Cache-based Document-level Statistical Machine Translation

Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 909–919, Edinburgh, Scotland, UK, July 27–31, 2011. c⃝2011 Association for Computational Linguistics Cache-based Document-level Statistical Machine Translation Zhengxian Gong1 Min Zhang2 Guodong Zhou1* 1 School of Computer Science and Technology Soochow University, Suzhou, China 215006 2 Institute for Infocomm Research, Singapore 138632 {zhxgong, gdzhou}@suda.edu.cn mzhang@i2r.a-star.edu.sg Abstract Statistical machine translation systems are usually trained on a large amount of bilingual sentence pairs and translate one sentence at a time, ignoring document-level information. In this paper, we propose a cache-based approach to document-level translation. Since caches mainly depend on relevant data to supervise subsequent decisions, it is critical to fill the caches with highly-relevant data of a reasonable size. In this paper, we present three kinds of caches to store relevant document-level infor- mation: 1) a dynamic cache, which stores bilin- gual phrase pairs from the best translation hypotheses of previous sentences in the test document; 2) a static cache, which stores rele- vant bilingual phrase pairs extracted from simi- lar bilingual document pairs (i.e. source documents similar to the test document and their corresponding target documents) in the training parallel corpus; 3) a topic cache, which stores the target-side topic words related with the test document in the source-side. In particu- lar, three new features are designed to explore various kinds of document-level information in above three kinds of caches. Evaluation shows the effectiveness of our cache-based approach to document-level translation with the perfor- mance improvement of 0.81 in BLUE score over Moses. Especially, detailed analysis and discussion are presented to give new insights to document-level translation. 1 Introduction During last decade, tremendous work has been done to improve the quality of statistical machine __________________ * Corresponding author. translation (SMT) systems. However, there is still a huge performance gap between the state-of-the- art SMT systems and human translators. Bond (2002) suggested nine ways to improve machine translation by imitating the best practices of human translators (Nida, 1964), with parsing the entire document before translation as the first priority. However, most SMT systems still treat parallel corpora as a list of independent sentence-pairs and ignore document-level information. Document-level information can and should be used to help document-level machine translation. At least, the topic of a document can help choose specific translation candidates, since when taken out of the context from their document, some words, phrases and even sentences may be rather ambiguous and thus difficult to understand. Anoth- er advantage of document-level machine transla- tion is its ability in keeping a consistent translation. However, document-level translation has drawn little attention from the SMT research community. The reasons are manifold. First of all, most of pa- rallel corpora lack the annotation of document boundaries (Tam, 2007). Secondly, although it is easy to incorporate a new feature into the classical log-linear model (Och, 2003), it is difficult to cap- ture document-level information and model it via some simple features. Thirdly, reference transla- tions of a test document written by human transla- …

Content truncated to 500 words