Pith. sign in

REVIEW 5 major objections 6 minor 41 references

NEWSCOPE claims that reranking retrieved news by sentence-level clusters yields more diverse perspectives on an event at roughly the same relevance as dense retrieval.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

NEWSCOPE adds sentence-level clustering and cluster-aware greedy reranking to dense news retrieval, reporting higher viewpoint diversity on two new benchmarks, at a small relevance cost.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Real resource contribution and a sensible pipeline, but the headline claim rests on circular metrics and the only independent diversity metric goes against it. the 5 major comments →

arxiv 2508.19758 v2 pith:IOX5PN2K submitted 2025-08-27 cs.CL cs.IR

Uncovering the Bigger Picture: Comprehensive Event Understanding Via Diverse News Retrieval

Yixuan Tang , Yuanyuan Shi , Yiqun Sun , Anthony Kum Hoe Tung This is my paper
classification cs.CL cs.IR
keywords diverse news retrievalsentence-level clusteringdiversity-aware rerankingevent-centric retrievalOPTICSevaluation metricsecho chambersLocalNews
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

News retrieval systems rank by topical relevance and therefore tend to return near-duplicate paragraphs that repeat the same quotes and angles. This paper proposes NEWSCOPE, a two-stage framework: a dense retriever gathers candidate paragraphs, then a sentence-level clustering and greedy reranker selects paragraphs that cover the most distinct sentence clusters, making the retrieved set reflect more perspectives on the same event. To back this up, the authors define three interpretable diversity metrics — average pairwise distance, positive cluster coverage, and information density ratio — and build two paragraph-level benchmarks, LocalNews and DSGlobal. Across retrieval depths from 5 to 50, NEWSCOPE reports higher diversity on those metrics than BM25, dense retrieval, MMR, and DkMIPS, with relevance scores close to the relevance-first baselines.

Core claim

The central claim is that semantic diversity in news retrieval can be modeled at the sentence level and traded off against relevance without much loss. NEWSCOPE treats sentences as atomic meaning units, clusters them with OPTICS into groups that stand for distinct informational aspects, and then greedily picks paragraphs that cover the largest number of still-uncovered clusters. Its GreedyPlus variant weights each cluster by its average relevance to the headline and adds a relevance term, so the final ranking jointly optimizes coverage and topical fit. The paper's evidence is that on both benchmarks, GreedyPlus roughly doubles Information Density Ratio at shallow depths and raises Positive C

What carries the argument

The load-bearing mechanism is the pairing of OPTICS density-based clustering (which needs no preset cluster count) with greedy cluster-coverage selection. Each paragraph is scored by how many still-uncovered sentence clusters it contains; GreedyPlus refines this by weighting each cluster with its average reranker relevance to the headline and adding a lambda-weighted query similarity term. The metrics C and I both presume that these sentence clusters are faithful proxies for distinct perspectives, which is the assumption on which the entire evaluation rests.

Load-bearing premise

The whole method and its main evaluation metrics assume that OPTICS clusters over sentence embeddings correspond to genuinely distinct news perspectives or event aspects; if clusters are arbitrary groupings, the reported diversity gains are artifacts of the clustering, not real viewpoint coverage.

