Research on Hierarchical Model of BERT-CNN-BiLSTM for Long Dialog Classification

Long dialog text classification is challenging due to the complexity and extended context. Existing methods often fail to capture nuanced semantics and long-range dependencies. Our research introduces a hybrid framework integrating Bidirectional Encoder Representations from Transformers (BERT) with convolutional neural networks (CNN) and Bidirectional Long Short-Term Memory (BiLSTM) networks. BERT's contextual embeddings capture token-level semantics, CNN extracts local features, and BiLSTM models sequential dependencies in long dialog. We evaluate the model on a real-world dialog dataset from China Unicom, which contains thousands of customer service interactions across diverse scenarios and complex contexts. Due to privacy restrictions, the dataset cannot be fully released, but we plan to provide a de-identified subset for academic use. Our findings demonstrate the effectiveness of this approach in enhancing classification performance and its potential applicability across various dialog systems.