Leveraging Discourse Structure for Extractive Meeting Summarization

TL;DR

This work tackles extractive meeting summarization by exploiting discourse structure through SDRT-based graphs, where each EDU is a node connected by labeled edges representing discourse relations. It introduces a pipeline with an EDU embedding module (MiniLM), an automatic graph generator, and a GNN-based classifier (Relational GCN and MixHop) to select salient EDUs, with a derived formula for relational message passing. Empirical results on AMI and ICSI show strong improvements over text-based and graph-based baselines across classification and summarization metrics, supported by extensive ablations on relation types, graph structure, and parsing methods, as well as a GPT-4-based ranking study. The findings highlight that graph structure often drives performance more than relation labels and reveal challenges in discourse parsing with limited annotated data, pointing to future work in robust discourse parsers and effective ranking for extractive summaries.

Abstract

We introduce an extractive summarization system for meetings that leverages discourse structure to better identify salient information from complex multi-party discussions. Using discourse graphs to represent semantic relations between the contents of utterances in a meeting, we train a GNN-based node classification model to select the most important utterances, which are then combined to create an extractive summary. Experimental results on AMI and ICSI demonstrate that our approach surpasses existing text-based and graph-based extractive summarization systems, as measured by both classification and summarization metrics. Additionally, we conduct ablation studies on discourse structure and relation type to provide insights for future NLP applications leveraging discourse analysis theory.

Figures (6)

• Figure 1: Overview of our model. A source meeting transcript is sent to both a sentence embedding model (UniLM), which generates EDU embeddings, and to a separate discourse graph generator. The EDU embeddings are then associated with nodes of the generate graph and the resulting graph structure is sent to a GNN-based node classifier which predicts whether an EDU should be included in a summary.
• Figure 2: A discourse graph for example \ref{['ex:decision']}. Node d$_1$ represents the first EDU of \ref{['d']} ("Okay"), and d$_2$, the second.
• Figure 3: Frequency of discourse relation types in AMI and ICSI predicted by Deep Sequential parser.
• Figure 4: Comparison of relations adjacent to extractive nodes and different discourse relationships in AMI and ICSI datasets.
• Figure 5: Impact of the sets of discourse relations on our RGCN classifier for both AMI and ICSI