Pith. sign in

REVIEW 4 major objections 8 minor 31 references

Multi-granular Training Strategies for Robust Multi-hop Reasoning Over Noisy and Heterogeneous Knowledge Sources

T0 review · 4 major / 8 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read AMKOR, a generative framework that fuses parametric and retrieved knowledge with probabilistic beam reasoning, reports state-of-the-art token-level F1 on four multi-hop QA datasets, averaging a 2.5-point gain over five baselines.

desk verdict A generic pipeline with an unreproducible headline result; the stated 2.5% improvement doesn't match the paper's own Table 1. read the letter →

arxiv 2502.05944 v1 pith:FFIFUSJJ submitted 2025-02-09 cs.CL

Jackson Coleman , Isaiah Lawrence , Benjamin Turner This is my paper · ORCID
classification cs.CL
keywords multi-hopquestionansweringmulti-sourceknowledgefusionprobabilisticbeamreasoningmulti-granularlearninglargelanguagemodelsretrieval-augmentedgenerationnoiserobustness
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

The paper proposes AMKOR, a generative framework for multi-source multi-hop question answering that fuses parametric knowledge from a large language model with retrieved external snippets at each reasoning step. It claims that searching over multiple reasoning trajectories with probabilistic beam reasoning, combined with a training loss over both intermediate steps and final answers, reduces cascading errors and improves accuracy. On four widely used multi-hop QA datasets, AMKOR reports token-level F1 scores above five comparison methods, with the largest margins on the deepest-reasoning datasets and an average improvement of 2.5 points. The paper also reports better tolerance to injected retrieval noise and comparable inference latency, so the central claim is that multi-source fusion, beam search over trajectories, and multi-granular training together produce the gains.

What carries the argument

The central mechanism is probabilistic beam reasoning over a dynamically fused knowledge representation. At each step, the model builds a fused embedding of parametric and retrieved knowledge through scaled dot-product attention, generates multiple candidate reasoning steps with a softmax over a learned projection, and keeps the trajectory with the highest joint probability. A multi-granular loss, combining a local loss over individual reasoning steps and a global loss over the final answer, trains both levels at once. This machinery is what the paper credits for mitigating cascading errors and for staying robust when retrieved knowledge is noisy or conflicting.

What would settle it

Reproduce Table 1 by running the five listed comparison methods and AMKOR on HotpotQA and MuSiQue with the same retriever, prompts, decoding settings, and answer extraction; if the reported baseline scores are unreachable or AMKOR's margin becomes statistically insignificant, the state-of-the-art claim is falsified. The absence of released code makes this reproduction the direct test.

Watch

Extended reading notes

Core claim

On the paper's own terms, AMKOR factorizes the joint probability of a reasoning trajectory and final answer as $P(a|T,q,K)$ times the product of per-step probabilities $P(t_i | t_{<i}, q, K)$. At each step it fuses the LLM's parametric knowledge with retrieved snippets using scaled dot-product attention, proposes $b$ candidate steps through a softmax, and selects the trajectory that maximizes the product of step probabilities. Training minimizes a weighted sum of a local loss over intermediate steps and a global loss over the final answer. The paper's discovery claim is that this machinery yields state-of-the-art token-level F1 on HotpotQA, 2WikiMQA, MuSiQue, and Bamboogle, and that the beam-reasoning component is the largest single contributor: removing it drops F1 on MuSiQue from 37.4 to 30.9.

Load-bearing premise

The claimed superiority over the five comparison methods rests on the assumption that those methods were configured and run under fair, comparable conditions; the paper does not report their hyperparameters or prompts, cites none of their papers, and releases no code, so a mismatch would erase the reported advantage.

Editorial extensions

If this is right

  • If the reported F1 gains hold, multi-hop QA systems can improve without extra retrieval traffic: AMKOR averages 2.3 retrievals per question with lower latency than the best tree-based baseline.
  • The ablation points to beam reasoning as the largest driver, so retrieval-augmented systems should expect more from trajectory search than from additional fusion or loss engineering.
  • The multi-granular loss is a transferable recipe: supervise both intermediate reasoning steps and final answers with a weighted sum, which can be applied to other stepwise generative reasoning tasks.
  • The noise-robustness results imply the framework can tolerate a high fraction of irrelevant context: at 40% injected noise on one dataset, AMKOR's F1 drops from 63.2 to 54.3, while the strongest baseline drops from 60.4 to 43.8.

