Universal Grammar and Lexis for Quick Ramp-Up of MT Systems

Introduction: The Boas Project

The Ecology of Boas

• the nature of data to be acquired;
• the requirements and preferences of the system for which knowledge is acquired; and
• the use of the acquisition tools.

• a morphological analysis grammar;
• a syntactic analysis grammar (determining, at least, phrase boundaries, dependency structures inside and between phrases, and grammatical functions);
• knowledge about a set of special "ecological" features (dates, acronyms, numbers, punctuation treatment, proper name recognition rules, orthographic conventions, etc.);
• a feature and structure transfer grammar between the SL and English; and
• a bilingual SL-English dictionary (including phrasals, proper names and closed-class items).

Defining Parameters for Boas

• language: Ukrainian
• parameter: gender
• domain: nouns, adjectives, possessives (head agreement), verbs in past tense
• range (parameter values): masculine, feminine, neuter
• realization: [gender markers in lexicon for nouns; inflection paradigms for adjectives, possessives and verbs in past tense]

• language: Hebrew
• parameter: gender
• domain: nouns, adjectives, possessive (non-first-person possessor agreement), finite verbs range (parameter values): masculine, feminine
• realization: [gender markers in lexicon for nouns; gender inflection paradigms for adjectives, possessives and verbs]

Unfortunately, work within this approach has not stressed the descriptive task of creating a comprehensive inventory of universal grammar parameters or even those for particular languages or language families. For Project Boas, it means that both the nature of the parameters it would be using and their inventory has to be developed in-house.

In order to define a set of parameters for Boas, it is essential to distinguish among the language phenomena that should be accorded the status of parameter and those that should be understood as parameter values or their realizations. Still other phenomena may remain, at least for the task at hand, outside the parameter system. We believe, with Dorr (1993), that parameters may be understood as building blocks of an interlingua in MT. We reserve judgment about whether every component of an interlingua is by definition parametric 2.

Thus, the parameter "lexical category" has a range of values {V, N, Adj, Adv, ...}. Any of these values may itself be considered a parameter. If viewed within a single language, their values are, ultimately, all words in the language which belong to the respective lexical categories. The realizations of these values are the specific forms of these words, which appear in text decorated with realizations of appropriate values of such morphological parameters as number, gender, case, etc.

An example of a syntactic parameter is head-modifier dependency, whose values include such pairs as "head: noun; modifier: adjective;" "head: verb; modifier: adverb," "head: noun; modifier: relative clause" and others. Realization options for these values involve word or constituent order rules (for instance, post or pre-posing) and agreement rules. Lexical parameters are viewed as language-independent lexical meanings (ontological concepts), such as tablefurniture. The values of this parameter are the word senses corresponding to this ontological concept across the inventory of languages. The realizations for these values are the words or phrases that express this meaning in each language, with a possibility of a lexical gap (a null value) included.

While Boas needs a complete list of all parameters in all languages, it is clear that each individual language uses a subset of that list. For instance, the parameter of gender used by Ukrainian, Hebrew and many other languages is not utilized by English. Indeed, there is no inflection for gender in any word of English (see, e.g., Quirk et al., 1985, pp.314ff). The pronominal forms he, she, his, her and hers, do stand for the English nouns which incorporate the meaning of maleness or femaleness, but this agreement is due to deictic anaphoric rules pertaining also to person and number parameter values.

So, it may appear that if English is a target language of an MT system, the gender parameter should be activated only for translating pronouns. Is it true then that in any particular MT situation only the parameter subset utilized by the target language is important? Why bother with determining the gender of a source language noun if there is no trace of this parameter value in the translation?

In fact, if the parameter of gender is active in the source language, it can be used not only for the purpose of transfer in MT but also to support the analysis of the source text before the bilingual step in machine translation. Thus, the gender of the antecedent of the Russian pronoun kotoryy in (1a,b) determines not only the form of the pronoun itself (which would not be reflected in an English translation) but, importantly, guides the dependency structure of the relative clause, resulting in quite different translations (2 a,b, respectively).

• (1a) Ruka cheloveka, kotoraya podderzhivala polku, soskol'znula.
• (1b) Ruka cheloveka, kotoryy podderzhival polku, soskol'znula.
• (2a) The hand of the man which propped up the shelf slid down.
• (2b) The hand of the man who propped up the shelf slid down.

