'This Crime is Not That Crime'—Classification and evaluation of four common crimes. (14th July 2022)

• Record Type: Journal Article
• Authors: Xu, Ke; Liu, Hangyu; Wang, Fang; Wang, Hansheng
• Is Part Of: Law, probability & risk. Volume 20:Number 3(2021)
• Page Start: 135
• Page End: 152
• Publication Date: 2022-07-14
• Subjects: Crime classification -- two-stage model -- deep learning model -- abnormal cases; Proximate cause (Law) -- Periodicals; Risk -- Periodicals; Law -- Mathematical models -- Periodicals; Law -- Methodology -- Periodicals; Probabilities -- Periodicals; Risk assessment -- Periodicals; Law -- Mathematical models; Law -- Methodology; Probabilities; Proximate cause (Law); Risk; Risk assessment; Periodicals; 340.1
• DOI: 10.1093/lpr/mgac006
• Languages: English
• ISSNs: 1470-8396

Abstract

Abstract: As the basis of criminal penalty, criminal conviction, integral to the protection of fundamental rights and freedom of people constitutes the basis and the core issue of criminal trials. Based on the data published on China Judgments Online, we proposed two types of classification models to apply the data of four common crimes from China Judgments Online and expounded their applications in identifying 'abnormal cases', defined as wrongly sentenced cases in this article. The two types of classification models we proposed are a two-stage model and two deep learning models. To construct the two-stage model, we first used three keyword-extraction models to extract the keywords and vectorize all the keywords, then used five classification models to build the two-stage model. For the deep learning models, we applied two different deep neural network models in the data to build the classifier. We then applied these two types of classification models to discover 'abnormal cases' in two steps. In the first step, we applied the two-stage model to extract the 'important words' which will significantly improve the probability of the two-stage model to classify cases into crimes of intentional injury. In the second step, we constructed a validation data set of cases whose verdicts are changed in the second instance rulings to test the 'important words' extracted in first step and the ability of the two-stage model and the two deep learning models to discover 'abnormal cases'. The results of this exercise show that: (1) 'important words' extracted in the first step are often associated with 'abnormal cases'; (2) these two types of classification models can effectively discover 'abnormal cases', but compared with the two deep learning models, the two-stage model (aka. Term Frequency-Inverse Document Frequency and Artificial Neural Network, the combination of a keyword extraction model and a classic machine-learning model) is more capable of discovering 'abnormal cases'.