Towards Fairness Assessment of Dutch Hate Speech Detection

Adriana Iamnitchi; Julie Bauer; Rishabh Kaushal; Thales Bertaglia

REVIEW 5 major objections 5 minor 13 references

Towards Fairness Assessment of Dutch Hate Speech Detection

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that counterfactual data augmentation—swapping social-group terms in Dutch sentences—improves the performance and group fairness of Dutch hate speech classifiers, while making counterfactual fairness worse for the most…

desk verdict Solid first step for Dutch counterfactual fairness, but the headline average CTF gain rests on a 34-template eval set and an unvalidated label mapping. read the letter →

arxiv 2506.12502 v1 pith:3IFM62RA submitted 2025-06-14 cs.CL cs.AIcs.CY

Julie Bauer , Rishabh Kaushal , Thales Bertaglia , Adriana Iamnitchi This is my paper

classification cs.CLcs.AIcs.CY

keywords counterfactualfairnesshatespeechdetectionDutchlanguageBERTjedataaugmentationgroupsocialtermstransformermodels

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

This paper asks whether Dutch hate speech detectors treat sentences targeting different social groups the same way, and whether that can be fixed. To answer, the authors build counterfactual sentences—rewrites in which a social-group term is swapped for another, such as 'Marokkanen' for 'Nederlanders'—using two LLM prompting methods, sentence log-likelihood (SLL), and manual group substitution (MGS). They fine-tune the Dutch BERT model BERTje on each augmented dataset and report that SLL and MGS improve hate speech detection (F1 from 0.75 to 0.79), sharply improve demographic parity difference (0.38 to 0.06 for SLL) and equalized odds difference (0.53 to 0.11), and lower average counterfactual token fairness from 0.24 to 0.10. The same augmentation worsens counterfactual fairness for toxic sentences, rising from 0.11 to 0.20 for SLL, a regression the paper flags as counterintuitive and unresolved. The paper's overall claim is that counterfactual data augmentation is a workable path toward fairer Dutch hate speech detection, with a known failure mode on the most harmful sentences.

What carries the argument

The load-bearing object is the counterfactual pair: an original sentence containing a social group term and a rewritten version with that term swapped for another group, everything else held fixed. On top of these pairs, the paper computes Counterfactual Token Fairness (CTF), the average absolute difference in predicted labels between originals and counterfactuals, plus demographic parity difference and equalized odds difference across identity-group categories. The second mechanism is fine-tuning: BERTje, a monolingual Dutch transformer, is fine-tuned on each counterfactually augmented corpus, and the fairness numbers are measured before and after that fine-tuning.

What would settle it

Construct a larger Dutch counterfactual evaluation set—for instance, 100 or more templates written and toxicity-annotated by native Dutch speakers, with no reliance on the appropriate/inappropriate versus offensive/violent mapping—and rerun the same five models on it. If the reported SLL and MGS gains in demographic parity, equalized odds, and average CTF do not reappear, or if the toxic-class CTF regression is much larger, the paper's central claim that counterfactual augmentation improves fairness in Dutch hate speech detection would fail; if they do reappear, the claim is corroborated.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that counterfactual data, even when generated with methods that visibly break Dutch grammar, can make a Dutch hate speech classifier fairer on average and across groups while slightly improving accuracy. The authors curate a Dutch Social Group Term list of 85 terms, generate over 100,000 counterfactual sentences, fine-tune BERTje on each generated corpus, and evaluate with Counterfactual Token Fairness (CTF), demographic parity difference, and equalized odds difference. Their headline numbers are that SLL and MGS fine-tuning lift F1 from 0.75 to 0.79, cut demographic parity difference from 0.38 to 0.06, cut equalized odds difference from 0.53 to 0.11, and reduce average CTF from 0.24 to 0.10, while CTF on the toxic class gets worse, from 0.11 to 0.20. The paper presents this toxic-class result as counterintuitive and calls for more research, while reporting the overall direction as improved fairness with maintained performance.

Load-bearing premise

The fairness evaluation rests on a hand-built test set of 34 English templates (17 toxic, 17 non-toxic) translated into Dutch, with toxicity labels carried over from the English source, and on a labeling rule that maps the model's four output classes to toxic/non-toxic; if those templates, labels, or the mapping do not represent real Dutch hate speech, the fairness conclusions do not follow.

