Scholarlink

<p>The evolution of pre-trained language models (PLMs) has significantly transformed the landscape of sentiment analysis, particularly in handling complex, noisy, informal, and short-text commonly found on social media. While numerous surveys have explored PLMs and sentiment analysis separately, few provide a focused evaluation of large language models (LLMs) in the context of sentiment classification across diverse, real-world datasets. This survey addresses that gap by systematically comparing traditional sentiment analysis techniques—such as lexicon-based, machine learning, and deep learning approaches—with modern large pre-trained language models, using text-based unimodal and multimodal benchmark dataset. Our survey offers a detailed knowledge of how transformer-based architectures, including models like BERT, RoBERTa, GPT<InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="41060_2025_805_Article_IEq1.gif" Format="GIF" Height="13" Rendition="HTML" Resolution="72" Type="Linedraw" Width="19" /> </InlineMediaObject> <EquationSource Format="TEX">\(-\)</EquationSource> <EquationSource Format="MATHML"><math> <mo>-</mo> </math></EquationSource> </InlineEquation>3.5, GPT-4, and LLaMA-2, are leveraged for sentiment classification tasks. Beyond analyzing performance improvements, this paper highlights practical aspects such as fine-tuning techniques, and domain adaptability. Furthermore, it sheds light on ongoing limitations, such as computational cost, interpretability issues, domain sensitivity, bias, fairness, and ethical concerns. Through a comprehensive and comparative approach, this study offers a valuable resource to identify the most appropriate tools and methods for sentiment classification in the dynamic and data-rich environment.</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

A Survey on the Impact of Pre-Trained Language Models in Sentiment Classification Task

  • 期刊论文

  • 发表日期：

    2025年5月29日

查看全文
  • Himanshu Gautam,
  • Abhishek Gaur,
  • Dharmendra Kumar Yadav

摘要

The evolution of pre-trained language models (PLMs) has significantly transformed the landscape of sentiment analysis, particularly in handling complex, noisy, informal, and short-text commonly found on social media. While numerous surveys have explored PLMs and sentiment analysis separately, few provide a focused evaluation of large language models (LLMs) in the context of sentiment classification across diverse, real-world datasets. This survey addresses that gap by systematically comparing traditional sentiment analysis techniques—such as lexicon-based, machine learning, and deep learning approaches—with modern large pre-trained language models, using text-based unimodal and multimodal benchmark dataset. Our survey offers a detailed knowledge of how transformer-based architectures, including models like BERT, RoBERTa, GPT \(-\) - 3.5, GPT-4, and LLaMA-2, are leveraged for sentiment classification tasks. Beyond analyzing performance improvements, this paper highlights practical aspects such as fine-tuning techniques, and domain adaptability. Furthermore, it sheds light on ongoing limitations, such as computational cost, interpretability issues, domain sensitivity, bias, fairness, and ethical concerns. Through a comprehensive and comparative approach, this study offers a valuable resource to identify the most appropriate tools and methods for sentiment classification in the dynamic and data-rich environment.