Low-Resource Languages & Dialects
A Benchmark and Labeling Framework for Community-Contingent Hate Speech Annotations, Replacing Single Gold Labels with Distributions of Plausible Ones
Generated automatically from the limitations stated in 8 papers (EMNLP, TheWebConference, NAACL), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
Every current hate speech dataset collapses a genuinely plural judgment — what counts as hateful depends on annotator identity, community norms, reclamation practices, and exposure to implicit forms — into one label, then trains and evaluates models against it. This means models cannot represent the disagreement they were built from, audits of pretraining data and model outputs in low-resource settings have no trustworthy ground truth to measure against, and moderation systems calibrated on one community's thresholds misfire (or stay silent) on another's counter-speech, quotes, and reclaimed slurs. Structurally, nobody today can state, for a given utterance, how hateful-label assignments vary across the populations that matter — so "biased model" and "ambiguous data" are permanently confounded.
Why it matters
Moderation and auditing systems that can be tuned to a named community's threshold with known error bars, and a clean separation of model bias from label ambiguity in any future hate speech evaluation.
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Re-annotate an existing multilingual/multimodal hate corpus with structured disagreement: multiple annotator pools stratified by community identity and region, each instance receiving a distribution over labels plus free-text rationale, with inter-pool divergence measured per item. Deliverable: a public dataset where the unit of truth is the distribution, and metrics (pool-divergence scores, agreement-by-subgroup) computed against it.
- 2
Train and evaluate label-distribution models — models outputting P(hateful | annotator profile) instead of a point label — and compare their calibration against single-label baselines on the misclassification cases already known to be systematic (disguised memes, communal words, reclaimed slurs). Measure calibration error and per-subgroup error gaps.
- 3
Audit downstream use: rerun a toxicity audit of one pretraining corpus under distributional labels versus gold labels, quantifying how much of the documented bias signal (e.g., occupation associations) survives when label subjectivity is explicitly modeled. Measure the change in flagged-content rates and which flagged items flip.
Have a different approach?
Describe how you would tackle this problem and we'll look for papers that already do it. Free; your text stays private.
Why it might fail
Stratified annotator pools may not actually capture "community" norms — self-identified membership may predict labeling behavior far more weakly than the framework assumes, collapsing the distributions into noise and leaving the single-label status quo justified.
Sub-problems
Learning Nuanced and Implicit Hate Speech Classifiers Under Weak Auxiliary Supervision
Current high-performing moderation models for implicit and target-specific hate speech depend on auxiliary metadata—such as explicitness tags, target-group annotations, and fine-grained annotator demographic profiles—that are absent in standard production corpora. When forced to train solely on aggregate binary labels, models conflate communal identity terms with toxic intent and fail on contextual nuances like reclaimed slurs or quoted counter-speech. Furthermore, hard-coded aggregation thresholds treat subjective socio-cultural disagreement as random noise, rendering classifiers non-transferable across diverse online communities.
Evidence
Each paper's own statement of the limitation, verbatim.
- Rethinking Implicit Hate Speech Detection: Focusing on Latent Hate Components via Dual-Process ArgumentationTheWebConference 2026
'Implicit hate' is socio-culturally contingent; label noise and inherited ambiguity mean thresholds may not transfer across communities, and reclaimed slurs, quoted speech, and counter-speech remain challenging.
- Causality Guided Representation Learning for Cross-Style Hate Speech DetectionTheWebConference 2026
Requires supervision beyond the hate label: training uses style (explicit/implicit) annotations and optional target-group labels (D = {(x, y, s, t)}), which many hate speech corpora lack
- Personalization up to a Point: Why Personalized Content Moderation Needs Boundaries, and How We Can Enforce ThemEMNLP 2025
The 'extreme annotator' definition (rating instances hateful-for-majority as non-hateful >50% of the time) is an arbitrary threshold affecting the findings
- AfriHate: A Multilingual Collection of Hate Speech and Abusive Language Datasets for African LanguagesNAACL 2025
Views on hate speech are subjective; even with diverse native annotators, all perspectives cannot be represented, and the authors explicitly forbid automated removal use
- MemeCLIP: Leveraging CLIP Representations for Multimodal Meme ClassificationEMNLP 2024
Annotation is subjective and prone to bias (final Fleiss' Kappa only 0.74-0.81 across tasks); models trained on it may exhibit biased predictions, and misclassification analysis shows hateful memes disguised as benign imagery are missed and communal words cause individual hate to be mislabeled as community-targeted
- Exploring Cross-Cultural Differences in English Hate Speech Annotations: From Dataset Construction to AnalysisNAACL 2024
The dataset is small (1,580 posts) and culture-specific post collection relied on fixed hate-related keywords from only Reddit and YouTube, potentially biasing collected posts.
