Systèmes de traduction automatique et levée d'ambiguïté: étude comparée de systèmes de TABR, TAS et TAN

• Autor(a) principal: Bacquelaine, Françoise
• Data de Publicação: 2018
• Tipo de documento: Livro
• Idioma: fra
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: https://hdl.handle.net/10216/119701
• topic: Humanidades; Humanities

Machine Translation (MT) has been a lively field of research eversince the invention of the computer. Rule-Based Machine Translation (RBMT) wasthe first option back in the 1940s-1950s; Statistical Machine Translation (SMT)appeared a few decades later and Neural Machine Translation (NMT) in the 21stcentury. This distinction is not strict since most MT systems are now hybrid, butnatural language ambiguity is a well-known pitfall, be it in human or machinetranslation.Two types of ambiguity can arise when using the rather common Englishword issue: "grammatical ambiguity" (noun or verb?), on the one hand, and"homographic and polysemic ambiguity (one word form with different sensesin the source language)" (Hutchins 2005: 17), on the other hand. The scope ofthis research is limited to three senses of the noun issue (1. An important topic orproblem for debate or discussion; 2. The action of supplying or distributing an itemfor use, sale, or official; 3. (formal or law) Children of one's own) and two sensesof the verb to issue (1. [WITH OBJECT] Supply or distribute (something) for useor sale; 2. [NO OBJECT] (issue from) Come, go, or flow out from). A sample ofsentences containing at least one example of usage was selected from the BritishNational Corpus in order to test and compare four English-French MT systems:SYSTRAN (free online RBMT), Google Translate (free online SMT), MT@EC(restricted access SMT) and free online Neural Machine Translation by LISA(University of Montreal). The outputs were compared to a human translation modelbased on translation memories (parallel corpora) in order to evaluate weaknessesand strengths of each system, compare the results and find out possible ways ofimproving MT output through hybridisation.Research results in cases like these are not just useful for theoreticallinguistics but can also be used to heighten awareness in human translators anddemonstrate that translators who are trained in computational linguistics can alsowork together with experts in artificial intelligence and machine translation.