An Efficient Coarse-to-Fine Facet-Aware Unsupervised Summarization Framework Based on Semantic Blocks

Proceedings of the 29th International Conference on Computational Linguistics, pages 6415–6425 October 12–17, 2022. 6415 An Efficient Coarse-to-Fine Facet-Aware Unsupervised Summarization Framework based on Semantic Blocks Xinnian Liang1∗, Jing Li2, Shuangzhi Wu3, Jiali Zeng3, Yufan Jiang3, Mu Li3, Zhoujun Li1† 1State Key Lab of Software Development Environment, Beihang University, Beijing, China 2 School of Information Renmin University of China, Beijing, China 3Tencent Cloud Xiaowei, Beijing, China {xnliang,lizj}@buaa.edu.cn; heylijing@126.com {frostwu,lemonzeng,garyyfjiang}@tencent.com,limugx@qq.com; Abstract Unsupervised summarization methods have achieved remarkable results by incorporating representations from pre-trained language mod- els. However, existing methods fail to consider efficiency and effectiveness at the same time when the input document is extremely long. To tackle this problem, in this paper, we pro- posed an efficient Coarse-to-Fine Facet-Aware Ranking (C2F-FAR) framework for unsuper- vised long document summarization, which is based on the semantic block. The semantic block refers to continuous sentences in the doc- ument that describe the same facet. Specifi- cally, we address this problem by converting the one-step ranking method into the hierar- chical multi-granularity two-stage ranking. In the coarse-level stage, we propose a new seg- ment algorithm to split the document into facet- aware semantic blocks and then filter insignif- icant blocks. In the fine-level stage, we select salient sentences in each block and then extract the final summary from selected sentences. We evaluate our framework on four long document summarization datasets: Gov-Report, BillSum, arXiv, and PubMed. Our C2F-FAR can achieve new state-of-the-art unsupervised summariza- tion results on Gov-Report and BillSum. In ad- dition, our method speeds up 4-28 times more than previous methods.1 1 Introduction The text summarization task aims to condense a document or a set of documents into several sen- tences and keep the primary information. Recent years, both supervised (Liu and Lapata, 2019; Liu and Liu, 2021; Liu et al., 2021b) and unsuper- vised (Zheng and Lapata, 2019; Dong et al., 2021b; Liang et al., 2021, 2022) methods have made sig- nificant improvements over short documents with the development of semantic representations from *Contribution during internship at Tencent Inc. †Corresponding Author 1https://github.com/xnliang98/c2f-far Pre-trained Language Models (PLMs). Due to the noise and complexity of the increased input and output length, long-form document summarization is still a challenge (Tay et al., 2021; Akiyama et al., 2021; Grail et al., 2021). Compared with super- vised one, unsupervised methods do not rely on large amounts of labeled data and have no limi- tation on input length. In addition, unsupervised methods can be easily adapted to data from differ- ent domains, types, and languages. In this paper, we focus on unsupervised extractive methods for long document summarization. Most unsupervised extractive methods are graph- based (Zheng and Lapata, 2019; Dong et al., 2021b; Liang et al., 2021, 2022). They represent docu- ment sentences as nodes in a graph, where the edge value is the similarity between sentences. Then, they measure the importance of each node via com- puting the degree centrality (Radev et al., 2000) or running PageRank (Brin and Page, 1998) algorithm. Liang et al. (2021) pointed …

Content truncated to 500 words