Pith. sign in

REVIEW 3 major objections 6 minor 33 references

This paper shows that a fully zero-shot pipeline combining lexical and semantic retrieval with a large-language-model pointwise reranker can retrieve the scientific source of a social media post at 64.39% MRR@5, with the LLM supplying most

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 ↗

T0 review · deepseek-v4-flash

2026-08-02 08:26 UTC pith:YSDDJK7G

load-bearing objection Solid shared-task ablation paper: zero-shot Qwen2.5-14B reranking adds ~10 MRR@5 points to a BM25+E5 RRF first stage on CheckThat! 2026 English, but the reranker comparison is confounded by candidate depth and the gold-label ceiling is unexamined. the 3 major comments →

arxiv 2607.24803 v1 pith:YSDDJK7G submitted 2026-07-06 cs.IR cs.CL

SciClaimSeekers at CheckThat! 2026: Retrieving Scientific Sources for Social Media Claims with LLM Reranking

Mohotarema Rashid , Nansu Baniya , Anirban Saha Anik , Xiaoying Song , Lingzi Hong This is my paper
classification cs.IR cs.CL
keywords scientific claim verificationevidence retrievalhybrid retrievalreciprocal rank fusionLLM rerankingpointwise cross-encodersocial mediaMRR@5
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 paper asks whether a retrieval system built entirely from off-the-shelf components can find the scientific paper behind a social media post, without any task-specific training. It reports that combining lexical retrieval, multilingual semantic retrieval, and a large instruction-tuned language model used as a pointwise reranker reaches about 64% MRR@5 on English posts, a 13.67-point improvement over the lexical baseline. The key claim is that the language-model reranker is the dominant component: it contributes roughly three-quarters of the total gain, about +10 MRR@5 points, whether it sits on top of the lexical ranking or the fused hybrid ranking. A reader should care because the result suggests that source attribution for scientific claims on social media can be handled by a zero-shot reranking pipeline rather than requiring a fine-tuned system.

Core claim

The central claim is that a 14-billion-parameter instruction-tuned language model, prompted to give each candidate paper a 0-10 relevance score for the post, is responsible for most of the system's accuracy. On the English test set the complete pipeline reaches 64.39% MRR@5 (57.97% Hit@1, 73.55% Hit@5), versus 50.21% for lexical retrieval and 55.77% for the fused hybrid first stage. The paper isolates the LLM reranker's contribution at roughly +10 MRR@5 points over either first-stage option, about three-quarters of the full +13.67-point gain over BM25. It also claims the reranker makes the first-stage fusion matter more: hybrid candidates plus LLM outperform BM25-plus-LLM by about 3.6 points

What carries the argument

The engine is a three-stage retrieve-then-rerank pipeline. Stage one runs a lexical sparse retriever (BM25) and a multilingual dense retriever in parallel, each returning 100 candidates per post. Stage two merges the two lists with Reciprocal Rank Fusion using k=60, producing a top-50 candidate set. Stage three scores each of those 50 (post, paper) pairs with a 14-billion-parameter instruction-tuned LLM used as a zero-shot pointwise cross-encoder, asking for an integer 0-10 relevance judgment with a specific instruction prompt; candidates are sorted by LLM score and ties broken by the fusion score. The mechanism that carries the argument is the LLM's scalar relevance judgment: it converts a

Load-bearing premise

The load-bearing premise is that the gold labels identify the correct source paper for every post and that this paper always sits inside the fixed 10,000-paper candidate pool; the reported scores count any absence from the top five as a miss, so label noise or an incomplete pool would change every number.

What would settle it

Take a random sample of posts where the system failed to place the correct paper in the top five and have independent annotators check whether the answer key is right and whether the true paper is even in the 10,000-paper candidate pool; if many failures are actually wrong or missing labels, the 64.39% number is not a fair measure of retrieval quality.

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

If this is right

  • A zero-shot LLM reranker can substitute for a fine-tuned reranker in scientific source attribution, supplying roughly three-quarters of the total observed gain without task-specific training.
  • The retrieval stage need not be perfect: the LLM applied to BM25 candidates alone reaches about 60.6% MRR@5 on the test set, and fusing in dense retrieval adds a further ~3.6 points after reranking.
  • First-stage recall compounds downstream: the hybrid fusion gain grows from about +1.5 points before reranking to about +3.6 points after reranking.
  • The design decisions transfer cleanly: development and test MRR@5 differ by only 0.03 points, so the tuning choices are not overfit to the development split.

