Domain Adaptation in MT Using Titles in Wikipedia as a Parallel Corpus: Resources and Evaluation

Domain Adaptation in MT using Wikipedia as a Parallel Corpus: Resources and Evaluation Gorka Labaka, I˜naki Alegria, Kepa Sarasola University of the Basque Country {gorka.labaka i.alegria kepa.sarasola}@ehu.eus Abstract This paper presents how a state-of-the-art Statistical Machine Translation system is enriched by using extra in-domain parallel corpora extracted from Wikipedia. We collect corpora from parallel titles and from parallel fragments in comparable articles from Wikipedia editions for English, Spanish and Basque. We carried out an evaluation with a double objective: to evaluate the quality of the extracted data and to evaluate the improvement from using domain-adaptation. We think this enrichment method can be very useful for languages with limited amount of parallel corpora, where in-domain data is crucial to improve the performance of MT systems. The experiments on the Spanish–English language pair improve a baseline trained on the Europarl corpus in more than 2 BLEU points when translating texts from the Computer Science domain. Keywords: machine translation, domain adaptation, Wikipedia 1. Introduction Domain adaptation has recently gained interest within sta- tistical machine translation (SMT) as a way to cope with the performance drop observed when testing conditions de- viate from training conditions. The basic idea is that in- domain training data can be exploited to adapt all compo- nents of an already developed system. However, previous work showed small performance gains after adaptation us- ing limited in-domain bilingual data (Bertoldi and Federico (2009), Daum´e III and Jagarlamudi (2011)). Our system is based on enriching a state-of-the-art SMT system with in-domain parallel corpora extracted from Wikipedia. We carried out an evaluation with a double ob- jective: to evaluate the quality of the extracted data and to evaluate the improvement from the domain-adaptation method. We think this kind of domain adaptation can be very use- ful for languages with limited amount of parallel corpora, where in-domain data is crucial to improve the performance of MT. As a previous step, comparable corpora were extracted from Wikipedia using a graph-based selection (Barr´on- Cede˜no et al., 2015). Then the titles of the collected articles were used as parallel sentences to improve a SMT baseline for the Spanish–English and Basque–English pairs. Finally, for the Spanish–English pair, fragments including parallel sentences are also extracted from the comparable articles in Wikipedia and their contribution to MT evalu- ated. The paper is organized as follows. Section 2 introduces several works on this subject and related topics. In Sec- tion 3 we summarize from Barr´on-Cede˜no et al. (2015) the methodology used to extract comparable corpora from Wikipedia. In Section 4 we describe how the parallel sub- corpora obtained from them are used to feed the SMT trans- lators, which are then evaluated. Finally, we draw the con- clusions and show how this work is being continued. 2. Related work Some effort has been devoted to improving MT using Wikipedia but generally domain-related parallel corpora from Wikipedia are not identified and only small units, mainly named-entities, are used. Wikipedia has been used mainly just with the aim of aug- menting dictionaries (Jones …

Content truncated to 500 words