A quantitative evaluation system based on a contract file document dynamic risk map model

Abstract

The application discloses a kind of quantification evaluation systems based on contract file document dynamic risk atlas model, it is related to contract file document evaluation field.The present application includes: data acquisition and pretreatment module, dynamic risk atlas construction module and quantification evaluation module;Data acquisition and pretreatment module: by optical character recognition OCR engine, character recognition is carried out to scanned contract PDF.The present application can dynamically evolve by defining real-time updateable risk entity and associated relationship, combined with triple update mechanism of time trigger, event trigger, rule trigger, so that risk atlas can evolve dynamically with internal and external environment changes, compared with traditional static rule engine, the mechanism can respond to legal regulations update, industry policy change and enterprise business adjustment in real time, reduce the problem that risk assessment model lags behind actual risk change.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of contract file document evaluation, in particular to a quantitative evaluation system based on a contract file document dynamic risk graph model. BACKGROUND

[0002] In the field of enterprise contract management, the traditional contract risk evaluation method mainly relies on a static rule engine or manual review, which has significant defects. Firstly, risk evaluation relies on the experience of reviewers, which is highly subjective and inefficient, and it is difficult to meet the rapid review needs of a large number of contracts. Secondly, the risk model is fixed and cannot dynamically respond to updates of laws and regulations, changes in industry policies and adjustments of business operations. Thirdly, only the contract text itself is reviewed in isolation, and there is a lack of correlation analysis of business operation risks and external environment risks in the contract performance process, resulting in a disconnection between the risk evaluation results and the actual situation. Although there are attempts to introduce knowledge graphs for contract risk analysis in the prior art, most of them are still at the stage of static entity relationship modeling, and a risk evaluation system with dynamic evolution capability has not been formed, which cannot meet the needs of enterprises for real-time, accurate and comprehensive risk control.

[0003] Therefore, we improve it and propose a quantitative evaluation system based on a contract file document dynamic risk graph model. SUMMARY

[0004] The application provides the following technical solutions:

[0005] The application is specifically as follows:

[0006] A quantitative evaluation system based on a contract file document dynamic risk graph model, comprising a data acquisition and preprocessing module, a dynamic risk graph construction module and a quantitative evaluation module.

[0007] The data acquisition and preprocessing module: a character recognition engine is used to recognize characters in a scanned contract PDF; a BERT pre-training model is used to perform semantic analysis on unstructured text, the text is divided according to sentences, the length of each sentence is less than or equal to 512 tokens, key entities such as contract subjects, objects, performance periods and breach of contract clauses are extracted, and a LAC model is used to extract logical relationships between entities; an ETL tool is used to structure the entity and relationship data and store them in a PostgreSQL database, and a relational data model comprising four tables of contracts, entities, clauses and relationships is established, wherein:

[0008] The contract subject storage format is: {subject ID, name, unified social credit code, legal representative};

[0009] The target storage format is: {target ID, name, specification, quantity, unit price};

[0010] The performance period storage format is: {clause ID, start date, end date, performance mode};

[0011] The breach of contract liability clause storage format is: {clause ID, breach of contract situation, breach of contract calculation method, dispute resolution method};

[0012] Dynamic risk map construction module: Neo4j graph database is used to store risk entities and associated relationships, including: entity definition unit, associated relationship unit and update unit;

[0013] Entity definition unit: define contract text risk entity, business operation risk entity and external environment risk entity;

[0014] Among them, the storage format of contract text risk entity is: {entity ID, clause ID, compliance score [0-100], balance of rights and obligations score [0-100]};

[0015] The storage format of business operation risk entity is: {entity ID, performance ability score [0-100], credit rating [AAA-E], 12 months' performance rate [0-100%], update timestamp}, wherein the performance ability score, credit rating and 12 months' performance rate support real-time dynamic update;

[0016] The storage format of external environment risk entity is: {entity ID, legal change risk score [0-100], industry supervision risk score [0-100], latest update time};

