BrightCookies at SemEval-2025 Task 9: Exploring Data Augmentation for Food Hazard Classification

Ali H\"urriyeto\u{g}lu; Bas H.M. van der Velden; Foteini Papadopoulou; Iris Hendrickx; Neris \"Ozen; Osman Mutlu

REVIEW 3 major objections 5 minor 15 references

BrightCookies at SemEval-2025 Task 9: Exploring Data Augmentation for Food Hazard Classification

T0 review · 3 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read Targeted minority-class augmentation with contextual word insertion improves BERT's fine-grained food hazard classification, but only for specific classes and models.

desk verdict Honest shared-task paper whose central significance claim collapses under multiple-testing correction; the observed gains are suggestive, not proven. read the letter →

arxiv 2504.20703 v1 pith:5I2KD726 submitted 2025-04-29 cs.CL

Foteini Papadopoulou , Osman Mutlu , Neris Özen , Bas H.M. van der Velden , Iris Hendrickx , Ali Hürriyetoğlu This is my paper · ORCID

classification cs.CL

keywords foodhazardclassificationdataaugmentationminorityclassescontextualwordinsertionBERTclassimbalanceSemEval-2025Task9explainableNLP

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 asks whether cheap word-level data augmentation on under-represented classes can help a multi-class food hazard classifier. It tests three operations—synonym replacement, random word swapping, and contextual word insertion—on top of TF-IDF machine learning models and four encoder-only transformers. The central positive result is narrow but concrete: for BERT, augmenting minority classes with contextual word insertion improved accuracy on minority hazard classes by about 6% over the baseline, and pairwise statistical tests showed significant gains in fine-grained categories. The paper is careful to report that no augmentation consistently improved overall F1-macro scores, and that minority-class gains came with slight losses on majority classes. A sympathetic reading is that targeted, cheap augmentation can help exactly where the task is hardest, but the effect is fragile and model-dependent.

What carries the argument

The load-bearing mechanism is an oversampling routine (Algorithm 1) that identifies minority classes by a support threshold, then distributes a fixed number of synthetic samples across each class's existing examples. The operation that carries the result is contextual word insertion (CW), which uses a masked BERT model to find likely words and inserts them into the title or text; synonym replacement and random word swapping serve as comparison operations. These augmented datasets are fed into fine-tuned encoder-only transformers, with BERT as the primary beneficiary, and the effect is measured through F1-macro, pairwise Kruskal-Wallis tests, confusion matrices, and SHAP explanations.

What would settle it

Take the augmented training set, have experts or a strong label model judge whether each generated sample still belongs to its original fine-grained class, then retrain BERT on only the verified subset and compare. If the 6% minority-class gain and the significant p-values disappear or shrink to insignificance, the reported effect is driven by label drift rather than by augmentation itself.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that contextual word insertion, applied only to minority classes, gives BERT a small but statistically significant edge on fine-grained food hazard and product classification. In the official test set, BERT with CW augmentation raised correct predictions on minority hazard samples from 261 to 277 (about 6%), while hazard-category minority predictions rose from 39 to 41; Kruskal-Wallis comparisons of three training runs found $P<0.05$ for the fine-grained hazard and product categories. The same benefit did not appear consistently for coarse categories or for other models. The paper interprets this as evidence that targeted augmentation of minority classes can improve transformer performance, provided the augmented text preserves the class label.

Load-bearing premise

The augmented sentences are assumed to keep the original hazard or product label, but no human or automatic check verifies this, and the paper's own examples show contextual insertion can alter meaning or add unrelated words; if these synthetic samples carry wrong labels, the apparent minority-class gains could be an artifact of label noise.

Editorial extensions

If this is right

Targeted minority-class augmentation with contextual word insertion is a viable, low-cost way to improve BERT on the fine-grained labels that matter most in food recall monitoring.
Augmentation of minority classes only—rather than the whole dataset—can shift performance toward rare hazards and products without architectural changes.
Overall F1-macro can stay flat while minority accuracy improves, so evaluations of augmentation should report minority-class and per-class metrics separately.
The benefit is not guaranteed across models or categories; each deployment should validate augmentation per model and label granularity before relying on it.

