Preface

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 computational linguistics

The success of modern software for natural language processing impresses our imagination. Programs for orthography and grammar correction, information retrieval from document databases, and translation from one natural language into 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 automatic translation and man-machine communication in unrestricted natural language, has not yet been achieved, though attempts to solve this problem already have a history of nearly 50 years.

This suggests that in order to solve the problem, developers of new software will need to use the methods and results of a fundamental science, in this 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 be solved in the beginning of this century by those who are sitting now on student benches. This book on computational linguistics models and their applications is targeted at these students, namely, at those students of the Latin American universities studying computer science and technology who are interested in the development of natural language processing software.

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

On the modern book market, there are many texts on Natural Language Processing (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 quickly developing discipline, the main methodological features of this book on computational linguistics 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 under a single cover. We also have rejected the idea to make our manual a reference book, 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 the following:

· 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 on our observations on the sources of the problems which we ourselves and our colleagues encountered while starting our careers in computational linguistics and which still trouble many programmers working in this field due to the lack of fundamental 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 computational linguistics that has been delivered by the authors since 1997 at the Center for Computing Research, National Polytechnic Institute, Mexico City. This course was focused on the basic set of ideas and facts from the fundamental science necessary for the creation of intelligent language processing tools, without going deeply into the details of specific algorithms or toy systems. The practical study of algorithms, architectures, and maintenance of real-world applied linguistic systems may be the topics of other courses.

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

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

The authors can be contacted at the following e-mail addresses: igor*cic.ipn.mx, gelbukh*gelbukh.com (gelbukh*cic.ipn.mx); see also www.Gelbukh.com (www.cic.ipn.mx/~gelbukh). The webpage for this book is www.Gelbukh.com/clbook.

Objectives and intended readers of the book

The main objectives of this book are to provide the students with few fundamentals of general linguistics, to describe the modern models of how natural languages function, and to explain how to compile the data—linguistic tables and machine dictionaries—necessary for the natural language processing systems, out of informally described facts of a natural language. Therefore, we want to teach the reader how to prepare all the necessary tools for the development of programs and systems oriented to automatic natural language processing. In order to repeat, we assume that our readers are mainly students in 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 Spanish language. However, this course is not a mere manual of Spanish. A broader basis for understanding the main principles is to be elucidated through some examples from English, French, Portuguese, and Russian. Many literature sources provide the reader with interesting examples for these languages. In our books, we provide analogous examples for Spanish wherever possible.

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

Neither it should be forgotten that the native language is studied in the school 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 or admissible language constructions from those not correct and not belonging to the language under investigation. Meantime, without a complete and correct description, computer makes errors quite unusual and illogical from a human point of view, so that the problem of text processing cannot be successfully solved.

Coordination with computer science

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

The objective of the book is to help the students in developing applied software systems and in choosing the proper models and data structures for these systems. We only reject the idea that the computer science’s tools of recent decades are sufficient for computational linguistics in theoretical aspects. Neither proper structuring of linguistic programs, nor object-oriented technology, nor specialized languages of artificial intelligence like Lisp or Prolog solve by themselves the problems of computational linguistics. All these techniques are just tools.

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

However, the significant part of the effort in the practical developing of 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 such tasks. Thus, again, we consider programming to be the everyday practical tool of 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 both fields can be more productive. Indeed, they do not need to invent things already well known in the adjacent field, just taking from the neighborhood what they prefer for their own purposes. Thus, we encourage our readers to deeply familiarize 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 are characteristic 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 www.Gelbukh.com/clbook. You can find there additional materials on the topics of this book, links to relevant Internet resources, and errata. Many publications by the authors of this book can be found at www.Gelbukh.com.

Acknowledgments

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. Galicia Haro, who performed a lot of administrative chores and paperwork, freeing our time 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 the main 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 of newspapers 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 for the needs of a developing human society. Modern natural languages are developing according to their own laws, in each epoch being an adequate tool for human communication, for expressing human feelings, thoughts, and actions. The structure and use of a natural language is based on the assumption that the participants of the conversation share a very similar experience and knowledge, as well as a manner of feeling, reasoning, and acting. The great challenge of the problem of intelligent automatic text processing is to use unrestricted natural language to exchange information with a creature of a totally different nature: the computer.

