事象駆動型と予測駆動型を融合した文脈解析手法

抄録

This paper describes an integrated parser for text understanding. In the integrated parser, processes of event-driven and expectation-driven understanding occur as an integral part of the parsing process. That is, mechanisms of event assimilation, event search, and 'top-down' inference are incorporated into a single module. The mechanisms of event assimilation and event search are used to find coherent relations between successive portions of a text (i.e., event-driven understanding). The 'top-down' inference mechanism using common sense knowledge is utilized to expect what is coming next (i.e., expectation-driven understanding). The common sense knowledge for expectation-driven understanding, called hierarchical event structure, is a property inheritance net that organizes a hierarchy of events. The hierarchical event structure is encoded explicitly into computational entities called actors which were proposed by Hewitt. When the integrated parser inputs the word associated with a particular event in the hierarchical event structure, it sends a message to the corresponding actor. The actor expects what kind of events are likely to occur in the text. Events extracted from the text are stored in an actor 'EVENT'. The knowledge for event-driven understanding, which is associated with lexical entries, is utilized to detect the coherent relation between events. Each lexical entry is also encoded into an actor, whereas the knowledge for event-driven understanding is encoded implicitly within the procedural information of the actor. When the actor 'EVENT' receives the message followed by the procedural information, it extracts the coherent relation between the stored events and the input sentence. The contextual information contained in the text is, thus, available to resolve anaphora and to select meanings of ambiguous words during the parsing of a sentence.