Coarser to Finer Level Document Classification through Recurrent Attention Mechanism using RL Agent

Document Classification is a Natural Language Processing task, which generally uses deep neural networks to extract features from full textual information. The extracted features may or may not be relevant for classification of a document. We propose a framework to address the classification of long documents from coarse level to finer level by combining recurrent attention mechanism. It constructs the discriminative features with fewer words. The main idea is to train the recurrent neural network which focuses its attention on distinct parts of the document. It includes reinforcement learning agent at the word level for emitting the next block location to be glimpsed. Convolutional neural network (CNN) is used to extract glimpsed features from focused words. Both sentence and document representation can be obtained from word level and sentence level respectively. Experiments conducted on our collected 5-class arXiv papers dataset, the proposed method surpass the existing methods with less observed words.