Evaluation of SIMPLE (Semantically Improved Matching Procedure for Language Enabled) conversational agents

Abstract

The effectiveness of user interactions with a conversational agent largely depends on the relevance of responses that the system is able to generate for a user utterance. Existing systems primarily employ syntactic templates (i.e. grammar rules and word matching) to indirectly extract meaning from user input. A different method of extracting meaning from input is to determine the semantic distance between the words in one sentence and the words in another. Such a distance metric is made possible by semantic networks, such as WordNet, that link words on relatedness. We present an approach to match user input utterances to agent responses based on a semantic distance metric using the WordNet lexical database and propose a number of uses for our approach in developing conversational agent systems.