Business contract key clause intelligent review and risk quantification method and device

Abstract

The invention relates to the technical field of artificial intelligence, in particular to a business contract intelligent review and risk quantification method and device, and the method comprises the steps: building and maintaining a business contract key term information base; obtaining and preprocessing a to-be-rechecked contract text; processing the text based on a bidirectional long-short term memory network and a conditional random field model, and extracting semantic representation; identifying key information through a multi-level attention mechanism; executing multi-label learning to classify and identify clause types and attributes; utilizing a dependency syntactic analysis technology to extract logical association and a responsibility chain among terms, and constructing a knowledge graph; identifying risk terms and generating risk prompts; business indexes are extracted, and risk open values are calculated; generating a rechecking report; the corresponding device comprises nine functional modules such as an information base management module, a text preprocessing module and a semantic representation extraction module, risk terms in a contract can be automatically recognized, a quantitative risk assessment result is provided, and contract auditing efficiency and accuracy are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence, in particular to a method and device for intelligent review and risk quantification of business contracts, which can be widely applied in enterprise legal management, business risk control, intelligent contract review and other fields. BACKGROUND

[0002] Business contracts are important legal basis for enterprise business activities, and various clauses contained therein are directly related to the interests and risks of enterprises. Traditional contract review mainly relies on manual inspection by legal personnel, which not only consumes time and effort, but also is prone to omission of important risk clauses due to human factors, especially when dealing with a large number of contracts or complex contracts, the quality and efficiency of review are difficult to guarantee.

[0003] There are some contract review tools on the market at present, but there are generally the following problems: first, the rule system based on keyword matching is difficult to adapt to diversified contract expression methods; second, there is a lack of deep understanding of the logical association between contract clauses; third, the risk assessment process is not transparent and lacks quantitative standards; fourth, it is difficult to effectively identify associated risks scattered in different clauses. With the development of artificial intelligence and natural language processing technology, the application of deep learning technology in the field of business contract review to realize intelligent, standardized and quantitative risk assessment has become an important direction of industry development. SUMMARY

[0004] The purpose of the present application is to provide a method and device for intelligent review and risk quantification of key clauses of business contracts, aiming to solve the problems of low efficiency, poor consistency and non-transparent risk assessment in traditional contract review, and to realize intelligent review and quantitative assessment of risks of business contracts.

[0005] The present application provides a method for intelligent review and risk quantification of key clauses of business contracts, which comprises:

[0006] establishing and maintaining a key clause information library of business contracts, wherein the key clause information library of business contracts comprises standard clause templates and risk assessment benchmarks;

[0007] obtaining a business contract text to be reviewed, and preprocessing the business contract text to obtain a structured text sequence;

[0008] processing the structured text sequence based on a bidirectional long short-term memory network and a conditional random field model to extract semantic representations of contract clauses, wherein the semantic dependency relationship between the pre- and post-context of the clauses is captured through bidirectional feature fusion;

[0009] According to the semantic representation, key information in the contract text is identified through a multi-level attention mechanism, wherein the multi-level attention mechanism comprises a word-level legal term enhanced attention, a sentence-level semantic key point identification attention, and a clause-level risk assessment attention.

[0010] Based on the key information, multi-label learning classification is performed to identify the clause type and clause attribute in the business contract, wherein the clause type comprises a business liability clause, a breach of contract clause, and a business indicator clause, and the clause attribute comprises a liability subject, a liability object, a liability limitation, and a time constraint.

[0011] Using dependency syntax analysis technology, logical associations and responsibility chains between contract clauses are extracted based on the clause type and clause attribute, and a contract knowledge graph is constructed.

[0012] According to the contract knowledge graph, risk clauses in the business contract are identified and compared with standard clauses in the key clause information base of the business contract to generate risk prompt content.

[0013] Business indicators, including transaction amount, payment period, and payment method, are extracted from the business contract, and risk exposure values are calculated based on the business indicators and the risk clauses.

[0014] A contract review report containing risk clauses, risk prompt content, and risk exposure values is generated and pushed to an auditor.

[0015] As a preferred embodiment, the establishment and maintenance of the key clause information base of the business contract specifically comprises:

[0016] Obtaining standard contract texts of multiple industries;

[0017] Analyzing the standard contract texts to extract standard clause content and logical relationships between clauses;

[0018] Based on the standard clause content, a mapping relationship between clause titles, liability limitation information, and clause risk levels is constructed;

[0019] According to historical audit data and legal expert knowledge, risk assessment benchmarks are set for different types of clauses;

[0020] The key clause information base of the business contract is updated regularly to adapt to changes in laws and regulations and industry standards.

[0021] As a preferred embodiment, the processing of the structured text sequence based on a bidirectional long short-term memory network and a conditional random field model specifically comprises:

[0022] The structured text sequence is input into a bidirectional long short-term memory network to obtain forward and backward feature representations of the text;

[0023] fusing the forward feature representation and the backward feature representation to generate a comprehensive feature representation;

[0024] performing sequence labeling through a conditional random field model based on the comprehensive feature representation to identify key elements in the contract text;

[0025] extracting contextual information for the key elements to form an enhanced semantic representation.

[0026] As a preferred embodiment, the multi-level attention mechanism specifically includes:

[0027] identifying professional terms in the contract text based on a predefined legal terminology dictionary and assigning higher attention weights to these terms;

[0028] analyzing sentence structures to identify key semantic nodes representing conditions, responsibilities, and consequences and dynamically adjusting the attention distribution of these nodes;

[0029] evaluating the risk level of clauses and assigning higher attention weights to high-risk clauses;

[0030] constructing a reference relationship graph between clauses to identify explicit references and implicit associations and ensuring that associated clauses are given sufficient attention.

[0031] As a preferred embodiment, the multi-label learning classification specifically includes:

[0032] performing multi-dimensional classification of clauses based on a hierarchical label system;

[0033] using a label dependency relationship model to capture logical associations between different labels;

[0034] addressing the label imbalance problem in legal text by assigning higher weights to rare but important label categories;

[0035] performing label conflict detection to identify and correct logical contradictions in the predicted label set based on legal logic rules.

