Proposal of a Conversational Mahjong Review Support System Based on Contribution Visualization for Discard Decisions by Deep Learning-Based AI

Abstract

In this study, we apply contribution visualization methods such as SHAP and Grad-CAM to a custom CNN-based mahjong discard decision model developed with reference to Suphx, in order to identify which features the AI focuses on during decision-making. By constructing a mechanism that explains the visualized results—such as heatmaps obtained from Grad-CAM and contribution scores from SHAP—in natural language using the ChatGPT API, we propose a conversational support system that enables users to understand and critically examine the AI’s reasoning process. Conventional review systems, such as NAGA, merely present numerical outputs such as discard choices and win rates and lack mechanisms to explicitly indicate the basis for such decisions. In contrast, our study visualizes and verbalizes the rationale behind the AI’s discard decisions, enabling users to engage in interactive review sessions with the AI. This bidirectional support system represents a novel approach. We also plan to evaluate its effectiveness and educational impact through comparison with conventional systems.