What would settle it

Take the candidate pool for a set of events, assign sentences to the same number of clusters at random, rerun GreedyPlus, and compare C and I to the reported values; if random clusters reproduce the gains, the cluster structure carries no information. Independently, have human annotators judge whether pairs of paragraphs drawn from different clusters cover the same or different aspects; agreement near chance would falsify the proxy.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Retrieval systems can expose readers to complementary angles by reranking a small candidate pool of 100 paragraphs, at roughly 1.2 seconds per event, without retraining.
  • The three metrics make viewpoint diversity measurable and interpretable, letting teams optimize for cluster coverage (C), information density (I), or pairwise semantic spread (D) without needing stance labels.
  • The lambda parameter provides a continuous relevance-diversity dial; at lambda=0.5 the reported F1 stays close to relevance-only retrieval while coverage metrics improve sharply.
  • The same pipeline and metrics work on both a regional and a global news benchmark, indicating the approach generalizes across event types and source distributions.
  • The clustering itself surfaces event aspects, as the appendix example shows five clusters capturing distinct details of a single flight-delay incident.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The sentence-cluster proxy could be validated against human-annotated event aspects; if clusters match reader-perceived angles, C becomes a true coverage measure and the same machinery could audit search engines for echo-chamber bias.
  • Applying this reranker inside retrieval-augmented generation could broaden the evidential base an LLM draws on; the paper only gestures at fact-checking as future work.
  • Because GreedyPlus down-weights clusters with low query similarity, it may systematically suppress minority or weakly relevant viewpoints; a fairness constraint could be tested to see whether coverage of controversial stances survives.
  • The approach likely transfers to non-news collections where 'aspect' means subtopic, such as scientific literature or legal documents, but the cluster semantics would need revalidation in each domain.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper proposes NEWSCOPE, a two-stage framework for event-centric diverse news retrieval. Stage I uses dense retrieval to get top-100 candidate paragraphs; Stage II clusters sentences with OPTICS, then greedily selects paragraphs to cover the resulting clusters (GreedySCS), optionally weighting cluster importance by relevance and adding a relevance term (GreedyPlus). The authors construct two paragraph-level benchmarks (LocalNews, DSGlobal), and introduce three diversity metrics: Average Pairwise Distance (D), Positive Cluster Coverage (C), and Information Density Ratio (I). They report that NEWSCOPE 'consistently outperforms strong baselines, achieving significantly higher diversity without compromising relevance.' The paper also includes ablations, a qualitative example, and an efficiency analysis.

Significance. The task formulation is timely and the two benchmarks may be useful resources. The idea of modeling diversity at sentence level rather than whole-document level is reasonable and could be a genuine contribution. The code and data are promised to be released, which is a strength. However, the central empirical claim — that NEWSCOPE achieves higher diversity without sacrificing relevance — rests on evaluation metrics (C and I) that are computed from the exact OPTICS sentence clusters that the greedy selection procedure is designed to cover. The only cluster-independent diversity metric (D) does not consistently favor NEWSCOPE, and relevance is lower than DenseRetr at most depths. No human judgment of perspective diversity, no independent aspect annotation, and no significance tests are provided. The paper's present form is therefore more a proposal of a method and benchmark than a validated demonstration of the claimed advantage. If the circularity concern is addressed (e.g., with human-annotated perspective labels and honest reporting of D), the contribution could be solid.

