A method and system for risk review of contract terms

Abstract

The present application relates to the technical field of natural language processing, in particular to a contract clause risk review method and system, comprising the following steps: obtaining a to-be-reviewed contract text, deconstructing the sentences of each contract clause to obtain a clause analysis feature set. The present application extracts the feature details of the syntactic tree depth, the reference distance, the logical nesting layer number and the professional term density by disassembling the contract clauses into basic semantic units sentence by sentence, uses the four types of index combinations in the clause analysis feature set to realize the measurement of the single clause analysis complexity score, and further aggregates the clause analysis complexity score to form specific indexes capable of quantifying the overall understandability risk of the contract. Meanwhile, the to-be-reviewed contract and the standard contract text library are executed for theme clustering and semantic weight distribution to clearly understand the actual focus of each theme, and the themes with larger semantic deviation are marked as missing risks.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of natural language processing, and particularly relates to a contract clause risk review method and system. BACKGROUND

[0002] The contract clause risk review method uses natural language processing technology to analyze the syntactic structure, reference relationship, logical nesting complexity and theme distribution characteristics of the contract clauses, quantitatively evaluates and identifies potential risks such as content missing and semantic ambiguity in the contract, so as to enable enterprises or individuals to quickly and accurately find hidden risks in the contract.

[0003] The prior art easily covers up the mutual influence between core risk indicators, and thus cannot reflect the individual differences of single clauses or special risk clauses, the importance of clauses is difficult to be clearly distinguished, and the phenomenon of high-risk clauses being ignored or risks being averaged may occur, especially when the semantic structure of the contract clauses is complex, the professional terms are dense or the reference relationship is complicated, which makes the overall risk determination tend to be subjective or fuzzy, and it is difficult to support the efficiency and accuracy of risk management decisions. Therefore, improvement is needed. SUMMARY

[0004] The purpose of the present application is to solve the problems in the prior art and provide a contract clause risk review method and system.

[0005] In order to achieve the above purpose, the present application adopts the following technical scheme, a contract clause risk review method, comprising the following steps:

[0006] Obtain the contract text to be reviewed, deconstruct the sentences of each contract clause to obtain a clause analysis feature set;

[0007] Based on the clause analysis feature set, a preset weight is assigned to the syntactic tree depth, reference distance, logical nesting layer number and professional term density of each recorded clause to generate a single clause analysis complexity score of each clause, and the single clause analysis complexity scores of all clauses in the contract are calculated to obtain a contract understandability risk index;

[0008] Based on the contract text to be reviewed and the standard contract text library of the same type, the word co-occurrence frequency of the contract text to be reviewed and the standard contract text library of the same type is calculated and clustered into themes to generate an actual theme distribution of the contract to be reviewed and a standard theme distribution, and the theme weights in the actual theme distribution of the contract to be reviewed are compared with the corresponding items in the standard theme distribution one by one to obtain a contract theme weight deviation list;

[0009] Filter the themes in the contract theme weight deviation list that exceed the preset threshold to determine the missing risk, determine the ambiguity risk when the contract understandability risk index exceeds the threshold, and generate a risk result.

[0010] Preferably, the clause parsing feature set acquisition step is:

[0011] Acquire the text of the contract under review, process each clause of the contract under review, split each clause into independent sentences, label each syntactic component in each sentence, analyze the direct dependency relationship between each syntactic component, determine the syntactic tree structure of each sentence according to the dependency relationship link, count the number of nodes in the longest path from the root node to the leaf node of the syntactic tree, and generate a syntactic tree depth value;

[0012] Based on the syntactic tree depth value, traverse the word sequence in each contract clause sentence by sentence, identify all pronouns and referents in the sentence, associate and pair the pronouns with the corresponding referents, record the actual lexical distance between each pair of associated pairs, take the number of intermediate lexical words in the pair as the lexical distance value, and form a set of lexical distance values of all pronoun-referent pairs in the clause;

[0013] Based on the set of lexical distance values, identify conditional logical operators and temporal logical operators in the contract clause sentence by sentence, expand and recursively each conditional logical operator and temporal logical operator identified respectively, count the number of nested layers of each operator in the sentence, and simultaneously count the frequency of each professional term in the preset list in the contract clause sentence. The syntactic tree depth value, the set of lexical distance values, the number of nested layers of the operator, and the frequency of the professional term are used as the clause feature dimensions to generate a clause parsing feature set.

[0014] Preferably, the single clause parsing complexity score acquisition step is:

[0015] Based on the clause parsing feature set, extract the syntactic tree depth value, the referential distance value, the number of logical nesting layers, and the professional term density value of each contract clause, respectively, and perform maximum and minimum difference scaling processing. The original values are mapped to the closed interval 0 to 1 in turn and recorded as standardized syntactic tree depth values, standardized referential distance values, standardized number of logical nesting layers, and standardized professional term density values.

[0016] According to the standardized syntactic tree depth value, the standardized referential distance value, the standardized number of logical nesting layers, and the standardized professional term density value, calculate the single clause parsing complexity value of each contract clause.