Reading between the lines

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

Beyond the paper, a natural next experiment is to add a label-preservation filter (human or LLM-based) to the augmented samples; if the gain survives filtering, the mechanism is robust, and if it vanishes, label noise was the driver.
Beyond the paper, the same oversampling recipe could transfer to other long-tailed regulatory corpora, such as medical device or cosmetic recall notices, with the same caveat that meaning-preservation must be checked per class.
Beyond the paper, the observed majority-class trade-off suggests a hybrid schedule—augment minority classes while down-weighting or partially up-weighting majority classes—might recover lost majority accuracy while keeping minority gains.
Beyond the paper, because the authors note future work on LLM-generated samples and quality verification, a direct comparison of contextual insertion against LLM paraphrasing on the same minority classes would locate where the cheap method's ceiling is.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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

Colleague,

The paper is a solid, honest shared-task system description whose load-bearing claim—statistically significant gains from contextual insertion for BERT on fine-grained minority classes—does not survive a minimal multiple-testing correction. The observed F1 gains (hazard 0.581→0.671, product 0.170→0.280) are real and could be practically useful, but the abstract's P<0.05 is not established.

What's actually new: the specific application of three standard augmentation techniques (synonym replacement, random swap, contextual insertion) to minority classes in the SemEval-2025 food hazard task, with a careful comparison across six ML models and four transformers. The negative result—augmentation doesn't consistently improve overall performance—is worth reporting. The paper does many things well: code is public, evaluation is against an external test set, hyperparameters and seeds are documented, and the confusion-matrix analysis honestly shows a majority/minority trade-off.

The soft spots, in order of importance. First, the significance claim. The p-values in Table 3 (0.0495, 0.0463) are at the floor of a 3-vs-3 Kruskal-Wallis test; with three runs per condition, complete separation of F1 scores gives H=3.857, p≈0.0495. Twelve comparisons are made with no multiple-testing correction. Under Holm-Bonferroni, the smallest p becomes ≈0.556, so nothing survives. The text in Section 6.1 also contradicts Table 3: it says product-category was significant with RW, but Table 3 lists RW p=0.3758; the significant coarse entry is SR (p=0.0463). That mismatch needs fixing. Second, the label-preservation concern is plausible: Section 4.3 and Algorithm 1 generate augmented samples without verification, and Table 1 shows semantically altered titles. This could introduce label noise, but it's not fatal—the test-set improvement is still a measured effect. The statistical overclaim is the bigger problem. Third, Table 2 shows single-run scores without variance for most models; that's minor for a system description, but a sentence acknowledging it would help.

Who gets value: researchers working on imbalanced text classification, data augmentation, or food safety NLP. It's a modest, honest contribution, not a breakthrough. I'd accept it for peer review conditional on revision: correct the significance analysis, fix the internal inconsistency, and soften the abstract to 'suggestive' rather than 'statistically significant.'

Referee Report

3 major / 5 minor

Summary. This paper describes the BrightCookies system for SemEval-2025 Task 9 (food hazard detection). The authors compare six classical ML classifiers and four encoder-only transformers on four classification targets (coarse/fine hazard and coarse/fine product). For each model family they train a baseline and three augmented variants in which minority classes are expanded by synonym replacement, random word swapping, or contextual word insertion. They report F1-macro on the official test set, confusion matrices, SHAP analyses, and a set of pairwise Kruskal-Wallis tests for BERT. The headline claims are that transformers outperform classical models, augmentation does not consistently improve overall performance, and that for BERT the fine-grained minority-class gains are statistically significant at P < 0.05.