Where Pith is reading between the lines

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

  • Editorial inference: the first-stage dense retriever and the instruction-tuned reranker are both multilingual, so the same zero-shot pipeline could be run on the German and French splits the paper leaves untested; those results would show whether the +10-point reranker gain is language-dependent.
  • Editorial inference: the 11-hour runtime on two GPU nodes is dominated by LLM prompt processing, so a smaller model distilled to imitate the 14B reranker's scores could retain much of the gain at a fraction of the cost—the paper names distillation as future work, but the expected trade-off is our inference.
  • Editorial inference: because reported scores depend on ground-truth labels and a fixed 10,000-paper pool, a human audit of failure cases is the natural way to tell whether misses are retrieval errors or label errors; if label errors are common, the true ceiling may be higher than 64% MRR@5.

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

3 major / 6 minor

Summary. The paper describes SciClaimSeekers, a three-stage retrieve-then-rerank system for the CLEF-2026 CheckThat! Task 1 on retrieving scientific sources for implicit social media claims. The pipeline combines Okapi BM25 and multilingual E5 dense retrieval, fuses the top-100 lists with Reciprocal Rank Fusion (k=60), and reranks the top-50 fused candidates with a zero-shot pointwise Qwen2.5-14B-Instruct cross-encoder. The authors report 64.39% MRR@5 on the official English test set, a +13.67-point improvement over the BM25 baseline, with the LLM reranker contributing roughly +10 points. They also present an ablation isolating each stage and claim that the LLM is the dominant source of gain, with a negligible 0.03-point dev-test gap.

Significance. The paper's contribution is a reproducible, open-source hybrid retrieval and LLM reranking pipeline that performs competitively in a shared task using only zero-shot models, without task-specific fine-tuning. The reported results come from the official CLEF evaluation, and the ablation is clearly structured, with a small dev-test gap suggesting that tuning decisions transfer. If the central claims hold, the work provides a practical blueprint for privacy-sensitive or offline deployment of LLM rerankers. However, the strength of the evidence is limited by an uncontrolled comparison between rerankers and the absence of any analysis of the gold-label ceiling or pool coverage, which makes the absolute MRR numbers and the attribution of gains less certain than the text implies.

major comments (3)
  1. [Section 4, Table 2] The comparison between rerankers is confounded by candidate depth. System 4 (bge-reranker-v2-m3) reranks the top-30 fused candidates, while System 6 (Qwen2.5-14B) reranks the top-50. The text attributes the large MRR difference (64.36% vs. 55.12% on dev) to the choice of reranker, but the candidate sets differ. To support the claim that the LLM is the dominant source of gain, the authors should run a controlled comparison at the same depth (e.g., both on top-30 or both on top-50) or explicitly acknowledge and analyze the confound.
  2. [Section 4, Eq. (2) and Table 2] All MRR/Hit scores are computed against the official gold pubkey labels, but the paper does not report oracle recall@50 or any pool-coverage analysis. If the true source is missing from the 10,000-paper candidate pool, or if gold labels are noisy, every number in Table 2 is biased and the measured LLM gain partially reflects first-stage ceiling rather than reranking ability. I request that the authors report oracle recall@50 for dev and test (fraction of queries whose gold paper appears in the hybrid RRF top-50) and, if possible, a basic label-quality audit or at least a reference to the task's annotation protocol. This is necessary to calibrate the +13.67-point improvement claim.
  3. [Section 4] The paper reports single-split MRR values without any statistical significance testing or confidence intervals. Given that several differences are small (e.g., BM25 vs. E5 dev MRR improvement of ~2 points), the central claim that the LLM contributes +10 MRR@5 points should be accompanied by a paired bootstrap or permutation test across the 3,905 dev queries. The test set is a single official split, so significance evidence from the dev set would materially strengthen the robustness of the findings.
minor comments (6)
  1. [Section 4, Eq. (2)] The equation appears mis-typeset: it reads "MRR = 1 |𝑄| |𝑄|∑ 𝑖=1 1/rank𝑖" and should be MRR = (1/|Q|) Σ_{i=1}^{|Q|} 1/rank_i.
  2. [Section 3.5] The sentence "We use k=60 recommended by prior research [28]. We chose this depth based on the findings from the development set" is ambiguous. Clarify whether "this depth" refers to the RRF constant k or the fusion depth (top-50), and state which values were explored on the dev set.
  3. [Table 2] The table would be clearer if the candidate depth for each reranker were included (e.g., "top-30" and "top-50" in the system names or a separate column).
  4. [Section 2] Typo: "our study is situated in intersection four major research streams" should be "at the intersection of four major research streams."
  5. [Section 6] The Limitations section mentions only the English-only evaluation and the runtime bottleneck. It would be appropriate to also note the absence of a pool-coverage or label-quality analysis as a limitation, as this affects the interpretation of the reported scores.
  6. [Abstract] "10.17 points over the unranked hybrid" is slightly misleading; the hybrid is ranked by RRF. "un-reranked hybrid" would be more precise.