Reading between the lines

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

  • The paper's error analysis shows knowledge omission (41.2% on MuSiQue) and conflicts (35.4%) outweigh reasoning inaccuracies, suggesting that future gains are more likely to come from retrieval and conflict resolution than from deeper reasoning, a direction the conclusion mentions only briefly.
  • The per-step probabilities AMKOR already computes could be repurposed as confidence signals for abstaining or flagging low-confidence answers; the paper does not explore this.
  • A stress test the paper leaves unrun is cross-dataset transfer, training on one multi-hop dataset and evaluating on another, which would separate mechanism-level gains from dataset-specific tuning.
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 / 8 minor

Summary. The paper proposes AMKOR, a generative LLM-based framework for multi-source multi-hop QA. It combines parametric and retrieved knowledge via attention-based fusion, explores reasoning trajectories by probabilistic beam reasoning, and optimizes a two-term loss over local reasoning steps and the final answer. The central claim is that AMKOR achieves state-of-the-art F1 on HotpotQA, 2WikiMQA, MuSiQue, and Bamboogle, with an average improvement of 2.5% over baselines. I read the paper as an empirical systems contribution. The stress-test concern about the experiments lands: none of the five baselines is cited, no model/configuration/split details are given, and Table 1 cannot be used to verify the headline 2.5% improvement. The circularity concern does not land, since Eqs. (1)-(8) are standard definitions and are not what generates the empirical numbers. The paper does not include code, error bars, or significance tests, so the empirical core is currently not reproducible.

Significance. If the empirical claims were substantiated, the paper would address an important problem: robust multi-hop reasoning from noisy, heterogeneous sources is a real bottleneck, and the proposed combination of fusion, beam search, and multi-granular losses is a reasonable design space to explore. The paper's strengths are the clear formal framing of the joint probability in Eq. (1) and the inclusion of human and noise-robustness evaluations, which are relevant for the claimed robustness. However, none of these strengths is backed by a reproducible artifact: there is no code release, no baseline citation, no hyperparameter report, and no statistical support. A central advertised number, the 2.5% average improvement, is contradicted by the paper's own Table 1. I therefore cannot regard the contribution as established.

major comments (4)
  1. [§4.2, Tables 1–7] The state-of-the-art claim rests entirely on comparisons that are not verifiable. None of CoT, OneR, IRCoT, FLARE, and ProbTree is cited, and the manuscript reports no base LLM, retriever, corpus version, dataset split, prompt, decoding configuration, or hyperparameters (including λ_local, λ_global, beam width, and number of retrieved snippets). No code is released, and no error bars or significance tests are provided. Under these conditions the 'significantly outperforming' claim cannot be checked or reproduced; this is load-bearing because the paper's only evidence for its central claim is these tables.
  2. [§1 vs. Table 1] The abstract and Section 1 state that AMKOR achieves an average improvement of 2.5% across datasets. Table 1 does not support this: against the strongest baseline per dataset, the margins are 2.8 (HotpotQA), 5.5 (2WikiMQA), 4.5 (MuSiQue), and 8.6 (Bamboogle) F1 points, averaging 5.35 points; against the mean baseline the margins are larger. The manuscript gives no formula or subset of numbers that yields 2.5%. The headline number is therefore internally inconsistent with the reported results.
  3. [§3.3–§3.4] The method is under-specified at the points the contributions depend on. Eq. (3) does not state how Q, K, V are derived from h_param and the snippet embeddings; Eq. (4) defines candidate probabilities but not how candidate steps are generated or how h_i^j is computed; and Eqs. (6)-(8) contain no contrastive term, despite the Introduction and contribution list promising contrastive learning. In addition, Eq. (7) requires ground-truth intermediate steps t_i, but the paper never states where such supervision comes from for any of the four datasets. These omissions make it impossible to implement, train, or ablate the proposed framework from the text.
  4. [§4.1] The evaluation protocol is not specified. The paper does not state which dataset splits or settings are used; for HotpotQA and MuSiQue, results differ substantially between distractor and fullwiki settings, and Section 4.5 introduces subsets (2-hop/3-hop/4-hop MuSiQue) without giving their sizes or construction. Without this information the F1 values in Tables 1-7 cannot be interpreted relative to published work, and the claimed robustness and complexity analyses cannot be reproduced.
minor comments (8)
  1. [Title] The title in the manuscript body contains a spacing typo, 'Multi-granular T raining Strategies'; this should be corrected.
  2. [Table captions] Several captions are set as 'T able 1' etc.; the extra space should be removed.
  3. [Abstract] The abstract says AMKOR 'establishes a new benchmark' for multi-source multi-hop QA, but the paper releases no benchmark dataset or leaderboard; this phrasing should be revised to describe a strong result.
  4. [§2.1] The related work on language models cites works on insect classification, bioinformatics, medical image registration, and product QA that are not integrated with multi-hop QA; the section should be tightened to the methods actually compared or built upon.
  5. [§4.4] The human evaluation reports averaged scores with no standard deviations or inter-annotator agreement, and it is not stated whether all methods were scored on the same 100 examples.
  6. [§4.1] Token-level F1 is used without defining the answer normalization; for standard datasets this choice can change scores nontrivially. The paper also does not state whether the official evaluation scripts were used.
  7. [§4.8] Table 7 reports robustness to noisy knowledge but does not describe how synthetic noise is injected into the retrieved knowledge; please specify the procedure.
  8. [§4.5] Table 4 lists only CoT, OneR, ProbTree, and AMKOR, omitting IRCoT and FLARE, which makes the claim that the gap 'widens' against baseline methods incomplete.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the method equations are definitional decompositions, but no fitted parameter is renamed as a prediction and no load-bearing claim reduces to a self-citation.

