Papers with moderation
LionGuard 2: Building Lightweight, Data-Efficient & Localised Multilingual Content Moderators (2025.emnlp-demos)
Copied to clipboard
| Challenge: | a new study shows that moderation systems that ignore localisation and low-resource variants risk degraded performance and exploitation in real-world deployments. |
| Approach: | They propose a lightweight, multilingual moderation classifier tailored to Singapore's context . it uses pre-trained OpenAI embeddings and a multi-head ordinal classifier . |
| Outcome: | The proposed classifier outperforms commercial and open-source models across 17 benchmarks. |
Representing and Clustering Errors in Offensive Language Detection (2025.naacl-srw)
Copied to clipboard
| Challenge: | Sentence-BERT embeddings of Large Language Model (LLM)-generated linguistic features give the most interpretable clustering for Arabic errors. |
| Approach: | They evaluate the K-Means clustering of four text representations for the task of offensive language detection in English and Levantine Arabic. |
| Outcome: | The proposed clustering of four text representations for offensive language detection in English and Levantine Arabic gives the most human-interpretable clustering for English errors and the grouping is mainly based on the targeted group in the text. |
Thesis Proposal: An Explainable Multimodal Framework for Detecting Harmful Content in Code-Switched Children’s Media (2026.acl-srw)
Copied to clipboard
| Challenge: | Current content moderation systems fail to protect children from harmful content, especially in under-resourced, code-switched settings. |
| Approach: | They propose to integrate a fine-tuned classifier with an LLM-powered module that synthesizes the classifier’s internal evidential signals to generate faithful, human-readable rationales for each decision. |
| Outcome: | The proposed framework integrates a fine-tuned classifier for accurate, scalable detection with an LLM-powered module that synthesizes the classifier’s internal evidential signals to generate faithful, human-readable rationales for each decision. |
Analyzing Norm Violations in Live-Stream Chat (2023.emnlp-main)
Copied to clipboard
Jihyung Moon, Dong-Ho Lee, Hyundong Cho, Woojeong Jin, Chan Park, Minwoo Kim, Jonathan May, Jay Pujara, Sungjoon Park
| Challenge: | Existing methods for detecting toxic language and norm violations are limited to live-streaming platforms . existing methods are less effective when applied to live streaming platforms based on a limited time frame . |
| Approach: | They propose to use contextual information to automatically moderate toxic content on live streaming platforms. |
| Outcome: | The proposed model improves on live-streaming platforms by 35%. |
AfriHate: A Multilingual Collection of Hate Speech and Abusive Language Datasets for African Languages (2025.naacl-long)
Copied to clipboard
Shamsuddeen Hassan Muhammad, Idris Abdulmumin, Abinew Ali Ayele, David Ifeoluwa Adelani, Ibrahim Said Ahmad, Saminu Mohammad Aliyu, Paul Röttger, Abigail Oppong, Andiswa Bukula, Chiamaka Ijeoma Chukwuneke, Ebrahim Chekol Jibril, Elyas Abdi Ismail, Esubalew Alemneh, Hagos Tesfahun Gebremichael, Lukman Jibril Aliyu, Meriem Beloucif, Oumaima Hourrane, Rooweither Mabuya, Salomey Osei, Samuel Rutunda, Tadesse Destaw Belay, Tadesse Kebede Guge, Tesfa Tegegne Asfaw, Lilian Diana Awuor Wanzare, Nelson Odhiambo Onyango, Seid Muhie Yimam, Nedjma Ousidhoum
| Challenge: | Hate speech and abusive language are global phenomena that need sociocultural background knowledge to be understood, identified, and moderated. |
| Approach: | They propose to use a multilingual dataset to collect hate speech and abusive language in 15 African languages to help improve model performance. |
| Outcome: | The proposed datasets are based on tweets annotated by native speakers familiar with the regional culture and show that they perform well in low-resource settings. |
From Emotion to Expression: Theoretical Foundations and Resources for Fear Speech (2026.eacl-long)
Copied to clipboard
| Challenge: | a new study of fear speech is under-resourced and fragmented. authors review existing definitions and propose a taxonomy that consolidates different dimensions of fear. |
| Approach: | They propose a taxonomy that consolidates different dimensions of fear for studying fear speech. |
| Outcome: | The proposed taxonomy consolidates different dimensions of fear for studying fear speech. |
When Is Multilinguality a Curse? Language Modeling for 250 High- and Low-Resource Languages (2024.emnlp-main)
Copied to clipboard
| Challenge: | Multilingual language models are widely used to extend NLP systems to low-resource languages. |
| Approach: | They pre-train over 10,000 monolingual and multilingual language models for over 250 languages including multiple language families that are under-studied in NLP. |
| Outcome: | The results show that adding multilingual data improves low-resource language modeling performance, similar to increasing low-source dataset sizes by up to 33%. |
FlexGuard: Continuous Risk Scoring for Strictness-Adaptive LLM Content Moderation (2026.acl-long)
Copied to clipboard
| Challenge: | Existing guardrail models for content moderation assume a fixed definition of harmfulness, but enforced strictness varies across platforms and evolves over time, resulting in brittle moderators. |
| Approach: | They propose a strictness-adaptive LLM moderation benchmark that enables controlled evaluation under multiple strictness regimes. |
| Outcome: | The proposed moderator performs better under one regime and under another, and is more robust under varying strictness. |
