Persistent Multi-Method XAI Architecture for Financial Sentiment Analysis

A novel framework for explainable AI in financial sentiment analysis conceptualizes XAI artifacts as enduring, searchable entities. The architecture archives LIME feature attributions, occlusion-derived word importance metrics, and saliency heatmaps in distributed storage compatible with S3, complete with structured metadata and summaries in natural language. This setup facilitates semantic searches across explanation histories and enables automatic reconstruction of indexes following system outages. Additionally, a retrieval-augmented generation (RAG) assistant supports the triangulation of explanations, allowing for the comparison of results from various XAI techniques on the same prediction through natural language interactions. The goal is to ensure that AI explanations are persistent, validated, and easily accessible for decision-makers in financial organizations.