full rationale

The derivation chain in Section 3 consists of definitional decompositions rather than fitted-input predictions: Eq. (1) is the probability chain rule, Eq. (3) defines a scaled dot-product attention fusion function, Eq. (4) defines candidate-step probabilities by softmax over a linear map, Eq. (5) selects a trajectory by maximizing the joint product, and Eq. (6) is a weighted sum of local and global losses. None of these equations takes a quantity fitted to data and then presents it as an independently derived prediction, and the empirical SOTA claims in Section 4 are not derived from these equations at all. There is also no load-bearing self-citation: the reference list contains no work by the present authors, no uniqueness theorem is invoked, and no ansatz is smuggled in via prior work. The reported 'average improvement of 2.5%' is not derivable from Table 1 and the baseline configurations are undocumented, but those are empirical reproducibility and falsifiability concerns, not circularity. The paper's central claim therefore does not reduce by construction to its inputs, so the honest finding is no significant circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical or mathematical entities. Its method rests on standard probabilistic identities and attention machinery, plus unverified domain assumptions about fusion, loss weighting, and baseline fairness. Free hyperparameters are numerous but all unreported, so the method cannot be reconstructed.

free parameters (4)
  • lambda_local = not reported
    Weight for the local reasoning loss in Eq. 6; no value or tuning procedure is given.
  • lambda_global = not reported
    Weight for the global answer loss in Eq. 6; no value or tuning procedure is given.
  • beam width b = not reported
    Number of candidate reasoning steps considered at each step in Section 3.3; no value is reported.
  • number of retrieved snippets s = not reported
    Number of external knowledge snippets fused per step in Section 3.2; no value or retrieval mechanism is described.
assumptions (4)
  • standard math The joint probability P(a,T|q,K) factorizes as in Eq. 1 via the chain rule.
    Eq. 1 is the probability chain rule applied to a trajectory; standard and correct.
  • domain assumption Scaled dot-product attention (Eq. 3) is an appropriate fusion operator for parametric and retrieved knowledge embeddings.
    The paper simply asserts Fuse uses attention, with no justification that heterogeneous embeddings are comparable or that attention resolves knowledge conflicts.
  • domain assumption Optimizing the weighted sum of local and global losses (Eq. 6) improves both intermediate reasoning and final answer quality.
    No analysis is given for why this particular combination is beneficial; the ablation only reports a table without statistical validation.
  • domain assumption The baseline comparisons in Tables 1, 3, and 4 are fair and computed under matched settings.
    Baselines are not cited and their configurations are undisclosed, so the comparison assumes a fair setup without evidence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-granular Training Strategies for Robust Multi-hop Reasoning Over Noisy and Heterogeneous Knowledge Sources." pith.science (2026). https://pith.science/paper/FFIFUSJJ

@misc{pith2026250205944,
  author       = {Pith},
  title        = {Pith review of: Multi-granular Training Strategies for Robust Multi-hop Reasoning Over Noisy and Heterogeneous Knowledge Sources},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FFIFUSJJ}},
  note         = {Machine review of arXiv:2502.05944}
}
read the original abstract

