Generalizing Word Lattice Translation

Proceedings of ACL-08: HLT, pages 1012–1020, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics Generalizing Word Lattice Translation Christopher Dyer∗, Smaranda Muresan, Philip Resnik∗ Laboratory for Computational Linguistics and Information Processing Institute for Advanced Computer Studies ∗Department of Linguistics University of Maryland College Park, MD 20742, USA redpony, smara, resnik AT umd.edu Abstract Word lattice decoding has proven useful in spoken language translation; we argue that it provides a compelling model for translation of text genres, as well. We show that prior work in translating lattices using finite state tech- niques can be naturally extended to more ex- pressive synchronous context-free grammar- based models. Additionally, we resolve a significant complication that non-linear word lattice inputs introduce in reordering mod- els. Our experiments evaluating the approach demonstrate substantial gains for Chinese- English and Arabic-English translation. 1 Introduction When Brown and colleagues introduced statistical machine translation in the early 1990s, their key in- sight – harkening back to Weaver in the late 1940s – was that translation could be viewed as an instance of noisy channel modeling (Brown et al., 1990). They introduced a now standard decomposition that distinguishes modeling sentences in the target lan- guage (language models) from modeling the rela- tionship between source and target language (trans- lation models). Today, virtually all statistical trans- lation systems seek the best hypothesis e for a given input f in the source language, according to ˆe = arg max e Pr(e|f) (1) An exception is the translation of speech recogni- tion output, where the acoustic signal generally un- derdetermines the choice of source word sequence f. There, Bertoldi and others have recently found that, rather than translating a single-best transcrip- tion f, it is advantageous to allow the MT decoder to consider all possibilities for f by encoding the alter- natives compactly as a confusion network or lattice (Bertoldi et al., 2007; Bertoldi and Federico, 2005; Koehn et al., 2007). Why, however, should this advantage be limited to translation from spoken input? Even for text, there are often multiple ways to derive a sequence of words from the input string. Segmentation of Chinese, decompounding in German, morpholog- ical analysis for Arabic — across a wide range of source languages, ambiguity in the input gives rise to multiple possibilities for the source word se- quence. Nonetheless, state-of-the-art systems com- monly identify a single analysis f during a prepro- cessing step, and decode according to the decision rule in (1). In this paper, we go beyond speech translation by showing that lattice decoding can also yield im- provements for text by preserving alternative anal- yses of the input. In addition, we generalize lattice decoding algorithmically, extending it for the first time to hierarchical phrase-based translation (Chi- ang, 2005; Chiang, 2007). Formally, the approach we take can be thought of as a “noisier channel”, where an observed signal o gives rise to a set of source-language strings f′ ∈ F(o) and we seek ˆe = arg max e max f′∈F(o) Pr(e, f′|o) (2) = …

Content truncated to 500 words