Large Language Model-Supported Label Expansion for Multi-Label Classification of Turkish E-Commerce Reviews

Abstract

This study addresses the multi-label text classification problem on Turkish customer reviews collected from e-commerce platforms. The goal is to automatically label each review with multiple categories simultaneously, such as price, delivery, product quality, and satisfaction. First, a baseline model was established using One-vs-Rest Logistic Regression on Term Frequency-Inverse Document Frequency (TF-IDF) representation as a classical approach. Subsequently, BERTurk and ELECTRA-Turkish pre-trained Transformer models were trained using multi-label fine-tuning. The distinctive aspect of the study is the semi-automatic expansion of the label pool using a large language model prior to model training and the enrichment of the final target outputs by normalizing the obtained labels. Experiments were evaluated using subset accuracy and micro/macro/weighted F1 metrics on the validation set. The results show that the TF-IDF-based method achieved 41.43% subset accuracy and 82.42% micro-F1, while Transformer-based models produced higher performance (BERTurk: 54.23% subset accuracy, 87.95% micro-F1, 63.51% macro-F1, 85.95% weighted F1; ELECTRA-Turkish: 51.50% subset accuracy, 87.05% micro-F1, 58.03% macro-F1, 84.90% weighted F1). © 2026 IEEE.