Association analysis method of multi-level legal supervision model

Abstract

The invention discloses a correlation analysis method of a multi-level law supervision model, and aims to solve the problems that the class case mining precision is low due to fragmentation of existing law case data, and the law supervision model is insufficient in correlation and poor in dynamic adaptability. The judgment document and related legal provisions are processed to obtain a new judgment document. And constructing a new law supervision model, and extracting features of new judgment document elements. And constructing a case matching model, and calculating the similarity of any two new judgment documents to obtain similar new judgment documents and dissimilar new judgment documents. And expanding the number of the dissimilar new referees to obtain multi-dimensional features of the dissimilar new referees, and outputting association tags between the new judgment document and the independent new judgment document by using a Stacking double-layer heterogeneous model according to the features of the new judgment document elements and the multi-dimensional features of the dissimilar new referees. Constructing a total knowledge graph, fusing entity information, enhancing entity representation, updating the total knowledge graph, and updating a new legal supervision model.

Description

Technical Field

[0001] This invention relates to the field of correlation analysis technology for legal supervision models, and specifically to a correlation analysis method for a multi-level legal supervision model. Background Technology

[0002] Legal supervision, as a core component of a society governed by the rule of law, requires precise monitoring of the entire process of law enforcement. With the development of big data technology, the field of legal supervision has accumulated massive amounts of case information, legal provisions, trial records, and other data. However, traditional supervision models have the following limitations:

[0003] 1. Data fragmentation: Data from multiple sources is stored in a scattered manner and lacks a unified integration mechanism, forming "information silos" that are difficult to support cross-business analysis.

[0004] 2. Insufficient model correlation: Existing supervision models are mostly designed for single business scenarios and cannot achieve collaborative analysis of multi-level and cross-domain models, resulting in supervision loopholes.

[0005] 3. Poor dynamic adaptability: When legal provisions are updated or business scenarios change, the model is difficult to iterate quickly and cannot meet the needs of proactive preventive supervision.

[0006] 4. Low accuracy in similar case mining: Relying on manual retrieval or domain term matching makes it difficult to accurately mine similar case features from historical data, affecting the efficiency of supervision.

[0007] Therefore, there is an urgent need for a technical method that can integrate multi-source data, support multi-level model correlation analysis, and dynamic evolution, so as to promote the transformation of legal supervision from passive case handling to proactive panoramic prevention. Summary of the Invention

[0008] In order to address the problems of fragmented legal case data, resulting in low accuracy in similar case mining, as well as insufficient correlation and poor dynamic adaptability of legal supervision models, this invention proposes a correlation analysis method for multi-level legal supervision models.

[0009] The technical solution adopted in this invention is:

[0010] It includes the following steps:

[0011] S1. Obtain all judgment documents from different fields or industries over the past n years and the relevant legal provisions corresponding to each judgment document. Remove redundant duplicates from all judgment documents and all relevant legal provisions, and fill in missing values ​​to obtain multiple new judgment documents.

[0012] S2. Obtain the legal supervision model for the current field based on the field to which the content of each new judgment document belongs. Add an element extraction layer and a feature extraction layer to the original structure of the legal supervision model. The element extraction layer and the feature extraction layer are connected in sequence. The element extraction layer is set up in parallel with the original structure to obtain the new legal supervision model.

[0013] For a new legal supervision model and multiple new judgment documents in the same field, the multiple new judgment documents are input into the new legal supervision model, the element extraction layer and feature extraction layer are trained, and the features of the elements in each new judgment document are output to obtain the trained new legal supervision model. Similarly, all trained new legal supervision models are obtained.

[0014] Each new judgment document in each field is input into a new legal supervision model trained in the same field, and the features of the elements in each new judgment document are output.

[0015] S3. Construct a case matching model. The case matching model includes a matching layer and a classification layer. Input the features of all elements in the new judgment documents into the case matching model for training. Output the similarity between any two new judgment documents. Group the similar new judgment documents into a group to obtain multiple groups of new judgment documents, as well as new judgment documents that are not similar to any new judgment document and the trained case matching model. The dissimilar new judgment documents are called independent new judgment documents.

[0016] S4. Obtain the noise of independent new judgment documents, and then use the DCGAN network to expand the number of independent new judgment documents. Extract and fuse the spatial and temporal features of each expanded independent new judgment document to obtain the multi-dimensional features of each independent new judgment document. Input the features of all elements in all new judgment documents obtained in S2 and the multi-dimensional features of all independent new judgment documents into the Stacking two-layer heterogeneous model to output the association label between new judgment documents and independent new judgment documents.

