Transfer in experience-guided machine translation

MT Summit VII Sept. 1999 Transfer in Experience-Guided Machine Translation Gang Zhao Junichi Tsujii Lernout & Hauspie Speech Products Information Science, University of Tokyo Language Engineering, UMIST Abstract Experience-Guided Machine Translation (EGMT) seeks to represent the transla- tors' knowledge of translation as experi- ences and translates by analogy. The trans- fer in EGMT finds the experiences most similar to a new text and its parts, seg- ments it into units of translation and trans- lates them by analogy to the experiences and then assembles them into a whole. A research prototype of analogical transfer from Chinese to English is built to prove the viability of the approach in the ex- ploration of new architecture of machine translation. The paper discusses how the experiences are represented and selected with respect to a new text. It describes how units of translation are defined, par- tial translation is derived and composed into a whole. 1 INTRODUCTION Translation is ‘negotiation’ [10] between two language systems in specific contexts. It involves two opera- tions: segmentation and transference. Segmentation [11, 30] breaks up the source text into units of transla- tion [8, 12, 21, 29, 30]. Transference establishes and combines translation equivalence. The knowledge of translation equivalence is not treated as it should be in dominant machine transla- tion architectures. The transfer systems [2, 14, 17, 23, 28]. based on theories of grammar, recognise the ne- cessity of descriptions of translation correspondences but tend to over-simplify the issue of units of transla- tion. They implicitly assume that the source-language constructs derived according to a monolingual gram- mar constitute units of translation for bilingual con- sideration [27]. In interlingua systems [4, 9, 18, 28], the translation is seen as a process of characterising the source text in terms of intermediate expression and verbalising the intermediate expression in the tar- get language. The underlying assumption is that the knowledge of professional translators is nothing dif- ferent from that of mere bilingual speakers, hence no need for contrastive knowledge. The difficulty in an expert-system approach to ma- chine translation comes from the modelling and pro- cessing the translators’ expertise. A formal and generic description of segmentation and transference is prob- lematic [3, 19, 24]. Generalisation may be made but with too many exceptions for effective knowledge pro- cessing. In addition, the justification of an equivalence in a particular context of translation involves multi- disciplinary factors. ‘Since these factors belong to a variety of different areas of life, there is a question whether a comprehensive account of translation in the form of a coherent and homogeneous theory can ever be achieved.’ [7] This paper explores the possibility of describing the translator’s knowledge in concrete terms, general- ising from the specific experience of translation at run time and translating by analogy to the most appro- priate experience. In this approach, the translation engine computes the similarity of a new text and ex- perience texts, identifies the best analogue (the most similar experience) and transfers the new text by the proposal from …

Content truncated to 500 words