Phrase Table Training for Precision and Recall: What Makes a Good Phrase and a Good Phrase Pair?

Proceedings of ACL-08: HLT, pages 81–88, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics Phrase Table Training For Precision and Recall: What Makes a Good Phrase and a Good Phrase Pair? Yonggang Deng∗ , Jia Xu+ and Yuqing Gao∗ ∗IBM T.J. Watson Research Center, Yorktown Heights, NY 10598, USA {ydeng,yuqing}@us.ibm.com +Chair of Computer Science VI, RWTH Aachen University, D-52056 Aachen, Germany xujia@cs.rwth-aachen.de Abstract In this work, the problem of extracting phrase translation is formulated as an information re- trieval process implemented with a log-linear model aiming for a balanced precision and re- call. We present a generic phrase training al- gorithm which is parameterized with feature functions and can be optimized jointly with the translation engine to directly maximize the end-to-end system performance. Multiple data-driven feature functions are proposed to capture the quality and confidence of phrases and phrase pairs. Experimental results demon- strate consistent and significant improvement over the widely used method that is based on word alignment matrix only. 1 Introduction Phrase has become the standard basic translation unit in Statistical Machine Translation (SMT) since it naturally captures context dependency and models internal word reordering. In a phrase-based SMT system, the phrase translation table is the defining component which specifies alternative translations and their probabilities for a given source phrase. In learning such a table from parallel corpus, two re- lated issues need to be addressed (either separately or jointly): which pairs are considered valid trans- lations and how to assign weights, such as proba- bilities, to them. The first problem is referred to as phrase pair extraction, which identifies phrase pairs that are supposed to be translations of each other. Methods have been proposed, based on syntax, that take advantage of linguistic constraints and align- ment of grammatical structure, such as in Yamada and Knight (2001) and Wu (1995). The most widely used approach derives phrase pairs from word align- ment matrix (Och and Ney, 2003; Koehn et al., 2003). Other methods do not depend on word align- ments only, such as directly modeling phrase align- ment in a joint generative way (Marcu and Wong, 2002), pursuing information extraction perspective (Venugopal et al., 2003), or augmenting with model- based phrase pair posterior (Deng and Byrne, 2005). Using relative frequency as translation probabil- ity is a common practice to measure goodness of a phrase pair. Since most phrases appear only a few times in training data, a phrase pair translation is also evaluated by lexical weights (Koehn et al., 2003) or term weighting (Zhao et al., 2004) as addi- tional features to avoid overestimation. The transla- tion probability can also be discriminatively trained such as in Tillmann and Zhang (2006). The focus of this paper is the phrase pair extrac- tion problem. As in information retrieval, precision and recall issues need to be addressed with a right balance for building a phrase translation table. High precision requires that identified translation candi- dates are accurate, while high recall wants as much valid phrase pairs as possible to …

Content truncated to 500 words