A public interest litigation case clue extraction and recommendation method based on a knowledge graph

Abstract

The application discloses a kind of based on knowledge graph's public interest litigation case clue extraction and recommendation method, comprising the following steps;Step one, case clue extraction;Using python crawler technology crawls case data, in case clue extraction, first using python crawler technology, from the typical website on Internet collection clue source data, also support user upload data simultaneously;By constructing case source knowledge graph, uniformly manages case source data, constructs typical case, laws and regulations, power list etc. Entity, and the mapping relationship between entities;Using pre-training BERT and graph embedding model to realize recommendation, not only based on semantic information, also integrated graph structure information, can realize more accurate, diversified case / laws and regulations / power list recommendation;Through the above steps, realize the automated case clue extraction and recommendation, greatly improve the efficiency and accuracy of public interest litigation judgment case.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of computers, and particularly relates to a public interest litigation case clue extraction and recommendation method based on a knowledge graph. BACKGROUND

[0002] Public interest litigation refers to litigation initiated by specific state organs, organizations and individuals against illegal acts. Public interest litigation aims to protect public interests, and its purpose is preventive. However, the prosecution subject of public interest litigation is increasingly diversified, and the plaintiff is not required to be an interest victim. Therefore, in public interest litigation, prosecutors often need to manually analyze a large amount of clues and make accurate judgments on the clues that need to be filed. With the development of social economy, the number of public interest litigation cases and clues has increased dramatically, and relying solely on manual analysis of case clues is inefficient and may result in missed or incorrect judgments. Therefore, by using related technologies in the fields of natural language processing and knowledge graphs, clues are extracted from case source information on the Internet, a case knowledge graph is constructed, and typical cases, laws and regulations, and power lists related thereto are recommended, which can effectively solve the problems of difficulty in discovering cases and long time in handling cases.

[0003] The existing Chinese invention patent with the publication number CN113220888B discloses a "case clue element extraction method based on Ernie model", which uses natural language processing technology to extract case clues from text data. However, this method only extracts case elements from semantics and does not construct a related knowledge base or make further recommendations to prosecutors. SUMMARY

[0004] In order to solve the defects in the prior art, the present application provides a case clue extraction and recommendation method based on a knowledge graph, which is used to extract case clues from a large amount of data, construct a public interest litigation case knowledge graph, complete the recommendation of related typical cases, laws and regulations, and power lists, and solve the problems of difficulty in discovering cases and long time in handling cases.

[0005] To achieve the above purpose, the present application provides the following technical solution: a public interest litigation case clue extraction and recommendation method based on a knowledge graph, comprising the following steps:

[0006] Step one, case clue extraction; python crawler technology is used to crawl case data. In case clue extraction, python crawler technology is first used to collect clue source data from typical websites on the Internet, and user uploaded data is also supported;

[0007] Step two, a model based on pre-trained Bert combined with conditional random field CRF is used to extract case clues from source data, completing clue extraction;

[0008] Step three, case source knowledge graph construction; adopt Neo4j graph database to construct case source knowledge graph containing three kinds of entities and two kinds of relations. In the construction of case source knowledge graph, three kinds of entities are set, i.e. typical cases, laws and regulations, and power list. Two kinds of relations are set, i.e. the correlation between typical cases and laws and regulations, and the correlation between laws and regulations and power list. The case source knowledge graph is constructed.

[0009] Step four, excellent case recommendation; adopt pre-trained SBERT combined with graph embedding model to recommend excellent cases. In the excellent case recommendation, pre-trained SBERT combined with graph embedding model is adopted to recommend typical cases related to case-making clues. Based on the relations in the knowledge graph, relevant laws and regulations and power list are recommended. In addition, typical cases, laws and regulations, and power list can also be recommended based on user input.

[0010] Preferably, in step one, python crawler technology is used to crawl case-making data. Specifically, request-html library in python is used to realize the crawler program. The target websites for crawling include microblog, BBS, comprehensive news platform, national / local media official website, and typical websites of key administrative websites.

[0011] Preferably, in step two, pre-trained Bert combined with conditional random field (CRF) model is used to extract case-making clues from source data. For the data obtained by crawling in step one, the data is first preprocessed, such as word segmentation and stop word removal. Then, the preprocessed data is trained by the model, and the trained data is used for extraction and prediction. In the model, the data first passes through a pre-trained Bert layer, then passes through a transfomer layer, and then passes through a bidirectional rnn and crf layer. Finally, the output is obtained through a fully connected layer. The extracted results include "illegal theme", "illegal behavior", "time", "place" and other information. After extracting the case-making clues, the case-making clues are scored in seven aspects, i.e. case source feature, time feature, place feature, name feature, contact way feature, clue aggregation degree, and regulation correlation degree. The case source feature score weight accounts for 40%, and the score weights of the other six features account for 10%. The final case-making index score is obtained. The clue with a case-making index score greater than 60 is determined as a case-making clue.

