First Steps towards Multi-Engine Machine Translation

We motivate our contribution to the shared MT task as a first step towards an integrated architecture that combines advantages of statistical and knowledge-based approaches. Translations were generated using the Pharaoh decoder with tables derived from the provided alignments for all four languages, and for three of them using web-based and locally installed commercial systems. We then applied statistical and heuristic algorithms to select the most promising translation out of each set of candidates obtained from a source sentence. Results and possible refinements are discussed. 1 Motivation and Long-term Perspective ”The problem of robust, efficient and reliable speech-to-speech translation can only be cracked by the combined muscle of deep and shallow processing approaches.” (Wahlster, 2001) Although this statement has been coined in the context of VerbMobil, aiming at translation for direct communication, it appears also realistic for many other translation scenarios, where demands on robustness, coverage, or adaptability on the input side and quality on the output side go beyond today’s technological possibilities. The increasing availability of MT engines and the need for better quality has motivated considerable efforts to combine multiple engines into one “super-engine” that is hopefully better than any of its ingredients, an idea pionieered in (Frederking and Nirenburg, 1994). So far, the larger group of related publications has focused on the task of selecting, from a set of translation candidates obtained from different engines, one translation that looks most promising (Tidhar and Küssner, 2000; Akiba et al., 2001; Callison-Burch and Flournoy, 2001; Akiba et al., 2002; Nomoto, 2004). But also the more challenging problem of decomposing the candidates and re-assembling from the pieces a new sentence, hopefully better than any of the given inputs, has recently gained considerable attention (Rayner and Carter, 1997; Hogan and Frederking, 1998; Bangalore et al., 2001; Jayaraman and Lavie, 2005). Although statistical MT approaches currently come out as winners in most comparative evaluations, it is clear that the achievable quality of methods relying purely on lookup of fixed phrases will be limited by the simple fact that for any given combination of topic, application scenario, language pair, and text style there will never be sufficient amounts of pre-existing translations to satisfy the needs of purely data-driven approaches. Rule-based approaches can exploit the effort that goes into single entries in their knowledge repositories in a broader way, as these entries can be unfolded, via rule applications, into large numbers of possible usages. However, this increased generality comes at significant costs for the acquisition of the required knowledge, which needs to be encoded by specialists in formalisms requiring extensive training to be used. In order to push the limits of today’s MT technology, integrative approaches will have to be developed that combine the relative advantages of