Scholarlink

<p>With the rapid development of human–computer interaction and natural language generation technology, chat dialogue summarization has attracted extensive attention from researchers, which aims to obtain significant information from chat history. The current chat dialogue summary models mainly focus on short dialogues due to the limitation of input text length. This paper proposes a RDPNet model, an effective multi-stage network for long chat dialogues. The method first uses a Retriever to select significant sentences from the long input dialogue, which not only can shorten the length of the source text, but also improve the structure of the chat. Secondly, a DialoGPT annotator is adopted to label the extracted important sentences so that it can further improve the structure of the dialogue. Finally, a large-scale pre-trained generation model ProphetNet is adopted to generate a concise dialogue summary. The experimental results demonstrate that RDPNet outperforms the state-of-the-art methods on three long chat summarization datasets DIALSUMM, SAMSum and TWEETSUMM and verify the effectiveness for long chat summarization.</p>

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

RDPNet: a multi-stage summary generation network for long chat dialogues

  • 期刊论文

  • 发表日期：

    2025年2月12日

查看全文
  • Shihao Yang,
  • Shaoru Zhang,
  • Huayu Zhang,
  • Chih-Cheng Hung,
  • Fengqin Yang,
  • Shuhua Liu

摘要

With the rapid development of human–computer interaction and natural language generation technology, chat dialogue summarization has attracted extensive attention from researchers, which aims to obtain significant information from chat history. The current chat dialogue summary models mainly focus on short dialogues due to the limitation of input text length. This paper proposes a RDPNet model, an effective multi-stage network for long chat dialogues. The method first uses a Retriever to select significant sentences from the long input dialogue, which not only can shorten the length of the source text, but also improve the structure of the chat. Secondly, a DialoGPT annotator is adopted to label the extracted important sentences so that it can further improve the structure of the dialogue. Finally, a large-scale pre-trained generation model ProphetNet is adopted to generate a concise dialogue summary. The experimental results demonstrate that RDPNet outperforms the state-of-the-art methods on three long chat summarization datasets DIALSUMM, SAMSum and TWEETSUMM and verify the effectiveness for long chat summarization.