WordPerfect text editor, perhaps the best grammar checker in the world at that time, was so imperfect and disorienting, that “nobody needs a program that’s wrong in more cases than it’s right.”

In the last few years, significant improvements have been made in grammarcheckers. For example, the grammar checker included in Microsoft Word is helpfulbut still very far from perfection.

Sometimes, rather simple operations can give helpful results by detectingsome very frequent errors. The following two classes of errors specific forSpanish language can be mentioned here:

· Absence of agreement between an article and the succeeding noun, in number and gender, like in *la gatos. Such errors are easily detectable within a very narrow context, i.e., of two adjacent words. For this task, it is necessary to resort to the grammatical categories for Spanish words.

· Omission of the written accent in such nouns as *articulo, *genero, *termino. Such errors cannot be detected by a usual spell checker taking the words out of context, since they convert one existing word to another existent one, namely, to a personal form of a verb. It is rather easy to define some properties of immediate contexts for nouns that never occur with the corresponding verbs, e.g., the presence of agreed articles, adjectives, or pronouns [38].

We can see, however, that such simplistic techniques fail in too many cases.For example, in combinations such as *las pruebas de evaluación numerosos, the disagreement between pruebas and numerosos cannot be detected by considering only the nearest context.

What is worse, a program based on such a simplistic approach would toofrequently give false alarms where there is no error in fact. For example, inthe correct combination las pruebas de evaluación numerosas, such a simplistic program would mention disagreement in number between the wordforms evaluación and numerosas.

In any case, since the author of the text is the only person thatdefinitely knows what he or she meant to write, the final decision must alwaysbe left up to the user, whether to make a correction suggested by the grammarchecker or to leave the text as it was.

Style checking

The stylistic errors are those violating the laws of use of correct wordsand word combinations in language, in general or in a given literary genre.

This application is the nearest in its tasks to normative grammars andmanuals on stylistics in the printed, oriented to humans, form. Thus, stylecheckers play a didactic and prescriptive role for authors of texts.

For example, you are not recommended to use any vulgar words or purelycolloquial constructions in official documents. As to more formal properties ofSpanish texts, their sentences should not normally contain ten prepositions de, and should not be longer than, let us say, twenty lines. With respect to Spanish lexicon, it is not recommended to use the English words parking and lobby instead of estacionamiento and vestíbulo, or to use the Americanism salvar in the meaning ‘to save in memory’ instead of guardar.

In the Spanish sentence La recolección de datos en tiempo real es realizada mediante un servidor, the words in boldface contain two stylistic anomalies: se realiza is usually better than es realizada, and such a close neighborhood of words with the same stem, like real and realizada, is unwanted.

In the Spanish sentence La grabación, reproducción y simulación de datos son funciones en todos los sistemas de manipulación de información, the frequency of words with the suffix -ción oversteps limits of a good style.

The style checker should use a dictionary of words supplied with theirusage marks, synonyms, information on proper use of prepositions, compatibilitywith other words, etc. It should also use automatic parsing, which can detectimproper syntactic constructions.

There exist style checkers for English and some other major languages,but mainly in laboratory versions. Meanwhile commercial style checkers areusually rather primitive in their functions.

As a very primitive way to assess stylistic properties of a text, somecommercial style checkers calculate the average length of words in the text,i.e., the number of letters in them; length of sentences, i.e., the number ofwords in them; length of paragraphs, i.e., the number of words and sentences.They can also use other statistical characteristics that can be easily calculatedas a combination of those mentioned.

The larger the average length of a word, sentence or paragraph, the moredifficult the text is to read, according to those simplest stylisticassessments. It is easy also to count the occurrences of prepositions de or nouns ending in -ción in Spanish sentences.

Such style checkers can only tell the user that the text is too complicated(awkward) for the chosen genre, but usually cannot give any specific suggestionsas to how to improve the text.

The assessment of deeper and more interesting stylistic properties, connected with the lexicon and the syntactic constructions, is still considered a task for thefuture.

References to words and wordcombinations

The references from any specific word give access to the set of wordssemantically related to the former, or to words, which can form combinationswith the former in a text. This is a very important application. Nowadays it isperformed with linguistic tools of two different kinds: autonomous on-linedictionaries and built-in dictionaries of synonyms.

Within typical text processors, the synonymy dictionaries are usuallycalled thesauri. Later we will see that this name corresponds poorly to thesynonymy dictionaries, since genuine thesauri usually include much moreinformation, for example, references to generic words, i.e., names of superclasses,and to specific words, i.e., names of subclasses.

References to various words or word combinations of a given naturallanguage have the objective to help the author of a text to create morecorrect, flexible, and idiomatic texts. Indeed, only an insignificant part ofall thinkable word combinations are really permitted in a language, so that theknowledge of the permitted and common combinations is a very important part oflinguistic competence of any author. For example, a foreigner might want toknow all the verbs commonly used with the Spanish noun ayuda, such as prestar or pedir, or with the noun atención, such as dedicar or prestar, in order to avoid combinations like pagar atención, which is a word-by-word translation of the English combination to pay attention. Special language-dependent dictionaries are necessary for this purpose (see, for example, Figure III.2).

Figure III.2. CrossLexica™, a dictionary of word combinations.

Within such systems, various complex operations are needed, such as automatedreduction of the entered words to their dictionary forms, search of relevantwords in the corresponding linguistic database, and displaying all of them in aform convenient to a non-linguist user. These operations are versatile andinclude both morphologic and syntactic issues [37].

Another example of a dictionary that provides a number of semanticrelations between different lexemes is EuroWordNet [55], a huge lexical resource reflecting diverse semantic links between lexemes of several European languages.

The ideological basis of EuroWordNet is the English dictionary WordNet [41]. English nouns, verbs, adjectives, and adverbs were divided into synonymy groups, or synsets. Several semantic relations were established between synsets: antonymy (reference to the “opposite” meaning), hyponymy (references to the subclasses), hyperonymy (reference to the superclass), meronymy (references to the parts), holonymy (reference to the whole), etc. Semantic links were established also between synsets of different parts of speech.

The classification hierarchy for nouns is especially well developedwithin WordNet. The number of hierarchical levels is in average 6 to 7,sometimes reaching 15. The upper levels of the hierarchy form the ontology, i.e., a presupposed scheme of human knowledge.

In essence, EuroWordNet is a transportation of the WordNet hierarchy toseveral other European languages, in particular to Spanish. The upper levels ofontology were obtained by direct translation from English, while for the otherlevels, additional lexicographic research turned out to be necessary. In thisway, not only links between synsets within any involved language weredetermined, but also links between synsets of a number of different languages.

The efforts invested to the WordNet and EuroWordNet were tremendous.Approximately 25´000 words were elaborated in several languages.

Information retrieval

Information retrieval systems (irs) are designed to search for relevant information in large documentary databases. This information can be of various kinds, with the queries ranging from “Find all the documents containing the word conjugar” to “Find information on the conjugation of Spanish verbs”. Accordingly, various systems use different methods of search.

The earliest irss were developed to search for scientific articles on a specific topic. Usually, the scientists supply their papers with a set of keywords, i.e., the terms they consider most important and relevant for the topic of the paper. For example, español, verbos, subjuntivo might be the keyword set of the article “On means of expressing unreal conditions” in a Spanish scientific journal.

These sets of keywords are attached to the document in the bibliographicdatabase of the irs, being physically kept together with the corresponding documents or separately from them. In the simplest case, the query should explicitly contain one or more of such keywords as the condition on what the article can be found and retrieved from the database. Here is an example of a query: “Find the documents on verbos and español”. In a more elaborate system, a query can be a longer logical expression with the operators and, or, not, e.g.: “Find the documents on (sustantivos or adjetivos) and (not inglés)”.

Nowadays, a simple but powerful approach to the format of the query isbecoming popular in irss for non-professional users: the query is still a set of words; the system first tries to find the documents containing all of these words, then all but one, etc., and finally those containing only one of the words. Thus, the set of keywords is considered in a step-by-step transition from conjunction to disjunction of their occurrences. The results are ordered by degree of relevance, which can be measured by the number of relevant keywords found in the document. The documents containing more keywords are presented to the user first.

In some systems the user can manually set a threshold for the number ofthe keywords present in the documents, i.e., to search for “at least m of n” keywords. With m = n, often too few documents, if any, are retrieved and many relevant documents are not found; with m = 1, too many unrelated ones are retrieved because of a high rate of false alarms.

Usually, recall and precision are considered the main characteristics of irss. Recall is the ratio of the number of relevant documents found divided by the total number of relevant documents in the database. Precision is the ratio of the number of relevant documents divided by the total number of documents found.

It is easy to see that these characteristics are contradictory in the generalcase, i.e. the greater one of them the lesser another, so that it is necessaryto keep a proper balance between them.

In a specialized irs, there usually exists an automated indexing subsystem, which works before the searches are executed. Given a set of keywords, it adds, using the or operator, other related keywords, based on a hierarchical system of the scientific, technical or business terms. This kind of hierarchical systems is usually called thesaurus in the literature on irss and it can be an integral part of the irs. For instance, given the query “Find the documents on conjugación,” such a system could add the word morfología to both the query and the set of keywords in the example above, and hence find the requested article in this way.

Thus, a sufficiently sophisticated irs first enriches the sets of keywords given in the query, and then compares this set with the previously enriched sets of keywords attached to each document in the database. Such comparison is performed according to any criteria mentioned above. After the enrichment, the average recall of the irs system is usually increased.

Recently, systems have been created that can automatically build sets ofkeywords given just the full text of the document. Such systems do not requirethe authors of the documents to specifically provide the keywords. Some of themodern Internet search engines are essentially based on this idea.

Three decades ago, the problem of automatic extraction of keywords wascalled automatic abstracting. The problem is not simple, even when it is solved by purely statistical methods. Indeed, the most frequent words in any business, scientific or technical texts are purely auxiliary, like prepositions or auxiliary verbs. They do not reflect the essence of the text and are not usually taken for abstracting. However, the border between auxiliary and meaningful words cannot be strictly defined. Moreover, there exist many term-forming words like system, device, etc., which can seldom be used for information retrieval because their meaning is too general. Therefore, they are not useful for abstracts.

The multiplicity of irss is considered now as an important class of the applied software and, specifically, of applied linguistic systems. The period when they used only individual words as keys has passed. Developers now try to use word combinations and phrases, as well as more complicated strategies of search. The limiting factors for the more sophisticated techniques turned out to be the same as those for grammar and style checkers: the absence of complete grammatical and semantic analysis of the text of documents. The methods used now even in the most sophisticated Internet search engines are not efficient for accurate information retrieval. This leads to a high level of information noise, i.e., delivering of irrelevant documents, as well as to the frequent missing of relevant ones.

The results of retrieval operations directly depend on the quality andperformance of the indexing and comparing subsystems, on the content of the terminological system or the thesaurus, and other data and knowledge used by thesystem. Obviously, the main tools and data sets used by an irs have the linguistic nature.

Topical summarization

In many cases, it is necessary to automatically determine what a givendocument is about. This information is used to classify the documents by theirmain topics, to deliver by Internet the documents on a specific subject to theusers, to automatically index the documents in an irs, to quickly orient people in a large set of documents,and for other purposes.

Such a task can be viewed as a special kind of summarization: to convey the contents of the document in a shorter form. While in “normal” summarization by the contents the main ideas of the document are considered, here we consider only the topics mentionedin the document, hence the term topical summarization.

Figure III.3. Classifier program determines the main topics of a document.

As an example, let us consider the system Clasitex™ that automatically determines the main topics of a document. A variant of its implementation, Classifier™, was developed in the Center of Computing Research, National Polytechnic Institute at Mexico City [46] (see Figure III.3). It uses two kinds of linguistic information:

· First, it neutralizes morphologic variations in order to reduce any word found in the text to its standard (i.e., dictionary) form, e.g., oraciones ® oración, regímenes ® régimen, lingüísticas ® lingüístico, propuesto ® proponer.

· Second, it puts into action a large dictionary of thesaurus type, which gives, for each word in its standard form, its corresponding position in a pre-defined hierarchy of topics. For example, the word oración belongs to the topic lingüística, which belongs in turn to the topic ciencias sociales, which in its turn belongs to the topic ciencia.

Then the program counts how many times each one of these topics occurredin the document. Roughly speaking, the topic mentioned most frequently is considered the main topic of the document. Actually, the topics in the dictionary have differentweights of importance [43, 45], so that the main topic is the one with the greatesttotal weight in the document.

Applied linguistics can improve this method in many possible ways. Forexample, in its current version, Clasitex does not count any pronouns found in the text, since it is not obvious what object a personal pronoun such as él can refer to.

What is more, many Spanish sentences contain zero subjects, i.e. implicitreferences to some nouns. This becomes obvious in English translation: Hay un libro. Es muy interesante Þ There is a book. It is very interesting Þ El libro es muy interesante. Thus, each Spanish sentence without any subject is implicitly an occurrence of the corresponding word, which is not taken into account by Clasitex, so that the gathered statictics is not completely correct.

Another system, TextAnalystÔ, for determining the main topics of the document and the relationships between words in the document was developed by MicroSystems, in Russia (see Figure III.4). This system is not dictionary-based, though it does have a small dictionary of stop-words (these are prepositions, articles, etc., and they should not be processed as meaningful words).

This system reveals the relationships between words. Words are considered related to each other if they co-occurred closely enough in the text, e.g., in thesame sentence. The program builds a network of the relationships between words.Figure III.4 shows the most important words found by TextAnalyst in theearly draft of this book, and the network of their relationships.

As in Clasitex, the degree of importance of a word, or its weight, isdetermined in terms of its frequency, and the relationships between words areused to mutually increase the weights. The words closely related to many of theimportant words in the text are also considered important.

Figure III.4. TextAnalyst program reveals the relationships between words.

In TextAnalyst, the list of the important words is used for the followingtasks:

· Compression of text by eliminating the sentences or paragraphs that contain the minimal number of important words, until the size of the text reaches the threshold selected by the user,

· Building hypertext by constructing mutual references between the most important words and from the important words to others to which they are supposedly related.

The TextAnalyst technology is based on a special type of a dynamic neuralnetwork algorithm. Since the Clasitex program is based on a large dictionary,it is a knowledge-based program, whereas TextAnalyst is not.

Automatic translation

Translation from one natural language to another is a very importanttask. The amount of business and scientific texts in the world is growingrapidly, and many countries are very productive in scientific and businessdomains, publishing numerous books and articles in their own languages. Withthe growth of international contacts and collaboration, the need fortranslation of legal contracts, technical documentation, instructions, advertisements,and other texts used in the everyday life of millions of people has become amatter of vital importance.

The first programs for automatic, or machine, translation were developedmore than 40 years ago. At first, there existed a hope that texts could be translatedword by word, so that the only problem would be to create a dictionary of pairsof words: a word in one language and its equivalent in the other. However, thathope died just after the very first experiments.

Then the ambitious goal was formulated to create programs which couldunderstand deeply the meaning of an arbitrary text in the source language,record it in some universal intermediate language, and then reformulate thismeaning in the target language with the greatest possible accuracy. It wassupposed that neither manual pre-editing of the source text nor manualpost-editing of the target text would be necessary. This goal proved to betremendously difficult to achieve, and has still not been satisfactorilyaccomplished in any but the narrowest special cases.

At present there is a lot of translation software, ranging from verylarge international projects being developed by several institutes or evenseveral corporations in close cooperation, to simple automatic dictionaries,and from laboratory experiments to commercial products. However, the quality ofthe translations, even for large systems developed by the best scientists, isusually conspicuously lower than the quality of manual human translation.

As for commercial translation software, the quality of translation itgenerates is still rather low. A commercial translator can be used to allowpeople quite unfamiliar with the original language of the document tounderstand its main idea. Such programs can help in manual translation oftexts. However, post-editing of the results, to bring them to the degree ofquality sufficient for publication, often takes more time than just manual translationmade by a person who knows both languages well enough.[4] Commercial translators are quite good for the texts of very specific, narrow genres, such as weather reports. They are also acceptable for translation of legal contracts, at least for their formal parts, but the paragraphs specifying the very subject of the contract may be somewhat distorted.

To give the reader an impression of what kind of errors a translationprogram can make, it is enough to mention a well-known example of themistranslation performed by one of the earliest systems in 1960s. It translatedthe text from Bible The spirit is willing, but the flesh is weak (Matt. 26:41) into Russian and then back into English. The English sentence then turned out to be The vodka is strong, but the meat is rotten [34]. Even today, audiences at lectures on automatic translation are entertained by similar examples from modern translation systems.

Two other examples are from our own experience with the popularcommercial translation package PowerTranslator by Globalink, one of the best in the market. The header of an English document Plans is translated into Spanish as the verb Planifica, while the correct translation is the Spanish noun Planes (see Figure III.5). The Spanish phrase el papel de Francia en la guerra is translated as the paper of France in the war, while the correct translation is the role of France in the war. There are thousands of such examples, so that nearly any automatically translated document is full of them and should be reedited.

Figure III.5. One of commercial translators.

Actually, the quality of translation made by a given program is not thesame in the two directions, say, from English to Spanish and from Spanish toEnglish. Since automatic analysis of the text is usually a more difficult taskthan generation of text, the translation from a language that is studied and described better has generally higher quality than translation into this language. Thus, the elaboration of Spanish grammars and dictionaries can improve the quality of the translation from Spanish into English.

One difficult problem in automatic translation is the word sensedisambiguation. In any bilingual dictionary, for many source words, dozens ofwords in the target language are listed as translations, e.g., for simpleSpanish word gato: cat, moneybag, jack, sneak thief, trigger, outdoor market, hot-water bottle, blunder, etc. Which one should the program choose in any specific case? This problem has proven to be extremely difficult to solve. Deep linguistic analysis of the given text is necessary to make the correct choice, on the base on the meaning of the surrounding words, the text, as a whole, and perhaps some extralinguistic information [42].

Another, often more difficult problem of automatic translation is restoringthe information that is contained in the source text implicitly, but which mustbe expressed explicitly in the target text. For example, given the Spanish textJosé le dio a María un libro. Es interesante, which translation of the second sentence is correct: He is interesting, or She is interesting, or It is interesting, or This is interesting? Given the English phrase computer shop, which Spanish translation is correct: tienda de computadora or tienda de computadoras? Compare this with computer memory. Is they are beautiful translated as son hermosos or son hermosas? Is as you wish translated as como quiere, como quieres, como quieren, or como queréis?[5] Again, deep linguistic analysis and knowledge, rather than simple word-by-word translation, is necessary to solve such problems.

Great effort is devoted in the world to improve the quality oftranslation. As an example of successful research, the results of the Translationgroup of Information Science Institute at University of South California can bementioned [53]. This research is based on the use of statistical techniques forlexical ambiguity resolution.

Another successful team working on automatic translation is that headedby Yu. Apresian in Russia [34]. Their research is conducted in the framework of the Meaning Û Text model.

Natural language interface

The task performed by a natural language interface to a database is tounderstand questions entered by a user in natural language and to provideanswers—usually in natural language, but sometimes as a formatted output.Typically, the entered queries, or questions, concern some facts about data contained in a database.

Since each database is to some degree specialized, the language of thequeries and the set of words used in them are usually very limited. Hence, thelinguistic task of grammatical and semantic analysis is much simpler than for other tasks related to natural language, such as translation.

There are some quite successful systems with natural language interfacesthat are able to understand a very specialized sublanguage quite well. Other systems, with other, usually less specialized sublanguages, are much less successful. Therefore, this problem does not have, at least thus far, a universal solution, most of the solutions being constructed ad hoc for each specific system.

The developers of the most popular database management systems usuallysupply their product with a formal query-constructing language, such as sql. To learn such a language is not too difficult, and this diminishes the need for a natural language interface. We are not aware of any existing commercial interface system that works with a truly unlimited natural language.

Nevertheless, the task of creating such an interface seems very attractivefor many research teams all over the world. Especially useful could be naturallanguage interfaces with speech recognition capabilities, which also wouldallow the user to make queries or give commands over a telephone line.

