Hierarchical Phrase Table Combination for Machine Translation

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 802–810, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics Hierarchical Phrase Table Combination for Machine Translation Conghui Zhu1 Taro Watanabe2 Eiichiro Sumita2 Tiejun Zhao1 1School of Computer Science and Technology Harbin Institute of Technology (HIT), Harbin, China 2National Institute of Information and Communication Technology 3-5 Hikari-dai, Seika-cho, Soraku-gun, Kyoto, Japan {chzhu,tjzhao}@mtlab.hit.edu.cn {taro.watanabe,Sumita}@nict.go.jp Abstract Typical statistical machine translation sys- tems are batch trained with a given train- ing data and their performances are large- ly influenced by the amount of data. With the growth of the available data across different domains, it is computationally demanding to perform batch training ev- ery time when new data comes. In face of the problem, we propose an efficient phrase table combination method. In par- ticular, we train a Bayesian phrasal inver- sion transduction grammars for each do- main separately. The learned phrase ta- bles are hierarchically combined as if they are drawn from a hierarchical Pitman-Yor process. The performance measured by BLEU is at least as comparable to the tra- ditional batch training method. Further- more, each phrase table is trained sepa- rately in each domain, and while compu- tational overhead is significantly reduced by training them in parallel. 1 Introduction Statistical machine translation (SMT) system- s usually achieve ’crowd-sourced’ improvements with batch training. Phrase pair extraction, the key step to discover translation knowledge, heav- ily relies on the scale of training data. Typi- cally, the more parallel corpora used, the more phrase pairs and more accurate parameters will be learned, which can obviously be beneficial to improving translation performances. Today, more parallel sentences are drawn from divergent do- mains, and the size keeps growing. Consequent- ly, how to effectively use those data and improve translation performance becomes a challenging is- sue. This joint work was done while the first author visited NICT. Batch retraining is not acceptable for this case, since it demands serious computational overhead when training on a large data set, and it requires us to re-train every time new training data is avail- able. Even if we can handle the large computation cost, improvement is not guaranteed every time we perform batch tuning on the newly updated train- ing data obtained from divergent domains. Tradi- tional domain adaption methods for SMT are also not adequate in this scenario. Most of them have been proposed in order to make translation sys- tems perform better for resource-scarce domain- s when most training data comes from resource- rich domains, and ignore performance on a more generic domain without domain bias (Wang et al., 2012). As an alternative, incremental learning may resolve the gap by incrementally adding da- ta sentence-by-sentence into the training data. S- ince SMT systems trend to employ very large scale training data for translation knowledge extraction, updating several sentence pairs each time will be annihilated in the existing corpus. This paper proposes a new phrase table combi- nation method. First, phrase pairs are extracted from …

Content truncated to 500 words