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

The necessity for intelligent automatic text processing arises mainly from the following two circumstances, both being connected with the quantity of the 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 problems in information exchange. The rapid development of computers in the last two decades has made possible the implementation of many ideas to solve the problems that one could not even imagine being solved automatically, say, 45 years ago, when the first computers appeared.

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

Linguistics and its structure

Linguistics is a science about natural languages. To be more precise, it covers 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 a whole (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 from comparative 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 each of the mentioned families—Romance, Germanic, and Slavonic—taken separately.

At the same time, it has noticed a number of similar words among these families. This finding has led to the conclusion that the mentioned families form a broader community of languages, which was called Indo-European languages. Several thousand years ago, the ancestors of the people now speaking Romance, Germanic, and Slavonic languages in Europe probably formed a common tribe or related tribes.

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

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

Spanish	English
constitución	constitution
revolución	revolution
investigación	?

Based on more complicated phonologic laws, it is possible even to predict the 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 usually only for motivation of purely synchronic models. History sometimes gives good suggestions for description of the current state of language, helping the researcher to 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, and some of them are rather similar in their functions to those of Latin. Inflected parts of speech, i.e., nouns, adjectives, participles, and pronouns, have different word endings for each case.

In English, there is only one oblique case, and it is applicable only to some 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 among linguistic applications of mathematics, we will follow the second, broader view on mathematical linguistics.

One of the branches of mathematical linguistics is quantitative linguistic. 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 applied linguistics 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 very broad 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. In our opinion, for a computer system or its part to be considered linguistic, it should 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 to linguistics. Though such word processors as Windows’ Notebook do deal with the processing of texts in natural language, we do not consider them linguistic software, since they are not sufficiently language-dependent: they can be used equally for processing of Spanish, English, or Russian texts, after some alphabetic adjustments.

Let us put another example: some word processors can hyphenate words according 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 libro ahora. 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 same core meaning. We intuitively consider them as different forms of some common entity. 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 dev...lv‑), 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., a specific member of the wordform set, is selected for this purpose. In the previous examples, 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 or language understanding systems have been undertaken by people without sufficient knowledge in theoretical linguistics. They hoped that they would succeed thanks to clever mathematical algorithms, fine programming in Assembler language, or just the speed of their computers. To our knowledge, all such attempts have failed. Even now it is still worthwhile to explain the necessity to have fundamental knowledge for those who would develop such systems, and thus to clarify why we decided to start a course in computational linguistics from 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 the structure 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 revolutionary engineering principles that did not even exist at the time when he or she studied in a university. Indeed, the underlying fundamentals of physics have remained the same.

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

The same situation can be observed with linguistics. Even experienced programmers who have taken solid courses on software systems for natural language 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 the program heavily relies on such tables and dictionaries as on its integral parts.

They will not even be able to understand the suggestions of professional linguists (who regrettably in many cases prove to know little about programming) and will not be able to explain to them the meaning of the data structures the program needs to use.

On the contrary, a good background in linguistics will allow the new specialists in computational linguistics to work productively in an interdisciplinary team, or just to compile all the necessary tables and dictionaries on their own.

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

Current state of applied research on Spanish

In our books, the stress on Spanish language is made intentionally and purposefully. For historical reasons, the majority of the literature on natural languages processing is not only written in English, but also takes English as the target language for these studies. In our opinion, this is counter-productive and thus it has become one of the causes of a lag in applied research 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 400 million, and Spanish is one of the official languages of the United Nations. As to the human-oriented way of teaching, Spanish is well described, and the Royal Academy of Madrid constantly supports orthographic [33] and grammatical research [30] and standardization. There are also several good academic-type dictionaries of Spanish, one the best of which being [28].

However, the lexicographic research reflected in these dictionaries is too human-oriented. Along with some historical information, these dictionaries provide 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 of research teams in computational linguistics. Several teams in this field oriented to Spanish work now in Barcelona and Madrid, Spain. However, even this is rather little for a country of the European Community, where unilingual and multilingual efforts are well supported by the governmental and international agencies. 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 in computational linguistics has been rather low in the past decades. Now, the team 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 in the 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, has announced the development of a natural language processing system for Spanish based on the idea of multistage processing. As usually with commercial developments, the details of the project are still rather scarce. We can only guess that a rather slow progress of the grammar checker of Word text processor for Windows is related somehow with these developments.