[0017] The association relationship unit: the logical dependency relationship within the text, the text-service association relationship and the external environment influence association are established, wherein the logical dependency relationship within the text is established by using the LAC model dependency syntax analysis to extract the conditional, causal and parallel relationship between clauses, and the confidence threshold is greater than or equal to 0.75; the text-service association relationship is established by using a pre-set rule engine to associate the contract clauses with the business indicators, the number of association rules is greater than or equal to 500, each rule contains the mapping relationship between the contract clause features and the business indicators, when the contract clause features and the business indicator data meet the rule triggering condition, the dynamic matching is realized through the rule engine; the external environment influence association is established through an API interface, the latest legal provisions are synchronized from the national legal regulations database at 00:00 every day, the TF-IDF+Cosine algorithm is used to calculate the matching degree of the legal provisions and the contract clauses, when the matching degree of the legal provisions and the contract clauses is less than 0.8, the 'compliance' risk entity of the corresponding contract clause is automatically marked as a to-be-evaluated state, and an artificial review process is triggered, the system pushes the to-be-evaluated task to the legal department, and the review personnel need to complete the clause re-evaluation and update the risk score within 2 working days;

[0018] The updating unit: three updating mechanisms of time triggering, event triggering and rule triggering are set;

[0019] The quantitative evaluation module: the multi-source data is input into a fusion model, the GCN model algorithm is used to mine the implicit association between risk entities, and a three-dimensional evaluation system containing risk occurrence probability, influence degree and transmission path complexity is constructed; the dimension weight is determined by using the AHP hierarchical analysis method, and the random forest model is generated by combining the historical case training, the overall risk quantitative score of the contract is generated, wherein the low risk is 1-3 points, including 3 points, the medium risk is 4-6 points, including 6 points, the high risk is 7 points, and the extremely high risk is more than 8 points, and the risk heat map and the transmission path diagram are displayed through a visual interface, the risk warning is supported by the email notification and the system pop-up window.

[0020] The specific steps of the data acquisition and preprocessing module are as follows:

[0021] SA1, start;

[0022] SA2, character recognition is performed on the contract scan by using an OCR engine, a text file is generated, and operation data is collected from an enterprise business system by using an ETL tool, and an association relationship with a contract principal ID is established;

[0023] SA3, adopt BERT pre-training model to perform named entity recognition on the text, and extract contract subject entity, object entity, performance deadline entity and breach of contract clause entity, specifically: divide the text by sentence, each sentence length ≤512 tokens, extract 768-dimensional embedding vector of each token through the model, use maximum pooling method to aggregate the embedding vectors of all tokens in the sentence into a sentence vector, and map the sentence vector to an entity label through a fully connected layer;

[0024] SA4, extract the logical relationship between entities through the dependency syntax analysis LAC model, and the confidence threshold is ≥0.75;

[0025] SA5, structure the extracted entity and relationship data and store it in the PostgreSQL database through the ETL tool, and establish a relational data model including four tables of contracts, entities, clauses and relationships;

[0026] SA6, end.

[0027] Wherein, the specific steps of the dynamic risk map construction module are as follows:

[0028] SB1, start;

[0029] SB2, import structured data into Neo4j graph database, create risk entity nodes, including: contract text risk entity, business operation risk entity and external environment risk entity;

[0030] SB3, establish the association relationship between entities through a pre-set rule engine, including: logical dependency relationship within the text, text-business association relationship and external environment influence association; the establishment of the association relationship between entities adopts a rule engine, and the specific rules are stored in the relationship_rules table, each rule includes: rule ID, trigger condition, association action and effective time;

[0031] SB4, set dynamic update trigger, including: time trigger, event trigger and rule trigger;

[0032] Time trigger: automatically update the 'last update time' attribute of the external environment risk entity at 2:00 am every day, and reacquire the latest legal provisions through the API interface to calculate the legal change risk score and the industry supervision risk score;

[0033] Event trigger: Subscribe to the update message of the external event source through the message queue. After the message consumer receives the event message, it parses the message content and updates the corresponding entity attribute. The APOC plug-in of the graph database is called to trigger the recalculation of the associated entity. When the contract subject is involved in a lawsuit, the "credit rating" and "record of litigation" attributes of the subject are updated within 30 minutes;

[0034] Rule trigger: The default threshold is 70%, and the threshold can be customized. When the enterprise financial indicators exceed the default threshold, the risk recalculation of all contracts with the enterprise as the subject is triggered;

[0035] SB5, end.