Editorial extensions

If this is right

If the claim holds, counterfactual augmentation is a ready-to-use recipe for Dutch hate speech systems: no architectural change, just add generated swapped-group sentences to the training data.
The finding that noisy SLL and MGS data help more than LLM-generated data suggests grammatical realism is not required for fairness gains, which changes how practitioners should spend generation effort.
The toxic-class CTF regression implies that average fairness metrics can hide worse behavior on the sentences that matter most, so deployment monitoring should report toxic-class fairness separately.
The released Dutch counterfactual datasets give later researchers a reusable evaluation and augmentation resource, filling a gap noted in the paper.

Reading between the lines

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

If the four-to-two label mapping drives the numbers, part of the apparent fairness gain could be an artifact of training data that shifts decisions between the 'inappropriate' and 'offensive' classes; testing with native-speaker labels on the same templates would separate artifact from genuine fairness improvement.
The near-zero non-toxic CTF after augmentation suggests the models learn to ignore social group terms in safe sentences but still key on them in toxic ones; a class-balanced augmentation scheme that generates only toxic counterfactuals might reduce the toxic-class regression.
The SLL and MGS data contain grammatical noise, so the fairness gain may come from regularization rather than from the counterfactual semantics; if so, other noise-injection methods could produce similar fairness at lower generation cost.

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

I'll keep it short. If you look at one thing, know this: the paper is the first to evaluate counterfactual fairness for Dutch hate speech detection, and it releases the first Dutch counterfactual datasets. That part is real and worth building on. But the headline "models perform better in terms of average counterfactual fairness" is fragile because the entire fairness evaluation is measured on 2,890 sentences built from 34 manually translated English templates (17 toxic, 17 non-toxic), with toxicity labels assumed to transfer from Dixon et al. and a four-class-to-binary mapping that is never validated. A human-annotation check could shift the numbers and possibly the rankings.

What the paper does well: it curates a Dutch Social Group Term list (85 terms), generates counterfactuals with four methods (LLMdef, LLMlist, SLL, MGS), and honestly documents the grammar and coherence failures of SLL and MGS in Dutch. That qualitative analysis is genuinely useful for anyone doing data augmentation in Dutch. The group fairness metrics (DPD and EOD) improve consistently across all four counterfactual models, with SLL showing the largest gains (DPD 0.38→0.06, EOD 0.53→0.11). The paper also flags in Section 7 that CTF for the toxic class worsens across all counterfactual models, which is the right kind of honesty.

The soft spots: the evaluation set is narrow. Thirty-four templates, even translated carefully, cannot represent Dutch hate speech. The binary mapping (appropriate/inappropriate → non-toxic; offensive/violent → toxic) is a reasonable default but it is not checked against human judgments, and all CTF numbers depend on it. There are also no error bars or significance tests anywhere, so the differences between models (e.g., SLL vs MGS on DPD) may or may not be meaningful. The abstract overstates the case: average CTF improves mostly because non-toxic CTF drops to near zero, while toxic-class CTF gets worse. The paper itself admits this, so it's a framing problem, not a hidden one.

The central claim—that counterfactual augmentation improves group fairness while roughly maintaining performance—is plausible and useful for the Dutch NLP community. The toxic-class CTF result is a genuine empirical finding worth exploring. I'd send this to peer review and ask for a validation of the eval set (even a small human-annotation study on the 34 templates), error bars or bootstrap confidence intervals, and a more nuanced abstract.

Referee Report

5 major / 5 minor

Summary. The paper evaluates counterfactual fairness of Dutch hate speech detection models. The authors curate an 85-term Dutch Social Group Terms list, generate counterfactual training data using LLM prompting (LLMdef, LLMlist), Sentence Log-Likelihood (SLL), and Manual Group Substitution (MGS), and fine-tune BERTje with each dataset. They evaluate classification performance on the IMSyPP test set and fairness on a synthetic Dutch counterfactual evaluation set, using Counterfactual Token Fairness (CTF), demographic parity difference, and equalized odds difference. Their main empirical claim is that SLL and MGS-based fine-tuning improve both performance and average fairness, while LLM-generated data does not improve performance.

