On Extracting Multiword NP Terminology for MT

Proceedings of the 13th Annual Conference of the EAMT, pages 205–212, Barcelona, May 2009 On Extracting Multiword NP Terminology for MT Svetlana Sheremetyeva LanA Consulting ApS Møllekrog, 4, Vejby 3210 Copenhagen, Denmark lanaconsult@mail.dk Abstract The paper addresses the issue of MT knowledge acquisition and describes a new hybrid methodology for automatic extraction of multi-word nominal termi- nology. The approach is based on statisti- cal techniques merged into a strongly lexicalized Constraint Grammar para- digm. It is targeted at intelligent output and computationally attractive properties. 1 Introduction The quality of machine translation output is to a large extent influenced by the comprehensive- ness of multiword term dictionaries where noun phrases (noun phrase terms) are more frequent than any other types of multiword expressions. Noun phrases (NPs) can often be translated into other languages irrespective of the context and contribute significantly to the robustness of MT systems by reducing the ambiguity inherent in word to word matching and text analysis. Multiword phrase databases are relevant for both RBMT and SMT systems; - in state-of-the- art statistical translation systems structural rela- tions between source and target sentences are captured by means of phrases instead of isolated words (Zens et al., 2002; Koehn et al., 2003). To build a two level hierarchy of phrases phrase driven SMT systems more and more focus on the linguistic concept of NP as the unit of de- composition (Hewavitharana et al, 2007). Creating multilingual MT resources is based on the acquisition of unilingual lexicons as the first and basic step (Pohl, 2006; Hewavitharana et al, 2007; Daille and Morin, 2008). ____________________________ © 2009 European Association for Machine Translation In many cases, especially for low resource lan- guages, lexical acquisition starts from the Eng- lish side independent of the translation direction. For example, (Hewavitharana et al, 2007) devel- oping an Arabic-to-English MT first extract Eng- lish NPs as the Arabic parsers available do not produce desired accuracy. The quality of mono- lingual (English, in particular) extraction, is thus of primary importance for the quality of MT. Another issue which matters a lot for practical MT systems is the speed of NP extraction proc- ess. It directly affects the costs of MT system development and maintenance. Despite a lot of research on extracting different kinds of multi- word phrases related to MT the problem still pre- sents a tough challenge (Piao et al., 2006). Ex- traction of NP phrases, in particular, is especially problematic as it normally involves parsing and often very expensive computationally. We suggest a new hybrid NP extraction meth- odology based on statistical techniques merged into a strongly lexicalized Constraint Grammar paradigm which features computationally attrac- tive properties and intelligent output. We illus- trate our approach on the example of the English language as its resources are most widely used in MT research and the quality of these resources is still to be improved. While testing our approach we got practical results for the patent domain corpus and saw how generic extraction procedures can take advantage of …

Content truncated to 500 words