[0036] As a preferred embodiment, the extraction of logical associations and responsibility chains between contract clauses specifically includes:

[0037] constructing a semantic dependency tree for each clause to identify the semantic structure within the clause;

[0038] extracting complete responsibility chains, including responsibility subjects, responsibility behaviors, responsibility objects, responsibility limitations, and consequences of breach of contract;

[0039] analyzing clause completeness to assess the absence of elements in the responsibility chain;

[0040] identifying logical consistency between clauses to detect responsibility conflicts, time conflicts, and scope conflicts.

[0041] As preferred, the calculation of the risk exposure value specifically comprises:

[0042] Obtaining the total transaction amount A of the contract;

[0043] Identifying the risk clauses in the contract and extracting the liquidated damages Li in each risk clause;

[0044] Determining the payment method P of the contract, wherein P takes the value of 1, 2, 3 or 4, corresponding to 30 / 60 / 90 days commercial paper payment, 3 days telegraphic transfer payment, 7 days commercial paper payment and 8-15 days commercial acceptance bill respectively;

[0045] Obtaining the payment period T of the contract;

[0046] According to the total transaction amount A, liquidated damages Li, payment method P and payment period T, calculating the risk exposure value K.

[0047] As preferred, the calculation formula of the risk exposure value K is:

[0048] K=A*(1+∑Li)*P*T

[0049] Wherein, A is the total transaction amount of the contract, Li is the liquidated damages of the i-th risk clause, P is the risk weight of the payment method, and T is the risk coefficient of the payment period.

[0050] As preferred, the generation of the contract review report containing risk clauses, risk prompt content and risk exposure value specifically comprises:

[0051] Sorting the identified risk clauses according to the risk level;

[0052] For each risk clause, generating a detailed explanation containing risk description, risk reason and improvement suggestion;

[0053] Converting the risk exposure value into a risk level, which is divided into low risk, medium risk and high risk;

[0054] Generating a visual risk analysis chart to intuitively display the risk distribution;

[0055] Integrating all risk information to form a structured review report.

[0056] A business contract key clause intelligent review and risk quantification device, comprising:

[0057] An information base management module for establishing and maintaining a business contract key clause information base, which contains standard clause templates and risk assessment benchmarks;

[0058] The text preprocessing module is configured to obtain a business contract text to be reviewed, and preprocess the business contract text to obtain a structured text sequence.

[0059] The semantic representation extraction module is configured to process the structured text sequence based on a bidirectional long short-term memory network and a conditional random field model, and extract a semantic representation of a contract clause, wherein a semantic dependency relationship of context before and after the clause is captured through bidirectional feature fusion.

[0060] The attention processing module is configured to identify key information in the contract text through a multi-level attention mechanism according to the semantic representation, wherein the multi-level attention mechanism includes a word-level legal term enhanced attention, a sentence-level semantic key point identification attention, and a clause-level risk assessment attention.

[0061] The multi-label classification module is configured to perform multi-label learning classification based on the key information to identify a clause type and a clause attribute in the business contract, wherein the clause type includes a business responsibility clause, a breach of contract clause, and a business indicator clause, and the clause attribute includes a responsibility subject, a responsibility object, a responsibility limitation, and a time constraint.

[0062] The semantic analysis module is configured to extract a logical association and a responsibility chain between contract clauses based on the clause type and the clause attribute by using a dependency syntax analysis technique, and construct a contract knowledge graph.

[0063] The risk identification module is configured to identify a risk clause in the business contract according to the contract knowledge graph, and compare the risk clause with a standard clause in the business contract key clause information base to generate risk prompt content.

[0064] The risk quantification module is configured to extract a business indicator including a transaction amount, a payment period, and a payment method from the business contract, and calculate a risk exposure value based on the business indicator and the risk clause.

[0065] The report generation module is configured to generate a contract review report containing the risk clause, the risk prompt content, and the risk exposure value, and push the report to an auditor.

[0066] The present application has the following beneficial effects:

[0067] 1. Improve the auditing efficiency: the deep learning technology is used to automatically identify the key clauses and risk points of the contract, which greatly improves the auditing efficiency and enables the legal personnel to focus on the judgment work with higher value.

[0068] 2. Enhance the auditing consistency: based on the standardized evaluation system, the human subjective differences are eliminated to ensure the consistency and replicability of the auditing standard.

[0069] 3. Deep understanding of contract semantics: Using bidirectional long short-term memory networks and multi-level attention mechanisms to achieve deep semantic understanding of contract texts, beyond simple keyword matching.

[0070] 4. Discovering implicit risk associations: Through dependency syntax analysis and contract knowledge graph construction, logical associations and potential risk points scattered in different clauses can be identified.

[0071] 5. Risk quantification and visualization: Convert risk assessment results into quantifiable indicators and provide intuitive risk displays to support decision-makers in making objective comparisons and judgments.

[0072] 6. Adaptability and scalability: The system can be customized according to different industries and enterprise characteristics, and continuously improve recognition accuracy through continuous learning. BRIEF DESCRIPTION OF DRAWINGS

[0073] Figure 1 The flowchart of the business contract key clause intelligent review and risk quantification method provided by the present application.

[0074] Figure 2 The structure diagram of bidirectional long short-term memory network and conditional random field model in the present application.

[0075] Figure 3 The hierarchical structure diagram of the multi-level attention mechanism in the present application.

[0076] Figure 4 The example diagram of dependency syntax analysis and responsibility chain extraction in the present application.

[0077] Figure 5 The process diagram of risk exposure value calculation in the present application.

[0078] Figure 6 The module structure diagram of the business contract key clause intelligent review and risk quantification device provided by the present application. DETAILED DESCRIPTION

[0079] Please refer to the accompanying Figures 1-6 , the present application will be further described in detail below in combination with the drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for illustration and explanation of the present application, and are not intended to limit the present application.

