CC BY
80
о
(N fü 2 Ф CL
fü £ ш .с сс
Е.Б. К03ЕРЕНК0
Синтаксическая многозначность и неоднозначность в перспективе машинного перевода
В статье рассматриваются вопросы выявления семантического содержания синтаксических структур английского и русского языков посредством их контрастивного сопоставления, а также с привлечением анализа параллельных текстов для решения задач машинного перевода. Особое внимание уделяется рассмотрению явлений синтаксической многозначности и неоднозначности, предлагаются механизмы представления многозначных синтаксических структур для лингвистических процессоров информационных систем. Предлагаемые решения основаны на принципе выражения функциональных значений через традиционные категориальные значения атомарных и распространенных языковых структур. Синтаксическая многозначность понимается как одновременная реализация более одного категориального значения в головной вершине некоторой синтаксической структуры. Явление синтаксической многозначности определяет схему возможного мновари-антного перевода (трансфера) языкового паттерна. Разработанная система многовариантных правил трансфера в дальнейшем уточняется с помощью методов машинного обучения.
Ключевые слова: машинный перевод, синтаксис, семантика, трансфер, линейные структуры, разрешение неоднозначности.
1. Introduction
The establishment of structures equivalence on the basis of functional semantics proved to be useful for developing the syntactic parse and transfer rules module for the English - Russian machine translation [21]. Generally, major efforts connected with natural language modeling lay emphasis at lexical semantics presentations and less attention is paid to the semantics of structures and establishment of functional similarity of language patterns as a core problem in multilingual systems design.
The set of functional meanings together with their categorial embodiments serves the source of constraints for the unification mechanism in the formal presentation of our grammar. The formalism developed employs feature-based parse, and head-feature inheritance for phrase structures which are singled out on the basis of functional identity in the source and target languages.
A constraint-based formalism which is called the Cognitive Transfer Grammar [21] was developed. It comprised 222 transferable phrase structures together with the transfer rules combined within the same pattern. Such patterns, or Cognitive Transfer Structures (CTS), served constitutional components of the declarative syntactical processor module and encoded both linear precedence and dependency relations within phrase structures.
However, the syntactic structures of natural languages in many cases tend to be polysemous and ambiguous which results in multiple possible transfers from one language to another. The systemic study of polysemous syntactic structures is carried out on the basis of functional transfer principle, and cross-lingual correspondences between the English and Russian languages are encoded in the multivariant cognitive transfer rules.
2. Polysemy of Syntactic Structures
By syntactic polysemy we mean the immediate realization of more than one categorial meaning within the head element of a language structure. The polysemous structures display variable manifestation of their categorial features depending on the functional role in the sentence. Consider such language phenomena as the Gerund, the Participle and the Infinitive. The Gerund comprises the features of both the Verb and the Noun, which affects the translation strategy when the appropriate means are to be chosen for representation of the English Gerund via the Russian language forms. The structures similar in category to the English Gerund are the Russian Verbal Nouns denoting "Activity", e.g. singing ^ penie, reading ^ chtenie, and both the English Gerund, and the Russian Verbal Noun allow direct object arguments if derived from transitive verbs. However, the direct transfer of the Gerund into the Russian Verbal Noun is the least probable translation variant of the three possible transfer schemes:
The Gerund (Eng) ^ Clause with the Finite Verb form (Rus) The Gerund (Eng) ^ Clause with the Infinitive The Gerund (Eng) ^ Verbal Noun.
This fact can be accounted for by the mechanisms employed in the Russian language for configuring sentential structures and is to be envisaged in the machine translation engine.
Consider the other most productive polysemous language structures which 53 comprise more than one categorial meaning:
The Participle ^ Verb + Adjective _
The Infinitive ^ Verb + Noun ==
Nominal Phrase as the Nominal Modifier ^ Noun + Adjective 5
Verbal Phrase as the Verbal Modifier ^ Verb + Adverb. ^
Thus we introduce the notion "polysemous syntactic structure" to deter- £ mine the set of possible transfer schemes for a given language structure. | When a polysemous structure is assigned specific categorial attributes real- | ized in this structure, the possible and preferable transfer schemes become predictable for the given structure.