Multi-source multi-hop question answering (QA) represents a challenging task in natural language processing due to the need for dynamic integration of heterogeneous knowledge sources and multi-step reasoning. Existing methods often suffer from cascading errors, insufficient handling of knowledge conflicts, and computational inefficiency. In this paper, we propose Adaptive Multi-source Knowledge-Oriented Reasoning (AMKOR), a generative framework that leverages large language models (LLMs) to dynamically fuse parametric and retrieved knowledge while exploring reasoning trajectories using probabilistic beam reasoning. AMKOR is further enhanced by a multi-granular learning strategy, optimizing both local reasoning steps and global answer accuracy. Experiments conducted on four widely-used multi-hop QA datasets, including HotpotQA and MuSiQue, demonstrate that AMKOR achieves state-of-the-art performance, significantly outperforming baseline methods on both reasoning accuracy and robustness. Additional analyses confirm its scalability, adaptability to noisy knowledge, and superior ability to handle complex multi-hop tasks. This work establishes a new benchmark for multi-source multi-hop QA by effectively combining reasoning quality and efficiency.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 22 canonical work pages

  1. [1]

    In: Findings of the Associ ation for Computational Linguistics: ACL 2023

    Zhou, Y., Shen, T., Geng, X., Tao, C., Xu, C., Long, G., Jiao , B., Jiang, D.: Towards robust ranker for text retrieval. In: Findings of the Associ ation for Computational Linguistics: ACL 2023. pp. 5387–5401 (2023)

    work page 2023

  2. [2]

    In : Proceedings of the AAAI Conference on Artificial Intelligence

    Zhou, Y., Shen, T., Geng, X., Tao, C., Shen, J., Long, G., Xu , C., Jiang, D.: Fine-grained distillation for long document retrieval. In : Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 19732–1 9740 (2024)

    work page 2024

  3. [3]

    In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A

    Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., X ia, F., Chi, E.H., Le, Q.V., Zhou, D.: Chain-of-thought prompting elicits reason ing in large language models. In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) Advances in Neural Information Processing Systems 3 5: Annual Conference on Neural Information Processing S...

    work page 2022

  4. [4]

    Findings of the Association for Computational Linguistics: ACL-IJCNLP 20 21 (2021)

    Zhou, Y., Geng, X., Shen, T., Pei, J., Zhang, W., Jiang, D.: Modeling event- pair relations in external knowledge graphs for script reas oning. Findings of the Association for Computational Linguistics: ACL-IJCNLP 20 21 (2021)

    work page 2021

  5. [5]

    CoRR abs/2406.14732 (2024)

    Bardhan, J., Xiao, B., Wang, D.Z.: TTQA-RS- A break-down p rompting approach for multi-hop table-text question answering with reasonin g and summarization. CoRR abs/2406.14732 (2024). https://doi.org/10.48550/ARXIV.2406.14732, https://doi.org/10.48550/arXiv.2406.14732

    arXiv doi:10.48550/arxiv.2406.14732 2024

  6. [6]

    CoRR abs/2401.02582 (2024)

    Zhang, D., Yang, J., Lyu, H., Jin, Z., Yao, Y., Chen, M., Luo, J.: Cocot: Contrastive chain-of-thought prompting fo r large multimodal models with multiple image inputs. CoRR abs/2401.02582 (2024). https://doi.org/10.48550/ARXIV.2401.02582, https://doi.org/10.48550/arXiv.2401.02582 12 J. Coleman et al

    arXiv doi:10.48550/arxiv.2401.02582 2024

  7. [7]

    In: Carpuat, M., de Marneffe, M., Ruíz, I.V.M

    Zhong, W., Gao, Y., Ding, N., Qin, Y., Liu, Z., Zhou, M., Wan g, J., Yin, J., Duan, N.: Proqa: Structural prompt-based pre-training for unified question an- swering. In: Carpuat, M., de Marneffe, M., Ruíz, I.V.M. (eds. ) Proceedings of the 2022 Conference of the North American Chapter of the Asso ciation for Com- putational Linguistics: Human Language Techn...

    work page 2022

  8. [8]

    InsectMamba: Insect Pest Classification with State Space Model

    Wang, Q., Wang, C., Lai, Z., Zhou, Y.: Insectmamba: Insect pest classification with state space model. arXiv preprint arXiv:2404.03611 (2 024)

    work page Pith review arXiv

Show all 31 references
  1. [9]

    In: Proceedings of the 60th Annual Meeting o f the Association for Computational Linguistics (Volume 1: Long Papers)

    Zhou, Y., Shen, T., Geng, X., Long, G., Jiang, D.: Claret: P re-training a correlation-aware context-to-event transformer for even t-centric generation and classification. In: Proceedings of the 60th Annual Meeting o f the Association for Computational Linguistics (Volume 1: Lo...

