Explainable Bert-Dtcn Framework for Customer Intent Classification in Customer Support Systems

Authors: Sachin Babulal Jadhav, Prof. (Dr.) D. B. Kshirsagar

Abstract: Customer-support chatbots increasingly operate as intent-understanding systems that must interpret short, noisy, and semantically overlapping service requests. Conventional intent-classification pipelines often rely on sparse lexical features or shallow classifiers, which may not sufficiently capture contextual relationships among tokens. This paper presents an explainable BERT-DTCN framework for customer intent classification in customer-support systems. The framework first employs Bidirectional Encoder Representations from Transformers (BERT) to obtain contextual token-level representations and then applies a Deep Temporal Convolution Network (DTCN) to learn temporal patterns through stacked dilated convolutional blocks. To improve transparency, a SHAP-based explanation layer is incorporated to identify tokens and phrases that contribute strongly to each intent prediction. The study is structured for the Bitext Customer Support Dataset containing 26,872 question-answer pairs distributed across 27 customer-service intents. Preliminary comparative evaluation under an 80:20 stratified split indicates that the proposed BERT-DTCN model improves macro F1-score over traditional TF-IDF, CNN, LSTM, and BERT-dense baselines while providing token-level explanation support for audit and error analysis.

DOI: http://doi.org/10.5281/zenodo.20641642