Intention detection in chatbots

Abstract

As chatbots evolved over the years, researchers realised they could be used in a number of different applications, ranging from customer service to social companions. This phenomenon led to companies investing resources into the development of their own chatbot. Although successful implementations can revolutionize their domain of application, their creation requires expertise and are often very time consuming to produce. This work explores and explains different methods and mechanisms used for making chatbots while focusing its attention on the intention detection task. A review of the state of the art models is conducted with the objective of determining the best performing intent classifier and look for ways to improve them. Another objective of this work is to find an effective goal-oriented architecture for chatbots and provide a working example. A series of experiments aimed at evaluating the performance of neural network models on intent classification has been conducted over multiple different dataset of varying characteristic. An attempt at improving classification by first identifying domains has also been attempted. In addition, the overall chatbot accuracy has been compared for different models taking the role of intent classifier on the hybrid architecture. The results have highlighted the fact that the performances are closely related to the quality and quantity of the training data. Despite this fact, the BERT model seems to perform better overall but it is not systematically better, especially over dataset with a high number of possible intent and limited utterances. When it comes to the chatbot architecture, everything depends on the complexity and requirement expected from the chatbot. Nevertheless, the explored hybrid architecture appears to be a good compromise between data requirement, scalability and performance. In the end, this paper concludes by discussing the assumptions made and the limitations of the systems studied as well as proposing different ideas for future works.