Factors for success and failure in MT

Factors for success (and failure) in Machine Translation — some lessons of the first 50 years of R&D Christian BOITET GETA, CLIPS, Institut IMAG (UJF & CNRS) BP 53, 38041 GRENOBLE cedex 9, France Christian.Boitet@imag.fr Fifth Machine Translation Summit, Luxemburg, 11—13 July 1995 Abstract Factors for success and failure in Machine Translation are extracted from an analysis of the 50 years history of MT. Situations are different in various types of MT (MT for watchers, for revisors, for translators, for authors). Success or failure may be judged at several levels, conceptual, engineering, operational, commercial, communicative, which concern more specifically researchers, developers, users, vendors and funders. For each category, we propose a set of rules for success and highlight pitfalls leading to almost certain failure. It is important to note that political factors linked with funding and decision-making processes may heavily influence the perception of MT in general and of specific systems or tools as successful or not. To sum up, MT is a complex scientific technology. Success can come only if man and machine are put in their right place considering the translational situation, if researchers and developers prefer pragmatic, hybrid solutions over strict adherence to elegant but simplistic theoretical designs, and if unrelated efforts using MT only as an umbrella or as a testbed are excluded from our consideration. Keywords Machine Translation (MT), Machine Assisted Translation (MAT), MT for watchers, MT for revisors, MT for translators, MT for authors, factors for success & failure in MT. Introduction When speaking of successes and failures in MT, we should distinguish between different types of MT, and between different criteria for judging systems. “Machine Aided Translation” (MAT) should be the correct term to cover all techniques and systems aiming at automating the translation process, whether the bulk of the translation process is done by a machine or by a human. In the context of this “MT Summit”, however, we will use “Machine Translation” (MT) instead. We consider four types of MT: - "MT for the watcher" aims at readers wanting to gain access to some information written in a foreign language, and prepared to accept possibly bad “rough” translations (perhaps with a slight brush-up) rather than nothing. We abbreviate this as MT-W or MT-W+ if some brush- up is performed on the rough machine outputs. - "MT for the revisor" aims at producing “raw” translations automatically, with a quality comparable to that of first drafts produced by humans, so that the professional translators, freed from that very boring and time-consuming task, can be promoted to revisors. We abbreviate this as MT-R, or MT-R- if the revision is only a brush-up. - "MT for the translator" aims at helping human translators do their job, by providing on-line dictionaries, thesaurii, translation memories, etc. We abbreviate this as MT-T. - "MT for the author" aims at (possibly monolingual) authors wanting to have their texts translated into one or several languages, and accepting to write under the control of the system, or to help the system …

Content truncated to 500 words