Go to ACL ARR 2025 May homepageA Multi-Model Collaborative Framework for Child Internet News Risk Management and Value Guidance
Abstract: With the widespread proliferation of the internet among children, residual toxic content and the absence of value-oriented guidance in online news have emerged as pressing safety challenges. This paper proposes a multi-model collaborative framework for children’s news rewriting — CRV‑LLM (Children’s Risk-control and Value-guidance Large Language Model) — designed to conduct in-depth risk identification and precise rewriting across four key dimensions: vocabulary, events, headlines, and values. CRV‑LLM integrates four lightweight risk detection models with a DeepSeek-R1-Distill-Qwen-32B rewriting model, achieving effective removal of potentially harmful information and embedding of positive value guidance, all while ensuring readability for young audiences. Experimental results demonstrate that CRV‑LLM outperforms mainstream models on core indicators such as safety and educational value, with a 62% improvement in inference efficiency. This work offers an efficient, scalable technical solution for the safe management of children’s online news content.
Paper Type: Long
Research Area: Generation
Research Area Keywords: interactive and collaborative generation,text-to-text generation, bias/toxicity,distillation,efficient models
Languages Studied: Chinese
Submission Number: 1633
Loading