Financial FAQ Question-Answering System Based on Question Semantic Similarity

In the current wave of digital transformation, Frequently Asked Questions (FAQ) answering systems have become a crucial technology to replace traditional manual customer service for efficiently addressing high-frequency issues. This paper focuses on two real business scenarios within the financial industry - banking and funds. We delve into and implement FAQ question-answering systems based on question semantic similarity, suitable for both cold start and domain adaptation phases. In the banking scenario, we confront the challenge of cold start problem. To mitigate the anisotropy issues associated with pre-trained models, we employ unsupervised SimCSE, which leverages dropout as data augmentation. In the fund scenario, where an ample labeled dataset is available for fine-tuning, we introduce the improved supervised CoSENT. CoSENT leverages unified optimization criteria throughout the training and prediction stages of SBERT. Experimental results indicate that CoSENT can achieve superior sentence embeddings. Starting from real-world scenarios, we propose a practical data accumulation process for FAQ question-answering systems, spanning from the cold start phase to fine-tuning domain-adapted models. In conclusion, the FAQ question-answering systems constructed in this paper can effectively adapt to the cold start and domain adaptation requirements in different business scenarios, providing valuable practical and theoretical references for enterprises in the process of digital transformation.