A fully unsupervised approach for mining parallel data from comparable corpora

A Fully Unsupervised Approach for Mining Parallel Data from Comparable Corpora Do Thi Ngoc Diep1,2, Laurent Besacier1, Eric Castelli2 (1) LIG Laboratory, CNRS/UMR-5217, Grenoble, France (2) MICA Center, CNRS/UMI-2954, Hanoi, Vietnam thi-ngoc-diep.do@imag.fr Abstract This paper presents an unsupervised method for extracting parallel sentence pairs from a comparable corpus. A trans- lation system is used to mine the compa- rable corpus and to detect parallel sen- tence pairs. An iterative process is im- plemented not only to increase the num- ber of extracted parallel sentence pairs but also to improve the overall quality of the translation system. A comparison be- tween this unsupervised method and a semi-supervised method is also pre- sented. The unsupervised method was tested in a hard condition: no available parallel corpus to bootstrap the process and the comparable corpus contained up to 50% of non parallel data. The experi- ments conducted show that the unsuper- vised method can be really applied in the case of lacking parallel data. While pre- liminary experiments are conducted on French-English translation, this unsuper- vised method is also applied successfully to a low e-resourced language pair (French-Vietnamese). 1 Introduction Over the past fifty years of development (Hutchins, 2001), machine translation (MT) has obtained good results when applied to several pairs of languages such as English, French, Italia, Japanese, etc. Many approaches for MT have been proposed, such as: rule-based (direct translation, interlingua-based, transfer-based), corpus-based (statistical, example-based) as well as hybrid approaches. However, research on © 2010 European Association for Machine Translation. statistical MT for low e-resourced languages always faces the challenge of getting enough data to support any particular approach. Statistical machine translation (SMT) uses sta- tistical method based on large parallel bilingual corpora of source and target languages to build a statistical translation model for source/target lan- guages and a statistical language model for target language. The two models and a search module are then used to decode the best translation (Brown et al, 1993; Koehn et al, 2003). Thus, a large parallel bilingual text corpus is a prerequi- site. However, such a corpus is not always avail- able, especially for low e-resourced languages. The most common methods to build parallel corpora consist in automatic methods which col- lect parallel sentence pairs from the Web (Resnik and Smith, 2003; Kilgarriff and Grefenstette, 2003), or alignment methods which extract paral- lel documents/sentences from two monolingual corpora (Koehn, 2005; Gale and Church, 1993, Patry and Langlais, 2005). There is also the method of extracting parallel sentence pairs from a comparable corpus (Zhao and Vogel, 2002; Fung and Cheung, 2004; Munteanu and Marcu, 2006). Abdul-Rauf and Schwenk (2009) present a semi-supervised extracting method requiring an initial parallel corpus in order to build a first SMT system that will be used during the semi- supervised extraction (see more in section 2.1). We assume that in the case of a low e-resourced language pair, even a small parallel corpus might not be available to start developing a SMT sys- tem. So, does a fully unsupervised method, …

Content truncated to 500 words