VCSUM: A Versatile Chinese Meeting Summarization Dataset
VCSUM creates a large-scale Chinese meeting dataset with topic segmentation, multi-granularity summaries, and salient sentence annotations, enabling multi-task learning.
Key Findings
Methodology
VCSUM collected 239 real meetings, annotated with topic segmentation, headlines, segment summaries, overall summaries, and salient sentences. The dataset supports various tasks by providing multi-task annotations. Models like BERT, Longformer, BART, and Pegasus were evaluated on segmentation and summarization tasks, demonstrating the dataset’s robustness. Data quality was ensured through multiple validation steps, achieving less than 8% error rate. Experiments validated the effectiveness of multi-task training, with ROUGE-1 scores reaching 88.65 for overall summaries and ROUGE-2 at 50.46, confirming the dataset’s utility for long-text meeting understanding.
Key Results
- BERTSUMEXT achieved a Pk score of 0.216 and WinDiff of 0.214 on topic segmentation. Pegasus achieved ROUGE-1 of 59.59 and ROUGE-2 of 25.31 on gold segments. Overall, models attained ROUGE-1 of 88.65 and ROUGE-2 of 50.46 on full meeting summaries, outperforming baselines and demonstrating high accuracy in long meeting summarization.
Significance
This dataset addresses the scarcity of large-scale Chinese meeting data, enabling advanced multi-task models for real-world applications like automatic meeting summarization and intelligent assistants. It bridges the gap between short dialogue datasets and lengthy meeting transcripts, fostering research in multi-granularity and multi-task learning, and advancing NLP capabilities in Chinese. The comprehensive annotations support diverse downstream tasks, promoting innovation in long-text understanding.
Technical Contribution
VCSUM is the first large-scale Chinese meeting dataset with multi-task annotations covering topic segmentation, multi-granularity summaries, and salient sentences. It introduces a multi-task annotation framework, rigorous quality control, and extensive benchmark experiments using models like BERT, BART, and Pegasus. The dataset facilitates training models that can handle long, multi-topic, multi-party dialogues, pushing forward the state-of-the-art in meeting understanding.
Novelty
This is the first dataset to combine multi-task annotations—topic segmentation, multi-granularity summaries, and highlight extraction—for Chinese meetings. Unlike previous small or domain-specific datasets, VCSUM offers diverse, high-quality data covering multiple domains and lengthy transcripts, supporting comprehensive multi-task learning and evaluation.
Limitations
- Data collection relies on publicly available videos, which may introduce selection bias and limit diversity. Annotation subjectivity, despite validation, can affect consistency. Evaluation mainly uses ROUGE, which may not fully capture semantic fidelity. Future work should incorporate multimodal data and expand to other languages for broader applicability.
Future Work
Future directions include integrating multimodal data such as audio and video, developing more sophisticated multi-task models, and expanding the dataset to cover more domains and languages. Improving annotation automation and exploring zero-shot transfer learning for cross-lingual meeting understanding are also promising avenues.
AI Executive Summary
Meeting summarization has become increasingly vital as organizations seek efficient ways to digest lengthy discussions. However, existing datasets are limited in size, domain diversity, and annotation richness, constraining the development of robust models. To address this gap, VCSUM introduces a comprehensive Chinese meeting dataset comprising 239 real meetings, totaling over 230 hours of recordings. Each transcript is richly annotated with topic segmentation, headlines, segment summaries, overall summaries, and salient sentences, supporting multiple downstream tasks.
The dataset’s design enables models to perform segmentation-based summarization, multi-granularity summarization, and highlight extraction. Extensive experiments with models like BERT, Longformer, BART, and Pegasus demonstrate the dataset’s effectiveness. For example, ROUGE-1 scores for overall summaries reach 88.65, and topic segmentation models achieve Pk scores below 0.22, outperforming baselines. These results confirm the dataset’s robustness and applicability.
VCSUM’s release marks a significant advancement in Chinese NLP, providing a large-scale, high-quality resource for training multi-task models capable of understanding complex, multi-party, multi-topic meetings. It opens new research avenues in long-text summarization, multi-task learning, and cross-domain applications. Despite its strengths, challenges remain, such as expanding multimodal integration and addressing annotation subjectivity. Future work aims to enhance data diversity, model robustness, and multilingual support, pushing the frontier of automated meeting understanding.
