Innovative Practice of Explainable AI in Hate Speech Detection on Vietnamese Social Media: An LLM Fine-Tuning Method Based on Chain-of-Thought Prompting

This article introduces an innovative study on hate speech detection for Vietnamese social media. By combining Chain-of-Thought (CoT) prompting and QLoRA fine-tuning technology, it achieves high-precision classification while extracting the reasoning basis and implicit statements behind model decisions, providing new ideas for the application of explainable AI in the content moderation field.

The development of social media has improved the efficiency of information dissemination, but traditional rule-based or shallow machine learning detection methods face challenges in Southeast Asian markets like Vietnam due to language uniqueness and complex cultural contexts. Existing systems only make binary judgments without interpretability, reducing credibility and leading to difficulties in appeal for wrongful bans and model optimization.

Selected the Qwen2.5-3B model (excellent multilingual support, balanced performance and efficiency); adopted QLoRA fine-tuning (4-bit quantization reduces memory usage, parameter isolation preserves general knowledge, fast iteration); designed Chain-of-Thought prompt templates to guide the model to first understand content, identify intent, generate reasoning, extract implicit statements, then make the final judgment.

1. Dual output mechanism: classification output (whether it is hate speech and its type) + explanation output (reasoning basis + implicit offensive statements); 2. Vietnamese adaptation: preprocess online slang/abbreviations/emojis, design grammar-adapted prompt templates, and introduce local cultural context; 3. Balance performance and interpretability: Chain-of-Thought improves transparency while enhancing classification performance.

1. Content moderation platforms: assist manual review, improve efficiency and accuracy; 2. Policy formulation and research: analyze hate speech propagation patterns, facilitate community management and anti-hate education; 3. Multilingual expansion: the methodology can be migrated to other languages to support global content security.

Current limitations: unstable reasoning consistency, insufficient robustness against adversarial examples, real-time performance needs optimization; Future directions: introduce constraint mechanisms to improve reasoning quality, strengthen adversarial example recognition, optimize response latency in high-concurrency scenarios.

Explainable AI is moving from academia to application, with great potential in the content security field. Through Chain-of-Thought prompting and efficient fine-tuning, this project endows AI with decision-making interpretability, which is expected to become a standard configuration for next-generation content moderation, helping to build a safer and fairer online environment.