Transformer-Based Sentiment Classification for Innovative Customer Feedback Analysis in Thai Cosmetic Industry

Abstract

This research aims to develop an effective sentiment analysis system for classifying Thai cosmetic reviews, which present unique linguistic challenges such as the lack of word boundaries, the use of domain-specific terminology, and the prevalence of sarcasm in online discourse. We propose a hybrid approach that combines domain-specific pre-processing with machine learning and deep learning architectures, constructing a specialized beauty corpus and extracting sentiment-bearing keywords. We compare the performance of three models: Support Vector Machine (SVM), Bidirectional LSTM (Bi-LSTM), and fine-tuned WangchanBERTa. The experimental results demonstrate that WangchanBERTa significantly outperforms both traditional approaches, achieving an F1-score of 0.9450, compared to 0.8652 for SVM and 0.7600 for Bi-LSTM. Error analysis of the SVM model reveals specific challenges of the Thai language, such as complex negation usage and temporal sentiment shifts, which are effectively addressed by the Transformer-based architecture. The proposed system offers scalable solutions for both large e-commerce platforms and small enterprises, enabling smart customer feedback analysis that enhances business productivity and supports data-driven resource management. This study showcases the effectiveness of integrating domain-specific natural language processing techniques with pre-trained language models, providing a robust benchmark for sentiment analysis in the beauty industry and contributing to technological innovation and economic growth in Thailand's digital economy.