Edit Distance: A New Data Selection Criterion for Domain Adaptation in SMT

Proceedings of Recent Advances in Natural Language Processing, pages 727–732, Hissar, Bulgaria, 7-13 September 2013. Edit Distance: A New Data Selection Criterion for Domain Adaptation in SMT Longyue Wang1, Derek F. Wong1, Lidia S. Chao1, Junwen Xing1, Yi Lu1, Isabel Trancoso2 1Natural Language Processing & Portuguese-Chinese Machine Translation La- boratory, University of Macau, Macau S.A.R., China 2L2F Spoken Language Systems Lab, INESC-ID, Lisboa, Portugal {mb15505,derekfw,lidiasc,mb15470,mb25435}@umac.mo, isabel.trancoso@inesc-id.pt Abstract This paper aims at effective use of training da- ta by extracting sentences from large general- domain corpora to adapt statistical machine translation systems to domain-specific data. We regard this task as a problem of filtering training sentences with respect to the target domain1 via different similarity metrics. Thus, we give new insights into when data selection model can best benefit the in-domain transla- tion. Based on the investigation of the state-of- the-art similarity metrics, we propose edit dis- tance as a new data selection criterion for this topic. To evaluate this proposal, we compare it with other methods on a large dataset. Com- parative experiments are conducted on Chi- nese-English travel dialog domain and the re- sults indicate that the proposed approach achieves a significant improvement over the baseline system (+4.36 BLEU) as well as the best rival model (+1.23 BLEU) using a much smaller training subset. This study may have a significant impact on mining very large corpo- ra in a computationally-limited environment. 1 Introduction A well-known problem of statistical machine translation (SMT) (Brown et al., 1993) is that the data-driven system is not guaranteed to perform optimally if the data for training and testing are not identically distributed. Domain adaptation for SMT has been explored at different component 1 It could be modeled by an in-domain corpus or text to be translated. levels: word level, phrase level, sentence level and model level. For example mining unknown words from comparable corpora (Daume III and Jagarlamudi, 2011), weighted phrase extraction (Mansour and Ney, 2012), mixing multiple mod- els (Civera and Juan, 2007; Foster and Kuhn, 2007; Eidelman et al., 2012), etc. Recently, data selection as a simple and effective way for this special task has attracted attention. Under the assumption that there exists a large general-domain corpus (general corpus) includ- ing sufficient domains, the task of data selection is to translate a domain-specific text using the optimized translation model (TM) or language model (LM) trained by less but more suitable data retrieved from the general corpus. To state it formally, R is an abstract model of target domain and sG is a sentence or a sentences pair in the general corpus G. The score of each sG is given by ( ) ( , ) G G Score s Sim s R  (1) which means if we could find a better function to measure the similarity between sG and R, G could be replaced by a new sub-corpus Gsub for training a domain-specific SMT system. We focus on two data selection criteria that have been explored for domain adaptation. One comes …

Content truncated to 500 words