[Paper Review] A Sliding-Window Approach to Automatic Creation of Meeting Minutes

Jia Jin Koay, Alexander Roustai+2 more|arXiv (Cornell University)|Apr 26, 2021

Topic Modeling38 references4 citations

TL;DR

This paper proposes a sliding-window approach using a fine-tuned BART summarizer to automatically generate meeting minutes from spoken transcripts. By processing transcripts in overlapping windows and extracting salient content, the method achieves strong performance on both human and automatic transcripts, outperforming baselines in extractive summarization and producing coherent, context-aware summaries without requiring annotated training data.

ABSTRACT

Meeting minutes record any subject matters discussed, decisions reached and actions taken at meetings. The importance of minuting cannot be overemphasized in a time when a significant number of meetings take place in the virtual space. In this paper, we present a sliding window approach to automatic generation of meeting minutes. It aims to tackle issues associated with the nature of spoken text, including lengthy transcripts and lack of document structure, which make it difficult to identify salient content to be included in the meeting minutes. Our approach combines a sliding window and a neural abstractive summarizer to navigate through the transcripts to find salient content. The approach is evaluated on transcripts of natural meeting conversations, where we compare results obtained for human transcripts and two versions of automatic transcripts and discuss how and to what extent the summarizer succeeds at capturing salient content.

Motivation & Objective

• To address the challenge of generating accurate meeting minutes from long, unstructured spoken transcripts common in virtual meetings.
• To overcome limitations of neural abstractive summarizers, such as positional bias and context length constraints, when applied to spoken meeting data.
• To evaluate the effectiveness of a sliding-window strategy in identifying salient content across diverse transcript types, including human and automatic transcriptions.
• To compare the proposed method with extractive and supervised abstractive baselines in terms of both automatic and human evaluation metrics.

Proposed method

• The method employs a sliding window over meeting transcripts, with configurable window size (W) and stride (S), to process local segments of the transcript.
• For each window, a fine-tuned BART-large-cnn model generates an abstractive summary of the local content, identifying salient utterances.
• The approach uses overlapping windows to ensure salient content is not missed due to window boundaries, with window and stride sizes tuned for optimal recall and precision.
• Local summaries are consolidated into a final meeting-level summary by merging and deduplicating selected utterances across windows.
• The system is evaluated on transcripts from the ICSI meeting corpus, including human transcriptions and two automatic versions from AMI ASR and Google Cloud Speech-to-Text.
• The method does not require any annotated training data, making it adaptable to diverse meeting domains and types.

Experimental results

Research questions

• RQ1 What are the optimal window and stride sizes for effective salience detection in spoken meeting transcripts?
• RQ2 Can a neural abstractive summarizer effectively identify salient content within local windows of spoken transcripts?
• RQ3 How can local summaries from overlapping windows be consolidated into a coherent, high-quality meeting-level summary?
• RQ4 How does the sliding-window approach compare to extractive and supervised abstractive baselines in terms of automatic and human evaluation?

Key findings

• The sliding-window approach with W=1024 and S=128 achieved the highest F-score (0.455) on the training set, balancing precision and recall.
• A small window and stride (W=128, S=128) yielded high recall (30% of utterances selected per meeting), but lower precision, while larger windows improved precision.
• The method outperformed TextRank and performed comparably to a supervised BERT summarizer in human evaluation, with 67% of utterances rated as highly relevant or relevant.
• The sliding-window summaries were perceived as more coherent and contextually connected than those from the supervised BERT model, with consecutive utterances more likely to share the same speaker or topic.
• The Google Cloud Speech-to-Text transcripts showed different utterance segmentation behavior compared to human and AMI ASR transcripts, affecting summary consistency.
• The approach successfully handled transcripts exceeding BART’s input length by processing them in overlapping windows, enabling end-to-end summarization without truncation.