Multi-Target Machine Translation with Multi-Synchronous Context-free Grammars

Human Language Technologies: The 2015 Annual Conference of the North American Chapter of the ACL, pages 293–302, Denver, Colorado, May 31 – June 5, 2015. c⃝2015 Association for Computational Linguistics Multi-Target Machine Translation with Multi-Synchronous Context-free Grammars Graham Neubig, Philip Arthur, Kevin Duh Graduate School of Information Science Nara Institute of Science and Technology 8916-5 Takayama-cho, Ikoma-shi, Nara, Japan {neubig,philip.arthur.om0,kevinduh}@is.naist.jp Abstract We propose a method for simultaneously translating from a single source language to multiple target languages T1, T2, etc. The mo- tivation behind this method is that if we only have a weak language model for T1 and trans- lations in T1 and T2 are associated, we can use the information from a strong language model over T2 to disambiguate the translations in T1, providing better translation results. As a spe- cific framework to realize multi-target transla- tion, we expand the formalism of synchronous context-free grammars to handle multiple tar- gets, and describe methods for rule extraction, scoring, pruning, and search with these mod- els. Experiments find that multi-target transla- tion with a strong language model in a similar second target language can provide gains of up to 0.8-1.5 BLEU points.1 1 Introduction In statistical machine translation (SMT), the great majority of work focuses on translation of a sin- gle language pair, from the source F to the tar- get E. However, in many actual translation situa- tions, identical documents are translated not from one language to another, but between a large num- ber of different languages. Examples of this abound in commercial translation, and prominent open data sets used widely by the MT community include UN documents in 6 languages (Eisele and Chen, 2010), European Parliament Proceedings in 21 languages 1Code and data to replicate the experiments can be found at http://phontron.com/project/naacl2015 Kyoto Protocol Kyoto Sanction Kyoto Treaty وتويك ةيقافتا 京都议定书 京都制裁 京都条约 ?? Target 2 Target 1 Source Figure 1: An example of multi-target translation, where a second target language is used to assess the quality of the first target language. (Koehn, 2005), and video subtitles on TED in as many as 50 languages (Cettolo et al., 2012). However, despite this abundance of multilingual data, there have been few attempts to take advantage of it. One exception is the multi-source SMT method of Och and Ney (2001), which assumes a situation where we have multiple source sentences, and would like to combine the translations from these sentences to create a better, single target translation. In this paper, we propose a framework of multi- target SMT. In multi-target translation, we translate F to not a single target E, but to a set of sentences E = ⟨E1, E2, . . . , E|E|⟩ in multiple target languages (which we will abbreviate T1, T2, etc.). This, in a way, can be viewed as the automated version of the multi-lingual dissemination of content performed by human translators when creating data for the UN, EuroParl, or TED corpora mentioned above. But what, one might ask, do we expect to gain by …

Content truncated to 500 words