Circularity Check

0 steps flagged

No circularity: all reported MRR numbers are measured against external official gold labels; no pipeline output is derived from its own inputs by construction.

full rationale

The paper's central quantitative claims—64.39% test MRR@5, +13.67 over BM25, and +10.17 for the LLM reranker—are all computed by applying Eq. (2) to system outputs and comparing them against the official CheckThat! 2026 gold pubkey labels, which are external to the pipeline. No parameter is fitted to the target metric and then reported as a prediction: the only dev-set tuning disclosed is the RRF k/fusion depth, and the resulting pipeline is evaluated on held-out test data (dev 64.36% vs. test 64.39%). The ablation attribution of the gain to the Qwen2.5 reranker is obtained by difference between independently evaluated system variants (BM25+LLM 60.73 vs. BM25 50.69; Hybrid+LLM 64.36 vs. Hybrid 54.19), not by construction. Citations in the paper are to prior work on standard components (BM25, E5, RRF, Qwen2.5) and to prior shared-task systems, but none is a self-citation and none is used to define the measured improvement. The concern that gold labels might be incomplete or that the 10,000-paper pool might miss true sources is a data-validity risk, not a circularity risk, because the labels are external rather than outputs of the proposed system. Therefore no circular step is present.

Axiom & Free-Parameter Ledger

4 free parameters · 3 axioms · 0 invented entities

The paper's results rest on standard pretrained models, the official task data, and a few hand-set hyperparameters. No new entities are introduced. The main reader-supplied inputs are the correctness of CLEF gold labels and zero-shot transfer of E5/Qwen; these are reasonable but unverified here.

free parameters (4)
  • RRF smoothing constant k = 60
    Eq. (1); Section 3.5 says k=60 is 'recommended by prior research' and also chosen based on development set findings. It is a hand-set hyperparameter.
  • First-stage retrieval depth = 100 per retriever
    Section 3.1: each of BM25 and dense retrieval returns top-100 candidates; no sensitivity analysis is reported.
  • Fusion depth = 50
    Section 3.5: fused top-50 candidates are passed to the LLM reranker; the text says the depth was chosen based on development-set findings.
  • Reranker candidate depth for cross-encoder comparison = 30 for bge-reranker vs 50 for Qwen
    Table 2 rows 4-6: the cross-encoder is evaluated on top-30 candidates while the LLM reranker is evaluated on top-50, confounding the comparison of reranker effectiveness.
axioms (3)
  • domain assumption CLEF-2026 Task 1 gold pubkey labels are correct and complete.
    All MRR/Hit numbers in Table 2 assume the official gold labels define ground truth (Section 4, Eq. 2). No label-quality analysis is provided.
  • domain assumption Zero-shot transfer of multilingual-E5-large-instruct and Qwen2.5-14B-Instruct to this retrieval task is adequate.
    No task-specific fine-tuning is used; the pipeline's effectiveness rests on pretrained-model generalization (Sections 3.4 and 3.6).
  • domain assumption The 10,000-paper fixed candidate pool contains the gold source for each post.
    Section 3.1 defines |D|=10,000 as the fixed pool; if a true source is absent, MRR@5 necessarily penalizes all systems regardless of retrieval quality.

pith-pipeline@v1.3.0-alltime-deepseek · 8135 in / 13100 out tokens · 136237 ms · 2026-08-02T08:26:59.707036+00:00 · methodology

0 comments
read the original abstract

Scientific claims often spread on social media faster than they can be verified, while posts rarely link to the original scholarly sources. To tackle this problem this paper presents system called SciClaimSeekers, a retrieval and reranking framework by combining BM25 and zero-shot multilingual E5 retrieval with Reciprocal Rank Fusion (k=60), followed by Qwen2.5-14B-Instruct pointwise reranking. The pipeline reaches 64.36% MRR@5 on the English development set a 13.67-point jump over BM25 and 10.17 points over the unranked hybrid and 64.39% on the official test set, in the CLEF-2026 CheckThat! Task 1 evaluation. Our experiment suggests that large pre-trained models, when combined into a careful pipeline, can be competitive with fine-tuned approaches on this task.

