Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Policies and Evaluation for Online Meeting Summarization

T0 review · 3 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Live meeting summarization is feasible, and the update policy—not the summarizer—controls the quality-latency trade-off.

desk verdict First systematic study of online meeting summarization; useful metrics but the headline comparisons are weakened by test-set selection and an overly idealized latency metric. read the letter →

arxiv 2502.03111 v1 pith:2QACEE4T submitted 2025-02-05 cs.CL cs.AIcs.LG

Felix Schneider (1) , Marco Turchi (1) , Alex Waibel (2) ((1) Zoom Communications , (2) Karlsruhe Institute of Technology) This is my paper · ORCID
classification cs.CLcs.AIcs.LG
keywords onlinemeetingsummarizationstreamingpoliciesExpectedLatencyintermediatesummaryqualityquality-latencytrade-offadaptivesegmentationAutoMin
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper makes the first systematic case that meeting summaries can be produced live, while the meeting is still underway, rather than only after it ends. The authors propose five policies—rules for when a summarizer reads input and writes output—layered on top of ordinary offline models, and introduce two metrics built for the online setting: Expected Latency, measuring how long a random moment must wait for the next summary, and R1-AUC, the area under the Rouge-1 curve over the meeting's progress. Evaluated on the AutoMin meeting corpus, online systems produce summaries that human raters score close to offline baselines, and the adaptive policies (choose how much to read based on the content) clearly beat fixed-length chunking in both final and intermediate quality. The implication is that online summarization is a tractable, deployable task whose main design lever is the update policy, not the summarizer itself.

What carries the argument

The central object is the policy: a control loop that at each step chooses READ (consume the next chunk of transcript), WRITE (emit a summary sentence), or ERASE (amend earlier output). Five policies instantiate this: length-based, model-based, sliding window, full rewriting, and fully incremental. Two purpose-built metrics carry the evaluation: Expected Latency (EL), computed as $\mathrm{EL} = \frac{1}{T}\sum_{t\in S\cup\{0\}}\frac{(N(t)-t)^2}{2}$, the expected wait from a uniform random time until the next write, and R1-AUC, the area under the Rouge-1 F1 curve as the meeting progresses. EL rests on the assumption that each emitted summary captures all content since the previous write; R1-AUC treats the user-visible summary at every moment as the system's output.

What would settle it

Take a meeting with known content-level timestamps. Build two online systems with identical write schedules—one that always writes all content since the last write, and one that deliberately defers half of it until the next write. If Expected Latency gives the two systems the same score while human annotators (or a content-alignment measure) judge the second system as noticeably later, the metric's core assumption is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that online meeting summarization—maintaining a useful summary as the meeting happens—is feasible and that the choice of update policy dominates the quality-latency trade-off. On the AutoMin 2023 test set, the best online systems score within about one ROUGE-1 point of the offline GPT-4 baseline (42.7 vs 43.7), and in human evaluation the model-based and sliding-window policies match or exceed the offline Bart-based Zoom-long system. The authors further claim that their two new metrics—Expected Latency, which reduces latency measurement to the gap between a random time point and the next write event, and Rouge-curve AUC, which integrates intermediate summary quality over the meeting—are expressive enough to distinguish systems that final Rouge alone cannot. The strongest systems are the adaptive ones: they dynamically select how much transcript to read per summary unit instead of using fixed chunk lengths.

Load-bearing premise

The Expected Latency metric assumes each summary the system writes contains everything important that happened since the last write; if summaries omit or hold back content, two systems with identical write schedules get the same latency score even though one is much later in practice.