The predominant categorial meaning of a polysemous syntactic structure (or syntaxeme) is determined by the syntactic function realized at a given moment. Thus the transfer scheme for a "stone wall" construction will be as follows:
Nounl + Noun 2 [Eng.] ^ Adjective + Noun 2 [Rus] The weight for this transformation will be higher than for the transformation: Nounl + Noun 2 [Eng] ^ Noun 2 + Nounl (Genitive ) [Rus] if the dictionary contains an Adjective as one of the possible translation equivalents for Nounl, that is the case when the dictionary is composed by various methods including acquisition of lexical units from parallel texts.
Judging by the function we establish the transfer field [21] within which the translation procedure will be carried out. The Functional Transfer Fields (FTF) support the possible paraphrasing variants and envisage the synonymous ways of conveying the same functional meaning across languages.
The syntactic polysemy manifests the action of the mechanism of the transposition of meaning stated by the Semiotic Universal Grammar [29]. The awareness of the mechanism action gives cues to designing the procedures for polysemous structures parse and transfer, the multiple syntactic-semantic data structures being associated with a given syntactic pattern.
Of special interest is the situation of the categorial shift in translating a syntactic pattern. The category of a syntactic pattern, i.e. phrase structure, is determined by the category of the head word of this phrase structure. Thus, when transfer employs conversion, and the category of the head word shifts to another category, the whole structure is assigned the feature of the new category. Thus a Nominal modifier of a Nominal Phrase becomes an Adjective in translation; a Verbal unit acting as a Verbal modifier becomes an Adverbial clause containing the Finite Verbal form. The latter case accords with the SUG principle of the Verb being the Sentence Nucleus [29].
54
3. Ambiguity of Syntactic Structures
We find it important to differentiate between polysemous and ambiguous syntactic structures. A polysemous structure implies possible realizations (U of meanings which are compatible within one language structure and can be i transferred to the structures of another language which are isofunctional to 5 the source language structure. An ambiguous syntactic structure presup-| poses alternative ways of interpretation, the meanings being incompatible ¿S within one language structure, thus we deal with ambiguity when we try to discern some Finite and Nonfinite verbal forms:
Gerund / Present Participle;
Infinitive / Present Simple;
Past Participle / Past Simple.
Ambiguous structures can be misleading to the parsing procedures and subsequent machine translation, as for example, the "garden path" is a well-known language phenomenon which may give incorrect parse at the early stage of analysis, that could be corrected only at the final stage:
The cars passed by the vessel drowned.
The possible interpretations for the sentence can be
The cars which were passed via the vessel drowned (the correct variant).
The cars which passed the vessel drowned.
However, the phrase
The new control system updated continuously displayed robust performance.
was analyzed and translated correctly by all the tested modern MT systems which comprise learning mechanisms within their framework. This can be explained by the presence of the broader context.
4. Functonal Semantics as the Basis for Contrastive Study
Translation activity involves the search for equivalence between structures of different languages. However, to establish whether the structures and units are equal or not, we need some general equivalent against which the language phenomena would be matched. In Contrastive Linguistics the notion of tertium comparationis is widely employed to denote this general equivalent, and in [2; 11] it was demonstrated that the approach based on the principle "from the meaning to the form" focusing on Functional Sytax would yield the necessary basis for equivalence search. We also consider important the typological studies of syntax as in [32] that offer an alternative to the standard generative view of syntax: syntactic phenomena are presented from a wide range of languages and the emphasis is laid on the major typological issues that syntactic theories must address.