    2022

  2. [10]

    CoRR abs/2306.07377 (2023)

    Nicholas, G., Bhatia, A.: Lost in translation: Large lan - guage models in non-english content analysis. CoRR abs/2306.07377 (2023). https://doi.org/10.48550/ARXIV.2306.07377, https://doi.org/10.48550/arXiv.2306.07377

    2023 doi arXiv

  3. [11]

    arXiv preprin t arXiv:2410.19732 (2024)

    Zhou, Y., Rao, Z., Wan, J., Shen, J.: Rethinking visual de pendency in long-context reasoning for large vision-language models. arXiv preprin t arXiv:2410.19732 (2024)

    2024 arXiv

  4. [12]

    CoRR abs/2202.03371 (2022), https://arxiv.org/abs/2202.03371

    Müller, M., Laurent, F.: Cedille: A large autoregressiv e french language model. CoRR abs/2202.03371 (2022), https://arxiv.org/abs/2202.03371

    2022 arXiv

  5. [13]

    arXiv preprint arXiv:2501.0137 7 (2025)

    Zhou, Y., Song, L., Shen, J.: Training medical large visi on-language models with abnormal-aware feedback. arXiv preprint arXiv:2501.0137 7 (2025)

    2025

  6. [14]

    CoRR abs/2408.10441 (2024)

    Chang, T.A., Arnett, C., Tu, Z., Bergen, B.K.: Gold- fish: Monolingual language models for 350 languages. CoRR abs/2408.10441 (2024). https://doi.org/10.48550/ARXIV.2408.10441, https://doi.org/10.48550/arXiv.2408.10441

    2024 doi arXiv

  7. [15]

    Ojo, J., Ogueji, K.: How good are commercial large langua ge models on african languages? In: Proceedings of the 4th Workshop on African Na tural Lan- guage Processing, AfricaNLP@ICLR 2023, Kigali, Rwanda, Ma y 1, 2023 (2023), https://openreview.net/pdf?id=MCgyGyRPEIU

    2023

  8. [16]

    CoRR abs/2401.04155 (2024)

    Liu, J., Yang, M., Yu, Y., Xu, H., Li, K., Zhou, X.: Large la n- guage models in bioinformatics: applications and perspect ives. CoRR abs/2401.04155 (2024). https://doi.org/10.48550/ARXIV.2401.04155, https://doi.org/10.48550/arXiv.2401.04155

    2024 doi arXiv

  9. [17]

    CoRR abs/2405.18774 (2024)

    Ma, M., Yang, Y.: Llama-reg: Using llama 2 for unsu- pervised medical image registration. CoRR abs/2405.18774 (2024). https://doi.org/10.48550/ARXIV.2405.18774, https://doi.org/10.48550/arXiv.2405.18774

    2024 doi arXiv

  10. [18]

    CoRR abs/2205.07634 (2022)

    Veres, C.: A precis of language models are not models of la nguage. CoRR abs/2205.07634 (2022). https://doi.org/10.48550/ARXIV.2205.07634, https://doi.org/10.48550/arXiv.2205.07634

    2022 doi arXiv

  11. [19]

    ICT Express 5(3), 182–186 (2019)

    Spirov, A.V., Myasnikova, E.M.: Linguistic modelling o f gene regulation: Trans- lation from the language of experiments to the language of mo delling. ICT Express 5(3), 182–186 (2019). https://doi.org/10.1016/J.ICTE.2018.10.008, https://doi.org/10.1016/j.icte.2018.10.008

    2019 doi

  12. [20]

    CoRR abs/2303.00077 (2023)

    Houghton, C.J., Kazanina, N., Sukumaran, P.: Beyond the limitations of any imaginable mechanism: large language models and psycho linguistics. CoRR abs/2303.00077 (2023). https://doi.org/10.48550/ARXIV.2303.00077, https://doi.org/10.48550/arXiv.2303.00077 AMKOR 13

    2023 doi arXiv

  13. [21]

    In: Findings of the Association for Comput ational Linguistics, ACL 2024, Bangkok, Thailand and virtual meeting, August 11- 16, 2024

    Zhou, Y., Li, X., Wang, Q., Shen, J.: Visual in-context le arning for large vision- language models. In: Findings of the Association for Comput ational Linguistics, ACL 2024, Bangkok, Thailand and virtual meeting, August 11- 16, 2024. pp. 15890– 15902. Association for Computat...