[0012] Preferably, in step three, Neo4j graph database is used to construct case source knowledge graph containing three kinds of entities and two kinds of relations. The three kinds of entities are typical cases, laws and regulations, and power list. The two kinds of relations are the correlation between typical cases and laws and regulations, and the correlation between laws and regulations and power list. Neo4j graph database technology can realize the storage, maintenance and visualization of knowledge graph data.

[0013] Preferably, the step four: adopting the pre-training BERT combined with the graph embedding model to recommend excellent cases; the case source data respectively pass through the pre-training BERT layer and the node2vec graph neural network layer, respectively obtain semantic embedding vectors and graph embedding vectors, and then combine the semantic embedding vectors and the graph embedding vectors of each node to obtain the final embedding vectors of the nodes; the cosine similarity of the extracted clue embedding vectors and the embedding vectors of each node in the knowledge graph is calculated, and the most relevant typical cases can be recommended from large to small according to the cosine similarity, and the relevant relationship in the knowledge graph is used to recommend the laws and regulations and the power list.

[0014] Compared with the prior art, the public interest litigation case clue extraction and recommendation method based on the knowledge graph has the following beneficial effects: the natural language processing technology is used to extract the case clues in the massive data, the case elements are extracted from the deep semantic, and the accuracy and efficiency of the case clue extraction are improved. The knowledge graph in the public interest litigation field is constructed, the entity mapping relationship in the public interest litigation field is constructed, the data is unified, and the semantic information is integrated. Meanwhile, the recommendation system based on the constructed knowledge graph is provided, after the case clues are extracted, the related typical cases, laws and regulations and power list can be quickly recommended, so that the prosecutor can quickly judge whether the case conditions are met, and the public interest litigation case efficiency is improved.

[0015] The natural language processing technology is used to extract the case clues from the network information, the public interest litigation case efficiency and accuracy are improved, the case source knowledge graph is constructed, the case source data is uniformly managed, the typical cases, laws and regulations, power list and the mapping relationship between entities are constructed, the pre-training BERT combined with the graph embedding model is adopted to realize the recommendation, the recommendation is not only based on the semantic information, but also integrates the graph structure information, so that more accurate and diversified case / law / regulation / power list recommendation can be realized, and through the above steps, the automatic case clue extraction and recommendation are realized, and the public interest litigation case judgment efficiency and accuracy are greatly improved. BRIEF DESCRIPTION OF DRAWINGS

[0016] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, together with the embodiments of the application, to explain the application, and do not constitute a limitation on the application, in the drawings:

[0017] Figure 1 The public interest litigation case clue extraction and recommendation process proposed in the application Figure 1 ;

[0018] Figure 2 The step flow chart proposed in the application;

[0019] Figure 3 The case clue extraction flow chart proposed in the application;

[0020] Figure 4 The public interest litigation case clue extraction and recommendation process proposed by the present application Figure 2

[0021] Figure 5 The public interest litigation case clue extraction and recommendation process proposed by the present application Figure 3 DETAILED DESCRIPTION

[0022] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0023] Please refer to Figures 1-5 The present application provides a technical solution: a public interest litigation case clue extraction and recommendation method based on a knowledge graph, comprising the following steps:

[0024] Step one, case clue extraction; adopt python crawler technology to crawl case data, in case clue extraction, first adopt python crawler technology to collect clue source data from typical websites on the Internet, and also support user uploaded data;

[0025] Step two, adopt a model based on pre-training Bert combined with conditional random field CRF to extract case clues from source data, complete clue extraction;

[0026] Step three, case source knowledge graph construction; adopt Neo4j graph database to construct a case source knowledge graph containing three kinds of entities and two kinds of relationships, in case source knowledge graph construction, set three kinds of entities of typical cases, laws and regulations, and power lists, set the correlation between typical cases and laws and regulations, set the correlation between laws and regulations and power lists two kinds of relationships, and construct a case source knowledge graph;

[0027] Step four, excellent case recommendation; adopt a combination of pre-training SBERT and graph embedding model to recommend excellent cases, in excellent case recommendation, adopt a combination of pre-training SBERT and graph embedding model to recommend typical cases related to case clues, and then recommend related laws and regulations and power lists based on the relationship in the knowledge graph. In addition, it also supports recommending typical cases, laws and regulations and power lists based on user input queries.

