Enterprise smart legal affair platform system based on generative language large model

Abstract

The invention discloses a hybrid enhanced enterprise smart law platform system based on a generative language large model. Four modules including a hybrid enhanced legal knowledge engine, a multi-modal legal document analysis module, a risk quantitative evaluation module and a compliance verification workflow work cooperatively. The hybrid enhanced legal knowledge engine integrates multi-source data, realizes real-time updating and semantic reasoning, and comprises map construction, a rule base and an incremental learning mechanism; the multi-modal legal document analysis module performs structured analysis on the heterogeneous document to generate a feature vector; the risk quantitative evaluation module is combined with Monte Carlo simulation and an analytic hierarchy process, quantifies the risk according to a compliance reference and analysis characteristics, and outputs a thermodynamic diagram and a report; a compliance verification workflow is driven by a finite-state machine, a verification module and a conflict detection module are integrated, a generative language large model is called to generate an improved scheme, and audit records are solidified and fed back for optimization. And the system runs according to the processes of analysis, supply rule, bias calculation and verification correction, so that the intelligence and accuracy of legal affair processing are improved.

Description

Technical Field

[0001] The present invention relates to the field of enterprise legal affairs system automation, and in particular to an enterprise intelligent legal affairs platform system based on a generative language large model. Background Art

[0002] With the development of the economy, the legal affairs faced by enterprises in the operation process are becoming more and more complex. To cope with numerous legal affairs, enterprises often build an internal legal affairs system for process-based response and handling. In the face of the continuous improvement and update of laws and regulations, the legal affairs system of enterprises also needs to be updated and adjusted in real time.

[0003] When the current enterprise legal affairs system faces legal affairs, its legal retrieval, case analysis, and risk avoidance plan deployment mainly rely on manual work, which not only has low efficiency, but also the subjectivity of human judgment in risk assessment will lead to a large difference between the evaluated plan and the actual situation.

[0004] The generative large language model can make up for the shortcomings of manual work in enterprise legal affairs due to its powerful ability to train through a large amount of data in a super-large-scale deep learning model and automatically generate coherent text, and can greatly improve the efficiency of the enterprise legal affairs system. Summary of the Invention

[0005] The present invention provides an enterprise intelligent legal affairs platform system based on a generative language large model. Based on the generative large language model, after the legal elements are extracted by the document parsing system, the knowledge engine provides benchmark rules, the risk module calculates the deviation risk, and finally the compliance workflow completes dynamic verification and correction. This system has efficient and accurate legal analysis capabilities, can improve the automation degree of legal affairs in the enterprise legal affairs department, reduce the enterprise compliance cost and enhance the risk prevention and control level.

[0006] To achieve the above technical objectives, the object of the present invention is achieved as follows: A hybrid enhanced enterprise intelligent legal affairs platform system based on a generative language large model, including four modules that cooperate with each other: a hybrid enhanced legal knowledge engine, a multi-modal legal document parsing module, a risk quantification and assessment module, and a compliance verification workflow.

[0007] The hybrid enhanced legal knowledge engine integrates legal knowledge graphs, dynamic rule bases and case reasoning algorithms to achieve multi-source data fusion and real-time updating of laws, regulations, precedents and contract templates, and supports semantic matching and logical reasoning; the multimodal legal document parsing module performs structured parsing of heterogeneous documents such as PDF / Word / scans based on deep learning, extracts clause semantics, rights and obligations, and signature information, and generates standardized feature vectors; the risk quantification assessment module combines Monte Carlo simulation and hierarchical analysis method to quantify risk levels such as contract breach and compliance deviation, and outputs visual heat maps and explainable reports; the compliance verification workflow drives the multi-node verification process through a finite state machine, calls smart contracts and cross-jurisdictional conflict detection algorithms, generates compliance improvement suggestions in real time, and records audit trails.

[0008] After the document parsing system extracts legal elements, the knowledge engine provides benchmark rules, the risk module calculates deviation risks, and finally the compliance workflow completes dynamic verification and correction.

[0009] The hybrid enhanced legal knowledge engine includes: a legal knowledge graph construction module and a dynamic legal rules library.

[0010] The legal knowledge graph construction module is based on ontology mapping technology to associate legal provisions, judicial interpretation texts and historical case databases, store entity relationships through RDF triples, and use graph neural networks to detect legal conflicts across jurisdictions;

[0011] The dynamic legal rule base is built by tracking real-time revised laws and industry compliance standards, using a bidirectional attention mechanism to match contract terms with the semantics of legal provisions. It then uses web crawlers to obtain announcements from the Ministry of Justice and the State Administration for Market Regulation, extracts document text through OCR recognition, and finally uses NLP-driven methods to extract key information and synchronize data in the legal rule base.

