Intent Understanding in a Virtual Agent

Abstract

This paper discusses the intent recognition system we have built. this system is to be used as part of a virtual agent that can help resolve end user queries. The end user queries are of different intents -- request for action, request for information, report of some issue, general greetings. Intent detection is a key component of the virtual agent o decide which type the query belongs to and to further invoke the appropriate action modules. The system uses a combination of machine learning and rules based techniques. The rules based component can be used in an unsupervised mode with only the dictionary databases to be loaded upfront. Classifier is a supervised block which requires training data. The system has a feedback based learning which enables the system's performance to improve with use. This paper brings out the architecture of the intent recognition system, alternate configurations, results obtained and conclusions. The key differentiator of this system is the ability to use this system for different domains with minimal supervision.