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

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

Conclusions

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

To develop such applications, specialists in computer science need to have adequate tools to investigate language with a view to its automatic processing. One of such tools is a deep knowledge of both computational linguistics 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 phenomena in mechanics. Later, Einstein’s theory of relativity incorporated the Newtonian theory as an extreme case, and in its turn for a long time excluded other methods of interpretation of a rather vast class of phenomena. Such exclusivity can be explained by the great power of purely mathematical description of natural phenomena in physics, where theories describe well-known facts and predict with good accuracy the other facts that have not yet been observed.

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

Let us give now a short retrospective of the development of general linguistics in the twentieth century. The reader should not be worried if he or she 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 were working in parallel. Most of them worked in Europe, and European structuralism kept a significant affinity to the research of the previous periods in its terminology and 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 linguist Noam 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 two most 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 very simple 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-terminal symbols 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 strings obtained earlier, we can construct a lot of fully grammatical and meaningful Spanish 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, and wide applicability of generative grammars. They became used not only for description 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 to facts of natural languages. At the beginning, this research stemmed from the basic 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 better accommodate 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:

Nested:

John

does see

Mary

_VP



Not nested:	Does	John	_N	see	_V	Mary	_N	?

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 natural languages proved to be rather complicated. The theory has lost its mathematical elegance, 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 within the 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 Chomsky and adopted by the newer theories. It is the subcategorization of verbs according to their ability to accept specific sets of complements. These complements 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 each valency. 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 both syntactic 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 N₂stand 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 N₂is an indirect complement.

We have induced only one of the possible subcategorization frames for the verb 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 roles stay 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 the situation is expressed by many languages in a quite standard form, namely the subject of the sentence. In Spanish, English, and several other languages (but not 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 was that 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 and then its value, e.g., “gender: Gen”. The rule above determines which values are to be equal. With each constituent, any number of features can be expressed, while different constituents within the same rule possibly can have different sets 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 the predicate 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 = 1^st, Num = singular, while in the second phrase, Pers = 3^rd, Num = plural on the both sides.

The constraint-based approach using the features was intensively used by the 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 determine features of the whole sentence, and the features of the noun determine the corresponding features of the whole noun phrase. Such formalism permits to easier specify the syntactic structure of sentences and thus facilitates syntactic analysis (parsing).

As it was already said, the interpretation in early generative grammars was 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 of natural languages.

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

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

As the first example of unification operations, let us compare feature sets 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 3^rd 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 some oversimplifications 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 for other 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 and syntactic valencies. For this verb, the correspondence is quite simple, but it is not so in general case.

The lower part of the table enumerates all possible options of representation for each syntactic valency at the syntactic level. The options operate with part-of-speech labels (N for a noun, V_inf 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 syntactic valencies in a government pattern. If all permutations of valencies of a specific verb are permitted in the language, then no information about word order is needed in this GP. Elsewhere, information about forbidden or permitted combinations is given explicitly, to make easier the syntactic analysis. For example, the English verb give permits only two word orders mentioned above.

Thus, government patterns are all-sufficient for language description and significantly 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 una canción 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 word agreement in many languages including Spanish can be accomplished easier by means 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 in the 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 the words constituting a sentence and of the order of these words in the sentence were separated from each other. Thus, the links between words and the order in which they appear in a sentence were proposed to be investigated apart, and relevant 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 between wordforms at the surface syntactic level determine links between corresponding labeled 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 syntactic level, directly ruling the personal name:

su ¬ hijo ® NAME ® Juan

Special rules of inter-level correspondence facilitate the transition to the 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 important for us only to claim that the mtt has been able to describe any natural language and any linguistic level in it.

Conclusions

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 satisfactory solutions for their tasks, especially for English, while other tasks and languages have 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 scientific domains for many purposes. Some of the most important ones among them are the following:

· 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 of each application. Later, some of these topics, with more deep explanations, can be touched upon once more.

Automatic hyphenation

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

The wrapping can be done only at specific positions within words, which generally, though not always, are syllable boundaries. For example, in Spanish one 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-produced texts through adjusting their right margins. It saves paper and at the same time preserves impression of smooth reading, just as without any hyphenation.

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

Usually, the linguistic information taken for such programs is rather limited. 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 detailed information about each word. The hyphenation can depend on the so-called morphemic 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 of typographic 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 the rules 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, when typing, one sometimes does not synchronize properly the movements of the hands and 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 correct words in a given natural language. It is supposed that most of the orthographic or 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 the help of the program. For example, if we try to insert into any English text the strings[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 can also 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 only one 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 orthographic correctors, while in English they are usually called spell checkers. In everyday practice, spell checkers are considered very helpful and are used by millions of users throughout the world. The majority of modern text editors are supplied now with integrated spell checkers. For example, Microsoft Word uses many spell checkers, a specific one for each natural language used in the text.

