A Coarse-to-Fine Text Matching Framework for Customer Service Question Answering.

Customer service question answering have recently seen increased interest in NLP due to their potential commercial values. However, existing methods are largely based on Deep Neural Networks (DNNs) that are computationally expensive and memory intensive, which hinder their deployment in many real-world scenarios. In addition, the customer service dialogue data is very domain-specific, and it is difficult to achieve a high matching accuracy without specific model optimization. In this paper, we propose CFTM, A Coarse-to-Fine Text Matching Framework, which consists of Fasttext coarse-grained classification, and Roformer-sim fine-grained sentence vector matching. This Coarse-to-Fine structure can effectively reduce the amount of model parameters and speed up system inference. We also use the CoSENT loss function to optimize the Roformer-sim model according to the characteristics of customer service dialogue data, which effectively improves the matching accuracy of the framework. We conduct extensive experiments on CHUZHOU and EIP customer service questioning datasets from KONKA. The result shows that CFTM outperforms baselines across all metrics, achieving a 2.5 improvement in F1-Score and a 30% improvement in inference time, which demonstrates that our CFTM gets higher response accuracy and faster interaction speed in customer service question answering.