Legal knowledge graph construction method and matched device and system

Abstract

The invention discloses a legal knowledge graph construction method and a matched device and system, and belongs to the technical field of legal sciences, the legal knowledge graph construction method comprises the following specific steps: step 1, multi-source data acquisition: acquiring multi-source data through a judicial private network security interface; and synchronously obtaining heterogeneous data of the legal provision database, the judgment document database and the law literature database. According to the method, the legal element recognition accuracy is remarkably improved by adopting the adversarial training triple extractor, the problem of regional judgment expression difference is effectively solved, the knowledge graph updating period is shortened to be within 24 hours based on a conflict resolution mechanism of authority and aging attenuation, the method adapts to legal revision and judicial interpretation change, and the method is high in practicability and easy to popularize. The FPGA hardware acceleration device realizes high concurrent processing, a built-in legal rule base automatically detects logic contradictions such as criminal responsibility age conflicts and the like, the error conduction risk is reduced, and authoritative technical support is provided for judicial decision making.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of legal science, and particularly relates to a legal knowledge graph construction method and a matching device and system. BACKGROUND

[0002] Legal data is scattered in heterogeneous sources such as judicial documents, regulation databases, and academic literature, and there are problems such as non-uniform terminology and large structural differences; traditional keyword retrieval cannot capture the logical relationship between legal concepts (such as the association between "joint liability" and "joint tort"); existing knowledge graph construction methods lack support for the hierarchical structure unique to the legal field (such as "article -> clause -> item"); and there is a lack of dynamic updating mechanism for case judgment highlights, resulting in poor timeliness of the knowledge graph. SUMMARY

[0003] The technical problem to be solved by the application is to overcome the shortcomings of the prior art and provide a legal knowledge graph construction method and a matching device and system.

[0004] The technical solution adopted to solve the above technical problem is as follows: a legal knowledge graph construction method, comprising the following specific steps:

[0005] Step 1: Multi-source data acquisition, through a judicial professional network security interface, synchronously acquiring heterogeneous data of a legal text database, a judicial document database, and a legal literature database;

[0006] Step 2: Hierarchical structure analysis, using a nested entity recognition model based on BERT-BiLSTM-CRF to perform hierarchical labeling on legal texts to identify chapter, section, article, clause, and item entities and their subordinate relationships;

[0007] Step 3: Judgment element extraction, constructing a judgment element triple extractor to extract subject behavior, legal characterization, and judgment result triples from judicial documents;

[0008] Step 4: Introducing an adversarial training mechanism, mixing documents from different regional courts in the training data to improve generalization ability;

[0009] Step 5: Dynamic knowledge fusion, when new data conflicts with the knowledge graph, calculating the authority weight of the conflict cases, and if the total weight of the conflict case set exceeds a threshold, triggering graph reconstruction;

[0010] Step 6: Cross-domain association construction, based on a legal responsibility transmission rule library, establishing association paths between civil law, criminal law, and administrative law entities.

[0011] By the technical solution, the legal element recognition accuracy is significantly improved by adopting the triplet extractor of the adversarial training, the regional judgment expression difference problem is effectively solved, the conflict resolution mechanism based on authority and time decay shortens the knowledge graph update cycle to 24 hours, adapts to legal revision and judicial interpretation change, the FPGA hardware acceleration device realizes high concurrency processing, the built-in legal rule library automatically detects logical contradictions such as criminal responsibility age conflicts, and the error transmission risk is reduced, thereby providing authoritative technical support for judicial decision-making.

[0012] Further, the nested entity recognition model adopts a legal domain pre-trained BERT model, the training data includes a provision interpretation annotation set issued by the Supreme People's Court, and the rule library in step six includes mapping rules of administrative punishment facts, civil tort liability constitutive elements and criminal crime constitutive elements.

[0013] By the technical solution, the accuracy of model training can be greatly improved, thereby realizing the construction of the knowledge graph.

[0014] Further, the triplet extractor in step three includes using a graph neural network to construct a causal relationship graph in the judgment reason, focusing on the legal element keywords in the judicial document through an attention mechanism, and the keyword library is certified by the Supreme People's Court Judicial Case Research Institute.