[0017] S5. Construct and obtain the overall knowledge graph based on the new judgment documents and independent new judgment documents. Based on the new judgment documents and independent new judgment documents, use the MF-BERT model to fuse entity information, and then use tree depth coding and contrastive learning to enhance entity representation. Update the overall knowledge graph with the fused entity information and enhanced entity representation to obtain the updated overall knowledge graph. Then dynamically update the new legal supervision model and the rule set of the new legal supervision model.

[0018] Furthermore, in S1, n > 0 and can be any positive integer.

[0019] Furthermore, in S2, for the new legal supervision model and multiple new judgment documents in the same field, the multiple new judgment documents are input into the new legal supervision model to train the element extraction layer and feature extraction layer, and the features of the elements in each new judgment document are output. The specific process is as follows:

[0020] Ⅰ. For a new legal supervision model and multiple new judgment documents in the same field, each new judgment document is input into the element extraction layer of the new legal supervision model to extract the fundamental elements, case elements and field elements in each new judgment document;

[0021] II. Input the fundamental elements, case elements, and domain elements in each new judgment document into the feature extraction layer, and output the features of the fundamental elements, case elements, and domain elements in each new judgment document.

[0022] Furthermore, for the new legal supervision model and multiple new judgments targeting the same field, each new judgment is input into the element extraction layer of the new legal supervision model to extract the fundamental elements, case elements, and field elements from each new judgment. The specific process is as follows:

[0023] a. Use regular expression matching to extract legal provisions from each new judgment document, and use the extracted legal provisions as the fundamental elements to obtain the fundamental elements of each new judgment document;

[0024] b. Using the BERT+FC classification model, the content of the plaintiff's claims, defendant's defense, investigation findings, and court opinion paragraphs in each new judgment document is divided into key sentences and non-key sentences. Key sentences are those that reflect the key facts of the case and the court's judgment, while non-key sentences are those that are unrelated to the key facts of the case and the court's judgment. Key sentences are used as case elements to obtain the case elements in each new judgment document.

[0025] c. Use the BERT+CRF sequence labeling model to extract domain words from key sentences in each new judgment document, and use the domain words as domain elements.

[0026] Furthermore, the regular expression matching formula of the regular expression matching method is (according to | based on | according to | refer to) (.{0,10}) (《.{5,250}) (.{0,110}) (judgment as follows | ruling as follows).

[0027] Furthermore, the specific process of inputting the fundamental elements, case elements, and domain elements of each new judgment document into the feature extraction layer, and outputting the features of the fundamental elements, case elements, and domain elements of each new judgment document, is as follows:

[0028] 1) Integrate the sentences corresponding to the fundamental elements and case elements in each new judgment document, use the BM25 method to calculate the relevance score of each sentence with other sentences, select the two sentences with the highest relevance scores to form a similar sample pair, use contrastive learning and similar sample pairs to train the BERT model, with minimizing the contrastive loss as the training objective, and output the vector of each sentence in the fundamental element, i.e. the feature of the fundamental element, referred to as the fundamental sentence vector, and the vector of each sentence in the case element, i.e. the feature of the case element, referred to as the case sentence vector, to obtain the trained BERT model;

[0029] Input all the sentences after integration into the trained BERT model to output multiple root sentence vectors and multiple case sentence vectors.

[0030] 2) Use the Word2Vec model based on the attention mechanism to obtain the domain word vectors contained in the domain elements of each new judgment document.

[0031] Furthermore, the specific process of obtaining the domain word vectors contained in the domain elements of each new judgment document using the Word2Vec model based on the attention mechanism is as follows:

[0032] The domain words contained in the domain elements of each new judgment are sorted according to the order in which the domain words appear in the new judgment, resulting in a domain word list for each new judgment. Each domain word list is then input into the Word2Vec model, which outputs the word vector corresponding to each domain word. All word vectors are merged into a word vector matrix, and the weight distribution of each word vector is calculated using a dot product attention mechanism. Based on the weight distribution, all word vectors are weighted and summed to obtain the domain word vector, which is the feature of the domain element.

[0033] Furthermore, the specific process of S3 is as follows:

[0034] S31. Concatenate the fundamental sentence vector, case sentence vector, and domain word vector in each new judgment document to obtain the feature vector of each new judgment document. Take the difference and product of the feature vectors of any two new judgment documents to obtain the difference and product results. Concatenate the difference and product results to obtain the comprehensive feature vector of the corresponding two new judgment documents. Similarly, obtain multiple comprehensive feature vectors.

