Hybrid machine translation architectures within and beyond the EuroMatrix project

Hybrid Machine Translation Architectures within and beyond the EuroMatrix project Andreas Eisele1,2, Christian Federmann2, Hans Uszkoreit1,2, Herv´e Saint-Amand1, Martin Kay1,3, Michael Jellinghaus1, Sabine Hunsicker1, Teresa Herrmann1, Yu Chen1 1: Saarland University, Saarbr¨ucken, Germany 2: DFKI GmbH, Saarbr¨ucken, Germany 3:Stanford University, USA Abstract. This paper presents two hybrid architectures combining rule- based and statistical machine translation (RBMT and SMT) technology. In the first case, several existing MT engines are combined in a multi- engine setup, and a decoder for SMT is used to select and combine the best expressions proposed by different engines. The other architecture uses lexical entries found using a combination of SMT technology with shallow linguistic processing, which are then included in a rule-based sys- tem. The first architecture has been implemented in the framework of the EuroMatrix project, where results from the recent evaluation campaign give important insights into the strengths and weaknesses of the approach relative to other participating RBMT and SMT systems. The second ar- chitecture has been developed in collaboration with the European Patent Office who uses the enhanced RBMT system in the framework of their European Machine Translation Project. 1 Introduction Recent work on statistical machine translation has led to significant progress in coverage and quality of translation technology [1, 2] but so far, most of this work has focused on translation into English, where relatively simple morphological structure and abundance of monolingual training data helped to compensate for the relative lack of linguistic sophistication of the underlying models. As SMT systems are trained on massive amounts of data, they are typically quite good at capturing implicit knowledge contained in co-occurrence statistics, which can serve as a shallow replacement for the world knowledge that would be required for the resolution of ambiguities and the insertion of information that happens to be missing in the source text but is required to generate well-formed text in the target language. In previous years, decades of work went into the implemen- tation of MT systems (typically rule-based) for frequently used language pairs1, and these systems quite often contain a wealth of linguistic knowledge about the languages involved, such as fairly complete mechanisms for morphological 1 See [3] for an extensive, regularly updated list of commercial MT systems 12th EAMT conference, 22-23 September 2008, Hamburg, Germany 27 and syntactic analysis and generation, as well as a large number of bilingual lexical entries spanning many application domains. It is an interesting challenge to combine the different types of knowledge into integrated systems that could then exploit both linguistic knowledge contained in the rules of one or several conventional MT system(s) and non-linguistic knowledge that can be extracted from large amounts of text. The EuroMatrix project (see www.euromatrix.net) has been exploring such combinations of rule-based and statistical knowledge sources, one of the approaches being an integration of existing rule-based MT systems into a multi-engine architecture. This paper describes several incarna- tions of such multi-engine architectures within the project. A careful analysis of the results will guide us in the choice …

Content truncated to 500 words