Bootstrapping Arabic-Italian SMT through Comparable Texts and Pivot Translation

Bootstrapping Arabic-Italian SMT through Comparable Texts and Pivot Translation Mauro Cettolo Nicola Bertoldi Marcello Federico FBK - Fondazione Bruno Kessler via Sommarive 18 - 38123 Povo, Trento, Italy {cettolo,federico,bertoldi}@fbk.eu Abstract This paper describes efforts towards the development of an Arabic to Italian SMT system for the news domain. Since only very little parallel data are available for this language pair, we investigated both the exploitation of comparable corpora and pivot translation. Experimental evaluation was conducted on a new benchmark devel- oped by extending two Arabic-to-English NIST evaluation sets. Preliminary results show potentials of both approaches with respect to performance achieved by a pop- ular state-of-the-art Web-based translation service. 1 Introduction Statistical machine translation (SMT) has shown to be a very competitive approach and almost lan- guage independent, at least in its basic setting. The concept that phrase-based SMT only needs one basic ingredient, parallel data, has propagated the belief that SMT systems can be easily ported across any language pair, given that sufficient par- allel data are available. Such optimistic view has been also supported by prominent research projects, such as EuroMatrix,1 and industrial ef- forts, mainly Google, which successfully devel- oped SMT systems for a large number of transla- tion directions. These initiatives however coped with SMT’s data hunger in different ways. The EuroMatrix project took advantage of an existing and very popular multilingual parallel corpus (Eu- roParl) covering all official EU languages, while Google could rely on its impressive in-house Web c⃝ 2011 European Association for Machine Translation. 1www.euromatrix.net mining and computing infrastructure, which are doubtlessly unaffordable by most research labs. Unfortunately, for most research labs, a major bottleneck that hinders the development of SMT systems for many language pairs is the lack of suf- ficient amounts of parallel data. In fact, parallel data are a scarce resource even for many socially and economically relevant language pairs,2 such as Arabic and Italian. In this work, we report on our efforts to set-up SMT baselines for news translation from Arabic to Italian, for which only a small amount of parallel texts was available. To compensate for the lack of data, we investigated and compared two alterna- tives: exploiting a fair amount of comparable data, automatically collected from the Web, and pivot translation through English. Results seem to re- ward the pivot translation and suggest further po- tential by combining the direct and pivot transla- tion methods. Experiments have been conducted on a new benchmark developed by the EUROMATRIXPLUS project,3 that extends two news data sets, em- ployed in the 2009 Arabic–English NIST MT eval- uation campaign. In particular, the Arabic side of the newswire evaluations sets was translated by experts both into French and into Italian.4 This benchmark was produced for the sake of fairly comparing direct versus pivot translation for two directions (Arabic–Italian/French) and two pivot languages (English and French), respectively. As of today, French has not been involved at all in our investigation: it will be considered in the near fu- 2See for example: www.translated.net/en/languages-that- matter 3www.euromatrixplus.net …

Content truncated to 500 words