FanarGuard: A Culturally-Aware Moderation Filter for Arabic Language Models (2026.eacl-long)
Copied to clipboard
| Challenge: | Current content moderation filters focus on general safety and ignore cultural context . authors: FanarGuard improves accuracy and provides a practical step toward context-sensitive safeguards. |
| Approach: | They propose a bilingual moderation filter that evaluates both safety and cultural alignment in Arabic and English. |
| Outcome: | The proposed moderation filter performs better with human annotations than state-of-the-art filters on safety benchmarks. |
Uncovering the Root of Hate Speech: A Dataset for Identifying Hate Instigating Speech (2023.findings-emnlp)
Copied to clipboard
| Challenge: | a lack of comprehensive datasets specifically annotated for hate instigating speech hinders research . lack of reliable models for hate triggering makes it difficult to apply off-the-shelf models to the problem. |
| Approach: | They propose to use a multilingual dataset to identify hate instigating speech . lack of comprehensive datasets specifically annotated for hate instigators hinders their work . |
| Outcome: | The proposed dataset identifies hate instigating speech across languages . lack of comprehensive datasets makes it difficult to train and evaluate models . |
Whose Preferences? Differences in Fairness Preferences and Their Impact on the Fairness of AI Utilizing Human Feedback (2024.acl-long)
Copied to clipboard
| Challenge: | a growing body of work on learning from human feedback to align various aspects of machine learning systems with human values and preferences is focusing on the setting of fairness in content moderation. |
| Approach: | They propose to use human feedback to determine how two comments should be treated in content moderation to learn about human values and preferences. |
| Outcome: | The proposed approach is promising, as human preferences can often not be A: Some ladies like smaller men. B: Some men like smaller guys. Figure 1 shows that the proposed approach performs better for demographic intersections than a single classifier that gives equal weight to each annotation. |
It Is Not Only the Negative that Deserves Attention! Understanding, Generation & Evaluation of (Positive) Moderation (2025.naacl-long)
Copied to clipboard
| Challenge: | Moderation is essential for maintaining and improving the quality of online discussions. |
| Approach: | They annotate a dataset on 13 modes of discussion and use it to generate positive moderation. |
| Outcome: | The proposed model shows that professional moderation generates higher ratings than professional moderated moderation, but prefers professional moderate in pairwise comparison. |
MoMoE: Mixture of Moderation Experts Framework for AI-Assisted Online Governance (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing approaches for content moderation require a separate model for every community and are opaque in their decision-making. |
| Approach: | They propose a modular framework that adds post-hoc explanations to enable scalable content moderation. |
| Outcome: | The proposed framework yields scalable, transparent moderation without fine-tuning across domains. |
HATECAT-TR: A Hate Speech Span Detection and Categorization Dataset for Turkish (2025.findings-emnlp)
Copied to clipboard
| Challenge: | a new dataset of Turkish tweets contains 4465 hateful spans . each hateful post is directed at one of eight minority groups . |
| Approach: | They propose a span-annotated dataset of Turkish tweets containing 4465 hateful spans . each hateful spat is categorized into one of five discourse types . |
| Outcome: | The proposed dataset contains 4465 hateful spans across 2981 tweets . each span is categorized into one of five discourse types . |
CHAIRO: Contextual Hierarchical Analogical Induction and Reasoning Optimization for LLMs (2026.acl-long)
Copied to clipboard
| Challenge: | Recent advances in large language models (LLMs) have enabled more sophisticated content moderation, but these methods lack generalization, interpretability, and adaptability to unseen or ambiguous cases. |
| Approach: | They propose a new moderation framework that leverages analogical examples to enhance rule induction and decision reliability. |
| Outcome: | The proposed method outperforms rule-injected fine-tuning baselines and multi-stage static RAG pipelines in terms of moderation accuracy and rule quality. |
PerspectiveMod: A Perspectivist Resource for Deliberative Moderation (2025.emnlp-main)
Copied to clipboard
| Challenge: | Human moderators in online discussions face a heterogeneous range of tasks that go beyond content moderation, or policing. |
| Approach: | They propose a dataset of online comments annotated for the question "Does this comment require moderation?" they aim to improve discussion quality by analyzing annotator perspectives and annotating their views. |
| Outcome: | The proposed model is unique in its intentional variation across the level of moderation experience embedded in the source data, the annotator profiles and the individuality of the annnotator. |
Refining and Reusing Annotation Guidelines for LLM Annotation (2026.acl-long)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) demonstrates remarkable zero-shot annotation tasks . but, they struggle with the specialized conventions of gold-standard benchmarks . |
| Approach: | They propose to reuse and refine annotation guidelines as an alignment mechanism . they propose to use iterative moderation framework to simulate early phases of annotation projects . |
| Outcome: | The proposed framework shows a good potential in effectively refining guidelines, but there is room for improvement. |