A Novel Approach for Constructing Conversational Agents using Sentence Similarity Measures
Zuhair A. Bandar, Keeley A. Crockett · 2008
This paper is concerned with the design of a novel approach for constructing Conversational Agents (CA) using sentence similarity measures. Most CA's scripts are organized into contexts consisting of a number of hierarchically organized rules. Each rule possesses a list of structural patterns of sentences and an associated response. User input is then matched against the patterns and the pre-determined response is sent as output. The huge overhead and maintenance of these CAs undoubtedly suggest scope for alternative approaches. In the proposed approach, sentence similarity measures are employed to examine scripts consisting of natural language sentences. This CA will be evaluated against an existing pattern-scripted CA using a set of domain-specific user utterances. Results indicate a substantial reduction of the proposed CA's knowledge-base and, thus, overall maintenance.