major comments (5)
  1. [§6.2, §4.2.2–4.2.3] The headline diversity metrics C and I are defined over the same OPTICS sentence clusters that GreedySCS and GreedyPlus explicitly maximize coverage of (Score(p)=|U∩Clusters(p)| and Score+(p)=Σ ClusterScore(c)). Consequently, C and I are close to the training objective, and high scores on these metrics are expected by construction. Appendix B does not resolve the concern: the five shown clusters are near-duplicate sentence paraphrases, not evidence that clusters correspond to distinct perspectives or aspects. An independent evaluation — e.g., human annotation of event aspects, or usage of an external diversity measure that does not use the OPTICS clustering — is needed before the 'higher diversity' claim can be accepted.
  2. [Tables 2 and 3, D column] The cluster-free diversity metric D contradicts the abstract's claim of consistent outperformance. On LocalNews, GreedyPlus D is 29.8 at top-5 (vs. DkMIPS 33.4), 30.2 at top-10 (vs. DkMIPS 34.6), 34.9 at top-20 (vs. DkMIPS 41.8), and 47.6 at top-50 (vs. MMR 51.3, DkMIPS 53.7). On DSGlobal, GreedyPlus D is below MMR at top-5 (33.8 vs. 35.2) and below both MMR and DkMIPS at top-20 (35.7 vs. 35.9, 37.2) and top-50 (47.1 vs. 51.8, 51.2). Thus the 'consistently outperforms' claim is not supported by the paper's own data unless D is argued to be less relevant, which would need a principled justification.
  3. [Tables 2 and 3, P/R/F1 columns] The statement 'without compromising relevance' is not supported by the reported relevance metrics. On LocalNews, GreedyPlus F1 is below DenseRetr at top-5 (27.5 vs. 29.3), top-10 (46.8 vs. 48.9), and top-20 (67.1 vs. 67.8). On DSGlobal, GreedyPlus F1 is below DenseRetr at top-5 (18.7 vs. 19.4), top-10 (33.0 vs. 34.9), and top-20 (55.0 vs. 57.9). The paper also uses the word 'significantly' but provides no confidence intervals, error bars, or significance tests. These deficits directly affect the abstract's central claim and must be addressed.
  4. [Algorithm 1] Algorithm 1's while condition is `while |S| > k or |U| > coverage_threshold`. As written, if k is the desired number of paragraphs and the selection starts with S empty, the condition `|S| > k` is false for any k ≥ 0, so the loop selects nothing. The intended condition is presumably `|S| < k` (with an appropriate stopping condition, likely `and` rather than `or` when coverage_threshold is used). This is not a minor typo: the algorithm defines the method, and the pseudocode as printed is unexecutable. The parameter `coverage_threshold` is also never defined or set in the main text.
  5. [§6.1, λ and evaluation protocol] The paper sets λ=0.5 and reports only that value in the main tables, with tuning in Appendix F. However, the main tables compare GreedyPlus to baselines using different λ settings where appropriate (MMR and DkMIPS also have λ). It is unclear whether the baseline hyperparameters were tuned on the same validation setup, and no variance across events is reported. Given the small numbers of events (103 and 147), the absence of per-event variance reporting or paired tests weakens the generality of all cross-system comparisons.
minor comments (6)
  1. [§4.2.2] The text after Algorithm 1 says 'The process continues until a stopping condition is met: either a predefined number of paragraphs is selected or sufficient cluster coverage is achieved.' The pseudocode's `or` and the condition `|S| > k` are inconsistent with this description; please fix both.
  2. [Abstract/§6.2] The phrase 'significantly higher diversity' appears in the abstract and Section 7, but no statistical significance is computed anywhere. Either add tests or replace 'significantly' with 'substantially.'
  3. [§6.2] The definitions of C and I do not state whether the cluster assignments come from clustering the candidate set only, or the entire corpus, and whether relevant paragraphs that are not retrieved contribute to the denominator. This affects the interpretability of the metrics.
  4. [Appendix A.3] There is a duplicated word: 'to to represent each event.' Also, the annotation procedure says 'three rounds' but the agreement rate is 98.7%; details on how disagreements were resolved and how 'auto-labeled as irrelevant' interacts with the LLM labels would be useful.
  5. [§7.3] The ablation claim that removing the diversity term 'slightly improves F1' is accurate for some depths, but in the LocalNews top-20 row the F1 improvement is from 67.1 to 68.3, which is small; please note this is not statistically tested.
  6. [Appendix B] The example clusters illustrate near-duplicate sentences, not distinct perspectives. Consider presenting clusters that are actually different angles or viewpoints, or explicitly acknowledge that OPTICS clusters often represent paraphrase groups rather than aspects.

Circularity Check

2 steps flagged

The headline diversity win is self-definitional: NEWSCOPE's reranker maximizes coverage of OPTICS clusters, and the headline diversity metrics C and I count coverage of those same clusters. The independent D metric shows NEWSCOPE is not consistently most diverse.

specific steps
  1. self definitional [Algorithm 1 / Section 4.2.2 (Greedy Cluster Selection) vs Section 6.2 (Evaluation Metrics, C and I)]
    ""Initialize S ← ∅, U ← C" and "Score (p) ← |U ∩ Clusters(p)|" (Algorithm 1); "C = # Covered Clusters / # Total Clusters in Relevant Paragraphs" and "I = # Covered Clusters / # Total Sentences in Retrieved Paragraphs" (Section 6.2)."

    GreedySCS's selection score is literally the count of uncovered sentence clusters covered by a paragraph, and U is initialized to all clusters C from the same OPTICS clustering used in evaluation. C and I are defined as the covered-cluster count divided by total relevant clusters/sentences. Thus a high C/I for GreedySCS is not independent evidence of viewpoint diversity; it is the greedy objective itself renamed as a metric. The paper supplies no human aspect annotation; Appendix B only shows near-duplicate sentence groups, so the chain 'clusters = perspectives -> coverage = diversity' is assumed, not tested.

  2. self definitional [Section 4.2.3 (Cluster-Based Weighting, GreedyPlus) vs Section 6.2 (Evaluation Metrics, C and I)]
    ""The diversity term encourages the selection of paragraphs that introduce novel content by covering high-quality clusters that have not yet been selected" (Section 4.2.3); "C = # Covered Clusters / # Total Clusters in Relevant Paragraphs" and "I = # Covered Clusters / # Total Sentences in Retrieved Paragraphs" (Section 6.2)."

