Domain Adaptation in Statistical Machine Translation using Factored Translation Models

Domain Adaptation in Statistical Machine Translation using Factored Translation Models Jan Niehues and Alex Waibel Institute for Anthropomatics Karlsruhe Institute for Technology Karlsruhe, Germany {jan.niehues,alex.waibel}@kit.edu Abstract In recent years the performance of SMT increased in domains with enough train- ing data. But under real-world conditions, it is often not possible to collect enough parallel data. We propose an approach to adapt an SMT system using small amounts of parallel in-domain data by introducing the corpus identifier (corpus id) as an ad- ditional target factor. Then we added fea- tures to model the generation of the tags and features to judge a sequence of tags. Using this approach we could improve the translation performance in two domains by up to 1 BLEU point when translating from German to English. 1 Introduction Statistical machine translation (SMT) is currently the most promising approach to machine transla- tion of large vocabulary tasks. The approach was first presented in Brown et al. (1993) and has been used in many translation systems since then. One drawback of this approach is that large amounts of training data are needed. Furthermore, the performance of the SMT system improves if this data is selected from a similar topic and from a similar genre. Since this is not possible for many real-world scenarios, one approach to overcome this problem is to use all available data to train a general system and to adapt the system using in- domain training data. Factored translation models as presented in Koehn and Hoang (2007) are able to tightly in- tegrate additional knowledge into a phrase-based statistical machine translation system. In most c⃝ 2010 European Association for Machine Translation. cases, the approach is used to incorporate linguis- tic knowledge, such as morphological, syntactic and semantic information. In contrast, we will use the approach to integrate domain knowledge into the system by introducing a corpus identifier (cor- pus id) tag. Using the corpus id as a target word factor en- ables us to adapt the SMT system by introducting two new types of features in the log-linear model in phrase-based SMT systems. First, we will use rela- tive frequencies to model the generation of the cor- pus id tags similar to the translation model features that are used to model the generation of the target words in a standard phrase-based system. Further- more, we can use features comparable to the word count and language model features to judge the generated sequence of corpus id tags. Moreover, using the general framework of factored transla- tion models leads to a simple integration of this ap- proach into state-of-the-art phrase-based systems. The remaining part of the paper is structured as follows: First, we present some related work in the next section. Afterwards, in Section 3, a motiva- tion for and an overview over the presented model is given. In the following two sections the new types of features are introduced. In Section 6 the approaches are evaluated and in the end a conclu- sion is given. 2 Related Work …

Content truncated to 500 words