pith. sign in

arxiv: 2605.26755 · v2 · pith:BSLQDZLKnew · submitted 2026-05-26 · 💻 cs.CL

SEEK: Semantic Evidence Extraction via Adaptive ChunKing for Multilingual Fact-Checking

Babu Kumar , Gaurav Kumar , Ayush Garg , Aditya Kishore , Jasabanta Patro This is my paper

Pith reviewed 2026-06-29 17:45 UTC · model grok-4.3

classification 💻 cs.CL
keywords multilingual fact-checkingsemantic evidence extractionadaptive chunkingveracity predictionevidence completenessLoRA adapterX-FACTRU22Fact
0
0 comments X

The pith

SEEK extracts coherent evidence chunks from full articles by identifying semantic topic transitions to support more reliable multilingual fact verification.

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

The paper proposes SEEK as a framework for building coherent evidence chunks from complete fact-checking articles rather than relying on search snippets or isolated sentences. It does this by detecting semantic topic transitions to keep local verification context intact. These chunks are encoded with a multilingual model and used to fine-tune LLMs via LoRA adapters for predicting claim veracity. Experiments on the X-FACT and RU22Fact datasets show gains in macro-f1 over multiple baselines, indicating that complete context helps avoid fragmented judgments. A reader would care because real-world fact-checking often involves long multilingual sources where missing context leads to errors.

Core claim

SEEK constructs coherent evidence chunks from full fact-checking articles by identifying semantic topic transitions and preserving local verification context. The constructed chunks are encoded using a multilingual encoder and then multilingual LLMs are finetuned using LoRA adapter for veracity prediction. Experiments on X-FACT and RU22Fact show that SEEK improves macro-f1 by up to 10% over semantic chunking, 19% over sentence chunking, and 20% over search-snippet baselines. Evidence completeness and significance analyses further show that SEEK preserves richer verification context and enables more reliable multilingual fact-checking.

What carries the argument

Adaptive chunKing framework that identifies semantic topic transitions to construct coherent evidence chunks from full articles while preserving local verification context.

If this is right

  • Coherent chunks from full articles preserve richer verification context than fragmented sentence or snippet methods.
  • Multilingual LLMs fine-tuned on these chunks achieve higher macro-f1 scores for veracity prediction.
  • Evidence completeness analyses confirm that the chunks support more reliable fact-checking across languages.
  • The method reduces dependence on incomplete external search results for evidence gathering.

Where Pith is reading between the lines

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

  • The same transition-based chunking could be tested on longer documents in other domains that require full-context reasoning.
  • If the gains hold, systems might shift from snippet retrieval to processing entire source articles as a default step.

Load-bearing premise

Identifying semantic topic transitions constructs coherent evidence chunks from full articles that preserve richer verification context than sentence-level or snippet-based approaches.

What would settle it

If SEEK-produced chunks yield no improvement or lower macro-f1 scores than sentence-level or search-snippet evidence on the same X-FACT and RU22Fact test sets, the benefit of adaptive semantic chunking would be refuted.

Figures

Figures reproduced from arXiv: 2605.26755 by Aditya Kishore, Ayush Garg, Babu Kumar, Gaurav Kumar, Jasabanta Patro.

Figure 1
Figure 1. Figure 1: High-level illustration of the proposed multi [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the SEEK and retrieval-based veracity prediction pipeline. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 5
Figure 5. Figure 5: Evidence completeness analysis on RU22FACT using an LLM-based evaluator. 7 Conclusion In this work, we proposed SEEK, a semantic evi￾dence extraction framework for multilingual fact￾checking. SEEK moves beyond short snippets, iso￾lated sentences, and fixed-size chunks by building coherent evidence passages from full web docu￾ments. It detects topic shifts and keeps boundary context, allowing the retrieved … view at source ↗
Figure 4
Figure 4. Figure 4: Evidence completeness analysis on X-FACT [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Example of topic-shift boundary detection [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Example chunks produced by the proposed Semantic Evidence Extraction with Adaptive chunKing(SEEK) [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Example chunks produced by the semantic chunking method on a Hindi fact-checking document. The [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Example chunks produced by the sentence chunking method on a Hindi fact-checking document. The [PITH_FULL_IMAGE:figures/full_fig_p014_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Illustration of the chunking strategies used for evidence retrieval. Sentence chunking groups consecutive [PITH_FULL_IMAGE:figures/full_fig_p016_11.png] view at source ↗
read the original abstract

