Problems of developing automated contract verification systems

This paper is dedicated to the development of an automated contract verification system and the challenges encountered during its implementation. The relevance of the research lies in addressing the tasks of optimizing the work of contract review experts. The study identifies key problems in this area, including: 1) the lack of a standardized risk assessment system, which complicates decision-making for decision-makers; 2) the lack of data for training the system, limiting its ability to effectively detect errors and correctly interpret legal terms; 3) insufficient control over the process when using large language models, which, despite their powerful analytical capabilities, do not always guarantee the accuracy and consistency of conclusions; 4) lack of consideration for context and semantic variability in legal texts, which complicates accurate classification and error detection, especially in cases where the same legal terms or phrases may have different meanings depending the context. As a result of the research, the following findings were obtained: 1) an original risk classification system for legal errors based on the fuzzy set theory; 2) algorithms for the operation of the legal document review system; 3) methods for generating data to fine-tune BERT models and adjust GPT models. Furthermore, the article proposes a combined approach using GPT and BERT models (poly-model system) to optimize the process: the GPT model processes the text and identifies errors based on predefined prompts, while the BERT model is used to assign risk levels to the errors and control the results of the GPT model, ensuring a more reliable and consistent analysis. The article also emphasizes the need to create open datasets of contracts based on the Russian law. Creating such resources will not only improve the quality of training for artificial intelligence systems but also enhance the accessibility and transparency of information for researchers and practitioners in the field of law. This will be an important step toward improving automated legal document analysis and increasing overall legal confidence when working with them.