[0012] The legal knowledge graph construction module dynamically updates knowledge nodes through an incremental learning mechanism. Its data source includes legal judgment documents of people's courts at all levels and the company's own contract library. The update frequency is synchronized with the triggering of legal revision announcement events.

[0013] The multimodal legal document parsing module (200) integrates OCR text recognition, a semantic segmenter based on the Transformer architecture, and a feature pyramid network to extract and fuse multimodal features of text, tables, and signature information in PDF, Word, and scanned documents, and generates a structured feature vector containing the semantics of clauses, rights and obligations, and legal effect.

[0014] The risk quantification and assessment module dynamically generates a three-dimensional risk heat map and an interpretable risk assessment report through the collaborative operation of the Monte Carlo risk simulation and the analytic hierarchy process weight configuration system, based on the deviation degree between the compliance benchmark value provided by the legal knowledge engine and the document parsing feature vector.

[0015] The collaborative operation process of the Monte Carlo risk simulation and the analytic hierarchy process weight configuration system is as follows:

[0016] S1. Input the standardized compliance threshold data such as legal provisions and case rules obtained from the legal knowledge engine, and the semantic vectors of contract clauses, the matrix of rights and obligations relationships, and the signature effectiveness identifier extracted by the multimodal parsing module.

[0017] S2. Measure the deviation between the document feature vector and the compliance benchmark value: where, b i is the i-th compliance benchmark value, ω i is the weight coefficient generated by the analytic hierarchy process, satisfying ∑ω i = 1;

[0018] S3. Construct a pairwise comparison matrix A of legal risk criteria, and calculate the weight vector W through eigenvalue decomposition:

[0019] A·W = λ max ·W, where, λ max is the largest eigenvalue of matrix A;

[0020] S4. Randomly sample the uncertain variables in the deviation degree D to generate N simulation results; statistically calculate the probability of occurrence P of risk events and the expected loss value E(L):

[0021] The compliance verification workflow realizes dynamic compliance verification through a multi-node verification architecture driven by a finite state machine, where:

[0022] The finite state machine defines a state set including an initial state, a clause verification state, a conflict detection state, and a correction suggestion state, and triggers state jumps according to the compliance benchmark value output by the legal knowledge engine;

[0023] The multi-node verification architecture integrates an intelligent contract validator and a cross-jurisdiction conflict detection module. The intelligent contract validator encodes the preset compliance rules into executable logic units to automatically verify the logical consistency of contract clauses. The cross-jurisdiction conflict detection module, based on the multi-region legal database in the legal knowledge graph, identifies legal jurisdiction conflicts in clauses through dependency syntactic analysis;

[0024] The compliance verification workflow calls upon a generative language model to generate, in real time, a compliance improvement plan containing a clause revision template, legal basis references, and implementation path instructions based on the risk heat map and conflict detection results of the risk quantification assessment module.

[0025] The system further generates audit trail records that meet standards, solidifies the timestamp of the verification process, operator identity and modification records through blockchain evidence storage technology, and supports digital signature verification;

[0026] The compliance verification workflow also deploys a dynamic feedback optimization mechanism to feed back the user's adoption results of the improvement plan to the legal knowledge engine, and update the state transition rules and conflict detection weights through the reinforcement learning algorithm.

[0027] Beneficial effects of the present invention: Based on a generative large language model, the present invention significantly improves the automation rate of the enterprise legal system and reduces the manpower investment of the legal department through the technical integration of legal knowledge graph + dynamic rule library + multimodal analysis + state machine decision chain. DETAILED DESCRIPTION

[0028] Embodiment 1:

[0029] The enterprise intelligent legal platform system based on the generative language big model is characterized by including four collaborative modules 400, including a hybrid enhanced legal knowledge engine (100), a multimodal legal document parsing module 200, a risk quantification assessment module 300, and a compliance verification workflow.