- Cheetah: Natural Language Generation for 517 African LanguagesACL 2024
No comprehensive analysis of biases, hateful speech, or socially undesirable content in pretraining data or model outputs, due to limited access to native speakers and automated tools for most supported languages
- JASMINE: Arabic GPT Models for Few-Shot LearningEMNLP 2023
No toxic/hateful content was filtered from pretraining data, so the models carry documented biases (e.g., people of African descent associated with low-wage occupations 100% of the time)
Nearest existing work
- “It’s Not Just Hate”: A Multi-Dimensional Perspective on Detecting Harmful Speech OnlineEMNLP 2022
- Exploring Cross-Cultural Differences in English Hate Speech Annotations: From Dataset Construction to AnalysisNAACL 2024
- Learning from the Worst: Dynamically Generated Datasets to Improve Online Hate DetectionACL 2021
- The Risk of Racial Bias in Hate Speech DetectionACL 2019
- HateDay: Insights from a Global Hate Speech Dataset Representative of a Day on TwitterACL 2025
- Hate Speech and Counter Speech Detection: Conversational Context Does MatterNAACL 2022
- Multi3Hate: Multimodal, Multilingual, and Multicultural Hate Speech Detection with Vision–Language ModelsNAACL 2025
- Directions for NLP Practices Applied to Online Hate Speech DetectionEMNLP 2022
- ModelCitizens: Representing Community Voices in Online SafetyEMNLP 2025
- AfriHate: A Multilingual Collection of Hate Speech and Abusive Language Datasets for African LanguagesNAACL 2025
- Hate Personified: Investigating the role of LLMs in content moderationEMNLP 2024
- Annotators with Attitudes: How Annotator Beliefs And Identities Bias Toxic Language DetectionNAACL 2022
- Latent Hatred: A Benchmark for Understanding Implicit Hate SpeechEMNLP 2021
- Oddballs and Misfits: Detecting Implicit Abuse in Which Identity Groups are Depicted as Deviating from the NormEMNLP 2024
- Hateful Word in Context ClassificationEMNLP 2024
Related open problems
Low-Resource Languages & Dialects
Reliability-Calibrated Annotation Protocols for Subjective and Low-Resource labeling Tasks
Across many NLP annotation efforts, gold standards rest on so few annotators—and on tasks so subjective—that reported inter-annotator agreement is often moderate at best (κ 0.18–0.46), and in some cases reliability fails outright (agreement too weak to include in evaluation at all). Because papers rarely distinguish "the task is ambiguous" from "the pool was too small," downstream users cannot judge how trustworthy the gold labels are, and resources with single-annotator or single-community provenance silently embed one worldview. Structurally, there is no standard way to report, audit, or correct for annotation reliability when it is low or unmeasured.
Low-Resource Languages & Dialects
Metric Inflation and Reliability Limits in IPA-Space Generation Evaluation
Multilingual systems increasingly evaluate generations in International Phonetic Alphabet (IPA) space to bridge script disparities, yet standard character-level metrics like chrF inflate scores because IPA relies heavily on shared Latin base characters. Furthermore, automatic grapheme-to-phoneme tools (such as Epitran) introduce systematic transcription errors and inventory mismatches that are masked when evaluating against synthetic or machine-translated benchmarks. Researchers currently cannot distinguish whether high IPA-level chrF gains reflect genuine semantic translation quality or superficial phonetic transcription artifacts.