Method and system for automatic substantive verification of text content generated by language models using a two-model architecture and a reference database

Abstract

The subject of the application, schematically presented in the figure, is a method and system for automatic verification of the content of text generated by language models using a dual-model architecture and a reference database. The method for automatic verification of the credibility of specialized document content involves splitting the content generation process into two independent processing paths. In the first path, the Generating Model creates a text input, which is then passed on in the second path to an independent Auditing Model, which decomposes the text into logical statements and verifies each extracted statement by comparing it with an external, closed reference database (Ground Truth). The system generates a final result for the user only for statements positively verified by the Auditing Model.