An Automatically Improving Method for Generating Descriptions of Financial Data Quality Grading with LLMs

Generating descriptions for financial data quality grades (e.g., poor, fair, excellent) enhances both data quality assessment and the trustworthiness of AI models. Traditionally, grading criteria have been manually compiled by humans, a process that is time-consuming and requires domain-specific expertise. In this work, we propose an automated, automatically improving framework for describing financial data quality grades at arbitrary levels. Specifically, we first train a financial classifier to categorize data into multiple quality grades, with the theoretical capability to support arbitrary grading levels. Then, a collected list of financial hypernyms is used to optimize the description for each financial grade using two search strategies. The quantitative results show that the financial knowledge–aware editor improves description accuracy and the QWK correlation score by over 10 points respectively on a hold-out test set, while the qualitative results indicate better performance in terms of informativeness and trustworthiness. We release the code and data here$^{1}$.

$^{1}https://github.com/code4nlp1713/code$