[0030] The hybrid-enhanced legal knowledge engine 100 integrates a legal knowledge graph, a dynamic rule base, and case-based reasoning algorithms to achieve multi-source data fusion and real-time updates of laws, regulations, precedents, and contract templates, supporting semantic matching and logical reasoning. The multimodal legal document parsing module 200 uses deep learning to perform structured parsing of heterogeneous documents such as PDFs, Word documents, and scanned documents, extracting clause semantics, rights and obligations, and signature information, generating standardized feature vectors. The risk quantification and assessment module 300 combines Monte Carlo simulation with the analytic hierarchy process to quantify the risk level of contract breaches and compliance deviations, generating visual heat maps and interpretable reports. The compliance verification workflow 400 uses a finite state machine to drive a multi-node verification process, invoking smart contracts and cross-jurisdictional conflict detection algorithms to generate compliance improvement recommendations in real time and record audit trails. After the document parsing system extracts legal elements, the knowledge engine provides benchmark rules, the risk module calculates deviation risks, and the compliance workflow performs dynamic verification and correction. This system offers efficient and accurate legal analysis capabilities, reducing corporate compliance costs and improving risk management.

[0031] Furthermore, the hybrid enhanced legal knowledge engine 100 includes a legal knowledge graph construction module 110 and a dynamic legal rule base 120. The legal knowledge graph construction module 110 uses ontology mapping technology to link legal provisions, judicial interpretation texts, and historical case databases, stores entity relationships using RDF triples, and utilizes graph neural networks to detect cross-jurisdictional legal conflicts. The dynamic legal rule base 120 tracks real-time revisions to laws and industry compliance standards, employs a bidirectional attention mechanism to match contract terms with the semantics of legal provisions, obtains announcements from the Ministry of Justice and the State Administration for Market Regulation through web crawlers, extracts document text through optical character recognition (OCR), and finally synchronizes data within the legal rule base through NLP-driven extraction of key information. The RDF triples unify the semantic framework to integrate heterogeneous multi-source data such as laws, regulations, judicial interpretations, and case law, breaking the fragmentation of traditional legal databases. Compared to traditional manual analysis, the application of graph neural networks improves the accuracy of cross-jurisdictional legal conflict detection. By replacing the manual process of reading, summarizing, analyzing, and updating rules through a crawler + OCR + NLP technology chain, the efficiency of rule base updates is significantly improved.

[0032] Furthermore, for structured announcement texts, the specific method for obtaining announcement documents from the Ministry of Justice and the State Administration for Market Regulation through web crawlers is to use Python's Scrapy or Requests library to simulate a browser access to the websites of the Ministry of Justice and the State Administration for Market Regulation, and crawl the HTML source code of the announcement page; for unstructured announcements, use PyPDF2PDF and python-docxWord to extract text; if it is a scanned copy, use OCR recognition to extract the file text.

[0033] Furthermore, the legal knowledge graph construction module 110 dynamically updates knowledge nodes through the incremental learning mechanism 113. Its data source includes legal judgment documents of people's courts at all levels and the company's own contract library. The update frequency is synchronized with the triggering of legal revision announcement events.

[0034] Furthermore, the incremental learning mechanism's technical process can be divided into three steps: monitoring data sources such as the Supreme People's Court and the National People's Congress website, identifying legal revision announcements through NLP; extracting entity relationships from newly added judgments, contracts, and other data; and minimizing interference with existing knowledge based on graph neural networks (GNNs) and regularization constraints. Specifically, the monitored data source can use the BERT-BiLSTM model to classify announcement text and determine whether to trigger an update:

[0035] P(update|T)=σ(W c BiLSTM(T)+b cAmong them, T is the announcement text vector, σ is the Sigmoid function, and an update is triggered when P ≥ 0.9; for entity relation triple extraction, sequence annotation of the pre-trained legal language model Legal-BERT can be used to extract triples (h, r, t) from the new data:

[0036] Score(h, r, t) = f θ (h) · W r · g Φ (t), where f θ and g Φ are entity encoders, W r is the relation matrix, and triples with scores higher than the threshold are retained; when updating the graph neural network, elastic weight consolidation constraints are used to protect existing knowledge from being overwritten: Among them, ζ total is the cross-entropy loss of the new data, F i is the Fisher information matrix of the parameter θ i in the historical data to measure its importance to old knowledge, λ is the trade-off coefficient to inhibit the forgetting of key old knowledge; the graph structure incremental update rule is for the new triple (h, r, t), update the adjacency matrix A and the node embedding Z: A new = A ∪ {(h, t)}, Z new = GNN(A new , Z, X), where X is the node feature matrix, and GNN uses a temporal graph convolutional network to capture the relevance of legal provisions evolving over time.

[0037] Furthermore, the multimodal legal document parsing module 200 performs multimodal feature extraction and fusion on the text, table, and signature information in PDFs, Word documents, and scanned documents by integrating OCR text recognition, a semantic segmenter based on the Transformer architecture, and a feature pyramid network, and generates a structured feature vector containing clause semantics, rights and obligations relationships, and legal effects.

