Translation differences and pragmatics-based MT

TRANSLATION DIFFERENCES AND PRAGMATICS-BASED MT Stephen Helmreich David Farwell Computing Research Laboratory New Mexico State University Las Cruces, New Mexico Abstract: This paper examines differences between two professional translations into English of the same Spanish newspaper article. Among other explanations for these differences, such as outright errors and free variation, we find a significant number of differences that can be explained on the basis of differing beliefs on the part of the translators about the subject matter. In particular, we attribute translation differ- ences to differing beliefs about the probability of future events and of rational or irrational behavior in response to such probability. We discuss the requirements for a pragmatics-based model of translation that would account for these differences. 1. Introduction For the past two years, we have been motivating a pragmatics-based approach to Machine Translation. We have tried to show that the next stage of Symbolic MT is to take the current state-of-the-art KBMT systems and embed them within a pragmatic framework which includes: • nested beliefs spaces representing the differing beliefs of both the participants in the translation process and participants in the events described in the text, • default reasoning about these events and participants to produce and compare possible interpreta- tions which make a coherent whole of all the clues in the text and the context, and • generation based on the most coherent interpretation, taking into account the beliefs of the target audience. We claim that translation (and therefore machine translation) is based on interpretation rather than mean- ing (that is, on an inferred coherent understanding of the intent of the author, rather than on a propositional content expressed by the text). Furthermore, the translation task itself is then defined in terms of re-creat- ing that interpretation within the belief space of the target language audience. Methodologically, we have been examining multiple translations of the same text for evidence that justifies the above premises (Farwell & Helmreich, 1993, 1995, 1996). In each of these papers, we have selected and analyzed particular examples which illustrate aspects of this approach to translation. In this paper, we continue this process, this time by looking at translation variants in an entire text. This approach also turns out to shed some interesting light on the recent ARPA MT evaluation methodology. Section 2 summarizes the results. Section 3 describes our methodology. Section 4 discusses the results in detail. Section 5 sets forth our approach to understanding some of the more significant translation variants, and Section 6 discusses some implications for machine translation and for MT evaluation. 43 2. Thesis If translations are governed by interpretations of the source language document, then it would be reason- able to expect to find evidence of this in examining translation variants throughout a complete text. In the text examined, we find clear evidence that about 45% of the translation variants can be ascribed to differ- ences in the two translators' beliefs. Moreover, a significant portion of these differences can be related to different overall …

Content truncated to 500 words