Enriching a statistical machine translation system trained on small parallel corpora with rule-based bilingual phrases

Proceedings of Recent Advances in Natural Language Processing, pages 90–96, Hissar, Bulgaria, 12-14 September 2011. Enriching a Statistical Machine Translation System Trained on Small Parallel Corpora with Rule-Based Bilingual Phrases V´ıctor M. S´anchez-Cartagena, Felipe S´anchez-Mart´ınez, Juan Antonio P´erez-Ortiz Transducens Research Group Departament de Llenguatges i Sistemes Inform`atics Universitat d’Alacant, E-03071, Alacant, Spain {vmsanchez,fsanchez,japerez}@dlsi.ua.es Abstract In this paper, we present a new hybridisa- tion approach consisting of enriching the phrase table of a phrase-based statistical machine translation system with bilingual phrase pairs matching structural transfer rules and dictionary entries from a shallow- transfer rule-based machine translation sys- tem. We have tested this approach on differ- ent small parallel corpora scenarios, where pure statistical machine translation systems suffer from data sparseness. The results obtained show an improvement in trans- lation quality, specially when translating out-of-domain texts that are well covered by the shallow-transfer rule-based machine translation system we have used. 1 Introduction Statistical machine translation (SMT) (Koehn, 2010) is currently the leading paradigm in machine translation research. SMT systems are very attrac- tive because they may be built with little human effort when enough monolingual and bilingual cor- pora are available. However, bilingual corpora large enough to build competitive SMT systems are not always easy to harvest, and they may not even exist for some language pairs. On the contrary, rule-based machine translation systems (RBMT) may be built without any parallel corpus; however, they need an explicit representation of linguistic in- formation whose coding by human experts requires a considerable amount of time. When both parallel corpora and linguistic infor- mation exist, hybrid approaches (Thurmair, 2009) may be followed in order to make the most of such resources. We focus on alleviating the data sparse- ness problem suffered by phrase-based statistical machine translation (PBSMT) systems (Koehn, 2010, ch. 5) when trained on small parallel cor- pora. We present a new hybrid approach which enriches a PBSMT system with resources from shallow-transfer RBMT. Shallow-transfer RBMT systems, which are described in detail below, do not perform a complete syntactic analysis of the input sentences, but rather work with much sim- pler intermediate representations. Hybridisation between shallow-transfer RBMT and SMT has not yet been explored. Existing hybridisation strategies involve more complex RBMT systems (Eisele et al., 2008) which are usually treated as black boxes; in contrast, our approach directly uses the RBMT dictionaries and rules. We provide an exhaustive evaluation of our hy- bridisation approach with two different language pairs: Breton–French and Spanish–English. While the first one suffers from actual resource scarceness, many different parallel corpora are available for the second one, which allows us to test our approach on different domains and check if it is able to im- prove the poor performance of PBSMT systems when translating texts from a domain not covered by the bilingual training data. The rest of the paper is organised as follows. Next section overviews the two systems we com- bine in our approach. Then, section 3 outlines related hybrid approaches, whereas our approach is described in section …

Content truncated to 500 words