Pith. sign in

REVIEW 4 major objections 7 minor 45 references

Query-Focused Event Summarization: A Dataset and Benchmark

T0 review · 4 major / 7 minor · reviewed 2026-07-14 · grok-4.5

Pith's one-line read A new dataset and two-stage method make query-focused event summaries practical for large thematic corpora of hundreds or thousands of documents.

desk verdict Solid resource paper that fills a real gap in large-scale query-focused event summarization; the LLM-annotation loop is a real but manageable soft spot, not a fatal circularity. read the letter →

arxiv 2607.11166 v1 pith:AX2W5F3T submitted 2026-07-13 cs.CL

classification cs.CL
keywords query-focusedeventsummarizationQFESQFESumadaptivethresholdingretrievalhierarchicalclusteringthematiccorpusmulti-documentcoreference
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

Existing query-focused summarization datasets and methods ignore large event-oriented corpora and break down when documents number in the hundreds or thousands. This paper defines Query-Focused Event Summarization (QFES), releases QFESum (8 events, 16 684 documents, 104 queries with relevance labels and reference summaries), and supplies a two-stage pipeline: adaptive-threshold retrieval followed by hierarchical event clustering. The pipeline first isolates query-relevant documents without a fixed cutoff, then extracts, clusters and de-duplicates concrete events so the final summary stays tightly focused on the query aspect of the thematic event. Experiments show consistent gains over strong retrieval and graph-based summarization baselines on both lexical and event-matching metrics, including human checks. A sympathetic reader cares because real users rarely want a generic multi-document summary; they want the slice of a long-running event that answers their particular question, and prior resources simply did not support that setting at scale.

What carries the argument

The two-stage RAT+SHC pipeline: RAT samples a document subset, learns a query-specific similarity threshold from BGE scores plus LLM judgments, then retrieves; SHC extracts query-conditioned events, partitions them by hierarchical BERTopic+HDBSCAN clustering ordered by query similarity, selects the most central event per cluster, and finally resolves coreferences.

What would settle it

Replace the LLM-derived relevance labels and extracted events with fully independent human annotations on a held-out thematic corpus and re-run the identical RAT+SHC pipeline; if the performance gap over the same baselines collapses or reverses, the central claim fails.

Watch

Extended reading notes

Core claim

On the newly constructed QFESum benchmark the combination of Query-Focused Retrieval with Adaptive Thresholding (RAT) and Query-Focused Summarization based on Hierarchical Clustering (SHC) produces higher-quality, more concise event summaries than BM25, DPR, RoBERTa-MNLI retrieval and TopicTLS, GraphRAG, FG-RAG or UnstructBase summarization, both when gold documents are given and when documents must first be retrieved.

Load-bearing premise

The bulk of document-relevance labels, event extractions and threshold calibrations rest on Qwen2.5-7B judgments that were only spot-checked by humans on 10 percent of the data.

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

4 major / 7 minor

Summary. The paper defines Query-Focused Event Summarization (QFES): given a large thematic corpus and a user query about one aspect of a shared event, first retrieve relevant documents then produce a query-focused event summary. It constructs QFESum from T17 and CRISIS (8 corpora, 16,684 documents, 104 queries), with human-refined queries and reference event summaries plus document-level relevance labels. It proposes a two-stage pipeline—RAT (dense retrieval with a sample-calibrated adaptive similarity threshold and LLM verification in a band) and SHC (query-focused event extraction, BERTopic then hierarchical HDBSCAN clustering with intra-cluster centrality selection, and GTE coreference filtering)—and reports consistent gains over BM25/DPR/RoBERTa-MNLI retrieval and TopicTLS/GraphRAG/FG-RAG/UnstructBase summarization on ROUGE, BLEU, METEOR, BERTScore, LLM event-alignment F1, and a small human event-match study, with ablations of hierarchy and retrieval.

