Scholarlink

The proliferation of online user-generated content has raised concerns about the presence of toxic language, which can be harmful and perpetuate societal biases. Sequence-to-sequence (seq2seq) language models have shown remarkable progress in natural language generation tasks, including text summarization. However, these models can inadvertently generate toxic or offensive content, reflecting biases present in their training data. This study proposes a novel approach to enhance the detoxification capabilities of seq2seq language models for dialogue summarization by leveraging Reinforcement Learning (RL) techniques. The methodology involves fine-tuning state-of-the-art models, such as FLAN-T5, BART, and GODEL, using a reward model based on a hate speech classifier. The fine-tuned models are evaluated for their ability to generate less toxic summaries while preserving the essential information from the input dialogues.

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

Enhancing Detoxification of Text Summaries with Seq2seq Language Models Using Reinforcement Learning Based Fine-Tuning

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • G. Bharathi Mohan,
  • R. Prasanna Kumar,
  • S. Akilesh Rao,
  • Y. Jaikanth,
  • G. Vishal Aadiith,
  • Elakkiya R

摘要

The proliferation of online user-generated content has raised concerns about the presence of toxic language, which can be harmful and perpetuate societal biases. Sequence-to-sequence (seq2seq) language models have shown remarkable progress in natural language generation tasks, including text summarization. However, these models can inadvertently generate toxic or offensive content, reflecting biases present in their training data. This study proposes a novel approach to enhance the detoxification capabilities of seq2seq language models for dialogue summarization by leveraging Reinforcement Learning (RL) techniques. The methodology involves fine-tuning state-of-the-art models, such as FLAN-T5, BART, and GODEL, using a reward model based on a hate speech classifier. The fine-tuned models are evaluated for their ability to generate less toxic summaries while preserving the essential information from the input dialogues.