Scholarlink

In the current information age, effectively understanding and categorizing large amounts of text data is paramount. Sequence classification, a core task in Natural Language Processing (NLP), addresses this need by assigning predefined categories to text sequences. This study explores prompt-based fine-tuning of Pretrained Language Models for Bengali sequence classification, addressing tasks such as sentiment analysis, bullying detection, and news categorization. We propose two prompting methods: Instruction Prompting, which provides carefully crafted instructions to guide the model through the task, and Label Information Prompting, which includes label information in the prompt to enhance the model’s understanding of the relationship between the text and the label. Our experiments demonstrate the potential of prompt-based fine-tuning in enhancing the performance of Bengali sequence classification tasks. Our method outperforms traditional fine-tuning methods with an increased accuracy of 2–4% across four publicly available datasets by employing only 20% of training data.

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

Prompt-Based Fine-Tuning of Bangla Language Models for Sequence Classification

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Asif Mahmud,
  • Md Shariar Hossain,
  • Md Rasel Meya,
  • Didarul Islam Didar,
  • Hasibul Islam,
  • Puja Dhar,
  • Poly Akter

摘要

In the current information age, effectively understanding and categorizing large amounts of text data is paramount. Sequence classification, a core task in Natural Language Processing (NLP), addresses this need by assigning predefined categories to text sequences. This study explores prompt-based fine-tuning of Pretrained Language Models for Bengali sequence classification, addressing tasks such as sentiment analysis, bullying detection, and news categorization. We propose two prompting methods: Instruction Prompting, which provides carefully crafted instructions to guide the model through the task, and Label Information Prompting, which includes label information in the prompt to enhance the model’s understanding of the relationship between the text and the label. Our experiments demonstrate the potential of prompt-based fine-tuning in enhancing the performance of Bengali sequence classification tasks. Our method outperforms traditional fine-tuning methods with an increased accuracy of 2–4% across four publicly available datasets by employing only 20% of training data.