Research On Multi-Label Classification Method Of Criminal Cases Laws And Crimes Based On Deep Learning

With the quick advancement of AI and big data technologies,a significant amount of data collected across a variety of industries has been used to address a wide range of practical issues with the deep learning and other techniques.In the judicial field,the hot research in the judicial field is legal intelligence,and judgment prediction is one of the crucial tasks.Judgment prediction is to take the factual description text of the case as input,process it by deep learning and other methods and get the corresponding trial results.Relying on artificial intelligence technology to achieve verdict prediction can help judicial officers handle cases efficiently and save labor costs effectively.Because of the "one-to-many" nature of law prediction and charge prediction,multi-label classification methods are used in this thesis to solve these two tasks.For domainality,many studies have ignored the importance of terminology,which affects the semantic expression,and for long and complex texts,some methods tend to take only the temporal context or spatial potential semantics,while feature fusion can actually reveal more detailed information.At the same time,the role of law articles on charge classification is less studied,and the incorporation of law article information can help improve the classification performance of charges.Besides,few methods consider the confusion of the obtained charges due to the similarity of the cases,which affects the accuracy of the charge classification.To address the above shortcomings,this thesis investigates a multi-label classification method for criminal cases based on deep learning for law and crime.It mainly consists of the following aspects of work.1.Describe the pertinent theoretical underpinnings of this thesis.Including techniques for text preprocessing,two algorithms for keyword extraction,deep learning theory,and theoretic underpinnings of multi-label text classification.2.Given a multi-label classification approach using neural networks called CLSA-AJP with hybrid feature extraction for regular bars.the algorithm autonomously constructs a legal professional vocabulary in preprocessing and adds it to the lexicon of the word separation tool for domain personalized word separation.Following that,we use a shortcut to connect the stacked convolutional blocks in order to abstract the discontinuous potential feature information in text space.In the meantime,we input the Bi LSTM network to mine the temporal contextual semantics of the case text,extract the temporal nonlinear correlation and long distance dependence of the data,and assign different weights through the self-attentive mechanism.Finally,we adaptively fuse the two parts of features to improve the text expression.The experimental data are compared to demonstrate the usefulness of the algorithm,and they demonstrate that the CLSA-AJP algorithm described in the research performs better than the traditional algorithm.The algorithm achieved 81.35%,70.64%,85.23%,74.97%,and 80.10% for macro-precision,macro-recall,micro-F1,macro-F1,and average F1 values on the Cail-small dataset,respectively,which were 1.14%,2.62%,4.97%,2.25%,and3.63%.3.For the merging of information from many categories,a multi-label classification technique called CLMI-CJP is presented.This algorithm incorporates the law article information and the charge keyword information obtained from the law article multi-label classification for charge classification based on extracting the mixed features of the long text of the case.In order to highlight the case content that is compliant with the law,the text is first preprocessed,and the basic case features are extracted using deep CNN and Bi LSTM networks.Next,the results of the law classification are encoded and incorporated into the basic case features using interactive attention.The charges’ keywords are then extracted using the TF-IDF and Text Rank algorithms,the average keyword phrases for each charge are pooled,the keyword table features of the charges are combined using the attention mechanism,the keyword table information is then incorporated into the case facts,and the case features combined with the information of both parties are used to predict the charges.In order to confirm the algorithm’s validity and correctness,the experimental comparison shows that the CLMI-CJP algorithm given in the thesis outperforms the classical algorithm.The algorithm achieved macro-F1,micro-F1 and average F1 values of 86.18%,79.60%,82.89% versus 82.90%,73.44%,78.17%on the Cail-small and CAC datasets,respectively,and compared to the more effective Fact-law algorithm,the average F1 values were 2.02% and 3.42% higher on the two datasets,respectively.According to experimental findings,the CLSA-AJP multi label classification algorithm concentrates on specifics and enhances classification performance by fusing spatial and temporal characteristics of legal phrases.The CLMI-CJP multi label classification algorithm of charges deals with charge ambiguity through the interaction of the law and the case,as well as the direction of the charge keyword,in order to improve the classification.In conclusion,utilizing the multi label approach to predict laws and charges has increased the classification effect,which can help more with the completion of intelligent judicial judgment prediction.