[0015] By the technical solution, the keywords can be effectively extracted, thereby improving the efficiency of constructing the causal relationship graph.

[0016] Further, in the authority weight calculation in step four, the court level coefficient is assigned as follows: the Supreme People's Court case = 1.0, the high court case = 0.8, the intermediate people's court case = 0.6, and the grassroots court case = 0.4, the time decay factor λ is related to the legal field, the civil law field takes 0.05-0.1, the criminal law field takes 0.15-0.2, and the following formula is used:

[0017] Weight W = alpha x court level coefficient + beta x e λ·Δt

[0018] Wherein, alpha + beta = 1, lambda belongs to [0.05, 0.2], and delta t is the difference between the current time and the case effective time.

[0019] By the technical solution, the case effective time can be avoided to be far away from the current time, and the change of the law can be avoided.

[0020] The application discloses a device matched with a legal knowledge graph construction method, which comprises a safe access module, a hardware acceleration unit and a dynamic storage cluster, the safe access module is integrated with a national secret SM4 encryption chip and supports data encryption transmission with a judgment document network and a legal information platform, the hardware acceleration unit is equipped with an FPGA acceleration card and is configured with a legal entity identification special circuit, and the FPGA acceleration card comprises a hardware calculation core of a BiLSTM-CRF model.

[0021] Through the technical scheme, the information confidentiality can be effectively improved, and the data processing capacity can be improved.

[0022] Further, the dynamic storage cluster stores legal entity relationships by using a Neo4j graph database and caches high-frequency access sub-graphs by using Redis, and a cache strategy is based on an LRU algorithm.

[0023] Through the technical scheme, the storage and reading speed can be greatly improved.

[0024] The application discloses a system of a legal knowledge graph construction method, which comprises an intelligent checking engine and a multi-role access interface, the intelligent checking engine is internally provided with a legal logic constraint rule library, comprises a criminal responsibility age rule, a litigation time limit rule and a legal conflict processing rule, and automatically detects logical contradictions in the knowledge graph, and generates a legal conflict analysis report when a conflict is detected.

[0025] Through the technical scheme, the logical contradictions can be avoided, and the accuracy of the knowledge graph can be greatly improved.

[0026] Further, the multi-role access interface comprises a judge end, a lawyer end and a public end, the judge end is responsible for providing a similar case recommendation function, matches historical cases based on judgment gist similarity, the lawyer end is responsible for opening a legal element analysis API, inputs a case fact and outputs a legal qualitative probability, and the public end is responsible for displaying a historical evolution path of a legal concept.

[0027] Through the technical scheme, the judge, the lawyer and the public can conveniently use the legal knowledge graph, and corresponding recommendations can be made according to different users.

[0028] The application has the following beneficial effects: the legal element recognition accuracy is significantly improved by adopting the triad extractor of the adversarial training, the regional judgment expression difference problem is effectively solved, the conflict resolution mechanism based on authority and time decay shortens the knowledge graph update cycle to 24 hours, adapts to legal revision and judicial interpretation change, the FPGA hardware acceleration device realizes high-concurrency processing, the built-in legal rule library automatically detects criminal responsibility age conflicts and other logical contradictions, the error transmission risk is reduced, and the application provides authoritative technical support for judicial decision-making. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0031] like Figure 1 As shown, this embodiment of a legal knowledge graph construction method and supporting apparatus and system includes the following specific steps:

[0032] Step 1: Multi-source data collection. Through the secure interface of the judicial intranet, heterogeneous data from legal provisions database, judgment document database and legal literature database are acquired simultaneously.

[0033] Step 2: Hierarchical structure parsing. Using a nested entity recognition model based on BERT-BiLSTM-CRF, the legal provisions are hierarchically labeled to identify chapters, sections, articles, clauses, items and their subordinate relationships.

[0034] Step 3: Judgment element extraction. Construct a judgment element triplet extractor to extract the subject's behavior, legal definition, and judgment result triplet from the judgment document.

