Pith. sign in

REVIEW 3 major objections 6 minor 32 references

Large Language Models for Toxic Language Detection in Low-Resource Balkan Languages

T0 review · 3 major / 6 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 Adding a short video description to prompts raises toxic-comment recall by an average of twelve percentage points in low-resource Balkan languages.

desk verdict A useful new benchmark for Balkan toxicity detection, held back by a context-augmentation effect that may partly reflect the authors' own annotations. read the letter →

arxiv 2506.09992 v2 pith:ILWNGS73 submitted 2025-06-11 cs.CL

Amel Muminovic , Amela Kadric Muminovic This is my paper · ORCID
classification cs.CL
keywords toxiclanguagedetectionlargemodelslow-resourcelanguagesSerbianCroatianBosniancontext-augmentedpromptingsocialmediamoderation
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 giving a large language model a little context makes it better at spotting toxic comments in Serbian, Croatian, and Bosnian, three languages with little labeled training data and weak moderation tools. It claims the answer is yes: adding a short, hand-written description of the video a comment replies to raised recall (the share of truly toxic comments caught) by about twelve percentage points on average across GPT-3.5 Turbo, GPT-4.1, Gemini 1.5 Pro, and Claude 3 Opus, and improved F1 by up to 0.10. The strongest overall configuration was Gemini 1.5 Pro with context, reaching F1 0.819 and accuracy 0.823, while zero-shot GPT-4.1 produced the fewest false alarms. If the result holds, moderation teams can get meaningful gains from prompt design alone, without fine-tuning.

What carries the argument

The load-bearing mechanism is the context-augmented prompt: the model receives the target comment plus a short, hand-written description of the YouTube or TikTok video it responds to, and is asked to output 0 or 1. The context is meant to supply the background needed to tell ordinary criticism from toxic remarks, for example that a comment about a "sick person" refers to a Serbian politician struggling in an interview. Fifteen such contexts were manually crafted, one per video, and paired with 4,500 comments in two prompting modes. The experimental design compares the same four models with and without this context, holding temperature at 0 and using fixed evaluation metrics, so the measured recall gain is attributed to the added context.

What would settle it

Generate contexts automatically from the videos' existing metadata or LLM summaries and rerun the same four models on the same 4,500 comments; if the average recall gain falls far below the reported twelve points, the effect comes from these particular hand-written contexts rather than from context itself.

Watch

Extended reading notes

Core claim

The paper's central finding is that context-augmented prompting consistently beats zero-shot prompting for toxicity detection in Serbian, Croatian, and Bosnian. Across all four models, adding a two-sentence description of the video raised recall by an average of 0.12 and improved F1 by 0.07 to 0.10, at the cost of sometimes more false positives. Gemini 1.5 Pro with context achieved the highest F1 score of 0.819 and accuracy of 0.823; GPT-4.1 zero-shot remained the precision leader at 0.940 with a false positive rate of 0.039. Context augmentation also cut false negatives substantially, by 331 comments for GPT-3.5 Turbo and 418 for Gemini 1.5 Pro. The paper interprets this as evidence that brief contextual information helps models disambiguate cultural references, slang, and sarcasm in a low-resource setting.

Load-bearing premise

Everything depends on the fifteen hand-written context snippets being neutral, representative descriptions rather than products of the authors' labeling choices; the same two people who labeled the comments also wrote the contexts.

Editorial extensions

If this is right

  • Adding a two-sentence video description to the prompt raises recall by about 0.12 on average across GPT-3.5 Turbo, GPT-4.1, Gemini 1.5 Pro, and Claude 3 Opus.
  • Context augmentation improves F1 scores by 0.07 to 0.10 and reduces missed toxic comments by hundreds: 331 fewer false negatives for GPT-3.5 Turbo and 418 fewer for Gemini 1.5 Pro.
  • Gemini 1.5 Pro with context is the strongest overall configuration at F1 0.819 and accuracy 0.823, while GPT-4.1 zero-shot is the best choice when false alarms are costly, with precision 0.940 and a false positive rate of 0.039.
  • The benefit appears across Serbian, Croatian, and Bosnian, with context improving F1 in every language for every model.
  • Prompt design alone, without fine-tuning, can yield meaningful moderation gains in a low-resource setting.