Figures

Figures reproduced from arXiv: 2607.24803 by Anirban Saha Anik, Lingzi Hong, Mohotarema Rashid, Nansu Baniya, Xiaoying Song.

Figure 1
Figure 1. Figure 1: SCiClaimSeekers System Architecture 2. Related Work Our study is situated in intersection four major research streams I) automated fact checking II) scientific claim verification III) social media based scientific discourse retrieval and IV) LLM based reranking systems. Early factchecking pipeline was shaped by FEVER benchmark [10] while SciFact extended the framework dedicated to scientific claims [11]. L… view at source ↗
Figure 1
Figure 1. Figure 1: After using the tool, the authors reviewed and edited the content as needed and take full [PITH_FULL_IMAGE:figures/full_fig_p006_1.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

33 extracted references · 5 canonical work pages · 2 internal anchors

  1. [1]

    Muhammed T, S

    S. Muhammed T, S. K. Mathew, The disaster of misinformation: A review of research in social media, International Journal of Data Science and Analytics 13 (2022) 271–285. doi: 10.1007/ s41060-022-00311-6

    2022

  2. [2]

    P. J. Sager, A. Kamaraj, B. F. Grewe, T. Stadelmann, Deep retrieval at CheckThat! 2025: Identifying scientific papers from implicit social media mentions via hybrid retrieval and re-ranking, 2025. doi:10.48550/arXiv.2505.23250.arXiv:2505.23250

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2505.23250 2025

  3. [3]

    Altay, A

    S. Altay, A. Acerbi, People believe misinformation is a threat because they assume others are gullible, New Media & Society 26 (2024) 6440–6461. doi:10.1177/14614448231153379

    work page doi:10.1177/14614448231153379 2024

  4. [4]

    Moravec, R

    P. Moravec, R. Minas, A. R. Dennis, Fake news on social media: People believe what they want to believe when it makes no sense at all, SSRN Scholarly Paper No. 3269541, 2018. doi:10.2139/ ssrn.3269541

    2018

  5. [5]

    Zhang, F

    S. Zhang, F. Ma, Y. Liu, W. Pian, Identifying features of health misinformation on social media sites: An exploratory analysis, Library Hi Tech 40 (2021) 1384–1401. doi: 10.1108/ LHT-09-2020-0242

    2021

  6. [6]

    Booth, A

    L. Booth, A. Aldaihani, J. Davidson, C. Wilson, C. Lawlor, P. Hong, M. E. Graham, Misinformation and readability of social media content on pediatric ankyloglossia and other oral ties, JAMA Otolaryngology–Head & Neck Surgery 151 (2025) 143–150. doi:10.1001/jamaoto.2024.4211

    arXiv 2025

  7. [7]

    J. M. Struß, S. Schellhammer, S. Dietze, V. V., V. Setty, T. Chakraborty, P. Nakov, A. Anand, P. Chungkham, S. Hafid, D. Sahnan, K. Todorov, The clef-2026 checkthat! lab: Advancing multilingual fact-checking, in: R. Campos, A. Jatowt, Y. Lan, M. Aliannejadi, C. Bauer, S. MacAvaney, A. Anand, Z. Ren, S. Verberne, N. Bai, M. Mansoury (Eds.), Advances in Inf...

    2026

  8. [8]

    J. M. Struß, S. Schellhammer, S. Dietze, V. V., V. Setty, T. Chakraborty, P. Nakov, A. Anand, P. Chungkham, S. Hafid, D. Sahnan, K. Todorov, Overview of the CLEF-2026 CheckThat! Lab: Advancing multilingual fact-checking, in: M. Hagen, M. Potthast, B. Stein, P. Schaer, E. Zangerle, S. MacAvaney, J. M. Struß, E. S. Salido, A. Barrón-Cedeño, Alba García Seco...