    2024

  14. [22]

    In: The Thir- teenth International Conference on Learning Representati ons (2025), https://openreview.net/forum?id=N1vYivuSKq

    Zhou, Y., Shen, J., Cheng, Y.: Weak to strong generalizat ion for large language models with multi-capabilities. In: The Thir- teenth International Conference on Learning Representati ons (2025), https://openreview.net/forum?id=N1vYivuSKq

    2025

  15. [23]

    https://doi.org/10.48550/ARXIV.2410.15512, https://doi.org/10.48550/arXiv.2410.15512

    Balepur, N., Gu, F., Ravichander, A., Feng, S., Boyd-Gra ber, J.L., Rudinger, R.: Reverse question answering: Can an LLM write a question so hard (or bad) that it can’t answer? CoRR abs/2410.15512 (2024). https://doi.org/10.48550/ARXIV.2410.15512, https://doi.org/10.48550/arXi...

    2024 doi arXiv

  16. [24]

    CoRR abs/2205.12665 (2022)

    Amouyal, S.J., Rubin, O., Yoran, O., Wolfson, T., Herzig , J., Be- rant, J.: QAMPARI: : An open-domain question answering benc h- mark for questions with many answers from multiple paragrap hs. CoRR abs/2205.12665 (2022). https://doi.org/10.48550/ARXIV.2205.12665, https://doi....

    2022 doi arXiv

  17. [25]

    (eds.) Proceedings of the 16th Conference o f the European Chapter of the Association for Computational Linguistics: Main Volume, EACL 2021, Online, April 19 - 23, 2021

    Tang, Y., Ng, H.T., Tung, A.K.H.: Do multi-hop question a nswering systems know how to answer the single-hop sub-questions? In: Merlo, P., Tiedemann, J., Tsarfaty, R. (eds.) Proceedings of the 16th Conference o f the European Chapter of the Association for Computational Lingui...

    2021 doi

  18. [26]

    In: Muresan, S., Nakov, P., Villavicencio, A

    Wang, R., Qian, Y., Feng, F., Wang, X., Jiang, H.: Co-vqa : Answering by interactive sub question sequence. In: Muresan, S., Nakov, P., Villavicencio, A. (eds.) Findings of the Association for Computational Lingu istics: ACL 2022, Dublin, Ireland, May 22-27, 2022. pp. 2396–240...

    2022 doi

  19. [27]

    Hwang, S., Lee, G.G.: Conversational QA dataset generat ion with answer re- vision. In: Calzolari, N., Huang, C., Kim, H., Pustejovsky, J., Wanner, L., Choi, K., Ryu, P., Chen, H., Donatelli, L., Ji, H., Kurohashi , S., Paggio, P., Xue, N., Kim, S., Hahm, Y., He, Z., Lee, T.K....

    2022

  20. [28]

    In: 5th International Conference on Behavi oral, Economic, and Socio-Cultural Computing, BESC 2018, Kaohsiung, Taiwan, N ovember 12-14,

    Chao, Z., Li, L.: The combination of context information to enhance simple ques- tion answering. In: 5th International Conference on Behavi oral, Economic, and Socio-Cultural Computing, BESC 2018, Kaohsiung, Taiwan, N ovember 12-14,

    2018

  21. [29]

    CoRR abs/2311.17502 (2023)

    Hu, X.: Enhancing answer selection in community questio n an- swering with pre-trained and large language models. CoRR abs/2311.17502 (2023). https://doi.org/10.48550/ARXIV.2311.17502, https://doi.org/10.48550/arXiv.2311.17502

    2023 doi arXiv

  22. [30]

    In: Hutte r, F., Kersting, K., Lijf- 14 J

    Zhang, S., Zhang, X., Lau, J.H., Chan, J., Paris, C.: Less is more: Rejecting un- reliable reviews for product question answering. In: Hutte r, F., Kersting, K., Lijf- 14 J. Coleman et al. fijt, J., Valera, I. (eds.) Machine Learning and Knowledge Di scovery in Databases - Euro...

    2020 doi

  23. [2018]

    pp. 109–114. IEEE (2018). https://doi.org/10.1109/BESC.2018.8697305, https://doi.org/10.1109/BESC.2018.8697305

    2018

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.