Tag
This paper critiques the reliance on Perspective API for toxicity measurement and argues for community-owned measurement infrastructure in NLP and LLM evaluation, releasing scores for 5.9 million text snippets to facilitate research after the API's shutdown.
This paper proposes SeTox, a search-augmented LLM framework for detecting implicit toxicity in Chinese neologisms by leveraging real-time web context and public consensus. Experiments show that even 3B-scale models outperform larger recent models on this task.
This paper analyzes toxic behavior on the decentralized Mastodon platform using NLP and machine learning, examining the challenges of moderation across independently operated servers and proposing insights for healthier online communities.
This paper presents a three-stage neural-symbolic pipeline for gaming toxicity detection, combining transformer ensembles with rule-based mediation, achieving top accuracy in the EEUCA 2026 shared task.
This survey synthesizes research on toxicity detection and detoxification for multilingual large language models, cataloging threat models, task formulations, detection approaches, and mitigation strategies, while identifying persistent challenges such as uneven language coverage and culturally contingent definitions of harm.
This paper investigates how LLMs' internal priors affect zero-shot annotation performance, finding that nearly two-thirds of errors resist prompt-based correction and introducing Definition-Specific Familiarity as a better predictor than memorization metrics.
The paper introduces CITA, a framework for generating implicit toxicity attacks in Chinese to evaluate and improve LLM toxicity detectors, finding high attack success rates across tested models.
This paper studies fairness in toxicity classification across three axes: ranking, calibration, and abstention. It compares ERM, reweighted ERM, and Group DRO methods with post-hoc interventions, finding that calibration disparity is a hidden fairness violation and that abstention itself can be unfair.
This paper presents a system for the EEUCA 2026 shared task on toxicity detection in gaming chat, achieving 4th place by fine-tuning Llama 3.1 8B with synthetic data augmentation. It highlights a 'validation trap' phenomenon where high validation scores do not correlate with test performance due to dataset distribution shifts.