[0080] Referring to Figure 1 , the business contract key clause intelligent review and risk quantification method provided by the present application includes the following steps:

[0081] Establish and maintain a business contract key clause information library: In one embodiment of the present invention, a business contract key clause information library needs to be established and maintained first, which contains standard clause templates and risk assessment benchmarks as a reference for subsequent contract review and risk quantification.

[0082] Specifically, the establishment process of the information library includes obtaining standard contract texts of multiple industries, such as sales contracts, procurement contracts, service contracts, etc. These standard contract texts are analyzed to extract standard clause content and logical relationships between clauses. Based on the extracted standard clause content, the mapping relationship of clause title, liability limitation information and clause risk level is constructed. At the same time, according to historical audit data and legal expert knowledge, risk assessment benchmarks are set for different types of clauses.

[0083] Preferably, the information library needs to be updated regularly to adapt to changes in laws and regulations and industry standards. For example, when new industry standards are released or laws and regulations are updated, the corresponding standard clause templates and risk assessment benchmarks are adjusted in time. Generally speaking, the update cycle of the information library can be set to be quarterly or semi-annual, and temporary update can be carried out in special cases.

[0084] Obtain the business contract text to be reviewed and preprocess: In this method, the business contract text to be reviewed is obtained and preprocessed to obtain a structured text sequence. Preprocessing is the basis for subsequent deep analysis, mainly including document parsing, text normalization, sentence segmentation, etc.

[0085] In specific implementation, first, contract documents in various formats (such as DOC, DOCX, PDF, etc.) are converted into pure text format. Then, text normalization processing is performed, including unifying punctuation marks, processing special characters, removing redundant spaces, etc. Then, sentence and paragraph segmentation is performed to identify the chapter structure of the contract, such as title, main text, appendix, etc. Finally, a structured text sequence is generated as the input for subsequent processing.

[0086] In addition, the preprocessing stage can also extract basic information of the contract, such as contract name, contracting parties, signing date, etc., which helps subsequent risk assessment and report generation.

[0087] Process text based on bidirectional long short-term memory network and conditional random field model Figure 2 The present invention uses bidirectional long short-term memory network (Bi-LSTM) and conditional random field (CRF) model to process structured text sequence and extract semantic representation of contract clauses. This is one of the core innovations of the present method, which realizes deep semantic understanding of contract text through deep learning technology.

[0088] In implementation, first, the structured text sequence is input into the bidirectional long short-term memory network to obtain the forward feature representation and the backward feature representation of the text. The forward processing captures the context before the text from the beginning to the end of the sentence; the backward processing captures the context after the text from the end to the beginning of the sentence. This bidirectional processing is particularly suitable for legal texts because the responsibilities and rights in legal provisions often depend on the complete understanding of the context.

[0089] The forward feature representation and the backward feature representation are fused to generate a comprehensive feature representation. The fusion process can adopt vector splicing, weighted average, etc. In this embodiment, vector splicing is preferred, and then dimension reduction is performed through a fully connected layer, which not only retains the information of the two directions but also controls the feature dimension.

[0090] Based on the comprehensive feature representation, sequence labeling is performed through a conditional random field model to identify the key elements in the contract text. The conditional random field model can consider the dependency between labels and is particularly suitable for the regularity of label transition in legal texts. For example, the "responsibility subject" label is usually followed by the "responsibility behavior" label, the "breach of contract condition" label is usually followed by the "breach of contract consequence" label, etc.

[0091] For the identified key elements, their context information is further extracted to form an enhanced semantic representation. This step combines the key elements with their context through window sliding, etc. to generate a more rich semantic representation.

[0092] In the Bi-LSTM model, important parameter settings include: the hidden layer dimension is 256, the word embedding dimension is 300, and the dropout rate (Dropout) is 0.3 to prevent overfitting. For the CRF layer, the label set includes {"responsibility subject", "responsibility object", "responsibility limit", "time constraint", "breach of contract condition", "breach of contract consequence", "other"}.

[0093] Multi-level attention mechanism identifies key information: reference Figure 3 The present application adopts a multi-level attention mechanism to identify key information in contract texts. Attention mechanism is an important technology in deep learning, which can make the model focus on the most important part of the text, improve processing efficiency and accuracy.

[0094] The multi-level attention mechanism of the present application includes three levels of word-level legal term enhanced attention, sentence-level semantic key point identification attention, and clause-level risk assessment attention.

[0095] Word-level legal terminology attention focuses on pre-defined legal terminology dictionaries to identify professional terms in contract texts, such as "liquidated damages", "force majeure", "exemption clause", etc., and gives higher attention weight. The terminology dictionary can contain thousands of commonly used legal terms and their importance scores. For example, "liquidated damages" may be assigned a base weight of 0.9, while "receipt" may be assigned a base weight of 0.5.

[0096] Sentence-level semantic key point attention focuses on analyzing sentence structure to identify key semantic nodes representing conditions, responsibilities and consequences, such as condition leading words ("if", "if"), responsibility allocation words ("should", "must"), consequence indicating words ("otherwise", "then") and dynamically adjusts the attention distribution of these nodes. This enables the model to capture the logical structure and semantic focus of the sentence.

[0097] Clause-level risk assessment attention assesses the risk level of the clause and gives higher attention weight to high-risk clauses. Risk assessment considers multiple dimensions such as responsibility allocation imbalance, clause detail, deviation from industry standards, etc. For example, a clause with payment conditions obviously biased towards the seller may be rated as high risk and receive higher attention weight.

[0098] In addition, the mechanism also constructs a reference relationship graph between clauses to identify explicit references (such as "according to Article X") and implicit associations (the same subject entity), ensuring that related clauses are given sufficient attention. This feature is particularly helpful in discovering potential risks between clauses that are logically related but scattered in different parts of the contract.

