Improving Narrative Classification and Explanation via Fine Tuned Language Models

Mohit Gupta; Rahul Bouri; Rishit Tyagi

REVIEW 5 major objections 4 minor 20 references

Improving Narrative Classification and Explanation via Fine Tuned Language Models

T0 review · 5 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A recall-biased BERT classifier plus GPT-4o refinement, and retrieval-grounded ReACT prompts, improve narrative labeling and explanation across five languages.

desk verdict The explanation-generation result is a genuine empirical nugget, but the abstract's classification claim is untested and the evaluation has a suspicious test-set leak; as written it should not be accepted. read the letter →

arxiv 2509.04077 v1 pith:HIF6DP7B submitted 2025-09-04 cs.CL

Rishit Tyagi , Rahul Bouri , Mohit Gupta This is my paper · ORCID

classification cs.CL

keywords narrativeclassificationmulti-labelimplicitdetectionfine-tunedBERTReACTpromptingsemanticsentenceretrievalauxiliaryknowledgebaseBERTScoreevaluation

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

At the center of this paper is a two-stage recipe for handling narratives that hide inside news articles rather than being stated outright. For classification, the authors fine-tune BERT with a deliberate recall bias—using focal loss and an adaptively lowered decision threshold—so that candidate narrative labels are rarely dropped, and then let GPT-4o prune and refine those candidates against the task's taxonomy. On a five-language corpus built around climate change and the Ukraine-Russia war, this moves dominant-narrative F1 from 0.251 with GPT-only prompting to 0.467, and sub-narrative F1 from 0.164 to 0.217. For explanation, the paper claims that what matters is not the size of the generation model but what the prompt can access: retrieving the article sentences most similar to the target narrative and pairing them with a structured taxonomy table inside a ReACT reasoning prompt consistently beats simpler prompts across Bulgarian, English, Hindi, Portuguese, and Russian. The authors interpret this as evidence that grounding in retrieved evidence plus a small knowledge base reduces hallucination and makes justifications more reliable, and they show the classification pipeline retains most of its gain when the refining LLM is the cheaper GPT-4o-mini.

What carries the argument

Two mechanisms carry the result. First, recall-prioritized classification: bert-base-uncased is trained with focal loss (to soften class imbalance) and an adaptively lowered threshold so false negatives are rare, then GPT-4o refines the candidate labels in two taxonomy-guided passes, first pruning false narrative labels, then assigning sub-narratives under the surviving main narratives. Second, retrieval-grounded explanation: sentences from the article are embedded and ranked by cosine similarity through a dual-pass scheme—top-five by similarity to the dominant narrative, plus any sentence above the fifth-ranked score for the sub-narratives—and the retrieved evidence plus a structured taxono

What would settle it

Take the same five-language news corpus and split it once into training, validation, and test folds; freeze the retrieval threshold, prompt template, and LLM variant using only the validation fold; then compute BERTScore F1 on the untouched test fold for the three ReACT configurations (plain, knowledge-base-only, knowledge-base-plus-retrieval). If the full configuration does not beat the knowledge-base-only configuration in at least one language on that held-out fold, the paper's central improvement claim is settled against.

Watch

Extended reading notes

Core claim

The paper's central claim is that narrative detection and narrative explanation improve when a small, recall-focused model proposes and a large model disposes, and when explanation prompts are given both textual evidence and definitional grounding. The authors report that the BERT-plus-GPT-4o classification pipeline reaches 0.467 F1 for dominant narratives and 0.217 for sub-narratives, well above the 0.251 and 0.164 achieved by prompting GPT-4o alone. For justification, adding semantically retrieved article sentences and a taxonomy knowledge base to a ReACT prompt outperforms both the simple prompt and the taxonomy-only prompt in every one of the five languages, with the full configuration s

Load-bearing premise

The load-bearing premise is that the reported F1 and BERTScore gains reflect generalization to articles the system has not already steered its configuration toward, rather than prompt or threshold choices made after seeing the scored labels.

Editorial extensions

If this is right

The BERT-then-GPT-4o pipeline shows that cheap, recall-heavy label candidates plus an LLM referee can beat direct LLM prompting on hierarchical multi-label narrative classification (0.467 vs 0.251 F1).
Explanation quality depends more on what the prompt can retrieve than on prompt complexity alone: the full retrieval-plus-knowledge configuration beat simple and knowledge-only prompts in all five languages.
Because the gain comes from grounding rather than a bigger generator, the approach supports replacing expensive LLMs with smaller variants for deployment; the paper reports near-comparable classification results with GPT-4o-mini.
For media analysis and fact-checking, the paper provides a concrete pattern that yields both a label and a short, evidence-anchored justification for why the label was chosen.

