Scholarlink

Text summarization and keyword extraction are essential for processing Vietnamese natural language processing (VNLP), a low-resource language with limited datasets and complex features like tonal structure and morphology. Multi-task learning (MTL) offers a promising approach by sharing knowledge across tasks, potentially improving performance in VNLP. This study evaluates single-task models such as TextRank, LexRank, and T5-small for summarization, and Bi-LSTM, T5-small and KeyBERT [18] for keyword extraction. A custom collection of 32,521 Vietnamese news articles, annotated with abstractive summaries and human-generated keywords, was employed in this study. Experiments on single-task settings show that transformer-based models like T5-small (ROUGE-L = 0.3390) and PhoBERT (F1-score = 0.3586) outperform traditional models such as TextRank and Bi-LSTM, with higher inference cost. To address the trade-off between accuracy and computational efficiency, we propose MTL strategies that employ shared encoder architectures and combined loss functions to jointly optimize both summarization and keyword extraction. A multi-task baseline is recommended at least ROUGE-L and F1-score of 0.3, with a reduction in inference time of at least 10–30% compared to separate single-task models. These findings support the potential of MTL as a unified and efficient approach to Vietnamese text processing and lay the groundwork for future development of real-world VNLP applications.

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

An Empirical Study of Deep Learning Approaches Toward a Multi-Task Learning Framework for Vietnamese Text Summarization and Keyword Extraction

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Dinh Thu Khanh,
  • Hoang Duc Trung,
  • Vu Duc Thi,
  • Le Minh Tuan,
  • Nguyen Long Giang

摘要

Text summarization and keyword extraction are essential for processing Vietnamese natural language processing (VNLP), a low-resource language with limited datasets and complex features like tonal structure and morphology. Multi-task learning (MTL) offers a promising approach by sharing knowledge across tasks, potentially improving performance in VNLP. This study evaluates single-task models such as TextRank, LexRank, and T5-small for summarization, and Bi-LSTM, T5-small and KeyBERT [18] for keyword extraction. A custom collection of 32,521 Vietnamese news articles, annotated with abstractive summaries and human-generated keywords, was employed in this study. Experiments on single-task settings show that transformer-based models like T5-small (ROUGE-L = 0.3390) and PhoBERT (F1-score = 0.3586) outperform traditional models such as TextRank and Bi-LSTM, with higher inference cost. To address the trade-off between accuracy and computational efficiency, we propose MTL strategies that employ shared encoder architectures and combined loss functions to jointly optimize both summarization and keyword extraction. A multi-task baseline is recommended at least ROUGE-L and F1-score of 0.3, with a reduction in inference time of at least 10–30% compared to separate single-task models. These findings support the potential of MTL as a unified and efficient approach to Vietnamese text processing and lay the groundwork for future development of real-world VNLP applications.