Skip to main content Skip to navigation

Tell Me a Story! Narrative-Driven XAI with Large Language Models

Project Overview

The document explores the role of generative AI, especially through explainable AI (XAI) techniques, in enhancing educational practices. It emphasizes the application of Large Language Models (LLMs) to generate narratives that elucidate AI predictions, thereby fostering greater understanding and trust among users in educational environments. Findings from multiple surveys indicate that these narrative-driven explanations significantly enhance user comprehension and decision-making, making AI technology more approachable for both novice users and data scientists. As a result, the integration of generative AI in education not only supports the learning process but also promotes informed engagement with AI systems, ultimately aiming to democratize access to AI-driven insights and tools in educational settings.

Key Applications

XAIstories, SHAPstories, CFstories

Context: Evaluating AI model predictions in educational settings such as student performance assessment and loan applications.

Implementation: The application of narrative generation using LLMs (specifically GPT-4) to create understandable explanations of AI predictions based on SHAP values and counterfactuals.

Outcomes: Improved comprehension of AI decisions by users, increased trust in AI systems, and enhanced decision-making abilities.

Challenges: Concerns about the narratives making incorrect assumptions and the potential for oversimplification.

Implementation Barriers

Technical and Usability Barrier

Current XAI methods lack the ability to produce natural language narratives that are easily understandable by non-experts. Non-expert users often find existing technical explanations (like SHAP values) difficult to interpret.

Proposed Solutions: Integrating LLMs to generate narratives that explain AI predictions in a user-friendly manner and using narrative-driven approaches (XAIstories) to present explanations in a more relatable and engaging format.

Project Team

David Martens

Researcher

James Hinns

Researcher

Camille Dams

Researcher

Mark Vergouwen

Researcher

Theodoros Evgeniou

Researcher

Contact Information

For information about the paper, please contact the authors.

Authors: David Martens, James Hinns, Camille Dams, Mark Vergouwen, Theodoros Evgeniou

Source Publication: View Original PaperLink opens in a new window

Project Contact: Dr. Jianhua Yang

LLM Model Version: gpt-4o-mini-2024-07-18

Analysis Provider: Openai

Let us know you agree to cookies