Significance. If the headline claims were fully supported, the paper would be a useful empirical data point for imbalanced multi-class classification in a safety-critical domain. Its strengths are the breadth of configurations (10 model types, 4 categories, 3 augmentations), public code and seed specification, detailed hyperparameter tables, and evaluation on the shared task's held-out test set. The SHAP error analysis is a useful complement. However, the core statistical claim is currently not established: the reported p-values are at the floor of a 3-vs-3 Kruskal-Wallis comparison and do not survive a minimal multiple-testing correction. The observed F1 gains are real numbers, but they should be described as suggestive rather than statistically verified.

major comments (3)

[Abstract, Section 6.1, Table 3] The central claim of a statistically significant improvement (P < 0.05) is not supported by the reported tests. Table 3 reports 12 pairwise Kruskal-Wallis comparisons (3 augmentations x 4 categories), each based on only three runs per condition. The fine-grained p-values (0.0463, 0.0495) are at the floor of what a 3-vs-3 asymptotic Kruskal-Wallis test can produce; complete separation of the six scores gives H = 3.857 and p approximately 0.0495. No multiple-testing correction is applied. Under Holm-Bonferroni, the smallest p-value (0.0463) becomes approximately 0.556, so none of the comparisons remain significant at alpha = 0.05. The abstract and conclusion should be revised to state that the gains are suggestive rather than statistically established, or the authors should provide a more powerful test with appropriate correction.
[Section 6.1, Table 3] The text misreports the contents of Table 3. It says that statistical significance was found in product-category with RW, but Table 3 lists p = 0.3758 for product-category RW; the only significant coarse-category entry is product-category SR with p = 0.0463. This is a factual inconsistency in the main results narrative and should be corrected.
[Section 4.3, Algorithm 1, Section 7] The augmentation procedure assumes that synthetic samples preserve the annotated labels, but this assumption is not verified. Section 4.3 says the modifications are made while maintaining the annotated classes, and Algorithm 1 copies the original label, yet no automatic or manual check is reported. The examples in Table 1 show contextual word insertion producing phrases such as 'may be judged unsafe' and 'his possible presence', which are semantically altered and could shift the label distribution. Since the paper's main claim is that targeted minority-class augmentation improves classification, possible label drift is a confound. Please provide label-preservation checks (e.g., sampling-based verification or a small human evaluation) or explicitly weaken the causal interpretation.

minor comments (5)

[Table 2] Only the BERT models are run with multiple seeds; all other entries in Table 2 are single runs. Without variance estimates, the cross-model comparisons (e.g., LRCW vs RFRW) should be interpreted cautiously, and the single-seed nature should be stated next to the table.
[Abstract, Section 6.2] The abstract's claim that contextual word insertion 'improved the accuracy of predictions for the minority hazard classes by 6%' is based on confusion-matrix counts (39 to 41 correct for hazard-category and 261 to 277 for hazard). This is a count increase rather than a formal accuracy metric; please clarify the wording.
[Table 6] The 75% row in the product-category block appears to contain only three numbers ('513.5 263.5 333.25') instead of four aligned columns. Please fix the formatting so the Initial and Augmented columns are clear.
[Section 6.3] The SHAP analysis draws conclusions from one correctly classified and one misclassified example per class. These examples are anecdotal; please add a caveat that they are illustrative rather than systematic.
[Section 7] The thresholds and numbers of added samples (200, 100, and 50) are described as a compromise after inspecting class distributions, but no sensitivity analysis is reported. A brief discussion of how results could change with these hyperparameters would strengthen the paper.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the augmentation experiments are evaluated on the official held-out test set and the claims do not reduce to fitted inputs.

full rationale

The paper's central claim is that word-level augmentation of minority classes can improve fine-grained F1-macro scores for BERT, measured against the task's official test set (Sections 5.3 and 6.1). The augmented training data are generated from the training split with standard libraries (nlpaug), and the comparison is between independently trained baseline and augmented models; no parameter fitted to the test set is renamed as a prediction, and no prediction is defined in terms of its own output. The cited dataset and task (Randl et al. 2024/2025) are external resources, not self-citations by the present authors, and no load-bearing argument reduces to a prior work by the same team. The weakness flagged by the skeptic—p-values from three-run Kruskal-Wallis tests that do not survive multiple-testing correction, and the unverified label preservation of augmented samples—concerns statistical robustness and data quality, not circularity. Accordingly, no circular steps are identified and the score is 0.