We do not expect to end up being able to cover all the languages of the world. Instead, we strive tocompile as many of the parameter sets as possible, with as many as possible attested realization options for a realistically large number of languages. A straightforward methodology for such an effort requires a parametrical exploration of a large number of potential target and source languages. In the next section, we explore how having only one target language and reusing computational resources developed in other projects simplifies this methodology and, in fact, makes it feasible.

Translation Environment Supported by Boas

These resources include a) the vocabulary of the generation lexicon which can serve as the list of lexical parameters for compiling the bilingual dictionary; b) a world model (ontology) providing the terms in which the senses of the English words and phrases are expressed (Boas uses the ontology from the Mikrokosmos project at NMSU CRL-see Mahesh and Nirenburg 1995); c) the structure and term definitions from the text meaning representation in Mikrokosmos (see, for instance, Onyshkevych and Nirenburg 1995), to help guide parameter elicitation; d) the set of English closed-class lexical items and morphemes; e) English grammar used in text synthesis, which provides the TL side of structural transfer rules in the runtime MT system (see Figure 1 above); and f) a set of "ecological" parameters and their realizations for English. While a complete description of the use of all of the above resources is beyond the scope of this paper, we will give a few brief illustrations.

The list of English word senses seeds the acquisition of the SL lexicon. The acquirer first simply translates all the word senses into SL and then adds SL features to the corresponding entries as needed. The result is an SL-TL transfer dictionary which also serves as the lexicon for SL analysis. The acquirer gets a lemma with all its senses:

Entry: table-n1
POS: noun
Sense: furniture

Entry: table-n2
POS: noun
Sense: diagram
and produces the following SL lexicon entries (the example is in Hebrew):

Entry: shulxan-n1
POS: noun
Gender: m (plural -ot)
Sense: table-n1

Entry: tavla-n1
POS: noun
Gender: f
Sense: table-n2.

In the examples, the senses are conveniently explained not in any specially designed lexicon/ontology notation, but rather through translation into English. Because each English translation is the entry head for a sense which is already explained in an ontology-based semantic metalanguage in the already existing Mikrokosmos lexicon, Expedition can benefit from richer semantic information than that acquired using Boas. We use the Mikrokosmos ontology as a search space to support word sense disambiguation. The method (suggested by Jim Cowie) depends on the bilingual dictionary of the kind illustrated above. Coarse grain-size lexical mappings of TL word senses to ontological concepts are established (for instance, Chihuahua and Poodle may be both linked to the ontological concept dog). The system, thus, knows that both Chihuahuas and Poodles have four legs, are carnivorous, domesticated, etc.

The disambiguation method uses such ontological constraints by computing a distance in the ontological space between ambiguous word senses on the one hand and the senses of other words in their context. SL syntactic information helps to guide the disambiguation process by providing additional constraints. Thus, closeness between senses of words belonging to the same syntactic unit is weighed more heavily than that across unit boundaries. The acquisition of the complete list of parameters in the single-TL environment is facilitated not only by the availability of the initial set of lexical parameters but also by the prominence of the syntactic and morphological parameters activated in English. Thus, for morphology and syntax, the existence of such comprehensive grammars of English as Quirk et al. (1985) allows a quick round-up of the major parameters. One cannot always limit oneself, however, to TL-induced acquisition as we have demonstrated in the previous section on the example of gender in English.

Source Language Knowledge Acquisition

Acquisition of descriptive knowledge about a language consists in Boas of a set of elicitation "episodes." No matter what particular form the actual elicitation techniques will take, it is clear that the content which needs to be elicited ultimately constitutes the necessary set of lexical meanings semantic structure dependencies characterizing the source language. The difference from the interlingua (knowledge-based) translation approaches is in the way the elicited knowledge is recorded. In Expedition, it will be recorded as bilingual, SL-English correspondences rather than in an abstract metalanguage. Thus, instead of representing data for computer programs, we concentrate on preparation of knowledge necessary to elicit information about a specific source language from a human user. The knowledge which needs to be prepared is largely parametric.

