Research On Intelligent Auxiliary Technology For Administrative Enforcement Documents

In recent years,under the background of intelligent judicial construction,various judicial fields,including administrative enforcement,are transforming into intelligence ones.And administrative agencies at all levels are committed to continuously promoting integration of law and artificial intelligence in the field of administrative enforcement.The increasing volume of administrative enforcement documents has put a great burden on administrative law enforcement personnel.How to assist law enforcement personnel to make punishment decision faster and more efficiently through intelligent assistive technology is one of the key points in the application research of administrative enforcement.For example,the cause of action determination technology for administrative enforcement documents enables the machine to automatically complete the encoding,selection and abstraction of judicial important information,and then convert it into the cause of action information that people can understand.While improving the efficiency of information processing,the cause of action determination technology can also assist administrative law enforcement personnel to grasp the core outline of administrative cases faster and more clearly,and thus make more precise punishment decisions.In addition,in the process of administrative enforcement,the recommendation of similar administrative enforcement documents also attracts much attention,which is also one of the focuses in administrative justice application research.The application of intelligent technology in the field of administrative enforcement requires a relatively complete construction of resources as the basic guarantee.However,there is currently a lack of exclusive datasets of administrative enforcement documents,which has greatly slowed down the integration between artificial intelligence and administrative enforcement.In addition,how to deal with the long-distance dependence phenomenon in administrative enforcement documents is the primary problem faced by the task of cause of action determination for administrative enforcement documents.And how to summarize the legal facts in the administrative enforcement documents and make a high-precision cause of case determination is also an urgent problem to be solved.Finally,for recommendation tasks of similar administrative enforcement documents,the modeling of long texts remains an unavoidable challenge.In addition,administrative enforcement documents contain a large amount of legal terminology,and it is difficult to fully utilize the legal information contained in administrative enforcement documents through traditional methods or existing pre-trained language models.Aiming at the above problems,in order to improve the resource informatization construction in the field of judicial administration,this thesis establishes an Administrative Enforcement Documents Dataset(AED).To solve the problem of long-distance dependence in the task of cause of action determination for administrative enforcement documents,a segmented model of the cause of action determination is proposed.To solve problems of longdistance dependence and legal information not being fully utilized in long texts,a recommendation model for administrative enforcement documents is proposed,which based on Graph Convolution Neural Network(GCN)and Pre-trained language model(PLM).Specifically,the main contributions of this thesis are as follows:(1)A dataset named AED that dedicated to administrative enforcement documents is constructed to promote the construction of information resources for judicial administration.Firstly,an original administrative enforcement documents dataset is constructed by the administrative enforcement cases published by the China’s Market Supervision Administration.Then,the professional legal team formulate the criteria for evaluating similar documents under the guidance of the document issued by the Supreme People’s Court of China on case search related to the Chinese legal system.And the similarity of 60 query documents and 7200 candidate documents is marked according to the criteria.Finally,the AED dataset is built.(2)A segmented cause of action model for administrative enforcement documents is proposed,which mainly uses text summarization technology to summarize the key legal information in administrative enforcement documents,and then abstracts the case of cause information of administrative enforcement documents.The proposed model mainly adopts a two-segment structure including an extractive summarization model and an abstractive cause of case model,which solves the long-distance dependency problem in the task of cause of action determination.The idea of contrast learning is also introduced to alleviate the exposure bias problem in the abstractive model.Among them,the extractive summarization model mainly consists of two parts,the BERT model and the document-level sentence encoder of Transformer structure,which is used to initially process administrative law enforcement documents to solve the problem of the long-distance dependency.And the extractive summarization model extracts key legal information from original law enforcement documents under the strategy of ensuring the integrity of legal information,which forms a preliminary summary as the training corpus for abstractive cause of case model.The cause of case model takes T5-PEGASUS as the main body,which is used to generate pseudo-optimal cause of case and candidate causes of case.And a contrast loss is constructed in the contrast scoring module based on the idea of contrast learning.Finally,the effectiveness of the algorithm is verified by experiments on the AED and CAIL2020 databases.(3)A recommendation model for similar administrative enforcement documents is proposed,which based on GCN and Pre-trained language model,focusing on the problem of long text modeling in the administrative law enforcement instrument recommendation process and the problem that the structural and legal information of administrative enforcement documents cannot be fully utilized.In this paper,the administrative law enforcement instrument retrieval technique is applied to the administrative law enforcement instrument recommendation task and the instruments matched by retrieval are recommended.In this thesis,an administrative enforcement documents retrieval method,which recommends the administrative enforcement documents matched by retrieval,is applied to the recommendation task for administrative enforcement documents.Firstly,this thesis constructs a keyword graph for a pair of administrative enforcement instruments based on legal keywords,and then combines Siamese Graph Convolution Network(SGCN),Global Graph Perception Attention(GGPA)and Graph Interaction Network(GIN)to obtain the graph-level interaction features of the document pair.After that,a text pre-trained module,which based on Lawformer that is a legal Pre-trained language model,is added to recommendation model to enable the model to learn more semantic information and legal knowledge.Finally,administrative enforcement document cause of case information is introduced as additional knowledge into the text Pretrained module to improve the recommendation accuracy.The experimental results on LeCaRD and AED datasets show that the proposed model has better performance compared with other advanced models.