    GreedyPlus's diversity term is the sum of relevance-weighted scores of yet-uncovered clusters in a paragraph, and the evaluation metrics C and I count exactly the number of covered clusters. Optimizing Score+ therefore directly optimizes C and I by construction. On the only cluster-free diversity metric, D (average pairwise distance), NEWSCOPE (GreedyPlus) is below DkMIPS and MMR in LocalNews at all depths (e.g., top-10 D=30.2 vs DkMIPS 34.6; top-20 D=34.9 vs DkMIPS 41.8) and below MMR on DSGlobal top-5 and top-50, so the 'consistently outperforming' claim in the abstract is not independently supported.

full rationale

The central circularity is that the paper's flagship diversity metrics C and I are computed from the same sentence clusters that the NEWSCOPE rerankers are explicitly designed to cover. Section 4.2.2 defines GreedySCS's score as |U ∩ Clusters(p)|, and Section 4.2.3 defines GreedyPlus's diversity term as the sum of ClusterScore over clusters(p) ∩ U; Section 6.2 then defines C and I as covered-cluster counts. So the reported 'significantly higher diversity' on C/I is a tautological consequence of the selection rule, not a measurement of an independently defined property. The paper's own Appendix B validates clusters only by listing near-duplicate sentence paraphrases, not by showing that clusters correspond to distinct news aspects or perspectives, so the premise that cluster coverage equals viewpoint coverage is assumed. The one non-circular diversity metric, D, does not support the abstract's consistency claim: on LocalNews, GreedyPlus's D is lower than DkMIPS at top-5/10/20/50, and on DSGlobal it is lower than MMR at top-5 and top-50. Relevance is also slightly below DenseRetr at most depths (e.g., LocalNews top-20 F1 67.1 vs DenseRetr 67.8). There is no machine-checked or externally validated support that would make the cluster coverage metrics independent. Self-citations to DkMIPS and QALinkPlus are present but not load-bearing for the main result. Overall, the central diversity claim reduces by construction on C/I, while the independent evidence is mixed; score 7 reflects that the headline claim is largely circular but not the entire paper (D, relevance, and qualitative examples retain some independent content).

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

No new physical entities are postulated. The key conceptual construct is 'sentence cluster as event aspect', listed as an axiom because it is unvalidated and underpins both the method and the main metrics.

free parameters (4)
  • lambda (λ) = 0.5
    Trade-off weight between diversity and relevance in GreedyPlus; tuned on the test benchmarks over values 1/8 to 8 (Appendix F), not selected by a validation split or theory.
  • candidate pool size = 100
    Number of top paragraphs passed from Stage I to Stage II. Chosen by authors; no sensitivity analysis.
  • OPTICS hyperparameters = not reported
    min_samples, xi and related parameters determine cluster granularity; needed to reproduce clustering and metrics but not specified in paper.
  • coverage_threshold = not reported
    Stopping condition in Algorithm 1; no value given.
axioms (5)
  • domain assumption Cosine similarity in the bilingual-embedding-large space captures paragraph-level relevance and semantic variation.
    Used in Stage I ranking, pairwise distance metric D, and cluster scoring. Not proven; depends on pretrained embeddings.
  • domain assumption OPTICS sentence clusters correspond to distinct perspectives or aspects of an event.
    This is the foundational premise of both re-ranking and the C/I metrics; no human validation that clusters equal aspects (Section 4.2.1, Appendix B only illustrative).
  • domain assumption GPT-4o-mini generated relevance labels are accurate ground truth.
    Query and relevance labels for LocalNews are LLM-generated with iterative refinement and only 100-sample human review; DSGlobal label provenance is not detailed.
  • domain assumption Google News Full Coverage provides a balanced, unbiased multi-source corpus.
    Data collection depends on this; authors tried to mitigate personalization via VPN and multiple conditions (Appendix A.1) but cannot guarantee coverage completeness.
  • domain assumption The top-100 candidate pool from dense retrieval contains sufficient diverse relevant content.
    Stage II can only rerank within this pool; if dense retrieval misses perspectives, they cannot be recovered. No recall audit beyond 80% coverage claim.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Uncovering the Bigger Picture: Comprehensive Event Understanding Via Diverse News Retrieval." pith.science (2026). https://pith.science/paper/IOX5PN2K