The task of development of natural language interfaces, though being lessdemanding to such branches of linguistics as morphology or syntax, are very demanding to such “deeper” branches of linguistics as semantics, pragmatics, and theory of discourse.

The specific problem of the interface systems is that they work not witha narrative, a monologue, but with a dialogue, a set of short, incomplete,interleaving remarks. For example, in the following dialogue:

User:      Are there wide high-resolution matrix printers in the store?

System:   No, there are no such printers in the store.

User:      And narrow?

it is difficult for the computer to understand the meaning of the last remark.

A rather detailed linguistic analysis is necessary to re-formulate thisuser’s question to Are there narrow high-resolution matrix printers in the store? In many cases, the only way for the computer to understand such elliptical questions is to build a model of the user’s current goals, its knowledge, and interests, and then try to guess what the computer itself would be asking at this point of the dialogue if it were the user, and in what words it would formulate such a question. This idea can be called analysis through synthesis.

Extraction of factual data from texts

Extraction of factual data from texts is the task of automatic generationof elements of a factographic database, such as fields, or parameters, based onon-line texts. Often the flows of the current news from the Internet or from aninformation agency are used as the source of information for such systems, andthe parameters of interest can be the demand for a specific type of a productin various regions, the prices of specific types of products, events involvinga particular person or company, opinions about a specific issue or a politicalparty, etc.

The decision-making officials in business and politics are usually toobusy to read and comprehend all the relevant news in their available time, sothat they often have to hire many news summarizers and readers or even toaddress to a special information agency. This is very expensive, and even inthis case the important relationships between the facts may be lost, since eachnews summarizer typically has very limited knowledge of the subject matter. Afully effective automatic system could not only extract the relevant facts muchfaster, but also combine them, classify them, and investigate theirinterrelationships.

There are several laboratory systems of that type for business applications,e.g., a system that helps to explore news on Dow Jones index, investments, andcompany merge and acquisition projects. Due to the great difficulties of thistask, only very large commercial corporations can afford nowadays the researchon the factual data extraction problem, or merely buy the results of suchresearch.

This kind of problem is also interesting from the scientific and technicalpoint of view. It remains very topical, and its solution is still to be foundin the future. We are not aware of any such research in the world targeted tothe Spanish language so far.

Text generation

The generation of texts from pictures and formal specifications is acomparatively new field; it arose about ten years ago. Some useful applicationsof this task have been found in recent years. Among them are multimedia systemsthat require a text-generating subsystem to illustrate the pictures throughtextual explanations. These subsystems produce coherent texts, starting fromthe features of the pictures.

Another very important application of systems of this kind is thegeneration of formal specifications in text form from quite formal technicaldrawings.

For example, compilation of a patent formula for a new device, often manypages long, is a boring, time-consuming, and error-prone task for a human. Thistask is much more suitable for a machine.

A specific type of such a system is a multilingual text generatingsystem. In many cases, it is necessary to generate descriptions and instructionsfor a new device in several languages, or in as many languages as possible.

Due to the problems discussed in the section on translation, the qualityof automatic translation of a manually compiled text is often very low.

Better results can be achieved by automatic generation of the requiredtext in each language independently, from the technical drawings andspecifications or from a text in a specific formal language similar to a programminglanguage.

Text generating systems have, in general, half of the linguistic problemsof a translation system, including all of the linguistic problems connected with the grammar and lexicon of the target language. This is still a vast set oflinguistic information, which is currently available in adequate detail foronly a few very narrow subject areas.

Systems of language understanding

Natural language understanding systems are the most general and complexsystems involving natural language processing. Such systems are universal inthe sense that they can perform nearly all the tasks of other language-relatedsystems, such as grammar and style checking, information retrieval, automatictranslation, natural language interface, extraction of factual data from texts,text generation, and so forth.

For example, automatic translation can be implemented as a textunderstanding system, which understands the text in the source language andthen generates a precise description of the learned information in the targetlanguage.

Hence, creation of a text understanding system is the most challengingtask for the joint efforts of computational linguistics and artificial intelligence.

To be more precise, the natural language processing module is only onepart of such a system. Most activities related to logical reasoning andunderstanding proper are concentrated in another its part—a reasoning module.These two modules, however, are closely interrelated and they should work intight cooperation.

The linguistic subsystem is usually bi-directional, i.e., it can both “understand,”or analyze, the input texts or queries, and produce, or generate, another text as the answer. In other words, this subsystem transforms a human utterance into an internal, or semantic, representation comprehensible to the reasoning subsystem, produces a response in its own internal format, and then transforms the response into a textual answer.

In different systems, the reasoning subsystem can be called theknowledge-based reasoning engine, the problem solver, the expert system, orsomething else, depending on the specific functions of the whole system. Itsrole in the whole system of natural language understanding is always veryimportant.

Half a century ago, Alan Turing suggested that the principal test ofintelligence for a computer would be its ability to conduct an intelligentdialogue, making reasonable solutions, giving advice, or just presenting therelevant information during the conversation.

This great goal has not been achieved thus far, but research in thisdirection has been conducted over the past 30 years by specialists inartificial intelligence and computational linguistics.

In order to repeat, the full power of linguistic science and the fullrange of linguistic data and knowledge are necessary to develop what we cancall a true language understanding system.

Related systems

There are other types of applications that are not usually considered systems of computational linguistics proper, but rely heavily on linguistic methodsto accomplish their tasks. Of these we will mention here two, both related topattern recognition.

Optical character recognition systems recognize the graphemes, i.e., letters, numbers, and punctuation marks, in a point-by-point image of an arbitrary text printed on paper, and convert them to the corresponding ascii codes. The graphemes can be in any font, typeface or size; the background of the paper can contain some dots and spots. An example of what the computer sees is given in Figure III.6.

A human being easily reads the following text Reglas de interpretaciónsemántica:  because he or she understands the meaning of the words. However, without understanding the meaning it is not possible to recognize, say, the first letter of the last string (is it a, s or g?), or the first letter(s) of the second line (is it r, i or m?).

Figure III.6. The image of a text, as the computer sees it.

Figure III.7. Several letters of the same text, as the computer sees them.

The first letters of the second line are shown separately in Figure III.7.One who does not know what the whole word means, cannot even say for sure thatthis picture represents any letters. However, one can easily read precisely thesame image, shown above, of the same letters in their complete context. Hence, it is obvious that the task of optical character recognition cannot be solvedonly by the methods of image recognition, without linguistic information.

The image recognition proper is beyond the competence of computational linguistics.However, after the recognition of an image even of much higher quality than theone shown in Figure III.6, some peculiar errors can still appear in thetextual representation of the image. They can be fixed by the operationssimilar to those of a spell checker. Such a specialized spell checker shouldknow the most frequent errors of recognition. For example, the lower-caseletter l is very similar to the figure 1, the letter n is frequently recognized as the pair ii, while the m can be recognized as iii or rn. Vice versa, the digraphs in, rn and ni are frequently recognized as m, and so forth.

Most such errors can be corrected without human intervention, on the baseof linguistic knowledge. In the simplest case, such knowledge is just adictionary of words existing in the language. However, in some cases a deeperlinguistic analysis is necessary for disambiguation. For example, only full parsingof the context sentence can allow the program to decide whether the picturerecognized as *danios actually represents the existing Spanish words darnos or damos.

A much more challenging task than recognition of printed texts is handwritingrecognition. It is translation into ascii form of the texts written by hand with a pen on paper or on the surface of a special computer device, or directly with a mouse on the computer screen. However, the main types of problem and the methods of solution for this task are nearly the same as for printed texts, at least in their linguistic aspect.

Speech recognition is another type of recognition task employing linguistic methods. A speech recognition system recognizes specific sounds in the flow of a speech of a human and then converts them into ascii codes of the corresponding letters. The task of recognition itself belongs both to pattern recognition and to phonology, the science bordering on linguistics, acoustics, and physiology, which investigates the sounds used in speech.

The difficulties in the task of speech recognition are very similar orquite the same as in optical character recognition: mutilated patterns, fusedpatterns, disjoint parts of a pattern, lost parts of the pattern, noisesuperimposing the pattern. This leads to even a much larger number ofincorrectly recognized letters than with optical character recognition, andapplication of linguistic methods, generally in the same manner, is even moreimportant for this task.


A short review of applied linguistic systems has shown that only verysimple tasks like hyphenation or simple spell checking can be solved on amodest linguistic basis. All the other systems should employ relatively deeplinguistic knowledge: dictionaries, morphologic and syntactic analyzers, and insome cases deep semantic knowledge and reasoning. What is more, nearly all ofthe discussed tasks, even spell checking, have to employ very deep analysis tobe solved with an accuracy approaching 100%. It was also shown that most of thelanguage processing tasks could be considered as special cases of the generaltask of language understanding, one of the ultimate goals of computationallinguistics and artificial intelligence.

IV.  Language as a Meaning Û Text Transformer

In this chapter, we will return to linguistics, to make a review of several viewpoints on natural language, and to select one of them as the base for further studies. The components of the selected approach will be defined, i.e., text and meaning. Then some conceptual properties of the linguistic transformations will be described and an example of these transformations will be given.

Possible points of view on natural language

One could try to define natural language in one of the following ways:

· The principal means for expressing human thoughts;

· The principal means for text generation;

· The principal means of human communication.

The first definition—“the principal means for expressing human thoughts”—touches upon the expressive function of language. Indeed, some features of the outer world are reflected in the human brain and are evidently processed by it, and this processing is just the human thought. However, we do not have any real evidence that human beings directly use words of a specific natural language in the process of thinking. Modes of thinking other than linguistic ones are also known. For example, mathematicians with different native languages can have the same ideas about an abstract subject, though they express these thoughts in quite different words. In addition, there are kinds of human thoughts—like operations with musical or visual images—that cannot be directly reduced to words.

As to the second definition—“the principal means for text generation”—thereis no doubt that the flow of utterances, or texts, is a very important resultof functioning of natural language.

However, communication includes not only generation (speaking), but alsounderstanding of utterances. This definition also ignores the starting point oftext generation, which is probably the target point of understanding.Generation cannot exist without this starting point, which is contents of the target utterance or the text in the whole. We can call these contents meaning.

Figure IV.1. The role of language in human communication.

As to the third definition—“the principal means of human communication,”—wecan readily agree that the communicative function is the main function ofnatural language.

Clearly, only persons living in contact with society really need alanguage for efficient communication. This definition is perhaps correct, butit does not touch upon two main aspects of communication in natural language,namely, speaking and understanding, and thus it does not try to define theseaspects, separately and in their interaction.

Thus, these definitions are not sufficient for our purposes. A betterdefinition should touch upon all useful components of the ones given above,such as text, meaning, generation, and understanding. Such definition will be given in the next section.

Language as a bi-directional transformer

The main purpose of human communication is transferring someinformation—let us call it Meaning[6]—from one person to the other. However, the direct transferring of thoughts is not possible.

Figure IV.2. Language functions like encoder / decoder in a communication channel.

Thus, people have to use some special physical representation of theirthoughts, let us call it Text.[7] Then, language is a tool to transform one of these representations to another, i.e. to transform Meanings to words when speaking, and the words to their Meaning when listening (see Figure IV.1).

It is important to realize that the communicating persons use the samelanguage, which is their common knowledge, and each of them has a copy of it inthe brain.

If we compare this situation with transferring the information over acommunication channel, such as a computer network, the role of language isencoding the information at the transmitting end and then decoding it at thereceiving end.[8] Again, here we deal with two copies of the same encoder/decoder (see Figure IV.2).

Thus, we naturally came to the definition of natural language as atransformer of Meanings to Texts, and, in the opposite direction, from Texts to Meanings (see Figure IV.3).

Figure IV.3. Language as a Meaning Û Text transformer.

This transformer is supposed to reside in human brain. By transformationwe mean some form of translation, so that both the Text and the correspondingMeaning contain the same information. What we specifically mean by these two concepts, Text and Meaning, will be discussed in detail later.

Being originally expressed in an explicit form by Igor Mel’čuk , this definition is shared nowadays by many other linguists. It permits to recognize how computer programs can simulate, or model, the capacity of the human brain to transform the information from one of these representations into another.

Essentially, this definition combines the second and the thirddefinitions considered in the previous section. Clearly, the transformation ofText into Meaning and vice versa is obligatory for any human communication,since it implies transferring the Meaning from one person to another using theText as its intermediate representation. The transformation of Meaning intoText is obligatory for the generation of utterances. To be more precise, in thewhole process of communication of human thoughts the definition 1 given earlieractually refers to Meaning, the definition 2 to Text, and the definition 3 toboth mentioned aspects of language.

Figure IV.4. Meaning Û Text many-to-many mapping.

With our present definition, language can be considered analogous to atechnical device, which has input and output. Some information, namely Text,being entered to its input, is transformed into another form with equivalent contents.

The new form at the output is Meaning. More precisely, we consider a bi-directional transformer, i.e., two transformers working in parallel but inopposite directions. Text is the result of the activity of one of thesetransformers, and Meaning, of the other.

Programmers can compare such a device with a compiler, let us say, a C++compiler, which takes a character file with the ascii text of the program in the input and produces somebinary code with machine instructions, as the output. The binary codecorresponds to the meaning of the program. However, a compiler usually cannottranslate the machine instructions back to a C++ program text.

As a mathematical analogy to this definition, we can imagine abi-directional mapping between one huge set, the set of all possible Texts, andanother huge set, the set of all possible Meanings (see Figure IV.4).

Figure IV.5. Metaphor of surface and deep structures.

The two sets, Texts and Meanings, are not quite symmetric in their properties.Only the Texts have an explicit expression, only they can be immediatelyobserved or directly transferred from one person to another, while the Meaningsreside in the brain of each person independently and cannot be immediatelyobserved or assessed.

This is similar to the front panel of an electrical device: the lightsand switches on its surface can be observed, while the electrical processes represented by the lights and controlled by the switches are deep[9] under the cover of the device, and can only be guessed by watching the lights or experimenting with the switches.

Another metaphor of surface and deep structures of language is shown in Figure IV.5.We can directly observe the surface of water, but in order to learn what leafis connected to what flower through common roots, we need to analyze what isunder the surface. There is much more below the surface than on top of it, andonly analysis of the deeper phenomena gives us understanding of the wholething.

Figure IV.6. Two levels of representation.

All this is often considered as surface and deep levels of the representation of utterances (see Figure IV.6). The man on the picture cannot see the meaning of the text immediately and has to penetrate below the surface of the text to find its meaning.

Thus, the set of Texts is considered the surface edge of the Meaning Û Text transformer, while the set of Meanings gives its deep edge. The Meaning corresponding to the given Text at the depth is also called its semantic representation.

The transformation of Meaning into Text is called synthesis of the Text. The transformation to the inverse direction, that is from Text into Meaning, is called analysis of Text. Thus, according to our basic definition, natural language is both analyzer and synthesizer of Texts, at the same time.

This definition uses the notions of Text and Meaning, although they havebeen neither defined nor described so far. Such descriptions will be given inthe following sections.

Text, what is it?

The empirical reality for theoretical linguistics comprises, in the firstplace, the sounds of speech. Samples of speech, i.e., separate words,utterances, discourses, etc., are given to the researchers directly and, forliving languages, are available in an unlimited supply.

Speech is a continuous flow of acoustic signals, just like music ornoise. However, linguistics is mainly oriented to the processing of naturallanguage in a discrete form.

The discrete form of speech supposes dividing the flow of the acousticsignals into sequentially arranged entities belonging to a finite set ofpartial signals. The finite set of all possible partial signals for a givenlanguage is similar to a usual alphabet, and is actually called a phonetic alphabet.

For representation of the sound of speech on paper, a special phonetictranscription using phonetic symbols to represent speech sounds was invented by scientists. It is used in dictionaries, to explain the pronunciation of foreign words, and in theoretical linguistics.

A different, much more important issue for modern computational linguisticsform of speech representation arose spontaneously in the human practice as thewritten form of speech, or the writing system.

People use three main writing systems: that of alphabetic type, of syllabictype, and of hieroglyphic type. The majority of humankind use alphabeticwriting, which tries to reach correspondence between letters and sounds ofspeech.

Two major countries, China and Japan,[10] use the hieroglyphic writing. Several countries use syllabic writing, among them Korea. Hieroglyphs represent the meaning of words or their parts. At least, they originally were intended to represent directly the meaning, though the direct relationship between a hieroglyph and the meaning of the word in some cases was lost long ago.

Letters are to some degree similar to sounds in their functions. In their origin, letters were intended to directly represent sounds, so that a text written in letters is some kind of representation of the corresponding sounds of speech. Nevertheless, the simple relationship between letters and sounds in many languages was also lost. In Spanish, however, this relationship is much more straightforward than, let us say, in English or French.

Syllabic signs are similar to letters, but each of them represents a whole syllable, i.e., a group of one or several consonants and a vowel. Thus, such a writing system contains a greater number of signs and sometimes is less flexible in representing new words, especially foreign ones. Indeed, foreign languages can contain specific combinations of sounds, which cannot be represented by the given set of syllables. The syllabic signs usually have more sophisticated shape than in letter type writing, resembling hieroglyphs to some degree.

In more developed writing systems of a similar type, the signs (called inthis case glyphs) can represent either single sounds or larger parts of words such as syllables, groups of syllables, or entire words. An example of such a writing system is Mayan writing (see Figure I.2). In spite of their unusual appearance, Mayan glyphs are more syllabic signs than hieroglyphs, and they usually represent the sounds of the speech rather than the meaning of words. The reader can become familiar with Mayan glyphs through the Internet site [52].

Currently, most of the practical tasks of computational linguistics are connected with written texts stored on computer media. Among written texts, those written inalphabetic symbols are more usual for computational linguistics than thephonetic transcription of speech.[11] Hence, in this book the methods of language processing will usually be applied to the written form of natural language.

For the given reason, Texts mentioned in the definition of languageshould then be thought of as common texts in their usual written form. Writtentexts are chains of letters, usually subdivided into separate words by spaces[12] and punctuation marks. The combinations of words can constitute sentences, paragraphs, and discourses. For computational linguistics, all of them are examples of Texts.[13]

Words are not utmost elementary units of language. Fragments of texts,which are smaller than words and, at the same time, have their own meanings,are called morphs. We will define morphs more precisely later. Now it is sufficient for us to understand that a morph can contain an arbitrary number of letters (or now and then no letters at all!), and can cover a whole word or some part of it. Therefore, Meanings can correspond to some specially defined parts of words, whole words, phrases, sentences, paragraphs, and discourses.

It is helpful to compare the linear structure of text with the flow ofmusical sounds. The mouth as the organ of speech has rather limited abilities.It can utter only one sound at a time, and the flow of these sounds can beadditionally modulated only in a very restricted manner, e.g., by stress,intonation, etc. On the contrary, a set of musical instruments can produceseveral sounds synchronously, forming harmonies or several melodies going inparallel. This parallelism can be considered as nonlinear structuring. Thehuman had to be satisfied with the instrument of speech given to him by nature.This is why we use while speaking a linear and rather slow method of acousticcoding of the information we want to communicate to somebody else.

The main features of a Text can be summarized as follows:

· Meaning. Not any sequence of letters can be considered a text. A text is intended to encode some information relevant for human beings. The existing connection between texts and meanings is the reason for processing natural language texts.

· Linear structure. While the information contained in the text can have a very complicated structure, with many relationships between its elements, the text itself has always one-dimensional, linear nature, given letter by letter. Of course, the fact that lines are organized in a square book page does not matter: it is equivalent to just one very long line, wrapped to fit in the pages. Therefore, a text represents non-linear information transformed into a linear form. What is more, the human cannot represent in usual texts even the restricted non-linear elements of spoken language, namely, intonation and logical stress. Punctuation marks only give a feeble approximation to these non-linear elements.

· Nested structure and coherence. A text consists of elementary pieces having their own, usually rather elementary, meaning. They are organized in larger structures, such as words, which in turn have their own meaning. This meaning is determined by the meaning of each one of their components, though not always in a straightforward way. These structures are organized in even larger structures like sentences, etc. The sentences, paragraphs, etc., constitute what is called discourse, the main property of which is its connectivity, or coherence: it tells some consistent story about objects, persons, or relations, common to all its parts. Such organization provides linguistics with the means to develop the methods of intelligent text processing.

