SA-NMS: Multi-document Summarization with Self-attention and Non-maximum Suppression Selection: Multi-document Summarization with Self-attention and Non-maximum Suppression Selection

In the extractive multiple documents summarization (MDS) field, most existing works input all related documents simultaneously to generate the summary. As a result, they usually suffer from the domination of one document (because of length or other factors) which jeopardizes the completeness of the summary. We propose a novel neural network model, Self-Attention with Non-Maximum Suppression (SA-NMS), inspired from the perspectives of human reading habits and object detection to solve the problem. Our model trains a self-attentive neural network at the single document level to get a sentence prominence score in its document, treating each document with the same weight to avoid the domination of one document. Then the NMS, an algorithm widely used in object recognition but first adopted in the field of NLP, is implemented as a strong filter to reduce redundancy. The experimental results show that our SA-NMS model not only is easy to train but also performs competitively on the small DUC-2004 dataset (50 pairs in total) and outperforms the extractive approaches on ROUGE scores on the large-scale Multi-News dataset.