Evaluation-Guided Pre-Editing of Source Text: Improving MT-Tractability of Light Verb Constructions

Proceedings of the 13th Annual Conference of the EAMT, pages 36–43, Barcelona, May 2009 Evaluation-guided pre-editing of source text: improving MT-tractability of light verb constructions Bogdan Babych Anthony Hartley Serge Sharoff Centre for Translation Studies University of Leeds, LS2 9JT, UK {b.babych,s.sharoff,a.hartley}@leeds.ac.uk Abstract This paper reports an experiment on eval- uating and improving MT quality of light-verb construction (LVCs) – com- binations of a ‘semantically depleted’ verb and its complement. Our method uses construction-level human evaluation for systematic discovery of mistranslated contexts and creating automatic pre-edit- ing rules, which make the constructions more tractable for Rule-Based Machine Translation (RBMT) systems. For rewrit- ten phrases we achieve about 40% reduc- tion in the number of incomprehensible translations into English from both French and Russian. The proposed meth- od can be used for enhancing automatic pre-editing functionality of state-of-the- art MT systems. It will allow MT users to create their own rewriting rules for fre- quently mistranslated constructions and contexts, going beyond existing systems’ capabilities offered by user dictionaries and do-not translate lists. 1 Introduction: Automatic rewriting functionality for MT systems Current state-of-the-art RBMT systems offer users customisable functionality for the transfer stage, in the form of user-definable do-not-trans- late lists and user dictionaries. However, the source language analysis capabilities of MT sys- tems remain largely inaccessible to users and such systems still do not offer any support for re- writing at the pre-editing stage. Rewriting offers a way of enhancing the comprehensibility of MT output by more efficiently exploiting existing transfer resources of an MT engine, and greatly extends coverage of construction in a customised way without any changes to the engine. © 2009 European Association for Machine Translation. In our paper we suggest that users will benefit from an integration of source-language rewriting capabilities into MT systems, and from their syn- ergies with existing tools – user dictionaries and do-not-translate lists. Such integration will offer users much greater flexibility, because certain phenomena are much better treated in a monolin- gual rewriting stage rather than within the dic- tionary. For example, rewriting can ‘repair’ in- tractable word order, or handle discontinuous multiword expressions in a much more prin- cipled way than a user dictionary. We describe an experiment which demon- strates the usefulness of source-rewriting func- tionality for state-of-the-art MT systems. For our evaluation-guided rewriting experiment we have chosen light-verb constructions (LVCs) – verb phrases consisting of a 'light', i.e., 'semantically depleted' verb and its object. Objects in such constructions are so-called logical predicates – such as names of actions, activities, states, prop- erties, relations – that put forward some situ- ational propositions. The relationship between verbs and complements can be described in terms of lexical functions like Oper1 (Mel’čuk, 1996). LVCs are very common in a variety of lan- guages, e.g., take action, take part, put pressure, make a decision} in English, оказывать давление ‘put pressure’ (lit. ‘make pressure’) in Russian, or tenir compte de ‘take into account’ (lit. ‘hold account of’) in French (Salkoff, 1999). These constructions are often …

Content truncated to 500 words