[0038] Furthermore, the core function of the semantic segmenter based on the Transformer architecture is to perform clause-level semantic segmentation and identify entities such as "obligations of Party A" and "rights of Party B" and their logical relationships for rights and obligations relationship positioning.

[0039] Example 2:

[0040] Furthermore, the collaborative operation process of the Monte Carlo risk simulation and analytic hierarchy process weight configuration system includes the following steps:

[0041] S1. Input the standardized compliance threshold data such as legal provisions and case rules obtained from the legal knowledge engine, and the contract clause semantic vectors, rights and obligations relationship matrices, and signature effectiveness identifications extracted by the multimodal parsing module;

[0042] S2. Measure the deviation between the document feature vector and the compliance benchmark value: where, b i is the i-th compliance benchmark value, and ω i is the weight coefficient generated by the analytic hierarchy process, satisfying ∑ω i = 1;

[0043] S3. Construct the pairwise comparison matrix A of the legal risk criterion, and calculate the weight vector W through eigenvalue decomposition: A·W = λ max ·W, where λ max is the largest eigenvalue of matrix A;

[0044] S4. Randomly sample the uncertain variables in the deviation degree D to generate N simulation results; count the occurrence probability P of the risk event and the expected loss value E(L):

[0045] Example 3:

[0046] Furthermore, the compliance verification workflow realizes dynamic compliance verification through a multi-node verification architecture driven by a finite state machine, where: the finite state machine defines a state set including an initial state, a clause verification state, a conflict detection state, and a correction suggestion state, and triggers state jumps according to the compliance benchmark values output by the legal knowledge engine; the multi-node verification architecture integrates an intelligent contract validator and a cross-jurisdiction conflict detection module, where the intelligent contract validator encodes the preset compliance rules into executable logic units to automatically verify the logical consistency of the contract clauses, and the cross-jurisdiction conflict detection module, based on the multi-region legal database in the legal knowledge graph, identifies legal jurisdiction conflicts in the clauses through dependency syntactic analysis; the compliance verification workflow calls a generative language large model to generate a compliance improvement plan including a clause revision template, a legal basis citation, and an implementation path description in real time according to the risk heat map of the risk quantification assessment module and the conflict detection results; the system further generates a standard-compliant audit trail record, solidifies the time stamp, operator identity, and modification records of the verification process through blockchain deposit technology, and supports digital signature verification; the compliance verification workflow also deploys a dynamic feedback optimization mechanism to feedback the adoption results of the improvement plan by the user to the legal knowledge engine, and updates the state transition rules and conflict detection weights through a reinforcement learning algorithm.

[0047] Furthermore, the formula chain of the compliance verification workflow is:

[0048] Contract terms → c i Calculation → State jump → Risk closed loop → Compliance final state (S4)

[0049] Specifically, the state set:

[0050] S = {S0 (initial state), S1 (clause verification state), S2 (conflict detection state), S3 (correction suggestion state), S4 (final state)}, clause compliance benchmark (Compliance with clause i, c i ∈[0,1]), the state jump rule can be Legal mapping examples are as follows:

[0051] c i <0.7→Terms violate the law,∑ω i c i <0.8→The overall risk of the contract exceeds the limit, and correction suggestions are generated.

Claims

1. A hybrid enhanced enterprise intelligent legal service platform system based on a generative language large model, characterized in that, It includes four collaborative modules (400), including a hybrid enhanced legal knowledge engine (100), a multimodal legal document parsing module (200), a risk quantification assessment module (300), and a compliance verification workflow. The hybrid enhanced legal knowledge engine (100) integrates legal knowledge graphs, dynamic rule bases, and case reasoning algorithms to achieve multi-source data fusion and real-time updating of laws, regulations, precedents, and contract templates, supporting semantic matching and logical reasoning. The multimodal legal document parsing module (200) performs structured parsing of heterogeneous documents such as PDF / Word / scans based on deep learning, extracts clause semantics, rights and obligations, and signature information, and generates standardized feature vectors. The risk quantification assessment module (300) combines Monte Carlo simulation with the analytic hierarchy process to quantify the risk levels of contract breach, compliance deviation, etc., and outputs a visual heat map and an interpretable report; the compliance verification workflow (400) drives the multi-node verification process through a finite state machine, calls smart contracts and cross-jurisdictional conflict detection algorithms, generates compliance improvement suggestions in real time, and records audit trails. After the document parsing system extracts legal elements, the knowledge engine provides benchmark rules, the risk module calculates deviation risks, and finally the compliance workflow completes dynamic verification and correction.

