Design and implementation of a chatbot in the context of customer support

Abstract

Customer support is perhaps one of the main aspects of the user experience for online services. However with the rise of natural language processing techniques, the industry is looking at automated chatbot solutions to provide quality services to an ever growing user base. This thesis presents a practical case study of such chatbot solution for the company GAMING1. First, an introduction to the market the company operates in is presented as well as a quick review of the field of conversational agents, highlighting the previous and current techniques used to develop chatbots. Then, the theory behind the techniques used is presented. Mainly deep learning techniques such as gated recurrent unit neural networks are discussed. Afterwards, a checklist of the issues solved by the chatbot is put on paper. Then a scalable software architecture for the chatbot is proposed and explained. A way of extracting ticket data as well as a quick dataset analysis are shown. A complete analysis of various neural network structures for user intent classifi- cation is shown alongside models for requesting a human operator if need be. The gated recurrent units were shown to be the most effective for classification whereas simpler models worked quite well for the human operator requester. Finally, a summary of performance metrics for the chatbot’s various submodules is shown. However since performance metrics are hard to interpret for dialogue systems, a series of practical test cases are presented as they show that the bot’s behaviour is more than satisfactory despite certain performance metrics remaining unsatisfactory.