[Paper Review] Dialogue Discourse-Aware Graph Model and Data Augmentation for Meeting Summarization

Xiachong Feng, Xiaocheng Feng+2 more|arXiv (Cornell University)|Dec 7, 2020

Topic Modeling25 references4 citations

TL;DR

This paper proposes DdaMS, a dialogue discourse-aware graph model that explicitly captures interaction relations between utterances in meetings using a relational graph encoder, and DdaDA, a data augmentation strategy that generates a 20x larger pseudo-summarization corpus by leveraging QA relations as pseudo-summaries. The approach achieves state-of-the-art performance on AMI and ICSI meeting summarization benchmarks by modeling discourse-aware interactions and pretraining on expanded data.

ABSTRACT

Meeting summarization is a challenging task due to its dynamic interaction nature among multiple speakers and lack of sufficient training data. Existing methods view the meeting as a linear sequence of utterances while ignoring the diverse relations between each utterance. Besides, the limited labeled data further hinders the ability of data-hungry neural models. In this paper, we try to mitigate the above challenges by introducing dialogue-discourse relations. First, we present a Dialogue Discourse-Dware Meeting Summarizer (DDAMS) to explicitly model the interaction between utterances in a meeting by modeling different discourse relations. The core module is a relational graph encoder, where the utterances and discourse relations are modeled in a graph interaction manner. Moreover, we devise a Dialogue Discourse-Aware Data Augmentation (DDADA) strategy to construct a pseudo-summarization corpus from existing input meetings, which is 20 times larger than the original dataset and can be used to pretrain DDAMS. Experimental results on AMI and ICSI meeting datasets show that our full system can achieve SOTA performance. Our codes will be available at: https://github.com/xcfcode/DDAMS.

Motivation & Objective

• To address the limitations of sequential modeling in meeting summarization, which fails to capture dynamic, non-linear interactions between utterances.
• To mitigate the data scarcity problem in meeting summarization, where labeled datasets are orders of magnitude smaller than in news summarization.
• To explore dialogue discourse relations as a structural inductive bias to improve representation learning and summary generation.
• To develop a scalable data augmentation strategy that leverages existing meeting structure to generate synthetic training data.
• To achieve state-of-the-art performance on standard meeting summarization benchmarks through joint modeling of discourse relations and data expansion.

Proposed method

• Construct a meeting graph with three vertex types: global, utterance, and discourse relation, to model interactions between utterances via defined dialogue discourse relations.
• Design a graph-to-sequence framework where utterance representations are updated through message passing over the relational graph, capturing long-range dependencies.
• Implement a dialogue discourse-aware data augmentation (DdaDA) strategy that identifies QA relations as pseudo-summaries and treats subsequent utterances as pseudo-meetings.
• Use the resulting 20x larger synthetic corpus to pretrain the DdaMS model, improving generalization and performance on low-resource settings.
• Integrate both word-level and utterance-level attention mechanisms to enhance representation learning and summary generation.
• Train the model using Adam with a fixed learning rate of 0.001, dropout rate of 0.5, and beam search with beam size 10 during inference.

Experimental results

Research questions

• RQ1 Can modeling dialogue discourse relations improve the performance of neural meeting summarization models?
• RQ2 How effective is a graph-based relational encoder in capturing non-linear, multi-turn interactions among utterances in meetings?
• RQ3 To what extent can dialogue discourse-aware data augmentation alleviate data scarcity in meeting summarization?
• RQ4 Does combining word-level and utterance-level attention mechanisms lead to better summary generation than using either alone?
• RQ5 How robust is the model when certain discourse relations are filtered out, and what is the impact on graph coherence and performance?

Key findings

• The DdaMS model achieves state-of-the-art performance on the AMI dataset, with ROUGE-1, ROUGE-2, and ROUGE-L scores of 51.42, 20.99, and 24.89, respectively.
• On the ICSI dataset, DdaMS achieves ROUGE-1, ROUGE-2, and ROUGE-L scores of 39.66, 10.09, and 17.53, outperforming all baselines.
• Ablation studies show that both word-level and utterance-level attention are crucial, with performance dropping by 1.15 ROUGE-1 points when word-level attention is removed.
• Filtering out even a small number of 'useless' discourse relations leads to performance degradation, indicating that graph coherence is sensitive to relation completeness.
• Case studies demonstrate that DdaMS better captures key concepts like 'fruit' and 'vegetable' by attending to multiple relevant utterances, unlike models relying only on dialogue acts.
• The DdaDA strategy successfully generates a 20x larger synthetic corpus, enabling effective pretraining and significantly boosting model generalization.