Comprehensive Document Summarization with Refined Self-Matching Mechanism

Abstract

Under the constraint of memory capacity of the neural network and the document length, it is difficult to generate summaries with adequate salient information. In this work, the self-matching mechanism is incorporated into the extractive summarization system at the encoder side, which allows the encoder to optimize the encoding information at the global level and effectively improves the memory capacity of conventional LSTM. Inspired by human coarse-to-fine understanding mode, localness is modeled by Gaussian bias to improve contextualization for each sentence, and merged into the self-matching energy. The refined self-matching mechanism not only establishes global document attention but perceives association with neighboring signals. At the decoder side, the pointer network is utilized to perform a two-hop attention on context and extraction state. Evaluations on the CNN/Daily Mail dataset verify that the proposed model outperforms the strong baseline models and statistical significantly.