[0017] Preferably, the contract understandability risk index acquisition step is:

[0018] Based on the single clause parsing complexity value corresponding to all clauses, calculate the contract understandability risk index.

[0019] Preferably, the obtaining step of the actual theme distribution of the to-be-reviewed contract and the standard theme distribution is:

[0020] The to-be-reviewed contract text and the standard contract text library are respectively subjected to paragraph segmentation, each paragraph is subjected to word-by-word segmentation processing, and a pre-trained semantic model is called to convert each word into a word vector, then the word vectors of all word units in each paragraph are subjected to weighted average to generate a paragraph vector, and a fixed theme keyword group is set according to the business type, and a corresponding theme center vector is generated by averaging the keyword word vectors of each group to obtain a paragraph vector set and a theme center vector set.

[0021] According to the paragraph vector set and the theme center vector set, the semantic weight value of each theme in the to-be-reviewed contract is calculated.

[0022] Based on the semantic weight value of each theme, the semantic weight value calculation is respectively completed for the to-be-reviewed contract and the standard contract text library, the corresponding weight vector is output in ascending order of theme number, and the actual theme distribution of the to-be-reviewed contract and the standard theme distribution are respectively constructed.

[0023] Preferably, the obtaining step of the contract theme weight deviation list is:

[0024] Based on the actual theme distribution of the to-be-reviewed contract and the standard theme distribution, the weight values corresponding to the same themes in the theme distributions are compared one by one, the difference absolute values between the weight values of each theme in the actual theme distribution of the to-be-reviewed contract and the weight values in the standard theme distribution are calculated, the difference absolute values of each theme are arranged in descending order, and a contract theme weight deviation list is formed.

[0025] Preferably, the obtaining step of the risk result is:

[0026] Based on the contract theme weight deviation list, a preset threshold is set, each theme in the contract theme weight deviation list is traversed, if the weight difference absolute value corresponding to a theme is greater than the preset threshold, the theme is marked as a missing risk theme, and all marked missing risk themes and weight difference absolute values are extracted to form a missing risk theme list.

[0027] Preferably, the obtaining step of the risk result further includes: based on the contract intelligibility risk index, setting an ambiguity risk threshold, judging the value size relationship between the contract intelligibility risk index and the ambiguity risk threshold, if the value of the contract intelligibility risk index is greater than the ambiguity risk threshold, it is determined that there is an ambiguity risk, and the missing risk theme list is combined to form a risk result.

[0028] The application also provides a risk review system, comprising:

[0029] A contract analysis module acquires a to-be-reviewed contract text, disassembles sentences of each contract clause to obtain a clause analysis feature set;

[0030] A complexity evaluation module assigns preset weights to a syntax tree depth, a reference distance, a logical nesting layer number and a professional term density of each clause recorded based on the clause analysis feature set, generates a single-clause analysis complexity score of each clause, and calculates a contract comprehensibility risk index by aggregating the single-clause analysis complexity scores of all clauses in the contract.

[0031] A theme deviation analysis module calculates a word co-occurrence frequency of a to-be-reviewed contract text and a same-type standard contract text library based on the to-be-reviewed contract text and the same-type standard contract text library, clusters themes into a theme distribution, generates an actual theme distribution of the to-be-reviewed contract and a standard theme distribution, and compares each theme weight in the actual theme distribution of the to-be-reviewed contract with a corresponding item in the standard theme distribution to obtain a contract theme weight deviation list.

[0032] A risk identification module screens themes in the contract theme weight deviation list that exceed a preset threshold, determines that the themes are missing risks, determines that the contract comprehensibility risk index exceeds a threshold, and determines that the contract comprehensibility risk index is an ambiguity risk, and generates a risk result.

[0033] Compared with the prior art, the present application has the following advantages and positive effects:

[0034] The present application disassembles contract clauses into basic semantic units sentence by sentence, extracts features such as syntax tree depth, reference distance, logical nesting layer number and professional term density, uses four types of indicators in the clause analysis feature set to measure a single-clause analysis complexity score, further aggregates the clause analysis complexity scores to form a specific indicator that can quantify the overall comprehensibility risk of a contract, performs theme clustering and semantic weight distribution on a to-be-reviewed contract and a standard contract text library, clearly defines the actual focus of each theme, and targets missing risks by positioning themes with greater semantic deviation. Compared with the prior art that only identifies content missing or semantic ambiguity, the present application can more comprehensively capture potential risks hidden in subtle text structures and theme differences, reduce risk omissions and misjudgments in the contract review process, improve the accuracy and objectivity of contract risk identification, reduce the probability of disputes caused by ambiguity and omissions in the contract execution process, and improve the overall credibility of contract clause risk assessment. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 The present application is a step schematic diagram. DETAILED DESCRIPTION

[0036] In order to make the purpose, technical scheme and advantages of the present application more clear, the present application is further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application.