[0099] The calculation of attention weights can use the following formula:

[0100] ,

[0101] where: is the attention weight of the t-th word; is the hidden state, representing the feature representation of the current word; is the context vector, representing the contextual information around the current word; , is the weight matrix, used to transform the hidden state and context vector respectively; is the attention vector, used to calculate the relevance score; is the bias term; is the hyperbolic tangent function, used to introduce non-linear transformation; is the normalization function, ensuring that the sum of all attention weights is 1.

[0102] In practical applications, when dealing with a clause such as "Party A shall pay the contract amount within 7 days after receiving Party B's notice", the model will give higher attention weight to keywords such as "shall", "within 7 days", and "pay" because these words indicate the responsible party, time limit, and responsible behavior.

[0103] The calculation of the clause importance score can be as follows:

[0104] ,

[0105] where: is the importance score of the i-th clause; is the word-level attention-based importance function, usually taking the weighted average of word-level attention weights; is the attention weight of each word in the i-th clause; is the clause type-based risk function, different types of clauses have different risk coefficients, for example, the risk coefficient of a breach of contract clause may be 1.5, while the risk coefficient of a general description clause may be 0.8; is the clause position-based importance function, basic clauses located in the front part of the contract are more important than clauses in the appendix; is the clause type; is the relative position of the clause in the contract.

[0106] In actual evaluation, for example, for a breach of contract clause appearing in the subject part of the contract, if it contains multiple high-attention-weight keywords such as "liquidated damages", "compensation", "immediate termination", etc., and the clause type is a high-risk type, the clause will get a higher importance score, and the system will focus on its analysis and evaluation.

[0107] Multi-label learning classifies and identifies clause types and attributes: In an embodiment of the present invention, based on the key information obtained from the previous steps, multi-label learning classification is performed to identify the clause types and clause attributes in the business contract. Clause types include business responsibility clauses, breach of contract clauses, and business indicator clauses, and clause attributes include responsible parties, responsible objects, responsibility limitations, and time constraints.

[0108] Specifically, multi-label learning classification classifies clauses in multiple dimensions based on a hierarchical label system. The first level is the clause type label, such as "payment clause", "delivery clause", "breach of contract", etc.; the second level is the clause attribute label, which is further divided into responsible party label, responsible object label, responsibility limitation label, and time constraint label. This hierarchical label system enables the model to understand the content of the clause from coarse to fine.

[0109] At the same time, the label dependency model is used to capture the logical association between different labels. For example, "default clause" usually appears with "limitation of liability", "payment clause" usually contains "time constraint" and other rules. The label dependency is constructed by calculating the co-occurrence frequency and conditional probability of labels through statistical analysis of historical data, and then corrected by expert knowledge.

[0110] For the label imbalance problem in legal text, higher weight is given to rare but important label categories. For example, "force majeure clause" has a low frequency of occurrence in contracts, but is very important, so it is given a higher weight. The weight setting can use the following formula:

[0111] ,

[0112] Where: wi is the weight of the ith label; wi is the weight of the ith label; wi is the weight of the ith label;

[0113] The formula for calculating the category weight is:

[0114] ,

[0115] Where: N is the total number of samples; N is the total number of samples; Ni is the number of samples of the ith category.

[0116] For example, in a dataset containing 10,000 training samples, if "force majeure clause" has only 100 samples, and there are a total of 20 clause categories, its category weight is . If the business importance weight of this category is set to 1.5, the maximum weight is .

[0117] In addition, label conflict detection is performed to identify and correct logical contradictions in the predicted label set based on legal logic rules. For example, the same responsibility cannot be assigned to multiple subjects, and the same behavior cannot be both required to be performed and allowed not to be performed. When a conflict is detected, the conflicting label is automatically corrected based on the prediction probability and rule priority.

[0118] In practical applications, such as identifying a clause like "the seller shall deliver the goods within 30 days after receiving the buyer's advance payment", the system will classify it as "delivery clause" (clause type), while identifying "the seller" as the responsible subject, "the goods" as the responsible object, "within 30 days" as the time constraint, and "after receiving the buyer's advance payment" as the responsibility limitation (clause attribute).

[0119] Extracting logical associations and chains of responsibility among contract clauses: Refer to Figure 4 The present application extracts logical associations and chains of responsibility among contract clauses based on the aforementioned identified clause types and attributes through dependency syntax analysis technology, and constructs a contract knowledge graph. This step is the key to deeply understand the internal logic of the contract and discover potential risks.

[0120] In specific implementation, first, a semantic dependency tree is constructed for each clause to identify the semantic structure within the clause. For example, for the expression "Party A shall pay within 7 days after receiving Party B's notice", the system can correctly identify "Party A" as the subject, "pay" as the predicate, "within 7 days" as the time limit, and "after receiving Party B's notice" as the conditional limit.

[0121] Then, the complete chain of responsibility is extracted, including the responsible subject, the responsible behavior, the responsible object, the responsibility limit, and the breach of contract consequences. The complete chain of responsibility should include all five elements, and any missing element may constitute a risk point. For example, if the clause specifies the responsibility but does not specify the consequences of breach of contract, the system will mark it as "risk of missing consequences"; if it specifies strict liability but does not set reasonable limit conditions, it will be marked as "risk of excessive liability".

[0122] Then analyze the completeness of the clause and evaluate the missing elements of the chain of responsibility. The completeness score can be calculated by the following formula:

[0123] ,

[0124] Where: Completeness(clause) is the completeness score of the clause, expressed in percentage; Present_elements is the number of responsibility chain elements actually present in the clause; Total elements is the total number of elements that should be included in the complete chain of responsibility (generally 5, i.e. responsible subject, responsible behavior, responsible object, responsibility limit, and breach of contract consequences).

[0125] For example, if a clause contains the responsible subject, the responsible behavior, the responsible object, and the time limit, but lacks the consequences of breach of contract, its completeness score is 4 / 5 x 100% = 80%. According to experience, clauses with a completeness score below 80% may have a higher risk and need special attention.

