Towards Using Web-Crawled Data for Domain Adaptation in Statistical Machine Translation

Towards Using Web-Crawled Data for Domain Adaptation in Statistical Machine Translation Pavel Pecina, Antonio Toral, Andy Way School of Computing Dublin City Universiy Dublin 9, Ireland {ppecina,atoral,away}@computing.dcu.ie Vassilis Papavassiliou, Prokopis Prokopidis, Maria Giagkou Institute for Language & Speech Processing Artemidos 6 & Epidavrou 151 25 Maroussi, Greece {vpapa,prokopis,mgiagkou}@ilsp.gr Abstract This paper reports on the ongoing work fo- cused on domain adaptation of statistical machine translation using domain-specific data obtained by domain-focused web crawling. We present a strategy for crawl- ing monolingual and parallel data and their exploitation for testing, language mod- elling, and system tuning in a phrase- -based machine translation framework. The proposed approach is evaluated on the domains of Natural Environment and Labour Legislation and two language pairs: English–French and English–Greek. 1 Introduction Performance of a statistical machine translation (SMT) system usually drops when it is applied on data of a different nature than that the system was trained on (Banerjee et al., 2010). As any other machine-learning application, SMT 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 a lot of special vocabulary that is not likely to be found in texts from other domains. Problems are also caused by divergence in style or genre, where the difference is not only in terminology but also in grammar. In order to achieve optimal performance, an SMT system must be trained on data from the same domain, of the same genre, and the same style as that it is applied on. For many domains, such training resources (monolingual and parallel data) are not available in large enough amounts to train a system of a sufficient quality. However, even small amounts of such data can be used to adapt © 2011 European Association for Machine Translation. an existing (general-domain) system to the partic- ular domain (Koehn and Schroeder, 2007). If the data is not available at all, a possible solution is to exploit publicly available data from the web. In this work, we present a strategy for crawling domain-specific texts from the web and their ex- ploitation for domain-adaptation in a phrase-based statistical machine translation (PB-SMT) frame- work. At the current stage, we focus on two re- sources: in-domain parallel data for parameter tun- ing and in-domain monolingual data for language model training. As part of our approach, we also create domain-specific test sets. The evaluation is carried out on the domains of Natural Environment (env) and Labour Legislation (lab) and two lan- guage pairs: English–French and English–Greek. The remaining part of the paper is organized as follows. After an overview of related work, we dis- cuss the possibility of adapting a general-domain SMT system to a specific domain by using various types of in-domain data. Then, we describe our baseline SMT system and the web-crawling strat- egy for monolingual and parallel data. Finally, we report on the results, …

Content truncated to 500 words