[0037] Referring to Figure 1 The present application provides a technical scheme, a contract clause risk review method, comprising the following steps:

[0038] Obtaining the text of the contract to be reviewed, deconstructing the sentences of each contract clause to obtain a clause analysis feature set;

[0039] Based on the clause analysis feature set, assigning a preset weight to the syntax tree depth, the reference distance, the number of logical nesting layers and the density of professional terms of each recorded clause, generating a single clause analysis complexity score for each clause, and calculating the single clause analysis complexity scores of all clauses in the contract to obtain a contract intelligibility risk index;

[0040] Based on the text of the contract to be reviewed and the standard contract text library of the same type, calculating the word co-occurrence frequency of the text of the contract to be reviewed and the standard contract text library of the same type and clustering into topics to generate the actual topic distribution of the contract to be reviewed and the standard topic distribution, comparing the weight of each topic in the actual topic distribution of the contract to be reviewed with the corresponding item in the standard topic distribution one by one, and obtaining a contract topic weight deviation list;

[0041] Screening topics with a topic weight deviation list exceeding a preset threshold, determining as missing risk, determining a contract intelligibility risk index exceeding a threshold as ambiguity risk, and generating a risk result.

[0042] The step of obtaining the clause analysis feature set is:

[0043] Obtaining the text of the contract to be reviewed, processing the text of the contract to be reviewed sentence by sentence, splitting each contract clause into independent sentences, marking each syntactic component in each sentence, and analyzing the direct dependency relationship between each syntactic component, determining the syntactic tree structure of each sentence according to the dependency relationship link, counting the number of nodes in the longest path from the root node to the leaf node of the syntactic tree, and generating a syntactic tree depth value;

[0044] Based on the syntactic tree depth value, traversing the word sequence in each contract clause sentence by sentence, identifying all referring words and referring entities in the sentence, associating and pairing the referring words with the corresponding referring entities, recording the actual word distance between each pair of associated pairs, taking the number of intermediate interval words in the pair as the word distance value, and forming a set of word distance values of all referring words-referring entities in the clause;

[0045] Based on the set of lexical distance numerical values, the conditional logical operators and time logical operators are identified for each sentence of the contract clause, the identified conditional logical operators and time logical operators are expanded one by one and recursively, the number of nested times of each operator in the sentence is counted, and the frequency of each professional term in the preset list in the contract clause sentence is counted, and the syntax tree depth value, the set of lexical distance numerical values, the number of nested times of the operator, and the frequency of the professional term are taken as the clause characteristic dimensions to generate the clause analysis characteristic set.

[0046] Specifically, after obtaining the to-be-reviewed contract text, first, according to the chapter clause number of the contract, such as "first", "1.1", "(one)" structured identifier, the complete contract text is accurately divided into independent contract clauses, then, in each independent contract clause, a rule-based sentence splitter is called, which uses full angle period ".", question mark "?", exclamation mark "!" as the main split symbol, splits the clause text into independent sentences, at the same time, in order to avoid errors in the division of the commonly used numbered list (such as "1.", "(a)") or the amount of money with decimal point in legal text, the splitter has built-in ignore rules, only when the above punctuation symbols are not followed by numbers or letters, the division operation is performed, after completing the sentence-by-sentence splitting, the system processes each independent sentence using a Chinese dependency syntax analysis model based on graph attention network, which has been trained on a corpus of more than 100,000 legal documents, and can accurately identify language structures specific to contract text, during analysis, the model first performs word segmentation on the sentence and labels the part of speech of each word, such as noun, verb or adjective, then the model calculates the dependency relationship probability between words to build a weighted directed acyclic graph to represent the syntax structure, where each word is a node and the dependency relationship is a directed edge, for example, in the sentence "the party B shall pay the full amount within five working days", the model will identify "payment" as the core verb, i.e. the root node of the syntax tree, and establish dependency links from "payment" to the subject "party B", the adverb "within five working days" and the object "amount", and from "amount" to the attribute "full amount", after the syntax tree structure is determined, the system uses a depth-first search algorithm to traverse all paths from the root node to the leaf node (i.e. the node with no out-degree), and calculates the number of nodes contained in each path, finally selects the path with the most nodes, and the length of the path is recorded, the number of nodes in this longest path is the syntax tree depth of the sentence, this process will be applied to each sentence in the contract to generate the syntax tree depth value.

[0047] After the generation step of the syntax tree depth value is completed, the system continues to traverse the lexical sequence in each contract clause at the level of each sentence, identifies and associates the pronouns in the sentence and the entities they point to. The model uses the encoder-decoder architecture of the Transformer and uses the BERT model pre-trained on a large amount of Chinese legal text as its underlying word vector representation layer. The model structure is specifically divided into an input layer, a span representation layer, and a scoring layer. When processing, the input of the model is a complete contract clause text. The model first tokenizes the text through the input layer and uses BERT to generate context-related word vectors for each word. Then, the span representation layer identifies all possible noun phrases and pronouns in the text as candidate reference items and generates a fixed-dimensional vector representation for each candidate span through a self-attention mechanism. Finally, in the scoring layer, the model calculates the association score between each pair of candidate reference items (a pronoun and a noun phrase appearing before it). The score is calculated by a feedforward neural network, and its input is the vector representation of the two candidate spans and the vector encoding of the distance between them. The model is trained through supervised learning on a legal contract corpus containing tens of thousands of annotated reference relationships. The training goal is to optimize the cross-entropy loss function so that the model outputs the highest association score for the correct reference relationship pair. For example, for the sentence "Party A must provide necessary technical documents, and the content should be true and valid", after reasoning, it will output an association pair that pairs the pronoun "it" with the entity "technical documents" it refers to. After completing the association of all pronouns and reference entities, the system will record the actual lexical distance between each successful association. The distance is calculated by counting the number of words between the reference entity and the pronoun. In the aforementioned example, "technical documents" and "it" are separated by two words, "and" and "and", so the lexical distance value of this pair is 2. This process is applied to all reference relationships identified in the clause to form a set of lexical distance values for all pronoun-reference entity pairs in the clause.

