High-performance intent classification in sparse supervised data conditions

Abstract

Intent recognition is the process of taking short messages, called utterances, and automatically classifying them as a specific intent from a set of possible intents. A model is trained using a number of sample utterances from each intent with the goal of being able to classify unseen utterances with high accuracy. The objective of this work is to build an intent recognition tool for chatbots that outperforms all other natural language understanding services available. In order to do this, we first created a corpus that could be used to test models in a wide variety of different conditions. Next we performed a comprehensive review of possible models by looking at relevant work in natural language processing and applying these techniques to intent recognition. Finally we present the Shallow Attention Extension that builds off of state-of-the-art language models to produce an intent recognition model with deep semantic understanding of English phrases. This model exhibits very high performance in a large set of different testing environments.