    2026

  9. [9]

    Schellhammer, S

    S. Schellhammer, S. Hafid, Y. S. Kartal, K. Boland, D. Dimitrov, K. Todorov, S. Dietze, Overview of the CLEF-2026 CheckThat! lab task 1 on source retrieval for scientific web claims, CLEF 2026, Jena, Germany, 2026

    2026

  10. [10]

    Thorne, A

    J. Thorne, A. Vlachos, C. Christodoulopoulos, A. Mittal, FEVER: A large-scale dataset for fact extraction and VERification, in: Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), Association for Computational Linguistics, 2018, pp. 809–819....

    2018

  11. [11]

    Wadden, S

    D. Wadden, S. Lin, K. Lo, L. L. Wang, M. van Zuylen, A. Cohan, H. Hajishirzi, Fact or fiction: Verifying scientific claims, in: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), Association for Computational Linguistics, 2020, pp. 7534–7550. doi:10.18653/v1/2020.emnlp-main.609

    doi:10.18653/v1/2020.emnlp-main.609 2020

  12. [12]

    Pradeep, S

    R. Pradeep, S. Sharifymoghaddam, J. Lin, RankVicuna: Zero-shot listwise document rerank- ing with open-source large language models, 2023. doi: 10.48550/arXiv.2309.15088. arXiv:2309.15088

    Pith/arXiv arXiv doi:10.48550/arxiv.2309.15088 2023

  13. [13]

    A. Rana, D. Khanna, T. Ghosal, M. Singh, H. Singh, P. S. Rana, RerrFact: Reduced evidence re- trieval representations for scientific claim verification, 2022. doi:10.48550/arXiv.2202.02646. arXiv:2202.02646

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2202.02646 2022

  14. [14]

    Hafid, S

    S. Hafid, S. Schellhammer, S. Bringay, K. Todorov, S. Dietze, SciTweets—a dataset and annotation framework for detecting scientific online discourse, in: Proceedings of the 31st ACM International Conference on Information & Knowledge Management (CIKM ’22), 2022, pp. 3988–3992. doi:10. 1145/3511808.3557693

    arXiv 2022

  15. [15]

    Hafid, Y

    S. Hafid, Y. S. Kartal, S. Schellhammer, V. Jacot, S. Bringay, S. Dietze, K. Todorov, Disambiguation of implicit scientific references on x, in: Proceedings of the 36th ACM Conference on Hypertext and Social Media, 2025, pp. 165–170

    2025

  16. [16]

    Ashbaugh, L

    C. Ashbaugh, L. Baumgärtner, T. Gress, N. Sidorov, D. Werner, AIRwaves at CheckThat! 2025: Retrieving scientific sources for implicit claims on social media with dual encoders and neural re-ranking, 2025. doi:10.48550/arXiv.2509.19509.arXiv:2509.19509

    work page doi:10.48550/arxiv.2509.19509 2025

  17. [17]

    Lee, S.-w

    D. Lee, S.-w. Hwang, K. Lee, S. Choi, S. Park, On complementarity objectives for hybrid retrieval, in: A. Rogers, J. Boyd-Graber, N. Okazaki (Eds.), Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Association for Computational Linguistics, 2023, pp. 13357–13368. doi:10.18653/v1/2023.acl-long.746

    work page doi:10.18653/v1/2023.acl-long.746 2023

  18. [18]

    Y. Ren, Y. Cao, P. Guo, F. Fang, W. Ma, Z. Lin, Retrieve-and-sample: Document-level event argument extraction via hybrid retrieval augmentation, in: A. Rogers, J. Boyd-Graber, N. Okazaki (Eds.), Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Association for Computational Linguistics, 2023, ...

    2023

  19. [19]

    L. Yang, J. Hu, M. Qiu, C. Qu, J. Gao, W. B. Croft, X. Liu, Y. Shen, J. Liu, A hybrid retrieval- generation neural conversation model, in: Proceedings of the 28th ACM International Conference on Information and Knowledge Management (CIKM ’19), 2019, pp. 1341–1350. doi: 10.1145/ 3357384.3357881

    arXiv 2019

  20. [20]

    Robertson, H

    S. Robertson, H. Zaragoza, The probabilistic relevance framework: BM25 and beyond, Foundations and Trends in Information Retrieval 3 (2009) 333–389. doi:10.1561/1500000019

    doi:10.1561/1500000019 2009

  21. [21]

    Karpukhin, B

    V. Karpukhin, B. Oğuz, S. Min, P. Lewis, L. Wu, S. Edunov, D. Chen, W.-t. Yih, Dense passage re- trieval for open-domain question answering, in: B. Webber, T. Cohn, Y. He, Y. Liu (Eds.), Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), Associa- tion for Computational Linguistics, 2020, pp. 6769–6781. doi:10.1...

