Skip to content
Open access

IRAIVI: Agentic Framework for Tamil Hate Speech Detection and Counter-Speech Generation Toward Responsible AI Intervention

2026 · IEEE Access · Vol 14, pp. 132668-132685 · 0 citations · 38 references

TL;DR

IRAIVI (Intelligent Responsible Agentic AI for Identifying Hate Speech, Verifying Legal References, and Intervention), a CrewAI-based prompt-driven agentic framework that orchestrates multiple agents, tasks, and a sequential workflow for hate speech detection, target identification, hate category classification, legal knowledge retrieval, and Tamil counter-speech generation is presented.

Abstract

Online gender-based hate speech affects women, men, and LGBTQIA+ communities and is difficult to moderate in low-resource Tamil because social media content often contains transliteration, code-mixing, dialectal variation, sarcasm, and culturally specific abusive expressions. This paper presents IRAIVI (Intelligent Responsible Agentic AI for Identifying Hate Speech, Verifying Legal References, and Intervention), a CrewAI-based prompt-driven agentic framework that orchestrates multiple agents, tasks, and a sequential workflow for hate speech detection, target identification, hate category classification, legal knowledge retrieval, and Tamil counter-speech generation. The framework integrates Tree-of-Thought (ToT) reasoning, Retrieval-Augmented Generation (RAG), Genetic Algorithm (GA)-based prompt optimization, and human-in-the-loop (HITL) escalation for uncertain cases. A Qwen-based self-instruction pipeline expanded 200 seed examples into 5,747 Alpaca-style instruction-input-output triplets. The dataset was collected from publicly available Tamil YouTube comments using the YouTube API and contained no personally identifiable information therefore, institutional ethics approval was not required. The proposed three-level annotation scheme includes hate or non-hate classification, target identification, and hate category (implicit or explicit), achieving strong annotation agreement with Cohen’s kappa (k = 0.87). Experiments were conducted using GPT-3.5, Gemini, Mistral-7B, Qwen-4B, Phi-2, and TinyLLaMA. GPT-3.5 achieved the highest classification F1-score of 0.91, while Qwen-4B achieved 0.85 among the evaluated open-source models. The ToT + GA configuration achieved a counter-speech quality score of 0.86 and a BERTScore of 0.92, demonstrating the effectiveness of the proposed framework for responsible Tamil hate speech detection and intervention.

Read PDF

Similar papers

Open access Sep 2026

VISH-GUARD: a multi-agent and LLM-powered framework for multilingual voice phishing detection

Voice phishing (vishing) has emerged as a major cybersecurity threat, leveraging persuasive speech and psychological manipulation to deceive victims in real time. Existing detection approaches remain limited by monolingual assumptions, unimodal analysis, and poor interpretability, reducing their effectiveness against...

Yasser Hmimou, Mohamed Tabaa, Azeddine Khiat et al. · 0 citations
Oct 2026

Evasion-hate: targeted augmentation for hate speech detection under prompt-induced surface-form shifts

Online hate speech detection is increasingly used as a decision-support component in content moderation, but most detectors are evaluated on clean held-out text. In practice, users can preserve harmful intent while changing surface form through spelling obfuscation, coded wording or code-switching, creating a robus...

Thanh Le, Thien Khai Tran · 0 citations
Open access Sep 2026

Detection of Hate and Dehumanizing Speech in Pashto Text Using Machine Learning Algorithms

The results indicate that distinguishing between hate speech and dehumanizing language remains challenging, particularly when both forms occur within the same comment, and demonstrate that appropriate text normalization and selecting an effective feature size can improve classification performance.

Shahid Nasiri, Nesar Ahmad Wasi · 0 citations

INFORMATION AND WEB TECHNOLOGIES

An NLP-driven adaptive framework that combines posterior-based phone scoring, reliability-weighted multi-view aggregation, calibrated uncertainty, feedback prioritization, longitudinal learner-state updating, controlled-to-spontaneous transfer, and expected-learning-gain task selection is converted.

Kuvandikov Jura Turunbayevich, Rohmonqulov Muhammadyusuf Egamberdi o'g'li · 0 citations
Open access Oct 2026

Algorithmic blind spot: (In)visible racism, multimodality and the limits of AI hate speech detection

Automated hate-speech detection has advanced rapidly, yet significant blind spots remain in how artificial intelligence recognises online abuse and processes its contextual and culturally situated meanings. This article examines why anti-Roma hate often evade moderation on social media. Drawing on insights from multi...

Petre Breazu, N. Katsos · 0 citations
Conference Aug 2026

Fusing Semantics and LLM-Driven Emotions by Gated Attention for Implicit Hate Speech Detection

Detecting implicit hate speech continues to be a significant challenge for natural language processing systems, as harmful intent is often concealed within polite, neutral, or coded language. Traditional text-based classifiers often struggle to identify these subtle forms of hostility because they lack the emotional co...

Yuan-Chih Yu · 0 citations

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.