Our work on descriptive knowledge acquisition in Boas is divided into three related parts: a) compiling the complete inventory of all parameters and their values, b) developing the elicitation episodes corresponding to this list, and c) implementing the elicitation techniques to acquire all realizations of parameter values in a given SL and all mappings of SL realizations into TL realizations.

A significant portion of the inventory of parameters has already been acquired in the property branch of the Mikrokosmos ontology: most of the "grammatical" meanings (as in grammatical semantics-cf. Frawley 1993, Raskin 1994) are already recorded and systematized there. This inventory is in the process of being expanded. Sources for this expansion include information about languages not used in the development of the Mikrokosmos ontology and literature on field linguistics. This expanded information includes, for instance, the honorific mode, as in Korean or Japanese, itself a value of the mode parameter, but having a range of values of its own that require considerable modifications in the English translations.

The inventory of parameter values is at this point much less complete. This inventory must include every grammatical meaning, for instance, each nominal case meaning which would include phenomena such as ergativity or the French partitive case, with de l'eau translating as "some water" and l'eau as "water" or "the water."

As far as elicitation techniques are concerned, some methodology has been adapted from field linguistics (see, for instance, Samarin 1967, Bouquiaux and Thomas 1992, Payne 1997). As the native speaker's input must be interpreted by a computational system, not a human, the field linguistics methodology is not applicable directly. Thus, in Comrie and Smith (1977), which is essentially a checklist of parameters for a field linguist, the existence and actual listing of the lexical categories in a SL is taken for granted, and the membership criteria never explored, a luxury that Boas cannot afford. Some recent attempts at automatic language knowledge acquisition (see, for instance, Knight 1996, Knight et al. 1995) are also of some relevance to our task. However, in the cited work the source of language information is not a native speaker, the scope of inquiry is more constrained, and the response range is more limited.

Boas experiments with new elicitation techniques, including asking the acquirers to consult a descriptive grammar of a language (if it is available) to derive comprehensive lists of phenomena, for instance, all the forms of the noun declension paradigm on a typical example with its translations into English, thus avoiding the need in a lengthy and tediously repetitive elicitation episode.

The inventory of parameters and values and the elicitation techniques in Boas are used and put to a test in the process of actual acquisition of the realizations of each parameter value in the SL. Thus, to return to the example of nominal case values, one has to a) elicit the noun inflection paradigms (if any); b) elicit prepositions (if any); c) combine prepositions and cases; d) elicit prepositional meanings; e) elicit meanings of preposition-case combinations (e.g., the Russian s dereva "from the tree," s derevom "with the tree," s derevo "the size of a tree"-see Nirenburg 1980); f) juxtapose these combinations with their parameter values. In the process of knowledge elicitation, the meanings can be expressed by the native speaker in a number of ways-ontologically, as English phrases, using pictures, diagrams, examples, etc. Multimodal representation, if made possible, improves the quality of acquisition by, among other things, breaking the tedium of the long sessions.

The most difficult issues in acquisition involve the transcategorial realization of values, such as the signalling of a noun case in the verb or non-standard clitics, or the lexical realizations in SL of grammatical parameters in TL, such as the possible absence of continuous tenses in a SL and the choice of a grammatical realization of such lexical values as "right now" in the SL as the present continuous marker on the corresponding verb.

Lexical acquisition proceeds as described in Section 4, aided by a special resource created for Boas/Expedition: continuing our work on significantly reducing the number of different senses in a lexicon entry by combining related senses in MRDs (see Nirenburg et al. 1995) and, more rarely, deleting the marginal ones, we have manually reduced a combined (Mikrokosmos and other sources) English lexicon of about 28,000 words to about 40,000 word senses, each of which serves as a lexical parameter for SL acquisition. In addition, frequency analyses of SL corpora will provide the requirements for adding lexical parameters from SL, not just TL.

Conclusion: Computational Field Linguistics?

Boas exemplifies the broad-coverage descriptive approach to NLP (see, for instance, Nirenburg and Raskin 1996) and adds to it a complementary new commitment to automating field-linguistic methodology. This goes hand in hand with the evolving reorientation of theoretical linguistics from selective theorizing in terms of atomistic rule postulation and testing back to the primary goal of linguistics, which is a theory-based language description.

A full evaluation of Boas, that is, the development of the first actual SL to English MT system over a six-month time interval, will take place within the next two years.

Acknowledgments

References