[0048] Based on the set of lexical distance values, the system then initiates a quantitative analysis of the logical structure and professional terms of each sentence of the contract clause. First, in order to identify and quantify the logical complexity, the system uses a method based on regular expressions and keyword matching to scan each sentence to find predefined conditional logical operators and temporal logical operators. The list of conditional logical operators includes patterns such as "if…then", "if…then", "unless", "provided that", etc., while the list of temporal logical operators contains time limiters such as "before", "after", "… within", "from…", etc. After identifying these operators, the system uses a stack-based data structure to calculate their nesting level. Specifically, when the analyzer encounters the starting part of a logical operator (such as "if"), it is pushed onto the stack. If another logical operator is encountered before the scope of the current operator ends (such as "unless"), it is also pushed onto the stack. At this time, the depth of the stack represents the nesting level of the latter. For example, in the sentence "if Party B breaches the contract, unless force majeure occurs, Party A has the right to claim compensation", the nesting level of "unless" is 1. The system records the maximum nesting level of each sentence. At the same time, the system counts the frequency of occurrence of professional terms in the contract clause sentence. This process relies on a pre-set list dynamically loaded according to the type of contract, which is developed by legal experts and business experts in the field. For example, for a "software procurement contract", the list contains about 150 professional terms such as "source code hosting", "acceptance test", "intellectual property", "force majeure", "confidentiality obligation", etc. The generation of this list is based on the analysis of more than 300 standard contracts of the same type, by extracting high-frequency noun phrases and screening through expert review. The system traverses the clause content word by word, accurately matches the terms in the list, and counts the total number of occurrences. Finally, the syntax tree depth value of each sentence, the set of lexical distance values of each clause, the maximum operator nesting number identified in each sentence, and the frequency of occurrence of professional terms counted in each clause are used as the original feature dimensions to describe the complexity of the clause, generating a clause analysis feature set.

[0049] The steps for obtaining the single-clause analysis complexity score are:

[0050] Based on the clause analysis feature set, the syntax tree depth value, the reference distance value, the logical nesting level value, and the professional term density value of each contract clause are extracted, and the maximum and minimum difference scaling processing is performed, respectively, to map the original values to the closed interval 0 to 1, and record them as standardized syntax tree depth value, standardized reference distance value, standardized logical nesting level value, and standardized professional term density value.

[0051] Based on the standardized syntactic tree depth, standardized reference distance, standardized logical nesting level, and standardized terminology density, the single-clause parsing complexity of each contract clause is calculated using the following formula:

[0052] ;

[0053] in, Indicates the first The numerical value of parsing a single clause of a contract. For the first The standardized syntax tree depth value in the text. For the first In this context, "standardization" refers to the numerical value of the distance. For the first The standardized logical nesting level value in the clause, For the first The density values ​​of standardized professional terms in the article, These are the preset risk weights for the four types of indicators mentioned above. For logical-referential collaborative risk coefficient, A constant that approaches zero is used to avoid the product under the square root being zero.

[0054] Specifically, based on the clause analysis feature set, the system first extracts four original risk indicators for each contract clause from the set, namely the syntactic tree depth value, the anaphora distance value, the logical nesting layer number value, and the professional term density value. The syntactic tree depth value is directly taken from the feature set, the anaphora distance value is obtained by calculating the average of the lexical distance between all identified anaphors and corresponding entities in the clause, the logical nesting layer number value is the maximum logical nesting depth identified in all sentences in the clause, and the professional term density value is calculated by dividing the total frequency of identified professional terms in the clause by the total token number of the clause. After extracting the four original values for all clauses, in order to eliminate the differences in dimensions and value ranges between different indicators, the system performs maximum and minimum difference scaling processing for each type of indicator within the scope of the entire contract. Specifically, taking the syntactic tree depth value as an example, the system first traverses all clauses in the current contract under review to find the maximum and minimum values of the syntactic tree depth value. For example, in a contract containing 50 clauses, the minimum value of the syntactic tree depth is 2 and the maximum value is 18. Then, the system traverses each clause again and applies a linear transformation to the specific syntactic tree depth value of the clause. The transformation is achieved by subtracting the global minimum value from the current clause value and dividing the resulting difference by the difference between the global maximum and minimum values. For example, for a clause with a syntactic tree depth of 10, the standardization calculation process is (10-2) divided by (18-2), resulting in 8 divided by 16, which is 0.5. This process is also applied to the anaphora distance value, the logical nesting layer number value, and the professional term density value. For each type of indicator, the maximum and minimum values within the entire contract are found, and the same scaling calculation is performed on the corresponding indicator value of each clause. All original values are mapped to the closed interval 0 to 1 in turn, and are recorded as standardized syntactic tree depth value, standardized anaphora distance value, standardized logical nesting layer number value, and standardized professional term density value, respectively.

