Scholarlink

<p>Abstractive summarization is an essential technique in natural language processing (NLP) that processes extensive textual data while retaining essential information. Pretrained transformer models, such as Pegasus-CNN/DailyMail, have gained popularity for this task due to their ability to reduce the training burden on large datasets. However, achieving high accuracy in downstream tasks requires fine-tuning these models with domain-specific datasets. This paper presents a fine-tuning strategy for the Pegasus-CNN/DailyMail model to generate high-quality summaries across diverse domain datasets. The approach utilizes two domain-specific datasets: SAMSum for conversational data and ArXiv for scientific articles. The fine-tuning process incorporated hyperparameter optimization, including adjustments to learning rate, batch size, beam size, and activation function, to enhance performance. The approach’s accuracy was measured with ROUGE indicators, showing significant performance gains. On the SAMSum dataset, the Pegasus-CNN/DailyMail model achieved ROUGE-1: 48.58, ROUGE-2: 24.23, and ROUGE-L: 45.89, closely matching the performance of PEGASUS 2B + SLiC (ROUGE-1: 54.37, ROUGE-2: 29.88, ROUGE-L: 45.89). Similarly, with the ArXiv dataset, it obtained ROUGE-1: 43.54, ROUGE-2: 22.34, and ROUGE-L: 42.23, demonstrating its efficiency in handling scientific summarization tasks. These results highlight the adaptability and robustness of the fine-tuning strategy for domain-specific 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

Pegasus-CNN/dailymail fine-tuning for domain-specific summarization: performance evaluation and hyperparameter optimization

  • 期刊论文

  • 发表日期：

    2025年7月8日

查看全文
  • Valli Mayil Velayutham,
  • Sandip Swarnakar

摘要

Abstractive summarization is an essential technique in natural language processing (NLP) that processes extensive textual data while retaining essential information. Pretrained transformer models, such as Pegasus-CNN/DailyMail, have gained popularity for this task due to their ability to reduce the training burden on large datasets. However, achieving high accuracy in downstream tasks requires fine-tuning these models with domain-specific datasets. This paper presents a fine-tuning strategy for the Pegasus-CNN/DailyMail model to generate high-quality summaries across diverse domain datasets. The approach utilizes two domain-specific datasets: SAMSum for conversational data and ArXiv for scientific articles. The fine-tuning process incorporated hyperparameter optimization, including adjustments to learning rate, batch size, beam size, and activation function, to enhance performance. The approach’s accuracy was measured with ROUGE indicators, showing significant performance gains. On the SAMSum dataset, the Pegasus-CNN/DailyMail model achieved ROUGE-1: 48.58, ROUGE-2: 24.23, and ROUGE-L: 45.89, closely matching the performance of PEGASUS 2B + SLiC (ROUGE-1: 54.37, ROUGE-2: 29.88, ROUGE-L: 45.89). Similarly, with the ArXiv dataset, it obtained ROUGE-1: 43.54, ROUGE-2: 22.34, and ROUGE-L: 42.23, demonstrating its efficiency in handling scientific summarization tasks. These results highlight the adaptability and robustness of the fine-tuning strategy for domain-specific summarization.