[0035] Step 4: Introduce an adversarial training mechanism to mix documents from courts in different regions into the training data to improve generalization ability;

[0036] Step 5: Dynamic knowledge fusion. When new data conflicts with the knowledge graph, the authority weight of the conflicting cases is calculated. If the total weight of the conflicting case set exceeds the threshold, the knowledge graph is reconstructed.

[0037] Step Six: Cross-domain Association Construction. Based on the legal responsibility transmission rule base, establish the association path between civil law, criminal law, and administrative law entities. By adopting a triplet extractor with adversarial training, the accuracy of legal element identification is significantly improved, effectively solving the problem of regional differences in judicial expression. Based on the conflict resolution mechanism of authority and time decay, the knowledge graph update cycle is shortened to within 24 hours, adapting to legal revisions and changes in judicial interpretations. FPGA hardware acceleration device enables high-concurrency processing. The built-in legal rule base automatically detects logical contradictions such as conflicts in the age of criminal responsibility, reducing the risk of erroneous transmission and providing authoritative technical support for judicial decision-making.

[0038] The nested entity recognition model adopts a legal field pre-training BERT model, the training data contains a provision interpretation annotation set issued by the Supreme People's Court, and the rule library in step six includes mapping rules of administrative punishment facts, civil tort liability constitutive elements and criminal crime constitutive elements, which can greatly improve the accuracy of model training, thereby realizing the construction of the knowledge graph.

[0039] The triad extractor in step three includes using a graph neural network to construct a causal relationship graph in the judgment reason, focusing on the legal requirement keywords in the judicial document through an attention mechanism, and the keyword library is certified by the Supreme People's Court Judicial Case Research Institute, which can effectively extract keywords and improve the efficiency of constructing the causal relationship graph.

[0040] In the authority weight calculation in step four, the court level coefficient is assigned as follows: Supreme People's Court case = 1.0, high court case = 0.8, intermediate people's court case = 0.6, and primary court case = 0.4, and the value of the time decay factor λ is related to the legal field, taking 0.05-0.1 in the civil law field and 0.15-0.2 in the criminal law field, and the following formula is used:

[0041] Weight W = alpha x court level coefficient + beta x e λ·Δt

[0042] Where alpha + beta = 1, lambda is [0.05, 0.2], and delta t is the difference between the current time and the case effective time, which can avoid the change of the law due to the long time of the case effective time.

[0043] A device for a legal knowledge graph construction method includes a secure access module, a hardware acceleration unit, and a dynamic storage cluster. The secure access module integrates a national SM4 encryption chip and supports data encryption transmission with the judicial document network and the law platform. The hardware acceleration unit is equipped with an FPGA acceleration card and configured with a legal entity recognition special circuit. The FPGA acceleration card contains a hardware computing core of the BiLSTM-CRF model, which can effectively improve the confidentiality of information and data processing capacity.

[0044] The dynamic storage cluster uses a Neo4j graph database to store legal entity relationships and caches high-frequency access sub-graphs through Redis. The cache strategy is based on the LRU algorithm, which can greatly improve the storage and reading speed.

[0045] A kind of legal knowledge graph construction method system, including intelligent checking engine and multi-role access interface, the built-in legal logic constraint rule library of the intelligent checking engine, include criminal responsibility age rule, litigation time limit rule, legal conflict processing rule, automatically detect the logic contradiction in knowledge graph, when detecting conflict, generate legal conflict analysis report, can avoid the appearance of logic contradiction, greatly improve the accuracy of knowledge graph.

[0046] The multi-role access interface includes judge end, lawyer end and public end, the judge end is responsible for providing case recommendation function, based on the similarity of judgment points matching historical cases, the lawyer end is responsible for opening legal element analysis API, input case fact and output legal qualitative probability, the public end is responsible for showing the historical evolution path of legal concept, can facilitate judge, lawyer and public to use legal knowledge graph, and according to the different of user and thus corresponding recommendation.

[0047] The above is only the preferred embodiment of the present application, and is not used to limit the protection scope of the present application.