[0036] The specific steps of the quantitative evaluation module are as follows:

[0037] SC1, start;

[0038] SC2, extract risk entity and associated relationship data from the graph database, and construct a feature matrix;

[0039] SC3, input the feature matrix into the trained GCN model algorithm to generate risk node embedding vectors;

[0040] SC4, based on the embedding vectors, construct a three-dimensional evaluation system containing risk occurrence probability, impact degree and transmission path complexity, and determine the dimension weight through AHP hierarchical analysis method. Combine the historical case training random forest model to generate the contract overall risk quantitative score:

[0041] SC5, calculate the final risk score according to the following formula: risk score = risk occurrence probability x 0.5396 x impact degree coefficient + impact degree basic value x 0.2970 + transmission path complexity index x 0.1634, wherein the impact degree basic value is determined by the grade corresponding value output by the gradient boosting tree model. The impact degree I level is 1.0, the II level is 1.5, the III level is 2.0, and the IV level is 2.5. The impact degree basic value is determined by the grade corresponding value output by the gradient boosting tree model;

[0042] SC6, generate the corresponding risk level according to the risk score. The risk level is divided into low, medium, high and extremely high;

[0043] SC7, store the evaluation results to the risk_assessment_results table, and output through the API interface;

[0044] SC8, end.

[0045] In SC3, the training parameters of the GCN model are:

[0046] Training dataset: contains 5000 historical contract risk cases, divided into training set, validation set and test set in the ratio of 7:2:1;

[0047] Model structure: input layer dimension is 128, hidden layer dimension is 64, output layer dimension is 32;

[0048] Training parameters: learning rate is 0.001, batch size is 64, optimizer is Adam, loss function is MSE loss;

[0049] Training rounds: 200 rounds;

[0050] Evaluation index: test set MSE is 0.021, risk level prediction accuracy is 87.3%.

[0051] In SC4, the specific construction of three-dimensional evaluation system is as follows:

[0052] Risk occurrence probability: calculated by random forest model, input features include clause compliance score, subject credit rating, historical compliance rate, output probability value range [0-1];

[0053] Influence degree: calculated by gradient boosting tree model, input features include contract amount, subject type, proportion of default liability clause liquidated damages, output influence degree level, influence degree level is divided into I-IV level;

[0054] Conduction path complexity: calculated by graph algorithm, by calculating the shortest path length between risk entities and the PageRank value of risk entities, the result is mapped to the range of [0-100] through linear normalization processing, and the conduction path complexity index is calculated.

[0055] In SC4, the weight of three-dimensional evaluation system is determined by AHP hierarchical analysis method, and the specific steps are as follows:

[0056] SD1, start;

[0057] SD2, construct a 3×3 judgment matrix A;

[0058] SD3, calculate the maximum eigenvalue of the judgment matrix λmax=3.009;

[0059] SD4, calculate the consistency index CI=0.0045;

[0060] SD5, query random consistency index RI=0.58;

[0061] SD6, calculate the consistency ratio CR=0.0077<0.1, pass the consistency test;

[0062] SD7, calculate the weight vector W = [0.5396, 0.2970, 0.1634];

[0063] SD8, end.

[0064] The quantitative evaluation module further comprises:

[0065] The visualization interface is configured to realize the risk heat map by using ECharts, and the color mapping rules are: 1-3 points are green, 4-6 points are yellow, 7 points are orange, and 8-10 points are red.

[0066] The early warning threshold is configured as: 4-6 points are yellow warning, 7 points are orange warning, and 8 points or more are red warning, which correspond to pop-up window notification, email notification of department heads, and email notification of company executives.

[0067] Compared with the prior art, the beneficial effects of the present application are:

[0068] 1. By defining real-time updateable risk entities and associated relationships, combined with the triple update mechanism of time trigger, event trigger and rule trigger, the risk map can dynamically evolve with the changes of internal and external environment. Compared with the traditional static rule engine, this mechanism can respond to legal regulations update, industry policy changes and enterprise business adjustment in real time, and reduce the problem that the risk assessment model lags behind the actual risk changes.

[0069] 2. Breakthrough the limitation of single contract text review, deeply associate the contract text risk with business operation data and external environment data, match the compliance of legal clauses and contract clauses in real time through external environment API interface, and the evaluation result is closer to the risk status in the actual operation of the enterprise.

[0070] 3. Use GCN graph neural network algorithm to mine the implicit association between risk entities, and combine AHP hierarchical analysis method to determine the weight of three-dimensional evaluation system. This method reduces the dependence on artificial experience and reduces the subjective bias of traditional manual review, improves the objectivity and accuracy of the evaluation result.