What differs our approach is the attention to the semantics of configura- 55 tions, i.e. the study of the way languages tend to arrange structures in order to convey certain meanings. And we focus on the linear patterns of the languages under study, since we assume that linearization is not a random == process but it is determined by the cognitive mechanisms of speech produc- 5 tion and the way they manifest themselves in syntactic potentials of a given ™ language. The research which gives evidence as to the mental control of | coordinate structures linearization was presented in [1]. g
The primary object of our contrastive language study was to establish | what particular language meanings are represented in the categorial-functional systems of the English and Russian languages. Categorial values embody the syntactic potentials of language units, i.e. their predictable behavior as syntactic structures (syntaxemes). Thus we can say that Category is the potential for Function, and multiple categorial values inflict multiple syntactic functions. However, when we analyze language in action, i.e. the utterances of written or sounding speech, the function of a particular language structure determines which of the potentials is implemented in this utterance, hence which of the possible categorial values of a polysemous syntactic structure is to be assigned to the parse result.
The fifteen major Functional Transfer Fields (FTF) singled out at the first stage of our development [21] comprised the basic groups of language configurations united on the principle of functional semantic similarity to the prototype functional structures and the possibility of language structures periphrasis within the same FTF. To illustrate the mechanism of polysemous structures transfer we take the Secondary Predication FTF and the Attribu-tiveness FTF.
The Secondary Predication FTF bearing the features of verbal modifiers for the Primary Predication structures (the non-inverted Finite Verb forms and tensed verbal phrase structures bearing the Tense - Aspect - Voice features) includes the nonfinite verbal forms and constructions, and subordinate clauses comprising the finite verbal forms. All these are united by the functional meanings they convey, e.g. qualification, circumstance, taxis (ordering of actions), etc.
The following schemes of transfer into Russian are applicable to the phrase: Feeling surprised seemed permanent.
"Gerund + Participle II + Finite Verbal Phrase" ^ " Sentence " ^ "Nominative Clause + Finite Verbal Phrase" (1) Or
"Verbal Noun Phrase + Finite Verbal Phrase" (2) The Participle in postposition to a Nominal Phrase most frequently would be transferred into a Russian Clause :
The material processed satisfied all the requirements.
"Nominal Phrase + Participle II + Finite Verbal Phrase" ^ "Sentence" ^
^"Nominal Phrase + Qualifying Clause + Finite Verbal Phrase" (1)
Or
"Nominal Phrase + Participle II + Finite Verbal Phrase" (2)
Attributiveness FTF: adjectives and adjectival phrases in all possible I forms and degrees comprise the semantic backbone of this field; included £ here are also other nominal modifiers, such as nominal language units and ¡5 structures (stone wall constructions, prepositional genitives - of - phrases), | and other dispersed language means which are isofunctional to the backbone units.
Consider the phrases of the kind: "a woman of means", "a man of talent". Possible contexts might be as follows:
She was a woman of means.
He was a man of talent.
The multivariant transfer would comprise the following Russian phrase structures:
(1) with the Genitive construction;
(2) with the Qualifying Clause;
(3) with the Preposition "s" (Russian):
"Nominal Phrase 1 + of + Nominal Phrase 2" ^
"Nominal Phrase 2 + Nominal Phrase l Genitive"
Or
"Nominal Phrase l + Qualifying Clause"
Or
"Nominal Phrase l + Prep "s" + Nominal Phrase 2 Instrumental"
The last variant would mean in Russian "a woman with means", "a man with talent".
We took into account the computational cost of the rule system which led us to a certain minimalism: we avoided introduction of abstract categories in rule design (having in mind the imperative known as Ockham's Razor: the notion that when presented with a choice of axioms or laws, or explanations, it is wise to choose the one that is the simplest). All the functional meanings were presented as feature - value structures based on traditional language categories.
5. Disambiguation Techniques: Rule-Based and Machine Learning Methods
The impact of differentiation between syntactic polysemy versus syntactic ambiguity consists in the following implementation decisions. An ambiguous structure is analyzed in alternative manner: each possible parse and transfer variant is presented as a separate rule, and constraints are introduced into the rule structure. A polysemous structure is assigned a multiple transfer scheme within one rule.
