Phrase Table Induction Using In-Domain Monolingual Data for Domain Adaptation in Statistical Machine Translation

Phrase Table Induction Using In-Domain Monolingual Data for Domain Adaptation in Statistical Machine Translation Benjamin Marie Atsushi Fujita National Institute of Information and Communications Technology 3-5 Hikaridai, Seika-cho, Soraku-gun, Kyoto, 619-0289, Japan {bmarie, atsushi.fujita}@nict.go.jp Abstract We present a new framework to induce an in- domain phrase table from in-domain monolin- gual data that can be used to adapt a general- domain statistical machine translation system to the targeted domain. Our method first compiles sets of phrases in source and target languages separately and generates candidate phrase pairs by taking the Cartesian product of the two phrase sets. It then computes in- expensive features for each candidate phrase pair and filters them using a supervised clas- sifier in order to induce an in-domain phrase table. We experimented on the language pair English–French, both translation directions, in two domains and obtained consistently better results than a strong baseline system that uses an in-domain bilingual lexicon. We also con- ducted an error analysis that showed the in- duced phrase tables proposed useful transla- tions, especially for words and phrases unseen in the parallel data used to train the general- domain baseline system. 1 Introduction In phrase-based statistical machine translation (SMT), translation models are estimated over a large amount of parallel data. In general, using more data leads to a better translation model. When no specific domain is targeted, general-domain1 par- allel data from various domains may be used to 1As in Axelrod et al. (2011), in this paper, we use the term general-domain instead of the commonly used out-of-domain because we assume that the parallel data may contain some in- domain sentence pairs. train a general-purpose SMT system. However, it is well-known that, in training a system to trans- late texts from a specific domain, using in-domain parallel data can lead to a significantly better trans- lation quality (Carpuat et al., 2012). Indeed, when only general-domain parallel data are used, it is un- likely that the translation model can learn expres- sions and their translations specific to the targeted domain. Such expressions will then remain untrans- lated in the in-domain texts to translate. So far, in-domain parallel data have been har- nessed to cover domain-specific expressions and their translations in the translation model. However, even if we can assume the availability of a large quantity of general-domain parallel data, at least for resource-rich language pairs, finding in-domain par- allel data specific to a particular domain remains challenging. In-domain parallel data may not exist for the targeted language pairs or may not be avail- able at hand to train a good translation model. In order to circumvent the lack of in-domain par- allel data, this paper presents a new method to adapt an existing SMT system to a specific domain by in- ducing an in-domain phrase table, i.e., a set of phrase pairs associated with features for decoding, from in- domain monolingual data. As we review in Sec- tion 2, most of the existing methods for inducing phrase tables are not designed, and may …

Content truncated to 500 words