[0071] 4. Realize the visual display of risk heat map and conduction path map by ECharts, support user interactive query of risk details, and combine threshold configuration to trigger pop-up window notification and email early warning mechanism, help enterprise management to quickly locate high-risk areas, develop targeted risk control strategies, and improve contract management efficiency and risk prevention and control response speed. BRIEF DESCRIPTION OF DRAWINGS

[0072] Figure 1 It is a schematic diagram of the overall architecture of the system of the present application;

[0073] Figure 2 It is a schematic diagram of the data acquisition and preprocessing module of the present application;

[0074] Figure 3 A dynamic risk map construction module flowchart of the present application is shown in Figure 1.

[0075] Figure 4 An update unit flowchart of the present application is shown in Figure 2.

[0076] Figure 5 A quantitative evaluation module flowchart of the present application is shown in Figure 3.

[0077] Figure 6 A risk entity and associated relationship schematic table of the present application is shown in Figure 4.

[0078] Figure 7 A risk assessment visualization and early warning mechanism schematic table of the present application is shown in Figure 5. DETAILED DESCRIPTION

[0079] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments of the present application.

[0080] Therefore, the following detailed description of the embodiments of the present application is not intended to limit the scope of the claimed present application, but only represents some embodiments of the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present application without making creative efforts are within the scope of protection of the present application.

[0081] It should be noted that the embodiments in the present application and the features and technical solutions in the embodiments can be combined with each other without conflict.

[0082] It should be noted that: similar reference numerals and letters represent similar items in the following drawings, and therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0083] In the description of the present application, it should be noted that the terms "upper", "lower", and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, or the orientation or positional relationship in which the product of the present application is usually placed, or the orientation or positional relationship commonly understood by those skilled in the art, and such terms are only for the convenience of describing the present application and simplifying the description, and are not intended to indicate or imply that the indicated device or element must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the present application. In addition, the terms "first", "second", and the like are only used to distinguish the description, and cannot be understood as indicating or implying relative importance.

[0084] The present application provides the following technical solutions:

[0085] Referring to Figures 1-7 A quantitative evaluation system based on a contract file document dynamic risk graph model, comprising: a data acquisition and preprocessing module, a dynamic risk graph construction module and a quantitative evaluation module;

[0086] The data acquisition and preprocessing module: the scanned contract PDF is subjected to character recognition by an optical character recognition (OCR) engine; a BERT pre-training model is used to perform semantic analysis on unstructured text, the text is segmented by sentence, the length of each sentence is ≤512 tokens, key entities such as contract subjects, objects, performance periods and breach of contract clauses are extracted, and logical relationships between entities are extracted by a dependency syntax analysis (LAC) model; the entity and relationship data are structured and stored in a PostgreSQL database by an ETL tool, and a relational data model comprising four tables of contracts, entities, clauses and relationships is established, wherein:

[0087] The storage format of the contract subject is: {subject ID, name, unified social credit code, legal representative};

[0088] The storage format of the object is: {object ID, name, specification, quantity, unit price};

[0089] The storage format of the performance period is: {clause ID, start date, end date, performance mode};

[0090] The storage format of the breach of contract clause is: {clause ID, breach of contract situation, breach of contract calculation method, dispute resolution method};

[0091] The dynamic risk graph construction module: a Neo4j graph database is used to store risk entities and associated relationships, including an entity definition unit, an associated relationship unit and an update unit;

[0092] The entity definition unit: defines contract text risk entities, business operation risk entities and external environment risk entities;

[0093] The storage format of the contract text risk entity is: {entity ID, clause ID, compliance score [0-100], balance of rights and obligations score [0-100]};

[0094] The storage format of the business operation risk entity is: {entity ID, performance capability score [0-100], credit rating [AAA-E], 12-month performance rate [0-100%], update timestamp}, wherein the performance capability score, credit rating and 12-month performance rate support real-time dynamic updating;

[0095] The storage format of the external environment risk entity is: {entity ID, legal change risk score [0-100], industry supervision risk score [0-100], latest update time};

