A Multi Granularity GCN Model For Statute Recommendation

In recent years,”Intelligent Courts” have developed rapidly.Take artificial in-telligence and big data technology as the core,it assists in filing judicial documents,recommending legal provisions involved in cases,analyzing charges of cases,and pre-dicting penalty term,which not only effectively improves the efficiency of judicial trial process,but also greatly improves the quality of trial,and promotes the modernization and intelligence of judicial trial process.This thesis focuses on the recommendation of legal articles,through analyzing and mining the semantic information in the description of the facts of the case to recom-mend the legal provisions consistent with the facts of the case.Recommendation of law articles is of great research value.It can not only help professional judicial per-sonnel to quickly screen out relevant laws and articles to improve work efficiency,but also provide legal consulting services for ordinary people to facilitate their access to professional legal advice and save consulting costs.The existing law recommendation work is based on the paradigm of classification framework,and the case description is divided into char level or word level features and input into feature encoders,such as RNN and Transformer.However,a single level of features can not represent the structural information of the document and can not cover specific information from other granularity,thus losing the richness of the multi-granularity semantics.In addition,the use of word segmentation tools to parse Chinese text is easy to cause semantic ambiguity,including word segmentation errors and ambiguity caused by different word segmentation results.More importantly,there are only slight differences between some cases,and how to distinguish these easily confused cases is a major difficulty in the recommendation of the law.In order to solve the above problems,this thesis proposes a method of law recom-mendation based on Graph Convolutional Network multi-granularity coding.Aiming at the shortcoming of single granularity coding in existing methods,this thesis adopts four-level coding structure,fully represents granularity features of chars,words,sen-tences and texts based on Attention Mechanism,and excavates potential semantic and structural information in texts.At the same time,the multi-granularity coding struc-ture can also alleviate the long-term dependency problem caused by longer text.To solve the problem of semantic ambiguity caused by word segmentation,this thesis uses Graph Convolutional Network to model all possible word segmentation results and ex-tract global lexical information to eliminate word segmentation ambiguity.In order to distinguish easily confusable cases,this thesis automatically selects words with clas-sification ability to construct a keyword glossary.Based on this glossary,keywords in case description are extracted and incorporated into the multi-granularity encoder as auxiliary features using Attention Mechanism.In the experimental analysis phase,this thesis compares the leading baseline mod-els in multiple fields on the CAIL2018 small open dataset.The experimental results show that the proposed model is higher than the baseline model in terms of accuracy,macro precision,macro recall and macro F1 score.Meanwhile,the effectiveness of each module is verified by ablation experiments in this thesis.Finally,a case study is given to prove that the model can distinguish confusion cases effectively.