Significance. This is a timely and useful contribution to non-English fairness evaluation. The paper releases the first Dutch counterfactual fairness datasets for hate speech, provides a manually curated Dutch SGT list, and is unusually honest about the limitations of the counterfactual generation methods, with explicit qualitative examples of grammatical and semantic failures. The open availability of the data (URL in Section 7) and the explicit reporting of negative results (toxic-class CTF worsening) are strengths that support the paper's integrity. If the methodological concerns about the evaluation set and metric definition are resolved, the empirical findings would be a valuable benchmark for future work on Dutch hate speech fairness.

major comments (5)

[Section 6, Eq. (3)] The CTF definition in Eq. (3) sums over all x in X and all x' in Xcf without restricting x' to be a counterfactual perturbation of x. In Garg et al. (2019), CTF is computed by pairing each original instance with its counterfactual variants; summing over all pairs of unrelated sentences would measure something closer to pairwise disagreement across the whole evaluation set and could artificially inflate the values. Please clarify the pairing (e.g., define Xcf(x) as the counterfactual set for x) and report how the 2,890 sentences were grouped for the computation.
[Section 6] The Dutch counterfactual evaluation dataset is constructed from 17 toxic and 17 non-toxic English templates that are manually translated, and every fairness result in Tables 4-6 depends on transferring the Dixon et al. (2018) toxicity labels to Dutch and on mapping the four IMSyPP classes to binary labels ('appropriate' and 'inappropriate' as non-toxic, 'offensive' and 'violent' as toxic). No evidence is provided that the translations preserve toxicity or that this mapping is justified for Dutch hate speech. A human annotation study of a sample of the 2,890 sentences and a sensitivity analysis of the label mapping (e.g., treating 'inappropriate' as toxic) are necessary to establish that the reported fairness improvements are not an artifact of these assumptions.
[Abstract and Section 7] The abstract's claim that models 'perform better in terms of hate speech detection, average counterfactual fairness and group fairness' is contradicted by the paper's own results for some methods: in Table 2, LLMdef and LLMlist reduce F1 to 0.61 from the baseline 0.75, and in Table 4 the toxic-class CTF worsens for every counterfactual model (e.g., from 0.11 for the baseline to 0.20 for SLL). The average CTF improvement is entirely driven by non-toxic templates. The abstract and conclusions should present this trade-off explicitly, and the 'models' claim should be restricted to the SLL and MGS variants.
[Tables 1, 8-12] The dataset counts in the main text and appendices are inconsistent. Section 3.2 reports 15,175 LLMdef posts, which matches Table 1, but the sum of Table 9 is 16,365; Table 1 lists 21,562 LLMlist posts but Table 10 sums to 21,349; Table 8 sums to 2,829 while its caption reports 2,825. These discrepancies undermine the reliability of the dataset documentation and must be reconciled.
[Section 3.2] The LLM-based counterfactual generation methods do not specify which large language model was used (e.g., GPT-3.5, GPT-4, Llama, or a specific version), nor the decoding parameters. Since the qualitative and downstream results depend on the model's behavior, this is a reproducibility gap that should be filled by naming the exact model and providing the generation configuration.

minor comments (5)

[Table 3] The header of the fifth column reads 'SSL' but should be 'SLL'.
[Section 2 and References] The in-text citation 'Vries (2024)' should be consistent with the reference list entry 'R Vries' by including the author's initials in the citation (e.g., de Vries, 2024) and formatting the reference properly.
[Section 6] The mapping from four IMSyPP classes to binary toxic/non-toxic is stated but not justified; adding a brief rationale or a citation to prior work that collapses these labels would help.
[Appendix C] The caption totals in Tables 8-12 do not match the row sums in several cases; please add a consistency check and correct the numbers.
[Section 3.2] The paper does not state which LLM was used for the LLMdef and LLMlist methods, which is essential for reproducibility; please name the model and version.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: fairness metrics are computed on a held-out template set, not fitted; the binary labeling and template translation are validity assumptions, not circular reductions.