[0096] The association relationship unit: establish the logical dependency relationship within the text, the text-service association relationship and the external environment influence association, wherein the establishment of the logical dependency relationship within the text adopts LAC model dependency syntax analysis to extract the conditional, causal and parallel relationship between clauses, and the confidence threshold is greater than or equal to 0.75; the establishment of the text-service association relationship is realized by a pre-set rule engine to associate the contract clauses with the business indicators, the number of association rules is greater than or equal to 500, each rule contains the mapping relationship between the contract clause features and the business indicators, when the contract clause features and the business indicator data meet the rule triggering condition, the dynamic matching is realized through the rule engine; the establishment of the external environment influence association is realized through an API interface, the latest legal provisions are synchronized from the national legal regulations database at 00:00 every day, the TF-IDF+Cosine algorithm is used to calculate the matching degree of the legal provisions and the contract clauses, when the matching degree of the legal provisions and the contract clauses is less than 0.8, the 'compliance' risk entity of the corresponding contract clause is automatically marked as a to-be-evaluated state, and an artificial review process is triggered, the system pushes the to-be-evaluated task to the legal department, and the review personnel need to complete the clause re-evaluation and update the risk score within 2 working days;

[0097] The update unit: set three kinds of update mechanisms of time trigger, event trigger and rule trigger;

[0098] The quantitative evaluation module: input the multi-source data into the fusion model, use the GCN model algorithm to mine the implicit association between risk entities, and construct a three-dimensional evaluation system including risk occurrence probability, influence degree and transmission path complexity; determine the dimension weight by AHP hierarchical analysis method, combine the historical case training random forest model to generate the contract overall risk quantitative score, wherein the low risk is 1-3 points including 3 points, the medium risk is 4-6 points including 6 points, the high risk is 7 points, and the extremely high risk is more than 8 points, and the risk heat map and the transmission path diagram are displayed through the visual interface, the risk warning is supported by the email notification and the system pop-up window.

[0099] The specific steps of the data collection and preprocessing module are as follows:

[0100] SA1, start;

[0101] SA2, perform character recognition on the contract scan through the OCR engine, generate a text file, and collect operation data from the enterprise business system through the ETL tool to establish the association relationship with the contract subject ID;

[0102] SA3, adopt the BERT pre-training model to perform named entity recognition on the text, and extract contract subject entity, object entity, performance deadline entity and breach of contract clause entity, specifically: divide the text by sentence, each sentence length ≤512 tokens, extract the 768-dimensional embedding vector of each token through the model, use the maximum pooling method to aggregate the embedding vectors of all tokens in the sentence into a sentence vector, and map the sentence vector to an entity label through a fully connected layer;

[0103] SA4, extract the logical relationship between entities through the dependency syntax analysis LAC model, and the confidence threshold is ≥0.75;

[0104] SA5, structure the extracted entity and relationship data and store it in the PostgreSQL database through the ETL tool, and establish a relational data model including four tables of contracts, entities, clauses and relationships;

[0105] SA6, end.

[0106] Among them, the specific steps of the dynamic risk map construction module are as follows:

[0107] SB1, start;

[0108] SB2, import structured data into the Neo4j graph database, and create risk entity nodes, including: contract text risk entities, business operation risk entities and external environment risk entities;

[0109] SB3, establish the association relationship between entities through a pre-set rule engine, including: logical dependency relationship within the text, text-business association relationship and external environment influence association; the establishment of the association relationship between entities adopts a rule engine, and the specific rules are stored in the relationship_rules table, each rule includes: rule ID, trigger condition, association action and effective time;

[0110] SB4, set a dynamic update trigger, including: time trigger, event trigger and rule trigger;

[0111] Time trigger: automatically update the 'last update time' attribute of the external environment risk entity at 2:00 am every day, and reacquire the latest legal provisions through the API interface to calculate the legal change risk score and the industry supervision risk score;

[0112] Event trigger: Subscribe to the update message of the external event source through the message queue. After the message consumer receives the event message, it parses the message content and updates the corresponding entity attribute. The APOC plug-in of the graph database is called to trigger the recalculation of the associated entity. When the contract subject is involved in a lawsuit, the "credit rating" and "record of litigation" attributes of the subject are updated within 30 minutes.

[0113] Rule trigger: The default threshold is 70%, and the threshold can be customized. When the enterprise financial indicators exceed the default threshold, the risk recalculation of all contracts with the enterprise as the subject is triggered.

[0114] SB5, end.

