A Deep Learning Approach to Financial Text Similarity Using FinBERT

In the field of finance, text similarity analysis plays a pivotal role in financial decision-making, market forecasting, and risk management. However, due to the specificity of financial texts, which often involve domain-specific terms and complex syntactic structures, conventional text encoding methods have shown limited effectiveness in this domain. This study adopts FinBERT, a model pre-trained on financial corpora, as the basis for text encoding to better capture the semantic information of financial texts. To further enhance feature representation, we combine Self Attention and Convolutional Neural Networks for capturing local text features. By leveraging the strengths of both CNN and self-attention mechanisms, the model achieves a more comprehensive understanding of text similarity. Additionally, this paper introduces a novel heuristic fusion function to integrate information. The heuristic fusion function employs a gating mechanism to control the fusion ratio, mitigating the introduction of noise during fusion. Experimental results on the Ant Financial ATEC dataset in this paper demonstrate that our approach achieves optimal performance, underscoring the effectiveness of our approach in financial text similarity tasks.