Textual OOD Detection for Intent Classification in the Banking Industry

Pierre-Emmanuel Diot, Paul O'Mahony

Abstract: With the growing number of online banks or digitalized services of traditional banks, the need for human contact in customer support drops drastically. Deep learning and natural language processing methods allow today to answer efficiently and precisely to customers' questions via chatbots and to get closer to the way a human would answer. However, it is important that these methods do not give wrong information and do not answer questions they are not capable of answering. In this report we evaluate several out-of-distribution detection methods to overcome these problems, and focus on the NLP task of intent classification in the banking domain.