In lieu of an abstract, here is a brief excerpt of the content:

180LANGUAGE, VOLUME 66, NUMBER 1 (1990) tence is, I maintain, obtainable only by considering it from within the context of general communicative competence. This surely requires attention to cultural and social knowledge networks from which speakers/writers operate. Department of Linguistics[Received 2 October 1989.] College of Arts and Sciences University of Oregon Eugene, OR 97403-1290 Machine translation: Linguistic characteristics of MT systems and general methodology of evaluation. By John Lehrberger and Laurent Bourbeau. (Linguistics Investigationes Supplementa, Studies in French & General Linguistics/Études en Linguistique Française et Générale, 15.) Amsterdam: Benjamins, 1988. Pp. 228. Reviewed by Jessie Pinkham, DePaul University Machine translation is a sound, well-documented, clearly-written introduction to the issues encountered in the evaluation of machine translation (MT) systems. Lehrberger and Bourbeau acquired their experience in MT as members of the group for research in automatic translation at the University of Montreal (TAUM), and their long-term involvement with the field is demonstrated by the comprehensive overview of the issues that they offer, illustrated with pertinent examples. The book is written with the MT system user—or perhaps the buyer—in mind, rather than the researcher, and its illustrations focus on translation from English to French. The reader with little linguistic or computational training should be able to follow the text without effort; readers trained in these fields may find the discussion superficial, however. The first project of the TAUM group at the University of Montreal was the TAUM-METEO system, which was designed to translate weather reports from English to French. This system, which is still in use, operates in a limited domain and translates a large percentage of the reports automatically, i.e. without any human correction—unlike the majority of existing MT systems, which require post-editing by humans. A subsequent prototype, dubbed TAUMAVIATION , was developed to extend these successes to aviation repair manuals . As it turned out, however, the text of these manuals did not constitute a sufficiently limited domain, and AVIATION did not perform as well as the METEO system. After an independent review of the project, the Canadian government discontinued funding for AVIATION (Isabelle & Bourbeau 1987). The authors relate their experiences as a guide to evaluating MT systems. This begins as a tour of possible designs, from direct to pivot-language-based systems on the one hand and from fully automatic systems to translation aids on the other. A direct translation is one which makes no attempt to analyze the structure of the text. In its simplest form, it is computerized word-for-word translation of the source language into the target language, with operations (e.g. adjective reordering and agreement) executed on the target-language output . The authors demonstrate that a direct approach cannot be successful; for REVIEWS181 example, they illustrate the difficulty in picking the correct translation for the simple sentence Fresh water enters the small reservoir (L'eau fraîche entre dans le petit réservoir): water is a homograph, i.e., it can be either a noun or a verb; enter can have more than one rendering, and will have to be followed by the preposition dans (in this case, but not others). Transfer from the source language to the target language requires considerable analysis. Systems are most often built to perform syntactic analysis, as well as some semantic analysis, of the source language, resulting in a representation called a pivot language. The pivot approach taken in the TAUMAVIATION system is described in some detail in the book. The most advanced design for MT systems relies on a universal semantic representation, usually called an interlingua. The interlingua replaces the pivot language in the previous design. Because the interlingua is claimed to be universal , the source language and the target language will in theory share an interlingua representation. The advantages of using an interlingua become obvious if one considers the creation of a multilingual translation system. The analysis of an English text into interlingua form could then serve as the semantic representation of any number of languages. However, L&B do not consider the interlingua design practical at present (34): Researchers generally agree that a deeper semantic analysis of the source texts is essential to further progress...

pdf

Share