Thus, we could say that linguistics studies human ways of linear encoding[14] of non-linear information.

Meaning, what is it?

Meanings, in contrast to texts, cannot be observed directly. As wementioned above, we consider the Meaning to be the structures in the humanbrain which people experience as ideas and thoughts. Since we do not know andcannot precisely represent those brain processes, for practical purposes wemust use a representation of Meaning, which is more suitable for manipulationin a computer. Thus, for our purposes, Meaning will be identified with thatrepresentation.

In the future, neurophysiological research will eventually discover whatsignals really correspond to meanings in our brain, and what structure thesesignals have, but for now those signals remain only vaguely understood. Hencewe take the pragmatic viewpoint that, if a representation we use allows acomputer to manipulate and respond to texts with an ability close to that of ahuman, then this representation is rather good for the real Meaning and fitsour purposes.

As it was described earlier, the task of language is to transforminformation from one representation, the Text, into another, the Meaning, andvice versa. A computer program that models the function of language mustperform the same task. In any application, there should be some other system ordevice that consumes the results of the transformation of texts, and producesthe information that is to be transformed into a text. The operations of such adevice or a system is beyond the scope of computational linguistics itself. Rather,such a system uses the linguistic module as its interface with the outer world (see Figure IV.7).

For a linguistic module of a system, the Meaning is a formal language ora format of information representation immediately understandable for, or executableon, the consumer of the information: the underlying expert or reasoning system,database, robot control system, etc. It is supposed that this underlying systemproduces its responses just in the same format. Thus, in practice, the formatof Meaning is already given to the developers of the linguistic module for anyspecific application system.

Usually, such systems are aware on the entities mentioned in the text, their states, properties, processes, actions, and relationships.

Besides, there exists other type of information in a text, such as beliefs,estimations, and intentions of its author. For example, in the Spanish sentence Creo que su esposa está aquí, the part reflecting the author’s belief is given in bold face. The author usually flavors any text with these elements. Words reflecting basic information, through some stylistic coloring, can additionally express author’s attitude. For example, the Spanish sentence Este hombrón no está trabajando ahora has the meaning ‘this man is not working now and I consider him big and coarse’.

Figure IV.7. Structure of an application system

with a natural language interface.

Perhaps only very formal texts like legal documents do not contain subjective attitude of the author(s) to the reflected issues. The advanced applicationsystem should distinguish the basic information delivered in texts from author’sbeliefs, estimations, and intentions.

Additionally, even a very formal text contains many explicit referencesand explanations of links between parts of a text, and these elements serve asa content table or a protocol of the author’s information about textstructuring. This is not the information about the relevant matters as such.Instead, this is some kind of meta-information about how the parts of the textare combined together, i.e., a “text about text.” We will not discuss suchinsertions in this book. Since properties, processes, and actions can berepresented as relationships between entities touched upon in Text, just thesefeatures are used to represent Meaning.

The entities, states, properties, processes, and actions are usuallydenoted by some names in semantic representation. These names can be compared with the names of variables and functions in a computer program. Such names have no parts of speech, so that a process or an action of, say, ‘development’ can be equally called in Spanish desarrollar or desarrollo, while the property of ‘small’ can be equally called pequeño or ser pequeño. Usually only one wordform is used for the name, so that the property itself is called, for instance, pequeño (neither pequeña nor pequeñas). Concerning the value plural of grammatical category of number, on the semantic level it is transformed to the notion of multiplicity and is represented by a separate element.

Two ways to represent Meaning

To represent the entities and relationships mentioned in the texts, thefollowing two logically and mathematically equivalent formalisms are used:

· Predicative formulas. Logical predicates are introduced in mathematical logic. In linguistics, they are used in conventional logical notation and can have one or more arguments. Coherence of a text is expressed in the first place by means of common arguments for the predicates involved. For example, the meaning of the sentence Las niñas pequeñas ven la flor roja is represented by the following conjunctive predicative formula:

VER(niña, flor) &

MUCHO(niña) &

PEQUEÑO(niña) &

SOLO(flor) &


In such representation, predicates SOLO and MUCHO have the meanings ‘number of the entities given by the argument is one’ and ‘number of the entities given by the argument is more than one,’ respectively. Arguments of the predicates that are not predicates by themselves are called terms. They are written in lowercase letters, in contrast to the predicates that are written in uppercase.

Figure IV.8. Semantic network for the sentence

Las niñas pequeñas ven la flor roja.

· Directed labeled graphs. The nodes of these graphs represent the terms or predicates, and the arrows connect predicates with their arguments, i.e., terms or other predicates. The arrows are marked with numeric labels according to the number of the corresponding argument (1st argument, 2nd, etc.). Though each predicate assigns to each its argument a specific semantic role, the numerical labels in the graph are used only to distinguish the arguments. For predicates denoting actions, the label 1 usually marks the agent, the label 2, patient or target, etc. Nevertheless, a label of such type does not make any semantic allusion, the enumeration being rather arbitrary. Thus, the graph representation is just equivalent to the predicate one rather than provides any additional information. The semantic representation in the form of directed labeled graph is often called semantic network. Figure IV.8 shows the semantic network representation for the example above.

These two representations are equivalent, and either of them can be used. Indeed, there exist a number of easy ways to encode any network linearly.

For human readers, in books and articles, the graph representation isespecially convenient. For internal structures of a computer program, theequivalent predicative representation is usually preferred, with special formalmeans to enumerate the referentially common arguments of various logical predicates.

The graph representation explicitly shows the commonality of the arguments,making it obvious what is known about a specific entity. In the drawing shownin the Figure IV.8, for example, it is immediately seen that there arethree pieces of information about the terms niña and flor, two pieces about the predicate VER, and one for predicate ROJO and SOLO.

Some scientists identify the representations of Meaning with the representation of human knowledge in general[15]. The human knowledge is of concern not only for natural language processing, but also for the task of transferring knowledge between computers, whether that knowledge is expressed by natural language or by some other means. For the transfer of knowledge, it is important to standardize methods of representation, so that the knowledge can be communicated accurately. Just for these purposes, the computer community is developing knowledge representation in both the linear and the graphical format. The accuracy and utility of any representation of knowledge should be verified in practice, i.e., in applications.

In the opinion of other scientists, the representations of Meaning and ofhuman knowledge can operate by the same logical structures, but the knowledgein general is in no way coincident with purely linguistic knowledge and evencan have different, non-discrete, nature. Hence, they argue, for the transitionfrom Meaning in its linguistic sense to the corresponding representation interms of general human knowledge, some special stage is needed. This stage isnot a part of language and can operate with tools not included in those of thelanguage proper.

Decomposition and atomization ofMeaning

Semantic representation in many cases turns out to be universal, i.e., common to different natural languages. Purely grammatical features of different languages are not usually reflected in this representation. For example, the gender of Spanish nouns and adjectives is not included in their semantic representation, so that this representation turned to be equal to that of English. If the given noun refers to a person of a specific sex, the latter is reflected on semantic level explicitly, via a special predicate of sex, and it is on the grammar of specific language where is established the correspondence between sex and gender. It is curious that in German nouns can have three genders: masculine, feminine, and neuter, but the noun Mädchen ‘girl’ is neuter, not feminine!

Thus, the semantic representation of the English sentence The littlegirls see the red flower it is the same as the one given above, despite the absence of gender in English nouns and adjectives. The representation of the corresponding Russian sentence is the same too, though the word used for red in Russian has masculine gender, because of its agreement in gender with corresponding noun of masculine.[16]

Nevertheless, the cases when semantic representations for two or moreutterances with seemingly the same meaning do occur. In such situations,linguists hope to find a universal representation via decomposition and evenatomization of the meaning of several semantic components.

In natural sciences, such as physics, researchers usually try to divideall the entities under consideration into the simplest possible, i.e., atomic,or elementary, units and then to deduce properties of their conglomerations from the properties of these elementary entities. In principle, linguistics hasthe same objective. It tries to find the atomic elements of meaning usuallycalled semantic primitives, or semes.

Semes are considered indefinable, since they cannot be interpreted interms of any other linguistic meanings. Nevertheless, they can be explained tohuman readers by examples from the extralinguistic reality, such as pictures,sound records, videos, etc. All other components of semantic representation should be then expressed through the semes.

In other words, each predicate or its terms can be usually represented inthe semantic representation of text in a more detailed manner, such as alogical formula or a semantic graph. For example, we can decompose


i.e., MATAR(x) is something like ‘causar cesar el vivir(x),’ or ‘cause stop living(x),’ where the predicates CESAR(x), VIVIR(y), and CAUSAR(z) are more elementary than the initial predicate MATAR(x).[17]

Figure IV.9 shows a decomposition of the sentence Juan mató aJosé enseguida = Juan causó a José cesar vivir enseguida in the mentioned more primitive notions. Note that the number labels of valencies of the whole combination of the primitives can differ from the number labels of corresponding valencies of the member primitives: e.g., the actant 2 of the whole combination is the actant 1 of the component VIVIR. The mark C in Figure IV.9 stands for the circumstantial relation (which is not a valency but something inverse, i.e., a passive semantic valency).

Figure IV.9. Decomposition of the verb MATAR into semes.

Over the past 30 years, ambitious attempts to find and describe a limitednumber of semes, to which a major part of the semantics of a natural languagewould be reduced, have not been successful.

Some scientists agree that the expected number of such semes is not muchmore than 2´000, but until now, this figure is still debatable. To comply withneeds of computational linguistics, everybody agreed that it is sufficient todisintegrate meanings of lexemes to a reasonable limit implied by the application.

Therefore, computational linguistics uses many evidently non-elementaryterms and logical predicates in the semantic representation. From this point of view, the translation from one cognate language to another does not need any disintegration of meaning at all.

Once again, only practical results help computational linguists to judgewhat meaning representation is the best for the selected application domain.

Not-uniqueness of Meaning Þ Text mapping: Synonymy

Returning to the mapping of Meanings to Texts and vice versa, we shouldmention that, in contrast to common mathematical functions, this mapping is notunique in both directions, i.e., it is of the many-to-many type. In thissection, we will discuss one direction of the mapping: from Meanings to Texts.

Different texts or their fragments can be, in the opinion of all or themajority of people, equivalent in their meanings. In other words, two or moretexts can be mapped to the same element of the set of Meanings. In Figure IV.4,the Meaning M is represented with three different Texts T, i.e., these three Texts have the same Meaning.[18]

For example, the Spanish adjectives pequeño and chico are equivalent in many contexts, as well as the English words small and little. Such equivalent words are called synonymous words, or synonyms, and the phenomenon is called synonymy of words. We can consider also synonymy of word combinations (phrases) or sentences as well. In these cases the term synonymous expressions is used.

The words equivalent in all possible contexts are called absolutesynonyms. Trivial examples of absolute synonymy are abbreviated and complete names of organizations, e.g. in Spanish onu º Organización de las Naciones Unidas. Nontrivial examples of absolute synonymy of single words are rather rare in any language. Examples from Mexican Spanish are: alzadura º alzamiento, acotación º acotamiento, coche º carro.

However, it is more frequent that the two synonyms are equivalent in theirmeanings in many contexts, but not all.

Sometimes the set of possible contexts for one such synonym covers thewhole set of contexts for another synonym; this is called inclusive synonymy. Spanish examples are querer > desear > anhelar: querer is less specific than desear which in turn is less specific than anhelar. It means that in nearly every context we can substitute desear or querer for anhelar, but not in every context anhelar can be substituted for querer or desear.

Most frequently, though, we can find only some—perhaps significant—intersectionof the possible sets of contexts. For example, the Spanish nouns deseo and voluntad are exchangeable in many cases, but in some cases only one of them can be used.

Such partial synonyms never have quite the same meaning. In some contexts, the difference is not relevant, so that they both can be used, whereas in other contexts the difference does not permit to replace one partial synonym with the other.

The book [24] is a typical dictionary of synonyms in printed form. Themenu item Language | Synonyms in Microsoft Word is a typical example of an electronic dictionary of synonyms. However, many of the words that it contains in partial lists are not really synonyms, but related words, or partial synonyms, with a rather small intersection of common contexts.

Not-uniqueness of Text Þ Meaning mapping: homonymy

In the opposite direction—Texts to Meanings—a text or its fragment canexhibit two or more different meanings. That is, one element of the surfaceedge of the mapping (i.e. text) can correspond to two or more elements of thedeep edge. We have already discussed this phenomenon in the section on automatictranslation, where the example of Spanish word gato was given (see page 72). Many such examples can be found in any Spanish-English dictionary. A few more examples from Spanish are given below.

· The Spanish adjective real has two quite different meanings corresponding to the English real and royal.

· The Spanish verb querer has three different meanings corresponding to English to wish, to like, and to love.

· The Spanish noun antigüedad has three different meanings:

– ‘antiquity’, i.e. a thing belonging to an ancient epoch,

– ‘antique’, i.e. a memorial of classic antiquity,

– ‘seniority’, i.e. length of period of service in years.

The words with the same textual representation but different meanings arecalled homonymous words, or homonyms, with respect to each other, and the phenomenon itself is called homonymy. Larger fragments of texts—such as word combinations (phrases) or sentences—can also be homonymous. Then the term homonymous expressions is used.

To explain the phenomenon of homonymy in more detail, we should resortagain to the strict terms lexeme and wordform, rather than to the vague term word. Then we can distinguish the following important cases of homonymy:

· Lexico-morphologic homonymy: two wordforms belong to two different lexemes. This is the most general case of homonymy. For example, the string aviso is the wordform of both the verb AVISAR and the noun AVISO. The wordform clasificación belong to both the lexeme CLASIFICACIÓN1 ‘process of classification’ and the lexeme CLASIFICACIÓN2 ‘result of classification,’ though the wordform clasificaciones belongs only to CLASIFICACIÓN2, since CLASIFICACIÓN1 does not have the plural form. It should be noted that it is not relevant whether the name of the lexeme coincides with the specific homonymous wordform or not.

Another case of lexico-morphologic homonymy is represented by two different lexemes whose sets of wordforms intersect in more than one wordforms. For example, the lexemes RODAR and RUEDA cover two homonymous wordforms, rueda and ruedas; the lexemes IR and SER have a number of wordforms in common: fui, fuiste, ..., fueron.

· Purely lexical homonymy: two or more lexemes have the same sets of wordforms, like Spanish REAL1 ‘real’ and REAL2 ‘royal’ (the both have the same wordform set {real, reales}) or QUERER1 ‘to wish,’ QUERER2 ‘to like,’ and QUERER3 ‘to love.’

· Morpho-syntactic homonymy: the whole sets of wordforms are the same for two or more lexemes, but these lexemes differ in meaning and in one or more morpho-syntactic properties. For example, Spanish lexemes (el) frente ‘front’ and (la) frente ‘forehead’ differ, in addition to meaning, in gender, which influences syntactical properties of the lexemes.

· Purely morphologic homonymy: two or more wordforms are different members of the wordform set for the same lexeme. For example, fáciles is the wordform for both masculine plural and feminine plural of the Spanish adjective FÁCIL ‘easy.’ We should admit this type of homonymy, since wordforms of Spanish adjectives generally differ in gender (e.g., nuevos, nuevas ‘new’).

Resolution of all these kinds of homonymy is performed by the human listeneror reader according to the context of the wordform or based on theextralinguistic situation in which this form is used. In general, the reader orlistener does not even take notice of any ambiguity. The corresponding mentaloperations are immediate and very effective. However, resolution of suchambiguity by computer requires sophisticated methods.

In common opinion, the resolution of homonymy (and ambiguity in general) is one of the most difficult problems of computational linguistics and must be dealt with as an essential and integral part of the language-understanding process.

Without automatic homonymy resolution, all the attempts to automatically“understand” natural language will be highly error-prone and have ratherlimited utility.

More on homonymy

In the field of computational linguistics, homonymous lexemes usuallyform separate entries in dictionaries. Linguistic analyzers must resolve thehomonymy automatically, by choosing the correct option among those described inthe dictionary.

For formal distinguishing of homonyms, their description in conventional dictionaries is usually divided into several subentries. The names of lexical homonymsare supplied with the indices (numbers) attached to the words in their standarddictionary form, just as we do it in this book. Of course, in text generation,when the program compiles a text containing such words, the indices areeliminated.

The purely lexical homonymy is maybe the most difficult to resolve sinceat the morphologic stage of text processing it is impossible to determine whathomonym is true in this context. Since morphologic considerations are useless,it is necessary to process the hypotheses about several homonyms in parallel.

Concerning similarity of meaning of different lexical homonyms, varioussituations can be observed in any language. In some cases, such homonyms haveno elements of meaning in common at all, like the Spanish REAL1 ‘real’ and REAL2 ‘royal.’ In other cases, the intersection of meaning is obvious, like in QUERER2 ‘to like’ and QUERER3 ‘to love,’ or CLASIFICACIÓN1 ‘process of classification’ and CLASIFICACIÓN2 ‘result of classification.’ In the latter cases, the relation can be exposed through the decomposition of meanings of the homonyms lexemes. The cases in which meanings intersect are referred to in general linguistics as polysemy.

For theoretical purposes, we can refer the whole set of homonymouslexemes connected in their meaning as vocable. For example, we may introduce the vocable {QUERER1, QUERER2, QUERER3}. Or else we can take united lexeme, which is called polysemic one.

In computational linguistics, the intricate semantic structures of various lexemes are usually ignored. Thus, similarity in meaning is ignored too.

Nevertheless, for purely technical purposes, sets of any homonymouslexemes, no matter whether they are connected in meaning or not, can be considered. They might be referred as pseudo-vocables. For example, the pseudo-vocable REAL = {REAL1, REAL2} can be introduced.

A more versatile approach to handle polysemy in computational linguisticshas been developed in recent years using object-oriented ideas. Polysemiclexemes are represented as one superclass that reflects the common part oftheir meaning, and a number of subclasses then reflect their semantic differences.

A serious complication for computational linguistics is that new sensesof old words are constantly being created in natural language. The older wordsare used in new meanings, for new situations or in new contexts. It has beenobserved that natural language has the property of self-enrichment and thus isvery productive.

The ways of the enrichment of language are rather numerous, and the mainof them are the following:

· A former lexeme is used in a metaphorical way. For example, numerous nouns denoting a process are used in many languages to refer also to a result of this process (cf. Spanish declaración, publicación, interpretación, etc.). The semantic proximity is thus exploited. For another example, the Spanish word estética ‘esthetics’ rather recently has acquired the meaning of heir-dressing saloon in Mexico. Since highly professional heir dressing really achieves esthetic goals, the semantic proximity is also evident here. The problem of resolution of metaphorical homonymy has been a topic of much research [51].

· A former lexeme is used in a metonymical way. Some proximity in place, form, configuration, function, or situation is used for metonymy. As the first example, the Spanish words lentes ‘lenses,’ espejuelos ‘glasses,’ and gafas ‘hooks’ are used in the meaning ‘spectacles.’ Thus, a part of a thing gives the name to the whole thing. As the second example, in many languages the name of an organization with a stable residence can be used to designate its seat. For another example, Ha llegado a la universidad means that the person arrived at the building or the campus of the university. As the third example, the Spanish word pluma ‘feather’ is used also as ‘pen.’ As not back ago as in the middle of ninth century, feathers were used for writing, and then the newly invented tool for writing had kept by several languages as the name of its functional predecessor.

· A new lexeme is loaned from a foreign language. Meantime, the former, “native,” lexeme can remain in the language, with essentially the same meaning. For example, English had adopted the Russian word sputnik in 1957, but the term artificial satellite is used as before.

· Commonly used abbreviations became common words, loosing their marking by uppercase letters. For example, the Spanish words sida and ovni are used now more frequently, then their synonymous counterparts síndrome de inmunodeficiencia adquirida and objeto volante no identificado.

One can see that metaphors, metonymies, loans, and former abbreviationsbroaden both homonymy and synonymy of language.

