A Model of the MT Lexicon for Verbs as an Interface between Syntactic Parsing and Semantic Representations

Machine translation (MT) deals with a large amount of data which represents sentence meanings constructed on the basis of the grammatical rules and the lexical knowledge of the source and target languages. As it is hardly possible to make an automated semantic analysis of the source language data, predominant methods in use have been transfer models such as those of phrase-based statistical MT and pattern-based MT, the main task of which is to replace the source language text with the target language text without understanding them. As it is reported that nearly half the text data consist of unanalyzable multiword like collocations, idioms, and phrasal verbs, it would be inevitable to resort to such a method. At the same time, however, it has been observed that MT output quality suffers to a significant extent when the sentences are not divided into proper segments. This paper proposes a model of the lexicon of verbs whose rich relational information contributes to parsing and understanding sentences. While other existing models of lexicons have no or little consideration for their roles of interaction with processes of syntactic parsing and semantic representation, the proposed model is designed for playing a role as an interface between the two processes in handling sentences with regular usages of verbs as well as idiomatic and metaphoric usages.