Effective Incorporation of Source Syntax into Hierarchical Phrase-based Translation

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 2064–2074, Dublin, Ireland, August 23-29 2014. Effective Incorporation of Source Syntax into Hierarchical Phrase-based Translation Tong Xiao†‡, Adri`a de Gispert§, Jingbo Zhu†‡, Bill Byrne§ † Northeastern University, Shenyang 110819, China ‡ Hangzhou YaTuo Company, Hangzhou 310012, China § University of Cambridge, CB2 1PZ Cambridge, U.K. {xiaotong,zhujingbo}@mail.neu.edu.cn {ad465,wjb31}@eng.cam.ac.uk Abstract In this paper we explicitly consider source language syntactic information in both rule extraction and decoding for hierarchical phrase-based translation. We obtain tree-to-string rules by the GHKM method and use them to complement Hiero-style rules. All these rules are then employed to decode new sentences with source language parse trees. We experiment with our approach in a state-of-the-art Chinese-English system and demonstrate +1.2 and +0.8 BLEU improvements on the NIST newswire and web evaluation data of MT08 and MT12. 1 Introduction Synchronous context free grammars (SCFGs) are widely used in statistical machine translation (SMT), with hierarchical phrase-based translation (Chiang, 2005) as the dominant approach. Hiero grammars are easily extracted from word-aligned parallel corpora and can capture complex nested translation re- lationships. Hiero grammars are formally syntactic, but rules are not constrained by source or target language syntax. This lack of constraint can lead to intractable decoding and bad performance due to the over-generation of derivations in translation. To avoid these problems, the extraction and application of SCFG rules is typically constrained by a source language span limit; (non-glue) rules are lexicalised; and rules are limited to two non-terminals which are not allowed to be adjacent in the source language. These constraints can yield good performing translation systems, although at a sacrifice in the ability to model long-distance movement and complex reordering of multiple constituents. By contrast, the GHKM approach to translation (Galley et al., 2006) relies on a syntactic parse on either the source or target language side to guide SCFG extraction and translation. The parse tree provides linguistically-motivated constraints both in grammar extraction and in translation. This allows for looser span constraints; rules need not be lexicalised; and rules can have more than two non-terminals to model complex reordering multiple constituents. There are also modelling benefits as more meaningful features can be used to encourage derivations with ”well-formed” syntactic tree structures. However, GHKM can have robustness problems in that translation relies on the quality of the parse tree and the diversity of rule types can lead to sparsity and limited coverage. In this paper we describe a simple but effective approach to introducing source language syntax into hierarchical phrase-based translation to get the benefits of both approaches. Unlike previous work, we do not resort to soft/hard syntactic constraints (Marton and Resnik, 2008; Li et al., 2013) or Hiero-style rule extraction algorithms for incorporating syntactic annotation into SCFGs (Zollmann and Venugopal, 2006; Zhao and Al-Onaizan, 2008; Chiang, 2010). We instead use GHKM syntactic rules to augment the baseline Hiero grammar and decoder. Our approach uses GHKM rules if possible and Hiero rules if not. We report performance on …

Content truncated to 500 words