[0126] Finally, identify the logical consistency among clauses, detect responsibility conflicts, time conflicts, and scope conflicts. For example, if one clause states "Party B is responsible for transportation" and another clause states "Party A assumes the responsibility for transportation", there is a conflict of responsibility; if one clause states "delivery within 10 days" and another clause states "inspection within 5 days", there may be a time conflict.

[0127] Through the above steps, a complete contract knowledge graph is constructed, which contains clause entities, relationships and attributes, and can intuitively show the internal structure and logical association of the contract.

[0128] Identifying risk clauses and generating risk prompts: In an embodiment of the present application, according to the constructed contract knowledge graph, the risk clauses in the commercial contract are identified, and compared with the standard clauses in the key clause information base of the commercial contract to generate risk prompt content.

[0129] The identification of risk clauses is based on multiple factors, including clause completeness score, responsibility allocation balance, deviation degree from industry standards, etc. For example, if a payment clause specifies a payment period that is significantly longer than the industry standard (e.g. 30 working days vs. industry standard 7 working days), the clause will be marked as a risk clause.

[0130] For each identified risk clause, the system compares it with the standard clauses in the information base to analyze the specific risk points. The comparison result generates risk prompt content, including risk description, risk level and improvement suggestion. For example, for the aforementioned clause with excessively long payment period, the system may generate the following risk prompt: "The payment period (30 working days) is significantly longer than the industry standard (7 working days), it is recommended to negotiate to shorten it to 7-10 working days to reduce the risk of funds."

[0131] The detail and professionalism of the risk prompt directly affect the decision-making efficiency of the auditors. Therefore, the prompt content should not only point out the risk, but also explain the risk reason and potential impact, and give reasonable improvement suggestions.

[0132] In practical applications, for example, when the system detects that an exemption clause is too broad, such as "the seller is not responsible for any indirect losses", it will generate a risk prompt: "The scope of exemption is too broad, it is recommended to specify the specific types and limits of indirect losses to avoid subsequent disputes. Risk level: medium-high."

[0133] Extracting business indicators and calculating risk exposure value: Referring to Figure 5 , the present application extracts business indicators from the commercial contract, including transaction amount, payment period and payment method, and calculates the risk exposure value based on these indicators and the aforementioned identified risk clauses. The risk exposure value is a quantitative assessment of the overall risk of the contract, providing an objective basis for decision-making.

[0134] In implementation, first, the total contract transaction amount A is obtained, which is usually specified in the main body of the contract or in a special price clause. Then the risk clauses in the contract are identified, and the liquidated damages Li in each risk clause is extracted. Next, the payment method P of the contract is determined, where P takes the values 1, 2, 3, or 4, corresponding to 30 / 60 / 90 days of commercial bill payment, 3 days of wire transfer payment, 7 days of commercial bill payment, and 8-15 days of commercial bill payment, respectively. Different payment methods have different risk levels, for example, wire transfer payment (P=2) has lower risk than commercial bill payment (P=4). Finally, the payment period T of the contract is obtained, which is usually specified in the payment clause.

[0135] According to the extracted total transaction amount A, liquidated damages Li, payment method P, and payment period T, the risk exposure value K is calculated, with the formula:

[0136] ,

[0137] Where: K is the risk exposure value, representing the maximum risk amount that the contract may face; A is the total contract transaction amount, in yuan; Li is the liquidated damages of the i-th risk clause, usually expressed as a percentage of the total contract amount; n is the total number of risk clauses; ∑Li is the sum of the liquidated damages of all risk clauses; P is the risk weight of the payment method, a dimensionless coefficient; T is the risk coefficient of the payment period, a dimensionless coefficient.

[0138] The risk weight P of the payment method can be set according to the enterprise's risk preference. Generally speaking:

[0139] P=1 (30 / 60 / 90 days of commercial bill payment): weight value 1.5;

[0140] P=2 (3 days of wire transfer payment): weight value 1.0;

[0141] P=3 (7 days of commercial bill payment): weight value 1.2;

[0142] P=4 (8-15 days of commercial bill payment): weight value 1.8;

[0143] These weight values are set based on actual business experience, with wire transfer payment having the lowest risk and commercial bill payment having the highest risk, so the weights range from 1.0 to 1.8.

[0144] The risk coefficient T of the payment period is related to the specific number of days and can be set using a piecewise function, for example:

[0145] T≤7 days: coefficient 1.0;

[0146] 7<T≤15 days: coefficient 1.2;

[0147] 15<T≤30 days: coefficient 1.5;

[0148] T>30 days: coefficient 2.0;

[0149] These coefficients reflect the general rule that the longer the payment period, the higher the risk. When the payment period exceeds 30 days, the risk coefficient will increase significantly.

[0150] Take a practical example: Suppose a sales contract transaction total amount A is 1 million yuan, containing two risk clauses, the liquidated damages are 5% and 3% of the total amount respectively, the payment method is commercial acceptance bill (P=4, weight 1.8), and the payment period is 45 days (T>30 days, coefficient 2.0). According to the formula, the risk exposure value K=1,000,000×(1+0.05+0.03)×1.8×2.0=3,888,000 yuan. This means that the maximum risk exposure of the contract is close to 390 million yuan, which is 3.9 times the transaction amount, belonging to high-risk contract.

[0151] The calculation result of risk exposure value K can be used to evaluate the overall risk level of the contract, which can be generally divided into three levels:

[0152] K<1.5 times of the transaction amount: low risk;

[0153] 1.5 times of the transaction amount≤K<3 times of the transaction amount: medium risk;

[0154] K≥3 times of the transaction amount: high risk;

[0155] Enterprises can adjust these thresholds according to their risk bearing capacity and compare them with risk tolerance to support decision making.

