Recommendations For Statutes Based On External Knowledge Of Statutes

Law article prediction aims to output relevant law articles given a fact description of a crime.Current works on law article recommendation task have the following limitations:1)It is difficult to distinguish similar law articles by using a single vector to select relevant law articles of cases from many candidate law articles;2)Most of the works focus on modeling single label samples,which has a gap with practical applications;3)Co-occurrence and other dependent relationships exist in the law.Such important features are ignored in the existing work.In the civil law system,the content of the law article is the main basis of the judge's judgment.It clarifies the meaning of the candidate labels of the law article prediction task,which is of great significance to the performance of the law article prediction.At present,there is little research on how to incorporate law article knowledge with law article prediction task to improve performance.In addition,current work that uses law article knowledge has above limitations as well.In order to alleviate the performance loss of the model caused by the above limitations,make up the gap between research and practical application,and further improve the model performance of the task of law article prediction by using the content of law article as external knowledge,we performed a lot of experimental research.Our work consists of the following parts:(1)This thesis proposes LAPF,a new framework of law article prediction that integrates external law article knowledge.The framework generates multiple different feature representations for the fact description information to retain richer feature information and uses specific vectors to determine whether the corresponding law articles are relevant law articles of the case.The framework uses the external knowledge of law articles to help the fact description to extract features,and further amplifies the feature differences between different law articles,making it easier to distinguish the easily confused law articles.The framework uses different vectors to carry out relevant law articles,which alleviates the performance loss caused by threshold selection in the multi-label classification task,so that multi-label samples can be well loaded into the model for training and learning,and meanwhile improves the model performance and application value.In addition,an attention mechanism is introduced in the framework to capture the dependencies between law articles,which further improves the performance of the model.(2)In order to verify the validity of the multi-representation law article prediction framework LAPF based on the external knowledge of law articles,we established the law article aware multi-representation law article prediction model(LAMM).The LAMM model encodes the content of the law article and generates a representation vector for each law article.The representation vector of the law article is used to interact with the fact description respectively to extract the local and global features of the fact description.To explore the influence of different encoding methods on the performance of the model,we encoded the content of laws based on sequential encoding and structured encoding respectively.Experiment results show that the LAMM model with the addition of law article contents improves on all performance indexes,and the performance improvement is more significant with the use of structured law article encoding.Extracting the dependency between law articles can improve the Accuracy and Recall of the law article prediction task.On the other hand,this approach gives the model some redundant noise,which decreases the Precision.(3)To compare different ways of interaction between law article information and fact descriptions,we further propose a law element aware multi-representation law article prediction model(LEMM)based on the LAPF framework.LEMM uses the representations of law article elements to interact with fact descriptions.The semantic alignment between elements of law articles and fact description can leverage the feature of fact descriptions.It contributes to distinguishing the different fact description representations which are extracted from interacting with law articles.The experiment result shows that the interaction between the law article elements and the fact description can make better use of the law article information,and the experimental performance is significantly improved.