Multilingual fact verification requires evidence that is both relevant and sufficiently complete for reliable factuality prediction. However, existing systems often rely on search snippets, sentence-level evidence, or locally segmented passages, which can miss decisive context and produce fragmented evidence. To overcome these limitations, we propose SEEK, a Semantic Evidence Extraction with an adaptive chunKing framework that constructs coherent evidence chunks from full fact-checking articles by identifying semantic topic transitions and preserving local verification context. The constructed chunks are encoded using a multilingual encoder and then multilingual LLMs are finetuned using LoRA adapter for veracity prediction. Experiments on X-FACT and RU22Fact show that SEEK improves macro-f1 by up to 10% over semantic chunking, 19% over sentence chunking, and 20% over search-snippet baselines. Evidence completeness and significance analyses further show that SEEK preserves richer verification context and enables more reliable multilingual fact-checking.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes SEEK, a Semantic Evidence Extraction framework using adaptive chunking to identify semantic topic transitions in full fact-checking articles, thereby constructing coherent evidence chunks that preserve richer verification context. These chunks are encoded via a multilingual encoder and used to fine-tune multilingual LLMs with LoRA adapters for veracity prediction. Experiments on the X-FACT and RU22Fact datasets report macro-F1 improvements of up to 10% over semantic chunking, 19% over sentence chunking, and 20% over search-snippet baselines, accompanied by analyses of evidence completeness and significance.

Significance. If the empirical gains hold under scrutiny, the method could meaningfully advance multilingual fact-checking by mitigating fragmentation in evidence sources. The focus on evidence completeness analyses represents a constructive addition to the literature on context-preserving retrieval for verification tasks.

major comments (3)
  1. [§3] §3 (Methodology): The description of semantic topic transition detection for adaptive chunking remains high-level with no explicit algorithm, similarity metric, or threshold parameters provided, which directly affects the reproducibility of the central claim that these chunks preserve richer verification context than baselines.
  2. [§4] §4 (Experiments): No dataset statistics (e.g., instance counts, language distributions, or class balance) are reported for X-FACT or RU22Fact, making it impossible to assess whether the reported macro-F1 gains of 10-20% are driven by the method or by dataset characteristics.
  3. [§4] §4 (Experiments): The paper provides no error analysis or breakdown of cases where SEEK fails to improve over baselines, which is load-bearing for validating the assumption that semantic topic transitions yield more reliable verification context.
minor comments (2)
  1. [Title] The title contains an inconsistent capitalization ('ChunKing') that should be standardized to 'Chunking'.
  2. [Abstract] Abstract and §1 would benefit from a brief statement of the number of languages covered in the multilingual experiments.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments, which highlight important aspects for improving the clarity and rigor of our work on SEEK. We address each major comment below and commit to revisions that enhance reproducibility and analysis without altering the core claims.

read point-by-point responses

  1. Referee: [§3] §3 (Methodology): The description of semantic topic transition detection for adaptive chunking remains high-level with no explicit algorithm, similarity metric, or threshold parameters provided, which directly affects the reproducibility of the central claim that these chunks preserve richer verification context than baselines.

    Authors: We agree that the current description in §3 is high-level and limits reproducibility. In the revised manuscript, we will expand this section to include the explicit algorithm for detecting semantic topic transitions, the similarity metric employed (cosine similarity over multilingual embeddings), and the specific threshold parameters used for chunk boundaries. This addition will directly support the claim regarding richer verification context. revision: yes

  2. Referee: [§4] §4 (Experiments): No dataset statistics (e.g., instance counts, language distributions, or class balance) are reported for X-FACT or RU22Fact, making it impossible to assess whether the reported macro-F1 gains of 10-20% are driven by the method or by dataset characteristics.