The mechanism of computational (contextual) reframing (CR) is being 57 designed for treatment of the two major bottlenecks: syntactic derivation history (for words in a secondary, tertiary, etc. syntactic function) and syn- _ tactic polysemy of structures. Reframing models the use of the same struc- == tural unit in different structural and/or lexical contexts, which results in the 5 difference of the meanings of this unit. The presentations for the syntactic ^ module rest on the basis of traditional word categories. Contextual correla- £ tions associated with each function of a structural unit are established via | stochastic data obtained from corpora study. |
The detailed exploration of language structures comprising Functional Transfer Fields within the systems of the English and Russian languages and incorporating the phenomena of syntactic polysemy and ambiguity within the rule bloc leads to development of the Multivariant Cognitive Transfer Grammar (MCTG) which serves the foundation for parse and transfer of language structures with special stress on the periphery language means, i.e. the least investigated structures. At present the MCTG rule set comprises 347 rules for parse and transfer.
In our case the parse for MT is designed on the principle of functional semantic priority. The systemic presentation of transferable structures establishes the nodes which can be transferred from the source language to the target language, unlike the monolingual grammars where the parse rules result in structures which in many cases should be reordered or split to be interpreted via the structures of another language.
Thus the grammar formalisms developed for a unilingual situation (phrase stucture rules systems for the English language) [6; 7; 14] would not give a transferable parse in many cases. For example, just one English phrase structure rule for simple sentence would suffice for grammar parse without translation, but for the English - Russian transfer a multiple structure of possible parses is required depending on the specific finite verbal form constituting the sentence. And to overcome this, an accurate scheme for all the particular verbal form cases is designed. Hence, a transferable grammar cannot be efficiently implemented by a mechanical composition of unilingual grammars: a semantic approach is required, and in our case, it is the employment of the functional transfer fields.
Since parse procedures sometimes may result in more than one possible structure, the rules and lexical entries are supplied with the probabilistic augmentations which serve for syntactic ambiguity resolution. The MCTG system of rules envisages the variants of transfer for the polysemous structures and separate alternative rules for ambiguous structures.
As natural language generates an infinite number of sequences, learning mechanisms are incorporated into the parse engine: information about unfamiliar words and structures can be inferred from the context. The data on
which the inference can be founded is accumulated by learning on parallel texts: a supervised algorithm is trained on a set of correct answers to the learning data, so that the induced model may result in more accurate decisions.
The lexical model employs the concise lexicon entries presenting cate-gorial, morphological and combinatorial information supplied with the statistical data for each lexical item characterizing its distribution.
At present the learning paradigm involved into deciding the tasks of disambiguation and acquiring new language patterns is: the probabilistic approach based on Bayesian methods. The new model under development comprises the features of probabilistic functional tree substitution grammar (PFTSG). The learning techniques under design are aimed at coping with the major corpus-based statistical grammar extraction flaw, namely, the overgeneration of rules. Thus we seek to develop the mechanisms of rule generalization without complicity increase.
We studied the existing results in the field of human cognitive mechanisms of language learning, as well as machine learning methods: there is substantial evidence that the way children learn their first language may be understood as information compression [9; 10]; the Optimality theory states the importance of grammatical architecture with the strict prioritization or ranking, rather than any scheme of numerical weighting.
Of particular interest for us was the study and comparison of various formal approaches, so that practical algorithmic solutions could be worked out, we adhere the strict lexicalism principle of the HPSG [25], i.e. word structure and phrase structure are governed by independent principles.
It is a well-known fact that the underlying tree representation plays a crucial role for the performance of a probabilistic model of grammar [18]. Probabilistic developments of the unification grammar formalism are given in [6; 7; 24]. A new probabilistic model for Combinatory Categorial Grammar is presented in [16].