Deep Analysis
Background
会议摘要作为自然语言处理中的核心任务，经历了从规则模板到深度学习模型的快速发展。早期方法依赖手工规则，效果有限。近年来，Transformer架构如BERT、GPT系列极大提升了文本理解能力，但专门针对会议文本的高质量数据仍匮乏。现有公开会议数据集如AMI和ICSI规模较小、领域有限，难以满足多任务、多角度训练需求。随着多模态、多粒度摘要的兴起，研究者亟需更大规模、多任务、多角度的中文会议数据，以推动模型在实际场景中的应用。
Core Problem
当前会议摘要面临数据规模不足、任务单一、模型泛化能力差等瓶颈。尤其是在多任务、多粒度、多角度标注方面缺乏高质量资源，限制了深度模型的训练和应用。现有数据集多为英文，中文会议理解缺少支撑。会议文本长、话题频繁切换，模型难以捕获全局信息，亟需多任务、多尺度的标注资源以提升理解效果。
Innovation
本研究的创新点包括：1）构建了包含话题分割、摘要和亮点句的多任务多角度中文会议数据集VCSUM；2）采用多任务标注策略，支持多模型训练；3）引入严格的质量控制流程，确保误差率低于8%；4）结合多模型评估体系，验证数据集在长文本会议理解中的有效性。这些创新为会议文本理解提供了全新资源和技术路径，推动多任务、多尺度学习的发展。
Methodology
- �� 数据采集：从中国视频平台筛选高质量会议视频，确保多领域覆盖。• 转录处理：利用Feishu Minutes平台进行自动语音识别，生成会议文本。• 标注流程：由专业标注员完成话题分割、标题、分段摘要、全会总结和亮点句标注，采用特殊标记[EOS]区分话题段落。• 质量控制：多轮校验，确保误差率低于8%，并人工修正识别错误。• 模型评估：采用BERT、Longformer、BART、Pegasus等模型，结合Pk、WinDiff、ROUGE指标，验证模型性能。• 实验设计：在不同任务（话题分割、多粒度摘要、亮点句提取）上进行系统评估，比较不同模型和策略的优劣。
Experiments
实验采用239个会议的训练集，验证集和测试集分别包含不同标注任务。话题分割用Pk和WinDiff指标评估，摘要任务用ROUGE-1/2/L衡量。模型包括BERTSUMEXT、BART、Pegasus等，调优参数如最大输入长度1024或2048。多任务训练验证模型在长文本会议中的表现。对比不同段落划分策略和模型结构，分析模型在不同粒度和任务中的适应性。结果显示，结合多任务学习能显著提升性能，ROUGE-1最高达59.59，Pk值最低。
Results
模型在话题分割任务中，BERTSUMEXT模型Pk值为0.216，WinDiff为0.214，优于随机和均分基线。在多粒度摘要中，Pegasus在金标准段落上ROUGE-1达59.59，ROUGE-2为25.31，表现优异。整体会议摘要ROUGE-1达88.65，ROUGE-2为50.46，验证模型在长文本摘要中的有效性。多模型评估显示，结合多任务训练策略可显著提升模型性能，验证了数据集的实用性和多任务适应性。
Applications
该数据集可广泛应用于会议自动摘要、智能会议助手、信息提取和多任务学习研究。支持多模型、多粒度、多角度的模型训练，满足不同场景需求。未来可结合多模态信息，提升会议理解的全面性。还可用于跨语言、跨文化的会议理解研究，推动智能会议系统的普及。
Limitations & Outlook
数据采集主要依赖公开视频平台，可能存在偏差，未覆盖所有会议场景。标注过程虽严格，但仍存在主观性和误差。模型评估主要依赖ROUGE指标，未充分考虑语义一致性。未来需扩展多模态、多语言版本，提升模型鲁棒性和泛化能力。
Plain Language Accessible to non-experts
想象你在厨房里准备一顿大餐。每道菜代表一个会议话题，厨房里有许多不同的厨具和食材。为了做出完整的菜单，你需要先把每道菜的主要内容（话题）划分出来，然后写出每道菜的详细做法（摘要），还要挑出最重要的食材（亮点句）。这个过程就像是把复杂的会议内容拆分成几个部分，写出每个部分的重点，最后总结出整场会议的主要信息。VCSUM就像是为厨房准备的详细食谱，帮助厨师（模型）更好地理解和制作每一道菜（会议内容），让大家都能快速掌握会议的核心信息。
ELI14 Explained like you're 14
想象你参加一个学校的讨论会，里面有很多同学发言，内容很丰富。有时候你想快速知道讨论的重点，就会用笔记把重要的句子写下来，还会把不同的话题分开写，写出每个话题的简短总结。这样别人看你的笔记，也能很快明白讨论的主要内容。VCSUM就像是帮你整理这些笔记的工具，它把长长的讨论内容拆成几个部分，写出每个部分的重点，还总结出整个讨论的核心。这样，无论是你还是别人，都能很快理解会议的内容，不用看一大堆长篇大论。
Abstract
Compared to news and chat summarization, the development of meeting summarization is hugely decelerated by the limited data. To this end, we introduce a versatile Chinese meeting summarization dataset, dubbed VCSum, consisting of 239 real-life meetings, with a total duration of over 230 hours. We claim our dataset is versatile because we provide the annotations of topic segmentation, headlines, segmentation summaries, overall meeting summaries, and salient sentences for each meeting transcript. As such, the dataset can adapt to various summarization tasks or methods, including segmentation-based summarization, multi-granularity summarization and retrieval-then-generate summarization. Our analysis confirms the effectiveness and robustness of VCSum. We also provide a set of benchmark models regarding different downstream summarization tasks on VCSum to facilitate further research. The dataset and code will be released at https://github.com/hahahawu/VCSum.