Rich Source-Side Context for Statistical Machine Translation

Proceedings of the Third Workshop on Statistical Machine Translation, pages 9–17, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics Rich Source-Side Context for Statistical Machine Translation Kevin Gimpel and Noah A. Smith Language Technologies Institute School of Computer Science Carnegie Mellon University Pittsburgh, PA 15213, USA {kgimpel,nasmith}@cs.cmu.edu Abstract We explore the augmentation of statistical ma- chine translation models with features of the context of each phrase to be translated. This work extends several existing threads of re- search in statistical MT, including the use of context in example-based machine translation (Carl and Way, 2003) and the incorporation of word sense disambiguation into a translation model (Chan et al., 2007). The context fea- tures we consider use surrounding words and part-of-speech tags, local syntactic structure, and other properties of the source language sentence to help predict each phrase’s transla- tion. Our approach requires very little compu- tation beyond the standard phrase extraction algorithm and scales well to large data sce- narios. We report significant improvements in automatic evaluation scores for Chinese- to-English and English-to-German translation, and also describe our entry in the WMT-08 shared task based on this approach. 1 Introduction Machine translation (MT) by statistical modeling of bilingual phrases is one of the most successful ap- proaches in the past few years. Phrase-based MT systems are straightforward to train from parallel corpora (Koehn et al., 2003) and, like the origi- nal IBM models (Brown et al., 1990), benefit from standard language models built on large monolin- gual, target-language corpora (Brants et al., 2007). Many of these systems perform well in competitive evaluations and scale well to large-data situations (NIST, 2006; Callison-Burch et al., 2007). End-to- end phrase-based MT systems can be built entirely from freely-available tools (Koehn et al., 2007). We follow the approach of Koehn et al. (2003), in which we translate a source-language sentence f into the target-language sentence ˆe that maximizes a linear combination of features and weights:1 ⟨ˆe, ˆa⟩ = argmax ⟨e,a⟩ score(e, a, f) (1) = argmax ⟨e,a⟩ M � m=1 λmhm(e, a, f) (2) where a represents the segmentation of e and f into phrases and a correspondence between phrases, and each hm is a R-valued feature with learned weight λm. The translation is typically found us- ing beam search (Koehn et al., 2003). The weights ⟨λ1, ..., λM⟩ are typically learned to directly mini- mize a standard evaluation criterion on development data (e.g., the BLEU score; Papineni et al., (2002)) using numerical search (Och, 2003). Many features are used in phrase-based MT, but nearly ubiquitous are estimates of the conditional translation probabilities p(ej i | f ℓ k) and p(f ℓ k | ej i) for each phrase pair ⟨ej i, f l k⟩ in the candidate sen- tence pair.2 In this paper, we add and evaluate fea- 1In the statistical MT literature, this is often referred to as a “log-linear model,” but since the score is normalized during nei- ther parameter training nor decoding, and is never interpreted as a …

Content truncated to 500 words