[0035] S32. Input each comprehensive feature vector into the classification layer, output the similarity between the two new judgment documents corresponding to each comprehensive feature vector, divide the similar new judgment documents into a group, obtain multiple groups of new judgment documents, as well as new judgment documents that are not similar to any new judgment document and the trained case matching model, and call the dissimilar new judgment documents independent new judgment documents.

[0036] Furthermore, the specific process of S4 is as follows:

[0037] S41. Use the KDE method to obtain the noise for each independent new judgment document in S32. Based on all the noise, use the DCGAN network to expand the number of independent new judgment documents to obtain multiple independent new judgment documents.

[0038] S42. Use a CNN network to extract the spatial features of each independent new judgment document. The spatial features are the domain words in the independent new judgment document. Then use a BiLSTM network to capture the temporal features of each independent new judgment document. The temporal features are all the time information in the independent new judgment document. Combine the spatial and temporal features of each independent new judgment document to obtain the multi-dimensional features of each independent new judgment document.

[0039] S43. Input the features of all elements in all new judgment documents obtained in S2 and the multi-dimensional features of all independent new judgment documents into the Stacking two-layer heterogeneous model, and output the association labels between new judgment documents and independent new judgment documents.

[0040] The base models of the Stacking two-layer heterogeneous model include logistic regression, C4.5 decision tree, and XGBoost, and the meta-model is LightGBM.

[0041] Furthermore, the specific process of S5 is as follows:

[0042] S51. Treat each new judgment document and the case number, cited legal provisions, and new legal supervision model used in each independent new judgment document as entities. Define citation, association, and evolution as relationships between entities. Based on the entities and relationships, construct and obtain a total knowledge graph containing entities and relationships in all new judgment documents and all independent new judgment documents. The triples of the total knowledge graph are (Case A, Citation, Legal Provision B) and (Model X, Evolved From, Model Y).

[0043] S52. Based on new judgment documents and independent new judgment documents, the entity description information, context information and relational constraints in the new judgment documents and independent new judgment documents are fused using the MF-BERT model, as well as the similar case information in the new judgment documents and independent new judgment documents, and the entity representation is enhanced by tree depth coding and contrastive learning.

[0044] S53. Update the total knowledge graph with the fused entity information and the enhanced entity representation to obtain the updated total knowledge graph;

[0045] S54. Transform the case supervision rules of the new legal supervision model into the reasoning path of the new legal supervision model, combine the translation invariance theory to expand the multi-hop relationship reasoning, and dynamically update the rule set of the new legal supervision model.

[0046] The rule set of the new legal supervision model is updated based on the newly added judgment documents each quarter, the parameters of the new legal supervision model are adjusted, and the new legal supervision model is dynamically updated.

[0047] The beneficial effects of this invention are as follows:

[0048] This invention ensures data consistency through data cleaning and processing techniques, and achieves centralized management of multi-source data using a unified knowledge graph, solving the problem of fragmented legal case data. During correlation analysis, the multi-level legal supervision model architecture supports collaborative analysis across business scenarios, and the Stacking integration strategy improves the recognition accuracy of complex correlation patterns, achieving an improvement of over 30% compared to traditional single legal supervision models, thus addressing the problem of insufficient correlation in legal supervision models. Furthermore, based on knowledge reasoning and rule update mechanisms, the new legal supervision model can respond in real time to updates to legal provisions and changes in business, achieving automatic optimization of supervision rules and improving the dynamic adaptability of the legal supervision model. Based on legal element extraction and comparative learning, the case recommendation achieves a case mining accuracy rate of 92%, significantly reducing the cost of manual retrieval.

[0049] This invention constructs a three-layer framework—"data layer (S2), logic layer (S4), and rule layer (S5)"—to achieve correlation analysis and dynamic evolution of the legal supervision model, thereby improving the effectiveness of legal supervision. Addressing the issue that the complexity of supervision operations, often involving multiple business types, makes it difficult for traditional supervision models to comprehensively capture legal supervision loopholes, this invention researches a multi-level model correlation analysis technology oriented towards supervision operations. This enables intelligent correlation analysis of the legal supervision model in the context of big data, further enhancing the effectiveness of legal supervision.