Returning to the description of all possible senses of homonymous words, we should admit that this problem does not have an agreed solution in lexicography. This can be proved by comparison of any two large dictionaries. Below, given are two entries with the same Spanish lexeme estante ‘rack/bookcase/shelf,’ one taken from the Dictionary of Anaya group [22] and the other from the Dictionary of Royal Academy of Spain (drae) [23].

estante (in Anaya Dictionary)

1. m. Armario sin puertas y con baldas.

2. m. Balda, anaquel.

3. m. Cada uno de los pies que sostienen la armadura de algunas máquinas.

4. adj. Parado, inmóvil.

estante (in drae)

1. a. p. us. de estar. Que está presente o permanente en un lugar. Pedro, ESTANTE en la corte romana.

2. adj. Aplícase al ganado, en especial lanar, que pasta constantemente dentro del término jurisdiccional en que está amillarado.

3. Dícese del ganadero o dueño de este ganado.

4. Mueble con anaqueles o entrepaños, y generalmente sinpuertas, que sirve para colocar libros, papeles u otras cosas.

5. Anaquel.

6. Cada uno de los cuatro pies derechos que sostienen la armadura del batán, en que juegan los mazos.

7. Cada uno de los dos pies derechos sobre que se apoya y gira el eje horizontal de un torno.

8. Murc. El que en compañía de otros lleva los pasos en las procesiones de Semana Santa.

9. Amér. Cada uno de los maderos incorruptiblesque, hincados en el suelo, sirven de sostén al armazón de las casas en lasciudades tropicales.

10. Mar. Palo o madero que se ponía sobre las mesas de guarnición para atar en él los aparejos de la nave.

One does not need to know Spanish to realize that the examples of the divergence in these two descriptions are numerous.

Some homonyms in a given language are translated into another language bynon-homonymous lexemes, like the Spanish antigüedad.

In other cases, a set of homonyms in a given language is translated intoa similar set of homonyms in the other language, like the Spanish plato when translated into the English dish (two possible interpretations are ‘portion of food’ and ‘kind of crockery’).

Thus, bilingual considerations sometimes help to find homonyms anddistinguishing their meanings, though the main considerations should be deservedto the inner facts of the given language.

It can be concluded that synonymy and homonymy are important andunavoidable properties of any natural language. They bring many heavy problemsinto computational linguistics, especially homonymy. 

Classical lexicography can help to define these problems, but theirresolution during the analysis is on computational linguistics.

Multistage character of the Meaning Û Text transformer

Figure IV.10. Levels of linguistic representation.

The ambiguity of the Meaning Û Text mapping in both directions, as well as the complicated structure ofentities on both ends of the Meaning Û Text transformer make it impossible to study this transformer without dividing the process of transformation into several sequential stages.

Existence of such stages in natural language is acknowledged by manylinguists. In this way, intermediate levels of representation of the information under processing are introduced (see Figure IV.10), as well as partial transformers for transition from a level to an adjacent (see Figure IV.11).

Two intermediate levels are commonly accepted by all linguists, withsmall differences in the definitions, namely the morphologic and syntacticones.

Figure IV.11. Stages of transformation.

In fact, classical general linguistics laid the basis for such a divisionbefore any modern research. We will study these levels later in detail.

Other intermediate levels are often introduced by some linguists so thatthe partial transformers themselves are divided into sub-transformers. Forexample, a surface and a deep syntactic level are introduced for the syntacticstage, and deep and a surface morphologic level for the morphologic one.

Thus, we can imagine language as a multistage, or multilevel,Meaning Û Text transformer (see Figure IV.12).

The knowledge necessary for each level of transformation is representedin computer dictionaries and computer grammars (see Fi­gure IV.13). A computer dictionary is a collection of information on each word, and thus it is the main knowledge base of a text processing system. A computer grammar is a set of rules based on common properties of large groups of words. Hence, the grammar rules are equally applicable to many words.

Since the information stored in the dictionaries for each lexeme is specifiedfor each linguistic level separately, program developers often distinguish amorphologic dictionary that specifies the morphologic information for each word, a syntactic dictionary, and a semantic dictionary, as in Fi­gure IV.13.

Figure IV.12. Interlevel processing.

In contrast, all information can be represented in one dictionary, giving for each lexeme all the necessary data. In this case, the dictionary entry for each lexeme has several zones that give the properties of this lexeme at the given linguistic level, i.e., a morphologic zone, syntactic zone, and semantic zone.

Clearly, these two representations of the dictionary are logically equivalent.

According to Fi­gure IV.13, the information about lexemes is distributedamong several linguistic levels. In Text, there are only wordforms. In analysis,lexemes as units under processing are involved at morphologic level. Then theyare used at surface and deep syntactical levels and at last disappeared atsemantic level, giving up their places to semantic elements. The latter elements conserve the meaning of lexemes, but are devoid of their purely grammaticalproperties, such as part of speech or gender. Hence, we can conclude that there is no level in the Text Þ Meaning transformer, which could be called lexical.

Translation as a multistage transformation

Fi­gure IV.13. The role of dictionaries and grammars in linguistic transformations.

The task of translation from one natural language to another is a goodillustration of multistage transformation of linguistic information.

Suppose there is a text in a language A that is to be translated into language B. As we have already argued, word-by-word translation leads to very poor and useless results. To translate the text with highest possible quality, the following stages of transformation are necessary:

· First stage of analysis starts from the source text in the language A and gives its morphologic representation specific for language A.

· Second stage of analysis starts from the morphologic representation and gives the syntactic representation specific for language A.

· Third stage of analysis starts from the syntactic representation and gives some level of semantic representation. The latter can be somewhat specific to language A, i.e., not universal, so that additional intra-level operations of “universalization” of semantic representation may be necessary.

The problem is that currently it is still not possible to reach the truesemantic representation, i.e., the true level of Meaning, consisting of the universal and thus standard set of semes. Therefore, all practical systemshave to stop this series of transformations at some level, as deep as possible,but not yet at that of universal Meaning.

· The transfer stage replaces the labels, i.e., of the conventional names of the concepts in language A, to the corresponding labels of language B. The result is the corresponding quasi-semantic level of representation in language B. In some cases, additional, more complex intra-level operations of “localization” are necessary at this stage.

· First stage of synthesis starts from the quasi-semantic representation with few features specific for the language B, and gives the syntactic representation quite specific for this language.

· Second stage of synthesis starts from the syntactic representation, and gives the morphologic representation specific for language B.

· Third stage of synthesis starts from the morphologic representation, and gives the target text in language B.

In the initial stages, the transformations go to the deep levels of thelanguage, and then, in the last stages, return to the surface, with the ultimateresult in textual form once more. The deeper the level reached, the smaller thedifference between the representations of this level in both languages A and B. At the level of Meaning, there is no difference at all (except maybe for the labels at semes). The deeper the level reached during the transformations, the smaller the differences that have to be ignored, and the better the quality of translation (see Figure IV.14). This scheme shows only the general idea of all these representations.

Figure IV.14. Translation as multistage transformation.

The given scheme works for an arbitrary pair of natural languages. However,if the two languages are very similar in their structure, the deeper stages ofthe transformation might not be necessary.

For example, if we translate from Spanish into Portuguese, then, becausethese two languages differ mainly in their lexicon, it can be sufficient to useonly the first stage of analysis and the last stage of synthesis, justreplacing each Spanish word by the corresponding Portuguese one on themorphologic level.

In Figure IV.14 this would correspond then to the “horizontal”transition directly on this level.

Two sides of a sign

The notion of sign, so important for linguistics, was first proposed in a science called semiotics. The sign was defined as an entity consisting of two components, the signifier and the signified. Let us first consider some examples of non-linguistic signs taken from everyday life.

· If you see a picture with a stylized image of a man in a wheelchair on the wall in the subway, you know that the place under the image is intended for disabled people. This is a typical case of a sign: the picture itself, i.e., a specific figure in contrasting colors, is the signifier, while the suggestion to assist the handicapped persons is the signified.

· Twenty Mexican pesos have two equally possible signifiers: a coin with a portrait of Octavio Paz and a piece of paper with a portrait of Benito Juárez. The signified of both of them is the value of twenty pesos. Clearly, neither of them has this value, but instead they denote it. Thus, they are two different but synonymous signs.

· Raising one’s hand (this gesture is the signifier) can have several different signifieds, for instance: (1) an intention to answer the teacher’s question in a classroom, (2) a vote for or against a proposal at a meeting, (3) an intention to call a taxi in the street, etc. These are three different, though homonymous, signs.

Linguistic sign

The notion of linguistic sign was introduced by Ferdinand de Saussure. By linguistic signs, we mean the entities used in natural languages, such as morphs, lexemes, and phrases.

Linguistic signs have several specific properties, the most obvious ofwhich is that they are to be combined together into larger signs and each onecan in turn consist of several smaller signs. Natural language can be viewed asa system of linguistic signs.

As another property of linguistic sign, its signifier at the surfacelevel consists of elementary parts, phonetic symbols in the phonetic transcriptionor letters in the written form of language. These parts do not have anysignified of their own: a letter has no meaning, but certain strings of lettersdo have it.

We have already mentioned other notation systems to represent words ofnatural language, such as hieroglyphic writing. Each hieroglyph usually has itsown meaning, so a hieroglyph is a linguistic sign. The gestures in the sign language for deaf people in some cases do have there own meanings, like hieroglyphs, and in other cases do not have any meaning of their own and serve like letters.

Linguistic sign in the mmt

In addition to the two well-known components of a sign, in the Meaning Û Text Theory yet another, a third component of a sign, is considered essential: a record about its ability or inability to combine with other specific signs. This additional component is called syntactics of the linguistic sign. For example, the ending morph ‑ar for Spanish infinitives has the same signified as -er and -ir. However, only one of them can be used in the wordform hablar, due to the syntactics of these three endings, as well as to the syntactics of the stem habl-. We can imagine a linguistic sign as it is shown on Figure IV.15.

Thus, syntactics of linguistic signs helps to choose one correct signfrom a set of synonymous signs, to be used in a specific context. For example,we choose the ending -ar for the stem habl- since it does accept just it. Beyond that, syntactics of linguistic signs helps us to disambiguate the signs, i.e., to decide which of the homonymous signs is used in a specific context.

For the extralinguistic example given above, the classroom is the context, in which we interpret the raising of one’s hand as the intention to say somethingrather than to call a taxi.

Similarly, the presence of the Spanish stem salt- is sufficient to interpret the -as ending as present tense, second person, singular in saltas, rather than feminine, plural as in the presence of the stem roj‑: rojas.

Linguistic sign in hpsg

In Head-driven Phrase Structure Grammar a linguistic sign, as usually, consists of two main components, a signifier and a signified. The signifier is defined as aphoneme string (or a sequence of such strings). Taking into account thecorrespondence between acoustic and written forms of language, such signifiercan be identified as conceptually coinciding with elements of Text in the mtt.

Figure IV.15. Syntactics of a linguistic sign.

As to the signified, an object of special type synsem is introduced for it. An object of this type is astructure (namely, a labeled tree called feature structure) with arcs representing features of various linguistic levels: morphologic, syntactic, and semantic, in a mixture. For a minimal element of the text, i.e., for a wordform, these features show:

· How to combine this wordform with the other wordforms in the context when forming syntactic structures?

· What logical predicate this word can be a part of? 

· What role this word can play in this predicate?

Simple considerations show that synsem in hpsg unites the properties of Meaning and syntactics of a sign as defined in the framework of the mtt, i.e., synsem covers syntactics plus semantics. Hence, if allrelevant linguistic facts are taken into consideration equally and properly bythe two approaches, both definitions of the linguistic sign, in hpsg and mtt, should lead to the same results.

Are signifiers given by nature or by convention?

The notion of sign appeared rather recently. However, the notionsequivalent to the signifier and the signified were discussed in science fromthe times of the ancient Greeks. For several centuries, it has been debatedwhether the signifiers of things are given to us by nature or by human convention.

The proponents of the first point of view argued that some words in anylanguage directly arose from onomatopoeia, or sound imitation. Indeed, we denote the sounds produced by a cat with verb mew or meow in English, maullar in Spanish, and miaukat’ in Russian. Hence, according to them, common signifiers lead to similar signifieds, thus creating a material link between the two aspects of the sign.

The opponents of this view objected that only a minority of words in anylanguage takes their origin from such imitation, all the other words beingquite arbitrary. For example, there is no resemblance between Spanish pared, English wall, and Russian stena, though all of them signify the same entity. Meanwhile, the phonetic similarity of the German Wand to the English wall, or the French paroi to the Spanish pared, or the Bulgarian stena to the Russian stena are caused by purely historic reasons, i.e., by the common origin of those pairs of languages.

The latter point of view has become prevalent, and now nobody looks formaterial links between the two sides of linguistic signs. Thus, a linguisticsign is considered a human convention that assigns specific meanings to somematerial things such as strings of letters, sequences of sounds, pictures inhieroglyphic writing or gestures in sign language.

Generative, mtt, and constraint ideas in comparison

In this book, three major approaches to linguistic description have beendiscussed till now, with different degree of detail: (1) generative approach developed by N. Chomsky, (2) the Meaning Û Text approach developed by I. Mel’čuk, and (3) constraint-based approach exemplified by the hpsg theory. In the ideal case, they produce equivalent results on identical language inputs. However, they have deep differences in the underlying ideas. In addition, they use similar terminology, but with different meaning, which may be misleading. In this section, we will compare their underlying ideas and the terminology. To make so different paradigms comparable, we will take only a bird’s-eye view of them, emphasizing the crucial commonalities and differences, but in no way pretending to a more deep description of any of these approaches just now.

Perhaps the most important commonality of the three approaches is thatthey can be viewed in terms of linguistic signs. All of them describe thestructure of the signs of the given language. All of them are used incomputational practice to find the Meaning corresponding to a given Text andvice versa. However, the way they describe the signs of language, and as a consequence the way those descriptions are used to build computer programs, is different. Generative idea. The initial motivation for the generative idea was the fact that describing the language is much more difficult, labor-consuming, and error-prone task than writing a program that uses such a description for text analysis. Thus, the formalism for description of language should be oriented to the process of describing and not to the process of practical application. Once created, such a description can be applied somehow.

Now, what is to describe a given language? In the view of generative tradition,it means, roughly speaking, to list all signs in it (in fact, this isfrequently referred to as generative idea). Clearly, for a natural language itis impossible to literally list all signs in it, since their number is infinite.Thus, more strictly speaking, a generative grammar describes an algorithm thatlists only the correct signs of the given language, and lists them all—in thesense that any given sign would appear in its output after a some time, perhapslong enough. The very name generative grammar is due to that it describes the process of generating all language signs, one by one at a time.

Figure IV.16. Generative idea.

There can be many ways to generate language signs. The specific kind ofgenerative grammars suggested by N. Chomsky constructs each sign gradually,through a series of intermediate, half-finished sign “embryos” of differentdegree of maturity (see Figure IV.16). All of them are built starting fromthe same “egg-cell” called initial symbol, which is not a sign of the given language. A very simple example of the rules for such gradual building is given on the pages 35 to 39; in this example, the tree structure can be roughly considered the Meaning of the corresponding string.

Where the infinity of generated signs comes from? At each step, called derivation, the generation can be continued in different ways, with any number of derivation steps. Thus, there exist an infinite number of signs with very long derivation paths, though for each specific sign its derivation process is finite.

However, all this generation process is only imaginable, and serves forthe formalism of description of language. It is not—and is not intended tobe—applied in practice for the generation of an infinitely long list oflanguage expressions, which would be senseless. The use of the description—oncecreated—for passing from Text to Meaning and vice versa is indirect. A programcalled parser is developed by a mathematician (not a linguist) by means of automatic “reversing” of the original description of the generative process.

Figure IV.17. Practical application of the generative idea.

This program can answer the questions: What signs would it generatethat have the given Text as the signifier? What signs would it generate that have the given Meaning as signified? (See Figure IV.17.)

The parser does not really try to generate any signs, but instead solvessuch an equation using the data structures and algorithms quite different fromthe original description of the generating process.

The result produced by such a black box is, however, exactly the same:given a Text, the parser finds such Meaning that the corresponding sign belongsto the given language, i.e., would be generated by the imaginable generation algorithm. However, the description of the imaginable generation process is much clearer than the description of the internal structures automatically built by the parser for the practical applications.

Meaning Û Text idea. As any other grammar, it is aimed at the practical application in language analysis and synthesis. Unlike generative grammar, it does not concentrate on enumeration of all possible language signs, but instead on the laws of the correspondence between the Text and the Meaning in any sign of the given language. Whereas for a given text, a generative grammar can answer the question Do any signs with such Text exist, and if so, what are their Meanings?, a the mtt grammar only guarantees the answer to the question If signs with such Text existed, what would be their Meanings?

In practice, the mtt models usually can distinguish existing signs from ungrammatical ones, but mainly as a side effect. This makes the mtt models more robust in parsing.

Another idea underlying the MTT approach is that linguists are goodenough at the intuitive understanding of the correspondence between Texts andMeanings, and can describe such correspondences directly. This allows avoidingthe complications of generative grammars concerning the reversion of rules.Instead, the rules are applied to the corresponding data structures directly aswritten down by the linguist (such property of a grammar is sometimes called typetransparency [47]). Direct application of the rules greatly simplifies debugging of the grammar. In addition, the direct description of the correspondence between Text and Meaning is supposed to better suite the linguistic reality and thus results in less number of rules.

Similarly to the situation with generative grammars, there can be manyways to describe the correspondence between Text and Meaning. The specific kindof the mtt grammars suggested by I. Mel’čuk describes such a correspondence gradually, through many intermediate, half-finished almost-Meanings, half-Meanings, half-Texts, and almost-Texts, as if they were located inside the same sign between its Meaning and Text (see Figure IV.18).

Since the mtt and the generative approach developed rather independently, by accident, they use similar terms in quite different and independent meanings. Below we explain the differences in the use of some terms, though these informal explanations are not strict definitions.

· In generative grammar (see Figure IV.16):

- Transformation: a term used in early works by N. Chomsky for a specific kind of non-context-free derivation.

- Deep structure, in the transformational grammar, is a half-finished sign with a special structure to which a transformation is applied to obtain a “readier” sign. It is nearer to the initial symbol than the surface structure.

Figure IV.18. Meaning Û Text idea.

- Surface structure is a half-finished sign obtained as the result of the transformation. It is nearer to the ready sign than the deep structure.

- Generation is used roughly as a synonym of derivation, to refer to the process of enumeration of the signs in the given language.

· In the mtt (see Figure IV.18):

- Transformation is sometimes used for equative correspondences between representations on different levels.

- Deep structure concerns the representation nearer to Meaning.

- Surface structure concerns the representation nearer to Text.

- Generation (of text) is used sometimes as a synonym of synthesis, i.e., construction of Text for the given Meaning.

Constraint-based idea. Similarly to the generative grammar, a constraint-based grammar describes what signs exist in the given language, however not by means of explicit listing (generation) of all such signs, but rather by stating the conditions (constraints) each sign of the given language must satisfy.

It can be viewed as if it specified what signs do not exist in the given language: if you remove one rule (generation option) from a generative grammar, it will generate less signs. If you remove one rule (constraint) from a constraint-based grammar, it will allow more signs (i.e., allow some signs that really are ungrammatical in the given language). Hence is the name constraint-based. (See also page 44.)

Since constraint-based grammars do not use the generation process shownon Figure IV.16, their rules are applied within the same sign rather thanto obtain one sign from another (half-finished) one.

This makes it similar to the mtt. Indeed, though the constraint-based approach was originated in the generative tradition, modern constraint-based grammars such as hpsg show less and less similarities with Chomskian tradition and more and more similarity—not in the formalism but in meaningful linguistic structures—with the mtt.

A constraint-based grammar is like a system of equations. Let us consider a simple mathematical analogy.

Each sheet of this book is numbered at both sides. Consider the side witheven numbers. Looking at the page number, say, 32, you can guess that it isprinted on the 16-th sheet of the book. Let what you see be Text and what youguess be Meaning; then this page number corresponds to a “sign” <32, 16>,where we denote <T, M> a sign with the Text T and Meaning M. In order todescribe such a “language”, the three approaches would use differentmathematical constructions (of course, in a very rough analogy):

· Generative grammar is like a recurrent formula: The sign <2, 1> (analogue of the initial symbol) belongs to this “language”, and if <x, y> belongs to it, then <x + 2, y + 1> belongs to it (analogue of a generation rule). Note that some effort is needed to figure out from this description how to find a sheet number by a page number.

