Scholarlink

Online customer reviews significantly influence potential buyers’ decisions and provide retailers with valuable insight into buyers’ opinions on their products. Sentiment analysis is the method used to extract opinions from reviews at the document, sentence, and aspect levels using natural language processing techniques. Traditional methods of Sentiment Analysis using Parts of Speech Tags (POS) and lexicons face challenges with polysemy, where words have multiple meanings based on context, leading to incorrect sentiment associations. This affects sentiment classification performance. To mitigate polysemy in POS-based sentiment analysis, we propose identifying sentiment-bearing words in reviews using contextual information to uncover the hidden overall sentiment. Our model uses BERT-based contextual embeddings to find such terms to mitigate polysemy. Furthermore, this approach incorporates supportive sentiment terms to reduce feature-level sparsity and combines them with Latent Dirichlet Allocation (LDA) for distinct sentiment word grouping. Compared to the baseline model, the proposed model shows impressive performance and improves the average sentiment classification accuracy by 2% across 5 datasets. This method has significant potential for sentiment analysis, achieving high classification performance.

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

Refined Sentiment Analysis Using POS Features and LDA: Mitigating Polysemy and Sparsity with BERT Contextual Embedding

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Thennakoon Mudiyanselage Anupama Udayangani Gunathilaka,
  • Yuefeng Li,
  • Jinglan Zhang,
  • Prabhashini Dhanushika Manage

摘要

Online customer reviews significantly influence potential buyers’ decisions and provide retailers with valuable insight into buyers’ opinions on their products. Sentiment analysis is the method used to extract opinions from reviews at the document, sentence, and aspect levels using natural language processing techniques. Traditional methods of Sentiment Analysis using Parts of Speech Tags (POS) and lexicons face challenges with polysemy, where words have multiple meanings based on context, leading to incorrect sentiment associations. This affects sentiment classification performance. To mitigate polysemy in POS-based sentiment analysis, we propose identifying sentiment-bearing words in reviews using contextual information to uncover the hidden overall sentiment. Our model uses BERT-based contextual embeddings to find such terms to mitigate polysemy. Furthermore, this approach incorporates supportive sentiment terms to reduce feature-level sparsity and combines them with Latent Dirichlet Allocation (LDA) for distinct sentiment word grouping. Compared to the baseline model, the proposed model shows impressive performance and improves the average sentiment classification accuracy by 2% across 5 datasets. This method has significant potential for sentiment analysis, achieving high classification performance.