Assumptions & free parameters 6 free parameters · 3 assumptions · 0 invented entities

The paper contributes no new mathematical objects. Its central intervention, minority-class augmentation, depends on hand-selected augmentation budgets, including thresholds and per-class sample counts, and on the unverified assumption that altered texts retain their original labels.

free parameters (6)

Coarse augmentation threshold tau = 200
Classes with fewer than 200 samples are augmented; chosen by hand as a compromise in Section 4.3, not derived from data.
Fine hazard augmentation threshold = 100
For the hazard category, classes with fewer than 100 samples are augmented; chosen ad hoc in Section 4.3.
Fine product augmentation threshold = 50
For the product category, classes with fewer than 50 samples are augmented; chosen ad hoc in Section 4.3.
Samples added per coarse minority class = 200
Two hundred synthetic samples are generated for each coarse minority class, as stated in Section 4.3.
Samples added per fine hazard class = 100
One hundred synthetic samples are generated for each fine hazard minority class, as stated in Section 4.3.
Samples added per fine product class = 50
Fifty synthetic samples are generated for each fine product minority class, as stated in Section 4.3.

assumptions (3)

domain assumption The annotated labels in the Food Recall Incidents dataset are correct and consistent.
The paper relies on expert labels for training and evaluation; no label noise is modeled in Section 3.
domain assumption Word-level augmentation using nlpaug preserves the semantic class of each sample.
Section 4.3 generates synthetic samples without validating whether the hazard or product label still applies; Table 1 shows examples where the meaning is altered.
domain assumption F1-macro computed on the leaderboard is the appropriate objective for hyperparameter selection.
Section 5.2 tunes hyperparameters to maximize category-level F1-macro on the development set.

how reviews work

0 comments

Cite this review

Pith. "Pith review of BrightCookies at SemEval-2025 Task 9: Exploring Data Augmentation for Food Hazard Classification." pith.science (2026). https://pith.science/paper/5I2KD726

@misc{pith2026250420703,
  author       = {Pith},
  title        = {Pith review of: BrightCookies at SemEval-2025 Task 9: Exploring Data Augmentation for Food Hazard Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5I2KD726}},
  note         = {Machine review of arXiv:2504.20703}
}

read the original abstract

This paper presents our system developed for the SemEval-2025 Task 9: The Food Hazard Detection Challenge. The shared task's objective is to evaluate explainable classification systems for classifying hazards and products in two levels of granularity from food recall incident reports. In this work, we propose text augmentation techniques as a way to improve poor performance on minority classes and compare their effect for each category on various transformer and machine learning models. We explore three word-level data augmentation techniques, namely synonym replacement, random word swapping, and contextual word insertion. The results show that transformer models tend to have a better overall performance. None of the three augmentation techniques consistently improved overall performance for classifying hazards and products. We observed a statistically significant improvement (P < 0.05) in the fine-grained categories when using the BERT model to compare the baseline with each augmented model. Compared to the baseline, the contextual words insertion augmentation improved the accuracy of predictions for the minority hazard classes by 6%. This suggests that targeted augmentation of minority classes can improve the performance of transformer models.

Figures

Figures reproduced from arXiv: 2504.20703 by the authors.

**Figure 2.** Confusion matrices comparing the perfor [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Visualizations of SHAP values for samples in the ground-truth classes of [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Distributions of hazard-category and product-category for classes occurrences [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]

**Figure 5.** Figure 5: Distributions of hazard and product for classes occurrences. The classes in the x-axis have been omitted due to the large number of classes and clearness of the chart [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]

**Figure 6.** Figure 6: Distributions of occurrences per country (left figure) and per year (right figure) published in the given [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references · 11 canonical work pages

[4]

Preprint, arXiv:1907.11692

Roberta: A robustly optimized bert pretraining ap- proach. Preprint, arXiv:1907.11692. Edward Ma

arXiv 1907
[5]

