Research On Statutes Recommendation Based On Prior Knowledge

In the current situation of increasing social dispute,a large number of judicial cases have flooded into the courts.And the limitations of judicial resources have caused judges to face severe work pressure.Under the background of comprehensively promoting the informatization construction of the people’s courts,a large number of judgment documents have been publicized.And the Artificial Intelligence of Law" has become the focus of many researches.Through the excavation and analysis of the judgment documents,the court can be provided with more intelligent technology to improve office efficiency.And it can also provide intelligent legal counseling services to the public and promote judicial access.The Statutes Recommendation is regarded as one of the techniques to achieve these goals.On one hand,it can recommend useful statutes for the judges and improve the efficiency of work;on the other hand,it can help people to cross the judicial gap and know the trial tendency in advance.And then,form the best litigation strategy.This thesis analyzes the Statutes Recommendation in detail,and introduces the data characteristics of the judgment documents,including a wide variety,more specific legal words and semi-verbalization.It is difficult to provide truly intelligent services only from the character matching level.LSTM(Long Short-Term Memory)and Attention Mechanism can mine the semantic feature of documents,and focus on the important information.In this thesis,Statutes Recommendation is completed by analyzing the semantic feature of the judgment documents using the LSTM and Attention Mechanism.First of all,this thesis proposes the pre-processing method for the characteristics of the judgment documents,including standardization of the statutes,standardization of the cases,word segmentation,establishment of legal stopwords and removing stopwords.In order to verify the effectiveness of adding the prior knowledge,this thesis proposes two methods for recommending the statutes.They are the methods based on the dynamicLSTM and the LDA-LSTM(Latent Dirichlet Allocation,Long Short-Term Memory).(1)Statutes Recommendation method based on dynamic-LSTM:The judge often refers to the content of the case when judging.And the cited statutes are actually related to the content.The method analyzes the content of the case and mines the internal relationship between the case and statutes.In this thesis,the variable-length sequence is processed by dynamic-LSTM,and the semantic vector of the judgment document is obtained by deep learning.Based on the semantic vector,multi-label classification is performed to complete the task of the Statutes Recommendation.(2)Statutes Recommendation method based on LDA-LSTM:By simulating human attention,the judge will additionally pay attention to the words related to the topic of the documents when reading the judgment documents.By introducing the topic vector of the judgment document as a priori knowledge,the Attention Mechanism is used to assign greater weight to the topic-related words when calculating the attention vector.Based on the attention vector,multi-label classification is performed to complete the task of Statutes Recommendation.Different from the traditional Bag-of-Words model like TFIDF to extract text features,the two methods mentioned above can extract the sequence feature of the documents.The LDA-LSTM model can also give the words related to the topic greater weight by adding the prior knowledge.In the experimental verification,this thesis designs a series of comparative experiments with the judgment documents of six Cause-of-Actions as the data set.The effectiveness of LDA-LSTM method which has introduced the prior knowledge is verified by comparing the experimental results of the two proposed methods and other methods.Then,the advantages and disadvantages of all methods are summarized.Finally,the further research work is prospected.