Finding decision jumps in text classification

Abstract

Text classification is one of the key problems in natural language processing (NLP), and in early years, it was usually accomplished by feature-based machine learning models. Recently, the deep neural network has become a powerful learning machine, making it possible to work with text itself as raw input for the classification problems. However, existing neural networks are typically end-to-end and lack explicit interpretation of the prediction. In this paper, we propose Jumper, a novel framework that models text classification as a sequential decision process. Generally, Jumper is a neural system that scans a piece of text sequentially and makes classification decisions at the time it wishes, which is inspired by the cognitive process of human text reading. In our framework, both the classification result and when to make the classification are part of the decision process, controlled by a policy network and trained with reinforcement learning. Experimental results of real-world applications demonstrate the following properties of a properly trained Jumper: (1) it tends to make decisions whenever the evidence is enough, therefore reducing total text reading by 30–40% and often finding the key rationale of the prediction; and (2) it achieves classification accuracy better than or comparable to state-of-the-art models in several benchmark and industrial datasets. We further conduct a simulation experiment with mock data, which confirms that Jumper is able to make a decision at the theoretically optimal decision position.