The amount of linguistic information necessary for spell checkers is much greater than for hyphenation. A simple but very resource-consuming approach operates 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 some presuppositions about the most common typographic and spelling errors. A deeper penetration into the correction problems requires a detailed knowledge of morphology, 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 quite evident tasks of correction of words, even taken separately, have not been yet solved. 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 adjacent words in the sentence or even the whole sentence are much more difficult tasks for computational linguists and software developers than just checking orthography.

Grammar errors are those violating, for example, the syntactic laws or the laws related to the structure of a sentence. In Spanish, one of these laws is the agreement between a noun and an adjective in gender and grammatical number. 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 the sentence. For example, it is rather difficult for a program to find the error in 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 the difficulty of such parsing, commercial grammar checkers are still rather primitive and rarely give the user useful assistance in the preparation of a text. 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 grammar checkers. For example, the grammar checker included in Microsoft Word is helpful but still very far from perfection.

Sometimes, rather simple operations can give helpful results by detecting some very frequent errors. The following two classes of errors specific for Spanish 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 too frequently give false alarms where there is no error in fact. For example, in the 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 that definitely knows what he or she meant to write, the final decision must always be left up to the user, whether to make a correction suggested by the grammar checker or to leave the text as it was.

Style checking

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

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

For example, you are not recommended to use any vulgar words or purely colloquial constructions in official documents. As to more formal properties of Spanish 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 their usage marks, synonyms, information on proper use of prepositions, compatibility with other words, etc. It should also use automatic parsing, which can detect improper syntactic constructions.

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

As a very primitive way to assess stylistic properties of a text, some commercial 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 of words in them; length of paragraphs, i.e., the number of words and sentences. They can also use other statistical characteristics that can be easily calculated as a combination of those mentioned.

The larger the average length of a word, sentence or paragraph, the more difficult the text is to read, according to those simplest stylistic assessments. 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 suggestions as 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 the future.

References to words and word combinations

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

Within typical text processors, the synonymy dictionaries are usually called thesauri. Later we will see that this name corresponds poorly to the synonymy dictionaries, since genuine thesauri usually include much more information, 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 natural language have the objective to help the author of a text to create more correct, flexible, and idiomatic texts. Indeed, only an insignificant part of all thinkable word combinations are really permitted in a language, so that the knowledge of the permitted and common combinations is a very important part of linguistic competence of any author. For example, a foreigner might want to know 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 automated reduction of the entered words to their dictionary forms, search of relevant words in the corresponding linguistic database, and displaying all of them in a form convenient to a non-linguist user. These operations are versatile and include both morphologic and syntactic issues [37].

Another example of a dictionary that provides a number of semantic relations 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 developed within 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 to several other European languages, in particular to Spanish. The upper levels of ontology were obtained by direct translation from English, while for the other levels, additional lexicographic research turned out to be necessary. In this way, not only links between synsets within any involved language were determined, 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 bibliographic database 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 is becoming 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 of the 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 general case, i.e. the greater one of them the lesser another, so that it is necessary to 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 of keywords given just the full text of the document. Such systems do not require the authors of the documents to specifically provide the keywords. Some of the modern Internet search engines are essentially based on this idea.

Three decades ago, the problem of automatic extraction of keywords was called 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 and performance of the indexing and comparing subsystems, on the content of the terminological system or the thesaurus, and other data and knowledge used by the system. 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 given document is about. This information is used to classify the documents by their main topics, to deliver by Internet the documents on a specific subject to the users, 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 mentioned in 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 occurred in the document. Roughly speaking, the topic mentioned most frequently is considered the main topic of the document. Actually, the topics in the dictionary have different weights of importance [43, 45], so that the main topic is the one with the greatest total weight in the document.

Applied linguistics can improve this method in many possible ways. For example, 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. implicit references 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 the same sentence. The program builds a network of the relationships between words. Figure III.4 shows the most important words found by TextAnalyst in the early draft of this book, and the network of their relationships.