Reading between the lines

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

  • If the central claim generalizes, the same two-sentence context trick could be tested in other low-resource languages, where even small recall gains translate into hundreds of comments hidden from users.
  • An ensemble of a high-precision zero-shot model and a high-recall context-augmented model is a natural next step, and the paper itself suggests it; our extension is that per-language threshold calibration would probably improve it further.
  • Because only 15 context snippets cover 4,500 comments, the measured gain is an average over many comments sharing one context; per-comment dynamic context could behave differently and deserves a direct test.
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

3 major / 6 minor

Summary. The paper introduces a manually labeled dataset of 4,500 YouTube and TikTok comments in Serbian, Croatian, and Bosnian and evaluates four LLMs (GPT-3.5 Turbo, GPT-4.1, Gemini 1.5 Pro, Claude 3 Opus) under zero-shot and context-augmented prompting. The central empirical claim is that adding a short hand-written video context raises recall by about 0.12 on average and improves F1 while sometimes increasing false positives, with Gemini 1.5 Pro context-augmented achieving the best overall balance. The paper also reports API costs, per-language F1 scores, qualitative failure cases, and practical recommendations such as threshold calibration and ensembling.

Significance. If the central result holds, the paper provides a practically useful demonstration that simple prompt changes can improve toxicity detection in underserved Balkan languages, and the released dataset is a valuable resource for a genuinely low-resource setting. The numerical tables are internally consistent: every row sums to 4,500 and the reported precision, recall, F1, and FPR values match the confusion-matrix counts. The annotation process reports high inter-annotator agreement, and the cost analysis is a useful addition. However, the main causal claim is currently supported only by a comparison in which the treatment contexts were authored by the same people who produced the gold labels and the treatment varies over only 15 video-level clusters; the evidence is therefore suggestive rather than conclusive.

major comments (3)
  1. [§III.B–III.C and §V.A] The context-augmentation condition is confounded with the annotators' labeling rationale. The 15 context snippets were manually written by the same two authors who assigned the gold labels, and the 'sick person' example in §V.A shows that the context supplies the precise interpretive link ('a Serbian politician struggling during an interview') that is needed to classify the comment as toxic. The context may thus re-inject gold-label information rather than measure the benefit of deployable context. The paper should either generate contexts from an independent source, compare against neutral video metadata (titles or descriptions), or explicitly analyze and bound this confound, ideally with a statement of whether contexts were written before or independently of labeling.
  2. [Table I and §IV.A] The treatment is applied at the level of only 15 videos, but all metrics are pooled over 4,500 comments, making the effective sample size for the context effect far smaller than the reported numbers suggest. The manuscript provides no per-video breakdown, confidence intervals, cluster-robust tests, or significance tests. The stated average recall improvement therefore has unknown precision and could be driven by a few unusually tailored contexts. A per-video analysis (including per-video recall gains) and a cluster bootstrap or mixed-effects model are required before the claim that context 'consistently increased recall' can be assessed.
  3. [§IV.G and Table III] The per-language F1 results are presented as evidence that context augmentation generalizes across Serbian, Bosnian, and Croatian, but these numbers inherit the same clustering problem and are not accompanied by per-language precision/recall counts, confusion matrices, or uncertainty estimates. Because the per-language F1 values cannot be derived from the aggregate counts in Table I, the computation should be documented, and the generalization claim should be limited to the observed descriptive trends unless supported by additional analysis.
minor comments (6)
  1. [Abstract and §III.A] The content categories listed in the abstract (music, politics, sports, modeling, influencer content, sexism, general topics) differ from the four categories given in §III.A (politics, music, sports, general/pop culture); please reconcile these lists and specify how the 15 videos map to categories.
  2. [Abstract and §V.A] The stated average recall gain of 'twelve percentage points' does not match Table I, where the gains are 0.144, 0.110, 0.182, and 0.109, averaging 0.136; either correct the number or explain the computation.
  3. [§III.C] Please state explicitly whether the context snippets were written before or independently of the annotation step, since the temporal ordering is directly relevant to the label-context confound.
  4. [§III.D] The paper states that all model outputs adhered strictly to the binary format, but it does not describe how any non-binary or malformed outputs would have been handled; please specify the parsing and exclusion criteria.
  5. [§IV.E and Table II] The cost table's column headings are easy to misread because the 500-request and 4,500-request sections are not clearly separated; a note confirming that the 4,500-request figures are scaled from the 500-request measurements would improve clarity.
  6. [Data Availability] The repository is described as containing instructions for obtaining the dataset, but the manuscript does not state whether the raw comments will be released directly or through a controlled-access process; please clarify the data-sharing terms.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the context-augmentation result is an empirical comparison, not a derivation from its own inputs.