· The mtt grammar is like an algorithm: given the page number x, its sheet number is calculated as x/2; given a sheet number y, its page number is calculated as 2 ´ y. Note that we have made no attempt to describe dealing with, or excluding of, odd page numbers x, which in fact do not belong to our “language.”

· Constraint-based grammar is like an equation or system of equations.  Just those signs belong to our “language,” for which x = 2y. Note that this description is the most elegant and simple, completely and accurately describes our “language,” and requires less reversing effort for practical application than the first one. However, it is more complex than the second one.

Constraint-based idea is a very promising approach adopted by the majorityof contemporaneous grammar formalisms. Probably with time, the linguisticfindings of the mtt will be re-formulated in the form of constraint-based rules, possibly by a kind of merging of linguistic heritage of the mtt and formalisms developed in frame of hpsg. However, for the time being we consider the mtt more mature and thus richer in detailed description of a vast variety of linguistic phenomena. In addition, this approach is most directly applicable, i.e., it does not need any reversing.

As to the practical implementation of hpsg parsers, it is still an ongoing effort at present.


The definition of language has been suggested as a transformer betweenthe two equivalent representations of information, the Text, i.e., the surfacetextual representation, and the Meaning, i.e., the deep semantic representation. This transformation is ambiguous in both directions: a homonymous Text corresponds to several different Meanings, and several synonymous Texts correspond to the same Meaning.

The description of the transformation process is greatly simplified byintroducing intermediate levels of information representation, of which themain are morphologic and syntactic. At each level, some of the problems arisingfrom synonymy and homonymy can be solved.

The general definitions of linguistic sign in Meaning Û Text Theory and in Head-driven Phrase Structure Grammar turned out to be in essence equivalent.

V.  Linguistic Models

Throughout the previous chapters, you have learned, on the one hand, that for many computer applications, detailed linguistic knowledge is necessary and, on the other hand, that natural language has a sophisticated structure, which is not easy to represent.

Thus, any application needs a description of language, i.e., theknowledge about its main properties. Such knowledge is organized in a model of language. The structure and degree of detail depend on the application’s needs.

Our objectives now are to discuss the problem of modeling incomputational linguistics. We observe the modeling in general, describe shortlythe neurolinguistic and psycholinguistic models, and then discuss thefunctional models of natural language, with a special emphasis on common featuresof these models.

What is modeling in general?

In natural sciences, we usually consider the system A to be a model of the system B if A is similar to B in some important properties and exhibits somewhat similar behavior in similar circumstances. According to this definition, it is unnecessary the physical nature of the modeling system A be the same as of the modeled system B. For example, B can be a technological aggregation in chemistry, while A can be a set of differential equations, i.e., a conceptual system of a mathematical nature. A mathematical model is usually the best if it really ensures sufficient approximation.

In the linguistic domain, we need a modeling system that, after receivinginformation of a linguistic nature in its input, exhibits in its output anacceptable degree of similarity to the results of natural language activity ofthe human brain. Such a definition follows the idea of Alan Turing mentionedabove.

Neurolinguistic models

Neurolinguistic models investigate the links between any external speech activity of human beings and the corresponding electrical and humoral activities of nerves in their brain.

It would be interesting to investigate in detail what part of the brainis activated when a human is preparing and producing an utterance or is tryingto understand an utterance just heard. Unfortunately, the problem to discoverthe way people think while speaking or understanding is tremendously difficult.Indeed, the unique objective way for a researcher to reveal the ways of thehuman thinking by neurophysiological methods is to synchronously investigateelectrical and humoral activities in a multiplicity of places in the humanbrain.

The brain of any human consists of a tremendous number of neural cells,or neurons. The neurons are not elementary units by themselves, since each ofthem is connected with hundreds of other neurons by their dendrites and axonsin a very complicated manner.

The dimensions of neurons, dendrites, and axons are so small that thetechniques of modern neurophysiology provide no practical opportunity toobserve each neuron separately. What is more, the intrusion of observing toolsinside the brain evidently changes the activity under observation. Tounderstand in detail the internal functioning of the brain, it would benecessary to observe vast groups of neurons activated in the multistage processand trace back a grandiose multiplicity of causal chains.

Then, after this tracing, the researcher would be able to express thedetected traces as rules and algorithms presumably used by the human brain.This is what might be called a neurolinguistic model of language processing.However, the extremely large number of possible combinations of signals tracedin such a way is well beyond the capability of any computer to handle.

Leaving aside the inadequacy of the modern neurophysical techniques, themathematical tools for such a modeling are insufficient either. We cannot evenhope to get close to the information activity of neuron groups in our brainwith rather simple mathematical models. According to the modern point of view,the neurons are complicated elements of logical type. Thus, a modeling systemshould contain elements of the switching circuitry. However, neither deterministicnor stochastic theory of approximation by such a circuitry has been developedwell nowadays.

About 30 years ago, neural networks were proposed as a tool of artificial intelligence research. They consist of elements referred to as formal neurons. These are standard logical computational units with rather limited abilities. There have been numerous studies attempting to apply neural networks to various tasks of artificial intelligence and particularly to language processing. If there are some statistical trends in phenomena under observation and modeling, these trends can be recognized and taken into account by this technique.

However, for computational linguistics the problem of how to choice theinput and output signals for modeling of the brain activity remains unclear. Ifwe propose some specific inner representations for these purposes, then theformal neurons will only model our inventions rather than real languageprocesses.

For this reason, without revolutionary changes in the research techniquesand new approaches for treatment of observable data, neurolinguistic models ofnatural language understanding are unlikely to give good results in the nearest future. At present, only very crude features of brain activity can be observed effectively, such as which areas within the brain show neuronal activity associated with human memory, general reasoning ability, and so forth.

Psycholinguistic models

Psycholinguistics is a science investigating the speech activity of humans, including perception and forming of utterances, via psychological methods. After creating its hypotheses and models, psycholinguistics tests them through psychological experiments. Therefore, psycholinguistics is similar to linguistics in its objects of investigation and similar to psychology in its methods.

Here is an example of a psycholinguistic experiment. The subject, i.e.,the person under observation, is given, one by one, a sequence of speciallyselected words as input stimuli. Then the subject is prompted to give the responseto each word as any other word first coming to his or her mind. The pairs“stimulus—response” are recorded: rosa—roja, padre—madre, mano—pie, etc. Based on such experiments, psycholinguists put forward a hypothesis that various psychological types of people demonstrate specific types of associative choices, so that they can give a true identification of the personality under investigation based exclusively on such tests.

In another experiment, the objectives can be nearer to linguistics assuch. The series of utterances with some syntactic ambiguities are given to thehuman subjects. The time required to the person under test for disambiguationis measured, when he or she selects at last one of possible hypotheses. On thegrounds of experimental statistics, a hypothesis is advanced how a human canunderstand such constructions and what is the humans’ way for the disambiguation.

Psycholinguistics tries also to describe the teaching of native and notnative language, social influence of a speech on humans, and so forth. In sucha way, psycholinguistics aims to explain some purely psychological features ofpeople concerning their speech behavior. Psychological features, in their turn,are closely related to the social behavior of a person. Two kinds ofactivities, both of very high complexity, are studied in parallel, and eachaspect of the science exploits techniques and suggestions from the other.

Therefore, psycholinguistics usually does not have linguistic ideas ofits own. It adopts them from various branches of linguistics and then uses themfor its own purposes without much critical review and feedback. Purelypsychological methods well adapted to the problems of linguistics as such havenot yet been proposed, so that these methods give little to generalized (holistic)models of language.

Functional models of language

In terms of cybernetics, natural language is considered as a black box for the researcher. A black box is a device with observable input and output but with a completely unobservable inner structure. In the framework of this type of model, language is thought to be an imaginary “speaking device”: the researcher asks the device some questions and records its answers.

The problem of the cybernetic modeling of natural language is moredifficult than in other cases, since there are two such boxes, the analyzingand synthesizing ones, working in opposite directions. The analyzing blockprocesses the utterances and the synthesizing block produces the reactions tothem.

A researcher observes the input of the analyzing block and the output ofthe synthesizing block, and tries to reconstruct the inner structure of eachblock separately. Unfortunately, the output of the analyzer is not directlyused as the input of the synthesizer. There is a block of reasoning in between,and its behavior is not described in linguistic terms, so that it is not soeasy to recognize either.

The main method of linguistics is to construct a model of naturallanguage, based on the observable input and output texts, and on the linguist’sintuition, or introspection. The linguists analyze their own intuition, putforward hypotheses, build models and test them on additional linguisticmaterial. In theoretical linguistics, the novel approaches can be testedagainst (compared with) intuitions of other linguists, while in computationallinguistics these approaches can be also tested through various applications.

In this way, linguists have proposed functional models of language. These models are intended to give the rules of conversion of the input linguistic information to the output information, without any attempt to directly reproduce the internal mechanisms of brain activity. No anthropomorphic features of the processing steps are searched for, and no direct emulation of brain’s responses is adduced. However, the ultimate results of all processing steps should be as near as possible to those of the human brain.

So far, functional models have proven to be the most successfullinguistic models, probably because they are based on real data with conceivable structure, easily accessible and available in unlimited supply, namely, on texts andrecorded speech.

Research linguistic models

There are still other models of interest for linguistics. They are calledresearch models. At input, they take texts in natural language, maybe prepared or formatted in a special manner beforehand. As an output, they produce other texts, usually strictly formatted and representing the contents of dictionaries, grammar tables, rules or anything similar to be used as a part of functional models.

As an example, we can collect all the agreed pairs like “article—noun” or“noun—adjective,” or all the prepositions occurring in an open, i.e., notprepared, text in natural language. As another example, we can extract from thetext of a dictionary those items of a given part of speech, which contain a predetermined combination of features.

Thus, research models are tools for constructing functional models. Theysimulate linguists in their research, whereas the functional models simulatehumans in the speech producing and understanding.

Common features of modern models of language

The modern models of language have turned out to possess several commonfeatures that are very important for the comprehension and use of these models.One of these models is given by the Meaning Û Text Theory already mentioned. Another model is that based on the Head-Driven Phrase Structure Grammar. The Chomskian approach within the Western linguistic tradition includes various other models different from hpsg.

Here are the main common features of all these models:

· Functionality of the model. The linguistic models try to reproduce functions of language without directly reproducing the features of activity of the brain, which is the motor of human language.

· Opposition of the textual/phonetic form of language to its semantic representation. The manual [9], depicting three different well-known syntactic theories (including one of the recent variant of the theory by N. Chomsky), notices: “Language ultimately expresses a relation between sound at one end of the linguistic spectrum and meaning at the other.” Just as the diffuse notion spectrum is somehow sharpened, we have the same definition of language as in the mtt. The outer, observable form of language activity is a text, i.e., strings of phonetic symbols or letters, whereas the inner, hidden form of the same information is the meaning of this text. Language relates two these forms of the same information.

· Generalizing character of language. Separate utterances, within a speech or a text, are considered not as the language, but as samples of its functioning. The language is a theoretical generalization of the open and hence infinite set of utterances. The generalization brings in features, types, structures, levels, rules, etc., which are not directly observable. Rather these theoretical constructs are fruits of linguist's intuition and are to be repeatedly tested on new utterances and the intuition of other linguists. The generalization feature is connected with the opposition competence vs. performance in Chomskian theory and to the much earlier opposition language vs. speech in the theory of Ferdinand de Saussure.

· Dynamic character of the model. A functional model does not only propose a set of linguistic notions, but also shows (by means of rules) how these notions are used in the processing of utterances.

· Formal character of the model. A functional model is a system of rules sufficiently rigorous to be applied to any text by a person or an automaton quite formally, without intervention of the model’s author or anybody else. The application of the rules to the given text or the given meaning always produces the same result. Any part of a functional model can in principle be expressed in a strict mathematical form and thus algorithmized.[19] If no ready mathematical tool is available at present, a new tool should be created. The presupposed properties of recognizability and algorithmizability of natural language are very important for the linguistic models aimed at computer implementation.

· Non-generative character of the model. Information does not arise or generated within the model; it merely acquires a form corresponding to other linguistic level. We may thus call the correspondences between levels equative correspondences. On the contrary, in the original generative grammars by Chomsky, the strings of symbols that can be interpreted as utterances are generated from an initial symbol, which has just abstract sense of a sentence. As to transformations by Chomsky in their initial form, they may change the meaning of an utterance, and thus they were not equative correspondences.

· Independence of the model from direction of transformation. The description of a language is independent of the direction of linguistic processing. If the processing submits to some rules, these rules should be given in equative (i.e., preserving the meaning) bi-directional form, or else they should permit reversion in principle.

· Independence of algorithms from data. A description of language structures should be considered apart from algorithms using this description. Knowledge about language does not imply a specific type of algorithms. On the contrary, in many situations an algorithm implementing some rules can have numerous options. For example, the mtt describes the text level separately from the morphologic and syntactic levels of the representation of the same utterance. Nevertheless, one can imagine an algorithm of analysis that begins to construct the corresponding part of the syntactic representation just as the morphologic representation of the first word in the utterance is formed. In the cases when linguistic knowledge is presented in declarative form with the highest possible consistency, implementing algorithms proved to be rather universal, i.e., equally applicable to several languages. (Such linguistic universality has something in common with the Universal Grammar that N. Chomsky has claimed to create.) The analogous distinction between algorithms and data is used with great success in modern compilers of programming languages (cf. compiler-compilers).

· Emphasis on detailed dictionaries. The main part of the description of any language implies words of the language. Hence, dictionaries containing descriptions of separate words are considered the principal part of a rigorous language description. Only very general properties of vast classes and subclasses of lexemes are abstracted from dictionaries, in order to constitute formal grammars.

Specific features of the Meaning Û Text Model

The Meaning Û Text Model was selected for the most detailed study in these books, and it is necessary now to give a short synopsis of its specific features.

· Orientation to synthesis. With the announced equivalence of the directions of synthesis and analysis, the synthesis is considered primary and more important for linguistics. Synthesis uses the entire linguistic knowledge about the text to be produced, whereas analysis uses both purely linguistic and extralinguistic knowledge, would it be encyclopedic information about the world or information about the current situation. That is why analysis is sometimes possible on the base of a partial linguistic knowledge. This can be illustrated by the fact that we sometimes can read a paper in a nearly unknown language, if the field and subject of the paper are well known to us. (We then heavily exploit our extralinguistic knowledge.) However, text analysis is considered more important for modern applications. That is why the generative grammar approach makes special emphasis on analysis, whereas for synthesis separate theories are proposed [49]. The Meaning Û Text model admits a separate description for analysis, but postulates that it should contain the complete linguistic and any additional extralinguistic part.

· Multilevel character of the model. The model explicitly introduces an increased number of levels in language: textual, two morphologic (surface and deep), two syntactic (surface and deep), and semantic one. The representation of one level is considered equivalent to that of any other level. The equative Meaning Þ Text processor and the opposite Text Þ Meaning processor are broken into several partial modules converting data from one level to the adjacent one. Each intermediate level presents the output of one module and, at the same time, the input of another module. The division of the model in several modules must simplify rules of inter-level conversions.

· Reinforced information-preserving character. The rules of correspondence between input and output data for modules withinthe mtt fully preserve information equivalence at all language levels.

· Variety of structures and formalisms. Each module has its own rules and formalisms in the mtt, because of significant variety of structures reflecting data on different levels (strings, trees, and networks, correspondingly). On each level, the mtt considers just a minimal possible set of descriptive features. On the contrary, the generative grammar tradition tries to find some common formalism covering the whole language, so that the total multiplicity of features of various levels are considered jointly, without explicit division to different levels.

· Peculiarities in deep and surface syntactic. The entities and syntactic features of these two levels are distinctly different in the mtt. Auxiliary and functional words of a surface disappear atthe depth. Analogously, some syntactic characteristics of wordforms are presentonly at the surface (e.g., agreement features of gender and number for Spanishadjectives), whereas other features, being implied by meaning, are retained onthe deeper levels as well (e.g., number for nouns). Such separation facilitatesthe minimization of descriptive means on each level. The notions of deep andsurface syntactic levels in Chomskian theory too, but as we could already see,they are defined there in a quite different way.

· Independence between the syntactic hierarchy of words and their order in a sentence. These two aspects of a sentence, the labeled dependency trees and the word order, are supposed to be implied by different, though interconnected, factors. Formally, this leads to the systematic use of dependency grammars on the syntactic level, rather than of constituency grammars. Therefore, the basic rules of inter-level transformations turned out to be quite different in the mtt, as compared to the generative grammar. The basic advantage of dependency grammars is seen in that the links between meaningful words are retained on the semantic level, whereas for constituency grammars (with the exception of HPSG) the semantic links have to be discovered by a separate mechanism.

· Orientation to languages of a type different from English. To a certain extent, the opposition between dependency and constituency grammars is connected with different types of languages. Dependency grammars are especially appropriate for languages with free word order like Latin, Russian or Spanish, while constituency grammars suit for languages with strict word order as English. However, the mtt is suited to describe such languages as English, French, and German too. Vast experience in operations with dependency trees is accumulated in frame of the mtt, for several languages. The generative tradition (e.g.,HPSG) moves to the dependency trees too, but with some reservations and in someindirect manner.

· Means of lexical functions and synonymous variations. Just the mtt has mentioned that the great part of word combinations known in any language is produced according to their mutual lexical constraints. For example, we can say in English heart attack and cordial greetings, but neither cardiac attack nor hearty greeting, though the meaning of the lexemes to be combined permit all these combinations. Such limitations in the combinability have formed the calculus of the so-called lexical functions within the mtt. The calculus includes rules of transformation of syntactic trees containing lexical functions from one form to another. A human can convey the same meaning in many possible ways. For example, the Spanish sentence Juan me prestó ayuda is equal to Juan me ayudó. Lexical functions permit to make these conversions quite formally, thus implementing the mechanism of synonymous variations. This property plays the essential role in synthesis and has no analog in the generative tradition. When translating from one language to another, a variant realizable for a specific construction is searched in the target language among synonymous syntactic variants. Lexical functions permit to standardize semantic representation as well, diminishing the variety of labels for semantic nodes.

· Government pattern. In contradistinction to subcategorization frames of generative linguistics, government patterns in the mtt directly connect semantic and syntactic valencies of words. Not only verbs, but also other parts of speech are described in terms of government patterns. Hence, they permit to explicitly indicate how each semantic valency can be represented on the syntactic level: by a noun only, by the given preposition and a noun, by any of the given prepositions and a noun, by an infinitive, or by any other way. The word order is not fixed in government patterns. To the contrary, the subcategorization frames for verbs are usually reduced just to a list of all possible combinations of syntactic valencies, separately for each possible order in a sentence. In languages with rather free word order, the number of such frames for specific verbs can reach a few dozens, and this obscures the whole picture of semantic valencies. Additionally, the variety of sets of verbs with the same combination of subcategorization frames can be quite comparable with the total number of verbs in such languages as Spanish, French or Russian.

· Keeping traditions and terminology of classical linguistics. The mtt treats the heritage of classical linguistics much more carefully than generative computational linguistics. In its lasting development, the mtt has shown that even the increased accuracy of description and the necessity of rigorous formalisms usually permits to preserve the existing terminology, perhaps after giving more strict definitions to the terms. The notions of phoneme, morpheme, morph, grammeme, lexeme, part of speech, agreement, number, gender, tense, person, syntactic subject, syntactic object, syntactic predicate, actant, circonstant, etc., have been retained. In the frameworks of generative linguistics, the theories are sometimes constructed nearly from zero, without attempts to interpret relevant phenomena in terms already known in general linguistics. These theories sometimes ignored the notions and methods of classical linguistics, including those of structuralism. This does not always give an additional strictness. More often, this leads to terminological confusion, since specialists in the adjacent fields merely do not understand each other.

Reduced models

We can formulate the problem of selecting a good model for any specificlinguistic application as follows.

