Sentence Selective Neural Extractive Summarization with Reinforcement Learning

Abstract

In this work we employed a common Recurrent Neural Network (RNN) based sequence model for single document summarization, composed of encoder-extractor hierarchical network architecture. We develop a sentence level selective encoding mechanism to select important feature before extracting sentences, and use a novel reinforcement learning based training algorithm to extend the sequence model. Besides, for single document extractive summarization task, most of researchers only pay attention to the main part of document. We analyze and explore the side information such as the headline and image caption in both CNN and Daily Mail news datasets. Empirical experiment results show the effect that our model outperforms the baseline model, and can be comparable with the state-of-the-art extractive systems when automatically evaluated in the ROUGE metric. The statistics analysis of the data set verifies our experiment results.