NICT@WMT09: Model Adaptation and Transliteration for Spanish-English SMT

Proceedings of the Fourth Workshop on Statistical Machine Translation , pages 105–109, Athens, Greece, 30 March – 31 March 2009. c⃝2009 Association for Computational Linguistics NICT@WMT09: Model Adaptation and Transliteration for Spanish-English SMT Michael Paul, Andrew Finch and Eiichiro Sumita Language Translation Group MASTAR Project National Institute of Information and Communications Technology Michael.Paul@nict.go.jp Abstract This paper describes the NICT statis- tical machine translation (SMT) system used for the WMT 2009 Shared Task (WMT09) evaluation. We participated in the Spanish-English translation task. The focus of this year’s participation was to in- vestigate model adaptation and transliter- ation techniques in order to improve the translation quality of the baseline phrase- based SMT system. 1 Introduction This paper describes the NICT statistical machine translation (SMT) system used for the shared task of the Fourth Workshop on Statistical Ma- chine Translation. We participated in the Spanish- English translation task under the Constrained Condition. For the training of the SMT engines, we used two parallel Spanish-English corpora pro- vided by the organizers: the Europarl (EP) cor- pus (Koehn, 2005), which consists of 1.4M paral- lel sentences extracted from the proceedings of the European Parliament, and the News Commentary (NC) corpus (Callison-Burch et al., 2008), which consists of 74K parallel sentences taken from ma- jor news outlets like BBC, Der Spiegel, and Le Monde. In order to adapt SMT systems to a specific do- main, recent research focuses on model adapta- tion techniques that adjust their parameters based on information about the evaluation domain (Fos- ter and Kuhn, 2007; Finch and Sumita, 2008a). Statistical models can be trained on in-domain and out-of-domain data sets and combined at run-time using probabilistic weighting between domain-specific statistical models. As the official WMT09 evaluation testset consists of documents taken from the news domain, we applied statistical model adaptation techniques to combine transla- tion models (tm), language models (lm) and dis- tortion models (dm) trained on (a) the in-domain NC corpus and (b) the out-of-domain EP corpus (cf. Section 2). One major problem in the given translation task was the large amount of out-of-vocabulary (OOV) words, i.e., source language words that do not oc- cur in the training corpus. For unknown words, no translation entry is available in the statistical trans- lation model (phrase-table). As a result, these OOV words cannot be translated. Dealing with languages with a rich morphology like Spanish and having a limited amount of bilingual resources make this problem even more severe. There have been several efforts in dealing with OOV words to improve translation quality. In ad- dition to parallel text corpora, external bilingual dictionaries can be exploited to reduce the OOV problem (Okuma et al., 2007). However, these ap- proaches depend on the coverage of the utilized external dictionaries. Data sparseness problems due to inflectional variations were previously addressed by applying word transformations using stemming or lemmati- zation (Popovic and Ney, 2005; Gupta and Fed- erico, 2006). A tight integration of morpho- syntactic information into the translation model was proposed by (Koehn and Hoang, 2007) …

Content truncated to 500 words