The formalism of the Multivariant Cognitive Transfer Grammar (MCTG) featuring the mechanism of the Probabilistic Functional Tree Substitution Grammar (PFTSG) is being developed to comprise the presentation facilities both for constituency and dependency, as well as disambiguation instrumentality.
6. Modern Tendencies in Machine Translation
The research in the NLP area is substantially stimulated by the fact that the market for MT grows mature in 2002-2003, and more corporations realize that the implementation of a customized machine translation system can promote a company in the competition for today's multilingual customers. New solutions and combinations of methods are being set forward.
The appearance of large parallel texts corpora promoted the statistical 59 methods of NLP which now augment the scheme of the principal existing approaches to MT design - direct translation, transfer and interlingua-based _ methods. A statistical machine translation was first introduced by [3; 31]. == The development of speech-to-speech translation systems [13; 19] substan- 5 tially stimulated the research in the field of MT. The existing computational ^ resources provided for today's MT systems allow to accumulate and recall £ previously corrected translations (Translation Memory and Example-based | Machine Translation) [8; 30]. A model for machine translation based on an | aligned text corpus is example-based machine translation, which means that the example-based translation employs the closest match in aligned corpora as a template for translation. The descriptions of example-based MT systems are given in [23; 27].
The advantages of this approach consist in fully or to a great extent automating the process of linguistic knowledge base formation. However, the bottleneck emerges on the other side: the rules automatically constructed lack accuracy and are overgenerated, which requires post-editing of rules and special efforts and techniques for exclusion of invalid and superfluous rules.
Currently, there are also research projects developing formal models of translation [3; 20; 33], and implemented systems, such as LMT (logical-based machine translation system from English to German) [22] of IBM company.
The present day tendency of comprising the rule-based and stochastic techniques within one framework leads to the emergence of projects featuring the strong sides of diverse approaches.
When considering the performance of the systems we relied rather on subjective evaluation of translations correctness than any tools giving numeric expression of translation quality. The task to develop such tools is of great complicity, though at present some interesting methods have been set forward [17].
Usually, major shortcomings of machine translation tend to be reduced in three ways: by narrowing the problem scope of the texts to be translated and tuning to a subject area; by employment of shallow approaches which give results relevant for informational purposes; by creating program instruments for human translators. The more specific and domain-oriented a system is, the more robust performance can be expected from it. Thus a high degree of precision show such systems as, for example, ENGSPAN, SPANAM [26] which were used in the Pan American Health Organization.
Since we lay emphasis on the synergy of methods, and the present research is focused on the rule system design, of special interest for us were the projects and implemented systems based on the translation memory
approach or including example-based and learning-based techniques as a primary feature or as one of their features, for they are giving the complementary solutions for the problems which cannot be treated efficiently by logical approach alone. Thus we studied the systems with computationally relevant presentations of contexts. We studied the MT projects available in the Internet, tested their performance (when possible) and compared them with traditional rule-based MT systems. The best-known developments of this group are as follows.
SDLX of the SDL International employs translation memory (TM) and supports all languages having the Latin, Arabic and Hebrew alphabets [28]. TM is the basic feature of the products launched by the company in 2003 : SDL insight and SDLX Translation Suite 2003.
DIPLOMAT machine translation system [33]. Rapid deployment is achieved through the use of Pangloss's example-based machine translation (EBMT) and transfer-based MT, within the Multi-Engine MT architecture, which uses a statistical target language model to help select between competing translations. This technology also makes the system's user-driven incremental improvement possible.
A combination of hand-made semantic descriptions and statistics was employed in Matador, a Spanish-English machine translation system, implemented following the Genereation-heavy Hybrid approach to Machine Translation (GHMT) [12; 15].
