Dependency Relations as Source Context in Phrase-Based SMT

Dependency Relations as Source Context in Phrase-Based SMT Rejwanul Haque a, Sudip Kumar Naskar a, Antal van den Bosch b, and Andy Way a Abstract. The Phrase-Based Statistical Machine Translation (PB-SMT) model has recently begun to include source context modeling, under the assumption that the proper lexical choice of an ambiguous word can be determined from the context in which it appears. Various types of lexical and syntactic features such as words, parts-of-speech, and supertags have been explored as effective source context in SMT. In this paper, we show that position-independent syntactic dependency relations of the head of a source phrase can be modeled as useful source context to improve target phrase selection and thereby improve overall performance of PB-SMT. On a Dutch—English translation task, by combining dependency relations and syntactic contextual features (part-of-speech), we achieved a 1.0 BLEU (Papineni et al., 2002) point improvement (3.1% relative) over the baseline. Keywords: phrase-based SMT, syntactic dependencies, memory-based learning 1 Introduction1 In log-linear phrase-based SMT (Koehn et al., 2003), the probability P(eI 1|fJ 1) of a target phrase eI 1 given a source phrase fJ 1 is modelled as a log-linear combination of features which typically consist of a finite set of translation features, and a language model (Och and Ney, 2002). The usual translation features involved in those models express dependencies between the source and target phrases, but not among the phrases in the source language themselves. Stroppa et al. (2007) were the first to show that incorporating source-language context using neighbouring words and part-of-speech tags had the potential to improve translation quality. Due to a strand of related work, source context modeling has now been shown to offer a new dimension to PB- SMT. By making use of similarity in the contexts of source phrases, information can be added that can positively influence the weighting and selection of target phrases. Approaches to include source context for proper selection of target phrases have been inspired by methods for word sense disambiguation (WSD), that employ rich context-sensitive features to determine the contextually most likely sense of a polysemous word. These contextual features may include lexical features of words appearing in the context and bearing sense- discriminatory information, position-specific neighbouring words (Giménez and Márquez, 2007; Stroppa et al., 2007), shallow and deep syntactic features of the sentential context (Gimpel and Smith, 2008) and full sentential context (Carpuat and Wu, 2007). Most of the work on syntactic features has made use of part-of-speech taggers (Stroppa et al., 2007), supertaggers (Haque et al., 2009) and shallow and deep syntactic parsers (Gimpel and Smith, 2008). In the present work, we explore how the local sentential context information from a dependency parse can be modeled as source context features to be integrated into a PB-SMT model. Copyright 2009 by Rejwanul Haque, Sudip Kumar Naskar, Antal van den Bosch, and Andy Way a CNGL, School of Computing, Dublin City University, Dublin 9, Ireland {rhaque, snaskar, away}@computing.dcu.ie bILK Research Group, Tilburg centre for Creative Computing, Tilburg University, Tilburg, The …

Content truncated to 500 words