Scholarlink

The prevalence of abusive language in online platforms poses a significant challenge to ensuring safe and inclusive digital spaces. The previous detection methods face problems of domain dependence, language barriers, and knowledge limitations. To address these issues, this study proposes a novel abusive language detection system based on retrieval-augmented large language model. The system integrates across-domain and across-language capabilities by adopting multilingual embedding model and retrieval-augmented generation technology. A diverse dataset covering multiple domains and languages ​​were built, and the retrieval-augmented large language model combines instruction-tuned large language models with specific knowledge retrieval to enhance adaptability and accuracy. Experimental results show that the system has excellent performance across domain and language boundaries, achieving robust detection of abusive language with minimal reliance on predefined rules or static knowledge bases. This work provides a scalable, context-aware, and adaptable solution to advance the state of the art in abusive language detection in modern digital environments.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Can Large Language Models Effectively Implement Abusive Language Detection Across Domains and Across Languages?

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Kai Huang,
  • Guozhu Jia,
  • Nana Han,
  • Ming Bai,
  • Zhe Chen

摘要

The prevalence of abusive language in online platforms poses a significant challenge to ensuring safe and inclusive digital spaces. The previous detection methods face problems of domain dependence, language barriers, and knowledge limitations. To address these issues, this study proposes a novel abusive language detection system based on retrieval-augmented large language model. The system integrates across-domain and across-language capabilities by adopting multilingual embedding model and retrieval-augmented generation technology. A diverse dataset covering multiple domains and languages ​​were built, and the retrieval-augmented large language model combines instruction-tuned large language models with specific knowledge retrieval to enhance adaptability and accuracy. Experimental results show that the system has excellent performance across domain and language boundaries, achieving robust detection of abusive language with minimal reliance on predefined rules or static knowledge bases. This work provides a scalable, context-aware, and adaptable solution to advance the state of the art in abusive language detection in modern digital environments.