As in Clasitex, the degree of importance of a word, or its weight, is determined in terms of its frequency, and the relationships between words are used to mutually increase the weights. The words closely related to many of the important 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 following tasks:

· 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 neural network 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 important task. The amount of business and scientific texts in the world is growing rapidly, and many countries are very productive in scientific and business domains, publishing numerous books and articles in their own languages. With the growth of international contacts and collaboration, the need for translation of legal contracts, technical documentation, instructions, advertisements, and other texts used in the everyday life of millions of people has become a matter of vital importance.

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

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

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

As for commercial translation software, the quality of translation it generates is still rather low. A commercial translator can be used to allow people quite unfamiliar with the original language of the document to understand its main idea. Such programs can help in manual translation of texts. However, post-editing of the results, to bring them to the degree of quality sufficient for publication, often takes more time than just manual translation made 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 translation program can make, it is enough to mention a well-known example of the mistranslation performed by one of the earliest systems in 1960s. It translated the 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 popular commercial 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 the same in the two directions, say, from English to Spanish and from Spanish to English. Since automatic analysis of the text is usually a more difficult task than 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 sense disambiguation. In any bilingual dictionary, for many source words, dozens of words in the target language are listed as translations, e.g., for simple Spanish 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 restoring the information that is contained in the source text implicitly, but which must be expressed explicitly in the target text. For example, given the Spanish text José 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 of translation. As an example of successful research, the results of the Translation group of Information Science Institute at University of South California can be mentioned [53]. This research is based on the use of statistical techniques for lexical ambiguity resolution.

Another successful team working on automatic translation is that headed by 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 to understand questions entered by a user in natural language and to provide answers—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 the queries and the set of words used in them are usually very limited. Hence, the linguistic 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 interfaces that 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 usually supply 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 attractive for many research teams all over the world. Especially useful could be natural language interfaces with speech recognition capabilities, which also would allow the user to make queries or give commands over a telephone line.

The task of development of natural language interfaces, though being less demanding 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 with a 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 this user’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 generation of elements of a factographic database, such as fields, or parameters, based on on-line texts. Often the flows of the current news from the Internet or from an information agency are used as the source of information for such systems, and the parameters of interest can be the demand for a specific type of a product in various regions, the prices of specific types of products, events involving a particular person or company, opinions about a specific issue or a political party, etc.

The decision-making officials in business and politics are usually too busy to read and comprehend all the relevant news in their available time, so that they often have to hire many news summarizers and readers or even to address to a special information agency. This is very expensive, and even in this case the important relationships between the facts may be lost, since each news summarizer typically has very limited knowledge of the subject matter. A fully effective automatic system could not only extract the relevant facts much faster, but also combine them, classify them, and investigate their interrelationships.

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, and company merge and acquisition projects. Due to the great difficulties of this task, only very large commercial corporations can afford nowadays the research on the factual data extraction problem, or merely buy the results of such research.

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

Text generation

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

Another very important application of systems of this kind is the generation of formal specifications in text form from quite formal technical drawings.

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

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

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

Better results can be achieved by automatic generation of the required text in each language independently, from the technical drawings and specifications or from a text in a specific formal language similar to a programming language.

Text generating systems have, in general, half of the linguistic problems of 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 of linguistic information, which is currently available in adequate detail for only a few very narrow subject areas.

Systems of language understanding

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

For example, automatic translation can be implemented as a text understanding system, which understands the text in the source language and then generates a precise description of the learned information in the target language.

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

To be more precise, the natural language processing module is only one part of such a system. Most activities related to logical reasoning and understanding proper are concentrated in another its part—a reasoning module. These two modules, however, are closely interrelated and they should work in tight 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 the knowledge-based reasoning engine, the problem solver, the expert system, or something else, depending on the specific functions of the whole system. Its role in the whole system of natural language understanding is always very important.

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

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

In order to repeat, the full power of linguistic science and the full range of linguistic data and knowledge are necessary to develop what we can call 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 methods to accomplish their tasks. Of these we will mention here two, both related to pattern 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 interpretación semántica: 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 that this picture represents any letters. However, one can easily read precisely the same image, shown above, of the same letters in their complete context. Hence, it is obvious that the task of optical character recognition cannot be solved only 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 the one shown in Figure III.6, some peculiar errors can still appear in the textual representation of the image. They can be fixed by the operations similar to those of a spell checker. Such a specialized spell checker should know the most frequent errors of recognition. For example, the lower-case letter 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 base of linguistic knowledge. In the simplest case, such knowledge is just a dictionary of words existing in the language. However, in some cases a deeper linguistic analysis is necessary for disambiguation. For example, only full parsing of the context sentence can allow the program to decide whether the picture recognized as *danios actually represents the existing Spanish words darnos or damos.