full rationale

The paper's central claim is that adding short, hand-written context snippets to LLM prompts raises recall and F1 for toxicity detection in Serbian, Croatian, and Bosnian. This is established by running four commercial LLMs over a fixed, independently manually labeled corpus and comparing outputs under zero-shot versus context-augmented prompts. The human labels are external to the models, and the context snippets are inputs to the prompts; no parameter is fitted from the labels to the model outputs, and no reported metric is defined in terms of the quantity it is said to predict. The recalled comment example (Section V.A) is an illustration of model behavior, not a fitted prediction. There is one self-citation, [21] (A. Muminovic, arXiv:2505.18927), but it appears only in a related-work sentence about GPT-based cyberbullying detection and is not load-bearing for any inference in this paper. The shared authorship of annotations and context snippets is a genuine validity and generalizability concern, and the paper itself acknowledges limited annotation diversity and static, manually crafted context; however, that is a potential bias in the experimental design, not equation-level circularity. No reduction of a predicted result to an input by construction is present.

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

No numeric free parameters or invented entities. The central empirical claim rests on data representativeness, annotation reliability, and the neutrality of author-written context prompts. These are domain assumptions rather than fitted constants.

assumptions (5)
  • domain assumption Definition of toxicity from Davidson et al. (2017) is applicable to Serbian, Croatian, and Bosnian online comments.
    Adopted in Sections I.A and III.B; all labels depend on this definition.
  • domain assumption The three languages are similar enough to pool for the main metrics.
    Section III.E pools all comments for precision, recall, F1, and accuracy; per-language F1 is reported separately, but the pooled analysis assumes joint treatment is valid.
  • domain assumption Manual annotation by two authors, both familiar with the languages and video content, is accurate and unbiased.
    Section III.B reports Cohen's kappa 0.87 and 94% agreement; however both labels and contexts come from the same two people, a risk acknowledged in Section V.C.
  • ad hoc to paper The 15 hand-crafted context snippets are neutral and representative of what a production moderator would provide.
    Section III.C introduces manually crafted contexts specifically for these videos; the context effect is measured only under this author-written context regime.
  • domain assumption LLM outputs strictly follow the requested binary format.
    Section III.D states all outputs adhered to the 0/1 format; no parsing failures or invalid outputs are reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Large Language Models for Toxic Language Detection in Low-Resource Balkan Languages." pith.science (2026). https://pith.science/paper/ILWNGS73

@misc{pith2026250609992,
  author       = {Pith},
  title        = {Pith review of: Large Language Models for Toxic Language Detection in Low-Resource Balkan Languages},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ILWNGS73}},
  note         = {Machine review of arXiv:2506.09992}
}
read the original abstract

Online toxic language causes real harm, especially in regions with limited moderation tools. In this study, we evaluate how large language models handle toxic comments in Serbian, Croatian, and Bosnian, languages with limited labeled data. We built and manually labeled a dataset of 4,500 YouTube and TikTok comments drawn from videos across diverse categories, including music, politics, sports, modeling, influencer content, discussions of sexism, and general topics. Four models (GPT-3.5 Turbo, GPT-4.1, Gemini 1.5 Pro, and Claude 3 Opus) were tested in two modes: zero-shot and context-augmented. We measured precision, recall, F1 score, accuracy and false positive rates. Including a short context snippet raised recall by about 0.12 on average and improved F1 score by up to 0.10, though it sometimes increased false positives. The best balance came from Gemini in context-augmented mode, reaching an F1 score of 0.82 and accuracy of 0.82, while zero-shot GPT-4.1 led on precision and had the lowest false alarms. We show how adding minimal context can improve toxic language detection in low-resource settings and suggest practical strategies such as improved prompt design and threshold calibration. These results show that prompt design alone can yield meaningful gains in toxicity detection for underserved Balkan language communities.

Figures

Figures reproduced from arXiv: 2506.09992 by the authors.