[0156] Generate contract review report: In the last step of the invention, generate a contract review report containing risk clauses, risk prompt content and risk exposure value, and push it to the auditors. The review report is a summary of the entire audit process, providing a comprehensive and intuitive risk assessment result for decision makers.

[0157] Specifically, the review report generation process first sorts the identified risk clauses by risk level to ensure that the most important risk points are presented first. Then for each risk clause, generate a detailed description containing risk description, risk cause and improvement suggestion. Then convert the risk exposure value into risk level, usually divided into low risk (K<threshold 1), medium risk (threshold 1≤K<threshold 2) and high risk (K≥threshold 2). The thresholds can be adjusted according to the enterprise's risk preference and industry characteristics.

[0158] To enhance the intuitiveness, the report can also generate visual risk analysis charts, such as risk heat maps, clause risk distribution charts, etc., to intuitively display the risk distribution. Finally, all risk information is integrated to form a structured review report, which facilitates quick understanding and decision-making by auditors.

[0159] The review report can be sent to relevant auditors through internal system push, email or other notification methods. For high-risk contracts, the system can also configure automatic reminders and tracking functions to ensure that risks are handled in a timely manner.

[0160] In practical applications, a typical review report may contain the following content:

[0161] Contract basic information (name, contracting parties, signing date, etc.);

[0162] Risk summary (overall risk level, main risk points);

[0163] Risk clause list (sorted by risk level);

[0164] Detailed analysis of each risk clause;

[0165] Risk exposure value calculation results;

[0166] Improvement suggestions;

[0167] Risk visualization charts.

[0168] For example, for the review report of the aforementioned sales contract, it may point out that the payment method and payment period are the main risk points, and suggest replacing the commercial draft with wire transfer payment or shortening the payment period to within 30 days to reduce the risk exposure value to an acceptable range.

[0169] Referring to Figure 6 The present application also provides an intelligent review and risk quantification device for key clauses of commercial contracts, which includes the following functional modules:

[0170] The information base management module 10 is used to establish and maintain the key clause information base of commercial contracts, which contains standard clause templates and risk assessment benchmarks. This module realizes the function of the first step in the aforementioned method and provides basic data support for the entire system.

[0171] The information base management module 10 includes a data acquisition unit, a data processing unit and a data updating unit. The data acquisition unit is responsible for obtaining standard contract texts and industry specifications from various channels; the data processing unit analyzes and structures the collected data; and the data updating unit is responsible for updating the information base content regularly or as needed.

[0172] The text preprocessing module 20 is configured to obtain a business contract text to be reviewed, and preprocess the business contract text to obtain a structured text sequence. The module is the basis for subsequent deep analysis, and ensures the standardization and consistency of the input data.

[0173] The text preprocessing module 20 includes a document parsing unit, a text normalization unit, and a structure recognition unit. The document parsing unit is responsible for converting contract documents of various formats into plain text; the text normalization unit performs operations such as punctuation unification and special character processing; and the structure recognition unit identifies the chapter structure and basic components of the contract.

[0174] The semantic representation extraction module 30 is configured to process the structured text sequence based on a bidirectional long short-term memory network and a conditional random field model, and extract semantic representations of contract clauses. The module is one of the core components of the device, and realizes deep semantic understanding of the contract text.

[0175] The semantic representation extraction module 30 includes a feature extraction unit, a feature fusion unit, and a sequence labeling unit. The feature extraction unit uses a Bi-LSTM network to extract forward and backward features of the text; the feature fusion unit fuses the features of the two directions; and the sequence labeling unit uses a CRF model to label the text and identify key elements.

[0176] The attention processing module 40 is configured to identify key information in the contract text through a multi-level attention mechanism based on the semantic representations. The module enables the system to focus on the most important parts of the contract, improving processing efficiency and accuracy.

[0177] The attention processing module 40 includes a word-level attention unit, a sentence-level attention unit, and a clause-level attention unit, which respectively implement word-level legal terminology enhanced attention, sentence-level semantic key point identification attention, and clause-level risk assessment attention in the aforementioned methods. In addition, it also includes a correlation analysis unit responsible for constructing a reference relationship graph between clauses.

[0178] The multi-label classification module 50 is configured to perform multi-label learning classification based on the key information, and identify the clause type and clause attribute in the business contract. The module realizes multi-dimensional understanding and classification of the contract clauses.

[0179] The multi-label classification module 50 includes a label system management unit, a classification processing unit, a weight adjustment unit, and a conflict detection unit. The label system management unit maintains a hierarchical label system; the classification processing unit performs multi-label classification tasks; the weight adjustment unit handles label imbalance problems; and the conflict detection unit is responsible for identifying and correcting label conflicts.

[0180] The semantic analysis module 60 is configured to extract logical associations and responsibility chains among contract clauses based on clause types and clause attributes by using a dependency syntax analysis technique, and construct a contract knowledge graph. The module deeply mines the internal logical structure of the contract, and provides a basis for risk identification.

[0181] The semantic analysis module 60 includes a dependency analysis unit, a responsibility chain extraction unit, an integrity evaluation unit, and a consistency checking unit. The dependency analysis unit constructs a semantic dependency tree; the responsibility chain extraction unit extracts a complete responsibility chain; the integrity evaluation unit evaluates the integrity of the responsibility chain; and the consistency checking unit checks the logical consistency among clauses.

[0182] The risk identification module 70 is configured to identify risk clauses in a business contract according to the contract knowledge graph, and compare the risk clauses with standard clauses in a key clause information library of the business contract, to generate risk prompt content. The module is the core of risk assessment, and directly affects the quality of review.

[0183] The risk identification module 70 includes a risk clause identification unit, a standard comparison unit, and a prompt generation unit. The risk clause identification unit identifies risk clauses based on various factors; the standard comparison unit compares the risk clauses with the standard clauses in the information library; and the prompt generation unit generates detailed risk prompt content.

[0184] The risk quantification module 80 is configured to extract business indicators, including transaction amount, payment period, and payment method, from the business contract, and calculate a risk exposure value based on the indicators and the risk clauses. The module converts risk assessment into quantifiable indicators, and supports objective decision-making.

