A Deep Prompt-Based Chain-of-Thought Approach to Harmful Euphemism Detection in Social Networks.

Area of Science:

• Cybersecurity and Natural Language Processing
• Computational Social Science

Background:

• Harmful euphemisms on social networks disrupt digital order and cause harm like cyberbullying.
• Existing research faces challenges: lack of annotated data, limited reasoning in lightweight models, and latency issues with Large Language Models (LLMs).

Purpose of the Study:

• To develop an effective and efficient method for detecting harmful euphemisms in Chinese social networks.
• To overcome limitations of current detection models, including data scarcity and computational constraints.

Main Methods:

• Collected a large corpus from social networks and created a fine-grained annotated harmful euphemism dataset.
• Designed a representation learning framework integrating prompt-based chain-of-thought reasoning and multi-head contrastive learning with LLM knowledge.
• Proposed a multi-dimensional semantic perception fusion framework with cross-channel dynamic fusion for implicit semantics and contextual knowledge.

Main Results:

• The proposed approach significantly outperforms state-of-the-art lightweight models in harmful euphemism detection.
• Achieved highly competitive performance compared to large-scale LLMs, with substantially lower inference latency and computational overhead.
• Demonstrated improved ability to capture implicit semantics and integrate external contextual knowledge.

Conclusions:

• The developed method provides a novel technical foundation for detecting harmful euphemisms on social networks.
• Offers a practical solution balancing high performance with efficiency, suitable for real-world deployment.
• Addresses key challenges in harmful euphemism detection, improving cyberspace governance.