Efficient opinion mining for imbalanced customer reviews in last-mile services

Abstract

Last-mile (LM) service manages the final stage of delivering products to customers in supply chains and logistics. Consumer opinion mining has recently become essential for providing high-level LM service quality. However, existing methods face challenges with domain-specific terminology and class imbalance. Therefore, we propose LM-BERT, a BERT-based text classification model specialized in LM service sentiment analysis. In addition, we introduce a teacher–student LM-BERT framework that alleviates data imbalance in online e-commerce reviews through high-confidence pseudo-labeling. After evaluating six Transformer models, KLUE-BERT was identified as the most suitable for our baseline. Experimental results demonstrate that domain-specific knowledge transfer improves performance by 1.78 % on seen data and 1.31 % on unseen data. Statistical verification and explainable artificial intelligence techniques were employed to confirm the reliability of our approach to enhance qualitative performance and expand domain knowledge. We also conducted an ablation study confirming that high-confidence pseudo-labeling (t = 0.99) outperforms the traditional resampling method. The proposed LM-BERT model can effectively support LM service quality evaluation and management based on the voice of the customer in e-commerce.