Editorial extensions

If this is right

  • Meeting platforms can offer live catch-up summaries to late joiners, with a freshness-versus-quality knob set by the policy.
  • Because policies wrap offline summarizers unchanged, any future improvement in the underlying model transfers directly to the online setting.
  • The proposed metrics give the field a common yardstick—Expected Latency for delay, R1-AUC for intermediate quality—so systems can be compared on one trade-off curve.
  • The adaptive policies' advantage over fixed chunking suggests that how the transcript is segmented is a primary quality lever, and the same segmentation could feed offline summarizers as well.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Expected Latency could be sharpened by content-level alignment—say, embedding similarity between each summary sentence and the source turn that first supplies it—breaking ties between systems that write on the same schedule but delay different content.
  • The READ/WRITE/ERASE policy loop is generic enough to extend beyond meetings to other live document-generation tasks, such as real-time news digests or action-item trackers, with EL and R1-AUC as portable metrics.
  • Because AutoMin lacks audio timestamps, the authors model each dialog turn as one time unit; validating EL on a corpus with real wall-clock timing (or ASR word timestamps) would test whether turn-count latency matches human-perceived delay.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper introduces the task of online meeting summarization, proposes several policies (length-based, model-based, sliding window, full rewriting, fully incremental) built on top of offline summarizers (Bart and GPT-4), and defines novel evaluation metrics: Expected Latency (EL), R1-AUC for intermediate summary quality, redundancy factor, and normalized erasure. The experiments on the AutoMin 2023 test set (12 meetings) with human evaluation lead to three claims: (1) online models can produce strong summaries, (2) the proposed metrics enable a detailed quality-latency analysis including intermediate outputs, and (3) adaptive policies outperform fixed scheduled ones.

Significance. If the claims hold, this is a useful first systematic study of online meeting summarization, establishing a task formulation, baseline policies, and evaluation metrics that the community can build on. The paper is commendable for publishing pseudocode, prompts, full per-system results, and a human evaluation that partially validates the automatic metrics. However, the central empirical claims are weakened by two load-bearing issues: the Expected Latency metric relies on an unvalidated coverage assumption that makes it a measure of write cadence rather than content-level delay, and the headline comparisons select the best chunk size per policy on the test set. The evidence for the third claim is also mixed, as only one of the two adaptive policies clearly outperforms the fixed-length baseline in the human evaluation.

major comments (3)
  1. [Section 6.2, Eq. (1)] The Expected Latency metric is derived under the explicit assumption that "the summarization output contains all relevant content from the source that appeared since the last output." Under that assumption, Eq. (1) depends only on the set of write timestamps S, not on what the summaries actually contain. Consequently, two systems with identical write schedules but vastly different content coverage receive the same EL, even though the paper defines latency as "the time between a piece of content appearing in the output and its appearance in the source." The paper then uses EL to compare systems in the quality-latency trade-off (e.g., sliding window Bart-256 EL 11.7 vs. length-based Bart-768 EL 25.9 in Table 4), so the latency axis can reflect "how often the system writes" rather than "how long until content appears." The assumption is stated but never validated or bounded, which undermines abstract claim (2). I recommend either validating the assumption on a sample via content alignment, or explicitly redefining EL as a cadence metric and providing a separate content-aware latency measure.
  2. [Table 1 and Section 7.1] The headline results in Table 1 select, for each policy and backend model, the chunk size with the best final ROUGE score on the test set. This is a form of test-set selection bias: each system is allowed to pick its most favorable hyperparameter after seeing the evaluation data, which inflates the reported quality and makes cross-policy comparisons (e.g., sliding window Bart-256 vs. length-based Bart-768 vs. model-based Bart-512) unfair. The full results in Table 4 partially mitigate this, but the main narrative and the human evaluation use only the best-per-policy rows. The paper should either report results for a fixed chunk size across all policies, or use a validation split for model selection and report the selected configurations.
  3. [Section 7.2, Table 3] Abstract claim (3), that "adaptive policies perform better than fixed scheduled ones," is not fully supported by the reported evidence. In the human evaluation of intermediate summaries (Table 3), the model-based policy receives an average score of 3.5, essentially tied with the length-based policy at 3.4, while only the sliding window policy (3.7) is clearly better. In the final-summary human evaluation (Table 2), the model-based system (4.4) is rated higher than the length-based system (3.6), but this rests on two annotators and no inter-annotator agreement or significance testing. Given the 12-meeting test set, the claim should be qualified as "some adaptive policies can outperform fixed ones based on automatic metrics and a limited human evaluation."
minor comments (5)
  1. [Section 4] The assumption of unit duration per dialog turn for latency calculation is stated in one sentence, but it has a direct effect on the absolute EL values and their interpretability; a brief discussion of how varying turn lengths (e.g., a long monologue vs. a one-word reply) would affect EL would be helpful.
  2. [Table 4] The LAAL column is reported for all systems but never used in the analysis or discussion; consider removing it or adding a note explaining why it is included.
  3. [Section 7.2] The description "The convex shape of the sliding window model's curve" is mathematically imprecise; a curve that rises quickly and then plateaus is typically concave or saturating, and the relationship between convexity and high AUC is not obvious. Please rephrase.
  4. [Author affiliation] There is a typo in the affiliation: "Karlsruhe Intstitute of Technology" should be "Karlsruhe Institute of Technology."
  5. [Section 7.1] There is a typo: "GPT-4 is still rated highest, but thesliding window Bart" should be "the sliding window Bart."

