ALEPH: an EBMT system based on the preservation of proportional analogies between sentences across languages

ALEPH: an EBMT system based on the preservation of proportional analogies between sentences across languages Yves Lepage, Etienne Denoual ATR – Spoken language communication labs Keihanna gakken tosi 619-0288 Kyoto, Japan {yves.lepage,etienne.denoual}@atr.jp Abstract We designed, implemented and assessed ALEPH, a pure example-based machine translation system. It strictly does not make any use of variables, templates or training, does not have any explicit transfer component, and does not require any preprocessing of the aligned examples. It relies on a specific operation: the reso- lution of analogical equations, that neutralizes transla- tion divergences in an elegant way. Starting only from theoretical results, a system that is state-of-the-art with the top IWSLT 2004 results could be built in six month time. Evaluated on the Unrestricted Data track of IWSLT 2004, our system achieved second place in CE, and third place in JE (with best BLEU for this latter track). For this year’s evaluation campaign, the features of the sys- tem allowed its immediate application to all possible language pairs in the C-STAR tracks. 1. Introduction We present a novel example-based machine translation system, ALEPH, which relies on an operation specific to language (proportional analogy). We evaluate this system on the tasks proposed during the previous IWSLT 2004 evaluation campaign in the Japanese-English and Chinese-English Unrestricted Data track. This evalua- tion shows that our system would have positioned itself in the top systems for this track (best BLEU in JE track, second position in CE track). This study demonstrates that it is possible to im- plement an EBMT system from scratch in a matter of months (0.5 man/year) and achieve reasonable results. Starting only from the theoretical principles of analogy on strings of characters, building a system that is state- of-the-art with the top IWSLT 2004 results could be achieved in as little as six month time. In the IWSLT 2005 evaluation campaign, this sys- tem competed in all C-STAR tracks. Indeed, an appeal- ing feature of this system is that it requires not train- ing whatsoever: the data are just loaded into memory at startup. As a consequence, it can be applied directly to any language pair for which there is sufficient available data. Following this introduction, the second section of this paper gives a rationale for choosing one and only one basic operation, proportional analogy, to process any sentence of any language. The third section ex- plains in detail the algorithm used to solve proportional analogies, while the fourth section shows its use as a blackbox function to perform translation. The fifth sec- tion recalls the experimental conditions of the Unre- stricted Data track of IWSLT 2004, gives the configu- ration details for the ALEPH system and ranks it among the participants of IWSLT 2004. The sixth section gives the results of this year’s campaign. The final section discusses the results obtained and future research. 2. Divergences across languages [1] quotes a study on a sample of 19, 000 sentences be- tween English and Spanish showing that one translation pair in three …

Content truncated to 500 words