A Multi-Domain Translation Model Framework for Statistical Machine Translation

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 832–840, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics A Multi-Domain Translation Model Framework for Statistical Machine Translation Rico Sennrich Institute of Computational Linguistics University of Zurich Binzm¨uhlestr. 14 CH-8050 Z¨urich sennrich@cl.uzh.ch Holger Schwenk and Walid Aransa LIUM, University of Le Mans 72085 Le Mans cedex 9, France lastname@lium.univ-lemans.fr Abstract While domain adaptation techniques for SMT have proven to be effective at im- proving translation quality, their practical- ity for a multi-domain environment is of- ten limited because of the computational and human costs of developing and main- taining multiple systems adapted to differ- ent domains. We present an architecture that delays the computation of translation model features until decoding, allowing for the application of mixture-modeling techniques at decoding time. We also de- scribe a method for unsupervised adapta- tion with development and test data from multiple domains. Experimental results on two language pairs demonstrate the effec- tiveness of both our translation model ar- chitecture and automatic clustering, with gains of up to 1 BLEU over unadapted sys- tems and single-domain adaptation. 1 Introduction The effectiveness of domain adaptation ap- proaches such as mixture-modeling (Foster and Kuhn, 2007) has been established, and has led to research on a wide array of adaptation techniques in SMT, for instance (Matsoukas et al., 2009; Shah et al., 2012). In all these approaches, adaptation is performed during model training, with respect to a representative development corpus, and the mod- els are kept unchanged when then system is de- ployed. Therefore, when working with multiple and/or unlabelled domains, domain adaptation is often impractical for a number of reasons. Firstly, maintaining multiple systems for each language pair, each adapted to a different domain, is costly in terms of computational and human resources: the full system development pipeline needs to be performed for all identified domains, all the mod- els are separately stored and need to be switched at runtime. This is impractical in many real applica- tions, in particular a web translation service which is faced with texts coming from many different do- mains. Secondly, domain adaptation bears a risk of performance loss. If there is a mismatch be- tween the domain of the development set and the test set, domain adaptation can potentially harm performance compared to an unadapted baseline. We introduce a translation model architecture that delays the computation of features to the de- coding phase. The calculation is based on a vec- tor of component models, with each component providing the sufficient statistics necessary for the computation of the features. With this framework, adaptation to a new domain simply consists of up- dating a weight vector, and multiple domains can be supported by the same system. We also present a clustering approach for un- supervised adaptation in a multi-domain environ- ment. In the development phase, a set of develop- ment data is clustered, and the models are adapted to each cluster. For each sentence that is being decoded, …

Content truncated to 500 words