Teaching contrastive linguistics for MT

Teaching Contrastive Linguistics for MT Paul Bennett Centre for Computational Linguistics UMIST P O Box 88 Manchester M60 1QD Great Britain Paul.Bennett@umist.ac.uk Abstract We discuss a framework for teaching con- trastive translation-oriented linguistics to students of machine translation. The frame- work embodies a classification of cross- linguistic differences and a demonstration of how the neutralization of such differences in abstract representations can make trans- fer modules simpler. We illustrate by means of examples from a range of languages, and then discuss a proposal for linguistic typol- ogy to which our system can be fruitfully related.1 1 Introduction Leaving aside computational aspects, what kind of linguistics should be taught to stu- dents of machine translation (MT)? To some extent, it should be the kind of linguistics taught as part of any introductory linguis- tics course, concentrating on morphology, lexis, syntax and semantics, though perhaps with a special emphasis on ambiguity and its types. But, in addition, there needs to be a great deal of attention paid to the ways in which languages differ from each other and how this creates problems for MT. How, though, are these linguistic differences to be approached?2 One possible answer would be the kind of course or module already taught in many linguistics departments, viz. on linguistic ty- pology or universal grammar, concentrating on the ways languages do and do not differ and perhaps on classifying languages into different types. This is the kind of approach found in textbooks such as Comrie (1981) and Croft (1990). For a number of reasons, however, such a course is not suitable. Firstly, typological work sometimes deals with relatively superficial aspects of lan- guage (such as case-marking and agree- ment) and often refers to ‘exotic’ languages with relatively few speakers which are of lit- tle relevance for MT. Secondly, it generally fails to provide the kind of explicit bilingual knowledge that is needed. Let’s just take one example: If a language uses a nonzero case marking for a direct object on the an- imacy/definiteness hierarchies, then it uses a nonzero case marking for di- rect objects higher on the hierarchies. (Croft 1990, p. 128) �� Thanks to Miriam Butt, Federico Gaspari and Piklu Gupta for comments on an earlier version. �� See Bennett (to appear) for some more general considerations on the relevance of linguistics for MT, in a volume aimed at translators; with a few excep- tions, points made in that paper are not repeated here. (The general idea here is that some combi- nation of animacy and definiteness triggers overt case-marking on an object – cf. Span- ish a.) This simply does not help in MT: the rules for case-marking in a particular lan- guage need to be stated irrespective of how they relate to these hierarchies, and in any case such a ‘surfacy’ feature is largely irrel- evant to translation. Instead, contrastive work which empha- sizes the main languages used in MT R&D (and those most often spoken by the stu- dents themselves) is preferable. It should …

Content truncated to 500 words