[0115] The specific steps of the quantitative evaluation module are as follows:

[0116] SC1, start;

[0117] SC2, extract risk entity and associated relationship data from the graph database, and construct a feature matrix;

[0118] SC3, input the feature matrix into the trained GCN model algorithm to generate risk node embedding vectors;

[0119] SC4, based on the embedding vectors, construct a three-dimensional evaluation system containing risk occurrence probability, impact degree and transmission path complexity, and determine the dimension weight through AHP hierarchical analysis method. Combine the historical case training random forest model to generate the contract overall risk quantitative score:

[0120] SC5, calculate the final risk score according to the following formula: risk score = risk occurrence probability x 0.5396 x impact degree coefficient + impact degree basic value x 0.2970 + transmission path complexity index x 0.1634, wherein the impact degree basic value is determined by the grade corresponding value output by the gradient boosting tree model. The impact degree I level is 1.0, II level is 1.5, III level is 2.0, IV level is 2.5. The impact degree basic value is determined by the grade corresponding value output by the gradient boosting tree model.

[0121] SC6, generate the corresponding risk level according to the risk score. The risk level is divided into low, medium, high and extremely high;

[0122] SC7, store the evaluation results to the risk_assessment_results table, and output through the API interface;

[0123] SC8, end.

[0124] In SC3, the training parameters of the GCN model are:

[0125] Training dataset: contains 5000 historical contract risk cases, divided into training set, validation set and test set in the ratio of 7:2:1;

[0126] Model structure: input layer dimension is 128, hidden layer dimension is 64, and output layer dimension is 32;

[0127] Training parameters: learning rate is 0.001, batch size is 64, optimizer is Adam, and loss function is MSE loss;

[0128] Training rounds: 200 rounds;

[0129] Evaluation index: test set MSE is 0.021, and risk level prediction accuracy is 87.3%.

[0130] In SC4, the specific construction of the three-dimensional evaluation system is as follows:

[0131] Risk occurrence probability: calculated by random forest model, input features include clause compliance score, subject credit rating, historical compliance rate, and output probability value range [0-1];

[0132] Influence degree: calculated by gradient boosting tree model, input features include contract amount, subject type, and proportion of default liability clause default fine, output influence degree level, influence degree level is divided into I-IV level;

[0133] Conduction path complexity: calculated by graph algorithm, by calculating the shortest path length between risk entities and the PageRank value of risk entities, the result is mapped to the range of [0-100] through linear normalization processing, and the conduction path complexity index is calculated.

[0134] In SC4, the weight of the three-dimensional evaluation system is determined by AHP hierarchical analysis method, and the specific steps are as follows:

[0135] SD1, start;

[0136] SD2, construct a 3x3 judgment matrix A;

[0137] SD3, calculate the maximum eigenvalue of the judgment matrix λmax=3.009;

[0138] SD4, calculate the consistency index CI=0.0045;

[0139] SD5, query random consistency index RI=0.58;

[0140] SD6, calculate the consistency ratio CR=0.0077<0.1, pass the consistency test;

[0141] SD7, calculate the weight vector W = [0.5396, 0.2970, 0.1634];

[0142] SD8, end.

[0143] The quantitative evaluation module further comprises:

[0144] The visual interface is configured to realize the risk heat map by using ECharts, and the color mapping rules are: 1-3 points are green, 4-6 points are yellow, 7 points are orange, and 8-10 points are red.

[0145] The early warning threshold is configured as: 4-6 points are yellow early warning, 7 points are orange early warning, and 8 points or more are red early warning, which correspond to pop-up window notification, email notification of department heads, and email notification of company executives, respectively.

[0146] In order for those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings.

[0147] It should be noted that the embodiments in the present application and the features and technical solutions in the embodiments can be combined with each other without conflict.

[0148] It should be noted that: similar reference numbers and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0149] The above embodiments are only used to illustrate the present application and not to limit the technical solutions described in the present application. Although the present application has been described in detail with reference to the above embodiments, the present application is not limited to the above specific embodiments, and any modification or equivalent replacement of the present application; all technical solutions and improvements that do not deviate from the spirit and scope of the present application are covered in the scope of the claims of the present application.