@misc{pith2026250819758,
  author       = {Pith},
  title        = {Pith review of: Uncovering the Bigger Picture: Comprehensive Event Understanding Via Diverse News Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IOX5PN2K}},
  note         = {Machine review of arXiv:2508.19758}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Access to diverse perspectives is essential for understanding real-world events, yet most news retrieval systems prioritize textual relevance, leading to redundant results and limited viewpoint exposure. We propose NEWSCOPE, a two-stage framework for diverse news retrieval that enhances event coverage by explicitly modeling semantic variation at the sentence level. The first stage retrieves topically relevant content using dense retrieval, while the second stage applies sentence-level clustering and diversity-aware re-ranking to surface complementary information. To evaluate retrieval diversity, we introduce three interpretable metrics, namely Average Pairwise Distance, Positive Cluster Coverage, and Information Density Ratio, and construct two paragraph-level benchmarks: LocalNews and DSGlobal. Experiments show that NEWSCOPE consistently outperforms strong baselines, achieving significantly higher diversity without compromising relevance. Our results demonstrate the effectiveness of fine-grained, interpretable modeling in mitigating redundancy and promoting comprehensive event understanding. The data and code are available at https://github.com/tangyixuan/NEWSCOPE.

Figures

Figures reproduced from arXiv: 2508.19758 by Anthony Kum Hoe Tung, Yiqun Sun, Yixuan Tang, Yuanyuan Shi.

Figure 1
Figure 1. Figure 1: Comparison of dense and diverse retrieval [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the NEWSCOPE framework. The two-stage pipeline combines efficient dense retrieval with sentence-level diversity-aware re-ranking to ensure both relevance and comprehensive perspective coverage. 2024; Huang et al., 2024a; Fabbri et al., 2019), which motivate the need for retrieval strategies that ensure balanced and diverse event representations. 2.2 Diverse Retrieval Traditional retrieval model… view at source ↗
Figure 3
Figure 3. Figure 3: Visualization of model performance across [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Positive vs. negative paragraph counts per [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Event distribution by number of positive para [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

41 extracted references · 33 canonical work pages · 6 internal anchors

  1. [1]

    online" 'onlinestring :=

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

  2. [2]

    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 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Shunit Agmon, Amir Gilad, Brit Youngmann, Shahar Zoarets, and Benny Kimelfeld. 2024. Finding convincing views to endorse a claim. CoRR, abs/2408.14974

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  4. [4]

    Mihael Ankerst, Markus M Breunig, Hans-Peter Kriegel, and J \"o rg Sander. 1999. Optics: Ordering points to identify the clustering structure. ACM Sigmod record, 28(2):49--60

    work page 1999

  5. [5]

    Jaime Carbonell and Jade Goldstein. 1998. The use of MMR , diversity-based reranking for reordering documents and producing summaries. In SIGIR, pages 335--336

    work page 1998

  6. [6]

    Laming Chen, Guoxin Zhang, and Eric Zhou. 2018. Fast greedy MAP inference for determinantal point process to improve recommendation diversity. In NeurIPS, pages 5627--5638

    work page 2018

  7. [7]

    Sihao Chen, Daniel Khashabi, Chris Callison - Burch, and Dan Roth. 2019. Perspectroscope: A window to the world of diverse perspectives. In ACL (3) , pages 129--134. Association for Computational Linguistics

    work page 2019

  8. [8]

    Ting-Chih Chen, Chia-Wei Tang, and Chris Thomas. 2024. Metasumperceiver: Multimodal multi-document evidence summarization for fact-checking. In ACL, pages 8742--8757

    work page 2024

  9. [9]

    Yuwei Chuai, Anastasia Sergeeva, Gabriele Lenzini, and Nicolas Pröllochs. 2024. Community fact-checks trigger moral outrage in replies to misleading posts on social media. CoRR, abs/2409.08829

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  10. [10]

    Andrew Estornell, Sanmay Das, and Yevgeniy Vorobeychik. 2020. Deception through half-truths. In AAAI, pages 10110--10117

    work page 2020

  11. [11]

    Fabbri, Irene Li, Tianwei She, Suyi Li, and Dragomir R

    Alexander R. Fabbri, Irene Li, Tianwei She, Suyi Li, and Dragomir R. Radev. 2019. Multi-news: A large-scale multi-document summarization dataset and abstractive hierarchical model. In ACL, pages 1074--1084

    work page 2019

  12. [12]