Figure 1
Figure 1. End-to-end pipeline for our toxicity detection study. Steps 1 and 2 show data collection from YouTube and TikTok; steps 3–6 cover cleaning, [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Precision versus recall for each model configuration; full model names [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 4
Figure 4. F1 scores for zero-shot and context-augmented configurations across [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 19 canonical work pages

  1. [1]

    Automated hate speech detection and the problem of offensive language,

    T. Davidson, D. Warmsley, M. Macy, and I. Weber, “Automated hate speech detection and the problem of offensive language,” Proc. Int. AAAI Conf. Web and Social Media (ICWSM), vol. 11, no. 1, pp. 512– 515, May 2017. doi:10.1609/icwsm.v11i1.14955

    doi:10.1609/icwsm.v11i1.14955 2017

  2. [2]

    The polarizing impact of political disinformation and hate speech: A cross-country configural narrative,

    P. N. Vasist, D. Chatterjee, and S. Krishnan, “The polarizing impact of political disinformation and hate speech: A cross-country configural narrative,” Inf. Syst. Front., pp. 1–26, Apr. 2023. doi:10.1007/s10796- 023-10390-w

    doi:10.1007/s10796- 2023

  3. [3]

    Anonymity and its role in digital aggression: A systematic review,

    M. Kim, M. Ellithorpe, and S. A. Burt, “Anonymity and its role in digital aggression: A systematic review,” Aggression and Violent Behavior, vol. 72, Art. no. 101856, Sep.–Oct. 2023. doi:10.1016/j.avb.2023.101856

    work page arXiv 2023

  4. [4]

    Hateful symbols or hateful people? Predictive features for hate speech detection on Twitter,

    Z. Waseem and D. Hovy, “Hateful symbols or hateful people? Predictive features for hate speech detection on Twitter,” in Proc. NAACL Student Research Workshop, San Diego, CA, USA, Jun. 2016, pp. 88–93. doi:10.18653/v1/N16-2013

    doi:10.18653/v1/n16-2013 2016

  5. [5]

    Data-efficient strategies for expanding hate speech detection into under-resourced languages,

    P. R ¨ottger, D. Nozza, F. Bianchi, and D. Hovy, “Data-efficient strategies for expanding hate speech detection into under-resourced languages,” in Proc. 2022 Conf. Empirical Methods in Natural Language Processing (EMNLP), Abu Dhabi, United Arab Emirates, Dec. 2022, pp. 5674–

    work page 2022

  6. [6]

    Automat- ing news comment moderation with limited resources: Benchmarking in Croatian and Estonian,

    R. Shekhar, M. Pranji ´c, S. Pollak, A. Pelicon, and M. Purver, “Automat- ing news comment moderation with limited resources: Benchmarking in Croatian and Estonian,” J. Lang. Technol. Comput. Linguist., vol. 34, no. 1, pp. 49–79, 2020. doi:10.21248/jlcl.34.2020.224

    work page doi:10.21248/jlcl.34.2020.224 2020

  7. [7]

    Facebook, Twitter struggling in fight against Balkan content violations,

    I. Jeremic and M. Stojanovic, “Facebook, Twitter struggling in fight against Balkan content violations,” Balkan Insight, Feb. 16, 2021. [On- line]. Available: https://balkaninsight.com/2021/02/16/facebook-twitter- struggling-in-fight-against-balkan-content-violations/

    work page 2021

  8. [8]

    Multilingual offensive language identification for low-resource languages,

    T. Ranasinghe and M. Zampieri, “Multilingual offensive language identification for low-resource languages,” ACM Trans. Asian Low- Resour. Lang. Inf. Process., vol. 21, no. 1, Art. no. 4, pp. 1–13, 2021. doi:10.1145/3457610

    work page doi:10.1145/3457610 2021

Show all 32 references
  1. [9]

    Beyond Balkanism: The scholarly politics of region making,

    D. Mishkova, “Beyond Balkanism: The scholarly politics of region making,” 1st ed. New York, NY , USA: Routledge, 2020, 292 pp. ISBN: 9780367666088

    2020

  2. [10]

    YouTube users by country 2025,

    World Population Review, “YouTube users by country 2025,” Tech/Business section, 2025. [Online]. Available: https://worldpopulationreview.com/country-rankings/youtube-users- by-country

    2025

  3. [11]

    TikTok user age, gender, and demographics (2025),

    F. Duarte, “TikTok user age, gender, and demographics (2025),” Exploding Topics, May 7, 2025. [Online]. Available: https://explodingtopics.com/blog/tiktok-demographics

    2025

  4. [12]

    Code- switching patterns can be an effective route to improve performance of downstream NLP applications: A case study of humour, sarcasm and hate speech detection,

    S. Bansal, V . Garimella, A. Suhane, J. Patro, and A. Mukherjee, “Code- switching patterns can be an effective route to improve performance of downstream NLP applications: A case study of humour, sarcasm and hate speech detection,” in Proc. 58th Annu. Meeting Assoc. Comput. Li...

