Research on the Construction and Optimization Algorithm of Large Language Model in Financial Text Analysis

This paper discusses the construction and tuning strategies for a large-scale model for financial literature parsing that combines a hierarchical attention network and a transformer framework. Initially, given the unique properties of financial data, the paper conceived a hierarchical attention architecture to enhance the model's ability to identify the core details of long-length content. Subsequently, by refining the transformer's multi-dimensional attention mechanism, the model's ability to parse financial terms and their contextual associations was enhanced. Furthermore, this paper proposed an improvement based on a self-regulating learning rate to accelerate the training process and improve the adaptability of the model. To confirm the feasibility of this strategy, the paper performed comprehensive simulations and deployed this model on multiple financial databases. The simulation results show that our concept surpasses conventional models in terms of accuracy, collection rate, and F1 statistics, especially in financial event identification and sentiment analysis tasks. Specifically, the model's accuracy increased by 3.7 percentage points and the collection rate jumped by 4.5 percentage points. This analysis contributes novel ideas to financial literature parsing and lays an academic foundation for the optimization of large-scale language models.