[0055] Formula: The benefit of the formula is that an interaction risk term is introduced, which is specifically used to quantify the synergistic amplification effect between logical structure complexity and anaphora ambiguity. In legal texts, when an ambiguous anaphor appears in a sentence with multiple nested logic, its ambiguity risk is much greater than the sum of the two independent risks. By calculating the geometric mean of the standardized logical nesting layer number and the standardized anaphora distance, and scaling by the logic-anaphora synergistic risk coefficient , this non-linear compound risk can be captured, allowing for a more realistic judgment of the potential understanding difficulty of the clause;

[0056] The steps for obtaining the risk weights are as follows: The pre-defined risk weights are determined using the Analytic Hierarchy Process (AHP) combined with expert scoring. First, an expert group of 15 legal experts and senior business managers with over 10 years of contract review experience is formed. For four risk dimensions—syntactic tree depth (D), reference distance (L), logical nesting (I), and terminology density (T)—experts are required to compare each risk dimension pairwise to determine its relative importance to contract ambiguity risk, and score them using a 1-9 scale. For example, if an expert considers logical nesting "slightly more important" than reference distance, the score is 3; otherwise, it's 1 / 3. After summing all expert scores, an independent judgment matrix is ​​constructed for each contract type (e.g., equipment procurement, software licensing, engineering construction). The maximum eigenvalue and its corresponding eigenvector of the matrix are calculated. After passing a consistency test (valid when the consistency ratio is less than 0.1), the normalized eigenvector is used as the risk weight for each indicator. For example, for a "Software License Agreement," the calculated weights might be: =0.15 (syntactic structure) =0.20 (referential relationship) =0.35 (logical nesting) =0.30 (technical term);

[0057] The steps to obtain it are: logical-referential collaborative risk coefficient. The numerical values ​​are determined through regression analysis of historical case data. First, a database containing 500 contract texts and their judgments that have historically led to legal disputes due to ambiguous clauses is collected. For each contract, the preliminary steps of this method are used to calculate the numerical values ​​of all clauses. , , , Numerical values, and record whether the clauses that cause disputes also have a high degree of logical nesting (e.g. >0.7) and long referential distance (e.g. Based on the characteristic of >0.7), a logistic regression model is constructed, with the dependent variable being "whether a dispute occurred" (yes=1, no=0), and the independent variables including independent terms and interaction terms of the four standardized indicators. By training the model on the dataset, the regression coefficients of the interaction terms are obtained, and these coefficients are used as... This method ensures the value of [the value]. The numerical value reflects the actual probability of this collaborative risk leading to negative outcomes in the real world. This was obtained through analysis of a database containing software and technology service contracts. The value is 0.65;

[0058] The acquisition step is that the parameter is a very small positive number set to prevent the operation of zero or negative number when calculating the square root, set to ;

[0059] The calculation process is:

[0060] According to the standardized syntax tree depth value, the standardized reference distance value, the standardized logical nesting layer value and the standardized professional term density value, and the parameters determined by the above method, the single clause analysis complexity value of the i-th contract clause is calculated as follows: The single clause analysis complexity value of the i-th contract clause Take the third clause (i.e. ) in a software license agreement as an example, the four standardized values obtained from the above step are as follows:

[0061] The standardized syntax tree depth value =0.50;

[0062] The standardized reference distance value =0.75;

[0063] The standardized logical nesting layer value =0.80;

[0064] The standardized professional term density value =0.60;

[0065] Substitute the determined weights and coefficients:

[0066] =0.15, =0.20, =0.35, =0.30;

[0067] =0.65;

[0068] = ;

[0069] The calculation process is as follows:

[0070] ;

[0071] ;

[0072] ;

[0073] ;

[0074] ;

[0075] ;

[0076] The result shows that the single clause resolution complexity value of the 3rd contract clause is 1.1885, which is a comprehensive quantitative indicator. A value greater than 1.0 usually means that the clause has a higher understanding difficulty and ambiguity risk, because it not only scores high on each basic risk indicator (the weighted sum is 0.685), but also has a synergistic risk due to its complex logical structure and ambiguous reference relationship (the synergistic risk item is 0.5035).

[0077] The steps for obtaining the contract understandability risk index are:

[0078] Based on the single clause resolution complexity values of all clauses, the contract understandability risk index is calculated, and the calculation formula is:

[0079]

[0080] Among them, represents the contract understandability risk index, represents the total number of clauses in the contract, is the single clause resolution complexity value of the th clause, represents the average value of all clause resolution complexity values, is the risk dispersion sensitivity coefficient, which is used to adjust the influence degree of the complexity distribution of the clauses on the overall risk.

