[3Xin4] Poster session 1

[3Xin4-47] Open the Black Box of AI

Saliency Map of DUI Sentencing and Legal XAI

〇Hsuan-Lei Shao¹, Wei-Hsin Wang², Sieh-Chuen Huang², Kuan-Ling Shen² (1.National Taiwan Normal Univ., 2.National Taiwan Univ.)

Keywords:Explainable AI, drunk driving, saliency map, sentencing prediction, textCNN

This article is to construct an AI model to predict drunk driving (DUI) sentencing cases in Taiwanese Judgments. We provide a textCNN model for the four-classification sentencing range with 72% accuracy and make it explainable AI (XAI) by visualized saliency maps. The method is to observe the” saliency value” by the final output differential by every word vector. We succeed in establishing a model which can input Chinese words and pick up” salient” words. More specifically speaking, phrases such” alcohol rate in his/her breath,” “highly dangerous,” and ”recidivist” have higher saliency values. They happen to echo the provisions of the Criminal Code (the DUI article §185-3 I, the sentencing article §57, and the recidivist §47). The result of this paper can be coherent with the legal domain knowledge, being the first step in the XAI approach to legal analytics.