Facilitating Translation Using Source Language Paraphrase Lattices

Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 420–429, MIT, Massachusetts, USA, 9-11 October 2010. c⃝2010 Association for Computational Linguistics Facilitating Translation Using Source Language Paraphrase Lattices Jinhua Du, Jie Jiang, Andy Way CNGL, School of Computing Dublin City University, Dublin, Ireland {jdu, jjiang, away}@computing.dcu.ie Abstract For resource-limited language pairs, coverage of the test set by the parallel corpus is an important factor that affects translation qual- ity in two respects: 1) out of vocabulary words; 2) the same information in an input sentence can be expressed in different ways, while current phrase-based SMT systems can- not automatically select an alternative way to transfer the same information. Therefore, given limited data, in order to facilitate trans- lation from the input side, this paper pro- poses a novel method to reduce the transla- tion difficulty using source-side lattice-based paraphrases. We utilise the original phrases from the input sentence and the correspond- ing paraphrases to build a lattice with esti- mated weights for each edge to improve trans- lation quality. Compared to the baseline sys- tem, our method achieves relative improve- ments of 7.07%, 6.78% and 3.63% in terms of BLEU score on small, medium and large- scale English-to-Chinese translation tasks re- spectively. The results show that the proposed method is effective not only for resource- limited language pairs, but also for resource- sufficient pairs to some extent. 1 Introduction In recent years, statistical MT systems have been easy to develop due to the rapid explosion in data availability, especially parallel data. However, in reality there are still many language pairs which lack parallel data, such as Urdu–English, Chinese– Italian, where large amounts of speakers exist for both languages; of course, the problem is far worse for pairs such as Catalan–Irish. For such resource- limited language pairs, sparse amounts of parallel data would cause the word alignment to be inac- curate, which would in turn lead to an inaccurate phrase alignment, and bad translations would re- sult. Callison-Burch et al. (2006) argue that lim- ited amounts of parallel training data can lead to the problem of low coverage in that many phrases en- countered at run-time are not observed in the train- ing data and so their translations will not be learned. Thus, in recent years, research on addressing the problem of unknown words or phrases has become more and more evident for resource-limited lan- guage pairs. Callison-Burch et al. (2006) proposed a novel method which substitutes a paraphrase for an un- known source word or phrase in the input sentence, and then proceeds to use the translation of that para- phrase in the production of the target-language re- sult. Their experiments showed that by translating paraphrases a marked improvement was achieved in coverage and translation quality, especially in the case of unknown words which previously had been left untranslated. However, on a large-scale data set, they did not achieve improvements in terms of auto- matic evaluation. Nakov (2008) proposed another way to use para- phrases in …

Content truncated to 500 words