[0081] Specifically, the formula is: The beneficial part of the formula is that the first part (i.e. ) reflects the average understanding difficulty of the contract, and the second part quantifies the risk brought by the dispersion degree of the complexity distribution among the clauses. This design is based on an important insight: a contract with extremely uneven clause complexity (e.g., most clauses are simple, but a few core clauses are exceptionally complex) is more risky than a contract where all clauses are at a moderate level of complexity, because the reviewer is likely to become fatigued when reviewing a large number of simple clauses, thereby ignoring a few “deeply buried” complex clauses. By introducing the standard deviation term adjusted by the risk dispersion sensitivity coefficient ;

[0082] The steps for obtaining the risk dispersion sensitivity coefficient The value of is set according to the importance of the contract and the risk tolerance strategy, it is not calculated, but determined by the risk management rules made by the legal and risk control departments of the enterprise in advance, which divides the contract into different levels according to its business impact, involved amount, severity of legal consequences and other factors, and specifies a value for each level, for example, the risk management policy of a company may stipulate that for "high risk" level contracts (such as merger and acquisition agreements, large financing agreements), 2.0 is set to amplify the impact of inconsistent clause complexity; for "medium risk" level contracts (such as important supplier contracts, technology licensing agreements), 1.2 is set; for "low risk" level contracts (such as standard confidentiality agreements, short-term lease contracts), 0.5 is set, when reviewing a new contract, the system first requires the user to select the risk level corresponding to the contract, and then calls the corresponding value from the rule library for calculation, for the software license agreement in this example, it is classified as a "medium risk" level, so 1.2 is taken as the value;

[0083] The acquisition step of is to count the total number of clauses in the contract In the first step of the method, that is, to obtain the text of the contract to be reviewed and to deconstruct it, the total number of clauses in the contract is automatically counted by identifying and counting the clause numbers (such as "Article 1", "1.1", "(1)" etc.) in the contract, which is a direct count value, for example, after initial processing, the system determines that the software license agreement currently being reviewed contains a total of 35 independent clauses, so ;

[0084] The acquisition step of is to acquire the single clause analysis complexity value of the th clause The single clause analysis complexity value of the th clause is the direct calculation output of the previous step, the system will calculate a value for each clause in the contract (from to ), thus forming a set of complexity values, and the average value of all clause analysis complexity values is calculated based on this set, which is calculated by adding the values of all 35 clauses and then dividing by the total number of clauses , i.e.

[0085] Calculation process:

[0086] ​​The single clause resolution complexity value corresponding to each clause calculated based on the previous steps , calculate the contract intelligibility risk index , in this example, the total number of contract clauses is known , risk dispersion sensitivity coefficient , by calculating all 35 clauses, the average value of all values is 0.75, and the sum of the squares of the differences between all clause values and the average value is calculated to be

[0087] Substitute the formula for calculation:

[0088]

[0089] ;

[0090] ;

[0091] ;

[0092] ;

[0093] ;

[0094] ;

[0095] The results show that the contract intelligibility risk index of this software license agreement is 1.1522, which reflects the average complexity of the contract (0.75) and the additional risk caused by the uneven complexity of the clauses (0.4022). The index will be compared with a preset ambiguity risk threshold (e.g. 0.9). Since 1.1522 is greater than 0.9, the contract will be determined to have a high ambiguity risk and trigger the corresponding warning, prompting the reviewer to conduct more detailed and in-depth manual review of the contract.

[0096] The steps for obtaining the actual subject distribution of the contract under review and the standard subject distribution are as follows:

[0097] The text of the contract under review and the standard contract text library are respectively divided into paragraphs, each paragraph is processed word by word, and a pre-trained semantic model is called to convert each word into a word vector. Then, the weighted average of all word vectors in each paragraph is calculated to generate a paragraph vector, and a fixed theme keyword group is set according to the business type, and the average of the keyword vectors in each group is calculated to generate a corresponding theme center vector, obtaining a paragraph vector set and a theme center vector set. ​

[0098] Based on the paragraph vector set and the topic center vector set, the semantic weight value of each topic in the contract under review is calculated using the following formula:

[0099] ;

[0100] in, For the contract under review, the first The semantic weight value of each topic, For the first The paragraph vector of each paragraph. For the first The theme center vector of each theme For the first The length of a paragraph (measured by the number of words). The total number of paragraphs. Total number of topics This is a similarity sensitivity index used to increase the weight significance of highly matched paragraphs;

[0101] Based on the semantic weight value of each topic, the semantic weight values ​​of the contracts to be reviewed and the standard contract text library are calculated respectively. The corresponding weight vectors are output in ascending order of topic number, and the actual topic distribution and standard topic distribution of the contracts to be reviewed are constructed respectively.

