Scholarlink

<p>Sentiment Analysis (SA) is a technique in Natural Language Processing (NLP) that uses computational methods to analyze the emotions expressed in user reviews. Analyzing opinions’ polarity can provide valuable insights into products based on user feedback, which can benefit producers. While many machine learning and deep learning models have demonstrated strong performance, they often face challenges in understanding complex relationships between entities, leading to missed nuances in real-world interactions. This limitation can impede a thorough understanding of user sentiment. In this article, we propose DLF-GAT, a novel Dynamic Loss Function Graph Attention Network architecture, which leverages XLM-RoBERTa embeddings for textual representation and is specifically designed to address the challenge of class imbalance in low-resource and multilingual datasets while effectively capturing complex semantic and relational dependencies in e-commerce platforms. Our model incorporates an innovative dynamic weighting mechanism that adaptively prioritizes underrepresented classes, and the attention mechanism within the graph network dynamically highlights the most influential nodes and relationships, enabling the model to capture intricate contextual and relational information with unprecedented precision. We evaluate our approach using two SA datasets: Twitter, which contains multilingual data, and DigiKala in the Persian language, which is classified as a low-resource language. Given the focus on Persian, the results demonstrate the effectiveness of DLF-GAT, achieving 70% accuracy on DigiKala and 83% on Twitter, thereby establishing it as a robust and novel approach for deepening sentiment understanding in complex, imbalanced, and multilingual contexts.</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

Dynamic loss function in graph attention networks for sentiment analysis on imbalanced datasets

  • 期刊论文

  • 发表日期：

    2026年5月7日

查看全文
  • Azar Fathipour Dehkordi,
  • Hamid Rastegari,
  • Faramarz Safi-Esfahani

摘要

Sentiment Analysis (SA) is a technique in Natural Language Processing (NLP) that uses computational methods to analyze the emotions expressed in user reviews. Analyzing opinions’ polarity can provide valuable insights into products based on user feedback, which can benefit producers. While many machine learning and deep learning models have demonstrated strong performance, they often face challenges in understanding complex relationships between entities, leading to missed nuances in real-world interactions. This limitation can impede a thorough understanding of user sentiment. In this article, we propose DLF-GAT, a novel Dynamic Loss Function Graph Attention Network architecture, which leverages XLM-RoBERTa embeddings for textual representation and is specifically designed to address the challenge of class imbalance in low-resource and multilingual datasets while effectively capturing complex semantic and relational dependencies in e-commerce platforms. Our model incorporates an innovative dynamic weighting mechanism that adaptively prioritizes underrepresented classes, and the attention mechanism within the graph network dynamically highlights the most influential nodes and relationships, enabling the model to capture intricate contextual and relational information with unprecedented precision. We evaluate our approach using two SA datasets: Twitter, which contains multilingual data, and DigiKala in the Persian language, which is classified as a low-resource language. Given the focus on Persian, the results demonstrate the effectiveness of DLF-GAT, achieving 70% accuracy on DigiKala and 83% on Twitter, thereby establishing it as a robust and novel approach for deepening sentiment understanding in complex, imbalanced, and multilingual contexts.