Improved Features and Grammar Selection for Syntax-Based MT

Proceedings of the Joint 5th Workshop on Statistical Machine Translation and MetricsMATR, pages 82–87, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics Improved Features and Grammar Selection for Syntax-Based MT Greg Hanneman and Jonathan Clark and Alon Lavie Language Technologies Institute Carnegie Mellon University Pittsburgh, PA 15213 USA {ghannema, jhclark, alavie}@cs.cmu.edu Abstract We present the Carnegie Mellon Univer- sity Stat-XFER group submission to the WMT 2010 shared translation task. Up- dates to our syntax-based SMT system mainly fell in the areas of new feature for- mulations in the translation model and im- proved filtering of SCFG rules. Compared to our WMT 2009 submission, we report a gain of 1.73 BLEU by using the new features and decoding environment, and a gain of up to 0.52 BLEU from improved grammar selection. 1 Introduction From its earlier focus on linguistically rich ma- chine translation for resource-poor languages, the statistical transfer MT group at Carnegie Mellon University has expanded in recent years to the in- creasingly successful domain of syntax-based sta- tistical MT in large-data scenarios. Our submis- sion to the 2010 Workshop on Machine Transla- tion is a syntax-based SMT system with a syn- chonous context-free grammar (SCFG), where the SCFG rules are derived from full constituency parse trees on both the source and target sides of parallel training sentences. We participated in the French-to-English shared translation task. This year, we focused our efforts on making more and better use of syntactic grammar. Much of the work went into formulating a more expan- sive feature set in the translation model and a new method of assigning scores to phrase pairs and grammar rules. Following a change of decoder that allowed us to experiment with systems using much larger syntactic grammars than previously, we also adapted a technique to more intelligently pre-filter grammar rules to those most likely to be useful. 2 System Overview We built our system on a partial selection of the provided French–English training data, us- ing the Europarl, News Commentary, and UN sets, but ignoring the Giga-FrEn data. After tokenization and some pruning of our training data, this left us with a corpus of approximately 8.6 million sentence pairs. We word-aligned the corpus with MGIZA++ (Gao and Vogel, 2008), a multi-threaded implementation of the standard word alignment tool GIZA++ (Och and Ney, 2003). Word alignments were symmetrized with the “grow-diag-final-and” heuristic. We automati- cally parsed the French side of the corpus with the Berkeley parser (Petrov and Klein, 2007), while we used the fast vanilla PCFG model of the Stan- ford parser (Klein and Manning, 2003) for the English side. These steps resulted in a parallel parsed corpus from which to extract phrase pairs and grammar rules. Phrase extraction involves three distinct steps. In the first, we perform standard (non-syntactic) phrase extraction according to the heuristics of phrase-based SMT (Koehn et al., 2003). In the second, we obtain syntactic phrase pairs using the tree-to-tree matching method of Lavie et al. (2008). Briefly, this method aligns nodes in …

Content truncated to 500 words