[0028] ​​In the application, preferably, the step one: adopting python crawler technology to crawl case data, wherein specifically, adopting request-html library in python to realize the crawler program, and the target website for crawling includes microblog, Baidu post bar, comprehensive news platform, national / local media official website, key field administrative website and other typical websites.

[0029] In the application, preferably, the step two: adopting pre-training Bert combined with conditional random field CRF model to extract case clues from source data; for the data obtained by crawling in the step one, first, the data is preprocessed, such as word segmentation and stop word removal, then the preprocessed data is trained by the model, and the trained data is used for extraction prediction; in the model, the data first passes through a pre-trained Bert layer, then passes through a transfomer layer, then passes through a bidirectional rnn and crf layer, and finally passes through a fully connected layer for output; finally, the extracted results include "illegal theme", "illegal behavior", "time", "place" and other information; after extracting the case clues, the case clues are scored from seven aspects of case source feature, time feature, place feature, name feature, contact way feature, clue aggregation degree and regulation correlation degree, wherein the case source feature score weight accounts for 40%, and the score weight of the other six features accounts for 10%, to obtain the final case index score; the clue with a case index greater than 60 points is determined as a case clue.

[0030] In the application, preferably, the step three: adopting Neo4j graph database to construct a case source knowledge graph containing three kinds of entities and two kinds of relationships; wherein the three kinds of entities are typical cases, laws and regulations, and power list; the two kinds of relationships are the correlation between the typical cases and the laws and regulations, and the correlation between the laws and regulations and the power list; adopting Neo4j graph database technology can realize the storage, maintenance and visualization of the knowledge graph data.

[0031] In the application, preferably, the step four: adopting pre-training BERT combined with graph embedding model to recommend excellent cases; the case source data passes through the pre-training BERT layer and the node2vec graph neural network layer respectively to obtain semantic embedding vectors and graph embedding vectors respectively, and then the semantic embedding vectors and the graph embedding vectors of each node are combined to obtain the final embedding vectors of the nodes; the cosine similarity of the embedding vectors of the extracted clues and the embedding vectors of each node in the knowledge graph is calculated, and the most relevant typical cases to the clues are recommended by sorting the cosine similarity from large to small, and then the laws and regulations and the power list are recommended through the correlation in the knowledge graph.

[0032] Embodiment one

[0033] A public interest litigation case clue extraction and recommendation method based on a knowledge graph, comprising the following steps:

[0034] Step one, case clue extraction; adopt python crawler technology to crawl case data, in case clue extraction, first adopt python crawler technology, collect clue source data from typical website on the Internet, also support user upload data;

[0035] Step two, adopt the model based on pre-training Bert combined with conditional random field CRF to extract case clues from source data, complete clue extraction;

[0036] Step three, case source knowledge graph construction; adopt Neo4j graph database to construct case source knowledge graph containing three kinds of entities and two kinds of relations, in case source knowledge graph construction, set three kinds of entities of typical cases, laws and regulations, power list, set the correlation between typical cases and laws and regulations, set the correlation between laws and regulations and power list two kinds of relations, construct case source knowledge graph;

[0037] Step four, excellent case recommendation; adopt the combination of pre-training SBERT and graph embedding model to recommend excellent cases, in excellent case recommendation, adopt the combination of pre-training SBERT and graph embedding model, recommend typical cases related to case clues, and recommend related laws and regulations and power list based on the relationship in the knowledge graph, in addition, also support to recommend typical cases, laws and regulations and power list based on user input query.

[0038] Embodiment two

[0039] A public interest litigation case clue extraction and recommendation method based on a knowledge graph, comprising the following steps:

[0040] Step one: adopt python crawler technology to crawl case data, wherein specifically, adopt request-html library in python to realize crawler program, the target website of crawling includes microblog, BBS, comprehensive news platform, national / local media official website, key field administrative website and other typical websites.

[0041] Step two: extract case clues from source data using a pre-trained Bert combined with conditional random field CRF model; for the data obtained by crawling in step one, first, the data is preprocessed, such as word segmentation and stop word removal, then the preprocessed data is trained through the model, and the trained data is used for extraction prediction; in the model, the data first passes through a pre-trained Bert layer, then passes through a transfomer layer, then passes through a bidirectional rnn and crf layer, and finally passes through a fully connected layer for output; the extracted results include "illegal theme", "illegal behavior", "time", "place" and other information; after extracting the case clues, the case clues are scored in seven aspects of case source characteristics, time characteristics, place characteristics, name characteristics, contact way characteristics, clue aggregation degree and regulation correlation degree, among which the case source characteristic score weight accounts for forty percent, and the score weight of the other six characteristics accounts for ten percent, to obtain the final case index score; the clue with a case index greater than 60 points is determined as a case clue.