    Mohsen Fayyaz, Ali Modarressi, Hinrich Schuetze, and Nanyun Peng. 2025. Collapse of dense retrievers: Short, early, and literal biases outranking factual evidence. arXiv preprint arXiv:2503.05037

    Pith/arXiv arXiv 2025

  13. [13]

    Fung, Kung-Hsiang Huang, Preslav Nakov, and Heng Ji

    Yi R. Fung, Kung-Hsiang Huang, Preslav Nakov, and Heng Ji. 2022. The battlefront of combating misinformation and coping with media bias. In KDD, pages 4790--4791

    work page 2022

  14. [14]

    Xuming Hu, Zhijiang Guo, Guanyu Wu, Lijie Wen, and Philip S. Yu. 2023. Give me more details: Improving fact-checking with latent retrieval. CoRR, abs/2305.16128

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  15. [15]

    Fabbri, Prafulla Kumar Choubey, Shafiq Joty, Caiming Xiong, and Chien-Sheng Wu

    Kung-Hsiang Huang, Philippe Laban, Alexander R. Fabbri, Prafulla Kumar Choubey, Shafiq Joty, Caiming Xiong, and Chien-Sheng Wu. 2024 a . Embrace divergence for richer insights: A multi-document summarization benchmark and a case study on summarizing diverse information from news articles. In NAACL-HLT, pages 570--593

    work page 2024

  16. [16]

    Qiang Huang, Yanhao Wang, Yiqun Sun, and Anthony Kum Hoe Tung. 2024 b . Diversity-aware k-maximum inner product search revisited. CoRR, abs/2402.13858

    Pith/arXiv arXiv 2024

  17. [17]

    Israa Jaradat, Haiqi Zhang, and Chengkai Li. 2024. On detecting cherry-picking in news coverage using large language models. CoRR, abs/2401.05650

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  18. [18]

    Zhengbao Jiang, Ji - Rong Wen, Zhicheng Dou, Wayne Xin Zhao, Jian - Yun Nie, and Ming Yue. 2017. Learning to diversify search results via subtopic attention. In SIGIR , pages 545--554. ACM

    work page 2017

  19. [19]

    Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6769--6781

    2020

  20. [20]

    Brendan King and Jeffrey Flanigan. 2023. Diverse retrieval-augmented in-context learning for dialogue state tracking. In ACL, pages 5570--5585

    work page 2023

  21. [21]

    Juhi Kulshrestha, Motahhare Eslami, Johnnatan Messias, Muhammad Bilal Zafar, Saptarshi Ghosh, Krishna P Gummadi, and Karrie Karahalios. 2019. Search bias quantification: investigating political bias in social media and web search. Information Retrieval Journal, 22:188--227

    work page 2019

  22. [22]

    Niklas Muennighoff, Nouamane Tazi, Loic Magne, and Nils Reimers. 2023. MTEB : Massive text embedding benchmark. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics (EACL), pages 2014--2037

    work page 2023

  23. [23]

    Stephen E Robertson, Steve Walker, Susan Jones, Micheline M Hancock-Beaulieu, Mike Gatford, et al. 1995. Okapi at TREC-3 . NIST Special Publication, pages 109--123

    work page 1995

  24. [24]

    Francisco-Javier Rodrigo-Ginés, Jorge Carrillo de Albornoz, and Laura Plaza. 2024. A systematic review on media bias detection: What is media bias, how it is expressed, and how to detect it. Expert Syst. Appl., 237(Part C):121641

    work page 2024

  25. [25]

    Daniel Russo, Stefano Menini, Jacopo Staiano, and Marco Guerini. 2024. Face the facts! evaluating rag-based fact-checking pipelines in realistic settings. CoRR, abs/2412.15189

    work page arXiv 2024

  26. [26]

    "Beware of deception": Detecting Half-Truth and Debunking it through Controlled Claim Editing

    Sandeep Singamsetty, Nishtha Madaan, Sameep Mehta, Varad Bhatnagar, and Pushpak Bhattacharyya. 2023. "beware of deception": Detecting half-truth and debunking it through controlled claim editing. CoRR, abs/2308.07973

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  27. [27]