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the Expected Latency metric rests on an acknowledged simplifying assumption rather than a circular reduction, and the central claims are grounded in external data and human evaluation.

full rationale

The paper's load-bearing claims are (1) online models can produce strong summaries, (2) the proposed metrics enable a quality-latency analysis that includes intermediate outputs, and (3) adaptive policies outperform fixed scheduled ones. None of these claims is equivalent to its inputs by construction. The Expected Latency metric in Section 6.2 is explicitly introduced as a proxy: 'That the summarization output contains all relevant content from the source that appeared since the last output. Given this simplification, calculating latency reduces to answering the question: “Sampling a random point in the meeting, how long do I have to wait for the next summary?”' This is an acknowledged modeling assumption that makes EL depend only on write timestamps, not a hidden derivation of content-level latency from the metric itself. The paper does not claim EL measures content-level delay under that assumption; it explicitly states the simplification. An untested or overly strong assumption is a correctness or validity risk, not circularity under the review rules. The R1-AUC metric for intermediate summaries is defined from ROUGE curves and then checked against human ratings, not fitted to the human scores; the human evaluation in Sections 7.1 and 7.2 provides an external sanity check on the ranking of policies. The policies themselves are parameter-free and applied post-training to an offline summarizer, so no fitted parameter is renamed as a prediction. The only self-citations are peripheral: the Zoom-long offline baseline and the deidentification-reversal preprocessing trick, both from Schneider and Turchi (2023). These are not load-bearing for the central adaptive-versus-fixed comparison, which is evaluated against the paper's own length-based systems and confirmed by human annotators. No uniqueness theorem is imported from the authors, and no known result is merely renamed. The derivation chain is therefore self-contained, with the caveat that EL's validity as a content-level latency measure depends on the stated coverage assumption, a concern about correctness rather than circularity.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

The results rest on several domain assumptions: AutoMin representativeness, unit-turn timing, EL content-coverage simplification, ROUGE-1 as the primary metric, and scoring partial summaries against the full reference. Chunk sizes are swept and the best per policy is selected post hoc, acting as a free parameter in the reported comparisons. No new entities are introduced.

free parameters (1)
  • Chunk size per policy (best selected for headline results) = Length-based Bart 768; Model-based Bart 512; Sliding Window Bart 256; Full Rewriting GPT-4 512; Full Incremental GPT-4…
    The paper tests chunk sizes 256, 512, 768, 1024, and 2048, then reports the system with the best final ROUGE score per policy in Table 1, which is a post hoc selection on the test set.
assumptions (5)
  • domain assumption The AutoMin 2023 test set (12 meetings) is representative of general meetings and has high-quality references.
    Section 4 states AutoMin's references are of high quality and it relies on Ghosal et al. 2023 for metric correlation.
  • domain assumption Expected Latency can be computed under the assumption that a summary contains all relevant content since the last write.
    Section 6.2 explicitly states this simplification, which reduces latency to WRITE event timings.
  • domain assumption Each dialog turn is assigned unit duration for latency calculation.
    Section 4 notes "we assume unit duration for each dialog turn for calculating latency" because AutoMin lacks timing info.
  • domain assumption ROUGE-1 F1 is the preferred automatic quality metric for this dataset.
    Section 6.1 cites Ghosal et al. 2023 for best human correlation and uses ROUGE-1 F1 throughout.
  • domain assumption Partial summaries at any time can be scored by ROUGE-1 F1 against the full final human reference.
    Section 7.2 computes Rouge for every output update against the full reference, treating the final summary as the target for partial ones.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Policies and Evaluation for Online Meeting Summarization." pith.science (2026). https://pith.science/paper/2QACEE4T