"eAccela BizLingo" is web-based English-Japanese and Japanese-English translation server software that enables all employees on a company intranet to translate documents, email, and web pages; when used in conjunction with eAccela BizSearch, it can serve as a crosslingual search system for web servers, file servers, or groupware containing multiple documents or emails in both English and Japanese; the developer is Fujitsu Software Corporation [5].
It is obvious that at the present moment the systems tend to comprise several complementary techniques: translation memory solutions are combined with modules which help to increase grammatical accuracy.
7. Conclusions
Our analysis and development experiments result in understanding the efficiency of what might be called the "exteriorization of meaning", i.e. accumulation of relevant data concerning the functional-categorial features of possible structural contexts and/or specific lexical contexts that help to disambiguate the parsed structures of the source language and decide on what particular meaning of a language structure is realized in the given text segment. Rather than invent a sophisticated antropocentric heuristics
for the rule-based disambiguation techniques via traditional linguistic pres- 61 entations, we need to design a synthetic mechanism comprising the core rule set and reliable learning methods.
The rule set applicable both for the transfer procedures and for acquiring == new linguistic data by corpora study should envisage the phenomena of syn- Ц tactic polysemy and ambiguity of structures. The solution employed in our ™ project is based on the Functional Transfer Fields approach grouping iso- | functional language structures, and the Multivariant Cognitive Transfer | Grammar (MCTG) comprising the rules which state the multiple equivalent | structures in the source and target languages. The MCTG linguistic rule set is being augmented by Probabilistic Functional Tree Substitution Grammar (PFTSG) features. Since the nodes of the MCTG have functional articulation, the trees and subtrees of the possible parses also have functional character, i.e. are tagged by functional values.
Further research and development is connected with the refinement of the existing presentations, inclusion of specific lexical-based rules into the grammar system, and excessive corpora-based experiments for extending the mechanisms of multiple transfer.
References
1. Лауфер Н.И. Линеаризация как компонент сочинительной конструкции // Моделирование языковой деятельности в интеллектуальных системах / Под ред. А.Е. Кибрика, А.С. Нариньяни. М., 1987. С. 167-176.
2. Мустайоки А. Функциональный синтаксис как основа сопоставления языков // Studia Slavica Finlandesia. Finnish Contributions to the 13th International Congress of Slavists, Ljubljana, August, 15-21, 2003. Ljubljana, 2003. P. 100-127.
3. A statistical approach to machine translation / Brown P.F. а. al. // Computational Linguistics. 1990. № 16. P. 79-85.
4. Alshawi H., Bangalore S., Douglas S. Automatic acquisition of hierarchical transduction models for machine translation // COLING/ACL-98. Montreal, 1998. P. 41-47.
5. ATLAS - Translation software of Fujitsu Company. URL: http:// www.fujitsu.com/services/software/translation/bizlingo/ (date of the application: 12.01.2016).
6. Briscoe T., Carroll J. Automatic extraction of subcategorization from corpora // Fifth Conference on Applied Natural Language Processing. Washington, D.C., 1997. P. 356-363.
7. Briscoe T., Carroll J. Generalized Probabilistic LR parsing of natural language (corpora) with unification-based grammars // Computational Linguistics. 1993. № 19 (1). Р. 25-59.
8. Brown R.D. Example-based machine translation in the Pangloss system // COLING-96. Copenhagen, 1996. P. 169-174.
9. Chater N. Reconciling simplicity and likelihood principles in perceptual organization // Psychological Review. 1996. № 103 (3). P. 566-581.
10. Chater N. The search for simplicity: a fundamental cognitive principle? // The Quarterly Journal of Experimental Psychology. 1999. № 52 (2). P. 273-302.
11. Chesterman A. Contrastive Functional Analysis. (Pragmatics & Beyond. New Series 47). Amsterdam/Philadelphia, 1998.
12. Dorr B., Habash N. Interlingua Approximation: A Generation-Heavy Approach // AMTA-2002. Interlingua Reliability Workshop. Tiburon, CA, 2002.
i 13. Frederking R., Rudnicky A.I., Hogan C. Interactive speech translation in the | DIPLOMAT project // Proceedings of the ACL-97 Spoken Language Transla-
£ tion Workshop. Madrid, 1997. P. 61-66.