[0042] Step three: build a case source knowledge graph containing three entities and two relationships using Neo4j graph database; the three entities are typical cases, laws and regulations, and power lists; the two relationships are the correlation between typical cases and laws and regulations, and the correlation between laws and regulations and power lists; using Neo4j graph database technology can realize the storage, maintenance and visualization of knowledge graph data.

[0043] Step four: recommend excellent cases by combining pre-trained BERT and graph embedding model; the case source data passes through the pre-trained BERT layer and the node2vec graph neural network layer respectively to obtain the semantic embedding vector and the graph embedding vector, and then combines the semantic embedding vector and the graph embedding vector of each node to obtain the final embedding vector of the node; calculate the cosine similarity between the embedding vector of the extracted clue and the embedding vector of each node in the knowledge graph, and recommend the most relevant typical case through the cosine similarity from large to small, and then recommend the laws and regulations and power lists through the correlation in the knowledge graph.

[0044] Embodiment three

[0045] A case clue extraction and recommendation method based on a knowledge graph, comprising the following steps:

[0046] Step one: use python crawler technology to crawl case data;

[0047] Use the request-html library in python to realize the crawler program, and the target websites for crawling include microblog, BBS, comprehensive news platform, national / local media official website, key field administrative website and other typical websites, which can grab the clue data that may become a case;

[0048] Step two: extract the case clues from the source data by using the pre-trained Bert combined with the conditional random field CRF model;

[0049] As shown in Figure 3 , the crawled data is predicted by the pre-trained Bert combined with the conditional random field CRF model, which is used to extract case clue entities including "illegal theme", "illegal behavior", "time", "place" and the like;

[0050] Step three: build a case source knowledge graph containing three entities and two relationships based on the Neo4j graph database;

[0051] The knowledge graph structure is as shown in Figure 4 , the function of building the case source knowledge graph is to uniformly manage data, build the correlation between typical case entities, legal regulation entities and power list entities, and form a graph structure, so as to facilitate the recommendation of the graph structure correlation as a supplement of semantic similarity;

[0052] Step four: recommend excellent cases by combining pre-trained BERT and graph embedding model;

[0053] The recommendation model is as shown in Figure 5 , which realizes recommendation by combining pre-trained BERT and graph embedding model, not only based on semantic information, but also integrated with graph structure information, which can realize more accurate and diversified case / legal regulation / power list recommendation.

[0054] The present application is a case clue extraction and recommendation method based on knowledge graph, including three modules of case clue extraction, case source knowledge graph construction and excellent case recommendation;

[0055] The process of the present application will be further described below in combination with specific examples and drawings:

[0056] First step: use python crawler technology to crawl case data; specifically, use the request-html library in python to realize the crawler program;

[0057] Second step: extract the case clues from the source data by using the pre-trained Bert combined with the conditional random field CRF model; specifically, the following steps are included:

[0058] 1) Preprocess the data by word segmentation, stop word removal, data annotation and the like;

[0059] 2) Build a training model; specifically, the model includes a pre-trained Bert layer, a transfomer layer, a bidirectional rnn and crf layer, and a full connection layer;

[0060] 3) The preprocessed data is trained through the model;

[0061] 4) The trained data is used to extract predictions; the extracted results include "illegal theme", "illegal behavior", "time", "location" and other information;

[0062] 5) After extracting the case clues, scoring is performed on the case clues from seven aspects of case source characteristics, time characteristics, location characteristics, name characteristics, contact method characteristics, clue aggregation degree, and regulation correlation degree, wherein the case source characteristic score weight accounts for 40%, and the score weights of the other six characteristics account for 10%, to obtain the final case index score; the clue with a case index greater than 60 points is determined as a case clue;

[0063] Step 3: A Neo4j graph database is used to build a case source knowledge graph containing three entities and two relationships; specifically, the following steps are included:

[0064] 1) Three entities of typical cases, laws and regulations, and power lists are constructed;

[0065] 2) Two relationships between typical cases and laws and regulations, and between laws and regulations and power lists are constructed;

[0066] 3) Neo4j is used to realize the storage, updating and visualization of the knowledge graph;