@misc{pith2026250203111,
  author       = {Pith},
  title        = {Pith review of: Policies and Evaluation for Online Meeting Summarization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2QACEE4T}},
  note         = {Machine review of arXiv:2502.03111}
}
read the original abstract

With more and more meetings moving to a digital domain, meeting summarization has recently gained interest in both academic and commercial research. However, prior academic research focuses on meeting summarization as an offline task, performed after the meeting concludes. In this paper, we perform the first systematic study of online meeting summarization. For this purpose, we propose several policies for conducting online summarization. We discuss the unique challenges of this task compared to the offline setting and define novel metrics to evaluate latency and partial summary quality. The experiments on the AutoMin dataset show that 1) online models can produce strong summaries, 2) our metrics allow a detailed analysis of different systems' quality-latency trade-off, also taking into account intermediate outputs and 3) adaptive policies perform better than fixed scheduled ones. These findings provide a starting point for the wider research community to explore this important task.

Figures

Figures reproduced from arXiv: 2502.03111 by the authors.

Figure 1
Figure 1. Rouge curves for the best-performing system for each policy. All curves are in Appendix [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Rouge F1 curves for all systems [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. AutoMeet: a proof-of-concept study of genAI to automate meetings in automotive engineering

    cs.CL 2025-07 conditional novelty 3.0 of 10

    Engineers in an automotive R&D department estimated that a fully integrated genAI meeting-documentation pipeline could cancel about one in five meetings, saving roughly 10.5% of working time, and named privacy control...

Reference graph

Works this paper leans on

41 extracted references · 28 canonical work pages · cited by 1 Pith paper

  1. [1]

    Evlampios Apostolidis, Eleni Adamantidou, Alexandros I Metsai, Vasileios Mezaris, and Ioannis Patras. 2021. Video summarization using deep neural networks: A survey. Proceedings of the IEEE, 109(11):1838--1863

    work page 2021

  2. [2]

    Naveen Arivazhagan, Colin Cherry, Isabelle Te, Wolfgang Macherey, Pallavi Baljekar, and George Foster. 2020. Re-translation strategies for long form, simultaneous, spoken language translation. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7919--7923. IEEE

    work page 2020

  3. [3]

    Abedelkadir Asi, Song Wang, Roy Eisenstadt, Dean Geckt, Yarin Kuper, Yi Mao, and Royi Ronen. 2022. An end-to-end dialogue summarization system for sales calls. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies: Industry Track, pages 45--53

    work page 2022

  4. [4]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gr...

    2020

  5. [5]

    Alexander R Fabbri, Wojciech Kry \'s ci \'n ski, Bryan McCann, Caiming Xiong, Richard Socher, and Dragomir Radev. 2021 a . Summeval: Re-evaluating summarization evaluation. Transactions of the Association for Computational Linguistics, 9:391--409

    work page 2021

  6. [6]

    Alexander Richard Fabbri, Faiaz Rahman, Imad Rizvi, Borui Wang, Haoran Li, Yashar Mehdad, and Dragomir Radev. 2021 b . Convosumm: Conversation summarization benchmark and improved abstractive summarization with argument mining. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Confe...

    work page 2021

  7. [7]

