A Hybrid Architecture for Out of Domain Intent Detection and Intent Discovery

Natural Language Processing is a set of computational methods that tries to process human language in different applications using linguistic analysis [Liddy, 2001]. With the advancement of deep learning approaches in recent years, research on NLP is developing rapidly. Studies related to NLP are divided into many categories: question answering, text summarization, topic modeling, sentiment analysis, etc [Eisenstein, 2019]. Among all these usages, task-oriented chatbots are a part of these categories that have taken much attention. Generally, these kinds of chatbots consist of 3 central units: Natural Language Understanding (NLU), Dialogue Management, and Natural Language Generation (NLG) [Galitsky, 2019]. The NLU unit is responsible for understanding users' intent and extracting related information that they enter so that the NLG unit can respond appropriately [Gupta et al., 2019]. In this article, we are going to propose a model to not only detect the intention of users but also check if their queries are in the domain of the chatbot's defined task and then cluster those unseen queries to map them to a pseudo label, so we can retrain our model to cover a broader domain. To clarify the problem, assume a customer who wants to book a train ticket from an assumptive origin to an assumptive destination. Then the customer may say something like "Book me a train ticket from my city to another city for 15th June at 2 pm." to the chatbot.