Significance. If the empirical claims hold under cleaner labels and fairer controls, the work is a useful contribution: existing QFS resources are not event-centric at this scale, and MDES resources lack queries and document–query labels. A public multi-corpus benchmark with relevance sets and event-level references, plus a practical retrieve-then-summarize design aimed at hundreds-to-thousands of documents, would be of clear interest to multi-document and event summarization. Strengths include public code/data, separate retrieval and gold-document summarization tables, ablations (Table 7), per-corpus breakdowns, and an explicit event-alignment evaluation (LLM and limited human). The main value is the task framing and dataset more than any single algorithmic novelty.

major comments (4)
  1. Section 2 (Query-Related Document Annotation) and Figures 7–9: gold Dr_i labels are produced primarily by Qwen2.5-7B using Pra prompts that ChatGPT-4o generates from the query plus the human reference summary Sr_i; only a 10% random sample is human-checked (claimed 97% accuracy). The same Qwen2.5-7B family is then used for RAT’s Illm judgments, SHC event extraction (Pee), and part of the evaluation stack. This creates a load-bearing circularity risk for the central claim that SHC (and RAT+SHC) outperforms document- or graph-based baselines: systematic model preferences over relevance phrasing or event form can shape gold sets, intermediate events, and LLM-F1 together. The paper should (i) report full human re-annotation or multi-annotator agreement on a larger stratified sample of document–query labels (not only 10%), (ii) re-run key tables with labels from a held-out model family or ful
  2. Section 3.2 and Appendix D.3 / Table 12: SHC’s BERTopic/HDBSCAN stage uses per-corpus min_dist, min_cluster_size, and min_samples (e.g., finan 0.005/3/2 vs libya 0.05/19/19), chosen so generated event counts stay near reference length. Baselines are largely run with default or fixed settings (TopicTLS uses fixed top-k clusters by dataset family; GraphRAG/FG-RAG/UnstructBase defaults). Length-matched, corpus-specific clustering knobs make the “method consistently outperforms baselines” claim (Tables 2–5, 9–11) hard to interpret as a pure algorithmic gain rather than output-length and density control. Provide a single global hyperparameter setting (or a fixed compression-ratio rule applied identically to all systems), report sensitivity, and re-evaluate primary tables under that protocol.
  3. Section 4.4 and Table 4: human event-level evaluation covers only eight Yemen queries and only SHC vs FG-RAG. That is too narrow to underwrite (a) superiority of SHC across QFESum and (b) that LLM-Pre/Rec/F1 (DeepSeek-v3 / Qwen3-32B) track human event matching in general—especially given 75% inter-annotator agreement on reference summary construction (Section 2). Expand human evaluation to multiple corpora and at least one additional baseline, and report annotator agreement on the event-match judgments themselves.
  4. Section 3.1 (adaptive threshold) and free parameters δ=0.03, θ=0.90, HDBSCAN R∈{2,4}, 10% sample for Acc(sj): the paper asserts robustness of RAT’s dynamic threshold but does not show sensitivity of retrieval F1 or downstream LLM-F1/ROUGE to these choices or to the LLM used for Illm. Because retrieval quality is argued to drive summarization (Table 5–6), a short sensitivity or leave-one-corpus-out calibration study is needed before the “adaptive thresholding consistently best” claim is load-bearing.
minor comments (7)
  1. Abstract and §1: “thematic corpusrefers” / spacing and hyphenation inconsistencies (e.g., Query-Focused vs query-focused); copy-edit throughout.
  2. Table 1 and prose: event counts in Stls vs E are not always aligned in the narrative example (BPoil “363 events” vs table “552”); clarify construction of E vs raw timeline entries.
  3. Figure 1–2 captions and red/yellow box legend are useful but low-resolution in text form; ensure camera-ready figures label RAT vs SHC stages and inputs/outputs unambiguously.
  4. §4.1 / Appendix D: list exact BGE and GTE model IDs and versions; “BGE” and “GTE” alone are under-specified for reproducibility despite the GitHub link.
  5. Related work (Appendix A) is solid but could more clearly contrast QFES with timeline summarization (T17/CRISIS) and multi-perspective news (NEWTS) in the main body rather than only in the introduction list of limitations.
  6. LLM-Score naming: Tables use LLM1-F1 / LLM2-F1 and LLM-F1 interchangeably; standardize and state which model is default when “LLM-F1” appears alone (DeepSeek-v3).
  7. Ethics/Limitations: note that reference summaries are derived from existing timeline summaries that may themselves embed editorial selection bias; this affects all systems equally but should be stated for dataset users.

