Integrating Intra-Speaker Topic Modeling and Temporal-Based Inter-Speaker Topic Modeling in Random Walk for Improved Multi-Party Meeting Summarization

Yun-Nung Chen, Florian Metze

Abstract

Summarizing multi-party meetings is a complex task due to the unstructured nature of spontaneous speech and the prevalence of errors in Automatic Speech Recognition (ASR). This paper introduces a graph-based summarization framework that leverages conversational dynamics by integrating intra-speaker topic consistency and inter-speaker temporal flow into a random walk algorithm. Experimental results on the SmartNotes corpus demonstrate that this approach significantly outperforms traditional lexical centrality baselines, achieving a 6-7% relative improvement in summary quality.

Introduction

Meeting summarization differs significantly from document summarization. In a multi-party setting, the context of who is speaking and when they are speaking is as important as the content itself. Traditional extractive methods often rely on lexical centrality, which treats utterances as isolated text units. This paper proposes a more nuanced approach that models the topical relationships within a speaker's contributions and the temporal flow between different speakers.

Methodology

The researchers represent a meeting as a graph where each utterance is a node. Importance scores are calculated and propagated through the graph using a PageRank-style random walk.

Topical Similarity via PLSA

To move beyond simple word matching, the model uses Probabilistic Latent Semantic Analysis (PLSA) to determine similarity based on underlying latent topics.

Modeling Conversational Dynamics

The core contribution of this work is the modification of graph edges to reflect conversational structure:

  • Intra-Speaker Consistency: Based on the assumption that a speaker maintains topical focus across their utterances, the model boosts the similarity scores between nodes belonging to the same participant.
  • Inter-Speaker Temporal Flow: To capture the logical progression of a discussion, the model assumes that utterances occurring close in time are likely related. This is implemented using a Gaussian distribution to weight the similarity of neighboring utterances.
Gaussian distribution used to model temporal proximity between utterances.

Experimental Setup

The system was evaluated using the SmartNotes corpus, which includes 10 multi-party meetings with both manual and ASR transcripts (the latter having a 44% word error rate). The performance was measured using ROUGE-1 and ROUGE-L F-measures, comparing the generated summaries against human-annotated "noteworthy utterances."

Results and Discussion

The experimental results confirm the effectiveness of integrating speaker and temporal dynamics.

Comparison of ROUGE scores across different model configurations for ASR and Manual transcripts.

Key Insights

  • Performance Improvement: The combined model (Random Walk + Intra + Inter) yielded a 6.0% to 7.2% relative improvement over the baseline across all transcript types.
  • ASR Resilience: Interestingly, the algorithm performed better on ASR transcripts than on manual ones. The authors suggest that ASR errors often occur on filler words, which inadvertently helps the algorithm focus on the "important" keywords that remain.
  • Synergy of Dynamics: Intra-speaker and inter-speaker modeling were found to be complementary, capturing different aspects of conversational structure to create a more coherent summary.

Conclusion

By incorporating speaker identity and temporal proximity into a graph-based random walk, this research provides a robust method for unsupervised meeting summarization. Future work will focus on developing a two-level graph model to capture the hierarchical structure of meetings, potentially further improving the identification of key discussion points.

Summary updated

BibTeX
@inproceedings{chen2012integrating,
  author = {Yun-Nung Chen and Florian Metze},
  title = {{Integrating Intra-Speaker Topic Modeling and Temporal-Based Inter-Speaker Topic Modeling in Random Walk for Improved Multi-Party Meeting Summarization}},
  booktitle = {Interspeech},
  year = {2012},
  doi = {10.21437/interspeech.2012-615},
}
Figure