Research On Psychological Attribution Analysis Of Violent Crime Based On Text Classification Technology

Social security situation of China is generally good.However,the social is still harmfulness because of the particularity of violent cases.The concept of "Intelligent Justice"has promoted the widespread application of artificial intelligence in the judicial field,which can provide a more automated method to study criminal psychological attribution.At present,research of attribution analysis relies on a large amount of professional knowledge.And there are few studies to analyze the psychological attribution of violent crimes by text classification.Due to the sensitivity of judicial data,there is a lack of public data sets.The description of criminal facts is not standardized,there are issues such as inaccurate word segmentation,and attribution feature engineering that cannot well map the relationship between the criminal facts and attribution types.In addition,there is a problem of uneven distribution of different types of samples,resulting in class-imbalance.To solve the above problems,in this paper,different types of features are fused to construct the psychological attribution system for violent crimes.And a new attribution method based on the imbalance data has been constructed for class-imbalance.The main research content and results of this paper are as follows:(1)Due to the particularity of judicial data and the lack of datasets,in this paper,the dataset of the psychological attribution of violent crimes has been established by criminal facts that can describe the process of criminal violence as the static factors.The content of"judgment determination content" and "accusation content" are extracted in the judgment documents as the research content.And according to the generalized three types of aggressive behavior,the attribution types of violent crimes can be divided into the premeditated,impulsive,and pathological type.The feature dictionaries,which can further quantify the psychological attribution types,has been constructed to improve the accuracy of word segmentation.(2)To solve the problems that the description of criminal facts is not standardized,the length of criminal facts is different,the single model cannot effectively extract key attribution features,and the feature engineering cannot well map the relationship between the criminal facts and attribution types,the paper proposes a method of criminal psychological attribution analysis based on multi-feature fusion.The text and numerical features are fused to obtain the attribution features comprehensively,to explore the psychological attribution factors of criminal violence.Firstly,to quantitatively analyze the relationship between criminal facts and attribution types,the frequency of keywords in criminal facts is calculated based on the predefined dictionaries,and numerical features are constructed to represent the degree of inclination of criminal facts in specific attribution types.Secondly,the semantic features of criminal facts are extracted based on the improved CNN(Convolutional neural network).And the global average pooling layer is introduced to replace the pooling layer and full connection layer,which can obtain important local semantic features and prevent overfitting.Finally,the concat technology is used to fuse numerical features and text features,to fully represent the attribution features of criminals.To verify the effect of the method,the experiments are conducted for compassion by the follows:CNN,LSTM(Long Short-Term Memory),SVM(support vector machines)based on TF-IDF and other machine learning methods.Experimental results show that the proposed method is better in accuracy,recall,precision,and F1 in the micro and macro scopes.And a radar chart has been drew based on the weight distribution obtained by softmax,and corrective education measures can be provided for relevant departments.(3)To solve the problem of class-imbalance in the above method,a method for assessing criminal psychological attribution based on unbalanced data was proposed to enhance the weight of minority features and improve ability of classification,starting from feature selection.Firstly,a new statistical method is constructed,B-TF-dIDF(Balanced-Term Frequency-distinguishing Inverse Document Frequency),to extract numerical features of keywords.The balanced factor is added to TF-IDF to balance the weight of different types of keywords,and the distinguishing factor is added to IDF to distinguish the attribution categories.Secondly,a hybrid neural network model composed of LSTM and CNN is constructed to fully represent the semantic features of criminal texts and balance the impact of different length samples on classification.Spatial Attention is introduced into the multiview feature maps to evaluate the impact of different semantic features on attribution results.Finally,the self-attention is constructed to fuse the numerical and test features,and the features with greater impact on attribution results are in higher weight.The experiment proved that the proposed model is better than other related methods in micro and macro scopes of various evaluation indicators.Additionally,the F1 of minority samples has increased by 6%-8%,indicating that the proposed method can reduce the impact of classimbalance.