A holistic model of the language facilitates describing the language as a whole system. However, when we concentrate on the objectives of a specific application system, we can select for our purposes only that level, or those levels, of the whole language description, which are relevant and sufficient for the specific objective. Thus, we can use a reduced model for algorithmization of a specific application.

Here are some examples of the adequate choice of such a reduced description.

· If we want to build an information retrieval system based on the use of keywords that differ from each other only by their invariant parts remaining after cutting off irrelevant suffixes and endings, then no linguistic levels are necessary. All words like México, mexicanos, mexicana, etc., can be equivalent for such a system. Other relevant groups can be gobierno, gobiernos, or ciudad, ciudades, etc. Thus, we can use a list containing only the initial substrings (i.e., stems or quasi-stems) like mexic-, gobierno-, ciudad-, etc. We also will instruct the program to ignore the case of letters. Our tasks can be solved by a simple search for these substrings in the text. Thus, linguistic knowledge is reduced here to the list of substrings mentioned above.

· If we want to consider in our system the wordforms dormí, duermo, durmió, etc., or será, es, fui, era, sido, etc. as equivalent keywords, then we must introduce the morphologic level of description. This gives us a method of how to automatically reduce all these wordforms to standard forms like dormir or ser.

· If we want to distinguish in our texts those occurrences of the string México that refer to the name of the city, from the occurrences that refer to name of the state or country, then we should introduce both morphologic and syntactic levels. Indeed, only word combinations or the broader contexts of the relevant words can help us to disambiguate such word occurrences.

· In a spell checker without limitations on available memory, we can store all wordforms in the computer dictionary. Nevertheless, if the memory is limited and the language is highly inflectional, like Spanish, French or Russian, we will have to use some morphologic representation (splitting words to stems and endings) for all the relevant wordforms.

· In grammar checkers, we should take morphologic and syntactic levels, in order to check the syntactic structures of all the sentences. The semantic level usually remains unnecessary.

· For translation from one natural language to another, rather distant, language, all the linguistic levels are necessary. However, for translation between two very similar languages, only morphologic and syntactic levels may be necessary. For the case of such very “isomorphic” languages as Spanish and Portuguese, the morphologic level alone may suffice.

· If we create a very simple system of understanding of sentences with a narrow subject area, a small dictionary, and a very strict order of words, we can reduce the dictionary to the set of strings reflecting initial parts of the words actually used in such texts and directly supply them with the semantic interpretations. In this way, we entirely avoid the morphologic and syntactic problems; only the textual and the semantic levels of representation are necessary.

· If we create a more robust system of text understanding, then we should take a full model of language plus a reasoning subsystem, for the complete semantic interpretation of the text.

However, to make a reasonable choice of any practical situation, we need to know the whole model.

Do we really need linguisticmodels?

Now let us reason a little bit on whether computer scientists really needa generalizing (complete) model of language.

In modern theoretical linguistics, certain researchers study phonology,the other ones morphology, the third ones syntax, and the fourth ones semanticsand pragmatics. Within phonology, somebody became absorbed in accentuation,within semantics, in speech acts, etc. There is no limit to the subdivision ofthe great linguistic science, as well as there is seemingly no necessity tooccupy oneself once more, after ancient Greeks, Ferdinand de Saussure and Noam Chomsky, with the philosophical question “What is natural language and what should its general model be?”

The main criteria of truth in theoretical linguistic research are itslogical character, consistency, and correspondence between intuitiveconceptions about the given linguistic phenomena of the theory’s author and ofother members of linguists’ community.

In this sense, the works of modern specialists in theoretical linguisticsseem to be just stages of inner development of this science. It often seemsunnecessary to classify them according to whether they support or correspond toany complete model.

The situation in computational linguistics is somewhat different. Herethe criterion of truth is the proximity of results of functioning of a programfor processing language utterances to the ideal performance determined bymental abilities of an average speaker of the language. Since the processingprocedure, because of its complexity, should be split into several stages, a completemodel is quite necessary to recommend what formal features and structures are tobe assigned to the utterances and to the language as a whole on each stage, andhow these features should interact and participate at each stage of linguistictransformations within computer. Thus, all theoretical premises and resultsshould be given here quite explicitly and should correspond to each other intheirs structures and interfaces.

Theoreticians tell us about the rise of experimental linguistics on thisbasis. It seems that in the future, experimental tests of the deepest resultsin all “partial” linguistic theories will be an inevitable element of evolutionof this science as a whole. As to computational linguistics, the computerizedexperimentation is crucial right now, and it is directly influenced by whatstructures are selected for language description and what processing steps arerecommended by the theory.

Therefore, the seemimgly philosophical problem of linguistic modelingturned out to be primordial for computational linguistics. Two linguisticmodels selected from their vast variety will be studied in this book in more detail.

Analogy in natural languages

Analogy is the prevalence of a pattern (i.e., one rule or a small set of rules) in the formal description of some linguistic phenomena. In the simplest case, the pattern can be represented with the partially filled table like the one on the page 20:





The history of any natural language contains numerous cases when a phonologic or morphologic pattern became prevailing and, by analogy, has adduced words with similar properties.

An example of analogy in Spanish phonology is the availability of the e before the consonant combinations sp‑, st‑, sn‑, or sf‑ at the beginning of words. In Latin, the combinations sp- and st‑ at the initial position were quite habitual: specialis, spectaculum, spiritus, statua, statura, etc.

When Spanish language was developed from Vulgar Latin, all such words had been considered uneasy in their pronunciation and have been supplied with e-: especial, espectáculo, espíritu, estatua, estatura, etc. Thus, a law of “hispanicizing by analogy” was formed, according to which all words with such a phonetic peculiarity, while loaned from any foreign language, acquire the e as the initial letter.

We can compile the following table of analogy, where the right column gives Spanish words after their loaning from various languages:

statura (Lat.)


sphaira (Gr.)


slogan (Eng.)


smoking (Eng.)


standardize (Eng.)


As another example, one can observe a multiplicity of nouns ending in ‑ción in Spanish, though there exist another suffixes for the same meaning of action and/or its result: ‑miento, ‑aje, ‑azgo, ‑anza, etc. Development of Spanish in the recent centuries has produced a great number of ‑ción-words derived by analogy, so that sometimes a special effort is necessary to avoid their clustering in one sentence for better style. Such a stylistic problem has been even called cacophony.

Nevertheless, an important feature of language restricts the law ofanalogy. If the analogy generates too many homonyms, easy understanding ofspeech is hampered. In such situations, analogy is not usually applied.

A more general tendency can be also observed. Lexicon and levels ofnatural language are conceptual systems of intricately interrelated subsystems.If a feature of some subsystem has the tendency to change and this hinders thecorrect functioning of another subsystem, then two possible ways for bypassingthe trouble can be observed. First, the innovation of the initiating subsystemcan be not accepted. Second, the influenced subsystem can also change itsrules, introducing in turn its own innovations.

For example, if a metonymic change of meaning gives a new word, and thenew word frequently occurs in the same contexts as the original one, then thiscan hinder the comprehension. Hence, either the novel or the original wordshould be eliminated from language. 

In modern languages, one can see the immediate impact of analogy in thefact that the great amount of scientific, technical, and political terms iscreated according to quite a few morphologic rules. For example, the Spanishverbs automatizar, pasteurizar, globalizar, etc., are constructed coming from a noun (maybe proper name) expressing a conception (autómata, Pasteur, globo, etc.) and the suffix -izar/-alizar expressing the idea of subjection to a conception or functioning according to it.

Computational linguistics directly uses the laws of analogy in the processingof unknown words. Any online dictionary is limited in its size so that manywords already known in the language are absent in it (say, because these wordsappear in the language after the dictionary was compiled). To “understand” suchwords in some way, the program can presuppose the most common and frequentproperties.

Let us imagine, for instance, a Spanish-speaking reader who meets theword internetizarán in a text. Basing on the morphologic rules, he or she readily reconstructs the infinitive of the hypothetical verb internetizar. However, this verb is not familiar either, whereas the word Internet could be already included in his or her mental dictionary. According to the analogy implied by ‑izar, the reader thus can conclude that internetizar means ‘to make something to function on the principles of Internet.’

A natural language processor can reason just in the same way. Moreover, when such a program meets a word like linuxizar it can suppose that there exists a conception linux even if it is absent in the machine dictionary. Such supposition can suggest a very rough “comprehension” of the unknown word: ‘to make something to function on the principles of  linux,’ even if the word linux is left incomprehensible.

Empirical versus rationalistapproaches

In the recent years, the interest to empirical approach in linguistic research has livened. The empirical approach is based on numerous statistical observations gathered purely automatically. Hence, it can be called statistical approach as well. It is opposed to the rationalist approach, which requires constructing a functional model of language on the base of texts and the researcher’s intuition. Throughout this book, we explain only the rationalist approach, both in the variants of the generative grammar, and of the mtt.

The empirical approach can be illustrated more easily on the example ofthe machine translation. A huge bilingual corpus of text is being taken, i.e.,two very long, equal in the meaning, and arranged in parallel, texts in twodifferent languages. Statistics is being gathered on text fragments going in nearlyequal places on the opposite sides of the bilingual. An attempt is being madeto learn how, for any fragment in one language (including those not yet met inthe corpus), to find a fragment in other language, which is equivalent to theformer in the meaning. The solution would give the translation of any text bythe empirical method.

It can be seen that such a method unites two types of models givenabove—research and functional ones. It is also obvious that it is impossible toaccumulate the statistics in general, without elaboration of some definitionsand specifications.

It is first necessary to determine what is the size of fragments to becompared, what are “nearly equal” places and what is the equivalence (or ratherquasi-equivalence) of the fragments in the two parallel texts. Hence, theanswer to these questions requires some elements of a rationalist model, as itwas exposed everywhere above.

It is difficult to deny the need of statistical observations in computationallinguistics. In particular, in any rationalist model we should take intoaccount those linguistic phenomena (lexemes, syntactic constructions, etc.)that can be met in texts most frequently. That is why we will spare the spacein this book for statistical methods in computational linguistics.

As to the empirical method just outlined, its real successes are stillunknown.

A common feature of rationalist and empiric methods is that both of thempresuppose natural language cognizable and algorithmizable. Linguists andphilosophers suggest sometimes the opposite point of view. They argue thatsince human beings usually reason without any limitations of logic, theirlanguage activity can also lack a logical and algorithmic basis.

As applied to the natural language, this pessimistic viewpoint howevercontradicts to the everyday human practice, as well as to the practice ofmodern computational linguistics. Humans can manage a new natural language inany age and in a rather rational manner, whereas computational linguistics hasalready managed some features of natural language, and the process of masteringis going on. Thus, we may neglect this pessimistic point of view.

Limited scope of the modern linguistic theories

Even the most advanced linguistic theories cannot pretend to cover allcomputational problems, at least at present. Indeed, all of them evidently havethe following limitations:

· Only the problems of morphology and syntax are under intensive elaboration in these theories, whereas semantics is investigated to a significantly lesser degree. The goal of atomization and quite consistent decomposition of semantic components remained unattained. The more limited problem of rational decomposition of word meaning, i.e., of the semantic representation for a given lexeme through the meaning of some other more simple ones, is not yet solved on a large scale in any language. This is the great problem of lexical semantics. It develops well, but computational linguistics considers this development too slow and lacking immediate connection with computations.

· Modern semantics cannot yet formalize its problems adjacent to pragmatics to the degree sufficient for applications. Indeed, there is no complete theory of links between the meaning of text and the goals of this text in a practical situation, as well as between the speaker’s intentions and the listener’s perception, though there are numerous valuable observations on these aspects. For example, computational linguistics cannot distinguish that the Spanish utterance ¿Dónde está la sal? is a real query for information about the salt, whereas ¿Podría usted pasarme la sal? is a request for the specific action given in a polite form. As another example, no automate can “comprehend” so far that the sentence Niños son niños is not a trivial tautology, but the idea that children have specific features of their own and thus should be treated properly. To cope with such intricacies, linguists should model the human world with respect to its customs, habits, etiquette, relations between generations, etc. This is an extralinguistic knowledge of encyclopedic type. Until now, computational linguistics and artificial intelligence do not know how to effectively distinguish, to assemble apart and then to combine the knowledge of purely linguistic and evidently encyclopedic type. What is more, a “dictionary,” which specifies all encyclopedic information needed for comprehension of texts rather simple for a human to understand, would be so huge that it is unlikely to be compiled with sufficient completeness in the nearest future.

· The results of the recent investigations mainly cover separate sentences, but not discourses. The complicated semantics of discourse, including the information on referential links between different entities, a target matter under explanation, current author’s estimations, and the author’s planning of the discourse, still waits its deeper research.

· It is well known in theoretical linguistics, that the set of wordforms comprising a sentence is chosen according to the main matter of this sentence, whereas the word order depends both on this wordform set (e.g., a preposition should precede the related noun) and on communicative structure of a text. This notion reflects what the author considers already known or presupposed at this stage of the communication process (i.e. topic) and what information he or she chooses to communicate right now (i.e. comment). In generative grammars, the variations of word order depending on communicative structure of texts had not been even noticed for a long time. The mtt and general linguistics as a whole give now a rather elaborated informal study of these problems. For example, this study explains the obvious difference in meaning between Spanish sentences Juan llegó ‘Juan came’ and Llegó Juan ‘It is Juan who came,’ where the same words go in different order. As a more complicated example using the same wordforms but in different order, the sentence En México se habla el español ‘In Mexico, Spanish is spoken’ turns to be unconditionally true, while the meaning of El español se habla en México ‘Spanish is spoken in Mexico’ is quite different and dubious, since Spanish is spoken not only in Mexico. In spite of all the theoretical advances, the global formalization of communicative structures is not yet attained. So far, these advances cannot be used for either text synthesis or analysis.

· The problem of how people learn natural language in their childhood remains unsolved. The idea of linguistic universalities once introduced in general linguistics has transformed now into the idea of the Universal Grammar by Chomsky. All languages are considered species of this grammar, with a finite set of generalized features supposedly adjustable to a specific “option” (Spanish, English, etc.). Newborn children are supposed to have the Universal Grammar in their brains, and their adaptation to a specific language is accomplished at childhood. However, the goal to discover the structure and laws of the Universal Grammar remains unattained until now. Thus, computational linguistics cannot propose any universal algorithms equally applicable to various languages.

Even proponents of the contemporary linguistic theories do not believethat all facts of languages can be interpreted through their favorite ideas, tosolve current problems of computational linguistics. Meanwhile, the science advances,maybe slower than we wish.

The readers of this book will be able to learn from it a multiplicity ofalready known linguistic facts and laws. In the same time, they can realizethat numerous interesting and very difficult problems, with which computationallinguistics is faced nowadays, stay yet unsolved. They are still awaiting a Chomskyof their own.


A linguistic model is a system of data (features, types, structures,levels, etc.) and rules, which, taken together, can exhibit a “behavior”similar to that of the human brain in understanding and producing speech andtexts. A functional linguistic model takes into account the observed languagebehavior of human beings rather than the physiological activity of the brain.This behavior is reflected in the texts or speech they produce in response tothe texts or speech they perceive.

So far, the direct modeling of the brain structures has failed, andseveral functional models were proposed for the sake of computationallinguistics. The modern functional models have many features in common. Theyare intended to be quite formal, have a dynamic and non-generative character,provide independence of linguistic algorithms from linguistic data, andconsider dictionaries as one of the main, inalienable parts of the model.

Theoretical approaches provide a solid basis for both holistic and reducedmodels of language oriented to applications. The degree of the reduction insuch a model heavily depends on the specific application.


This section contains some review questions recommended to the readers to verifytheir correct understanding of the contents of the book, and the problemsrecommended for exams.

Review questions

The following questions can be used to check whether the reader has understood and remembered the main contents of the book. The questions are also recommended for the exam on this course of Computational Linguistics The questions marked with the sign ° are the most important ones.

1. Why is automatic processing of natural language important for the humankind?

2. Why are theoretical aspects of linguistics necessary for computational linguistics?

3. How are related the methods of computational linguistics and of artificial intelligence?

4. How is coordinated computational linguistics with computer science?

5. What is general linguistics?

6. What aspects of natural language do phonology, morphology, syntax, semantic, and pragmatic study?

7. What is historical linguistics? Contrastive linguistics? Sociolinguistics?

8. What is dialectology? What is Mexican Spanish with respect to Spanish language?

9. What is lexicography? Why is it important for nl processing?

10. What are the narrower and the broader comprehension of mathematical linguistics?

11. What is computational linguistics? How is it related with applied linguistics?


12. What is the structuralist approach in general linguistics?

13. What are constituents? What is constituency tree?

14. What mathematical means were proposed by Noam Chomsky? What purposes can they serve for?

15. What example of context-free grammar to generate simple sentences do you know?

16. What are transformation grammars?

17. What are valencies in linguistics? What is the difference between syntactic and semantic valencies?

18. What are subcategorization frames and how they describe valencies of verbs?

19. What are constraints in computational linguistics?

20. What is the notion of head in Head-driven Phrase Structure Grammar?

21. What is the idea of unification in computational linguistics?

22. Why should language be viewed as a transformer?

23. Why should this transformer be considered to contain several stages of transformation?

24. What is meant by Text in the Meaning Û Text Theory?

25. What is meant by Meaning in the Meaning Û Text Theory?

26. What are the main levels of language representation? Which levels are called surface ones, and which deep one? Why?

27. What are dependency tree in computational linguistics?

28. What are the two methods of information representation on semantic level? What are the semantic labels? Are they words?

29. What are government patterns in the Meaning Û Text Theory? How they describe syntactic and semantic valencies?

30. What are the main applications and classes of applications of computational linguistics?

31. What linguistic knowledge is used in hyphenation programs? Spell checkers? Grammar checkers? Style checkers?

32. What linguistic knowledge is used in information retrieval systems? In what a way does this knowledge influence the main characteristics of information retrieval systems?

33. How can we determine automatically the theme of the document?

34. How is the linguistic knowledge used in automatic translation? What are the main stages of automatic translation? Are all of these stages always necessary?

35. What is automatic text generation?

36. What are specifics of natural language interfaces?

37. What is extraction of factual data from texts?

38. What is language understanding? What linguistic knowledge should it employ? What are the main difficulties in creation of systems for language understanding?

39. What is EuroWordNet?

40. Do optical character recognition and speech recognition require linguistic knowledge?

41. What is modeling in general?

42. What kinds of linguistic modeling do you know? What are research linguistic models used for?

43. What are functional models in linguistics? What are their common features?

44. Are the Meaning Û Text Theory and Head-driven Phrase Structure Grammar functional models?

45. What are specific features of the Meaning Û Text model?

46. What are holistic and reduced models? Is the most detailed and broad model always the better one?

47. What aspects of language are not covered by modern linguistic models?


48. Word, wordform, and lexeme, what is the difference between them? When can we use each of them?

49. What is synonymy? What kinds of synonyms exist? Can synonymy be avoided in natural language?

50. What is homonymy? What kinds of homonyms exist? Can homonymy be avoided in natural language?

51. What are metaphoric and metonymic methods of creation of new words in natural language?

52. What are the specifics of computer-based dictionaries?

53. What is analogy in linguistics? How can we use it for nl processing?

54. What is empirical approach in linguistics? To what kind of problems can it be applied?


55. What is a sign? What is a linguistic sign?

56. What is the syntactics of a linguistic sign in the Meaning Û Text Theory?

57. What is the structure of the linguistic sign in Head-driven Phrase Structure Grammar?

58. Are there any commonalities in the linguistic description between generative, Meaning Û Text, and constraint-based approaches?

Problems recommended for exams

In this section, each test question is supplied with a set of four variants of the answer, of which exactly one is correct and the others are not.

