Improving Translation Selection with Supersenses

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 3114–3123, Osaka, Japan, December 11-17 2016. Improving Translation Selection with Supersenses Haiqing Tang1, Deyi Xiong1∗, Oier Lopez de Lacalle2 and Eneko Agirre2 Soochow University, Suzhou, China1 University of the Basque Country, Donostia, Spain2 hqtang@stu.suda.edu.cn, dyxiong@suda.edu.cn {oier.lopezdelacalle, e.agirre}@ehu.eus Abstract Selecting appropriate translations for source words with multiple meanings still remains a chal- lenge for statistical machine translation (SMT). One reason for this is that most SMT systems are not good at detecting the proper sense for a polysemic word when it appears in different contexts. In this paper, we adopt a supersense tagging method to annotate source words with coarse-grained ontological concepts. In order to enable the system to choose an appropriate translation for a word or phrase according to the annotated supersense of the word or phrase, we propose two translation models with supersense knowledge: a maximum entropy based model and a supersense embedding model. The effectiveness of our proposed models is validated on a large-scale English-to-Spanish translation task. Results indicate that our method can signifi- cantly improve translation quality via correctly conveying the meaning of the source language to the target language. 1 Introduction Phrase-based SMT has achieved better performance than word-based SMT. One of the reasons is that continuous phrases, rather than single words, are used as translation units so that useful context in- formation can be captured for selecting appropriate translations. Even so, when translating sentences containing ambiguous words, which have multiple meanings, the state-of-the-art phrase-based SMT is still suffering from inaccurate lexical choice which makes translations unable to correctly convey the meaning of source sentences. Recent studies show that in order to improve translation quality, one must correctly identify the most likely senses of source-side ambiguous words when selecting target transla- tion (Gao et al., 2013; Zou et al., 2013; Zhang et al., 2014). One common approach to deal with ambiguity is to incorporate a word sense disambiguation (WSD) system into SMT system. At first, Carpuat and Wu (2005) attempt to use the senses of source ambiguous words predicted by a standard formulation of WSD directly in a word-based SMT but results are dis- appointing. They are skeptical of the assumption that WSD systems are useful for SMT. Instead of the standard WSD task, Vickrey et al. (2005) propose a novel formulation of WSD for SMT: directly pre- dicting possible target translation candidates as senses for ambiguous source words. This reformulated WSD has been shown to help SMT by several subsequent studies, including later work by Carpuat and Wu (2007). Following this WSD reformulation for SMT, they integrate the WSD training, where sense definitions are drawn automatically from all phrasal translation candidates rather than from a predefined sense inventory into a phrase-based SMT. In addition to WSD, topic model (Blei et al., 2003) is yet another technique used to detect most likely senses of source words. Various topic-specific lexicon translation models are proposed to improve trans- lation quality. These models can be …

Content truncated to 500 words