Efficient Minimum Error Rate Training and Minimum Bayes-Risk Decoding for Translation Hypergraphs and Lattices

Proceedings of the 47th Annual Meeting of the ACL and the 4th IJCNLP of the AFNLP, pages 163–171, Suntec, Singapore, 2-7 August 2009. c⃝2009 ACL and AFNLP Efficient Minimum Error Rate Training and Minimum Bayes-Risk Decoding for Translation Hypergraphs and Lattices Shankar Kumar1 and Wolfgang Macherey1 and Chris Dyer2 and Franz Och1 1Google Inc. 1600 Amphitheatre Pkwy. Mountain View, CA 94043, USA {shankarkumar,wmach,och}@google.com 2Department of Linguistics University of Maryland College Park, MD 20742, USA redpony@umd.edu Abstract Minimum Error Rate Training (MERT) and Minimum Bayes-Risk (MBR) decod- ing are used in most current state-of-the- art Statistical Machine Translation (SMT) systems. The algorithms were originally developed to work with N-best lists of translations, and recently extended to lat- tices that encode many more hypotheses than typical N-best lists. We here extend lattice-based MERT and MBR algorithms to work with hypergraphs that encode a vast number of translations produced by MT systems based on Synchronous Con- text Free Grammars. These algorithms are more efficient than the lattice-based versions presented earlier. We show how MERT can be employed to optimize pa- rameters for MBR decoding. Our exper- iments show speedups from MERT and MBR as well as performance improve- ments from MBR decoding on several lan- guage pairs. 1 Introduction Statistical Machine Translation (SMT) systems have improved considerably by directly using the error criterion in both training and decoding. By doing so, the system can be optimized for the translation task instead of a criterion such as like- lihood that is unrelated to the evaluation met- ric. Two popular techniques that incorporate the error criterion are Minimum Error Rate Train- ing (MERT) (Och, 2003) and Minimum Bayes- Risk (MBR) decoding (Kumar and Byrne, 2004). These two techniques were originally developed for N-best lists of translation hypotheses and re- cently extended to translation lattices (Macherey et al., 2008; Tromble et al., 2008) generated by a phrase-based SMT system (Och and Ney, 2004). Translation lattices contain a significantly higher number of translation alternatives relative to N- best lists. The extension to lattices reduces the runtimes for both MERT and MBR, and gives per- formance improvements from MBR decoding. SMT systems based on synchronous context free grammars (SCFG) (Chiang, 2007; Zollmann and Venugopal, 2006; Galley et al., 2006) have recently been shown to give competitive perfor- mance relative to phrase-based SMT. For these systems, a hypergraph or packed forest provides a compact representation for encoding a huge num- ber of translation hypotheses (Huang, 2008). In this paper, we extend MERT and MBR decoding to work on hypergraphs produced by SCFG-based MT systems. We present algorithms that are more efficient relative to the lattice al- gorithms presented in Macherey et al. (2008; Tromble et al. (2008). Lattice MBR decoding uses a linear approximation to the BLEU score (Pap- ineni et al., 2001); the weights in this linear loss are set heuristically by assuming that n-gram pre- cisions decay exponentially with n. However, this may not be optimal in practice. We employ MERT to select these weights by optimizing …

Content truncated to 500 words