Claims

1. A quantitative evaluation system based on a contract file document dynamic risk map model, characterized in that, Comprise: Data acquisition and preprocessing module, dynamic risk map construction module and quantitative evaluation module; Data acquisition and preprocessing module: through the optical character recognition OCR engine to scan the contract PDF character recognition; using BERT pre-training model for semantic analysis of unstructured text, according to the sentence segmentation text, each sentence length ≤512 token, extraction key entity includes: contract subject, object, performance period and breach of contract liability clause, and through the dependency syntax analysis LAC model to extract the logical relationship between entities; through ETL tool to structure the entity and relationship data storage to PostgreSQL database, to establish a relational data model including contracts, entities, clauses, relationships four tables, wherein: The storage format of the contract subject is: {subject ID, name, unified social credit code, legal representative}; The storage format of the object is: {object ID, name, specification, quantity, unit price}; The storage format of the performance period is: {clause ID, start date, end date, performance mode}; The storage format of the breach of contract liability clause is: {clause ID, breach of contract situation, breach of contract calculation method, dispute resolution method}; Dynamic risk map construction module: using Neo4j graph database to store risk entities and associated relationships, including: entity definition unit, associated relationship unit and update unit; Entity definition unit: define contract text risk entity, business operation risk entity and external environment risk entity; Among them, the storage format of the contract text risk entity is: {entity ID, clause ID, compliance score [0-100], balance of rights and obligations score [0-100]}; The storage format of the business operation risk entity is: {entity ID, performance ability score [0-100], credit rating [AAA-E], 12 months of performance rate [0-100%], update timestamp}, wherein, the performance ability score, credit rating, 12 months of performance rate support real-time dynamic update; The storage format of the external environment risk entity is: {entity ID, legal change risk score [0-100], industry supervision risk score [0-100], latest update time}. The association relationship unit: establish the logical dependency relationship within the text, the text-service association relationship and the external environment influence association, wherein the establishment of the logical dependency relationship within the text adopts LAC model dependency syntax analysis to extract the conditional, causal and parallel relationship between clauses, and the confidence threshold is greater than or equal to 0.75; the establishment of the text-service association relationship is realized by a preset rule engine to associate the contract clauses with the business indicators, the number of association rules is greater than or equal to 500, each rule contains the mapping relationship between the contract clause features and the business indicators, when the contract clause features and the business indicator data meet the rule triggering condition, the dynamic matching is realized through the rule engine; the establishment of the external environment influence association is realized through an API interface, the latest legal provisions are synchronized from a national legal regulation database at 00:00 every day, the TF-IDF+Cosine algorithm is used to calculate the matching degree of the legal provisions and the contract clauses, when the matching degree of the legal provisions and the contract clauses is less than 0.8, the 'compliance' risk entity of the corresponding contract clause is automatically marked as a to-be-evaluated state, and an artificial review process is triggered, the system pushes the to-be-evaluated task to a legal department, and a reviewer needs to complete clause reevaluation and update the risk score within 2 working days; The updating unit: three updating mechanisms of time triggering, event triggering and rule triggering are set; The quantitative evaluation module: multi-source data is input into a fusion model, a GCN model algorithm is used to mine the implicit association between risk entities, a three-dimensional evaluation system containing risk occurrence probability, influence degree and conduction path complexity is constructed; the dimension weight is determined by an AHP hierarchical analysis method, a random forest model is generated by combining historical case training, the contract overall risk quantitative score is generated, wherein the low risk is 1-3 points including 3 points, the medium risk is 4-6 points including 6 points, the high risk is 7 points, and the extremely high risk is 8 points and above, and the risk heat map and the conduction path diagram are displayed through a visual interface, the risk warning is supported by email notification and system pop-up window; The specific steps of the dynamic risk map construction module are as follows: SB1, start; SB2, import structured data into a Neo4j graph database, and create risk entity nodes, including: contract text risk entities, business operation risk entities and external environment risk entities; SB3, establish the association relationship between entities through a preset rule engine, including: the logical dependency relationship within the text, the text-service association relationship and the external environment influence association; the establishment of the association relationship between entities adopts the rule engine, and the specific rules are stored in the relationship_rules table, each rule contains: rule ID, triggering condition, association action and effective time; SB4, set a dynamic updating trigger, including: time triggering, event triggering and rule triggering; Time triggering: automatically update the 'latest update time' attribute of the external environment risk entity at 2:00 am every day, and reacquire the latest legal provisions through an API interface, calculate the legal change risk score and the industry supervision risk score; Event trigger: Subscribe to the update message of external event sources through the message queue. After receiving the event message, the message consumer parses the message content and updates the corresponding entity attributes. The APOC plug-in of the graph database is called to trigger the recalculation of associated entities. When a lawsuit event of a contract subject is monitored, the "credit rating" and "lawsuit record" attributes of the subject are updated within 30 minutes. Rule trigger: The default threshold is 70%, and the threshold can be customized. When the enterprise financial indicators exceed the default threshold, the risk recalculation of all contracts with the enterprise as the subject is triggered. SB5, end. The specific steps of the quantitative evaluation module are as follows: SC1, start; SC2, extract risk entity and associated relationship data from the graph database to construct a feature matrix; SC3, input the feature matrix into the trained GCN model algorithm to generate risk node embedding vectors; SC4, based on the embedding vectors, construct a three-dimensional evaluation system including risk occurrence probability, impact degree, and transmission path complexity. Determine the dimension weight through AHP hierarchical analysis method, and combine historical case training to generate random forest model to generate contract overall risk quantitative score: SC5, calculate the final risk score according to the following formula: risk score = risk occurrence probability x 0.5396 x impact degree coefficient + impact degree basic value x 0.2970 + transmission path complexity index x 0.1634, wherein the impact degree basic value is determined by the grade corresponding value output by the gradient boosting tree model, the impact degree I level is 1.0, the II level is 1.5, the III level is 2.0, and the IV level is 2.