    Authors: We concur that dataset statistics are necessary for contextualizing the results. We will add a dedicated subsection or table in §4 reporting instance counts, language distributions, and class balances for both X-FACT and RU22Fact to allow readers to evaluate the reported macro-F1 improvements appropriately. revision: yes

  3. Referee: [§4] §4 (Experiments): The paper provides no error analysis or breakdown of cases where SEEK fails to improve over baselines, which is load-bearing for validating the assumption that semantic topic transitions yield more reliable verification context.

    Authors: We acknowledge the importance of error analysis for validating the method's assumptions. In the revision, we will incorporate an error analysis subsection in §4 that examines and categorizes cases where SEEK does not outperform the baselines, including qualitative examples and quantitative breakdowns to assess when semantic topic transitions provide reliable context. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper proposes an empirical pipeline for multilingual fact-checking: adaptive chunking via semantic topic transitions to form evidence chunks, followed by multilingual encoding and LoRA fine-tuning of LLMs for veracity prediction. Claims rest on experimental macro-F1 gains versus external baselines (semantic chunking, sentence chunking, search snippets) on X-FACT and RU22Fact. No equations, derivations, fitted parameters renamed as predictions, or load-bearing self-citations appear in the abstract or described method; the central result is an observable performance comparison on independent datasets and is therefore self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No details available from abstract to populate free parameters, axioms, or invented entities.

pith-pipeline@v0.9.1-grok · 5697 in / 916 out tokens · 36085 ms · 2026-06-29T17:45:15.852623+00:00 · methodology

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 · 20 canonical work pages · 4 internal anchors

  1. [1]

    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.all mid.sentence after.sentence after.block STRINGS urlintro eprinturl eprintpr...

  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]

    Isabelle Augenstein, Christina Lioma, Dongsheng Wang, Lucas Chaves Lima, Casper Hansen, Christian Hansen, and Jakob Grue Simonsen. 2019. https://doi.org/10.18653/v1/D19-1475 M ulti FC : A real-world multi-domain dataset for evidence-based fact checking of claims . In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing an...

    work page doi:10.18653/v1/d19-1475 2019

  4. [4]

    Recep Firat Cekinel, Pinar Karagoz, and C a g r C \"o ltekin. 2024. https://aclanthology.org/2024.lrec-main.368/ Cross-lingual learning vs. low-resource fine-tuning: A case study with fact-checking in T urkish . In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pag...

    2024

  5. [5]

    Jiangui Chen, Ruqing Zhang, Jiafeng Guo, Yixing Fan, and Xueqi Cheng. 2022. https://doi.org/10.1145/3477495.3531827 GERE : Generative evidence retrieval for fact verification . In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '22, pages 2184--2189, New York, NY, USA. Association for ...

    work page doi:10.1145/3477495.3531827 2022

  6. [6]

    Matthijs Douze, Alexandr Guzhva, Chengqi Deng, Jeff Johnson, Gergely Szilvasy, Pierre-Emmanuel Mazaré, Maria Lomeli, Lucas Hosseini, and Hervé Jégou. 2024. http://arxiv.org/abs/2401.08281 The Faiss library . arXiv preprint

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  7. [7]

    Zhijiang Guo, Michael Schlichtkrull, and Andreas Vlachos. 2022. https://doi.org/10.1162/tacl_a_00454 A survey on automated fact-checking . Transactions of the Association for Computational Linguistics, 10:178--206

    work page doi:10.1162/tacl_a_00454 2022

  8. [8]

    Ashim Gupta and Vivek Srikumar. 2021 a . https://doi.org/10.18653/v1/2021.acl-short.86 X -fact: A new benchmark dataset for multilingual fact checking . In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 675--6...

    work page doi:10.18653/v1/2021.acl-short.86 2021

  9. [9]

    Ashim Gupta and Vivek Srikumar. 2021 b . https://aclanthology.org/2021.acl-short.86 X-fact: A new benchmark dataset for multilingual fact checking . In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 675--682, ...