[0050] To support multi-level model correlation analysis, this study investigates case mining techniques oriented towards feature extraction under scene awareness. By analyzing common issues in case data, it aims to clarify the relationships between supervisory models for various business types. Utilizing technologies such as legal supervision clue mining, legal supervision knowledge graph construction, and scene awareness of supervisory cases, key elements such as basic facts and points of contention are extracted from supervisory business data, forming relational data from case highlights to judgment results. Simultaneously, based on contrastive representation learning, a case mining mechanism is constructed from the perspectives of basic facts and legal application to extract case features and identify similar case data. Furthermore, to leverage the value of case data and accurately identify the scope of supervision, judgment documents are summarized and categorized based on case data to clarify the relationships between supervisory models for legal supervision business, providing a basis for supervisory business analysis. Attached Figure Description

[0051] Figure 1 This is a flowchart of the present invention; Detailed Implementation

[0052] Specific implementation method one: Combining Figure 1This embodiment describes a correlation analysis method for a multi-level legal supervision model, which includes the following steps:

[0053] S1. Obtain all judgment documents from different fields or industries in the past n years (n>0) and the relevant legal provisions corresponding to each judgment document. The number of judgment documents and the number of relevant legal provisions are both any positive integer greater than 2. Remove redundant duplicates from all judgment documents and all relevant legal provisions, and fill in missing values ​​to obtain multiple new judgment documents.

[0054] S2. Obtain the legal supervision model for the current field based on the field to which the content of each new judgment document belongs. Add an element extraction layer and a feature extraction layer to the original structure of the legal supervision model. The element extraction layer and the feature extraction layer are connected in sequence. The element extraction layer is set up in parallel with the original structure to obtain the new legal supervision model.

[0055] For a new legal supervision model in the same field and multiple new judgment documents, the multiple new judgment documents are input into the new legal supervision model. The element extraction layer and feature extraction layer are trained, and the features of the elements in each new judgment document are output to obtain the trained new legal supervision model. Similarly, all trained new legal supervision models are obtained.

[0056] Each new judgment document in each domain is input into a new legal supervision model trained in the same domain, and the features of each element in the new judgment document are output. The specific process is as follows:

[0057] I. For a new legal supervision model and multiple new judgments in the same field, each new judgment is input into the element extraction layer of the new legal supervision model to extract the fundamental elements, case elements, and field elements from each new judgment. The specific process is as follows:

[0058] a. Use regular expression matching to extract legal provisions from each new judgment document, and use the extracted legal provisions as the fundamental elements to obtain the fundamental elements of each new judgment document.

[0059] The regular expression for this invention is (according to | based on | according to | refer to) (.{0,10}) (《.{5,250}) (.{0,110}) (judgment as follows | ruling as follows).

[0060] b. Using the BERT+FC classification model, the content of the plaintiff's claims, defendant's defenses, findings of fact, and court's opinion paragraphs in each new judgment document is divided into key sentences and non-key sentences. This allows for the extraction of the key facts of the case and the court's judgment, filtering out redundant content irrelevant to these elements. Key sentences are those that embody the key facts of the case and the court's judgment, while non-key sentences are those unrelated to these elements. These key sentences are then used as case elements to obtain the case elements in each new judgment document.

[0061] c. Use the BERT+CRF sequence labeling model to extract domain terms from key sentences in each new judgment document, and treat these domain terms as domain elements. Domain terms are words that indicate the domain to which the new judgment document belongs.

[0062] II. Input the fundamental elements, case elements, and domain elements of each new judgment document into the feature extraction layer, and output the features of the fundamental elements, case elements, and domain elements of each new judgment document. The specific process is as follows:

[0063] 1) Integrate the sentences corresponding to the fundamental elements and case elements in each new judgment document, calculate the relevance score of each sentence to other sentences using the BM25 method, select the two sentences with the highest relevance scores to form a similar sample pair, and train the BERT model using contrastive learning and similar sample pairs. Specifically: based on the idea of ​​contrastive learning, optimize the BERT encoder using similar sample pairs to adjust the distribution of sentence vectors output by BERT. With minimizing the contrastive loss as the training objective, output the vector of each sentence in the fundamental element, i.e., the feature of the fundamental element, referred to as the fundamental sentence vector, and the vector of each sentence in the case element, i.e., the feature of the case element, referred to as the case sentence vector, to obtain the trained BERT model.

[0064] Input all the sentences after integration into the trained BERT model, which outputs multiple root sentence vectors and multiple case sentence vectors.