[0185] The risk quantification module 80 includes an indicator extraction unit, a weight setting unit, and a calculation processing unit. The indicator extraction unit extracts relevant business indicators from the contract; the weight setting unit determines the weights of the indicators; and the calculation processing unit calculates the risk exposure value according to a formula.

[0186] The report generation module 90 is configured to generate a contract review report containing risk clauses, risk prompt content, and risk exposure values, and push the report to a reviewer. The module is the output end of the entire review process, and directly faces the end user.

[0187] The report generation module 90 includes a content organization unit, a visualization processing unit, and a push management unit. The content organization unit organizes report content and sorts the content according to risk levels; the visualization processing unit generates risk analysis charts; and the push management unit is responsible for sending the report to relevant reviewers.

[0188] To better understand the embodiments of the present application, a specific application example is described below.

[0189] Suppose there is a sales contract containing the following clauses:

[0190] "Article 5: Party A shall pay 70% of the total contract amount to Party B within 30 working days after receiving and passing the quality inspection of Party B's goods."

[0191] The system of the present application processes this clause as follows:

[0192] 1. In the text preprocessing stage, the system identifies this clause as an independent text unit and performs normalization processing.

[0193] 2. In the semantic representation extraction stage, the text unit is processed by a Bi-LSTM-CRF model to extract semantic representations. Forward processing captures the semantic flow of "Party A shall receive... ", and backward processing captures the semantic flow of "... pay 70% of the total contract amount". The two are integrated to form a complete semantic representation.

[0194] 3. In the attention processing stage, the word-level attention mechanism identifies "30 working days" and "70%" as key information points; the sentence-level attention identifies "should be within..." as a responsibility indication structure; and the clause-level attention identifies this clause as a payment clause and assesses its risk level.

[0195] 4. In the multi-label classification stage, the system classifies this clause as "payment clause" (clause type), identifies "Party A" as the responsible subject, "30 working days" as the time constraint, and "pay 70%" as the responsibility content (clause attribute).

[0196] 5. In the semantic analysis stage, the system constructs a semantic dependency tree for this clause and extracts the responsibility chain: subject (Party A) → action (payment) → object (70% of the contract amount) → condition (receipt and quality inspection) → time limit (within 30 working days). The analysis shows that the responsibility chain is complete and has no missing elements.

[0197] 6. In the risk identification stage, the system compares this clause with standard clauses and finds two risk points: the payment period (30 working days) is significantly longer than the industry standard (7 working days); and the one-time payment proportion (70%) is higher than the industry safety level (50%). Based on this, a risk prompt is generated: "The payment period is too long and the one-time payment proportion is too high, it is recommended to adjust to phased payment or shorten the payment period."

[0198] 7. In the risk quantification stage, assuming that the total contract transaction amount A is 1,000,000 yuan, the liquidated damages Li of this risk clause is 5% of the total amount, the payment method P is 30-day commercial bill payment (P=1, weight 1.5), and the payment period T is 30 working days (coefficient 1.5). According to the formula K=A×(1+ΣLi)×P×T, the risk exposure value K is calculated as K=1,000,000×(1+0.05)×1.5×1.5=2,362,500 yuan.

[0199] 8. Report generation phase, the system generates a review report containing the above risk terms, risk warnings and risk exposure values, and evaluates the contract risk level as "high risk" (because the K value exceeds the preset threshold of 2,000,000 yuan), and pushes it to the reviewer for final decision.

[0200] As can be seen from the above examples, the present application can automatically identify the risk terms in the contract and provide quantitative risk assessment, greatly improving the efficiency and accuracy of contract review.

[0201] The technical solutions of the present application are not limited to the above embodiments, and can also have other variations and improvements. For example:

[0202] 1. Model training, a transfer learning strategy can be used to pre-train the model on general legal texts and fine-tune it on specific types of business contracts to improve the generalization ability of the model.

[0203] 2. Risk assessment, risk assessment standards and weight settings can be customized according to different industries and enterprise characteristics to enhance the adaptability of the system.

[0204] 3. User interaction, an interactive review interface can be designed to allow reviewers to adjust risk parameters and view the impact of different decisions, improving decision transparency.

[0205] 4. Data security, data desensitization and encryption techniques can be used to ensure the security of contract information and meet the requirements of enterprises for data protection.

[0206] The above description is only for the preferred embodiments of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modifications, equivalent replacements, improvements, etc. within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. Intelligent review and risk quantification method for key terms of business contracts, characterized by: include: Establish and maintain a business contract key terms information database, which includes standard terms templates and risk assessment benchmarks; Obtaining a business contract text to be reviewed, and preprocessing the business contract text to obtain a structured text sequence; Processing the structured text sequence based on a bidirectional long short-term memory network and a conditional random field model to extract semantic representations of contract terms, wherein the semantic dependencies between the context of the terms are captured through bidirectional feature fusion; Based on the semantic representation, key information in the contract text is identified through a multi-level attention mechanism, wherein the multi-level attention mechanism includes word-level legal term enhancement attention, sentence-level semantic key point identification attention, and clause-level risk assessment attention; Based on the key information, multi-label learning classification is performed to identify the clause types and clause attributes in the business contract, wherein the clause types include business liability clauses, breach of contract clauses, and business indicator clauses, and the clause attributes include responsible parties, responsible objects, liability limitations, and time constraints; Using dependency parsing technology, based on the clause types and clause attributes, the logical associations and responsibility chains between contract clauses are extracted to construct a contract knowledge graph. Identify risk clauses in the business contract based on the contract knowledge graph, compare them with standard clauses in the business contract key clause information database, and generate risk warning content; extracting business indicators from the business contract, including transaction amount, payment period, and payment method, and calculating a risk exposure value based on the business indicators and the risk clauses; Generate a contract review report containing risk clauses, risk warnings and risk exposure values ​​and send it to the reviewer.

