Machine learning-based large model explainable RAG generation method

Abstract

The application discloses a large model explainable RAG generation method based on machine learning, comprising the following steps: performing semantic retrieval in a knowledge base based on a user input question to obtain relevant candidate document segments; constructing a hierarchical bidirectional propagation network by using an improved hierarchical relevance propagation algorithm, and respectively calculating forward and reverse relevance weight matrixes of generated content and the document segments; fusing the forward and reverse relevance weight matrixes to obtain a bidirectional relevance fusion matrix; adopting a moth flame optimization algorithm to dynamically optimize the fusion matrix with the accuracy of generated content and traceability transparency as target, and determining a final relevance threshold parameter set; generating a mapping labeling result based on the final optimized fusion matrix, and outputting final generated content and a traceability explanation report. The application realizes accurate and transparent explanation of the source of generated content, and significantly enhances the credibility and explainability of the large model generation result.

Images