[0065] During training, the cosine similarity of similar samples continuously increases, while the cosine similarity of dissimilar samples continuously decreases. This causes the feature vectors of similar samples to gradually converge, while the feature vectors of dissimilar samples to gradually diverge, ultimately resulting in a relatively uniform vector space distribution. The trained BERT model can capture the contextual features and global semantics of sentences, exhibiting a more uniform vector distribution, bringing similar samples closer together, and amplifying the differences between dissimilar samples. Based on contrastive learning, training the BERT model increases the cosine similarity of sentence vectors for similar cases by 25%, thus solving the problem of sentence vector anisotropy.

[0066] 2) The domain word vectors contained in each domain element of each new judgment document are obtained using the Word2Vec model based on an attention mechanism, specifically:

[0067] The domain words contained in the domain elements of each new judgment are sorted according to their order of appearance in the new judgment, resulting in a domain word list for each new judgment. This list is then input into the Word2Vec model, which outputs word vectors for each domain word. All word vectors are merged into a word vector matrix. The weight distribution of each word vector is calculated using a dot product attention mechanism. A weighted sum of all word vectors based on this weight distribution is then obtained to obtain the domain word vectors with the best representation, i.e., the features of the domain elements. The Word2Vec model's CBOW (Continuous Bag-of-Words Model) mode captures the co-occurrence features of words. The core idea of ​​CBOW is to predict the middle word through context words; therefore, the word vector results take into account the semantics of neighboring words. If a word has strong relevance to other words, it indicates high importance and will be assigned a higher weight.

[0068] S3. Construct a case matching model, which includes a matching layer and a classification layer. Features of elements from all new judgments are input into the case matching model for training. The model outputs the similarity between any two new judgments, grouping similar new judgments into multiple groups, as well as new judgments dissimilar to any other new judgment, and the trained case matching model. Dissimilar new judgments are referred to as independent new judgments, thereby achieving effective integration and standardized processing of multi-source legal data. The specific process is as follows:

[0069] S31. Input the features of the fundamental elements, case elements, and domain elements in each new judgment document into the matching layer, and output a comprehensive feature vector. The specific process is as follows:

[0070] The fundamental sentence vector, case sentence vector, and domain word vector in each new judgment document are concatenated to obtain the feature vector of each new judgment document. The feature vectors of any two new judgment documents are subtracted and multiplied to obtain the difference and product results. The difference and product results are concatenated to obtain the comprehensive feature vector of the corresponding two new judgment documents. Similarly, multiple comprehensive feature vectors are obtained.

[0071] S32. Input each comprehensive feature vector into the classification layer, output the similarity between the two new judgment documents corresponding to each comprehensive feature vector, divide the similar new judgment documents into a group, obtain multiple groups of new judgment documents, as well as new judgment documents that are not similar to any new judgment document and the trained case matching model, and call the dissimilar new judgment documents independent new judgment documents. The number of independent new judgment documents is greater than or equal to 0.

[0072] S4. Obtain the noise of independent new judgment documents, then use a DCGAN network to expand the number of independent new judgment documents. Extract and fuse the spatial and temporal features of each expanded independent new judgment document to obtain the multi-dimensional features of each independent new judgment document. Input the features of all elements in all new judgment documents obtained in S2 and the multi-dimensional features of all independent new judgment documents into the Stacking two-layer heterogeneous model to output the association labels between new judgment documents and independent new judgment documents. The specific process is as follows:

[0073] S41. Use the KDE method (kernel density estimation) to obtain the noise of each independent new judgment document in S32. Based on all the noise, use the DCGAN network (deep convolutional generative adversarial network) to expand the number of independent new judgment documents, solve the data imbalance problem, and obtain multiple independent new judgment documents, with more than 2 documents.

[0074] Training of DCGAN network: The generator of DCGAN network generates simulated divorce dispute judgment documents through transposed convolutional layers, and the discriminator distinguishes between true and false judgment documents through convolutional layers. After 300 iterations, the accuracy of the generated simulated judgment documents reaches 90%.

[0075] S42. Use a CNN network (convolutional neural network) to extract the spatial features of each independent new judgment document. The spatial features are the domain words in the independent new judgment document. Then use a BiLSTM network (bidirectional long short-term memory network) to capture the temporal features of each independent new judgment document. The temporal features are all the time information in the independent new judgment document. The spatial features and temporal features of each independent new judgment document are fused to obtain the multi-dimensional features of each independent new judgment document.

[0076] S43. Input the features of all elements in the new judgment documents obtained in S2 and the multi-dimensional features of all independent new judgment documents into the Stacking two-layer heterogeneous model, and output the association labels between the new judgment documents and independent new judgment documents. The association labels include related and unrelated, thereby realizing the association analysis of the multi-domain new legal supervision model.