1. Why automatic natural language processing (npl) is important for the humankind?

A. Because npl takes decisions in place of humans.

B. Because npl facilitates humans to prepare, to read, and to search through many texts.

C. Because npl permits humans not to read any texts by themselves.

D. Because npl facilitates humans to use computers.

2. Why theoretical aspects of linguistics are necessary for computational linguistics?

A. Because they help to prepare good user’s manuals for products of computational linguistics.

B. Because they help to evaluate the performance of computational linguistics products.

C. Because they help to gather statistics of various language elements.

D. Because they help to comprehend general structure of languages.

3. How does computational linguistics (cl) coordinate with artificial intelligence (ai)?

A. cl is a part of ai.

B. ai is a part of cl.

C. cl does not coordinate with ai at all.

D. cl and ai have many tools in common.

4. How does computational linguistics (cl) coordinate with computer science (cs)?

A. cl is a part of cs.

B. cs is a part of cl.

C. cs is a tool for cl.

D. cl is a tool for cs.

5. What does general linguistics study?

A. Laws of orthography.

B. Laws and structures of languages.

C. Rules of good word usage.

D. Rules of good style.

6. What does phonology study?

A. Sounds of music.

B. Sounds uttered by animals.

C. Sounds forming words for their distinguishing.

D. Sounds of alarm.

7. What does morphology study?

A. How to combine words to sentences.

B. How to combine sounds or letters to words.

C. How to form abbreviations.

D. How to form composed words like rascacielos.

8. What does syntax study?

A. How to combine parts of words to words.

B. How to compose a text of paragraphs.

C. How to compose a paragraph of sentences.

D. How to combine words to phrases and sentences.

9. What does semantics study?

A. How humans think.

B. How humans code the meaning in their brains.

C. How humans perceive the outer world.

D. How human express their wishes.

10. What does historical linguistics study?

A. Biographies of eminent linguists.

B. Theories of language origin in the prehistoric times.

C. Evolution of different languages in the course of time.

D. History of development of grammars.

11. What does contrastive linguistics study?

A. Controversies between different linguistic theories.

B. Differences between various languages.

C. Antonyms like pequeño–grande ‘small–big’.

D. Similarities of various non-cognate languages in their structures.

12. What part of linguistics studies peculiarities of Spanish in the Yucatan peninsula?

A. Historical linguistics.

B. Dialectology.

C. Sociolinguistics.

D. Comparative linguistics.

13. What does lexicography study?

A. Rules of orthography.

B. Rules of good word usage.

C. Pronunciation of words.

D. Description of all words in languages.

14. What does computational linguistics study?

A. How to count words and other linguistic elements in texts.

B. How to create programs for automatic text processing.

C. How to teach a foreign language with the aid of a computer.

D. How to prepare correct text with the aid of a computer.

15. How is computational linguistics (cl) related with applied linguistics (al)?

A. cl is a part of al.

B. al is a part of cl.

C. al is equal to cl.

D. al and cl are independent branches of linguistics.

16. What are constituents in linguistic description?

A. Arbitrary parts of a sentence.

B. Contiguous groups of words in a sentence.

C. Words with the most important meaning in a sentence.

D. All words in a sentence except for auxiliary ones.

17. What does a constituency tree contain as its nodes?

A. Various words.

B. Various grammatical categories.

C. Various sentences.

D. Various word endings.

18. What mathematical means did Chomsky propose?

A. Hierarchy of generative grammars.

B. Algorithms of language analysis.

C. Normative grammars for several languages.

D. Modified version of English grammar.

19. What can transformation grammars describe?

A. How to shorten context-free sentences.

B. How to repeat context-free sentences.

C. How to transit from a context-free sentence to its negative or interrogative version.

D. How to generate context-free sentences.

20. What is valency in linguistics?

A. A label at a word.

B. A link from one word to another.

C. A prepositional phrase.

D. A part of a labeled tree.

21. What is the difference between syntactic and semantic valencies?

A. Syntactic valencies link some words into pairs, while semantic valencies link other pairs.

B. Syntactic and semantic valencies link the same pairs of words, but in opposite directions.

C. Syntactic valencies describe the links between the same words as semantic valencies, but on different levels.

D. Syntactic and semantic valencies are essentially the same.

22. What is the notion of head in Head-driven Phrase Structure Grammar?

A. The principal subconstituent.

B. The center subconstituent.

C. The leftmost subconstituent.

D. The constituent that covers all its subconstituents.

23. What is unification in computational linguistics?

A. Standardization of features of wordforms.

B. Reducing wordforms to their dictionary forms.

C. Revealing similarities of features of different wordforms and uniting feature sets.

D. Uniting structures of several sentences into a common structure.

24. What is dependency tree in computational linguistics?

A. The same as constituency tree.

B. A labeled hierarchy of immediate links between wordforms in a sentence.

C. Hierarchy of meanings represented by words in a sentence.

D. Hierarchy of features assigned to wordforms in a sentence.

25. What applications of computational linguistics are the most developed now?

A. Grammar checking.

B. Spell checking.

C. Style checking.

D. Language understanding.

26. What applications of computational linguistics are the least developed now?

A. Grammar checking.

B. Language understanding.

C. Style checking.

D. Information retrieval.

27. What linguistic knowledge is used for automatic hyphenation?

A. How to use various fonts for different words.

B. What letters are vowels and consonants.

C. How to use lowercase and uppercase letters in writing.

D. How to combine usual words and numbers in a text.

28. What linguistic knowledge is used for spell checking?

A. How to use lowercase and uppercase letters in writing.

B. What are the laws of morphologic variations for words in this language.

C. What are rules of hyphenation in this language.

D. What words can stay adjacent in a text.

29. What linguistic knowledge is sufficient for grammar checking?

A. What syntactical constructions are correct in this language.

B. What words are supposedly correct in this language.

C. What phrases are commonly used in this language.

D. What words can stay adjacent in a text.

30. What linguistic knowledge is used for style checking?

A. What punctuation marks are appropriate in such a context.

B. How to arrange parts of a text in a proper sequence.

C. What words are more appropriate in a context.

D. How to split a text to adequate parts.

31. What linguistic knowledge is used in information retrieval systems?

A. Inverse files of terms.

B. Dictionaries and thesauri, consisting of terms.

C. Quick search algorithms.

D. Keyword sets at each document.

32. What are the main characteristics of information retrieval systems?

A. Response time.

B. Recall and precision.

C. Necessary size of memory for delivery.

D. User-friendliness of the interface.

33. How can we better determine automatically the theme of the document?

A. By considering the “hottest” themes for the present moment.

B. By considering the origin of the document.

C. By considering the words, which the document uses.

D. By considering the addressee of the document.

34. What is automatic text generation?

A. Deriving a text from some formal specifications.

B. Selecting entries in a preliminary prepared list of phrases.

C. Generating phrases basing on statistics of various language elements.

D. Derivation process of some generative grammar.

35. What is extraction of factual data from a text?

A. Determining what is the time of creation and the size of the text file.

B. Determining what audience this text is oriented to.

C. Determining qualitative and quantitative features of events, persons or things, which are touched upon in the text.

D. Determining linguistic means used in the text.

36. What is language understanding by a computer?

A. Transforming text into a binary code.

B. Transforming text into a graph representation.

C. Transforming text into a form that conserves the meaning and is directly usable by purposeful automata.

D. Transforming text into a table representation.

37. What are the main difficulties in creation of systems for language understanding?

A. Modern computers are insufficiently quick to solve the problem.

B. The ways of coding of meaning in texts are very complicated and are not sufficiently investigated.

C. Modern computers have insufficiently big memory to solve the problem.

D. Syntactic analysis gives too many variants.

38. What is WordNet (EuroWordNet)?

A. A usual dictionary, but in electronic form.

B. A thesaurus with a developed network of semantic links.

C. An electronic dictionary of synonyms.

D. An electronic dictionary in which we can find the part—whole links between words.

39. What linguistic knowledge does optical character recognition require?

A. How to use lowercase and uppercase letters in writing.

B. What strings of letters are correct words in writing.

C. What are rules of hyphenation in this language.

D. What words can stay adjacent in a text.

40. What linguistic knowledge does speech recognition require?

A. What variations of intonation do exist in this language.

B. What variations of logical stress do exist in this language.

C. What sequences of sounds are correct words in speech.

D. What words can stay adjacent in a speech in this language.

41. What is natural language?

A. Principal means for expressing human thoughts.

B. Principle means for text generation.

C. Bi-directional transformer Meaning Û Text.

D. Principal means of human communication.

42. What is a model in general?

A. It is an important part of the modeled system.

B. It imitates the most important features of the modeled system.

C. It includes the modeled system as the most important part.

D. It is connected with the modeled system within a system of higher rank.

43. What is the reduced model of a language?

A. It reflects all linguistic levels, but to different degree.

B. It models linguistic levels most important for the applied system.

C. It models surface linguistic levels.

D. It models deep linguistic levels.

44. What aspect of language is the least explored by modern linguistics?

A. Morphology.

B. Syntax.

C. Phonology.

D. Semantics.

45. What is a lexeme?

A. A set of letters.

B. A string of letters.

C. A set of wordforms with the same meaning.

D. A common meaning of several wordforms.

46. What entity forms the entry in a common vocabulary?

A. A word.

B. A wordform.

C. A lexeme.

D. A morph.

47. How many word occurrences are in the sentence Yo te amo, pero tú no me contesta como yo ‘I love you but you do not return me my love’?

A. Twelve.

B. Ten.

C. Nine.

D. Seven.

48. How many wordforms are in the sentence Yo te amo, pero tú no me contesta como yo ‘I love you but you do not return me my love’?

A. Twelve.

B. Ten.

C. Nine.

D. Seven.

49. How many lexemes are there in the sentence Yo te amo, pero tú no me contestas como yo ‘I love you but you do not return me my love’?

A. Twelve.

B. Ten.

C. Nine.

D. Seven.

50. What pair of the following ones consists of synonyms?

A. escoger, optar ‘choose, opt’.

B. tener, obtener ‘have, obtain’.

C. fuerza, debilidad ‘power, weakness’.

D. árbol, manzana ‘tree, apple’.

51. What are synonyms?

A. The signifieds are different, but the signifiers are equal.

B. The signifiers are different, but the signifieds are equal.

C. The signifieds are different, and the signifiers are different.

D. The signifieds are equal, and the signifiers are equal.

52. What are homonyms?

A. The signifieds are different, but the signifiers are equal.

B. The signifiers are different, but the signifieds are equal.

C. The signifieds are different, and the signifiers are different.

D. The signifieds are equal, and the signifiers are equal.

53. By what method used in order to enrich natural languages the Spanish words escuela and teatro have acquired the meaning ‘corresponding building’?

A. By metaphor.

B. By metonymy.

C. By loaning from other language.

D. By assigning a new meaning to an old word at random.

54. How many components does a linguistic sign have?

A. One

B. Two

C. Three

D. More than three.


In this section we list recommended and additional literature for more profound study of the topic of this book, as well as bibliographic references. The reader is also encouraged to study other publications of the authors of this book, which can be found on

Recommended literature

1. Allen, J. Natural Language Understanding. The Benjamin / Cummings Publ., Amsterdam, Bonn, Sidney, Singapore, Tokyo, Madrid, 1995.

2. Cortés García, U., J. Béjar Alonso, A. Moreno Ribas. Inteligencia Artificial. Edicions upc, Barcelona, 1993.

3. Grishman, R. Computational linguistics. An introduction. Cambridge University Press, 1986.

4. Jurafsky, D., J. H. Martin. Speech and Language Processing: An Introduction to Natural Language Processing, Computational Linguistics, and Speech Recognition. Prentice-Hall, 2000; see

5. Manning, C., H. Schütze, Foundations of Statistical Natural Language Processing, MIT Press, 1999;

6. Mel’čuk, I. A. Dependency Syntax: Theory and Practice. State University of New York Press, NY, 1988.

7. Pollard, C., and I. A. Sag. Head-driven Phrase Structure grammar. csli Publ., Stanford. University of Chicago Press, Chicago and London, 1994.

8. Sag, I. A., and T. Wasow. Syntactic theory: Formal Introduction. csli Publ., Stanford University of Chicago Press, Chicago and London, 1999; see also

9. Sell, P. Lectures on Contemporary Syntactic Theories. csli Publ., Stanford, 1985.

Additional literature

 10. Baeza-Yates, R., B. Ribeiro-Neto. Modern Information Retrieval. Addison Wesley Longman and ACM Press, 1999.

 11. Beristáin, Helena. Gramática estructural de la lengua española. UNAM. Limusa Noriega Editores. México, 1996.

 12. Chomsky, N. Syntactic Structures. The Hague: Mouton, 1957.

 13. Fillmore, C. J. The case for case. In: Bach, E., and B.T. Halms (eds.), Universals in linguistic theory. NY, Chicago, San Francisco, 1968

 14. Fuentes, J. L. Gramática moderna de la lengua española. Biográfica Internacional, Colombia, 1988.

 15. Gartz, Irene. Análisis de las estructuras del español. Editorial Trillas, México, 1991.

 16. Leroy, M. Las grandes corrientes de la lingüística, Fondo de Cultura Económica, 1982.

 17. Mel’čuk, I. A. Dependency in Linguistic Description. Benjamin Publ. Company, 1999.

 18. Saussure, Ferdinand de. Curso de lingüística general. Editorial Fontamara, México, 1996.

 19. Steele, J. (ed). Meaning–Text Theory. Linguistics, Lexicography, and Implications. University of Ottawa Press, Ottawa, 1990.

General grammars and dictionaries

 20. Criado de Val, M. Gramática española. Madrid, 1958.

 21. Cuervo, R. J. Diccionario de construcción y régimen de la lengua castellana. Instituto de Cara y Cuervo. Bogotá, 1953.

 22. Diccionario De La Lengua. Grupo Anaya. diccionario/diccionar.htm.

 23. Diccionario de la Real Academia Española. Real Academia Española. Edición en cd-rom, Espasa-Calpe, Madrid.

 24. Diccionario Océano de sinónimos y antónimos. Océano Grupo Editorial, Barcelona, 1997.

 25. Gilli Gaya, S. Curso superior de sintaxis española. Barcelona, 1964.

 26. Lara Ramos, L. F. (dir.). Diccionario del Español Usual en México. El Colegio de México, 1996.

 27. Martínez Amador, E. Diccionario gramatical y de dudas del idioma. Barcelona, 1953.

 28. Moliner, María. Diccionario de uso del español. Edición cd-rom. Editorial Gredos, 1998.

 29. Moneva Puyol, J. M. Gramática castellana. Barcelona—Madrid—Buenos Aires—Río de Janeiro, 1936.

 30. Orozco, Elena. La gramática desencuadernada. México, Eds. El Caballito, 1985.

 31. Seco, R. Manual de gramática española. Madrid, 1968.

 32. Real Academia Española. Esbozo de una Nueva Gramática de la Lengua Española. Madrid, 1973.

 33. Real Academia Española. Ortografía de la Lengua Española. Madrid, 1999.


 34. Apresian, Yu. D. et al. Linguistic support of the system ETAP-2 (in Russian). Nauka, Moscow, Russia, 1989.

 35. Beekman, G. “Una mirada a la tecnología del mañana”. En Computación e informática hoy. Addison Wesley Iberoamericana, 1996.

 36. Benveniste, E. Problemas de lingüística general. 7a ed., México, Editorial Siglo XXI, 1978.

 37. Bolshakov, I. A. “Multifunction thesaurus for Russian word processing”. Proc. of 4th Conference on Applied Natural Language Processing, Stuttgart, October 13-15,1994. pp. 200-202.

 38. Bolshakov, I., S. N. Galicia-Haro. “Algoritmo para corregir ciertos errores ortográficos en español relacionados al acento”. Computación y Sistemas, Vol. 1, No. 2, 1997, pp. 57-62.

 39. Compendio Xcaret de la escritura jeroglífica maya descifrada por Yuri V. Knórosov. Universidad de Quintana Roo, el grupo Xcaret y el sello Vía Láctea, 1999.

 40. Inteligencia artificial. Conceptos, técnicos y aplicaciones. Marcombo. Boixareu Editores. Barcelona—México, 1987.

 41. Fellbaum, C. (ed.). WordNet as Electronic Lexical Database. mit Press, 1998.

 42. Gelbukh, A. Using a Semantic Network Dictionary in Some Tasks of Disambiguation and Translation. Technical report, Serie Roja, N 36. CIC, IPN, 1998, ISBN 970-18-1894-6, 23 pp.

 43. Galicia-Haro, S. N., A. Gelbukh, I. A. Bolshakov. “Three Mechanisms of Parser Driving for StructureDisambiguation”. Proc. CICLing-2001, Computational Linguistics and Intelligent Text Processing.Lecture Notes in Computer Science, N 2004, Springer-Verlag, 2001, pp. 190–192.

 44. Gelbukh A., G. Sidorov, and A. Guzmán-Arenas. “A Method of Describing Document Contents through Topic Selection”, Proc. of SPIRE’99, International Symposium on String Processing and Information Retrieval, Cancun, Mexico, September 22-24, 1999, IEEE Computer Society Press, pp. 73–80.

 45. Gelbukh A., G. Sidorov, and A. Guzmán-Arenas. “Use of a weighted topic hierarchy for text retrieval and classification”. In: Václav Matoušek et al. (eds.). Text, Speech and Dialogue. Proc. tsd-99. Lecture Notes in Artificial Intelligence, No. 1692, Springer-Verlag, 1999, pp. 130–135.

 46. Guzmán-Arenas A. “Finding the main themes in a Spanish document”. Journal Expert Systems with Applications, Vol. 14, No. 1/2. Jan/Feb 1998, pp. 139-148.

 47. Hausser, Ronald. Foundations of computational linguistics: man-machine communication in natural language. Springer Verlag, 1999.

 48. Lara, L.F., et al. Investigaciones lingüísticas en lexicografía. El Colegio de México, México, 1989.

 49. McKeown, Kathleen. Text generation. Cambridge University Press, Cambridge, 1985.

 50. Mel’čuk, I. A. Experience in theories of Meaning Û Text linguistic models (in Russian). Nauka, Moscow, Russia, 1974.

 51. Pustejovsky, J. The generative lexicon. mit Press. Cambridge, Massachusetts - London, 1995.

 52. Rabbit in the Moon (information about Mayan culture): www. halfmoon. org.

 53. Translation group of the Information Science Institute of the University of South California, GAZELLE. html.

 54. Tveter, D. R. The Pattern Recognition Basis of Artificial Intelligence. IEEE Computer Society. Los Alamitos, CA, 1998.

 55. Vossen, P. (ed.). EuroWordNet General Document. Version 3: www.hum.uva.ln/~ewn.


Some Spanish-oriented groups and resources

Here we present a very short list of groups working on Spanish, with their respective urls, especially the groups in Latin America. The members of the ritos network ( / Ritos2) are marked correspondingly. In addition, we give a few urls of dictionaries or directories of Spanish resources, as well as sites of general linguistic interest.

Scientific groups and associations

· AMPLN, Mexican Association for Natural Language Processing:

· SEPLN, Spanish Society for Natural Language Processing:

· Center for Computing Research (cic, Centro de Investigación en Computación), National Polytechnic Institute (ipn, Instituto Politécnico Nacional), Laboratory of Natural Language, Mexico City, Mexico [ritos]: www. cic. ipn. mx,; gelbukh @ gelbukh . com (gelbukh @ cic. ipn. mx).

· National Institute for Astrophysics, Optics and Electronics (inaoe, Instituto Nacional de Astrofísica, Óptica y Electrónica), Aurelio López López and Manuel Montes y Gómez’s group, Puebla, Mexico: www.  inaoep.  mx, cseg.  inaoep.  mx  / ~allopez, ~mmontesg; allopez @ GISC1. inaoep.  mx,

· Institute for Investigation in Applied Mathematics and Systems (iimas, Instituto de Investigación en Matemáticas Aplicadas y en Sistemas), National Autonomous University of Mexico (unam) [ritos], Luis Pineda Cortés’s group, Mexico City, Mexico:  www. iimas. unam. mx, leibniz. iimas. unam. mx / ~luis, luis @ leibniz. iimas. unam. mx.

· Benemérita Universidad Autónoima de Puebla, the group of Héctor Himénez Salazar.

· National Autonomous University of Mexico (unam) [ritos], Linguistic Engineering Group led by Gerardo Sierra Martínez:

· Center for Linguistic and Literature Studies (Centro de Estudios Lingüísticos y Literarios), El Colegio de México, Luis Fernando Lara’s group, Mexico City, Mexico: www. colmex. mx / centros / cell / default. htm, lara@colmex. mx.

