Scholarlink

Sentiment analysis in Vietnamese e-commerce is crucial for understanding customer feedback and improving business strategies, yet it faces unique challenges due to the language’s linguistic complexity and the informal nature of online reviews. This paper addresses the critical gap in optimized sentiment analysis models for Vietnamese by conducting a comprehensive evaluation of state-of-the-art pre-trained language models, including PhoBERT, viBERT, XLM-RoBERTa, and their parameter-efficient variants. We fine-tune these models on the ViShopSent dataset of Vietnamese e-commerce reviews, carefully annotated for sentiment polarity. Our experiments demonstrate that XLM-RoBERTa with cross-validation achieves an accuracy of 91.54%, while parameter-efficient fine-tuning using LoRA with 2 unfrozen layers maintains competitive performance at 91.53% accuracy with 44.5 M trainable parameters (16.5% of the full model). The results reveal important insights into model behavior with Vietnamese text, including challenges with informal language, code-mixing, and domain adaptation. The paper also introduces ViShopSent as a valuable benchmark for future research in Vietnamese sentiment analysis.

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

Fine-Tuning Pre-trained Language Models for Vietnamese Sentiment Classification in E-Commerce

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Quang-Vinh Pham,
  • Ngoc-Hieu Dang,
  • Phi-Long Tran,
  • Quang-Hung Le

摘要

Sentiment analysis in Vietnamese e-commerce is crucial for understanding customer feedback and improving business strategies, yet it faces unique challenges due to the language’s linguistic complexity and the informal nature of online reviews. This paper addresses the critical gap in optimized sentiment analysis models for Vietnamese by conducting a comprehensive evaluation of state-of-the-art pre-trained language models, including PhoBERT, viBERT, XLM-RoBERTa, and their parameter-efficient variants. We fine-tune these models on the ViShopSent dataset of Vietnamese e-commerce reviews, carefully annotated for sentiment polarity. Our experiments demonstrate that XLM-RoBERTa with cross-validation achieves an accuracy of 91.54%, while parameter-efficient fine-tuning using LoRA with 2 unfrozen layers maintains competitive performance at 91.53% accuracy with 44.5 M trainable parameters (16.5% of the full model). The results reveal important insights into model behavior with Vietnamese text, including challenges with informal language, code-mixing, and domain adaptation. The paper also introduces ViShopSent as a valuable benchmark for future research in Vietnamese sentiment analysis.