Discourse-Aware Unsupervised Summarization of Long Scientific Documents

1 May 2020  ·  Yue Dong, Andrei Mircea, Jackie C. K. Cheung ·

We propose an unsupervised graph-based ranking model for extractive summarization of long scientific documents. Our method assumes a two-level hierarchical graph representation of the source document, and exploits asymmetrical positional cues to determine sentence importance. Results on the PubMed and arXiv datasets show that our approach outperforms strong unsupervised baselines by wide margins in automatic metrics and human evaluation. In addition, it achieves performance comparable to many state-of-the-art supervised approaches which are trained on hundreds of thousands of examples. These results suggest that patterns in the discourse structure are a strong signal for determining importance in scientific articles.

PDF Abstract
Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Unsupervised Extractive Summarization arXiv Summarization Dataset PacSum ROUGE-1 38.57 # 4
ROUGE-2 10.93 # 4
ROUGE-L 34.33 # 3
Unsupervised Extractive Summarization arXiv Summarization Dataset HipoRank ROUGE-1 39.34 # 2
ROUGE-2 12.56 # 2
ROUGE-L 34.89 # 2
Unsupervised Extractive Summarization Pubmed HipoRank ROUGE-1 43.58 # 1
ROUGE-2 17.00 # 1
ROUGE-L 39.31 # 1
Unsupervised Extractive Summarization Pubmed PacSum ROUGE-1 39.79 # 3
ROUGE-2 14.00 # 4
ROUGE-L 36.09 # 3