An approach to conversational agent design using semantic sentence similarity

This paper presents a novel framework for constructing a Semantic-Based Conversational Agent (SCAF). Traditional conversational agents (CA) interpret scripts using structural patterns of sentences, which require the script writer to consider every possible permutation that a user may send as input to the CA. This is a time-consuming process, which takes no consideration of semantic content, working solely with the structural form of the sentence. Furthermore, this has proven to be a high maintenance task that can produce some unforeseen consequences when modifying or introducing new patterns into a script. This invariably results in the script writer reassessing the entire script to prevent such occurrences. Different script writers possess differing levels of skill and as such this can prove to be an exasperating task. The proposed SCAF interprets scripts consisting of natural language sentences by means of a semantic sentence similarity measure. User input is measured semantically against the natural language sentences of the context in order to respond with an appropriate output. Such scripting is effortless and alleviates the burden of the traditional pattern-scripted methodologies. Evaluation of the framework has highlighted its potential and shown improvements on traditional CAs.