2. The intelligent review and risk quantification method for key terms of business contracts according to claim 1 is characterized by: The establishment and maintenance of the business contract key terms information database specifically includes: Obtain standard contract texts for multiple industries; Analyze the standard contract text to extract the content of standard clauses and the logical relationship between clauses; Based on the content of the standard clauses, a mapping relationship between the clause title, liability limitation information, and clause risk level is constructed; Establish risk assessment benchmarks for different types of clauses based on historical review data and legal expert knowledge; The information database of key terms of the business contracts is regularly updated to adapt to changes in laws, regulations and industry standards.

3. The intelligent review and risk quantification method for key terms of business contracts according to claim 1 is characterized by: The processing of the structured text sequence based on the bidirectional long short-term memory network and the conditional random field model specifically includes: Inputting the structured text sequence into a bidirectional long short-term memory network to obtain forward feature representation and backward feature representation of the text; Fusing the forward feature representation and the backward feature representation to generate a comprehensive feature representation; Based on the comprehensive feature representation, sequence labeling is performed using a conditional random field model to identify key elements in the contract text; For the key elements, their context information is extracted to form an enhanced semantic representation.

4. The intelligent review and risk quantification method for key terms of business contracts according to claim 1 is characterized in that: The multi-level attention mechanism specifically includes: Based on a predefined legal terminology dictionary, professional terms in the contract text are identified and assigned higher attention weights; Analyze sentence structure, identify key semantic nodes representing conditions, responsibilities, and consequences, and dynamically adjust attention allocation to these nodes; Assess the risk level of the clauses and assign higher attention weight to high-risk clauses; Build a reference relationship map between terms, identify explicit references and implicit associations, and ensure that related terms receive sufficient attention.

5. The intelligent review and risk quantification method for key terms of business contracts according to claim 1 is characterized in that: The multi-label learning classification specifically includes: Based on a hierarchical labeling system, the clauses are classified into multiple dimensions; Use the label dependency model to capture the logical associations between different labels; To address the label imbalance problem in legal texts, we assign higher weights to rare but important label categories; Perform label conflict detection to identify and correct logical contradictions in the predicted label set based on legal logic rules.

6. The intelligent review and risk quantification method for key terms of business contracts according to claim 1 is characterized in that: The logical relationship and responsibility chain between the extraction contract clauses specifically include: Construct a semantic dependency tree for each clause to identify the semantic structure within the clause; Extract the complete responsibility chain, including responsible parties, responsible behaviors, responsible objects, liability limitations and consequences of breach of contract; Analyze the completeness of clauses and assess the missing elements of the responsibility chain; Identify logical consistency between clauses and detect responsibility conflicts, timing conflicts, and scope conflicts.

7. The intelligent review and risk quantification method for key terms of business contracts according to claim 1 is characterized in that: The calculation of the risk exposure value specifically includes: Get the total contract transaction amount A; Identify the risk clauses in the contract and extract the liquidated damages in each risk clause; Determine the payment method P for the contract, where P can be 1, 2, 3, or 4, corresponding to 30 / 60 / 90-day commercial bill payment, 3-day wire transfer payment, 7-day commercial bill payment, and 8-15-day commercial acceptance bill payment, respectively; Get the payment period T of the contract; Calculate the risk exposure value K based on the total transaction amount A, liquidated damages Li, payment method P and payment period T.

8. The intelligent review and risk quantification method for key terms of business contracts according to claim 7 is characterized in that: The calculation formula of the risk exposure value K is: K=A*(1+ΣLi)*P*T Among them, A is the total contract transaction amount, Li is the penalty for breach of contract of the i-th risk clause, P is the risk weight of the payment method, and T is the risk coefficient of the payment period.

9. The intelligent review and risk quantification method for key terms of business contracts according to claim 1 is characterized in that: The generation of a contract review report containing risk clauses, risk warnings and risk exposure values ​​specifically includes: Rank the identified risk clauses by risk level; For each risk item, generate a detailed description including risk description, risk cause and improvement suggestions; Convert risk exposure values ​​into risk levels, which are classified into low risk, medium risk, and high risk; Generate visual risk analysis charts to intuitively display risk distribution; Integrate all risk information into a structured review report.

10. Intelligent review and risk quantification device for key terms of business contracts, characterized by: include: An information base management module is used to establish and maintain an information base of key terms of business contracts, which includes standard terms templates and risk assessment benchmarks; A text preprocessing module is used to obtain the business contract text to be reviewed, preprocess the business contract text, and obtain a structured text sequence; a semantic representation extraction module for processing the structured text sequence based on a bidirectional long short-term memory network and a conditional random field model to extract the semantic representation of the contract terms, wherein the semantic dependencies between the context of the terms are captured through bidirectional feature fusion; an attention processing module, configured to identify key information in the contract text based on the semantic representation through a multi-level attention mechanism, wherein the multi-level attention mechanism includes word-level legal term enhancement attention, sentence-level semantic key point identification attention, and clause-level risk assessment attention; a multi-label classification module for performing multi-label learning classification based on the key information to identify clause types and clause attributes in the business contract, wherein the clause types include business liability clauses, breach of contract clauses, and business indicator clauses, and the clause attributes include responsible parties, responsible objects, liability limitations, and time constraints; A semantic analysis module is used to extract the logical associations and responsibility chains between contract clauses based on the clause types and clause attributes using dependency parsing technology, and to construct a contract knowledge graph; A risk identification module, configured to identify risk clauses in a business contract based on the contract knowledge graph, compare the risk clauses with the standard clauses in the business contract key clause information database, and generate risk warning content; a risk quantification module, configured to extract business indicators from the business contract, including transaction amount, payment period, and payment method, and calculate a risk exposure value based on the business indicators and the risk clauses; The report generation module is used to generate a contract review report containing risk clauses, risk warning content and risk exposure values, and push it to the reviewer.

Images