A.4 Transformer Models Details In this section, we explain the encoder-only trans- former models’ details and architectures we used in the experiments

Library Version URL Transformers 4.49.0 https://huggingfac e.co/docs/transfor mers/index PyTorch 2.6.0 https://pytorch.or g/ SpaCy 3.8.4 https://spacy.io/ Scikit-learn 1.6.0 https://scikit-lea rn.org/stable/ Pandas 2.2.3 https://pandas.pyd ata.org/ Optuna 4.2.1 https://optuna.org / NumPy 2.0.2 https://numpy.org/ NLP AUG 1.1.11 https://nlpaug.rea dthedocs....

work page 2019
[8]

Preprint, arXiv:1706.03762

Attention is all you need. Preprint, arXiv:1706.03762. Benjamin Warner, Antoine Chaffin, Benjamin Clavié, Orion Weller, Oskar Hallström, Said Taghadouini, Alexis Gallagher, Raja Biswas, Faisal Ladhak, Tom Aarsen, Nathan Cooper, Griffin Adams, Jeremy Howard, and Iacopo Poli

arXiv
[9]

Preprint, arXiv:2412.13663

Smarter, better, faster, longer: A modern bidirectional encoder for fast, memory efficient, and long context finetuning and inference. Preprint, arXiv:2412.13663. Jason Wei and Kai Zou

arXiv
[10]

EDA: Easy data augmen- tation techniques for boosting performance on text classification tasks. In Proceedings of the 2019 Con- ference on Empirical Methods in Natural Language Processing and the 9th International Joint Confer- ence on Natural Language Processing (EMNLP- IJCNLP), pages 6382–6388, Hong Kong, China. As- sociation for Computational Linguisti...

work page 2019
[11]

In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38–45, Online

Trans- formers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38–45, Online. Association for Computational Linguistics. World Health Organization

work page 2020
[12]

Ta- ble 11 shows some sample titles and text from the dataset along with their annotated classes

A Dataset and Experiments Details A.1 Dataset Details In this section, tables and figures related to the statistics of the provided dataset are presented. Ta- ble 11 shows some sample titles and text from the dataset along with their annotated classes. Table 10 presents the number and the names of the annotated classes. Figure 4 and Figure 5 show the dist...

work page 2024
[14]

With bold, we indicate the higher score per column

B.3 Official Submitted Models Since only one submission was allowed during the evaluation phase, the predictions of the models that Model hazard- category product- category hazard productST1 ST2 SVMbase 0.644 0.692 0.436 0.250 0.670 0.363 SVMCW 0.641 0.675 0.402 0.240 0.657 0.343 SVMSR 0.646 0.699 0.435 0.259 0.674 0.364 SVMRW 0.646 0.690 0.432 0.253 0.67...

work page 2024

Show all 15 references

[1995]

Wordnet: a lexical database for english. Commun. ACM, 38(11):39–41. Korbinian Randl, Manos Karvounis, George Marinos, John Pavlopoulos, Tony Lindgren, and Aron Henriks- son. 2024a. Food recall incidents. Korbinian Randl, John Pavlopoulos, Aron Henriksson, and Tony Lindgren. 20...

2024
[2019]

BERT: Pre-training of deep bidirectional transformers for language under- standing. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Tech- nologies, V olume 1 (Long and Short Papers), pages 417...

2019
[2020]

Preprint, arXiv:1910.01108

Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. Preprint, arXiv:1910.01108. Connor Shorten, Taghi M. Khoshgoftaar, and Borko Furht

1910 arXiv
[2023]

In Proceedings of the 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1421–1427, Toronto, Canada

NLP-LTU at SemEval-2023 task 10: The impact of data aug- mentation and semi-supervised learning techniques on text classification performance on an imbalanced dataset. In Proceedings of the 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1421–1427, Tor...