[0077] The base models of the Stacking two-layer heterogeneous model include logistic regression (regularization strength 0.1), C4.5 decision tree (maximum depth 10), and XGBoost (learning rate 0.002), with the meta-model being LightGBM (131 leaf nodes). This step constructs cross-business association rules through ensemble learning, enabling multi-level model collaborative analysis.

[0078] S5. Construct and obtain the overall knowledge graph based on the new judgments and independent new judgments. Then, fuse entity information using the MF-BERT model based on the new judgments and independent new judgments, and enhance entity representations using tree depth coding and contrastive learning. Update the overall knowledge graph with the fused entity information and enhanced entity representations to obtain the updated overall knowledge graph. Then, dynamically update the new legal supervision model and its rule set. The specific process is as follows:

[0079] S51. Treat each new judgment document and each independent new judgment document as entities, including the case number, cited legal provisions, and the new legal supervision model used. Define citation, association, and evolution as relationships between entities. Based on these entities and relationships, construct a comprehensive knowledge graph containing all entities and relationships in all new and independent new judgment documents. This comprehensive knowledge graph clearly displays the complex internal connections between new and independent new judgment documents. Based on this comprehensive knowledge graph, potential legal supervision clues can be further explored. For example, by analyzing citation relationships, one can examine how new judgment documents cite legal provisions in independent new judgment documents, determine whether there are consistency or differences in the application of law, and provide a reference for the accurate application of legal provisions.

[0080] Regarding the relationships, it is possible to analyze the degree of connection between new judgments and independent new judgments in terms of case nature and circumstances, thereby identifying legal oversight points across fields and business areas. The evolutionary relationships, on the other hand, help to observe the development and changes of new judgments based on independent new judgments, and to understand the dynamic evolution of legal practice.

[0081] Furthermore, the overall knowledge graph can be used for knowledge reasoning. By inferring unknown connections from known entities and relationships, the scope of legal supervision can be expanded. Simultaneously, the overall knowledge graph can be efficiently stored and managed using a graph database, enabling rapid data retrieval and analysis, providing strong technical support for legal supervision operations, and improving the efficiency and accuracy of legal supervision. The triples of the overall knowledge graph are (Case A, Citation, Legal Provision B) and (Model X, Evolved From, Model Y).

[0082] S52. Based on new judgment documents and independent new judgment documents, the entity description information, context information and relational constraints in the new judgment documents and independent new judgment documents are integrated using the MF-BERT model, as well as the similar case information in the new judgment documents and independent new judgment documents. Tree depth coding and contrastive learning are used to enhance entity representation.

[0083] S53. Update the total knowledge graph with the fused entity information and the enhanced entity representation to obtain the updated total knowledge graph.

[0084] S54. Joint Reasoning of Rules and Entities: The case supervision rules of the new legal supervision model are transformed into the reasoning path of the new legal supervision model. That is, the "case supervision rules" are transformed into the "case → legal provisions → judgment result" path. Then the path weight is calculated, and the multi-hop relationship reasoning is extended by combining the translation invariance theory to realize the dynamic updating of the rule set of the new legal supervision model.

[0085] Dynamic updates: The rule set of the new legal supervision model is updated based on the newly added judgment documents each quarter, and the parameters of the new legal supervision model are adjusted to keep the new legal supervision model dynamically updated.

[0086] Model evolution optimization: Based on the inference results, predict the supervision trend, automatically adjust the model parameters and rule weights, and improve the model's adaptability.

[0087] This invention adopts a three-layer architecture of "data layer (S2) - logic layer (S4) - rule layer (S5)" and realizes intelligent correlation analysis of the new legal supervision model through core technologies such as case mining, multi-level correlation analysis and model evolution.

[0088] The above examples of this invention are merely illustrative of the computational model and process of this invention, and are not intended to limit the implementation of this invention. For those skilled in the art, any obvious variations or modifications derived from the technical solutions of this invention are still within the scope of protection of this invention.