5. The impact degree basic value is determined by the grade corresponding value output by the gradient boosting tree model; SC6, generate the corresponding risk level according to the risk score, and the risk level is divided into low, medium, high and extremely high; SC7, store the evaluation results to the risk_assessment_results table, and output through the API interface; SC8, end; In SC3, the training parameters of the GCN model are as follows: Training data set: contains 5000 historical contract risk cases, divided into training set, validation set and test set in the ratio of 7:2:1; Model structure: input layer dimension is 128, hidden layer dimension is 64, and output layer dimension is 32; Training parameters: learning rate is 0.001, batch size is 64, optimizer is Adam, and loss function is MSE loss; Training rounds: 200 rounds; Evaluation index: test set MSE is 0.021, and risk level prediction accuracy is 87.3%; In SC4, the specific construction of the three-dimensional evaluation system is as follows: Risk occurrence probability: calculated by random forest model, input features include clause compliance score, subject credit rating, and historical compliance rate, and output probability value range [0-1]; Impact degree: calculated by gradient boosting tree model, input features include contract amount, subject type, and default liability clause default fine ratio, output impact degree level, and impact degree level is divided into I-IV level. Conduction path complexity: through graph algorithm calculation, by calculating the shortest path length between risk entities and the PageRank value of risk entities, the results are mapped to the range of [0-100] through linear normalization processing, and the conduction path complexity index is calculated comprehensively; In the SC4, the weight of the three-dimensional evaluation system is determined by the AHP hierarchical analysis method, and the specific steps are as follows: SD1, start; SD2, construct a 3*3 judgment matrix A; SD3, calculate the maximum eigenvalue of the judgment matrix λmax=3.009; SD4, calculate the consistency index CI=0.0045; SD5, query the random consistency index RI=0.58; SD6, calculate the consistency ratio CR=0.0077<0.1, pass the consistency test; SD7, calculate the weight vector W=[0.5396, 0.2970, 0.1634]; SD8, end; The quantitative evaluation module further comprises: The visual interface is configured to use ECharts to realize the risk heat map, and the color mapping rules are: 1-3 points for green, 4-6 points for yellow, 7 points for orange, and 8-10 points for red; The early warning threshold is configured as: 4-6 points for yellow warning, 7 points for orange warning, and 8 points or more for red warning, which correspond to pop-up window notification, email notification of department responsible person, and email notification of company executives.

2. The system for quantitative evaluation based on a dynamic risk map model of contract file documents according to claim 1, characterized in that, The specific steps of the data collection and preprocessing module are as follows: SA1, start; SA2, perform character recognition on the contract scan through the OCR engine, generate a text file, and collect operation data from the enterprise business system through the ETL tool to establish an association relationship with the contract subject ID; SA3, use the BERT pre-training model to perform named entity recognition on the text, and extract contract subject entities, subject entities, performance period entities, and breach of contract clause entities, specifically: divide the text by sentence, and the length of each sentence is ≤512 tokens, extract the 768-dimensional embedding vector of each token through the model, use the maximum pooling method to aggregate the embedding vectors of all tokens in the sentence into a sentence vector, and map the sentence vector to an entity label through a fully connected layer; SA4, extract the logical relationship between entities through the dependency syntax analysis LAC model, and the confidence threshold is ≥0.75; SA5, structure the extracted entity and relationship data and store it in the PostgreSQL database through the ETL tool to establish a relational data model including four tables of contracts, entities, clauses, and relationships; SA6, end.

Images