Unsupervised Mutual Learning of Discourse Parsing and Topic Segmentation in Dialogue

TL;DR

The paper tackles unsupervised joint modeling of discourse parsing and topic segmentation in dialogue by introducing a unified representation that links rhetorical and topic structures. It proposes two linguistically grounded hypotheses, Local Discourse Coupling and Global Topology Constraint, and presents an unsupervised mutual learning framework (UMLF) with three modules to enable bidirectional reinforcement between structures. A differentiable alignment loss over matrices such as $A^{top}$ and $A^{rhe}$ guides mutual learning, followed by decoding with TextTiling and Eisner for topic segmentation and discourse parsing, respectively. Evaluations on five benchmark datasets show UMLF consistently outperforms strong PLM baselines and yields competitive gains on large language models, demonstrating robust cross-scale applicability and practical potential for dialogue systems.

Abstract

In dialogue systems, discourse plays a crucial role in managing conversational focus and coordinating interactions. It consists of two key structures: rhetorical structure and topic structure. The former captures the logical flow of conversations, while the latter detects transitions between topics. Together, they improve the ability of a dialogue system to track conversation dynamics and generate contextually relevant high-quality responses. These structures are typically identified through discourse parsing and topic segmentation, respectively. However, existing supervised methods rely on costly manual annotations, while unsupervised methods often focus on a single task, overlooking the deep linguistic interplay between rhetorical and topic structures. To address these issues, we first introduce a unified representation that integrates rhetorical and topic structures, ensuring semantic consistency between them. Under the unified representation, we further propose two linguistically grounded hypotheses based on discourse theories: (1) Local Discourse Coupling, where rhetorical cues dynamically enhance topic-aware information flow, and (2) Global Topology Constraint, where topic structure patterns probabilistically constrain rhetorical relation distributions. Building on the unified representation and two hypotheses, we propose an unsupervised mutual learning framework (UMLF) that jointly models rhetorical and topic structures, allowing them to mutually reinforce each other without requiring additional annotations. We evaluate our approach on two rhetorical datasets and three topic segmentation datasets. Experimental results demonstrate that our method surpasses all strong baselines built on pre-trained language models. Furthermore, when applied to LLMs, our framework achieves notable improvements, demonstrating its effectiveness in improving discourse structure modeling.

Figures (10)

• Figure 1: An Illustration of Rhetorical and Topic Structures in Medical Consulting. (Q-Elab: Question-Elaboration, Narr: Narration, QAP: Question-Answer-Pair, Exp: Explanation, Cont: Continuation, Alter: Alternation, Comm: Comment.)
• Figure 2: The unified representation of rhetorical and topic structures for dialogues.
• Figure 3: An example of rhetorical structure and topic structure promoting each other with 11 utterances. The common structure simultaneously integrates both rhetorical and topic structures while ensuring consistency throughout.
• Figure 4: The overall architecture of our unsupervised mutual learning framework. 1) Rhetorical and Topic Representation Module: encode the topic and rhetorical structure information of dialogue utterances into two score matrices. 2) Mutual Learning Module: two aggregators separately integrate local rhetorical information and global topic information into the other structure matrix and align these two fused matrices with a common matrix under the unified representation. 3) Parsing Module: Obtain the topic structure and rhetorical structure by two parsing algorithms.
• Figure 5: Rhetoric-assisted topic analysis using Local Rhetorical Intensity on the TIAGE dataset. Rhetorical Intensity denotes the degree of clustering in the rhetorical matrix based on the numerical values of its cells. A higher rhetorical intensity facilitates clearer segmentation of the rhetorical matrix into distinct blocks.