Circularity Check

0 steps flagged · score 1.0 of 10

No load-bearing circularity; empirical QFES results rest on a newly constructed dataset evaluated against external baselines and limited human checks, with only mild shared-LLM self-reference that does not force the reported scores by construction.

full rationale

This is an empirical dataset-and-benchmark paper, not a first-principles derivation. The central claims are that RAT+SHC outperform listed retrieval (BM25, DPR, RoBERTa-MNLI) and summarization (TopicTLS, GraphRAG, FG-RAG, UnstructBase) baselines on ROUGE/BLEU/METEOR/BERTScore/LLM-event-F1/human event-matching when run on the newly annotated QFESum (8 corpora, 16k+ docs, 104 queries). Dataset construction starts from external T17/CRISIS timeline summaries, uses human annotators for reference event sets E and query-focused Sr_i (75% agreement, third-annotator resolution), then LLM-assisted document labeling (ChatGPT-4o prompt generation + Qwen2.5-7B sentence judgments) with a 10% human verification sample claiming 97% accuracy. RAT and SHC themselves call the same Qwen family for threshold calibration, event extraction and coreference, and LLM-based evaluation also uses DeepSeek/Qwen. This creates a mild self-referential risk of shared model bias, but it does not reduce any reported metric to an input by algebraic identity, fitted-parameter renaming, or uniqueness theorem. Baselines operate under different extraction/graph regimes (some also use Qwen or GPT variants), gold-document and retrieved-document tables are reported separately, and a small human event-matching study on Yemen queries is provided. No self-citation of a uniqueness result, no ansatz smuggled via prior author work, and no equation that is tautological with its inputs. Per the criteria the paper is self-contained against external baselines; the shared-LLM usage is a correctness/confound concern, not circularity of the enumerated kinds. Score remains low (1) solely for the documented self-use of the annotation/extraction model.

Assumptions & free parameters 5 free parameters · 4 assumptions · 4 invented entities

Empirical NLP paper; load-bearing elements are modeling choices, thresholds, embedding assumptions and LLM reliability rather than mathematical axioms. Free parameters control retrieval aggressiveness and summary length; domain assumptions concern embedding geometry and event extractability; the main invented entities are the task, dataset and two modules.

free parameters (5)
  • δ (similarity band for LLM verification)
    Fixed at 0.03; documents inside [τ*−δ, τ*+δ] are re-checked by LLM. Directly affects precision/recall of RAT.
  • coreference threshold θ
    Fixed at 0.90 (GTE similarity); controls final summary redundancy.
  • HDBSCAN radii R_high / R_mid
    Set to 2 and 4 respectively; determine how finely high- and mid-similarity event clusters are split.
  • per-corpus BERTopic/HDBSCAN hyperparameters (min_dist, min_cluster_size, min_samples)
    Explicitly varied across the eight corpora (Table 12) to keep generated event counts roughly comparable to references; constitutes length-control fitting.
  • sample fraction for adaptive threshold (10 %)
    Used to estimate τ*; choice affects stability of the learned threshold.
assumptions (4)
  • domain assumption BGE (and GTE) cosine similarity is a sufficiently faithful proxy for query–document and event–event semantic relatedness.
    Underpins both the primary retrieval ranking and the final coreference filter (Sections 3.1–3.2).
  • domain assumption Query-relevant content can be reliably extracted as discrete event sentences by a 7B LLM given the designed prompt.
    Event Extraction stage of SHC and the bulk of document annotation rest on this (Section 3.2, Appendix D).
  • ad hoc to paper A 10 % random sample plus LLM binary judgments yields an accurate adaptive similarity threshold for the whole corpus.
    Core of the RAT adaptive-thresholding procedure (Section 3.1).
  • ad hoc to paper Intra-cluster average cosine centrality identifies the single most representative event of a cluster.
    Key-event selection rule inside SHC (Section 3.2).
invented entities (4)
  • QFES task
    purpose: Formalize query-focused summarization specialized to large thematic event corpora with document-level relevance labels.
    Defines the problem the dataset and method address; not previously standardized.
  • QFESum dataset independent evidence
    purpose: Provide 8 thematic corpora, 104 queries, gold relevant-document sets and reference event summaries.
    Primary empirical contribution; constructed via human+LLM pipeline from T17/CRISIS.
  • RAT (Query-Focused Retrieval with Adaptive Thresholding)
    purpose: Retrieve query-relevant documents without a fixed global similarity cutoff.
    First stage of the proposed pipeline.
  • SHC (Query-Focused Summarization based on Hierarchical Clustering)
    purpose: Extract, hierarchically cluster, select and de-duplicate key events into a concise query-focused summary.
    Second stage of the proposed pipeline.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Query-Focused Event Summarization: A Dataset and Benchmark." pith.science (2026). https://pith.science/paper/AX2W5F3T

