Scholarlink

<p>Abstractive summarization task aims to generate a concise and accurate summary that effectively captures the key information from the input document. In recent years, large language models (LLMs) have emerged as foundational models for developing domain-specific summarization systems. While these models are effective, they also tend to inherit the hallucinations produced by LLMs. Furthermore, the fundamental inconsistency between training objectives and inference protocols in summarization systems have yet to be effectively resolved. Addressing both two issues, we introduce <b>H</b>eterogeneous <b>I</b>nformation-Driven <b>Con</b>trastive optimization for abstractive summarization (HICON). The proposed method integrates the classic Tensor-Product Representation into abstractive summarization, enabling the capture of syntactic information from the source document. By integrating syntactic information with semantic features, the summarization model gains a fine-grained understanding of the document, thus generating more faithful summaries. Considering the inconsistency problem, HICON selects the generated summaries as training examples and incorporates contrastive learning loss alongside generative loss to bridge the gap between training and inference. Experimental results shows that our method achieves new state-of-the-art results on the CNN/Daily Mail dataset, with a R-1 score of 49.22. Further analysis reveals that the summaries produced by HICON exhibit greater relevance to the source document, enhanced fluency, and fewer errors.</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

Heterogeneous Information-Driven Contrastive Optimization for Abstractive Summarization

  • 期刊论文

  • 发表日期：

    2026年2月16日

查看全文
  • Peipei Wang,
  • Jingan Lu,
  • Shuai Gong

摘要

Abstractive summarization task aims to generate a concise and accurate summary that effectively captures the key information from the input document. In recent years, large language models (LLMs) have emerged as foundational models for developing domain-specific summarization systems. While these models are effective, they also tend to inherit the hallucinations produced by LLMs. Furthermore, the fundamental inconsistency between training objectives and inference protocols in summarization systems have yet to be effectively resolved. Addressing both two issues, we introduce Heterogeneous Information-Driven Contrastive optimization for abstractive summarization (HICON). The proposed method integrates the classic Tensor-Product Representation into abstractive summarization, enabling the capture of syntactic information from the source document. By integrating syntactic information with semantic features, the summarization model gains a fine-grained understanding of the document, thus generating more faithful summaries. Considering the inconsistency problem, HICON selects the generated summaries as training examples and incorporates contrastive learning loss alongside generative loss to bridge the gap between training and inference. Experimental results shows that our method achieves new state-of-the-art results on the CNN/Daily Mail dataset, with a R-1 score of 49.22. Further analysis reveals that the summaries produced by HICON exhibit greater relevance to the source document, enhanced fluency, and fewer errors.