A much more challenging task than recognition of printed texts is handwriting recognition. 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 or quite the same as in optical character recognition: mutilated patterns, fused patterns, disjoint parts of a pattern, lost parts of the pattern, noise superimposing the pattern. This leads to even a much larger number of incorrectly recognized letters than with optical character recognition, and application of linguistic methods, generally in the same manner, is even more important for this task.

Conclusions

A short review of applied linguistic systems has shown that only very simple tasks like hyphenation or simple spell checking can be solved on a modest linguistic basis. All the other systems should employ relatively deep linguistic knowledge: dictionaries, morphologic and syntactic analyzers, and in some cases deep semantic knowledge and reasoning. What is more, nearly all of the discussed tasks, even spell checking, have to employ very deep analysis to be solved with an accuracy approaching 100%. It was also shown that most of the language processing tasks could be considered as special cases of the general task of language understanding, one of the ultimate goals of computational linguistics 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”—there is no doubt that the flow of utterances, or texts, is a very important result of functioning of natural language.

However, communication includes not only generation (speaking), but also understanding of utterances. This definition also ignores the starting point of text 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,”—we can readily agree that the communicative function is the main function of natural language.

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

Thus, these definitions are not sufficient for our purposes. A better definition 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 some information—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 their thoughts, 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 same language, which is their common knowledge, and each of them has a copy of it in the brain.

If we compare this situation with transferring the information over a communication channel, such as a computer network, the role of language is encoding the information at the transmitting end and then decoding it at the receiving 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 a transformer 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 transformation we mean some form of translation, so that both the Text and the corresponding Meaning 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 third definitions considered in the previous section. Clearly, the transformation of Text into Meaning and vice versa is obligatory for any human communication, since it implies transferring the Meaning from one person to another using the Text as its intermediate representation. The transformation of Meaning into Text is obligatory for the generation of utterances. To be more precise, in the whole process of communication of human thoughts the definition 1 given earlier actually refers to Meaning, the definition 2 to Text, and the definition 3 to both mentioned aspects of language.

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

With our present definition, language can be considered analogous to a technical 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 in opposite directions. Text is the result of the activity of one of these transformers, 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 some binary code with machine instructions, as the output. The binary code corresponds to the meaning of the program. However, a compiler usually cannot translate the machine instructions back to a C++ program text.

As a mathematical analogy to this definition, we can imagine a bi-directional mapping between one huge set, the set of all possible Texts, and another 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 immediately observed or directly transferred from one person to another, while the Meanings reside in the brain of each person independently and cannot be immediately observed or assessed.

This is similar to the front panel of an electrical device: the lights and 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 leaf is connected to what flower through common roots, we need to analyze what is under the surface. There is much more below the surface than on top of it, and only analysis of the deeper phenomena gives us understanding of the whole thing.

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 have been neither defined nor described so far. Such descriptions will be given in the following sections.

Text, what is it?

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

Speech is a continuous flow of acoustic signals, just like music or noise. However, linguistics is mainly oriented to the processing of natural language in a discrete form.

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

For representation of the sound of speech on paper, a special phonetic transcription 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 linguistics form of speech representation arose spontaneously in the human practice as the written form of speech, or the writing system.

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

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 in this 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 in alphabetic symbols are more usual for computational linguistics than the phonetic 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 language should then be thought of as common texts in their usual written form. Written texts 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 of musical 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 be additionally modulated only in a very restricted manner, e.g., by stress, intonation, etc. On the contrary, a set of musical instruments can produce several sounds synchronously, forming harmonies or several melodies going in parallel. This parallelism can be considered as nonlinear structuring. The human 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 acoustic coding 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 we mentioned above, we consider the Meaning to be the structures in the human brain which people experience as ideas and thoughts. Since we do not know and cannot precisely represent those brain processes, for practical purposes we must use a representation of Meaning, which is more suitable for manipulation in a computer. Thus, for our purposes, Meaning will be identified with that representation.