    Nikhil Garg, Benoit Favre, Korbinian Reidhammer, and Dilek Hakkani-T \"u r. 2009. Clusterrank: A graph based method for meeting summarization. In Interspeech, Brighton (UK)

    work page 2009

  8. [8]

    Tao Ge, Lei Cui, Baobao Chang, Sujian Li, Ming Zhou, and Zhifang Sui. 2016. News stream summarization using burst information networks. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 784--794

    work page 2016

Show all 41 references
  1. [9]

    Tirthankar Ghosal, Ond r ej Bojar, Marie Hled \' kov \'a , Tom Kocmi, and Anna Nedoluzhko. 2023. https://aclanthology.org/2023.inlg-genchal.19 Overview of the second shared task on automatic minuting ( A uto M in) at INLG 2023 . In Proceedings of the 16th International Natural...

    2023

  2. [10]

    Bogdan Gliwa, Iwona Mochol, Maciej Biesek, and Aleksander Wawer. 2019. Samsum corpus: A human-annotated dialogue dataset for abstractive summarization. In Proceedings of the 2nd Workshop on New Frontiers in Summarization, pages 70--79

    2019

  3. [11]

    J Gu, G Neubig, K Cho, and VOK Li. 2017. Learning to translate in real-time with neural machine translation. In European Chapter of the Association for Computational Linguistics (EACL) Conference, 2017. The Association for Computational Linguistics

    2017

  4. [12]

    Adam Janin, Don Baron, Jane Edwards, Dan Ellis, David Gelbart, Nelson Morgan, Barbara Peskin, Thilo Pfau, Elizabeth Shriberg, Andreas Stolcke, et al. 2003. The icsi meeting corpus. In 2003 IEEE International Conference on Acoustics, Speech, and Signal Processing, 2003. Proceed...

    2003

  5. [13]

    Wessel Kraaij, Thomas Hain, Mike Lincoln, and Wilfried Post. 2005. The ami meeting corpus. In Proc. International Conference on Methods and Techniques in Behavioral Research

    2005

  6. [14]

    Shamit Lal, Shivam Duggal, and Indu Sreedevi. 2019. Online video summarization: Predicting future to better summarize present. In 2019 IEEE Winter Conference on applications of computer vision (WACV), pages 471--480. IEEE

    2019

  7. [15]

    Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58t...

    2020

  8. [16]

    Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74--81

    2004

  9. [17]

    Zhengyuan Liu and Nancy F Chen. 2022. Dynamic sliding window modeling for abstractive meeting summarization. In INTERSPEECH, pages 5150--5154

    2022

  10. [18]

    Gabriel Murray, Giuseppe Carenini, and Raymond Ng. 2010. Generating and validating abstracts of meeting conversations: a user study. In Proceedings of the 6th international natural language generation conference

    2010

  11. [19]

    Anna Nedoluzhko, Muskaan Singh, Marie Hled \' kov \'a , Tirthankar Ghosal, and Ond r ej Bojar. 2022. https://aclanthology.org/2022.lrec-1.340 ELITR minuting corpus: A novel dataset for automatic minuting from multi-party meetings in E nglish and C zech . In Proceedings of the ...

    2022

  12. [20]

    Jan Niehues, Ngoc-Quan Pham, Thanh-Le Ha, Matthias Sperber, and Alex Waibel. 2018. Low-latency neural speech translation. arXiv preprint arXiv:1808.00491

    2018 arXiv

  13. [21]

    Andrei Olariu. 2014. Efficient online summarization of microblogging streams. In Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, volume 2: Short Papers, pages 236--240

    2014

  14. [22]

    OpenAI. 2023. Gpt-4 technical report. https://cdn.openai.com/papers/gpt-4.pdf

    2023

  15. [23]

    Sara Papi, Marco Gaido, Matteo Negri, and Marco Turchi. 2022 a . https://doi.org/10.18653/v1/2022.findings-emnlp.11 Does simultaneous speech translation need simultaneous models? In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 141--153, Abu Dhab...

