From the real world to real words: the METEO case

From the Real World to Real Words: The METEO case Philippe Langlais, Thomas Leplus Simona Gandrabur and Guy Lapalme RALI Universit´e de Montr´eal http://rali.iro.umontreal.ca Abstract. Machine Translation (MT) is the focus of extensive scientific in- vestigations driven by regular evaluation campaigns, but which are mostly oriented towards a somewhat artificial task: translating news articles into En- glish. In this paper, we investigate how well current MT approaches deal with a real-world task. We have rationally reconstructed one of the only MT systems in daily production use: the METEO system. We show how a combination of a sentence-based memory approach, a phrase-based statisti- cal engine and a neural-network rescorer can give results comparable to those of the current system while offering a faster development cycle and arguably better customization possibilities. 1 Introduction Machine Translation is a field nowadays strongly anchored in a paradigm of performance. Eval- uation exercises such as those conducted within the TIDES project are pushing system designers to constantly improve their systems. The shared task usually consists in translating news articles excerpts from a foreign language into English. While this is certainly a challenging issue, real life applications of machine translation in a pro- duction setting (i.e. without human revision) will likely be more targeted than newspaper articles. ist.More focused evaluation exercises do ex- Within the IWSLT workshop (Akiba et al., 2004), the main objective was to provide an evaluation framework for spoken language translation technologies. The shared task con- sisted in translating sentences from the Basic Travel Expression Corpus (BTEC) which gath- ers sentences believed to be useful for a tourist in a foreign country. In the Verbmobil project (Wahlster, 2000), transcriptions of spontaneous speech from several narrow domains such as ap- pointment scheduling were translated from Ger- man into English. In this study, we focus on an even more con- crete task and one of the greatest successes of machine translation: the English to French trans- lation of weather forecasts issued by Environ- ment Canada1, which we call here the METEO task. Machine translation of weather forecasts helps guarantee that the translations are produced in a timely manner given the very short life of a weather report. Weather reports are issued every 6 hours. We specifically chose this task because there already exists a fully operational rule-based translation system designed for it, who’s per- formance was carefully measured (Macklovitch (1985)), and because a very large bitext corpus of previously published weather forcasts is avail- able. Thus, we were able to build a variety of corpus-based MT systems targeted on this spe- cific task and compare their respective perfor- mances. 1The current reports are available at http://meteo. ec.gc.ca/forecast/textforecast_f.html 166 EAMT 2005 Conference Proceedings Langlais, Leplus, Grandrabur, Lapalme In the mid seventies, a group of linguists and computer scientists of Universit´e de Montr´eal (TAUM group) developed an English to French weather report machine translation system which became known as the TAUM-METEO sys- tem. Spin-offs of that research system, namely METEO-1 and METEO-2 (Grimaila and Chan- dioux, 1992), have …

Content truncated to 500 words