Graph Propagation for Paraphrasing Out-of-Vocabulary Words in Statistical Machine Translation

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 1105–1115, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics Graph Propagation for Paraphrasing Out-of-Vocabulary Words in Statistical Machine Translation∗ Majid Razmara1 Maryam Siahbani1 Gholamreza Haffari2 Anoop Sarkar1 1 Simon Fraser University, Burnaby, BC, Canada {razmara,msiahban,anoop}@sfu.ca 2 Monash University, Clayton, VIC, Australia reza@monash.edu Abstract Out-of-vocabulary (oov) words or phrases still remain a challenge in statistical machine translation especially when a limited amount of parallel text is available for training or when there is a domain shift from training data to test data. In this paper, we propose a novel approach to finding translations for oov words. We induce a lexicon by constructing a graph on source language monolingual text and employ a graph propagation technique in order to find translations for all the source language phrases. Our method differs from previous approaches by adopting a graph propagation approach that takes into account not only one-step (from oov directly to a source language phrase that has a translation) but multi-step paraphrases from oov source language words to other source language phrases and eventually to target language transla- tions. Experimental results show that our graph propagation method significantly improves per- formance over two strong baselines under intrin- sic and extrinsic evaluation metrics. 1 Introduction Out-of-vocabulary (oov) words or phrases still re- main a challenge in statistical machine translation. SMT systems usually copy unknown words verba- tim to the target language output. Although this is helpful in translating a small fraction of oovs such as named entities for languages with same writ- ing systems, it harms the translation in other types of oovs and distant language pairs. In general, copied-over oovs are a hindrance to fluent, high quality translation, and we can see evidence of this in automatic measures such as BLEU (Papineni et al., 2002) and also in human evaluation scores such as HTER. The problem becomes more se- vere when only a limited amount of parallel text is available for training or when the training and test data are from different domains. Even noisy trans- lation of oovs can aid the language model to better ∗This research was partially supported by an NSERC, Canada (RGPIN: 264905) grant. The third author was sup- ported by an early career research award from Monash Uni- versity to visit Simon Fraser University. re-order the words in the target language (Zhang et al., 2012). Increasing the size of the parallel data can re- duce the number of oovs. However, there will al- ways be some words or phrases that are new to the system and finding ways to translate such words or phrases will be beneficial to the system. Re- searchers have applied a number of approaches to tackle this problem. Some approaches use pivot languages (Callison-Burch et al., 2006) while oth- ers use lexicon-induction-based approaches from source language monolingual corpora (Koehn and Knight, 2002; Garera et al., 2009; Marton et al., 2009). Pivot language techniques tackle this problem by taking advantage …

Content truncated to 500 words