    doi:10.18653/v1/2020.emnlp-main 2020

  22. [22]

    Khattab, M

    O. Khattab, M. Zaharia, ColBERT: Efficient and effective passage search via contextualized late interaction over BERT, 2020. doi:10.48550/arXiv.2004.12832.arXiv:2004.12832

    Pith/arXiv arXiv doi:10.48550/arxiv.2004.12832 2020

  23. [23]

    Nogueira, Z

    R. Nogueira, Z. Jiang, R. Pradeep, J. Lin, Document ranking with a pretrained sequence-to-sequence model, in: T. Cohn, Y. He, Y. Liu (Eds.), Findings of the Association for Computational Linguistics: EMNLP 2020, Association for Computational Linguistics, 2020, pp. 708–718. doi:10.18653/v1/ 2020.findings-emnlp.63

    doi:10.18653/v1/ 2020

  24. [24]

    W. Sun, L. Yan, X. Ma, S. Wang, P. Ren, Z. Chen, D. Yin, Z. Ren, Is ChatGPT good at search? investigating large language models as re-ranking agents, 2024. doi: 10.48550/arXiv.2304. 09542.arXiv:2304.09542

    Pith/arXiv arXiv doi:10.48550/arxiv.2304.09542 2024

  25. [25]

    X. Ma, L. Wang, N. Yang, F. Wei, J. Lin, Fine-tuning LLaMA for multi-stage text retrieval, in: Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’24), 2024, pp. 2421–2425. doi:10.1145/3626772.3657951

    arXiv 2024

  26. [26]

    Zhuang, Z

    H. Zhuang, Z. Qin, K. Hui, J. Wu, L. Yan, X. Wang, M. Bendersky, Beyond yes and no: Improving zero-shot LLM rankers via scoring fine-grained relevance labels, in: K. Duh, H. Gómez, S. Bethard (Eds.), Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 2: Sho...

    doi:10.18653/v1/2024.naacl-short.31 2024

  27. [27]

    L. Wang, N. Yang, X. Huang, L. Yang, R. Majumder, F. Wei, Multilingual E5 text embeddings: A technical report, 2024. doi:10.48550/arXiv.2402.05672.arXiv:2402.05672

    Pith/arXiv arXiv doi:10.48550/arxiv.2402.05672 2024

  28. [28]

    G. V. Cormack, C. L. A. Clarke, S. Büttcher, Reciprocal rank fusion outperforms Condorcet and individual rank learning methods, in: Proceedings of the 32nd International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’09), 2009, pp. 758–759. doi:10.1145/1571941.1572114

    arXiv 2009

  29. [29]

    A. Yang, B. Yang, B. Hui, B. Zheng, B. Yu, C. Zhou, C. Li, C. Li, D. Liu, F. Huang, et al., Qwen2.5 technical report, 2024. doi:10.48550/arXiv.2412.15115.arXiv:2412.15115

    Pith/arXiv arXiv doi:10.48550/arxiv.2412.15115 2024

  30. [30]

    Thakur, N

    N. Thakur, N. Reimers, A. Rücklé, A. Srivastava, I. Gurevych, BEIR: A heterogeneous benchmark for zero-shot evaluation of information retrieval models, in: Proceedings of the Neural Information Pro- cessing Systems Track on Datasets and Benchmarks (NeurIPS 2021), 2021.arXiv:2104.08663

    Pith/arXiv arXiv 2021

  31. [31]

    Sennrich, B

    R. Sennrich, B. Haddow, A. Birch, Neural machine translation of rare words with subword units, in: Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Association for Computational Linguistics, 2016, pp. 1715–1725. doi:10. 18653/v1/P16-1162

    2016

  32. [32]

    Reimers, I

    N. Reimers, I. Gurevych, Sentence-BERT: Sentence embeddings using Siamese BERT-networks, in: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), Association for Computational Linguistics, 2019, pp. 3982–3992. doi:10.18653/v1/D19-1410

    doi:10.18653/v1/d19-1410 2019

  33. [33]

    Bruch, S

    S. Bruch, S. Gai, A. Ingber, An analysis of fusion functions for hybrid retrieval, ACM Transactions on Information Systems 42 (2023) 20:1–20:35. doi:10.1145/3596512

    doi:10.1145/3596512 2023