Reading between the lines

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

This retrieval-plus-taxonomy ReACT template should transfer to other hierarchical labeling tasks—for example propaganda techniques, framing categories, or hate-speech subtypes—where a taxonomy and source text exist; the paper does not test this transfer, but its mechanism is not narrative-specific.
The dual-pass dynamic-threshold retrieval used for sub-narrative evidence is a reusable evidence-selection rule; a natural standalone test would apply it to claim verification or summarization, where selected sentences must cover both a main claim and its supporting details.
The paper's strongest empirical claim—consistent gains in all five languages—would be settled most cleanly by a re-run in which prompt variants and thresholds are frozen on a validation fold before any test label is seen; if the advantage of the semantic-search component shrinks under that protocol, the reported gain is partly a product of configuration choices made after seeing the scored labels.

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.

Desk Editor's Note

read the letter

Two things to know. First, the paper's strongest defensible result is in Table 3: adding semantic retrieval on top of a ReACT prompt with the taxonomy knowledge base improves BERTScore F1 for narrative justifications across all five languages, with a consistent step-up from the simple prompt. That is a useful, reproducible-looking empirical finding for the SemEval-2025 shared task, and the authors describe the retrieval and prompting design in enough detail that someone could re-implement it. Second, the abstract's headline claim—that auxiliary knowledge in prompts improves classification accuracy—is not supported by any experiment in the paper. Table 2 only varies whether BERT pre-filters the labels before GPT-4o; it never ablates the taxonomy prompt. So the classification half of the central claim is a non-sequitur.

What the paper does well: the method section is clear and honest about design choices (focal loss for imbalance, recall-oriented thresholding, dual-pass retrieval, ReACT with taxonomy in the Action step). The explanation result is new for this benchmark and is the paper's real contribution.

Where it falls down: - The classification claim is untested, as above. - The evaluation is not rigorous: only an 80-20 split is mentioned, and Section 5.1 says the results were 'compiled after the task was complete,' which strongly suggests the official test set was used for configuration selection. That is a load-bearing leak if true, and it's at minimum undocumented. - The identical F1 scores between GPT-4o-mini and GPT-4o across most rows in Table 2 (e.g., 0.251 for Narrative Overall both) are implausible and unexplained. That alone would make me want to see the per-example predictions. - They claim precision and recall improvements for justification but only report BERTScore F1. - No code, data, or prompts are released, which hurts reproducibility.

Who this is for: readers using LLM pipelines for narrative analysis in shared-task settings. The explanation-generation recipe (semantic retrieval + ReACT + taxonomy in Action) is worth trying. The classification pipeline, as presented, should not be cited as evidence that taxonomy prompting improves classification.

Recommendation: This deserves a serious referee—the explanation result is worth checking—but the paper needs major revision before acceptance: re-run the classification ablation, disclose the test-set split, explain the identical F1s, and release code/prompts. As it stands, I'd reject it.

Referee Report

5 major / 4 minor

Summary. The paper addresses SemEval-2025 Task 10 on multilingual narrative characterization, with two components: multi-label classification of narratives/sub-narratives (Subtask 2) and generation of evidence-based justifications (Subtask 3). For classification, the authors fine-tune BERT-base-uncased with focal loss, then use a two-stage GPT-4o pipeline with a narrative taxonomy to refine predictions. For explanation, they retrieve semantically similar sentences and use a ReACT-style prompt augmented with a taxonomy-based auxiliary knowledge base. Results are reported as F1 scores for classification (Table 2) and BERTScore F1 for justification (Table 3). The abstract claims that integrating auxiliary knowledge in prompts improves both classification accuracy and justification reliability.

Significance. If the claims were fully supported, the paper would provide a practical pipeline for a shared task, combining fine-tuned encoders with LLM post-processing and retrieval-augmented reasoning. The focus on grounded, multilingual justifications is relevant, and the use of a structured taxonomy as an auxiliary knowledge source is a reasonable design idea. However, the current evidence base is too thin: the classification half of the central claim is not tested by any reported ablation, and the explanation results lack statistical grounding. The contribution is therefore presently at the level of a system description rather than a validated method.

major comments (5)