    Ronit Singhal, Pransh Patwa, Parth Patwa, Aman Chadha, and Amitava Das. 2024. Evidence-backed fact checking using rag and few-shot in-context learning with llms. CoRR, abs/2408.12060

    Pith/arXiv arXiv 2024

  28. [28]

    Timo Spinde, Christin Jeggle, Magdalena Haupt, Wolfgang Gaissmaier, and Helge Giese. 2022. How do we raise media bias awareness effectively? effects of visualizations to communicate bias. Plos one, 17(4):e0266204

    work page 2022

  29. [29]

    Aniruddh Sriram, Fangyuan Xu, Eunsol Choi, and Greg Durrett. 2024. Contrastive learning to improve retrieval for real-world fact checking. CoRR, abs/2410.04657

    Pith/arXiv arXiv 2024

  30. [30]

    Yandong Sun, Yixuan Tang, and Anthony K. H. Tung. 2023. Qalinkplus: Text enrichment with QA data. IEEE Data Eng. Bull. , 47(4):115--128

    work page 2023

  31. [31]

    Crowd Intelligence for Early Misinformation Prediction on Social Media

    Megha Sundriyal, Harshit Choudhary, Tanmoy Chakraborty, and Md. Shad Akhtar. 2024. Crowd intelligence for early misinformation prediction on social media. CoRR, abs/2408.04463

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  32. [32]

    Yixuan Tang, Weilong Huang, Qi Liu, Anthony K. H. Tung, Xiaoli Wang, Jisong Yang, and Beibei Zhang. 2017. Qalink: Enriching text documents with relevant q & a site contents. In CIKM , pages 1359--1368. ACM

    work page 2017

  33. [33]

    Yixuan Tang, Jincheng Wang, and Anthony Kum Hoe Tung. 2025. The missing parts: Augmenting fact verification with half truth detection. In EMNLP . Association for Computational Linguistics

    work page 2025

  34. [34]

    Nandan Thakur, Nils Reimers, Johannes Daxenberger, and Iryna Gurevych. 2020. Augmented sbert: Data augmentation method for improving bi-encoders for pairwise sentence scoring tasks. arXiv e-prints, pages arXiv--2010

    work page 2020

  35. [35]

    Nandan Thakur, Nils Reimers, Andreas R \"u ckl \'e , Abhishek Srivastava, and Iryna Gurevych. 2021. BEIR: a heterogeneous benchmark for zero-shot evaluation of information retrieval models. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks

    work page 2021

  36. [36]

    James Thorne, Andreas Vlachos, Christos Christodoulopoulos, and Arpit Mittal. 2018. FEVER: a large-scale dataset for fact extraction and verification. In NAACL-HLT , pages 809--819

    work page 2018

  37. [37]

    Zhengjia Wang, Danding Wang, Qiang Sheng, Juan Cao, Silong Su, Yifan Sun, Beizhe Hu, and Siyuan Ma. 2023. Understanding news creation intents: Frame, dataset, and method. CoRR, abs/2312.16490

    Pith/arXiv arXiv 2023

  38. [38]

    Long Xia, Jun Xu, Yanyan Lan, Jiafeng Guo, and Xueqi Cheng. 2016. Modeling document novelty with neural tensor network for search result diversification. In SIGIR , pages 395--404. ACM

    work page 2016

  39. [39]

    Jun Yu, Sunil Mohan, Duangmanee Putthividhya, and Weng - Keen Wong. 2014. Latent dirichlet allocation based diversified retrieval for e-commerce search. In WSDM , pages 463--472. ACM

    work page 2014

  40. [40]

    Xiaoying Zhang, Hongning Wang, and Hang Li. 2023. Disentangled representation for diversified recommendations. In WSDM , pages 490--498. ACM

    work page 2023

  41. [41]

    Fabbri, Junru Liu, Ryo Kamoi, Xiaoxin Lu, Caiming Xiong, Jieyu Zhao, Dragomir Radev, Kathleen R

    Yusen Zhang, Nan Zhang, Yixin Liu, Alexander R. Fabbri, Junru Liu, Ryo Kamoi, Xiaoxin Lu, Caiming Xiong, Jieyu Zhao, Dragomir Radev, Kathleen R. McKeown, and Rui Zhang. 2024. Fair abstractive summarization of diverse perspectives. In NAACL-HLT, pages 3404--3426

    work page 2024

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.