Dynamic Terminology Integration Methods in Statistical Machine Translation

Dynamic Terminology Integration Methods in Statistical Machine Translation M¯arcis Pinnis Tilde, Vienibas gatve 75a, Riga, Latvia University of Latvia, 19 Raina Blvd., Riga, Latvia marcis.pinnis@tilde.lv Abstract In this paper the author presents meth- ods for dynamic terminology integration in statistical machine translation systems using a source text pre-processing work- flow. The workflow consists of exchange- able components for term identification, inflected form generation for terms, and term translation candidate ranking. Auto- matic evaluation for three language pairs shows a translation quality improvement from 0.9 to 3.41 BLEU points over the baseline. Manual evaluation for seven lan- guage pairs confirms the positive results; the proportion of correctly translated terms increases from 1.6% to 52.6% over the baseline. 1 Introduction In professional translation services, correct and consistent handling of terminology is an important indicator of translation quality. However, pure sta- tistical machine translation (SMT) systems, such as, Moses (Koehn et al., 2007) in a general sce- nario cannot ensure correct and consistent han- dling of terminology, because statistics of large amounts of data are difficult to control if not con- strained by means of, e.g., bilingual term collec- tions or translation model or language model adap- tation techniques. In cases where the context is too ambiguous (e.g., if an SMT system receives just a short translation segment or the SMT system’s models are limited in the possibilities to analyse larger context) or when external knowledge is re- c⃝ 2015 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. quired, it can be impossible for an SMT system to guess the correct translation. In the localisation industry customers often pro- vide their own term collections that have to be strictly used during translation to ensure correct and consistent usage of terminology. Obviously, such collections may contain term translations that are rated as unlikely (in certain contexts) by an SMT system’s models or they may even be miss- ing in the models at all if custom adaptation of the models using the customers’ provided data is not performed. If such SMT systems would be inte- grated in localisation service workflows, it would not be possible to ensure high terminology transla- tion quality in the SMT suggestions. Therefore, ef- fective methods that can benefit from custom term collections are necessary. Researchers have tried to address the terminol- ogy integration challenge directly by using in- domain term collections and indirectly by tackling the broader challenge of domain adaptation. Sig- nificant research efforts have been focussed on us- ing in-domain parallel and monolingual corpora (that contain in-domain terminology) to perform SMT system translation and language model adap- tation to specific domains (to name but a few, Koehn & Schroeder (2007), Bertoldi & Federico (2009), Hildebrand et al. (2005), and many others). Terminology integration has been also indirectly addressed by research on multi-word unit integra- tion in SMT. E.g., Bouamor et al. (2012) showed that for French-English it is enough to simply add multi-word unit pairs to the parallel …

Content truncated to 500 words