Enriching SMT Training Data via Paraphrasing

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 803–810, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Enriching SMT Training Data via Paraphrasing Wei He1, Shiqi Zhao12, Haifeng Wang2, Ting Liu1 1Research Center for Social Computing and Information Retrieval, Harbin Institute of Technology {whe,tliu}@ir.hit.edu.cn 2Baidu {zhaoshiqi,wanghaifeng}@baidu.com Abstractx This paper proposes a novel method to resolve the coverage problem of SMT system. The method generates paraphrases for source-side sentences of the bilingual parallel data, which are then paired with the target-side sentences to generate new parallel data. Within a statis- tical paraphrase generation framework, we employ an object function, named Sentence Novelty, to select paraphrases which having the most novel information to the bilingual training corpus of the SMT model. Meanwhile, the context is considered via a language model in the source language to ensure the fluency and accuracy of paraphrase substitution. Com- pared to a state-of-the-art phrase based SMT system (Moses), our method achieves an im- provement of 1.66 points in terms of BLEU on a small training corpus which simulates a re- source-poor environment, and 1.06 points on a training corpus of medium size. 1 Introduction Current statistical machine translation (SMT) systems learn how to translate by analyzing bi- lingual parallel corpora. Generally speaking, high-quality translations can be produced when ample training data is available. Previous studies have indicated that the translation quality can be improved by 2 points of BLEU (Papineni et al., 2002) when the size of the parallel data is doubled (Koehn et al., 2003). However, for the so called low density language pairs that do not This work was partially done when the first author was visiting Baidu. have large-scale parallel corpora, limited amount of training data usually leads to a problem of low coverage in that many phrases encountered at run-time have not been observed in the training data. According to Callison-Burch et al. (2006), for a training corpus containing 10,000 words, translations will have been learned for only 10% of the unigrams in the test set. For a training cor- pus containing 100,000 words this increases to 30%. This problem becomes more serious for higher-order n-grams, and for morphologically richer languages. To overcome the coverage problem of SMT, besides the efforts of mining larger parallel cor- pora from various resources, some researchers have investigated to use paraphrasing approaches. The studies can be classified into two categories by the target of paraphrasing: (1) paraphrasing the input source sentences; (2) paraphrasing the training corpus. In the first category, the pro- posed approaches mainly focus on handling n- grams that are unknown to the SMT model. Cal- lison-Burch et al. (2006) and Marton et al. (2009) paraphrase unknown terms in the input sentences using phrasal paraphrases extracted from bilin- gual and monolingual corpora. Mirkin et al. (2009) rewrite unknown terms with entailments and paraphrases acquired from WordNet. Onishi et al. (2010) and Du et al. (2010) build paraph- rase lattices for input sentences and select the best translations using a …

Content truncated to 500 words