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

As it was described earlier, the task of language is to transform information from one representation, the Text, into another, the Meaning, and vice versa. A computer program that models the function of language must perform the same task. In any application, there should be some other system or device that consumes the results of the transformation of texts, and produces the information that is to be transformed into a text. The operations of such a device 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 or a format of information representation immediately understandable for, or executable on, the consumer of the information: the underlying expert or reasoning system, database, robot control system, etc. It is supposed that this underlying system produces its responses just in the same format. Thus, in practice, the format of Meaning is already given to the developers of the linguistic module for any specific 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 application system should distinguish the basic information delivered in texts from author’s beliefs, estimations, and intentions.

Additionally, even a very formal text contains many explicit references and explanations of links between parts of a text, and these elements serve as a content table or a protocol of the author’s information about text structuring. 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 text are combined together, i.e., a “text about text.” We will not discuss such insertions in this book. Since properties, processes, and actions can be represented as relationships between entities touched upon in Text, just these features are used to represent Meaning.

The entities, states, properties, processes, and actions are usually denoted 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, the following 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) &

ROJO(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 (1^st argument, 2^nd, 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 is especially convenient. For internal structures of a computer program, the equivalent predicative representation is usually preferred, with special formal means 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 shown in the Figure IV.8, for example, it is immediately seen that there are three 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 of human knowledge can operate by the same logical structures, but the knowledge in general is in no way coincident with purely linguistic knowledge and even can have different, non-discrete, nature. Hence, they argue, for the transition from Meaning in its linguistic sense to the corresponding representation in terms of general human knowledge, some special stage is needed. This stage is not a part of language and can operate with tools not included in those of the language proper.

Decomposition and atomization of Meaning

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 little girls 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 more utterances with seemingly the same meaning do occur. In such situations, linguists hope to find a universal representation via decomposition and even atomization of the meaning of several semantic components.

In natural sciences, such as physics, researchers usually try to divide all 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 has the same objective. It tries to find the atomic elements of meaning usually called semantic primitives, or semes.

Semes are considered indefinable, since they cannot be interpreted in terms of any other linguistic meanings. Nevertheless, they can be explained to human 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 in the semantic representation of text in a more detailed manner, such as a logical formula or a semantic graph. For example, we can decompose

MATAR(x) ® CAUSAR(MORIR(x)) ® CAUSAR(CESAR(VIVIR(x))),

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 a José 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 limited number of semes, to which a major part of the semantics of a natural language would be reduced, have not been successful.

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

Therefore, computational linguistics uses many evidently non-elementary terms 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 judge what 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 should mention that, in contrast to common mathematical functions, this mapping is not unique in both directions, i.e., it is of the many-to-many type. In this section, 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 the majority of people, equivalent in their meanings. In other words, two or more texts 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 absolute synonyms. 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 their meanings in many contexts, but not all.

Sometimes the set of possible contexts for one such synonym covers the whole 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—intersection of 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. The menu 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 can exhibit two or more different meanings. That is, one element of the surface edge of the mapping (i.e. text) can correspond to two or more elements of the deep edge. We have already discussed this phenomenon in the section on automatic translation, 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 are called 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 resort again 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 listener or reader according to the context of the wordform or based on the extralinguistic situation in which this form is used. In general, the reader or listener does not even take notice of any ambiguity. The corresponding mental operations are immediate and very effective. However, resolution of such ambiguity 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 rather limited utility.

More on homonymy

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

For formal distinguishing of homonyms, their description in conventional dictionaries is usually divided into several subentries. The names of lexical homonyms are supplied with the indices (numbers) attached to the words in their standard dictionary 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 are eliminated.

The purely lexical homonymy is maybe the most difficult to resolve since at the morphologic stage of text processing it is impossible to determine what homonym 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, various situations can be observed in any language. In some cases, such homonyms have no 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 homonymous lexemes 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 homonymous lexemes, 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 linguistics has been developed in recent years using object-oriented ideas. Polysemic lexemes are represented as one superclass that reflects the common part of their meaning, and a number of subclasses then reflect their semantic differences.

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

The ways of the enrichment of language are rather numerous, and the main of 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 abbreviations broaden 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 sin puertas, 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 incorruptibles que, hincados en el suelo, sirven de sostén al armazón de las casas en las ciudades 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 by non-homonymous lexemes, like the Spanish antigüedad.

In other cases, a set of homonyms in a given language is translated into a 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 and distinguishing their meanings, though the main considerations should be deserved to the inner facts of the given language.