[Abstract and Section 5.2, Table 2] The abstract claims that integrating auxiliary knowledge in prompts improves classification accuracy, but no experiment in the paper varies the taxonomy/knowledge-base variable for classification. Table 2 compares GPT-only against BERT+GPT; the taxonomy is present in the GPT post-processing step in both BERT+GPT conditions and is not ablated. The large improvement from 0.251 to 0.467 is attributable to BERT pre-filtering, not to the taxonomy. This is load-bearing: the headline claim is unsupported by the presented evidence. An ablation with and without the taxonomy prompt in the GPT post-processing step is required.
[Section 5.1 and Section 4.1.1] The paper describes only an 80-20 train-validation split, yet reports final F1 scores without specifying whether they are computed on a held-out test set or on the official SemEval test labels. Section 5.1 says results were 'compiled after the task was complete,' which suggests the test set may have been used for configuration selection. No error bars, confidence intervals, or significance tests are provided; Table 3 differences, e.g., 0.6018 vs. 0.6114, may be within noise. The authors must clearly describe the exact data split, which labels were used for development, which for final evaluation, and report variance or significance.
[Table 2] The GPT-4o-mini and GPT-4o narrative rows are identical for all three Narrative categories (0.227, 0.301, 0.251), and sub-narrative scores are nearly identical (0.156 vs. 0.158, etc.). This is implausible if these are different models with any stochasticity, and it suggests possible copying, a reporting error, or an unreported deterministic configuration. At minimum, this requires an explanation and, ideally, raw prediction outputs or repeated runs to establish reliability.
[Section 5.2 and Table 3] The text states that the framework 'demonstrated superior performance across all three evaluation metrics—BERT F1, Precision, and Recall,' but Table 3 reports only BERTScore F1. Precision and recall are never shown. Additionally, no comparison is made with official SemEval baseline systems or with a simple fine-tuned BERT/generative baseline without the proposed retrieval and knowledge augmentation. The claimed consistent improvement across languages needs to be backed with the actual metric values and ideally per-language breakdowns with variance.
[Section 4.2.1] The dual-pass retrieval uses the dominant narrative as a query and then adds sub-narrative sentences that exceed the threshold set by the 5th-ranked sentence. This dynamic threshold is vaguely defined: it is unclear whether the threshold is a similarity value or a rank-based cutoff, how it behaves across languages/domains, and what happens when fewer than five sentences exist. Since the retrieval component is central to the claimed improvement in Table 3, its behavior should be specified precisely or an ablation varying the number of retrieved sentences should be reported.

minor comments (4)

[Section 4.2.2] The phrase 'we experimented with few-shot prompting but found that ReACT prompting yielded more structured...' is not supported by any quantitative comparison. Please either report these experiments or remove the claim.
[Section 5.1] The sentence 'The approach with the highest F1 score was chosen as the objective' is unclear: is this a description of model selection or of the evaluation metric used by the task organizers? Please rephrase.
[Overall] There is no code or data availability statement. Given the reproducibility concerns raised above, including code and/or anonymized outputs would be valuable.
[Related Work] References are somewhat scattered; e.g., the survey on hallucination (Huang et al., 2025) and the XAI survey (Zhao et al., 2024) are cited generally but not tied to the specific design choices. Some related-work citations appear to be placeholders rather than substantive discussion.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the pipeline is evaluated against external SemEval-2025 gold labels, and no prediction reduces to a fitted parameter or self-citation.

full rationale

The paper's derivation chain is empirical and externally grounded. Classification fine-tunes bert-base-uncased on the task's training labels with focal loss and threshold tuning, then refines predictions through a GPT-4o post-processing step; the reported F1 numbers in Table 2 are compared against the shared task's gold labels. Explanation generation in Section 4.2 retrieves sentences using the given dominant narrative and sub-narratives and then structures a ReACT justification with the provided taxonomy; the evaluation in Table 3 uses BERTScore against reference explanations. There are no self-citations by the authors, no imported uniqueness theorem, no fitted parameter later relabeled as a prediction, and no ansatz smuggled in via citation. The abstract's claim that auxiliary knowledge improves classification is not directly demonstrated by Table 2, which lacks a with/without-taxonomy ablation, but that is an experimental omission rather than a circular reduction. The use of the target narrative as a retrieval query is inherent to Subtask 3 (the label is an input), so the explanation is not predicting the label from its own output. Section 5.1's 'compiled after the task was complete' is ambiguous about test-set selection but does not constitute a circular derivation. Score 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

No new entities are introduced. The taxonomy table comes from the task, and the retrieval and reasoning components are existing tools. The free parameters are primarily unreported hyperparameters and thresholds that the central performance claims depend on but that are not documented.

free parameters (5)

