Scholarlink

Large language models, such as ChatGPT, have shown great promise in various natural language processing tasks, including abstractive summarization. However, these models are prone to generating hallucinated content—information not found in the original source. To tackle this issue, we introduce a mathematical framework for quantifying hallucinations and assess the effectiveness of different fine-tuning methods for mitigating this problem. We focus on fine-tuning ChatGLM due to its proficiency in both English and Chinese, as well as its ability to run on consumer-grade graphics cards. Our experiments, conducted on the XL-Sum datasets, utilize BLEU and ROUGE metrics to evaluate performance. The results reveal that P-Tuning is the most effective method for reducing hallucinations while maintaining high-quality summaries. This study not only enhances the reliability of abstractive summarization but also suggests the potential for reducing hallucinations in other NLP tasks through fine-tuning methods.

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

Parameter-Efficient Fine-Tuning of ChatGLM to Mitigate Hallucinations in Chinese Abstractive Summarization

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Yongjian Huang,
  • Simin Wu

摘要

Large language models, such as ChatGPT, have shown great promise in various natural language processing tasks, including abstractive summarization. However, these models are prone to generating hallucinated content—information not found in the original source. To tackle this issue, we introduce a mathematical framework for quantifying hallucinations and assess the effectiveness of different fine-tuning methods for mitigating this problem. We focus on fine-tuning ChatGLM due to its proficiency in both English and Chinese, as well as its ability to run on consumer-grade graphics cards. Our experiments, conducted on the XL-Sum datasets, utilize BLEU and ROUGE metrics to evaluate performance. The results reveal that P-Tuning is the most effective method for reducing hallucinations while maintaining high-quality summaries. This study not only enhances the reliability of abstractive summarization but also suggests the potential for reducing hallucinations in other NLP tasks through fine-tuning methods.