Dialogue Summarization with Static-Dynamic Structure Fusion Graph

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 13858–13873 July 9-14, 2023 ©2023 Association for Computational Linguistics Dialogue Summarization with Static-Dynamic Structure Fusion Graph Shen Gao1∗, Xin Cheng2∗, Mingzhe Li3, Xiuying Chen4, Jinpeng Li2, Dongyan Zhao2,5,6†, Rui Yan 7,8† 1 School of Computer Science and Technology, Shandong University 2 Wangxuan Institute of Computer Technology, Peking University 3Ant Group 4 Computational Bioscience Research Center, KAUST 5 National Key Laboratory of General Artificial Intelligence 6BIGAI, Beijing, China 7 Gaoling School of Artificial Intelligence, Renmin University of China 8 Engineering Research Center of Next-Generation Intelligent Search and Recommendation, Ministry of Education {shengao,zhaody}@pku.edu.cn, chengxin1998@stu.pku.edu.cn, ruiyan@ruc.edu.cn Abstract Dialogue summarization, one of the most challenging and intriguing text summarization tasks, has attracted increasing attention in re- cent years. Since dialogue possesses dynamic interaction nature and presumably inconsistent information flow scattered across multiple ut- terances by different interlocutors, many re- searchers address this task by modeling dia- logue with pre-computed static graph struc- ture using external linguistic toolkits. How- ever, such methods heavily depend on the re- liability of external tools and the static graph construction is disjoint with the graph represen- tation learning phase, which could not make the graph dynamically adapt to the downstream summarization task. In this paper, we propose a Static-Dynamic graph-based Dialogue Sum- marization model (SDDS)*, which fuses prior knowledge from human expertise and implicit knowledge from a PLM, and adaptively adjusts the graph weight, and learns the graph struc- ture in an end-to-end learning fashion from the supervision of summarization task. To verify the effectiveness of SDDS, we conduct exten- sive experiments on three benchmark datasets (SAMSum, MediaSum, and DialogSum) and observe significant improvement over strong baselines. 1 Introduction Dialogue summarization, aiming at distilling the salient information from a dialogue context into a concise summary, is one of the most chal- lenging and intriguing tasks in text summariza- tion (Gurevych and Strube, 2004; Feng et al., 2021a; Cheng et al., 2023a). It can help people quickly capture the highlights of a semi-structured and multi-participant dialogue without reviewing the complex dialogue context (Feng et al., 2022) *The first two authors contributed equally. †Corresponding Author. *Code available at https://github.com/Hannibal046/ SDDS Input: Dialogue Context Output: Generated Summary A: I got my concert ticket B: Whooaa that's great! C: I will see you there A: Yes for sure. C: Who you're going with? i will ask D A got a ticket for concert. C and B are going, too. A got a ticket for concert. C is going, too. I got my concert ticket and will see you there Our Framework Static Graph Existing Method Plain Document Summarization Fusion Dynamic Graph Figure 1: Traditional summarization methods and exist- ing dialogue summarization methods can not understand the dialogue structure comprehensively while our frame- work could adaptively learn dialogue structure with a dynamic graph module. and has many real-world applications (Liu et al., 2019; Zhang et al., 2021). Since dialogue is the most fundamental and spe- cially privileged …

Content truncated to 500 words