Combining domain and topic adaptation for SMT

Combining Domain and Topic Adaptation for SMT Eva Hasler1 ehasler@ed.ac.uk Barry Haddow1 bhaddow@inf.ed.ac.uk Philipp Koehn1,2 pkoehn@inf.ed.ac.uk 1ILCC, School of Informatics, University of Edinburgh 2Center for Language and Speech Processing, Johns Hopkins University Abstract Recent years have seen increased interest in adapting translation models to test domains that are known in advance as well as using latent topic representations to adapt to unknown test domains. However, the relationship between domains and latent topics is still somewhat un- clear and topic adaptation approaches typically do not make use of domain knowledge in the training data. We show empirically that combining domain and topic adaptation approaches can be beneficial and that topic representations can be used to predict the domain of a test document. Our best combined model yields gains of up to 0.82 BLEU over a domain-adapted translation system and up to 1.67 BLEU over an unadapted system, measured on the stronger of two training conditions. 1 Introduction Domain adaptation is a very active area of research in statistical machine translation (SMT) and there is a large and growing body of work on different techniques to adapt translation and language models (Foster and Kuhn, 2007; Matsoukas et al., 2009; Foster et al., 2010; Sennrich, 2012) to specific target domains that are usually known in advance, for example the news do- main. An extension of the standard domain adaptation task is multi-domain adaptation where a translation system is adapted to several known target domains (see for example Cui et al. (2013)). In cases where the target domains are not assumed to be known, dedicated domain classifiers can be trained and used to automatically predict the target domain and choose an appropriate model based on the prediction (Banerjee et al., 2010). Recently, there has been increased work on the application of topic modelling to translation model adaptation (Gong et al., 2011; Su et al., 2012; Eidelman et al., 2012; Hewavitharana et al., 2013; Hasler et al., 2014a) in an attempt to move away from the notion of a domain as the source of a corpus. Topic adaptation techniques build on the assumption that the origin of sentences and documents is unknown at test time, and the unsupervised nature of topic models is useful for detecting structure across corpus boundaries in training sets to adapt to diverse test sets. While domain adaptation techniques rely on a given, hard clustering of the data, topic adaptation aims to induce a soft clustering that is more suited to the task. While topic models are very useful for detecting and grouping the semantic differences in documents, making use of our knowledge about corpus boundaries in the training data could potentially help to adapt more specifially to style or genre on top of adapting to topics. By predicting the domain label of test documents, we can combine both approaches to translate un- labelled documents from different genres and topics. We show that domain and topic adaptation can be complementary and that finding the right balance between the two could lead …

Content truncated to 500 words