    2021

  10. [10]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. https://www.microsoft.com/en-us/research/publication/lora-low-rank-adaptation-of-large-language-models/ Lora: Low-rank adaptation of large language models . In ICLR 2022

    2022

  11. [11]

    Kung-Hsiang Huang, ChengXiang Zhai, and Heng Ji. 2022. https://aclanthology.org/2022.coling-1.86/ CONCRETE : Improving cross-lingual fact-checking with cross-lingual retrieval . In Proceedings of the 29th International Conference on Computational Linguistics, pages 1024--1035, Gyeongju, Republic of Korea. International Committee on Computational Linguistics

    2022

  12. [12]

    Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.550 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, Online. Ass...

    work page doi:10.18653/v1/2020.emnlp-main.550 2020

  13. [13]

    C. Kiss. 2025. https://doi.org/10.1007/s10791-025-09638-7 Max--min semantic chunking of documents for rag application . Discover Computing, 28(117)

    work page doi:10.1007/s10791-025-09638-7 2025

  14. [14]

    LangChain. 2024. Langchain semantic chunking. https://python.langchain.com/docs/how_to/semantic_chunker/. Accessed: 2024-07-04

    2024

  15. [15]

    Preslav Nakov, David Corney, Maram Hasanain, Firoj Alam, Tamer Elsayed, Alberto Barrón-Cedeño, Paolo Papotti, Shaden Shaar, and Giovanni Da San Martino. 2021. https://doi.org/10.24963/ijcai.2021/619 Automated fact-checking for assisting human fact-checkers . In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-2...

    work page doi:10.24963/ijcai.2021/619 2021

  16. [16]

    Venktesh, Erik Martin, Henrik Vatndal, Vinay Setty, and Avishek Anand

    Kevin Nanhekhan, V. Venktesh, Erik Martin, Henrik Vatndal, Vinay Setty, and Avishek Anand. 2025. https://doi.org/10.1007/978-3-031-88717-8_28 Flashcheck: Exploration of efficient evidence retrieval for fast fact-checking . In Advances in Information Retrieval, volume 15575 of Lecture Notes in Computer Science, pages 385--399, Cham. Springer

    work page doi:10.1007/978-3-031-88717-8_28 2025

  17. [17]

    Rrubaa Panchendrarajan and Arkaitz Zubiaga. 2024. https://doi.org/10.1016/j.nlp.2024.100066 Claim detection for automated fact-checking: A survey on monolingual, multilingual and cross-lingual research . Natural Language Processing Journal, 7:100066

    work page doi:10.1016/j.nlp.2024.100066 2024

  18. [18]

    Peng et al

    Q. Peng et al. 2025 a . https://arxiv.org/abs/2505.10740 Semeval-2025 task 7: Multilingual and crosslingual fact-checked claim retrieval . In Proceedings of the 19th International Workshop on Semantic Evaluation (SemEval-2025)

    work page arXiv 2025

  19. [19]

    Qiwei Peng, Robert Moro, Michal Gregor, Ivan Srba, Simon Ostermann, Marian Simko, Juraj Podrouzek, Mat \'u s Mesar c \'i k, Jaroslav Kop c an, and Anders S gaard. 2025 b . https://aclanthology.org/2025.semeval-1.323/ S em E val-2025 task 7: Multilingual and crosslingual fact-checked claim retrieval . In Proceedings of the 19th International Workshop on Se...

