The GREYC machine translation system for the IWSLT 2007 evaluation campaign

The GREYC Machine Translation System for the IWSLT 2007 Evaluation Campaign Yves Lepage and Adrien Lardilleux GREYC, University of Caen Basse-Normandie, France Yves.Lepage@info.unicaen.fr, Adrien.Lardilleux@etu.info.unicaen.fr Abstract The GREYC machine translation (MT) system is a slight evo- lution of the ALEPH machine translation system that partic- ipated in the IWLST 2005 campaign. It is a pure example- based MT system that exploits proportional analogies. The training data used for this campaign were limited on purpose to the sole data provided by the organizers. However, the training data were expanded with the results of sub-sentential alignments. The system participated in the two classical tasks of translation of manually transcribed texts from Japanese to English and Arabic to English. 1. Introduction This paper gives a sketch of the GREYC machine transla- tion system that participated in the IWSLT 2007 evaluation campaign. This system is an evolution of the ALEPH ma- chine translation system that participated in the IWSLT 2005 evaluation campaign [1]. In particular, the core engine has been redesigned. It is still basically the same example-based machine translation system, but it now adds an alignment preprocessing phase. This system participated in the two classical tasks of IWLST 2007: the translation of read speech from Japanese into English and from Arabic into English in the so-called clean conditions, i.e., manually transcribed texts. Our main goal is to assess how much can be achieved by example- based techniques using the minimum amount of data and to show progress year after year. In this respect, we did not use any data outside of the training data sets provided by the or- ganizers: 40,000 aligned utterances in Japanese-English and 20,000 in Arabic-English. No use was made of the develop- ment sets and no use was made of any extra resources or tools like the ones listed on the IWSLT 2007 resources page.1 It was shown in previous experiments and during a pre- vious participation to IWSLT that the system was unable to translate in the absence of sufficient data because the core en- gine did not compile the data in any respect. 20,000 aligned sentences were shown to be too small to get any transla- tion, and consequently, for the previous participation in the Japanese-English IWSLT 2005 track, 140,000 extra aligned sentences from the BTEC corpus were used in addition to the 1http://iwslt07.itc.it/menu/resources.html training data provided by the organizers. [2] reports a fall in BLEU from 0.53 with a training corpus of 160,000 aligned sentences to 0.42 when using only a quarter of the same train- ing corpus. The challenge we addressed this year was pre- cisely to cope with the provided data alone. Our answer was to inflate the training data by adding sub-sentential align- ment data obtained from the training data using example- based techniques. An example in Section 4.3 will explain why sub-sentential alignments may increase the coverage of sentences translated by analogy. 2. Preprocessing of the data We applied two basic preprocessing methods on the training data as well as on the test …

Content truncated to 500 words