2023
[2024]

In 2024 3rd Edition of IEEE Delhi Section Flagship Conference (DELCON), pages 1–8

Explor- ing deep learning methods for text augmentation to handle imbalanced datasets in natural language pro- cessing. In 2024 3rd Edition of IEEE Delhi Section Flagship Conference (DELCON), pages 1–8. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Man- dar Joshi, Danqi Chen,...

2024
[2025]

In Proceedings of the 19th International Workshop on Semantic Evaluation (SemEval-2025), Vienna, Aus- tria

SemEval- 2025 task 9: The food hazard detection challenge. In Proceedings of the 19th International Workshop on Semantic Evaluation (SemEval-2025), Vienna, Aus- tria. Association for Computational Linguistics. Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf

2025
[2026]

Then, these models were trained in both train and dev sets and provided their predictions on the test set

were submitted and were found to have the bestF1- macro scores on the development set for each cate- gory are: RoBERTa base for hazard-category with 0.880 F1-macro score, RoBERTa RW for product-category with 0.750 F1-macro score, BERT CW for hazard with 0.682F1-macro score, BE...

1994

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.

[1] [4]

Preprint, arXiv:1907.11692

Roberta: A robustly optimized bert pretraining ap- proach. Preprint, arXiv:1907.11692. Edward Ma

arXiv 1907

[2] [5]

A.4 Transformer Models Details In this section, we explain the encoder-only trans- former models’ details and architectures we used in the experiments

Library Version URL Transformers 4.49.0 https://huggingfac e.co/docs/transfor mers/index PyTorch 2.6.0 https://pytorch.or g/ SpaCy 3.8.4 https://spacy.io/ Scikit-learn 1.6.0 https://scikit-lea rn.org/stable/ Pandas 2.2.3 https://pandas.pyd ata.org/ Optuna 4.2.1 https://optuna.org / NumPy 2.0.2 https://numpy.org/ NLP AUG 1.1.11 https://nlpaug.rea dthedocs....

work page 2019

[3] [8]

Preprint, arXiv:1706.03762

Attention is all you need. Preprint, arXiv:1706.03762. Benjamin Warner, Antoine Chaffin, Benjamin Clavié, Orion Weller, Oskar Hallström, Said Taghadouini, Alexis Gallagher, Raja Biswas, Faisal Ladhak, Tom Aarsen, Nathan Cooper, Griffin Adams, Jeremy Howard, and Iacopo Poli

arXiv

[4] [9]

Preprint, arXiv:2412.13663

Smarter, better, faster, longer: A modern bidirectional encoder for fast, memory efficient, and long context finetuning and inference. Preprint, arXiv:2412.13663. Jason Wei and Kai Zou

arXiv

[5] [10]

EDA: Easy data augmen- tation techniques for boosting performance on text classification tasks. In Proceedings of the 2019 Con- ference on Empirical Methods in Natural Language Processing and the 9th International Joint Confer- ence on Natural Language Processing (EMNLP- IJCNLP), pages 6382–6388, Hong Kong, China. As- sociation for Computational Linguisti...

work page 2019

[6] [11]

In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38–45, Online

Trans- formers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38–45, Online. Association for Computational Linguistics. World Health Organization

work page 2020

[7] [12]

Ta- ble 11 shows some sample titles and text from the dataset along with their annotated classes

A Dataset and Experiments Details A.1 Dataset Details In this section, tables and figures related to the statistics of the provided dataset are presented. Ta- ble 11 shows some sample titles and text from the dataset along with their annotated classes. Table 10 presents the number and the names of the annotated classes. Figure 4 and Figure 5 show the dist...

work page 2024

[8] [14]

With bold, we indicate the higher score per column

B.3 Official Submitted Models Since only one submission was allowed during the evaluation phase, the predictions of the models that Model hazard- category product- category hazard productST1 ST2 SVMbase 0.644 0.692 0.436 0.250 0.670 0.363 SVMCW 0.641 0.675 0.402 0.240 0.657 0.343 SVMSR 0.646 0.699 0.435 0.259 0.674 0.364 SVMRW 0.646 0.690 0.432 0.253 0.67...

work page 2024