@misc{pith2026260711166,
  author       = {Pith},
  title        = {Pith review of: Query-Focused Event Summarization: A Dataset and Benchmark},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AX2W5F3T}},
  note         = {Machine review of arXiv:2607.11166}
}
read the original abstract

A thematic corpus is a collection of semantically coherent documents that collectively describe different aspects of a shared thematic event. Such a corpus typically contains hundreds or even thousands of documents. While users' interests in a thematic event often span multiple dimensions, Query-Focused Summarization (QFS) aims to generate summaries tailored to users' queries. However, existing QFS datasets lack event-oriented summarization, and most QFS methods struggle with large-scale corpora. To address these challenges, we propose the Query-Focused Event Summarization (QFES) task and construct the QFESum dataset, which contains 8 thematic events, 16,684 documents, and 104 queries. Furthermore, we introduce a two-stage QFES framework consisting of Query-Focused Retrieval with Adaptive Thresholding (RAT) and Query-Focused Summarization based on Hierarchical Clustering (SHC). Experimental results on QFESum show that RAT and SHC consistently outperform the baselines, demonstrating their effectiveness for QFES. The dataset and code are publicly available at https://github.com/sarcasm-hcy02/QFES-QFESum.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 19 linked inside Pith

  1. [1]

    The k-means algorithm: A comprehensive survey and performance evaluation.Electronics, 9(8):1295, 2020

    Mohiuddin Ahmed, Raihan Seraj, and Syed Mohammed Shamsul Islam. The k-means algorithm: A comprehensive survey and performance evaluation.Electronics, 9(8):1295, 2020

  2. [2]

    Newts: A corpus for news topic-focused summarization.arXiv preprint arXiv:2205.15661, 2022

    Seyed Ali Bahrainian, Sheridan Feucht, and Carsten Eickhoff. Newts: A corpus for news topic-focused summarization.arXiv preprint arXiv:2205.15661, 2022

  3. [3]

    Meteor: An automatic metric for mt evaluation with improved correlation with human judgments

    Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. InProceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65–72, 2005

  4. [4]

    Topic concentration in query focused summarization datasets

    Tal Baumel, Raphael Cohen, and Michael Elhadad. Topic concentration in query focused summarization datasets. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 30, 2016

  5. [5]

    Longformer: The long-document transformer.arXiv preprint arXiv:2004.05150, 2020

    Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer.arXiv preprint arXiv:2004.05150, 2020

  6. [6]

    Bge m3-embedding: Multi-lingual, multi- functionality, multi-granularity text embeddings through self-knowledge distillation.arXiv preprint arXiv:2402.03216, 2024

    Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. Bge m3-embedding: Multi-lingual, multi- functionality, multi-granularity text embeddings through self-knowledge distillation.arXiv preprint arXiv:2402.03216, 2024

  7. [7]

    Overview of duc 2005

    Hoa Trang Dang. Overview of duc 2005. InProceedings of the document understanding conference, volume 2005, pages 1–12, 2005

  8. [8]

    Overview of the tac 2008 update summarization task

    Hoa Trang Dang, Karolina Owczarzak, et al. Overview of the tac 2008 update summarization task. InTAC, 2008

Show all 45 references
  1. [9]

    Webcites: Attributed query-focused summarization on chinese web search results with citations

    Haolin Deng, Chang Wang, Li Xin, Dezhang Yuan, Junlang Zhan, Tian Zhou, Jin Ma, Jun Gao, and Ruifeng Xu. Webcites: Attributed query-focused summarization on chinese web search results with citations. InProceedings of the 62nd Annual Meeting of the Association for Computational...

  2. [10]

    From local to global: A graph rag approach to query-focused summarization

    Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Truitt, Dasha Metropolitansky, Robert Osazuwa Ness, and Jonathan Larson. From local to global: A graph rag approach to query-focused summarization. arXiv preprint arXiv:2404.16130, 2024

  3. [11]

    Retrieval-augmented generation for large language models: A survey.arXiv preprint arXiv:2312.10997, 2(1), 2023

    Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yixin Dai, Jiawei Sun, Haofen Wang, and Haofen Wang. Retrieval-augmented generation for large language models: A survey.arXiv preprint arXiv:2312.10997, 2(1), 2023

  4. [12]

    Examining the state-of-the-art in news timeline summarization.arXiv preprint arXiv:2005.10107, 2020

