A CORPUS-BASED LOCAL CONTEXT ANALYSIS FOR SPOKEN DIALOGS

Grammatical and semantic constraints are effective for interpreting or understanding linguistic expressions. However, they appear to be inad equate for selecting among several candidates, all of which may be rel atively correct or inadequate grammatically or semantically. Clearly, we humans interpret a linguistic expression contextually even if there are many potential interpretations. This paper introduces an example- based local context analysis method using tagged corpora to deal with contextual selection of linguistic expressions, taking into account th e cohesive nature of spoken dialogues. This method performs calculatio ns for similarity scores between linguistic expressions and for likeli hood scores to select the most suitable expression. Both illocutionary force-based and morpho-syntactic classifications are considered, alon g with the frequencies of existing sets of neighboring linguistic expr essions stored in an example database. An experimental processing unit which performs such local context analysis has been implemented in a bidirectional (English and Japanese) translation prototype system, and has shown its applicability to the selection of context-dependent tra nslation candidates. This local context analysis mechanism can be used with conventional translation systems without contextual processing t o raise translation accuracy.