full rationale

The paper's derivation chain is not circular. The fairness evaluation dataset (Section 6) is a held-out set of 2,890 sentences built from 34 manually translated templates; it is not used to train any model, and no parameter is fitted to it. The binary mapping from four model classes to toxic/non-toxic is an explicit modeling choice that affects the absolute CTF/DPD/EOD values, but it is not defined in terms of the conclusions, and Table 4 reports toxic and non-toxic CTF separately, allowing the reader to see that the average improvement is driven by the non-toxic half. The SLL/MGS/LLM counterfactual data are generated from the IMSyPP training posts, and the fine-tuned models are then evaluated on the separate template-based counterfactual set; this is a standard train/eval split rather than a self-fulfilling construction. The only self-citation (Bertaglia et al., 2021) appears in a related-work list and does not carry any load-bearing premise. No uniqueness theorem or ansatz is imported from the authors' prior work. Therefore the empirical claims, while dependent on the unvalidated label-transfer assumption from Dixon et al., are not circular.

Assumptions & free parameters 4 free parameters · 4 assumptions · 2 invented entities

The central claims rest on a set of manual curation choices (SGT list, template translations, label mapping) and on the assumption that English-derived fairness definitions and toxicity labels transfer to Dutch. None of these are parameter-free derivations; they are modeling decisions that directly shape every reported number. The paper does not provide external validation for any of them.

free parameters (4)

Toxic/non-toxic label mapping for the four IMSyPP classes = appropriate and inappropriate mapped to non-toxic; offensive and violent mapped to toxic
This mapping (Section 6) defines the binary labels used in every CTF, DPD, and EOD calculation. If the mapping is wrong, every fairness number changes.
Selection of 85 Dutch Social Group Terms = 85 terms across 7 categories, with only one adjective conjugation form and no plural nouns
The SGT list (Appendix A) determines which posts enter the baseline data (2,649 posts) and which substitutions are allowed. The choice to exclude plural forms and most conjugation variants is a manual curation decision that shapes all downstream data.
Threshold for SLL counterfactual acceptance = counterfactual accepted if log-likelihood >= original sentence
Section 3.3 states only counterfactuals with higher or equal log-likelihood are kept. The paper does not report what fraction were rejected or how sensitive the results are to this threshold.
Template toxicity labels from Dixon et al. 2018 = 17 toxic and 17 non-toxic templates
The fairness evaluation set is built from these templates, translated to Dutch. The paper assumes the English toxicity labels transfer to the Dutch translations without re-validation by Dutch annotators.

assumptions (4)

domain assumption Counterfactual fairness requires equal prediction distributions across counterfactual variants of sensitive attributes.
This definition (Section 1) imports the causal fairness framework of Kusner et al. 2017 into text classification without a causal graph; the paper treats sentence-level token substitution as the counterfactual intervention.
domain assumption The IMSyPP labels 'offensive' and 'violent' constitute hate speech, and 'appropriate' and 'inappropriate' do not.
Stated in Section 3.1 and used again in Section 6 to convert four-way predictions into binary toxic/non-toxic. This is a contested labeling choice that directly affects all fairness metrics.
domain assumption Sentence log-likelihood from GPT-2 is a valid proxy for the plausibility of a Dutch counterfactual.
Section 3.3 uses GPT-2 to score Dutch sentences despite GPT-2 being primarily an English model. The paper's own qualitative results (Section 4.1) show that SLL selects sentences with incorrect Dutch grammar, undermining this proxy's validity.
domain assumption A fair model should treat structurally identical sentences with different group terms equally.
This is the core normative premise of the paper (Section 1 example with Moroccans vs Dutch). It is assumed, not argued for, and it underpins the entire CTF evaluation.

invented entities (2)

Dutch Social Group Terms list (85 terms)
purpose: A curated inventory of Dutch identity terms used to filter the dataset and to define counterfactual substitutions.
The list is manually curated by the authors (Section 3.1, Appendix A) and has no external validation, such as inter-annotator agreement or a comparison with an independently constructed Dutch group term list.
Dutch counterfactual evaluation dataset (2,890 sentences)
purpose: A held-out template-based set used to compute CTF, DPD, and EOD fairness metrics.
The dataset is constructed by the authors from 34 translated English templates and the SGT list. The toxicity labels are inherited from Dixon et al. 2018 without Dutch re-annotation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Towards Fairness Assessment of Dutch Hate Speech Detection." pith.science (2026). https://pith.science/paper/3IFM62RA