    Demian Gholipour Ghalandari and Georgiana Ifrim. Examining the state-of-the-art in news timeline summarization.arXiv preprint arXiv:2005.10107, 2020

  5. [13]

    Bertopic: Neural topic modeling with a class-based tf-idf procedure.arXiv preprint arXiv:2203.05794, 2022

    Maarten Grootendorst. Bertopic: Neural topic modeling with a class-based tf-idf procedure.arXiv preprint arXiv:2203.05794, 2022

  6. [14]

    Lightrag: Simple and fast retrieval-augmented generation

    Zirui Guo, Lianghao Xia, Yanhua Yu, Tu Ao, and Chao Huang. Lightrag: Simple and fast retrieval-augmented generation. arXiv preprint arXiv:2410.05779, 2024

  7. [15]

    Fg-rag: Enhancing query-focused summarization with context- aware fine-grained graph rag.arXiv preprint arXiv:2504.07103, 2025

    Yubin Hong, Chaofan Li, Jingyi Zhang, and Yingxia Shao. Fg-rag: Enhancing query-focused summarization with context- aware fine-grained graph rag.arXiv preprint arXiv:2504.07103, 2025

  8. [16]

    From moments to milestones: Incremental timeline summarization leveraging large language models

    Qisheng Hu, Geonsik Moon, and Hwee Tou Ng. From moments to milestones: Incremental timeline summarization leveraging large language models. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7232–7246, 2024

  9. [17]

    Dense passage retrieval for open-domain question answering

    Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick SH Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. Dense passage retrieval for open-domain question answering. InEMNLP (1), pages 6769–6781, 2020

  10. [18]

    Aquamuse: Automatically generating datasets for query-based multi-document summarization.arXiv preprint arXiv:2010.12694, 2020

    Sayali Kulkarni, Sheide Chammas, Wan Zhu, Fei Sha, and Eugene Ie. Aquamuse: Automatically generating datasets for query-based multi-document summarization.arXiv preprint arXiv:2010.12694, 2020

  11. [19]

    Natural questions: a benchmark for question answering research.Transactions of the Association for Computational Linguistics, 7:453–466, 2019

    Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, et al. Natural questions: a benchmark for question answering research.Transactions of the Association for Computat...

  12. [20]

    Simple is effective: The roles of graphs and large language models in knowledge-graph-based retrieval-augmented generation.arXiv preprint arXiv:2410.20724, 2024

    Mufei Li, Siqi Miao, and Pan Li. Simple is effective: The roles of graphs and large language models in knowledge-graph-based retrieval-augmented generation.arXiv preprint arXiv:2410.20724, 2024

  13. [21]

    Towards general text embeddings with multi-stage contrastive learning.arXiv preprint arXiv:2308.03281, 2023

    Zehan Li, Xin Zhang, Yanzhao Zhang, Dingkun Long, Pengjun Xie, and Meishan Zhang. Towards general text embeddings with multi-stage contrastive learning.arXiv preprint arXiv:2308.03281, 2023. 12 CHENYU HU AND BANG W ANG

  14. [22]

    Rouge: A package for automatic evaluation of summaries

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

  15. [23]

    Deepseek-v3 technical report.arXiv preprint arXiv:2412.19437, 2024

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report.arXiv preprint arXiv:2412.19437, 2024

  16. [24]

    Learning to rank utterances for query-focused meeting summarization.arXiv preprint arXiv:2305.12753, 2023

    Xingxian Liu and Yajing Xu. Learning to rank utterances for query-focused meeting summarization.arXiv preprint arXiv:2305.12753, 2023

  17. [25]

    Roberta: A robustly optimized bert pretraining approach.arXiv preprint arXiv:1907.11692, 2019

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach.arXiv preprint arXiv:1907.11692, 2019

  18. [26]

    hdbscan: Hierarchical density based clustering.J

    Leland McInnes, John Healy, Steve Astels, et al. hdbscan: Hierarchical density based clustering.J. Open Source Softw., 2(11):205, 2017

  19. [27]

    Reinforcement replaces supervision: Query focused summarization using deep reinforcement learning

    Swaroop Nath, Pushpak Bhattacharyya, and Harshad Khadilkar. Reinforcement replaces supervision: Query focused summarization using deep reinforcement learning. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 15770–15789, 2023

  20. [28]

    Socratic pretraining: Question-driven pretraining for controllable summarization

    Artidoro Pagnoni, Alex Fabbri, Wojciech Kry´sci´nski, and Chien-Sheng Wu. Socratic pretraining: Question-driven pretraining for controllable summarization. InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages ...

