Cut the noise: Mutually reinforcing reordering and alignments for improved machine translation

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 1275–1284, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics Cut the noise: Mutually reinforcing reordering and alignments for improved machine translation Karthik Visweswariah IBM Research India v-karthik@in.ibm.com Mitesh M. Khapra IBM Research India mikhapra@in.ibm.com Ananthakrishnan Ramanathan IBM Research India anandr42@gmail.com Abstract Preordering of a source language sentence to match target word order has proved to be useful for improving machine transla- tion systems. Previous work has shown that a reordering model can be learned from high quality manual word alignments to improve machine translation perfor- mance. In this paper, we focus on further improving the performance of the reorder- ing model (and thereby machine transla- tion) by using a larger corpus of sentence aligned data for which manual word align- ments are not available but automatic ma- chine generated alignments are available. The main challenge we tackle is to gen- erate quality data for training the reorder- ing model in spite of the machine align- ments being noisy. To mitigate the effect of noisy machine alignments, we propose a novel approach that improves reorder- ings produced given noisy alignments and also improves word alignments using in- formation from the reordering model. This approach generates alignments that are 2.6 f-Measure points better than a baseline su- pervised aligner. The data generated al- lows us to train a reordering model that gives an improvement of 1.8 BLEU points on the NIST MT-08 Urdu-English eval- uation set over a reordering model that only uses manual word alignments, and a gain of 5.2 BLEU points over a standard phrase-based baseline. 1 Introduction Dealing with word order differences between source and target languages presents a significant challenge for machine translation systems. Failing to produce target words in the correct order results in machine translation output that is not fluent and is often very hard to understand. These problems are particularly severe when translating between languages which have very different structure. Phrase based systems (Koehn et al., 2003) use lexicalized distortion models (Al-Onaizan and Pa- pineni, 2006; Tillman, 2004) and scores from the target language model to produce words in the cor- rect order in the target language. These systems typically are only able to capture short range re- orderings and the amount of data required to po- tentially capture longer range reordering phenom- ena is prohibitively large. There has been a large body of work showing the efficacy of preordering source sentences using a source parser and applying hand written or auto- matically learned rules (Collins et al., 2005; Wang et al., 2007; Ramanathan et al., 2009; Xia and Mc- Cord, 2004; Genzel, 2010; Visweswariah et al., 2010). Recently, approaches that address the prob- lem of word order differences between the source and target language without requiring a high qual- ity source or target parser have been proposed (DeNero and Uszkoreit, 2011; Visweswariah et al., 2011; Neubig et al., 2012). These methods use a small corpus of manual word alignments …

Content truncated to 500 words