Claims

1. A method for constructing a legal knowledge graph, characterized in that, The specific steps include the following: Step 1: Multi-source data collection. Through the secure interface of the judicial intranet, heterogeneous data from legal provisions database, judgment document database and legal literature database are acquired simultaneously. Step 2: Hierarchical structure parsing. Using a nested entity recognition model based on BERT-BiLSTM-CRF, the legal provisions are hierarchically labeled to identify chapters, sections, articles, clauses, items and their subordinate relationships. Step 3: Judgment element extraction. Construct a judgment element triplet extractor to extract the subject's behavior, legal definition, and judgment result triplet from the judgment document. Step 4: Introduce an adversarial training mechanism to mix documents from courts in different regions into the training data to improve generalization ability; Step 5: Dynamic knowledge fusion. When new data conflicts with the knowledge graph, the authority weight of the conflicting cases is calculated. If the total weight of the conflicting case set exceeds the threshold, the knowledge graph is reconstructed. Step Six: Cross-domain association construction: Based on the legal responsibility transmission rule base, establish the association path between civil law, criminal law and administrative law entities.

2. The method for constructing a legal knowledge graph according to claim 1, characterized in that, The nested entity recognition model adopts a pre-trained BERT model in the legal domain. The training data includes the annotation set of articles published by the Supreme People's Court. The rule base in step six includes mapping rules for administrative penalty facts, elements of civil tort liability and elements of criminal offenses.

3. The method for constructing a legal knowledge graph according to claim 2, characterized in that, The triplet extractor in step three includes using a graph neural network to construct a causal relationship graph in the reasoning of the judgment, and using an attention mechanism to focus on legal elements keywords in the judgment document. The keyword library has been certified by the Judicial Case Research Institute of the Supreme People's Court.

4. The method for constructing a legal knowledge graph according to claim 3, characterized in that, In step four, the authority weight calculation uses the following court level coefficients: Supreme People's Court cases = 1.0, Higher People's Court cases = 0.8, Intermediate People's Court cases = 0.6, and Basic People's Court cases = 0.

4. The statute of limitations decay factor λ is related to the legal field, ranging from 0.05 to 0.1 in civil law and from 0.15 to 0.2 in criminal law, specifically using the following formula: Weight W = α × Court Level Coefficient + β × e λ·Δt Where α+β=1, λ∈[0.05,0.2], and Δt is the difference between the current time and the effective time of the case.

5. The supporting apparatus for the legal knowledge graph construction method according to claim 4, characterized in that, It includes a secure access module, a hardware acceleration unit, and a dynamic storage cluster. The secure access module integrates a national cryptographic SM4 encryption chip, which supports encrypted data transmission with the China Judgments Online and the Faxin platform. The hardware acceleration unit is equipped with an FPGA acceleration card and is configured with a dedicated circuit for legal entity recognition. The FPGA acceleration card contains a hardware-based computing core of the BiLSTM-CRF model.

6. The supporting apparatus for the legal knowledge graph construction method according to claim 5, characterized in that, The dynamic storage cluster uses the Neo4j graph database to store legal entity relationships and caches frequently accessed subgraphs through Redis, with the caching strategy based on the LRU algorithm.

7. The system for constructing a legal knowledge graph according to claim 6, characterized in that, It includes an intelligent verification engine and a multi-role access interface. The intelligent verification engine has a built-in legal logic constraint rule library, which includes rules on the age of criminal responsibility, statute of limitations, and rules for handling legal conflicts. It automatically detects logical contradictions in the knowledge graph and generates a legal conflict analysis report when a conflict is detected.

8. The system for constructing a legal knowledge graph according to claim 7, characterized in that, The multi-role access interface includes a judge's side, a lawyer's side, and a public side. The judge's side is responsible for providing a similar case recommendation function and matching historical cases based on the similarity of the key points of the judgment. The lawyer's side is responsible for opening up the legal elements analysis API, inputting case facts and outputting the probability of legal characterization. The public side is responsible for displaying the historical evolution path of legal concepts.

Images