Encouraging Consistent Translation Choices

2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 417–426, Montr´eal, Canada, June 3-8, 2012. c⃝2012 Association for Computational Linguistics Encouraging Consistent Translation Choices Ferhan Ture,1 Douglas W. Oard,2,4 Philip Resnik3,4 1Department of Computer Science 2College of Information Studies 3Department of Linguistics 4Institute for Advanced Computer Studies University of Maryland, College Park, MD 20740 USA fture@cs.umd.edu, oard@umd.edu, resnik@umd.edu Abstract It has long been observed that monolingual text exhibits a tendency toward “one sense per dis- course,” and it has been argued that a related “one translation per discourse” constraint is op- erative in bilingual contexts as well. In this pa- per, we introduce a novel method using forced decoding to confirm the validity of this con- straint, and we demonstrate that it can be ex- ploited in order to improve machine translation quality. Three ways of incorporating such a preference into a hierarchical phrase-based MT model are proposed, and the approach where all three are combined yields the greatest improve- ments for both Arabic-English and Chinese- English translation experiments. 1 Introduction In statistical Machine Translation (MT), the state-of- the-art approach is to translate phrases in the context of a sentence and to re-order those phrases appro- priately. Intuitively, it seems as if it should also be possible to draw on information outside of a single sentence to further improve translation quality. In this paper, we challenge the conventional approach of translating each sentence independently, and ar- gue that it can indeed also be beneficial to consider document-scale context when translating text. Mo- tivated by the success of a “one sense per discourse” heuristic in Word Sense Disambiguation (WSD), we explore the potential benefit of leveraging a “one translation per discourse” heuristic in MT. The paper is organized as follows. We begin with related work in Section 2. Next, we provide new confirmation that the hypothesized one-translation- per-discourse condition does indeed often hold, based on a novel analysis using forced decoding (Section 3). We incorporate this idea into a hierarchi- cal MT framework by adding three new document- scale features to the translation model (Section 4). We then present experimental results demonstrat- ing solid improvements in translation quality ob- tained by leveraging these features, both for Arabic- English (Ar-En) and Chinese-English (Zh-En) trans- lation (Section 5). Conclusions and future work are presented in Section 6. 2 Related work Exploiting discourse-level context has to date received only limited attention in MT re- search (e.g., (Gim´enez and M`arquez, 2007; Liu et al., 2010; Carpuat, 2009; Brown, 2008; Xiao et al., 2011)). Exploratory analysis of reference trans- lations by Carpuat (2009) motivates a hypothesis that MT systems might benefit from the “one sense per discourse” heuristic, first introduced by Gale et al. (1992), which has proven to be effective in the context of WSD (Yarowsky, 1995). Carpuat’s ap- proach was to do post-processing on the translation output to impose a “one translation per discourse” constraint where the system would otherwise have made a different choice. A …

Content truncated to 500 words