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

Classical lexicography can help to define these problems, but their resolution 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 of entities 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 many linguists. 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, with small differences in the definitions, namely the morphologic and syntactic ones.

Figure IV.11. Stages of transformation.

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

Other intermediate levels are often introduced by some linguists so that the partial transformers themselves are divided into sub-transformers. For example, a surface and a deep syntactic level are introduced for the syntactic stage, 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 represented in 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 specified for each linguistic level separately, program developers often distinguish a morphologic 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 distributed among several linguistic levels. In Text, there are only wordforms. In analysis, lexemes as units under processing are involved at morphologic level. Then they are used at surface and deep syntactical levels and at last disappeared at semantic level, giving up their places to semantic elements. The latter elements conserve the meaning of lexemes, but are devoid of their purely grammatical properties, 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 good illustration 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 true semantic representation, i.e., the true level of Meaning, consisting of the universal and thus standard set of semes. Therefore, all practical systems have 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 the language, and then, in the last stages, return to the surface, with the ultimate result in textual form once more. The deeper the level reached, the smaller the difference 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 of the transformation might not be necessary.

For example, if we translate from Spanish into Portuguese, then, because these two languages differ mainly in their lexicon, it can be sufficient to use only the first stage of analysis and the last stage of synthesis, just replacing each Spanish word by the corresponding Portuguese one on the morphologic 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 of which is that they are to be combined together into larger signs and each one can in turn consist of several smaller signs. Natural language can be viewed as a system of linguistic signs.

As another property of linguistic sign, its signifier at the surface level consists of elementary parts, phonetic symbols in the phonetic transcription or letters in the written form of language. These parts do not have any signified of their own: a letter has no meaning, but certain strings of letters do have it.

We have already mentioned other notation systems to represent words of natural language, such as hieroglyphic writing. Each hieroglyph usually has its own 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 sign from 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 something rather 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 a phoneme string (or a sequence of such strings). Taking into account the correspondence between acoustic and written forms of language, such signifier can 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 a structure (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 all relevant linguistic facts are taken into consideration equally and properly by the 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 notions equivalent to the signifier and the signified were discussed in science from the times of the ancient Greeks. For several centuries, it has been debated whether 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 any language 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 any language takes their origin from such imitation, all the other words being quite 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 for material links between the two sides of linguistic signs. Thus, a linguistic sign is considered a human convention that assigns specific meanings to some material things such as strings of letters, sequences of sounds, pictures in hieroglyphic writing or gestures in sign language.

Generative, mtt, and constraint ideas in comparison

In this book, three major approaches to linguistic description have been discussed 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 that they can be viewed in terms of linguistic signs. All of them describe the structure of the signs of the given language. All of them are used in computational practice to find the Meaning corresponding to a given Text and vice 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 is frequently referred to as generative idea). Clearly, for a natural language it is impossible to literally list all signs in it, since their number is infinite. Thus, more strictly speaking, a generative grammar describes an algorithm that lists only the correct signs of the given language, and lists them all—in the sense that any given sign would appear in its output after a some time, perhaps long 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 of generative grammars suggested by N. Chomsky constructs each sign gradually, through a series of intermediate, half-finished sign “embryos” of different degree of maturity (see Figure IV.16). All of them are built starting from the 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 for the formalism of description of language. It is not—and is not intended to be—applied in practice for the generation of an infinitely long list of language expressions, which would be senseless. The use of the description—once created—for passing from Text to Meaning and vice versa is indirect. A program called 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 generate that 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 solves such an equation using the data structures and algorithms quite different from the 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 belongs to 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 guar

Selection of topics

Linguistics and its structure

Word, what is it?

Current state of applied research on Spanish

A simple context-free grammar

The linguistic research after Chomsky: Valencies and interpretation

The Meaning Û Text Theory: Multistage transformer and government patterns

The Meaning Û Text Theory: Dependency trees

The Meaning Û Text Theory: Semantic links

IV. Language as a Meaning Û Text Transformer

Not-uniqueness of Meaning Þ Text mapping: Synonymy

Not-uniqueness of Text Þ Meaning mapping: homonymy

Multistage character of the Meaning Û Text transformer

Translation as a multistage transformation

Linguistic sign in the mmt

Linguistic sign in hpsg

The linguistic research after Chomsky:
Valencies and interpretation

The Meaning Û Text Theory:
Multistage transformer and government patterns