Event Sequence Model for Semantic Analysis of Time and Location in Dialogue System

摘要

It is important for a natural language dialogue system to interpret relations among event concepts appearing in a dialogue. The more complex a dialog becomes, the more essential it becomes for a natural language dialogue system to perform this kind of interpretation. Traditionally, many studies have focused on this problem. Some dialogue systems supported such semantic analysis by using rules and/or models designed for particular scenes involving specific type of dialogue and/or specific problem solving. However, these frameworks require system developers to reconstruct those rules/models even if a slight change is added to the targeted scene. In many cases, their rules/models heavily depend on specific type of dialogue/problem solving, and they do not have high reusability and modularity. Since those rules/models have scene-depending design, they cannot be used to incrementally construct a bigger rule or model. In this research, we focus on a set of event concepts which are usually expected to occur sequentially. In a dialogue, a spoken event concept enables the listeners to guess a sequence of events. The sequence may sometimes be logically inferred, and it may be understood based on general common sense. We believe that a concept model of sequential events can be designed for each bigger event concept that consists of a series of smaller events. Using the sequentiality of the events in the model, a dialogue system can analyze time and location of each event in a dialogue. In this paper, we design a structure of the event sequence model and propose a framework for analyzing time and location of event concepts appearing in a dialogue. We implemented this framework in a dialogue system, and designed some event sequence models. We confirmed that this system could analyze time and location of sequential events without scene-depending rules.