Automatic Tune Set Generation for Machine Translation with Limited Indomain Data

Automatic Tune Set Generation for Machine Translation with Lim- ited In-domain Data Jinying Chen, Jacob Devlin, Huaigu Cao, Rohit Prasad, and Premkumar Natarajan Raytheon BBN Technologies 10 Moulton Street, Cambridge 02138, USA {jchen, jdevlin, hcao, rprasad, pnatarajan}@bbn.com Abstract Many effective adaptation techniques for statistical machine translation crucially rely on in-domain development sets to learn model parameters. In this paper we present a novel method that automatically generates the matching tune set for Ara- bic-to-English MT with limited in- domain data1. This technique improves our MT system over two baselines (tuned on data from the same domain but differ- ent genres) by 1.2 and 3.5 BLEU points using significantly less tuning data (1/6 and 1/2 of the baselines). Lexical and morphological features contribute to the success of our method in different ways. Generating tune sets using length distri- bution also improves the system signifi- cantly. Finally, our method obtains com- petitive results in experiments where in- genre tune sets are available. 1 Introduction Adapting statistical machine translation (SMT) systems to different domains is a well-known and challenging problem. Many effective techniques developed for SMT adaptation crucially rely on in-domain development sets to learn model pa- rameters or interpolation weights. For example, Koehn and Schroeder (2007); Ueffing et al., (2007); Matsoukas et al. (2009); Foster et al., (2010), to name a few. However, in some situa- tions, in-domain data can be so limited and in a © 2012 European Association for Machine Translation. 1 This paper is based upon work supported by the DARPA MADCAT Program (Approved for Public Release, Distri- bution Unlimited). The views expressed are those of the authors and do not reflect the official policy or position of the Department of Defense or the U.S. Government. few cases, no matching tune sets are available. Our problem falls into this category. Most existing work in domain adaptation for SMT focused on language models, translation models, lexicons and parallel training data (Koehn and Schroeder, 2007; Lü et al., 2007; Wu et al., 2008; Matsoukas et al., 2009; Foster et al., 2010). Tune set adaptation shares a belief with other adaptation techniques that using training data similar to the test set (in domain, topic, and style) plays a critical role in SMT performance. A unique feature of this problem, however, is the high demands of matching quality. Many param- eters in the SMT system are estimated using the tune set, so negative effects caused by noise (e.g., mismatch in topic and translation style) can be propagated easily. In fact, a large number of SMT domain adaptation techniques also adopted a general framework that requires a tune set to learn model parameters (Ueffing et al., 2007) or interpolation weights for data from different do- mains (Koehn and Schroeder, 2007; Matsoukas et al., 2009; Foster et al., 2010). In this paper we present a highly effective method that automatically generates matching tune sets for an Arabic-to-English MT task with considerably limited in-domain data. Our method is based on the nearest neighbor approach and a …

Content truncated to 500 words