Adaptive classification threshold(s) = not reported
Section 4.1.2: threshold tuned adaptively to minimize false negatives; no values or validation procedure given.
Number of retrieved sentences (top-5) = 5
Section 4.2.1: top 5 sentences retrieved by cosine similarity; hard-coded without sensitivity analysis.
Dual-pass retrieval threshold = similarity of 5th-ranked sentence from first pass
Section 4.2.1: threshold set dynamically by the 5th-ranked sentence; no analysis of this choice.
GPT-4o/4o-mini prompt templates and temperatures = not reported
Sections 4.1.3 and 4.2.2: prompts are described at a high level; no exact prompts, temperatures, or seeds, so the exact behavior is not reproducible.
Focal loss hyperparameters = not reported
Section 4.1.2: focal loss is used but gamma/alpha values are not given, and the loss function is only cited.

assumptions (4)

domain assumption The SemEval-2025 Task 10 annotations (dominant and sub-narratives) are correct ground truth.
Sections 3 and 4 depend entirely on these labels for training and evaluation.
domain assumption BERTScore is a valid proxy for explanation justification quality.
Section 5 uses BERTScore as the only evaluation metric for explanations, without human evaluation or task-specific metrics.
domain assumption The taxonomy table definitions provide factual grounding that reduces hallucinations.
Section 4.2.3 asserts this, but no hallucination metric or human evaluation is reported.
domain assumption The 80-20 split and training procedure generalize to the official test set.
Section 4.1.1 describes only a train-validation split; the official test set and its separation from prompt/feature tuning are never described.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Improving Narrative Classification and Explanation via Fine Tuned Language Models." pith.science (2026). https://pith.science/paper/HIF6DP7B

@misc{pith2026250904077,
  author       = {Pith},
  title        = {Pith review of: Improving Narrative Classification and Explanation via Fine Tuned Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HIF6DP7B}},
  note         = {Machine review of arXiv:2509.04077}
}

read the original abstract

Understanding covert narratives and implicit messaging is essential for analyzing bias and sentiment. Traditional NLP methods struggle with detecting subtle phrasing and hidden agendas. This study tackles two key challenges: (1) multi-label classification of narratives and sub-narratives in news articles, and (2) generating concise, evidence-based explanations for dominant narratives. We fine-tune a BERT model with a recall-oriented approach for comprehensive narrative detection, refining predictions using a GPT-4o pipeline for consistency. For narrative explanation, we propose a ReACT (Reasoning + Acting) framework with semantic retrieval-based few-shot prompting, ensuring grounded and relevant justifications. To enhance factual accuracy and reduce hallucinations, we incorporate a structured taxonomy table as an auxiliary knowledge base. Our results show that integrating auxiliary knowledge in prompts improves classification accuracy and justification reliability, with applications in media analysis, education, and intelligence gathering.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 12 canonical work pages

[1]

Lu Cao, Xinyue Liu, and Hong Shen. 2021. Adaptable focal loss for imbalanced text classification. In International Conference on Parallel and Distributed Computing: Applications and Technologies, pages 466--475. Springer

work page 2021
[2]

Chris Culy and Susanne Z Riehemann. 2003. The limits of n-gram translation evaluation metrics. In Proceedings of Machine Translation Summit IX: Papers

work page 2003
[3]

Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171--4186

2019
[4]

Xiaocheng Feng, Bing Qin, and Ting Liu. 2018. A language-independent neural network for event detection. Science China Information Sciences, 61:1--12

work page 2018
[5]

Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, et al. 2025. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Information Systems, 43(2):1--55

work page 2025
[6]

Zhao Jingling, Zhang Huiyun, and Cui Baojiang. 2014. Sentence similarity based on semantic vector model. In 2014 Ninth International Conference on P2P, Parallel, Grid, Cloud and Internet Computing, pages 499--503. IEEE

work page 2014
[7]

Negar Mokhberian, Andrés Abeliuk, Patrick Cummings, and Kristina Lerman. 2020. https://doi.org/10.1007/978-3-030-60975-7_16 Moral Framing and Ideological Bias of News , page 206–219. Springer International Publishing

doi:10.1007/978-3-030-60975-7_16 2020
[8]

Jakub Piskorski, Tarek Mahmoud, Nikolaos Nikolaidis, Ricardo Campos, Alípio Jorge, Dimitar Dimitrov, Purificação Silvano, Roman Yangarber, Shivam Sharma, Tanmoy Chakraborty, Nuno Guimarães, Elisa Sartori, Nicolas Stefanovitch, Zhuohan Xie, Preslav Nakov, and Giovanni Da San Martino. 2025. SemEval -2025 task 10: Multilingual characterization and extraction...