    2020 doi

  5. [13]

    A multitask, multilingual, multimodal evaluation of ChatGPT on reasoning, halluci- nation, and interactivity,

    Y . Bang, S. Cahyawijaya, N. Lee, W. Dai, D. Su, B. Wilie, H. Lovenia, Z. Ji, T. Yu, W. Chung, Q. V . Do, Y . Xu, and P. Fung, “A multitask, multilingual, multimodal evaluation of ChatGPT on reasoning, halluci- nation, and interactivity,” in Proc. 13th Int. Joint Conf. Natural...

    2023 doi

  6. [14]

    Using Convolutional Neural Networks to Classify Hate-Speech,

    B. Gamb ¨ack and U. K. Sikdar,“Using Convolutional Neural Networks to Classify Hate-Speech,” in Proceedings of the First Workshop on Abusive Language Online, Vancouver, BC, Canada, Aug. 2017, pp. 85–90. doi: 10.18653/v1/W17-3013

    2017 doi

  7. [15]

    Toxic speech detection using traditional machine learning models and BERT and fastText em- bedding with deep neural networks,

    P. Malik, A. Aggrawal, and D. K. Vishwakarma, “Toxic speech detection using traditional machine learning models and BERT and fastText em- bedding with deep neural networks,” in Proc. 5th Int. Conf. Computing Methodologies and Communication (ICCMC), Erode, India, 2021, pp. 1254...

    2021

  8. [16]

    A comparative study of using pre-trained language models for toxic comment classification,

    Z. Zhao, Z. Zhang, and F. Hopfgartner, “A comparative study of using pre-trained language models for toxic comment classification,” in Companion Proc. Web Conf. (WWW), New York, NY , USA, 2021, pp. 500–507. [Online]. Available: https://doi.org/10.1145/3442442.3452313

    2021

  9. [17]

    A machine learning ap- proach to identify toxic language in the online space,

    L. Kaati, A. Shrestha, and N. Akrami, “A machine learning ap- proach to identify toxic language in the online space,” in Proc. IEEE/ACM Int. Conf. Advances in Social Networks Analysis and Mining (ASONAM), Istanbul, Turkey, 2022, pp. 396–402, doi: 10.1109/ASONAM55673.2022.10068619

    2022

  10. [18]

    Towards generalisable hate speech detection: A review on obstacles and solutions,

    W. Yin and A. Zubiaga, “Towards generalisable hate speech detection: A review on obstacles and solutions,” PeerJ Comput. Sci., vol. 7, p. e598, Jun. 2021. [Online]. Available: https://doi.org/10.7717/peerj-cs.598

    2021 doi

  11. [19]

    Watch your lan- guage: Investigating content moderation with large language mod- els,

    D. Kumar, Y . A. AbuHashem, and Z. Durumeric, “Watch your lan- guage: Investigating content moderation with large language mod- els,” Proc. Int. AAAI Conf. Web and Social Media (ICWSM), vol. 18, no. 1, pp. 865–878, May 2024. [Online]. Available: https://doi.org/10.1609/icwsm.v...

    2024 doi

  12. [20]

    Probing LLMs for hate speech detection: strengths and vulnerabilities,

    S. Roy, A. Harshvardhan, A. Mukherjee, and P. Saha, “Probing LLMs for hate speech detection: strengths and vulnerabilities,” in Findings of the Association for Computational Linguistics: EMNLP 2023, Singapore, 2023, pp. 6116–6128, doi: 10.18653/v1/2023.findings-emnlp.407

    2023 doi

  13. [21]

    Moderating harm: Benchmarking large language models for cyberbullying detection in YouTube comments,

    A. Muminovic, “Moderating harm: Benchmarking large language models for cyberbullying detection in YouTube comments,” arXiv preprint arXiv:2505.18927, May 2025. [Online]. Available: https://arxiv.org/abs/2505.18927

    2025 arXiv

  14. [22]

    Chain-of-thought prompting elicits reasoning in large language models,

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, B. Ichter, F. Xia, E. H. Chi, Q. V . Le, and D. Zhou,“Chain-of-thought prompting elicits reasoning in large language models,” in Proc. 36th Int. Conf. Neural Inf. Process. Syst. (NeurIPS), Red Hook, NY , USA: Curran Associates Inc., 20...

