A Method of Analyzing Japanese Speech Act Types

A method of analyzing Japanese utterances is developed for a new dialogue translation method, which is called the Intention Translation method. The analyzing method consists of two stages, (i) extraction of surface speech act types from input utterances, and (ii) extraction of less language-dependent speech act types from surface types. In the first stage, input utterances are analyzed in the unification-based lexicosyntactic, syntactico-semantic way. The syntactico-semantic way permits integrated descriptions of information from various sources, and the lexico-syntactic way provides this analysis method with modularity. This allows descriptions of complex constraints on the uses of predicate constituents and these constraints, especially on the uses of honorific predicate constituents, make it possible to analyze ellipsis related to discourse participants. This first stage is used as the analysis part of NADINE (Natural Dialogue Interpretation Expert) system. In the second stage, surface speech act types are analyzed by using plan recognition inference to obtain less language-dependent speech act types. The plan recognition inference uses a special kind of plan schemata, speech act schemata. The inference is extended to use the unification of surface speech act types with decompositions of schemata instead of simple pattern-matching. This allows bidirectional information flow between decompositions and surface speech act types and then makes it possible to allow supplementation of some ellipsis with expectation of speech acts from higher level plan.