Inducing translation templates for example-based machine translation

MT Summit VII Sept. 1999 Inducing Translation Templates for Example-Based Machine Translation Michael Carl Institut für Angewandte Informationsforschung, Martin-Luther-Straße 14. 66111 Saarbrücken, Germany, carl@iai.uni-sb.de Abstract This paper describes an example-based ma- chine translation (EBMT) system which relays on various knowledge resources. Mor- phologic analyses abstract the surface forms of the languages to be translated. A shal- low syntactic rule formalism is used to per- colate features in derivation trees. Trans- lation examples serve the decomposition of the text to be translated and determine the transfer of lexical values into the tar- get language. Translation templates deter- mine the word order of the target language and the type of phrases (e.g. noun phrase, prepositional phase, ...) to be generated in the target language. An induction mecha- nism generalizes translation templates from translation examples. The paper outlines the basic idea underlying the EBMT sys- tem and investigates the possibilities and limits of the translation template induc- tion process. 1 Introduction Until the end of the eighties MT was strongly dom- inated by rule-based systems which deduce transla- tions of natural language texts based on a bilingual lexicon and a grammar. Linguistic theory and gram- mar formalisms have evolved and are ready at hand to support this approach. Unification and feature struc- tures are such computational means which have influ- enced linguistic theories, formalisms and conception of MT systems. However, much human effort is re- quired since, both, the grammar and the lexicon are manually entered into the machine. In 1990 a new epoch started when Brown et al. (1990) presented a paper on statistics-based MT. Translation was no longer considered as deduction but rather as a stochastic process which relies on a number of ran- dom variables. Given a well evolved stochastic the- ory and powerful hardware, random variables could be computed from a sufficiently large set of example translations. Since then machine learning technologies have evol- ved dramatically in such a way that not only statistics but also other learning techniques have entered the MT domain. Artificial neural networks were used for MT (McLean, 1992) or to transfer a source language parse tree into a target language parse tree (Wang and Waibel, 1995). Hiroshi et al. (1996) describe an MT system which makes use of genetic algorithms to optimize translation rules. Other approaches mix more or less shallow linguistic analysis with statis- tics (Knight et al., 1994) or with symbolic inductive methods (Sato and Nagao, 1990; Güvenir and Tunc 1996; Güvenir and Cicekli, 1998; Collins and Cun- ningham, 1997; Collins, 1999). Knight et al. (1994) ranks semantic analyzes of an input sentence by sta- tistical means to filter out least probable analyses. Sato and Nagao (1990) replace subtrees in a tree bank of syntactically analyzed source-target language equiv- alences to obtain target language parse trees. Güvenir and Tunc (1996), Güvenir and Cicekli (1998) Collins and Cunningham (1997) and Collins (1999) in- duce translation templates to map a generalized input sentence into the target language. Different as they are, these approaches have …

Content truncated to 500 words