    2022

  15. [23]

    Moderating new waves of online hate with chain- of-thought reasoning in large language models,

    N. Vishwamitra et al., “Moderating new waves of online hate with chain- of-thought reasoning in large language models,” in Proc. IEEE Symp. Security and Privacy (SP), San Francisco, CA, USA, 2024, pp. 788–806, doi: 10.1109/SP54263.2024.00181

    2024

  16. [24]

    Improving covert toxicity detection by retrieving and generating references,

    D.-H. Lee, H. Cho, W. Jin, J. Moon, S. Park, P. R ¨ottger, J. Pujara, and R. K.-w. Lee, “Improving covert toxicity detection by retrieving and generating references,” in Proc. 8th Workshop on Online Abuse and Harms (WOAH 2024), Mexico City, Mexico, 2024, pp. 266–274, doi: 10.1...

    2024 doi

  17. [25]

    Abusive speech detection in Serbian using machine learning,

    D. Joki ´c, R. Stankovi ´c, and B. ˇSandrih Todorovi ´c, “Abusive speech detection in Serbian using machine learning,” in Proc. 1st Int. Conf. Natural Language Processing and Artificial Intelligence for Cyber Security, Lancaster, UK, 2024, pp. 153–163. [Online]. Available: htt...

    2024

  18. [26]

    Early dialectal diversity in South Slavic II,

    F. Kortlandt, “Early dialectal diversity in South Slavic II,” Studies in Slavic and General Linguistics, vol. 30, pp. 215–235, 2003. [Online]. Available: http://www.jstor.org/stable/40997749

    2003

  19. [27]

    Tviterasi, tvitera ˇsi or twittera ˇsi? Pro- ducing and analysing a normalised dataset of Croatian and Serbian tweets,

    M. Mili ˇcevi´c and N. Ljube ˇsi´c, “Tviterasi, tvitera ˇsi or twittera ˇsi? Pro- ducing and analysing a normalised dataset of Croatian and Serbian tweets,” Slovenˇsˇcina 2.0, vol. 4, no. 2, pp. 156–188, Sep. 2016, doi: 10.4312/slo2.0.2016.2.156-188

    2016 doi

  20. [28]

    SRBerta—A transformer language model for Serbian Cyrillic legal texts,

    M. Bogdanovi ´c, J. Koci ´c, and L. Stoimenov, “SRBerta—A transformer language model for Serbian Cyrillic legal texts,” Languages, vol. 9, no. 1, p. 24, 2024, doi: 10.3390/info15020074

    2024 doi

  21. [29]

    Development of a multilingual model for machine sentiment analysis in the Ser- bian language,

    D. Draskovic, D. Zecevic, and B. Nikolic, “Development of a multilingual model for machine sentiment analysis in the Ser- bian language,” Mathematics, vol. 10, no. 18, p. 3236, 2022, doi: 10.3390/math10183236

    2022 doi

  22. [30]

    Detecting ugly and derogatory words in Serbian language using a web browser extension,

    A. Milakovi ´c, V . Jocovi ´c, J. Cincovi ´c, M. Mi ´covi´c, U. Radenkovi ´c, and D. Dra ˇskovi´c, “Detecting ugly and derogatory words in Serbian language using a web browser extension,” in Proc. 2024 32nd Telecom- munications Forum (TELFOR), Belgrade, Serbia, 2024, pp. 1–4, ...

    2024

  23. [31]

    Fine-tuning GPT-3.5 for hate speech detection in Albanian YouTube comments: Challenges and analysis,

    H. Kopani and R. Llapushi, “Fine-tuning GPT-3.5 for hate speech detection in Albanian YouTube comments: Challenges and analysis,” in Bridging Horizons in Artificial Intelligence, Robotics, Cybersecurity, Smart Cities, and Digital Economy (ICITTBT 2024), K. Dhoska and E. Spaho,...

    2024 doi

  24. [5691]

    doi:10.18653/v1/2022.emnlp-main.383

    2022 doi

Pith tools

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