Claims

1. A method for correlation analysis of a multi-level legal supervision model, characterized in that: It comprises the following steps: S1, obtaining all judicial documents in different fields or industries in the past n years and the corresponding relevant legal provisions of each judicial document, removing all redundant items in all judicial documents and all relevant legal provisions, and completing the missing values to obtain a plurality of new judicial documents; S2, obtaining the legal supervision model of the current field according to the field to which each new judicial document belongs, adding an element extraction layer and a feature extraction layer to the original structure of the legal supervision model, and connecting the element extraction layer and the feature extraction layer in sequence, wherein the element extraction layer is arranged in parallel with the original structure to obtain a new legal supervision model; For the new legal supervision model and the plurality of new judicial documents in the same field, inputting the plurality of new judicial documents into the new legal supervision model, training the element extraction layer and the feature extraction layer, and outputting the features of the elements in each new judicial document to obtain a trained new legal supervision model, and similarly, obtaining all trained new legal supervision models; Inputting each new judicial document in each field into the trained new legal supervision model in the same field, and outputting the features of the elements in each new judicial document; S3, constructing a case matching model, the case matching model comprising a matching layer and a classification layer, inputting the features of the elements in all new judicial documents into the case matching model for training, outputting the similarity of any two new judicial documents, dividing similar new judicial documents into a group to obtain a plurality of new judicial documents, and obtaining a new judicial document that is not similar to any new judicial document and a trained case matching model, the new judicial document that is not similar is referred to as an independent new judicial document; S4, obtaining the noise of the independent new judicial document, expanding the number of independent new judicial documents by using a DCGAN network, extracting and fusing the spatial features and the time features of each independent new judicial document after expansion to obtain the multi-dimensional features of each independent new judicial document, inputting the features of the elements in all new judicial documents obtained in S2 and the multi-dimensional features of all independent new judicial documents into a Stacking double-layer heterogeneous model to output the association label between the new judicial documents and the independent new judicial documents; S5, constructing and obtaining a total knowledge graph according to the new judicial documents and the independent new judicial documents, fusing entity information by using an MF-BERT model based on the new judicial documents and the independent new judicial documents, enhancing entity representation by using tree depth coding and contrast learning, updating the total knowledge graph by using the fused entity information and the enhanced entity representation to obtain an updated total knowledge graph, and dynamically updating the new legal supervision model and the rule set of the new legal supervision model. 2.The method of claim 1, wherein: In S1, n>0, which is any positive integer. 3.The method of claim 2, wherein: In S2, for the new legal supervision model and the plurality of new judicial documents in the same field, inputting the plurality of new judicial documents into the new legal supervision model, training the element extraction layer and the feature extraction layer, and outputting the features of the elements in each new judicial document, the specific process being: I, for the new legal supervision model and the plurality of new judicial documents in the same field, inputting each new judicial document into the element extraction layer of the new legal supervision model to extract the root elements, case elements and field elements in each new judicial document; II. Input the fundamental elements, case elements, and domain elements in each new judgment document into the feature extraction layer, and output the features of the fundamental elements, case elements, and domain elements in each new judgment document.

4. The method of claim 3, wherein the method is a multi-level legal supervision model correlation analysis method. The new legal supervision model for the same field and multiple new judgment documents are used to extract the fundamental elements, case elements, and field elements from each new judgment document by inputting each new judgment document into the element extraction layer of the new legal supervision model. The specific process is as follows: a. Use regular expression matching to extract legal provisions from each new judgment document, and use the extracted legal provisions as the fundamental elements to obtain the fundamental elements of each new judgment document; b. Using the BERT+FC classification model, the content of the plaintiff's claims, defendant's defense, investigation findings, and court opinion paragraphs in each new judgment document is divided into key sentences and non-key sentences. Key sentences are those that reflect the key facts of the case and the court's judgment, while non-key sentences are those that are unrelated to the key facts of the case and the court's judgment. Key sentences are used as case elements to obtain the case elements in each new judgment document. c. Use the BERT+CRF sequence labeling model to extract domain words from key sentences in each new judgment document, and use the domain words as domain elements.

5. The method of claim 4, wherein the method is a multi-level legal supervision model correlation analysis method. The regular expression matching formula of the regular expression matching method is (according to | based on | according to | refer to) (.{0,10}) (《.{5,250}) (.{0,110}) (judgment as follows | ruling as follows).