    2025

  20. [20]

    Qu et al

    R. Qu et al. 2025. https://aclanthology.org/2025.findings-naacl.114/ Is semantic chunking worth the computational cost? Findings of the Association for Computational Linguistics: NAACL 2025

    2025

  21. [21]

    Dorian Quelle, Calvin Yixiang Cheng, Alexandre Bovet, and Scott A. Hale. 2025. https://doi.org/10.1140/epjds/s13688-025-00520-6 Lost in translation: using global fact-checks to measure multilingual misinformation prevalence, spread, and evolution . EPJ Data Science, 14(1):22

    work page doi:10.1140/epjds/s13688-025-00520-6 2025

  22. [22]

    Michael Schlichtkrull, Zhijiang Guo, and Andreas Vlachos. 2023. https://proceedings.neurips.cc/paper_files/paper/2023/file/cd86a30526cd1aff61d6f89f107634e4-Paper-Datasets_and_Benchmarks.pdf Averitec: A dataset for real-world claim verification with evidence from the web . In Advances in Neural Information Processing Systems, volume 36, pages 65128--65167....

    2023

  23. [23]

    James Thorne, Andreas Vlachos, Christos Christodoulopoulos, and Arpit Mittal. 2018. http://arxiv.org/abs/1803.05355 FEVER : a large-scale dataset for fact extraction and verification . Updated version of NAACL2018 paper. https://doi.org/10.48550/arXiv.1803.05355

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1803.05355 2018

  24. [24]

    Sentence Transformers. 2021. sentence-transformers/all-minilm-l6-v2. https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2. Accessed: 2026-05-17

    2021

  25. [25]

    UncleCode. 2024. Crawl4ai: Open-source llm friendly web crawler & scraper. https://github.com/unclecode/crawl4ai. GitHub repository. Please use the commit hash you're working with

    2024

  26. [26]

    Viswanathan et al

    V. Viswanathan et al. 2025. https://arxiv.org/abs/2509.11492 Claimiq at checkthat! 2025: Comparing prompted and fine-tuned language models for verifying numerical claims

    work page arXiv 2025

  27. [27]

    Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2024. Multilingual e5 text embeddings: A technical report. arXiv preprint arXiv:2402.05672

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  28. [28]

    liar, liar pants on fire

    William Y. Wang. 2017. “liar, liar pants on fire”: A new benchmark dataset for fake news detection. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (ACL), pages 422--426. Association for Computational Linguistics

    2017

  29. [29]

    Yirong Zeng, Xiao Ding, Yi Zhao, Xiangyu Li, Jie Zhang, Chao Yao, Ting Liu, and Bing Qin. 2024. https://arxiv.org/abs/2403.16662 Ru22fact: Optimizing evidence for multilingual explainable fact-checking on russia-ukraine conflict . arXiv preprint arXiv:2403.16662

    work page arXiv 2024

  30. [30]

    Hengran Zhang, Ruqing Zhang, Jiafeng Guo, Maarten de Rijke, Yixing Fan, and Xueqi Cheng. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.422 From relevance to utility: Evidence retrieval with feedback for fact verification . In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 6373--6384, Singapore. Association for Computa...

    work page doi:10.18653/v1/2023.findings-emnlp.422 2023

  31. [31]

    Shengyu Zhang, Linfeng Dong, Xiaoya Li, Sen Zhang, Xiaofei Sun, Shuhe Wang, Jiwei Li, Runyi Hu, Tianwei Zhang, Fei Wu, and Guoyin Wang. 2025. http://arxiv.org/abs/2308.10792 Instruction tuning for large language models: A survey

    work page arXiv 2025

  32. [32]

    Liwen Zheng, Chaozhuo Li, Xi Zhang, Yu-Ming Shang, Feiran Huang, and Haoran Jia. 2024 a . https://doi.org/10.18653/v1/2024.findings-acl.551 Evidence retrieval is almost all you need for fact verification . In Findings of the Association for Computational Linguistics: ACL 2024, pages 9274--9281, Bangkok, Thailand. Association for Computational Linguistics

    work page doi:10.18653/v1/2024.findings-acl.551 2024

  33. [33]

    Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. 2024 b . http://arxiv.org/abs/2403.13372 Llamafactory: Unified efficient fine-tuning of 100+ language models . In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), Bangkok, Thailand. As...

    work page internal anchor Pith review Pith/arXiv arXiv 2024