Shallow Semantic Trees for SMT

Proceedings of the 6th Workshop on Statistical Machine Translation, pages 316–322, Edinburgh, Scotland, UK, July 30–31, 2011. c⃝2011 Association for Computational Linguistics Shallow Semantic Trees for SMT Wilker Aziz, Miguel Rios and Lucia Specia Research Group in Computational Linguistics University of Wolverhampton Stafford Street, Wolverhampton, WV1 1SB, UK {w.aziz, m.rios, l.specia}@wlv.ac.uk Abstract We present a translation model enriched with shallow syntactic and semantic information about the source language. Base-phrase la- bels and semantic role labels are incorporated into an hierarchical model by creating shal- low semantic “trees”. Results show an in- crease in performance of up to 6% in BLEU scores for English-Spanish translation over a standard phrase-based SMT baseline. 1 Introduction The use of semantic information to improve Statis- tical Machine Translation (SMT) is a very recent re- search topic that has been attracting significant at- tention. In this paper we describe our participation in the shared translation task of the 6th Workshop on Statistical Machine Translation (WMT) with a sys- tem that incorporates shallow syntactic and semantic information into hierarchical SMT models. The system is based on the Moses toolkit (Hoang et al., 2009; Koehn et al., 2007) using hierarchi- cal models informed with shallow syntactic (chunks) and semantic (semantic role labels) information for the source language. The toolkit SENNA (Collobert et al., 2011) is used to provide base-phrases (chunks) and semantic role labels. Experiments with English-Spanish and English- German news datasets show promising results and highlight important issues about the use of seman- tic information in hierarchical models as well as a number of possible directions for further research. The remaining of the paper is organized as fol- lows: Section 2 presents related work; Section 3 de- scribes the method; Section 4 presents the results ob- tained for the English-Spanish and English-German translation tasks; and Section 5 brings some conclu- sions and directions for further research. 2 Related Work In hierarchical SMT (Chiang, 2005), a Synchronous Context Free Grammar (SCFG) is learned from a parallel corpus.The model capitalizes on the recur- sive nature of language replacing sub-phrases by an unlabeled nonterminal. Hierarchical models are known to produce high coverage rules, once they are only constrained by the word alignment. Neverthe- less the lack of specialized vocabulary also leads to spurious ambiguity (Chiang, 2005). Syntax-based models are hierarchical models whose rules are constrained by syntactic informa- tion.The syntactic constraints have an impact in the rule extraction process, reducing drastically the number of rules available to the system. While this may be helpful to reduce ambiguity, it can lead to poorer performance (Ambati and Lavie, 2008). Motivated by the fact that syntactically constrain- ing a hierarchical model can decrease translation quality, some attempts to overcome the problems at rule extraction time have been made. Venugopal and Zollmann (2006) propose a heuristic method to relax parse trees known as Syntax Augmented Ma- chine Translation (SAMT). Significant gains are ob- tained by grouping nonterminals under categories when they do not span across syntactic constituents. Hoang and Koehn (2010) propose a soft …

Content truncated to 500 words