Scholarlink

Customer feedback on online ordering and delivery platforms is crucial to business performance. Food delivery companies (FDS) use client feedback to improve satisfaction. A lot of English-language research has focused on sentiment analysis (SA). Despite its growing use as an Internet writing language, Arabic sentiment analysis has received little research, and there are few datasets or lexicons available for it. This study uses Arabic FDS reviews for in-depth emotion mining. We utilize RNN, mBERT, RoBERTa, and AfriBERT to demonstrate the effectiveness of the recommended technique in assessing FDS sentiment. The study uses oversampling methods like Generative Pre-trained Transformer to classify emotions from an imbalanced dataset. We assess accuracy, F1-score, recall, and precision for emotion classification. The examination compares strategies with and without dataset preprocessing. Comprehensive trials demonstrate the efficiency of the framework and the superiority of transformer models over standard deep learning models. The AfriBERTa model outperforms baseline norms with an F1 score of 0.92 and an accuracy of 0.919.

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

How Can Large Language Models and Data Augmentation Improve Customer Satisfaction on Food Delivery Services?

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Nouri Hicham,
  • Nassera Habbat

摘要

Customer feedback on online ordering and delivery platforms is crucial to business performance. Food delivery companies (FDS) use client feedback to improve satisfaction. A lot of English-language research has focused on sentiment analysis (SA). Despite its growing use as an Internet writing language, Arabic sentiment analysis has received little research, and there are few datasets or lexicons available for it. This study uses Arabic FDS reviews for in-depth emotion mining. We utilize RNN, mBERT, RoBERTa, and AfriBERT to demonstrate the effectiveness of the recommended technique in assessing FDS sentiment. The study uses oversampling methods like Generative Pre-trained Transformer to classify emotions from an imbalanced dataset. We assess accuracy, F1-score, recall, and precision for emotion classification. The examination compares strategies with and without dataset preprocessing. Comprehensive trials demonstrate the efficiency of the framework and the superiority of transformer models over standard deep learning models. The AfriBERTa model outperforms baseline norms with an F1 score of 0.92 and an accuracy of 0.919.