14. Grover C., Carroll J., Briscoe T. The Alvey Natural Language Tools Grammar (4-th Release). Technical Report, 1993, Computer Laboratory. University of Cambridge, 1993.
15. Habash N., Dorr B. Handling Translation Divergences: Combining Statistical and Symbolic Techniques in Generation-Heavy Machine Translation // AMTA-
2002. Tiburon, California, USA, 2002.
16. Hockenmaier Ju. Data and Models for Statistical Parsing with Combinatory Categorial Grammar: Ph.D. thesis. University of Edinburgh, 2003. URL: www.ircs.upenn.edu/~juliahr/Dissertation/thesis.pdf (date of the application: 12.01.2016).
17. Hovy E., King M., Popescu-Belis A. Principles of Context-Based Machine Translation Evaluation // Machine Translation. 2002. Vol. 16. P. 1-33.
18. Johnson M. PCFG Models of Linguistic Tree Representations // Computational Linguistics. 1998. 24(4). P. 613-632.
19. Kay M., Gawron J., Norvig P. Verbmobil: A Translation System for Face-to-Face Dialog. CSLI. Stanford University, 1992.
20. Knight K., Al-Onaizan Y. Translation with finite-state devices // Machine Translation and the Information Soup / Farwell D., Gerber L., Hovy E.H. (eds.). Berlin, 1998. P. 421-437.
21. Kozerenko E.B. Cognitive Approach to Language Structure Segmentation for Machine Translation Algorithms // Proceedings of the International Conference on Machine Learning, Models, Technologies and Applications, June, 23-26,
2003, Las Vegas, USA. CSREA Press, 2003. P. 49-55.
22. Michael C. McCord. Design of LMT: a prolog-based machine translation system // Computational Linguistics. 1989. № 15(1). P. 33-52.
23. Nagao M. A framework of a mechanical translation between Japanese and English by analogy principle // Artificial and Human Intelligence / A. Elithorn and R.B. Banerji (eds.). Edinburgh, 1984. P. 173-180.
24. Osborne M., Briscoe T. Learning Stochastic Categorial Grammars // Proceedings of CoNLL97: Computational Natural Language Learning / T.M. Ellison, ed. Somerset, NJ, 1998. P. 80-87.
25. Pollard C., Sag I.A. Head-Driven Phrase Structure Grammar. Chicago, 1994.
26. Rico C. Translator Training in the European Higher Education Area. Curriculum Design for the Bologna Process. A Case Study // The Interpreter and Translator Trainer. 2010. Vol. 4 (1). P. 89-114.
27. Sato S. CTM: An example-based translation aid system // COLING-14. Nantes, 1992. P. 1259-1263
28. SDL-Trados Machine Translation Studio. URL: http://www.translationzone.com/ 63 products/trados-studio/ (date of the application: 12.01.2016).
29. Shaumyan S. Categorial Grammar and Semiotic Universal Grammar // Proceedings of The International Conference on Artificial Intelligence, IC-AI'03. Las « Vegas, Nevada, 2003. P. 623-629. |
30. Sumita E., Iida H. Experiments and prospects of example-based machine trans- S lation // ACL-91. Berkeley, CA, 1991. P. 185-192. 1
31. The mathematics of statistical machine translation: Parameter estimation / ^ Brown P.F., S.A. Della Pietra, V.J. Della Pietra, R.L. Mercer // Computational | Linguistics. 1993. № 19 (2). P. 263-311. |
32. Valin, van Jr. R.D. Syntax: Structure, Meaning, and function. Cambridge, 1998.
33. Wu D., Wong H. Machine translation with a stochastic grammatical channel // COLING/ACL-98. Montreal, 1998. P. 1408-1414.