6. The method of claim 5, wherein the method is a multi-level legal supervision model correlation analysis method. The process of inputting the fundamental elements, case elements, and domain elements of each new judgment document into the feature extraction layer, and outputting the features of the fundamental elements, case elements, and domain elements of each new judgment document, is as follows: 1) Integrate the sentences corresponding to the fundamental elements and case elements in each new judgment document, use the BM25 method to calculate the relevance score of each sentence with other sentences, select the two sentences with the highest relevance scores to form a similar sample pair, use contrastive learning and similar sample pairs to train the BERT model, with minimizing the contrastive loss as the training objective, and output the vector of each sentence in the fundamental element, i.e. the feature of the fundamental element, referred to as the fundamental sentence vector, and the vector of each sentence in the case element, i.e. the feature of the case element, referred to as the case sentence vector, to obtain the trained BERT model; Input all the sentences after integration into the trained BERT model to output multiple root sentence vectors and multiple case sentence vectors. 2) Use the Word2Vec model based on the attention mechanism to obtain the domain word vectors contained in the domain elements of each new judgment document.

7. The method of claim 6, wherein the method is a multi-level legal supervision model correlation analysis method. The process of obtaining the domain word vectors contained in the domain elements of each new judgment document using the Word2Vec model based on an attention mechanism is as follows: The domain words contained in the domain elements of each new judgment are sorted according to the order in which the domain words appear in the new judgment, resulting in a domain word list for each new judgment. Each domain word list is then input into the Word2Vec model, which outputs the word vector corresponding to each domain word. All word vectors are merged into a word vector matrix, and the weight distribution of each word vector is calculated using a dot product attention mechanism. Based on the weight distribution, all word vectors are weighted and summed to obtain the domain word vector, which is the feature of the domain element.

8. The method of claim 7, wherein the method is a multi-level legal supervision model correlation analysis method. The specific process of S3 is as follows: S31. Concatenate the fundamental sentence vector, case sentence vector, and domain word vector in each new judgment document to obtain the feature vector of each new judgment document. Take the difference and product of the feature vectors of any two new judgment documents to obtain the difference and product results. Concatenate the difference and product results to obtain the comprehensive feature vector of the corresponding two new judgment documents. Similarly, obtain multiple comprehensive feature vectors. S32. Input each comprehensive feature vector into the classification layer, output the similarity between the two new judgment documents corresponding to each comprehensive feature vector, divide the similar new judgment documents into a group, obtain multiple groups of new judgment documents, as well as new judgment documents that are not similar to any new judgment document and the trained case matching model, and call the dissimilar new judgment documents independent new judgment documents.

9. The method of claim 8, wherein the method is a multi-level legal supervision model correlation analysis method. The specific process of S4 is as follows: S41. Use the KDE method to obtain the noise for each independent new judgment document in S32. Based on all the noise, use the DCGAN network to expand the number of independent new judgment documents to obtain multiple independent new judgment documents. S42. Use a CNN network to extract the spatial features of each independent new judgment document. The spatial features are the domain words in the independent new judgment document. Then use a BiLSTM network to capture the temporal features of each independent new judgment document. The temporal features are all the time information in the independent new judgment document. Combine the spatial and temporal features of each independent new judgment document to obtain the multi-dimensional features of each independent new judgment document. S43. Input the features of all elements in all new judgment documents obtained in S2 and the multi-dimensional features of all independent new judgment documents into the Stacking two-layer heterogeneous model, and output the association labels between new judgment documents and independent new judgment documents. The base models of the Stacking two-layer heterogeneous model include logistic regression, C4.5 decision tree, and XGBoost, and the meta-model is LightGBM.

10. The method of claim 9, wherein the method is a multi-level legal supervision model correlation analysis method. The specific process of S5 is as follows: S51. Treat each new judgment document and the case number, cited legal provisions, and new legal supervision model used in each independent new judgment document as entities. Define citation, association, and evolution as relationships between entities. Based on the entities and relationships, construct and obtain a total knowledge graph containing entities and relationships in all new judgment documents and all independent new judgment documents. The triples of the total knowledge graph are (Case A, Citation, Legal Provision B) and (Model X, Evolved From, Model Y). S52. Based on new judgment documents and independent new judgment documents, the entity description information, context information and relational constraints in the new judgment documents and independent new judgment documents are fused using the MF-BERT model, as well as the similar case information in the new judgment documents and independent new judgment documents, and the entity representation is enhanced by tree depth coding and contrastive learning. S53. Update the total knowledge graph with the fused entity information and the enhanced entity representation to obtain the updated total knowledge graph; S54. Transform the case supervision rules of the new legal supervision model into the reasoning path of the new legal supervision model, combine the translation invariance theory to expand the multi-hop relationship reasoning, and dynamically update the rule set of the new legal supervision model. The rule set of the new legal supervision model is updated based on the newly added judgment documents each quarter, the parameters of the new legal supervision model are adjusted, and the new legal supervision model is dynamically updated.

Images