· Autonomous University of Tlaxcala, Mexico; the group led by Heriberto Cuayáhuitl Portilla.

· Computer Science Department, University of Chile, Ricardo Baeza Yates’s group, Santiago de Chile, Chile: www. dcc. uchile. cl, www. dcc. uchile. cl  /  ~rbaeza,

· Department of Languages and Information Systems (lsi, Departament de Llenguatges i Sistemes Informàtics), Polytechnic University of Catalonia (upc, Universitat Politécnica de Catalunya), Horacio Rodriguez Hontoria’s group:,

· Facultad de Ingeniería, Instituto de Computación, Montevideo, Uruguay [ritos]: www. fing. edu. uy.

· Group of Data Structures and Computational Linguistics (Grupo de Estructuras de Datos y Lingüística Computacional), Spain: protos. dis. ulpgc. es.

· Research Group on Automatic Text Generation and Discourse Processing, Spain: www. ucm. es / info / atg.

· Reference Center of Linguistic Engineering (Centre de Referència en Enginyeria Lingüística): www. cesca. es / crel.

· Spanish Society of Natural Language Processing (Sociedad Española para el Procesamiento del Lenguaje Natural): gplsi. dlsi. ua. es / sepln.

· Universidad Autónoma de Madrid (uam), Facultad de Filosofía y Letras, Laboratorio de Lingüística Informática.

· Universidad Nacional de Educación a Distancia (uned), Madrid, Spain  [ritos]: www. uned. es.

· Universidad Mayor de San Simón, Cochabamba, Bolivia [ritos]: www. umss. edu. bo.

· Universidade da Coruña, A Coruña, Spain [ritos]: www., brisaboa @

· Universitat de Barcelona (UB), Toni Martí’s group.

· Pontifícia Universidade Católica do Rio Grande do Sul, Porto Alegre, Brazil [ritos]: www. pucrs. br, vera @ kriti. inf. pucrs. br.

Resources and directories of resources

· Association for Computational Linguistics:

· Archive of linguistic papers: / ~radev / u /db / acl.

· Compilers and languages (cole): coleweb. dc. fi.

· Dictionaries and encyclopedias: www. ucm. es / BUCM / cps / eng / 0411. htm.

· Homepage of Spanish language (Página de la Lengua Española): www. latintop. com/ espannol.

· Homepage of Spanish language (Página del Idioma Español): www. el-castellano. com.

· The Linguist List:

· Virtual Center Cervantes (Centro Virtual Cervantes): cvc. cervantes. es / portada. htm.

Some International Conferences

The conferences with the Proceedings published as issues as the journal Lecture Notes in Computer Science, Springer-Verlag, are marked in boldface.

· ACL (Association for Computational Linguistics), www.aclweb. org.

· CICLing (Computational Linguistics and Intelligent Text Processing),

· COLING (Computational Linguistics), see research/ilash/iccl.

· DEXA (Databases and Expert Systems Applications), www.dexa. org.

· NLDB (Applications of Natural Language to Information Systems),

· NLPKE (Natural Language and Knowledge Engineering).

· RANLP (Recent Advances in Natural Language Processing), see, for example,

· SEPLN (Spanish Society for Natural Language Processing),

· TSD (Text, Speech and Dialogue), see, for example, tsd2004.

Additional materials and resources can be found on the webpage of this book, Other publications by the authors of this book can be found at

English-Spanish dictionary of terminology

actant actuante

ambiguity ambigüedad

analysis análisis

analyzer analizador

cfg see context-free grammars

circonstant circunstante

coherence coherencia

constituency constituyencia

constituency tree árbol de constituyencia

constituent constituyente

consonant consonante

context-free grammars gramáticas libres de contexto

deep structure estructura profundo

dependency dependencia

dependency tree árbol de dependencia

dialectology dialectología

discourse discurso

Generalized Phrase Structure Grammar Gramática Generalizada de Estructura de Frase

generation generación

generative grammars gramáticas generativas

government pattern patrón de rección (régimen)

gpsg see Generalized Phrase Structure Grammar

grammar checking revisión de gramática

grammatic case caso gramatical

Head-driven Phrase Structure Grammar Gramática de Estructura de Frase Manejada por Núcleo

historic linguistics lingüística histórica

homonym homónimo

homonymy homonimía

hpsg see Head-driven Phrase Structure Grammar

hyphenation división de palabras con guiones

information retrieval recuperación de información

irs, information retrieval system sistema de recuperación de información

interpretation interpretación

level of representation nivel de representación

lexeme lexema

lexicography lexicografía

linear structure of text estructura lineal de texto

linguistic sign signo lingüístico

logical predicate predicado lógico

main topic of a document tema principal del documento

mathematical linguistics lingüística matemática

mathematical logic lógica matemática

meaning significado

Meaning Û Text model modelo Significado Û Texto

Meaning Û Text Theory Teoría Significado Û Texto

morph morfo

morphology morfología

morphosyntactic morfosintáctico

mtt see Meaning Û Text Theory

natural language interface interfaz del lenguaje natural

natural language understanding comprensión de lenguaje natural

nested structure estructura anidada de texto

noun sustantivo

noun phrase sintagma nominal

onomatopoeia onomatopeya

optical character recognition reconocimiento óptico de caracteres

oriented labeled graph grafo orientado (dirigido) con etiquetas

orthographic error error ortográfico

parser  analizador sintáctico (párser)

part of speech categoría gramatical

phonetic alphabet alfabeto fonético

phonetic symbol símbolo fonético

phonetic transcription transcripción fonética

phonology fonología

phrase frase, sintagma

phrase structure estructura de frase

polysemic polisemántico

polysemy polisemia

pragmatics pragmática

psycholinguistics psicolingüística

reference referencia

search engine motor de búsqueda

semantic network red semántica

semantic representation representación semántica

semantics semántica

seme sema

semiotics semiótica

sides of linguistic sign lados del signo lingüístico

sign signo

signified significado

signifier significante

sociolinguistcs lingüística sociológica

speech recognition reconocimiento del habla

spelling error  error de mecanografía, de ortografía

spell checking revisión ortográfica

style checking revisión estilística

sublanguage sublenguaje

synonym sinónimo

synonymy sinonimia

syntactic analyzer analizador sintáctico

syntactic predicate predicado sintáctico

syntactic structure estructura sintáctica

syntactics sintáctica

syntax sintaxis

synthesis síntesis

synthesizer sintetizador

term término

text texto

text preparation preparación del texto

transformational grammars gramáticas transformacionales

translation traducción

unification unificación

typographic error see spelling error

valency valencia

vowel vocal

wordform forma (gramatical) de la palabra

Index of illustrations

Figure I.1. Structure of linguistic science.                                         18

Figure I.2. The ancient Mayan writing system was deciphered with statistical methods. 24

Figure II.1. Example of constituency tree.                                          37

Figure II.2. Example of a dependency tree.                                       49

Figure III.1. Alternatives for the word *caió.                                     57

Figure III.2. CrossLexica™, a dictionary of word combinations.       62

Figure III.3. Classifier program determines the main topics of a document. 67

Figure III.4. TextAnalyst program reveals the relationships between words. 69

Figure III.5. One of commercial translators.                                      72

Figure III.6. The image of a text, as the computer sees it.                  79

Figure III.7. Several letters of the same text, as the computer sees them. 79

Figure IV.1. The role of language in human communication.            84

Figure IV.2. Language functions like encoder / decoder in a communication channel. 85

Figure IV.3. Language as a Meaning Û Text transformer.               86

Figure IV.4. Meaning Û Text many-to-many mapping.                     87

Figure IV.5. Metaphor of surface and deep structures.                     88

Figure IV.6. Two levels of representation.                                         89

Figure IV.7. Structure of an application system with a natural language interface. 95

Figure IV.8. Semantic network for the sentence  Las niñas pequeñas ven la flor roja. 97

Figure IV.9. Decomposition of the verb MATAR into semes.           101

Figure IV.10. Levels of linguistic representation.                             110

Figure IV.11. Stages of transformation.                                           111

Figure IV.12. Interlevel processing.                                                 112

Figure IV.13. The role of dictionaries and grammars in linguistic transformations. 113

Figure IV.14. Translation as multistage transformation.                  115

Figure IV.15. Syntactics of a linguistic sign.                                    118

Figure IV.16. Generative idea.                                                         121

Figure IV.17. Practical application of the generative idea.             122

Figure IV.18. Meaning Û Text idea.                                               124

Index of authors, systems, andterminology


actant: 40, 41, 48, 101, 141

ai: See artificial intelligence

ambiguity: 23, 73, 105

resolution: 23, 105

analogy: 20, 145

analysis: 77, 90

analysis through synthesis: 75

analyzer: 90

Anaya dictionary: 108

Apresian, Yu.: 7, 47, 73

article: 20, 26, 37, 46

artificial intelligence: 11, 77, 131

automatic abstracting: 65


Beristáin, H.: 31

black box: 133

Bloomfield, L.: 34


cacophony: 146

cfg: See context-free grammar


grammar: 54, 58

spell: 53, 55

style: 54, 60

Chomsky, N.: 8, 34, 38, 52, 120, 136, 144

circonstant: 41, 101, 141

Clasitex: 67, 68

Classifier: 67

coherence: 93

communicative structure: 150

compression of text: 69

computer dictionary: 111

computer grammar: 111

computer science: 10

Conceptual Graphs: 98

connectivity: 93

consonant: 26, 55, 91

constituency approach: 34, 50

constituency tree: 36, 49, 50

constituent: 34

constraints: 42, 125

context-free grammar: 35

CrossLexica™: 62


daughter: 44

deduced element: 51

deep structure: 51, 124

dependency approach: 50

dependency tree: 49

derivation: 121

descriptivism: 10

diachrony: 19

dialectology: 22


importance: 137

morphologic: 112

semantic: 112

syntactic: 112

directed labeled graph: 97

discourse: 74, 92, 93, 150

drae dictionary: 108


elliptical question: 75

empirical approach: 147

encyclopedic information: 137, 150

equative correspondences: 136


orthographic: 53, 56

typographic: 53, 55

EuroWordNet: 63

extraction of data: 11, 54, 75


family of languages

Germanic: 19

Romance: 19

Slavonic: 19, 21

feature structure: 118

Filmore, C.: 42

Fuentes, J.L.: 31

fundamental science: 28


Gartz, I.: 31

Generalized Phrase Structure Grammar: 44

generation: 54, 76, 77, 124, 125

generative grammar: 8, 23, 35, 36, 120, 136, 148

glyphs: 91

government pattern: 47, 140

gpsg: See Generalized Phrase Structure Grammar

grammar checking: 54, 58

grammatical case: 21, 27

grammatical category: 35


handwriting recognition: 80

head: 44

head daughter: See head

head principle: 45

Head-driven Phrase Structure Grammar: 8, 12, 44, 118, 120, 125

hieroglyphs: 91

holonymy: 63

homonyms: 104, 116

homonymy: 104

lexical: 105

lexico-morphologic: 104

morphologic: 105

morpho-syntactic: 105

hpsg: See Head-driven Phrase Structure Grammar

hyperonymy: 63

hypertext: 70

hyphenation: 53, 54

hyponymy: 63


information retrieval: 54, 63

information retrieval system: 63

IRS: See Information Retrieval System


Knorozov, Yu.: 23

Knowledge Interchange Format: 98



ergative: 21

inflectional: 21

language generation: 125

Lara Ramos, L.F.: 30, 174

letter: 93

letters: 91

levels of representation: 110

deep: 89

surface: 89

lexeme: 27, 101, 104, 116, 141

polysemic: 106

lexicography: 22

linear structure of text: 93

linguistic typology: See contrastive linguistics

linguistic universalities: 151

linguistics: 17

applied: 23

comparative: See historical linguistics

computational: 24, 25

contrastive: 20

engineering: 24

general: 17

historical: 19

mathematical: 23, 38

quantitative: 23

statistical: 23

loan word: 20

López López, A.: 173


main topic of a document: 66

mathematical logic: 37, 51, 96

meaning: 84

Meaning: 94

Meaning Û Text Theory: 7, 47, 49, 73, 86, 120, 122, 134

Mel’čuk, I.A.: 7, 13, 47, 52, 86, 120, 123

Melchuk, I.A: See Mel’čuk, I.A.

merhonymy: 63

metaphor: 107

metonymy: 107

model: 129

functional: 133

holistic: 141

Meaning Û Text: 47

neurolinguistic: 130

reduced: 142

research: 134

morph: 92, 116, 141

morphology: 6, 18, 42, 47, 57, 74, 111

mtt: See Meaning Û Text Theory


natural language

interface: 54, 73

processing: 25, 29, 98, 147

understanding: 54, 77, 131

nested structure: 36, 37, 39

neural networks: 131

noise: 66

non-uniqueness: 102, 103


onomatopoeia: 119

optical character recognition: 54, 78


parser: 122

part of speech: 20, 47, 134, 141

phonetic alphabet: 90

phonetic symbols: 90

phonetic transcription: 90

phonology: 17, 80

phrase: 34, 40, 42, 92

phrase structure approach: 34

Pineda Cortés, L.A.: 30, 173

Pollard, C.: 8

polysemy: 106

PowerTranslator: 71

pragmatics: 18, 74, 149

precision: 65


logical: 96, 98, 101, 118

syntactic: 21, 42, 44, 141

prescriptivism: 10

primitive: 100

principle: 45

production rule: 35

productivity: 107

psycholinguistics: 22, 131


rationalist approach: 147

recall: 65

references: 61

relationships: 95

relevance: 64


morphologic: 113, 143

semantic: 50, 51, 89, 96, 98, 99, 100, 101, 114, 127

syntactic: 51, 114, 139

restored element: See deduced element

Rodriguez Hontoria, H.: 174


Sag, I.: 8

Saussure, F. de: 19, 34, 116, 144

semantic network: 98

semantic role: 98

semantics: 18, 30, 40, 41, 45, 47, 50, 54, 74, 77, 89, 96, 99, 106, 107, 112, 127, 138, 139, 143, 149

seme: 100

semiotics: 116

sign: 116

linguistic: 116, 117, 118

signified: 116

signifier: 116

sociolinguistics: 22

speech recognition: 54, 80

spell checking: 53, 55

statistical approach: 147

structuralism: 34

style checking: 54, 60

subcategorization frame: 40, 41, 42, 49, 140

sublanguage: 22, 74

surface element: 51

surface structure: 124, 125

syllabic writing: 91

synchrony: 19

synonyms: 102, 116

synonymy: 102

absolute: 102

inclusive: 103

partial: 103

synset: 63

syntactic object: 21

syntactic structure: 36

syntactic subject: 21

syntactics: 117

syntax: 18

synthesis: 90

synthesizer: 90


term: 97

Tesnière, L.: 50

Text: 90

text preparation: 53

TextAnalyst: 68

topical summarization: 67

transformation: 38, 51, 124

transformational grammar: 38

translation: 54, 70, 113

type transparency: 123

typo: See typographic error


underspecification: 46

unification: 45

Universal Grammar: 151


valency: 40, 47, 101

semantic: 41, 42, 140

syntactic: 140, 141

valency filler: 40

vocable: 106

vowel: 26, 55, 91

Vulgar Latin: 145


word: 26, 28

word occurrence: 27, 142

wordform: 27, 104

WordNet: 63


Zholkovskij, A.: See Žolkovsky, A.

Zholkovsky, A.: See Žolkovsky, A.

Žolkovsky, A.: 7

zone of a dictionary entry: 112

morphologic: 112

semantic: 112

syntactic: 112

Can computers meaningfully process human language? If this is difficult, why? If this is possible, how? This book introduces the reader to the fascinating science of computational linguistics and automatic natural language processing, which combines linguistics and artificial intelligence.

The main part of the book is devoted to the explanation of the innerworking of a linguistic processor, a software module in charge of translatingnatural language input into a representation directly usable traditional artificialintelligence applications and, vice versa, of translating their answer intohuman language.

Overall emphasis in the book is made on a well-elaborated, though—for anumber of historical reasons—so far little-known in the literaturecomputational linguistic model called Meaning Û Text Theory. For comparison, other models and formalismsare considered in detail.

The book is mainly oriented to researchers and students interested in applicationsof natural language processing techniques to Spanish language. In particular,most of the examples given in the book deal with Spanish languagematerial—which is a feature of the book distinguishing it from other books onnatural language processing. However, our main exposition is sufficientlygeneral to be applicable to a wide range of languages.

Specifically, it was taken into account that many readers of the bookwill be Spanish native speakers. For them, some comments on the Englishterminology, as well as a short English-Spanish dictionary of technical termsused in the book, were included. Still, reading the book in English will helpSpanish-speaking readers to become familiar with the style and terminology usedin the scientific literature on the subject.


was born in Moscow, Russia, in 1934. He obtained his M.Sc. degree in physics in 1956 from the Department of physics of the Moscow State “Lomonossov” University, Ph.D. degree in information technology in 1961 from the VYMPEL Institute, Moscow, Russia, and D.Sc. degree in computer science in 1966 from the same institute. He received the National Award of USSR in Science and Technology in 1989. Since 1996, he works for the Natural Language and Text Processing Laboratory of the Computing Research Center, National Polytechnic Institute, Mexico City. He is National Researcher of Mexico of excellence level III, author of more than 200 publications on theory of radars, theory of probability, and computational linguistics.



was born in Moscow, Russia, in 1962. He obtained his M.Sc. degree in mathematics in 1990 from the Department of mechanics and mathematics of the Moscow State “Lomonossov” University and Ph.D. degree in computer science in 1995 from the All-Russian Institute for Scientific and Technical Information. Since 1997 he is the head of the Natural Language and Text Processing Laboratory of the Computing Research Center, National Polytechnic Institute, Mexico City. He is academician of the Mexican Academy of Sciences, National Researcher of Mexico of excellence level I, distinguished lecturer of the ACM, founder of the Mexican Association for Natural Language Processing and the CICLing international conference series, author of more than 250 publications on computational linguistics. Currently he is Distinguished Visiting Professor at Chung-Ang University, Seoul, Korea.


[1] The term programmed is sometimes used for such manuals, this term being related to the structure of the manual rather than to a computer program.

[2] Guiones in the Spanish version.

[3] In all cases, an initial asterisk marks strings containing errors.

[4] According to the tests conducted by the publishing house of the Russian edition of PC World.

[5] In the variant of Spanish spoken in Spain. Consider also como vosotros queréis vs. como vosotras queréis.

[6] We use capitalization to distinguish the terms Meaning and Text in their specific sense used in the Meaning Û Text Theory from the conventional usage of these words.

[7] For the sake of the argument, here we consider speech, as well as written text, to be a kind of Text.

[8] It is not surprising that later in this book, levels of information representation will appear, in direct analogy with the modern computer network protocols.

[9] See also discussion of the terms surface and deep in comparison with their usage in generative grammars on the page 132.

[10] In fact, Japanese language uses a mixture of hieroglyphic and syllabic symbols, though the use of syllabic symbols is limited.

[11] This does not mean that the discussed methods are not applicable to phonetic transcription or, on the other hand, to hieroglyphs. However, just for simplification we will choose only the representation by letters.

[12] In some writing systems, like Japanese, words are not separated by spaces in the written text. Of course, this does not mean that these languages do not have words, but the word boundaries are not reflected in writing. As opposed to Japanese, Vietnamese separates all the syllables.

[13] In Western tradition including hpsg, Text in the given sense is called list of phoneme strings, or simply phonetic representation of a linguistic sign.

[14] And also compression, which increases the non-linearity of the underlying structure.

[15] Within the computer community, efforts are under way to develop knowledge representation both in a linear format (kif = Knowledge Interchange Format), and in a graphical format (cg = Conceptual Graphs).

[16] In Russian: Malen’kie devochki vidjat krasnyj cvetok; the two last words are singular masculine.

[17] Some researchers consider the predicate TO LIVE(x) elementary.

[18] This is shown by the three bold arrows in Figure IV.4.

[19] Formality of any model is quite natural from the programming point of view.

This is a plain text version of the ORIGINAL FILE. Please see that file instead.