Towards a new generation of conversational agents based on sentence similarity

Abstract

The Conversational Agent (CA) is a computer program that can engage in conversation using natural language dialogue with a human participant. Most CAs employ a pattern-matching technique to map user input onto structural patterns of sentences. However, every combination of utterances that a user may send as input must be taken into account when constructing such a script. This chapter was concerned with constructing a novel CA using sentence similarity measures. Examining word meaning rather than structural patterns of sentences meant that scripting was reduced to a couple of natural language sentences per rule as opposed to potentially 100s of patterns. Furthermore, initial results indicate good sentence similarity matching with 13 out of 18 domain-specific user utterances as opposed to that of the traditional pattern matching approach. © 2009 Springer Netherlands.