work page 2025

Show all 20 references

[9]

Libo Qin, Qiguang Chen, Yuhang Zhou, Zhi Chen, Yinghui Li, Lizi Liao, Min Li, Wanxiang Che, and Philip S. Yu. 2024. https://api.semanticscholar.org/CorpusID:269005862 Multilingual large language model: A survey of resources, taxonomy and frontiers . ArXiv, abs/2404.04925

2024 arXiv
[10]

Korbinian Randl, John Pavlopoulos, Aron Henriksson, and Tony Lindgren. 2024. https://doi.org/10.18653/v1/2024.findings-acl.459 CICL e: Conformal in-context learning for largescale multi-class food risk classification . In Findings of the Association for Computational Linguisti...

2024 doi
[11]

Pedro L Rodriguez and Arthur Spirling. 2022. Word embeddings: What works, what doesn’t, and how to tell the difference for applied research. The Journal of Politics, 84(1):101--115

2022
[12]

Brenda Santana, Ricardo Campos, Evelin Amorim, Al \' pio Jorge, Purifica c \ a o Silvano, and S \'e rgio Nunes. 2023. A survey on narrative extraction from textual data. Artificial Intelligence Review, 56(8):8393--8435

2023
[13]

Bhaskarjit Sarmah, Dhagash Mehta, Benika Hall, Rohan Rao, Sunil Patel, and Stefano Pasquali. 2024. Hybridrag: Integrating knowledge graphs and vector retrieval augmented generation for efficient information extraction. In Proceedings of the 5th ACM International Conference on ...

2024
[14]

Chi Sun, Xipeng Qiu, Yige Xu, and Xuanjing Huang. 2019. How to fine-tune bert for text classification? In China national conference on Chinese computational linguistics, pages 194--206. Springer

2019
[15]

Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824--24837

2022
[16]

Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. 2023 a . Tree of thoughts: Deliberate problem solving with large language models. Advances in neural information processing systems, 36:11809--11822

2023
[17]

Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2023 b . React: Synergizing reasoning and acting in language models. In International Conference on Learning Representations (ICLR)

2023
[18]

Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. 2019 a . Theoretically principled trade-off between robustness and accuracy. In International conference on machine learning, pages 7472--7482. PMLR

2019
[19]

Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. 2019 b . Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675

2019 arXiv
[20]

Haiyan Zhao, Hanjie Chen, Fan Yang, Ninghao Liu, Huiqi Deng, Hengyi Cai, Shuaiqiang Wang, Dawei Yin, and Mengnan Du. 2024. Explainability for large language models: A survey. ACM Transactions on Intelligent Systems and Technology, 15(2):1--38

2024

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Lu Cao, Xinyue Liu, and Hong Shen. 2021. Adaptable focal loss for imbalanced text classification. In International Conference on Parallel and Distributed Computing: Applications and Technologies, pages 466--475. Springer

work page 2021

[2] [2]

Chris Culy and Susanne Z Riehemann. 2003. The limits of n-gram translation evaluation metrics. In Proceedings of Machine Translation Summit IX: Papers

work page 2003

[3] [3]

Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171--4186

2019

[4] [4]

Xiaocheng Feng, Bing Qin, and Ting Liu. 2018. A language-independent neural network for event detection. Science China Information Sciences, 61:1--12

work page 2018

[5] [5]

Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, et al. 2025. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Information Systems, 43(2):1--55

work page 2025

[6] [6]

Zhao Jingling, Zhang Huiyun, and Cui Baojiang. 2014. Sentence similarity based on semantic vector model. In 2014 Ninth International Conference on P2P, Parallel, Grid, Cloud and Internet Computing, pages 499--503. IEEE

work page 2014

[7] [7]

Negar Mokhberian, Andrés Abeliuk, Patrick Cummings, and Kristina Lerman. 2020. https://doi.org/10.1007/978-3-030-60975-7_16 Moral Framing and Ideological Bias of News , page 206–219. Springer International Publishing

doi:10.1007/978-3-030-60975-7_16 2020

[8] [8]

Jakub Piskorski, Tarek Mahmoud, Nikolaos Nikolaidis, Ricardo Campos, Alípio Jorge, Dimitar Dimitrov, Purificação Silvano, Roman Yangarber, Shivam Sharma, Tanmoy Chakraborty, Nuno Guimarães, Elisa Sartori, Nicolas Stefanovitch, Zhuohan Xie, Preslav Nakov, and Giovanni Da San Martino. 2025. SemEval -2025 task 10: Multilingual characterization and extraction...

work page 2025