Customer Query Classification Based on DistilBERT and TextCNN

Customer service representatives handle customer inquiries and resolve issues. However, due to the high volume of inquiries, it is challenging for customer service representatives to provide timely services. To address this issue, natural language processing (NLP) and deep learning techniques are applied to analyze customer queries within a customer query management system. This study explores a case involving a Thai company's customer question classification using several deep learning approaches. Initially, the text is preprocessed through tokenization and text vectorization using Word2Vec, Universal Sentence Encoder (USE), RoBERTa, and DistilBERT. Question classification uses machine learning and deep learning techniques, including CNN, Bi-LSTM, SVM, TextCNN, and Fine-tuning RoBERTa/ DistilBERT. The experimental results show that the DistilBERT encoding technique and TextCNN classifier yield the best performance, achieving an accuracy of 91.59%, recall of 91.59%, and F1-score of 91.61%.