[0067] Step 4: Excellent cases are recommended by combining pre-trained BERT and graph embedding model; specifically, the following steps are included:

[0068] 1) The case source data is subjected to a pre-trained BERT layer to obtain semantic embedding vectors of nodes;

[0069] 2) The case source data is subjected to a graph neural network node2vec layer to obtain graph structure embedding vectors of graph nodes;

[0070] 3) The semantic embedding vectors and the graph structure embedding vectors of the graph nodes are spliced to obtain the final embedding vectors of the graph nodes;

[0071] 4) The final embedding vectors of the extracted case clues in step 1 and the final embedding vectors of all other nodes in the knowledge graph are subjected to cosine similarity calculation, and the calculation results are sorted from high to low to obtain the final similar typical case recommendation results;

[0072] Based on the typical case recommendation results obtained in the above steps, relevant laws and regulations and power lists in the knowledge graph are found to obtain the recommendation results of the laws and regulations and the power lists.

[0073] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the exact construction and arrangement described. It is intended, therefore, to cover all modifications and adaptions that fall within the scope of the claims and their equivalents.

Claims

1. A public interest litigation case clue extraction and recommendation method based on a knowledge graph, characterized by, Comprise the following steps; Step one, case clues extraction; adopt python crawler technology to crawl the case data, in the case clue extraction, first adopt python crawler technology, collect clue source data from the typical website on the Internet, also support user upload data; Step two, adopt the model based on pre-training Bert combined with conditional random field CRF to extract case clues from source data, complete the clue extraction; the step two: adopt the model based on pre-training Bert combined with conditional random field CRF to extract case clues from source data; for the data obtained by crawling in step one, first, the data is segmented, and the stop words are removed, then, the pretreated data is trained through the model, and the trained data is used for extraction prediction; in the model, the data first passes through a pre-trained Bert layer, then passes through a transfomer layer, then passes through a bidirectional rnn and crf layer, and finally passes through a full connection layer for output; finally, the extracted results include "illegal theme", "illegal behavior", "time", "place" information; after extracting the case clues, the case clues are scored through seven aspects of case source characteristics, time characteristics, place characteristics, name characteristics, contact way characteristics, clue aggregation degree and regulation correlation degree, wherein the case source characteristic score weight accounts for 40%, and the other six characteristic score weights account for 10%, to obtain the final case index score; the clue whose case index is greater than 60 points is determined as a case clue; Step three, case source knowledge graph construction; adopt Neo4j graph database to construct case source knowledge graph containing three kinds of entities and two kinds of relations, in the construction of case source knowledge graph, set three kinds of entities of typical cases, laws and regulations, power list, set the correlation between typical cases and laws and regulations, set the correlation between laws and regulations and power list two kinds of relations, construct case source knowledge graph; the step three: adopt Neo4j graph database to construct case source knowledge graph containing three kinds of entities and two kinds of relations; wherein the three kinds of entities are typical cases, laws and regulations, and power list; two kinds of relations are the correlation between typical cases and laws and regulations, and the correlation between laws and regulations and power list; adopt Neo4j graph database technology can realize the storage, maintenance and visualization of knowledge graph data; Step four, excellent case recommendation; adopt the combination of pre-training SBERT and graph embedding model to recommend excellent cases. In the excellent case recommendation, the combination of pre-training SBERT and graph embedding model is adopted to recommend typical cases related to the case clues, and then based on the relationship in the knowledge graph, relevant laws and regulations and power lists are recommended. In addition, it also supports recommending typical cases, laws and regulations and power lists based on user input queries; step four: adopt the combination of pre-training BERT and graph embedding model to recommend excellent cases; the case source data is respectively subjected to the pre-training BERT layer and the node2vec graph neural network layer to obtain the semantic embedding vector and the graph embedding vector respectively, and then the semantic embedding vector and the graph embedding vector of each node are combined to obtain the final embedding vector of the node; the cosine similarity of the embedding vector of the extracted clue and the embedding vector of each node in the knowledge graph is calculated, and the most relevant typical cases can be recommended by sorting the cosine similarity from large to small, and then the relevant laws and regulations and power lists are recommended through the related relationship in the knowledge graph.

2. The public interest litigation case clue extraction and recommendation method based on a knowledge graph according to claim 1, characterized in that: The step one: adopt the python crawler technology to crawl the case data, wherein specifically, the request-html library in python is used to realize the crawler program, and the target website to be crawled includes microblog, BBS, comprehensive news platform, national / local media official website, and typical website of key field administration website.

Images