False Positive Intent Detection Framework for Chatbot Annotation

Abstract

For chatbots answering thousands of user queries daily, it requires huge annotation efforts or explicit signals from users to identify incorrect chatbot predictions. Identification of such False Positives is key to improving chatbot accuracy and is a challenging problem due to the high cost and limited explicit signals from users. In this paper, we present a framework for automatically detecting False Positive intents in an employee chatbot through implicit feedback by capturing specific user behavior using techniques such as detection of repeated queries and leveraging on active learning sampling strategies to find cases where the chatbot might have provided an incorrect response. Using this approach within the bank, annotators can prioritize their efforts and detect False Positive intent approximately three times better than manual screening of random chatbot dialogues. This framework can be reused across different chatbot applications.