@misc{pith2026250612502,
  author       = {Pith},
  title        = {Pith review of: Towards Fairness Assessment of Dutch Hate Speech Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3IFM62RA}},
  note         = {Machine review of arXiv:2506.12502}
}

read the original abstract

Numerous studies have proposed computational methods to detect hate speech online, yet most focus on the English language and emphasize model development. In this study, we evaluate the counterfactual fairness of hate speech detection models in the Dutch language, specifically examining the performance and fairness of transformer-based models. We make the following key contributions. First, we curate a list of Dutch Social Group Terms that reflect social context. Second, we generate counterfactual data for Dutch hate speech using LLMs and established strategies like Manual Group Substitution (MGS) and Sentence Log-Likelihood (SLL). Through qualitative evaluation, we highlight the challenges of generating realistic counterfactuals, particularly with Dutch grammar and contextual coherence. Third, we fine-tune baseline transformer-based models with counterfactual data and evaluate their performance in detecting hate speech. Fourth, we assess the fairness of these models using Counterfactual Token Fairness (CTF) and group fairness metrics, including equality of odds and demographic parity. Our analysis shows that models perform better in terms of hate speech detection, average counterfactual fairness and group fairness. This work addresses a significant gap in the literature on counterfactual fairness for hate speech detection in Dutch and provides practical insights and recommendations for improving both model performance and fairness.

Figures

Figures reproduced from arXiv: 2506.12502 by the authors.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

13 extracted references · 12 canonical work pages

[3]

InProceedings of the 5th Workshop on Online Abuse and Harms (WOAH 2021), pages 92–101

Improving counterfactual genera- tion for fair hate speech detection. InProceedings of the 5th Workshop on Online Abuse and Harms (WOAH 2021), pages 92–101. Wietse De Vries, Andreas van Cranenburgh, Arianna Bisazza, Tommaso Caselli, Gertjan van Noord, and Malvina Nissim

work page 2021
[6]

In Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society, pages 219–226

Counterfactual fairness in text classification through robustness. In Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society, pages 219–226. Moritz Hardt, Eric Price, and Nati Srebro

work page 2019
[9]

Accessed: 2024-09-04

Hate speech detection and trends - imsypp multilingual hate speech database. Accessed: 2024-09-04. Dana Pessach and Erez Shmueli

work page 2024
[10]

People Make Better Edits: Measuring the Efficacy of LLM-Generated Counterfactually Augmented Data for Harmful Language Detection

People make better edits: Measuring the efficacy of llm-generated counterfactually aug- mented data for harmful language detection.arXiv preprint arXiv:2311.01270. Guðbjartur Sigurbergsson and Leon Derczynski

work page Pith review arXiv
[12]

InCOLING 2020-28th International Conference on Computational Linguis- tics, Proceedings of the Conference

Generating plausible counterfactual explanations for deep transformers in financial text classification. InCOLING 2020-28th International Conference on Computational Linguis- tics, Proceedings of the Conference. Wenjie Yin and Arkaitz Zubiaga

work page 2020
[13]

heteroseksueel

Towards gener- alisable hate speech detection: a review on obstacles and solutions.PeerJ Computer Science, 7:e598. A Categories of Social Group Terms in the Dutch Context Table 7 maps the social group terms from the Dutch context into seven categories. Table 7: Categorization of Social Group Terms. Category Values nationality turk, marokkaan, mocro, neder...

work page 1971
[2018]

InProceed- ings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society, pages 67–73

Measuring and mitigat- ing unintended bias in text classification. InProceed- ings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society, pages 67–73. Paula Fortuna and Sérgio Nunes

work page 2018
[2019]

Fabio Del Vigna, Andrea Cimino, Felice Dell’Orletta, Marinella Petrocchi, and Maurizio Tesconi

Bertje: A Dutch BERT model.arXiv preprint arXiv:1912.09582. Fabio Del Vigna, Andrea Cimino, Felice Dell’Orletta, Marinella Petrocchi, and Maurizio Tesconi

arXiv 1912

Show all 13 references

[2020]

InProceedings of the International Conference on Language Resources and Evaluation: LREC 2020, pages 3498–3508

Offensive language and hate speech detection for dan- ish. InProceedings of the International Conference on Language Resources and Evaluation: LREC 2020, pages 3498–3508. European Language Resources As- sociation. Malliga Subramanian, Veerappam- palayam Easwaramoorthy Sathisku...

