Speaker-Aware Discourse Parsing on Multi-Party Dialogues

Proceedings of the 29th International Conference on Computational Linguistics, pages 5372–5382 October 12–17, 2022. 5372 Speaker-Aware Discourse Parsing on Multi-Party Dialogues Nan Yu1, Guohong Fu1,2∗, Min Zhang3 1School of Computer Science and Technology, Soochow University, China 2Institute of Artificial Intelligence, Soochow University, China 3Institute of Computing and Intelligence, Harbin Institute of Technology (Shenzhen), China nyu@stu.suda.edu.cn,{ghfu,minzhang}@suda.edu.cn Abstract Discourse parsing on multi-party dialogues is an important but difficult task in dialogue sys- tems and conversational analysis. It is believed that speaker interactions are helpful for this task. However, most previous research ignores speaker interactions between different speakers. To this end, we present a speaker-aware model for this task. Concretely, we propose a speaker- context interaction joint encoding (SCIJE) ap- proach, using the interaction features between different speakers. In addition, we propose a second-stage pre-training task, same speaker prediction (SSP), enhancing the conversational context representations by predicting whether two utterances are from the same speaker. Ex- periments on two standard benchmark datasets show that the proposed model achieves the best- reported performance in the literature. We will release the codes of this paper to facilitate fu- ture research1. 1 Introduction Discourse parsing on multi-party dialogues aims to identify the discourse relations between utter- ances in dialogues, which has received increasing attention in the natural language processing (NLP) community (Shi and Huang, 2019; He et al., 2021; Liu and Chen, 2021; Yang et al., 2021). Unlike traditional text-level discourse parsing based on the Rhetorical Structure Theory (RST) (Mann and Thompson, 1988) and the Penn Discourse Tree- Bank (PDTB) (Prasad et al., 2008), this task is per- formed based on the Segmented Discourse Relation Theory (SDRT) (Asher et al., 2003). It represents a multi-party dialogue by a discourse dependency tree (Afantenos et al., 2015). Figure 1 shows an example. The leaf nodes are utterances, and the arcs indicate the discourse relations between utter- ances. Each utterance is referred as an elementary discourse unit (EDU) in SDRT discourse parsing. ∗Corresponding author. 1https://github.com/yunan4nlp/SA-DPMD u1[A: Anyone have wood?] u2[A: I can spare sheep, ore or wheat?] u3[B: I’ve got wood,] u4[B: trade for ore?] u5[C: Nope sry.] u6[A: Deal.] Elab QAP Q-Elab QAP QAP Figure 1: An example of a discourse dependency tree. u1, u2, u3, u4, u5 refer to EDUs. “Q-Elab”, “QAP”, “Q- Elab”, and “Elab” refer to discourse relations. “A”, “B”, and “C” are three speakers. A multi-party dialogue has several aspects that make its discourse parsing more challenging than that of a written text created by one author. It in- volves multiple speakers who interact with each other in different roles during turn shifting and make contributions to the interactions with multiple potential threads (Afantenos et al., 2015). There- fore, in addition to conversational contexts, speaker interactions are also important cues in determining the discourse structure of a multi-party dialogue. Most current research for discourse parsing on multi-party dialogues focuses on conversational context modeling with different methods. A pio- neer study by Afantenos et al. (2015) adopts a statis- tical model for this task, using …

Content truncated to 500 words