  21. [29]

    Richard Yuanzhe Pang, Alicia Parrish, Nitish Joshi, Nikita Nangia, Jason Phang, Angelica Chen, Vishakh Padmakumar, Johnny Ma, Jana Thompson, He He, et al. Quality: Question answering with long input texts, yes! InProceedings of the 2022 Conference of the North American Chapter...

  22. [30]

    Bleu: a method for automatic evaluation of machine translation

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. InProceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311–318, 2002

  23. [31]

    Scaling multi-document event summarization: Evaluating compression vs

    Adithya Pratapa and Teruko Mitamura. Scaling multi-document event summarization: Evaluating compression vs. full-text approaches.arXiv preprint arXiv:2502.06617, 2025

  24. [32]

    The probabilistic relevance framework: Bm25 and beyond.Foundations and Trends® in Information Retrieval, 3(4):333–389, 2009

    Stephen Robertson, Hugo Zaragoza, et al. The probabilistic relevance framework: Bm25 and beyond.Foundations and Trends® in Information Retrieval, 3(4):333–389, 2009

  25. [33]

    Chatgpt and open-ai models: A preliminary review.Future Internet, 15(6):192, 2023

    Konstantinos I Roumeliotis and Nikolaos D Tselikas. Chatgpt and open-ai models: A preliminary review.Future Internet, 15(6):192, 2023

  26. [34]

    Review on query-focused multi-document summarization (qmds) with comparative analysis.ACM Computing Surveys, 56(1):1–38, 2023

    Prasenjeet Roy and Suman Kundu. Review on query-focused multi-document summarization (qmds) with comparative analysis.ACM Computing Surveys, 56(1):1–38, 2023

  27. [35]

    Learning to rank salient content for query-focused summarization.arXiv preprint arXiv:2411.00324, 2024

    Sajad Sotudeh and Nazli Goharian. Learning to rank salient content for query-focused summarization.arXiv preprint arXiv:2411.00324, 2024

  28. [36]

    Towards verifiable text generation with evolving memory and self-reflection

    Hao Sun, Hengyi Cai, Bo Wang, Yingyan Hou, Xiaochi Wei, Shuaiqiang Wang, Yan Zhang, and Dawei Yin. Towards verifiable text generation with evolving memory and self-reflection. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 8211–...

  29. [37]

    Timeline summarization from relevant headlines

    Giang Tran, Mohammad Alrifai, and Eelco Herder. Timeline summarization from relevant headlines. InEuropean Conference on Information Retrieval, pages 245–256. Springer, 2015

  30. [38]

    Attention is all you need.Advances in neural information processing systems, 30, 2017

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

  31. [39]

    Exploring neural models for query-focused summarization

    Jesse Vig, Alexander Richard Fabbri, Wojciech Kry´sci´nski, Chien-Sheng Wu, and Wenhao Liu. Exploring neural models for query-focused summarization. InFindings of the association for computational linguistics: NAACL 2022, pages 1455–1468, 2022

  32. [40]

    Manifold-ranking based topic-focused multi-document summarization

    Xiaojun Wan, Jianwu Yang, Jianguo Xiao, et al. Manifold-ranking based topic-focused multi-document summarization. In IJCAI, volume 7, pages 2903–2908, 2007

  33. [41]

    Squality: Building a long-document summarization dataset the hard way

    Alex Wang, Richard Yuanzhe Pang, Angelica Chen, Jason Phang, and Samuel Bowman. Squality: Building a long-document summarization dataset the hard way. InProceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 1139–1156, 2022

  34. [42]

    Unstructured evidence attribution for long context query focused summarization

    Dustin Wright, Zain Muhammad Mujahid, Lu Wang, Isabelle Augenstein, and David Jurgens. Unstructured evidence attribution for long context query focused summarization. In Christos Christodoulopoulos, Tanmoy Chakraborty, Carolyn Rose, and Violet Peng, editors,Proceedings of the ...

  35. [43]

    Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025

  36. [44]

    Bertscore: Evaluating text generation with bert.arXiv preprint arXiv:1904.09675, 2019

    Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert.arXiv preprint arXiv:1904.09675, 2019. QUERY-FOCUSED EVENT SUMMARIZATION: A DATASET AND BENCHMARK 13

  37. [45]

    The arrest and charges against Gadhafi

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

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.