2020
[2021]

In Proceedings of the 5th Workshop on Online Abuse and Harms (WOAH 2021), pages 191–200, Online

Abusive language on social media through the legal looking glass. In Proceedings of the 5th Workshop on Online Abuse and Harms (WOAH 2021), pages 191–200, Online. Association for Computational Linguistics. Tommaso Caselli, Arjan Schelhaas, Marieke Weultjes, Folkert Leistra, Hy...

2021
[2022]

In Proceedings of the Third Workshop on Threat, Ag- gression and Cyberbullying (TRAC 2022), October, Gyeongju, Republic of Korea, pages 37–42

The role of context in detecting the target of hate speech. In Proceedings of the Third Workshop on Threat, Ag- gression and Cyberbullying (TRAC 2022), October, Gyeongju, Republic of Korea, pages 37–42. Ilia Markov, Ine Gevers, and Walter Daelemans

2022
[2023]

In7th Workshop on Online Abuse and Harms, WOAH 2023, co-located with ACL 2023, pages 69–84

Benchmarking offensive and abusive language in Dutch tweets. In7th Workshop on Online Abuse and Harms, WOAH 2023, co-located with ACL 2023, pages 69–84. Association for Computational Linguis- tics, ACL Anthology. Simon Caton and Christian Haas

2023
[2024]

InCompanion Proceedings of the ACM Web Conference 2024, pages 1538–1545

Llm-guided counterfactual data genera- tion for fairer ai. InCompanion Proceedings of the ACM Web Conference 2024, pages 1538–1545. Nanlir Sallau Mullah and Wan Mohd Nazmee Wan Zainon

2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [3]

InProceedings of the 5th Workshop on Online Abuse and Harms (WOAH 2021), pages 92–101

Improving counterfactual genera- tion for fair hate speech detection. InProceedings of the 5th Workshop on Online Abuse and Harms (WOAH 2021), pages 92–101. Wietse De Vries, Andreas van Cranenburgh, Arianna Bisazza, Tommaso Caselli, Gertjan van Noord, and Malvina Nissim

work page 2021

[2] [6]

In Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society, pages 219–226

Counterfactual fairness in text classification through robustness. In Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society, pages 219–226. Moritz Hardt, Eric Price, and Nati Srebro

work page 2019

[3] [9]

Accessed: 2024-09-04

Hate speech detection and trends - imsypp multilingual hate speech database. Accessed: 2024-09-04. Dana Pessach and Erez Shmueli

work page 2024

[4] [10]

People Make Better Edits: Measuring the Efficacy of LLM-Generated Counterfactually Augmented Data for Harmful Language Detection

People make better edits: Measuring the efficacy of llm-generated counterfactually aug- mented data for harmful language detection.arXiv preprint arXiv:2311.01270. Guðbjartur Sigurbergsson and Leon Derczynski

work page Pith review arXiv

[5] [12]

InCOLING 2020-28th International Conference on Computational Linguis- tics, Proceedings of the Conference

Generating plausible counterfactual explanations for deep transformers in financial text classification. InCOLING 2020-28th International Conference on Computational Linguis- tics, Proceedings of the Conference. Wenjie Yin and Arkaitz Zubiaga

work page 2020

[6] [13]

heteroseksueel

Towards gener- alisable hate speech detection: a review on obstacles and solutions.PeerJ Computer Science, 7:e598. A Categories of Social Group Terms in the Dutch Context Table 7 maps the social group terms from the Dutch context into seven categories. Table 7: Categorization of Social Group Terms. Category Values nationality turk, marokkaan, mocro, neder...

work page 1971

[7] [2018]

InProceed- ings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society, pages 67–73

Measuring and mitigat- ing unintended bias in text classification. InProceed- ings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society, pages 67–73. Paula Fortuna and Sérgio Nunes

work page 2018

[8] [2019]

Fabio Del Vigna, Andrea Cimino, Felice Dell’Orletta, Marinella Petrocchi, and Maurizio Tesconi

Bertje: A Dutch BERT model.arXiv preprint arXiv:1912.09582. Fabio Del Vigna, Andrea Cimino, Felice Dell’Orletta, Marinella Petrocchi, and Maurizio Tesconi

arXiv 1912