    2022 doi

  16. [24]

    Sara Papi, Marco Gaido, Matteo Negri, and Marco Turchi. 2022 b . Over-generation cannot be rewarded: Length-adaptive average lagging for simultaneous speech translation. In Proceedings of the Third Workshop on Automatic Simultaneous Translation, pages 12--17

    2022

  17. [25]

    Minh Quang Pham, Sathish Reddy Indurthi, Shamil Chollampatt, and Marco Turchi. 2023. Select, prompt, filter: Distilling large language models for summarizing conversations. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 12257--12265

    2023

  18. [26]

    Felix Schneider and Marco Turchi. 2023. https://aclanthology.org/2023.inlg-genchal.14 Team zoom @ A uto M in 2023: Utilizing topic segmentation and LLM data augmentation for long-form meeting summarization . In Proceedings of the 16th International Natural Language Generation ...

    2023

  19. [27]

    Royal Sequiera, Luchen Tan, and Jimmy Lin. 2018. Overview of the trec 2018 real-time summarization track. In TREC

    2018

  20. [28]

    Antoine Tixier, Polykarpos Meladianos, and Michalis Vazirgiannis. 2017. Combining graph degeneracy and submodularity for unsupervised extractive summarization. In Proceedings of the workshop on new frontiers in summarization, pages 48--58

    2017

  21. [29]

    Gokhan Tur, Andreas Stolcke, Lynn Voss, John Dowding, Beno \^ t Favre, Raquel Fern \'a ndez, Matthew Frampton, Michael Frandsen, Clint Frederickson, Martin Graciarena, et al. 2008. The calo meeting speech recognition and understanding system. In 2008 IEEE Spoken Language Techn...

    2008

  22. [30]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30

    2017

  23. [31]

    Alex Waibel, Michael Bett, Michael Finke, and Rainer Stiefelhagen. 1998. Meeting browser: Tracking and summarizing meetings. In Proceedings of the DARPA broadcast news workshop, pages 281--286

    1998

  24. [32]

    Klaus Zechner. 2002. Automatic summarization of open-domain multiparty dialogues in diverse genres. Computational Linguistics, 28(4):447--485

    2002

  25. [33]

    Tianyi Zhang, Faisal Ladhak, Esin Durmus, Percy Liang, Kathleen McKeown, and Tatsunori B Hashimoto. 2023. Benchmarking large language models for news summarization. arXiv preprint arXiv:2301.13848

    2023 arXiv

  26. [34]

    Yusen Zhang, Ansong Ni, Ziming Mao, Chen Henry Wu, Chenguang Zhu, Budhaditya Deb, Ahmed Awadallah, Dragomir Radev, and Rui Zhang. 2022. Summn: A multi-stage summarization framework for long input dialogues and documents: A multi-stage summarization framework for long input dia...

    2022

  27. [35]

    Bin Zhao and Eric P Xing. 2014. Quasi real-time summarization for consumer videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2513--2520

    2014

  28. [36]

    Ming Zhong, Pengfei Liu, Yiran Chen, Danqing Wang, Xipeng Qiu, and Xuan-Jing Huang. 2020. Extractive summarization as text matching. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6197--6208

    2020

  29. [37]

    Ming Zhong, Yang Liu, Yichong Xu, Chenguang Zhu, and Michael Zeng. 2022. Dialoglm: Pre-trained model for long dialogue understanding and summarization. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 11765--11773

    2022

  30. [38]

    Ming Zhong, Da Yin, Tao Yu, Ahmad Zaidi, Mutethia Mutuma, Rahul Jha, Ahmed Hassan, Asli Celikyilmaz, Yang Liu, Xipeng Qiu, et al. 2021. Qmsum: A new benchmark for query-based multi-domain meeting summarization. In Proceedings of the 2021 Conference of the North American Chapte...

    2021

  31. [39]

    Chenguang Zhu, Ruochen Xu, Michael Zeng, and Xuedong Huang. 2020. A hierarchical network for abstractive meeting summarization with cross-domain pretraining. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 194--203

    2020

  32. [40]

    URL: " 'urlintro :=

    ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year eprint doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before...

  33. [41]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.