Parallel Relationship Graph to Improve Multi-Document Summarization

Menghua Lu,Lijia Liang,Gongshen Liu
DOI: https://doi.org/10.1007/978-3-031-15931-2_52
2022-01-01
Abstract:Multi-document summarization (MDS) is an important branch of information aggregation. Compared with the single-document summary (SDS), MDS faces the problem of large search space, redundancy and complex cross-document relation. In this paper, we propose an abstractive MDS model based on Transformer, which considers the parallel information of documents with the graph attention network. During decoding, our model can utilize graph information to guide the summary generation. In addition, combined with the pre-trained language model, our model can further improve the summarization performance. Empirical results on the Multi-News and WikiSum datasets show that our model brings substantial improvements over several strong baselines, and ablation studies verify the effectiveness of our key mechanisms.
What problem does this paper attempt to address?