A hybrid text classification approach for chatbots

Abstract

Chatbots are preferred in many fields due to their ability to provide fast and uninterrupted customer service at all hours. Most chatbots work by classifying input text and responding accordingly. In this work, a hybrid chatbot approach is presented by combining a commercial system and a deep learning-based text classification model. Additionally, active learning-based label correction and data expansion approaches are used to increase chatbot performance and keep it up-to-date. In offline tests, the hybrid method made three times fewer errors than the methods it was composed of. Online evaluations performed after deployment show that the hybrid method was able to preserve its superiority and that keeping the dataset up-to-date had positive contributions.