Elamur Research⌕
☰
home / papers / 2023 / Abstractive Meeting Summarization: A Survey

Abstractive Meeting Summarization: A Survey

Paper · 2023 · Virgile Rennard, Guokan Shang, Julie Hunter et al.

Summary

This TACL survey reviews automatic summarization of multi-party meetings. This area benefits most from abstractive methods, which write new sentences instead of copying them, but it is also hard to handle. The authors explain how meetings differ from news or scientific text. Transcripts are long, disfluent and low in information density. Several speakers may hold conflicting views, meeting formats vary (decision-making, brainstorming, information sharing), and judgments of what deserves a summary are subjective. Readers also prefer abstractive over extractive summaries of conversations. The paper then examines the field's resources: meeting corpora such as AMI and ICSI, evaluation metrics, and summarization systems. It groups systems into interpretation-focused approaches, which exploit dialogue acts, discourse structure or multimodal cues, and generation-focused approaches built on pre-trained language models. The main conclusion is that fine-tuning models trained on text or simpler dialogue will not deliver commercial-quality tools, so meeting-specific methods are required. A closing research agenda makes the survey a practical starting point for building conversation summarizers.

Key points

  • Meeting inputs are much longer than news articles. An AMI transcript averages 4,757 tokens with a 322-token summary, while a CNN/DailyMail article averages 781 tokens with a 56-token summary.
  • In the AMI corpus, four participants (PM, ME, UI, ID) role-play a design team building a TV remote control. This produces spontaneous but rather "well-behaved" multimodal interactions. ICSI contains 75 naturally occurring meetings.
  • Motivation: American employees reportedly spend about 6 hours per week in meetings and managers about 23 (Rogelberg et al., 2007). Videoconferencing during COVID-19 increased this load.
  • Factual consistency is a major open problem. Roughly 30% of seq2seq summaries contain fabricated facts. Existing consistency metrics work at the sentence level, which suits long transcripts poorly.
  • Proposed future directions include:
    • multi-task learning
    • weak or zero supervision
    • prompting LLMs such as GPT-3, PaLM and ChatGPT, which had not yet been applied to meetings
    • spoken-language models in the style of DialoGPT and DialogLM
    • incorporating commonsense knowledge

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓École PolytechniquePalaiseau · France14 works · 898 citations🏢LinagoraToulouse · France6 works · 133 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

🧭 Topics 1Research subjects and languages the work is about.Summarization · 232 works

Cite

Cite as Virgile Rennard, Guokan Shang, Julie Hunter and Michalis Vazirgiannis. 2023. Abstractive Meeting Summarization: A Survey. Transactions of the Association for Computational Linguistics.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.1162/tacl_a_00578

@article{rennard2023abstractive,
  title = {{Abstractive Meeting Summarization: A Survey}},
  author = {Virgile Rennard, Guokan Shang, Julie Hunter and Michalis Vazirgiannis},
  journal = {Transactions of the Association for Computational Linguistics},
  year = {2023},
  doi = {10.1162/tacl\_a\_00578},
  url = {https://doi.org/10.1162/tacl_a_00578}
}

References

6 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Moreover, Goyal et al. (2022) pointed out that this fundamental paradigm shift brings new and serious challenges for evaluation: both existing reference-based and reference-free automatic metrics cannot reliably evaluate zero-shot summaries. Because prompt-based models are not restrictively trained… find
  2. James Allen and Mark Core. 1997. Draft of DAMSL: Dialog act markup in several layers. https://www.cs.rochester.edu/research/cisd/resources/damsl Nicholas Asher. 1993. Reference to Abstract Objects in Discourse. Dordrecht, Boston, and London: Kluwer. https://doi.org/10 .1007/978-94-011-1715-9 ↗
  3. Nicholas Asher, Julie Hunter, Mathieu Morey, Benamara Farah, and Stergos Afantenos. 2016. Discourse structure and dialogue acts in multiparty dialogue: The STAC corpus. In Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC’16), pages 2721–2727, Portorož, Sl…
  4. Siddhartha Banerjee, Prasenjit Mitra, and Kazunari Sugiyama. 2015. Generating abstractive summaries from meeting transcripts. In Proceedings of the 2015 ACM Symposium on Document Engineering, pages 51–60. https://doi.org/10 .1145/2682571.2797061 find
  5. Iz Beltagy, Matthew E. Peters, and Arman Cohan. 2020. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150.
  6. James C. Bezdek, Robert Ehrlich, and William Full. 1984. FCM: The fuzzy c-means clustering algorithm. Computers & Geosciences, 10(2–3):191–203. https://doi.org/10 .1016/0098-3004(84)90020-7 \#\# Conclusion find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/rennard-2023-abstractive-meeting-summarization-survey/