View Dialogue in 2D: A Two-stream Model in Time-speaker Perspective for Dialogue Summarization and beyond

Proceedings of the 29th International Conference on Computational Linguistics, pages 6075–6088 October 12–17, 2022. 6075 View Dialogue in 2D: A Two-stream Model in Time-speaker Perspective for Dialogue Summarization and Beyond Keli Xie, Dongchen He, Jiaxin Zhuang, Siyuan Lu, Zhongfeng Wang∗ School of Electronic Science and Engineering, Nanjing University, Nanjing, China {klxie, dc.he, kkchong, sylu}@smail.nju.edu.cn, zfwang@nju.edu.cn Abstract Existing works on dialogue summarization of- ten follow the common practice in document summarization and view the dialogue, which comprises utterances of different speakers, as a single utterance stream ordered by time. How- ever, this single-stream approach without spe- cific attention to the speaker-centered points has limitations in fully understanding the di- alogue. To better capture the dialogue infor- mation, we propose a 2D view of dialogue based on a time-speaker perspective, where the time and speaker streams of dialogue can be ob- tained as strengthened input. Based on this 2D view, we present an effective two-stream model called ATM to combine the two streams. Ex- tensive experiments on various summarization datasets demonstrate that ATM significantly surpasses other models regarding diverse met- rics and beats the state-of-the-art models on the QMSum dataset in ROUGE scores. Be- sides, ATM achieves great improvements in summary faithfulness and human evaluation. Moreover, results on machine reading compre- hension datasets show the generalization ability of the proposed methods and shed light on other dialogue-based tasks. Our code will be publicly available online.1 1 Introduction Dialogue summarization is a task aiming to gener- ate a succinct and coherent summary of the given dialogue, which has been explored in many appli- cations, such as automatic meeting summarization, and drawn the attention of many researchers. With the development of Transformer-based pre- trained models (Vaswani et al., 2017; Lewis et al., 2020; Zhang et al., 2020a), remarkable progress has been made in text summarization, especially in document summarization (El-Kassas et al., 2021). However, dialogue summarization is still quite chal- lenging partly due to the structural characteristics ∗ Corresponding author 1https://github.com/shakeley/View2dSum of dialogue text (Feng et al., 2021a). A major dif- ference is that a document is often organized with a unified narrative perspective, while a dialogue includes many speakers that bring diversity and switches of narrative perspectives (Kryscinski et al., 2021). Moreover, the information from different speakers is scattered (Liu et al., 2021), which poses a challenge to the summarizer as the dialogue sum- mary is often speaker-centered that focuses on the speaker’s actions and opinions (Xu and Lapata, 2021; Zhong et al., 2021b). Current works on dialogue summarization usu- ally view the dialogue as a single utterance stream ordered by time (Zhong et al., 2021a), like prior studies on document summarization do (Lin and Ng, 2019). This single-stream approach, however, has limitations in fully using the information about dialogue (Lei et al., 2021b). Firstly, the utterances of different speakers are interlaced in the single stream, which may weaken the semantic continuity from each speaker’s perspective. Besides, to obtain a concise summary instead of a laundry list, we need to both notice the …

Content truncated to 500 words