Domain Adaptation of Statistical Machine Translation using Web-Crawled Resources: A Case Study

Domain Adaptation of Statistical Machine Translation using Web-Crawled Resources: A Case Study 1Faculty of Mathematics and Physics Charles University in Prague Czech Republic pecina@ufal.mff.cuni.cz Pavel Pecina1, Antonio Toral2, Vassilis Papavassiliou3, Prokopis Prokopidis3, Josef van Genabith2 2School of Computing Dublin City Universiy Dublin 9, Ireland {atoral,josef}@computing.dcu.ie 3Institute for Language and Speech Processing, Athena RIC Athens, Greece {vpapa,prokopis}@ilsp.gr Abstract We tackle the problem of domain adapta- tion of Statistical Machine Translation by exploiting domain-specific data acquired by domain-focused web-crawling. We de- sign and evaluate a procedure for auto- matic acquisition of monolingual and par- allel data and their exploitation for train- ing, tuning, and testing in a phrase-based Statistical Machine Translation system. We present a strategy for using such resources depending on their availability and quan- tity supported by results of a large-scale evaluation on the domains of Natural En- vironment and Labour Legislation and two language pairs: English–French, English- -Greek. The average observed increase of BLEU is substantial at 49.5% relative. 1 Introduction Recent advances of Statistical Machine Transla- tion (SMT) have improved Machine Translation (MT) quality to such an extent that it can be suc- cessfully used in industrial processes (Flournoy and Duran, 2009). However, this mostly happens in very specific domains for which ample train- ing data is available (Wu et al., 2008). Using in-domain1 data for training has a substantial ef- fect on the final translation quality: SMT, as any other machine-learning application, is not guaran- teed to perform optimally if the data for training and testing are not identically (and independently) distributed, which is often the case in practice. The main problem is usually vocabulary coverage: spe- cific domain texts typically contain vocabulary that is not likely to be found in texts from other do- mains (Banerjee et al., 2010). Other problems can be caused by divergence in style or genre where the difference is not only in lexis but also in grammar. © 2012 European Association for Machine Translation. 1In this work, in-domain always refers to the domain of test data. In order to achieve optimal performance, an SMT system should be trained on data from the same domain, genre, and style as it is applied to. For many domains, though, in-domain data of a size sufficient to train a full system is hard to find. Recent experiments have shown that even small amounts of such data can be used to adapt a sys- tem to the domain of interest (Koehn et al., 2007). In this work, we present a strategy for automatic web-crawling and cleaning of domain-specific data. Further, our exhaustive experiments, car- ried out for the Natural Environment (env) and Labour Legislation (lab) domains and English– French (EN–FR) and English–Greek (EN–EL) lan- guage pairs (in both directions), demonstrate how the crawled data improves SMT quality. After an overview of related work, we discuss the possibility of adapting a general-domain SMT system by using various types of in-domain data. Then, we present our web-crawling procedure fol- lowed by a description of a series of experiments …

Content truncated to 500 words