Prediction Model for Drunk Driving Sentencing: Applying TextCNN to Chinese Judgement Texts

Drunk driving cases often arouse public concern in Taiwanese society. According to 2013 amended Paragraph 1 of Article 185–3, the Taiwan Criminal Code, a drunk person may face up to a maximum sentence of two years in prison if his/her exhalation contains alcohol of 0.25 mg per liter or more, or blood alcohol concentration is 0.05% or more. The huge volume of “drunk driving” cases becomes a considerable workload for the court and therefore it may be worthwhile developing an automatic sentencing supportive system. This research attempts to train a deep-learning model to predict sentences by inputting the section of “recidivist/facts of the judgement.” The TextCNN (Convolutional Neural Networks) model reached a 73% accuracy rate in four-category sentencing prediction. This research suggests that adopting two kinds of pre-processing methods and a well-trained model directly to unstructured judgement texts without word segmentation can result in a good performance. It opens the possibility of applying different machine learning techniques to legal texts.