2. The hybrid enhanced enterprise intelligent legal service platform system based on the generative language large model according to claim 1, characterized in that: The hybrid enhanced legal knowledge engine (100) includes: a legal knowledge graph construction module (110) and a dynamic legal rule library (120). The legal knowledge graph construction module (110) is based on ontology mapping technology to associate legal and regulatory provisions, judicial interpretation texts and historical case databases, stores entity relationships through RDF triples, and uses graph neural networks to achieve cross-jurisdictional legal conflict detection; The dynamic legal rule base (120) is achieved by tracking real-time revised laws and industry compliance standards, using a two-way attention mechanism to match contract terms with the semantics of legal provisions, obtaining announcement documents from the Ministry of Justice and the State Administration for Market Regulation through web crawlers, extracting document texts through OCR recognition, and finally synchronizing data in the legal rule base through NLP-driven extraction of key information.

3. The enterprise intelligent legal service platform system based on the generative language large model according to claim 2, wherein: The legal knowledge graph construction module (110) dynamically updates knowledge nodes through an incremental learning mechanism (113). Its data source includes legal judgment documents of people's courts at all levels and the company's own contract library. The update frequency is synchronized with the triggering of legal revision announcement events.

4. The enterprise intelligent legal service platform system based on the generative language large model according to claim 1, characterized in that: The multimodal legal document parsing module (200) integrates OCR text recognition, a semantic segmenter based on the Transformer architecture, and a feature pyramid network to extract and fuse multimodal features of text, tables, and signature information in PDF, Word, and scanned documents, and generates a structured feature vector containing the semantics of clauses, rights and obligations, and legal effect.

5. The enterprise intelligent legal service platform system based on the generative language large model according to claim 1, characterized in that: The risk quantification assessment module (300) dynamically generates a three-dimensional risk heat map and an interpretable risk assessment report based on the compliance benchmark value provided by the legal knowledge engine and the deviation degree of the document parsing feature vector through the coordinated operation of the Monte Carlo risk simulation and the hierarchical analysis method weight configuration system.

6. The enterprise intelligent legal service platform system based on the generative language large model according to claim 5, characterized in that: The collaborative operation process of the Monte Carlo risk simulation and analytic hierarchy process weight configuration system is as follows: S1. Input the standardized compliance threshold data such as legal provisions and case rules obtained from the legal knowledge engine, and the semantic vectors of contract clauses, the matrix of rights and obligations relationships, and the signature effectiveness identifier extracted by the multimodal parsing module; S2. Measure the deviation between the document feature vector and the compliance benchmark value: where b i is the i-th compliance benchmark value, and ω i is the weight coefficient generated by the analytic hierarchy process, satisfying ∑ω i = 1; S3. Construct the pairwise comparison matrix A of the legal risk criteria, and calculate the weight vector W through eigenvalue decomposition: A·W = λ max ·W, where λ max is the largest eigenvalue of matrix A; S4. Randomly sample the uncertainty variables in the deviation degree D to generate N simulation results; count the probability of occurrence P of the risk event and the expected loss value E(L):

7. The enterprise intelligent legal service platform system based on the generative language large model according to claim 1, wherein: The compliance verification workflow (400) realizes dynamic compliance verification through a multi-node verification architecture driven by a finite state machine, where: The finite state machine defines a state set including an initial state, a clause verification state, a conflict detection state, and a correction suggestion state, and triggers state jumps according to the compliance benchmark value output by the legal knowledge engine; The multi-node verification architecture integrates an intelligent contract validator and a cross-jurisdiction conflict detection module. The intelligent contract validator encodes preset compliance rules into executable logic units to automatically verify the logical consistency of contract clauses. The cross-jurisdiction conflict detection module, based on the multi-region legal database in the legal knowledge graph, identifies legal jurisdiction conflicts in clauses through dependency syntactic analysis; The compliance verification workflow calls a generative language large model to generate a compliance improvement plan including clause revision templates, legal basis references, and implementation path descriptions in real time according to the risk heat map of the risk quantification assessment module and the conflict detection results; The system further generates a standard-compliant audit trail record, solidifies the time stamp, operator identity, and modification records of the verification process through blockchain evidence storage technology, and supports digital signature verification; The compliance verification workflow also deploys a dynamic feedback optimization mechanism, feeds back the adoption result of the improvement plan by the user to the legal knowledge engine, and updates the state transition rules and conflict detection weights through a reinforcement learning algorithm.

Images