[0102] Specifically, for the contract text under review and the standard contract text library, the system first performs paragraph segmentation based on blank lines or explicit chapter titles (such as "Article 1" or "Article 1") in the text, dividing the continuous text stream into independent paragraph units. For each segmented paragraph, the system calls a Jieba word segmentation tool integrated with a legal professional vocabulary, which contains more than 50,000 legal and business terms, to ensure accurate recognition of professional compound words such as "intellectual property" and "force majeure" during the word segmentation phase. After word segmentation, the system applies a pre-trained Law-BERT model on a large Chinese legal document corpus. This model can convert each segmented word into a 768-dimensional context-dependent word vector. Subsequently, to extract a single vector representing the semantic meaning of the entire paragraph from the set of word vectors, a weighted average method based on inverse document frequency (IDF) is used. The IDF value of each word is pre-calculated based on its frequency of occurrence in a background corpus containing 500,000 contracts of various types. The IDF value of common words is lower, while the IDF value of key professional words is higher. The calculation process of the paragraph vector is as follows: multiply the word vector of each word in the paragraph by its corresponding IDF weight, then add all the weighted vectors element by element, and finally divide the result vector by the sum of all word IDF weights. At the same time, the system loads a list of fixed theme word groups predefined by legal experts based on the type of contract under review. For example, for a "software development contract," the list contains eight themes such as "intellectual property" (key words: source code, ownership, copyright, patent), "payment conditions" (key words: fee, payment, invoice, acceptance, milestone), and "confidentiality obligations" (key words: confidential information, disclosure, confidentiality period). The system uses the aforementioned Law-BERT model to obtain the word vectors of all key words under each theme and generates a 768-dimensional theme center vector representing the core semantic meaning of the theme by calculating the average of these vectors. Finally, the vectorization process is completed for all paragraphs and all themes, resulting in a set of paragraph vectors and a set of theme center vectors.

[0103] Formula: The benefits of the formula are as follows: first, the cosine similarity between the paragraph vector and the theme center vector captures the relevance of the paragraph and the theme at the semantic level, surpassing traditional keyword frequency statistics; second, the introduction of paragraph length as a weight makes the contribution of more detailed content paragraphs to the theme weight greater, which aligns with the intuition that longer paragraphs usually carry more important information; third, the introduction of the similarity sensitivity index effectively distinguishes core relevant paragraphs from marginal relevant paragraphs by nonlinear amplification of high similarity, making the calculated theme weight more prominent in highlighting the core issues of the contract.

[0104] The acquisition step of the paragraph vector of the first paragraph is The acquisition step of the paragraph vector of the first paragraph is The paragraph vector of the first paragraph is the direct output of the previous step, specifically, it is a single 768-dimensional vector obtained by converting each word unit to a 768-dimensional word vector through the Law-BERT model after segmenting the first paragraph, and then weighting and averaging the word vectors according to the pre-calculated IDF value. The vector represents the core meaning of the entire paragraph in the semantic space, for example, a paragraph discussing software delivery and acceptance will have a paragraph vector in the vector space close to the vector positions of "delivery", "acceptance criteria", "test", etc.

[0105] The acquisition step of the theme center vector of the first theme is The theme center vector of the first theme is also derived from the previous step, which is generated according to the pre-set theme phrase of a specific contract type. The system will find all the keywords associated with the first theme, for example, the "intellectual property" theme is associated with the keywords "source code", "ownership", "copyright", and "patent". Then, the Law-BERT model is called to obtain the 768-dimensional word vectors of these four keywords, and finally, a 768-dimensional theme center vector is obtained by calculating the average of the four vectors, which can represent the core semantics of the abstract theme "intellectual property".

[0106] The acquisition step of the paragraph length of the first paragraph is The paragraph length of the first paragraph is obtained by directly counting the total number of word units contained in the paragraph after segmenting the paragraph, for example, if a paragraph is segmented into 125 word units, the value of

[0107] The acquisition step of the total number of paragraphs is The total number of paragraphs is obtained by counting the total number of independent paragraph units generated after segmenting the entire contract text, for example, a contract is segmented into 68 paragraphs, and the value of

[0108] The acquisition step of the total number of themes is ​​​​​​​is determined by the fixed list of subject vectors corresponding to the selected business type, which is pre-defined by domain experts and contains all the core subjects that the contract of this type should have, for example, when reviewing "software development contract", the system loads a list containing eight subjects: "intellectual property", "payment conditions", "confidentiality obligations", "liability for breach of contract", "technical support", "project management", "acceptance criteria", "dispute resolution", therefore The value of is 8;

[0109] The acquisition step is that the similarity sensitivity index is an empirical parameter, whose purpose is to adjust the influence of cosine similarity on the maximum weight, and its value is determined by grid search experiments on a benchmark dataset containing 500 contracts manually annotated by legal experts for each subject weight, in the experiment, The value of ranges from 1.0 to 4.0 with a step of 0.2, and for each value, the system calculates the root mean square error (RMSE) between the subject distribution calculated by the system and the distribution annotated by experts, and finally selects the value that makes the RMSE minimum as the optimal parameter, research shows that when is set to 2.5, the system can best reproduce the judgment of experts, both amplifying the influence of core paragraphs and not completely ignoring the contribution of secondary paragraphs, so is set to 2.5;

[0110] Calculation process:

[0111] According to the paragraph vector set and the subject center vector set, the semantic weight value of the th subject in the contract under review is calculated Take the weight of the first subject "intellectual property" (i.e. ) as an example, assuming that the contract has paragraphs, the total number of subjects (subject 1: intellectual property, subject 2: payment conditions), and the following data is obtained from the previous steps:

[0112] Paragraph 1 ( ): length , cosine similarity with subject 1 is 0.9, and cosine similarity with subject 2 is 0.2;

[0113] Paragraph 2 ( ): length , cosine similarity with subject 1 is 0.3, and cosine similarity with subject 2 is 0.8;

[0114] Paragraph 3 ( ): length Cosine similarity to Topic 1 is 0.7, and cosine similarity to Topic 2 is 0.5;

[0115] First, calculate the numerator, i.e. the total affinity of all paragraphs to Topic 1:

[0116] ;

[0117] ;

[0118] ;

[0119] Next, calculate the denominator, i.e. the total affinity of all paragraphs to all topics, which requires first calculating the total affinity of all paragraphs to Topic 2:

[0120] ;

[0121] ;

[0122] ;

[0123] The denominator is the sum of the total affinities of the two topics:

[0124] ;

[0125] Finally, calculate :

[0126] ;

[0127] The result shows that the semantic weight value of the topic "intellectual property" in the pending contract is about 0.5093, meaning that about 50.93% of the semantic content in the contract is related to intellectual property. This value will become a component in the actual topic distribution vector of the pending contract, which will be used for subsequent comparison with the standard topic distribution.

[0128] Based on the semantic weight value of each topic, the system will perform a complete semantic weight value calculation process on the currently reviewed pending contract and a large standard contract text library. For the pending contract, the system calls the aforementioned calculation formula to calculate a corresponding semantic weight value for each topic (from Topic 1 to Topic M) in the preset list, for example, for a software development contract containing eight topics, the system will calculate The eight values are then arranged in ascending order of the subject number (e.g. 1 for intellectual property, 2 for payment terms) to form an eight-dimensional vector, which is the actual subject distribution of the contract under review, depicting the content distribution of the specific contract in each core issue. At the same time, the system performs the same operation on the standard contract text library, which is a collection of hundreds of high-quality sample contracts for the same type of business (e.g. software development contract). The system calculates the subject distribution vector for each standard contract in the library, then averages the vectors element by element to obtain a final average subject distribution vector representing the "ideal" or "regular" content layout of the type of contract. This averaged vector is defined as the standard subject distribution, reflecting how a well-structured and complete contract should allocate its content in each subject in this business scenario. Finally, the system outputs two vectors of the same dimension and in the same order: the actual subject distribution of the contract under review and the standard subject distribution.

[0129] The steps to obtain the contract subject weight deviation list are as follows:

[0130] Based on the actual subject distribution of the contract under review and the standard subject distribution, the absolute difference between the weight value of each subject in the actual subject distribution and the weight value in the standard subject distribution is calculated. The absolute difference of each subject is arranged in descending order to form the contract subject weight deviation list.

[0131] Specifically, based on the two vectorized data of the actual subject distribution of the under-review contract and the standard subject distribution, the system starts the item-by-item comparison and deviation calculation process. The two vectors have the same dimension and subject order. For example, for a software development contract, both vectors may contain eight subject weight values arranged in the fixed order of "intellectual property", "payment conditions", "confidentiality obligations", "liability for breach of contract", etc. The system first extracts the numerical values at the same index position in the two vectors, i.e., the weight of the same subject in the under-review contract and the standard template, and then calculates the absolute value of the difference between the two weight values. This operation will be performed for all subjects one by one. For example, if the weight of the "intellectual property" subject in the actual subject distribution of the under-review contract is 0.22, and the weight in the standard subject distribution is 0.30, the absolute value of the difference is |0.22-0.30|, i.e., 0.08. If the weight of the "payment conditions" subject is 0.25 in the under-review contract and 0.23 in the standard contract, the absolute value of the difference is |0.25-0.23|, i.e., 0.02. After completing the difference calculation for all eight subjects, the system will obtain a set containing eight difference absolute values, each difference quantifies the deviation between the content length of the under-review contract on the corresponding subject and the industry standard template. Finally, the system arranges these subjects and their corresponding difference absolute values as a pair of data in descending order according to the numerical value of the difference absolute value, with the subject with the largest difference at the forefront, to generate a structured contract subject weight deviation list.

[0132] The risk result obtaining step is:

[0133] Based on the contract subject weight deviation list, a preset threshold is set, and each subject in the contract subject weight deviation list is traversed. If the weight difference absolute value corresponding to a subject is greater than the preset threshold, the subject is marked as a missing risk subject, and all marked missing risk subjects and weight difference absolute values are extracted to form a missing risk subject list.

[0134] Based on the contract intelligibility risk index, an ambiguity risk threshold is set, and the numerical size relationship between the contract intelligibility risk index and the ambiguity risk threshold is judged. If the numerical value of the contract intelligibility risk index is greater than the ambiguity risk threshold, it is determined that there is an ambiguity risk. Combined with the missing risk subject list, the risk result is formed.

[0135] Specifically, based on the contract theme weight deviation list, the system then proceeds to the judgment of missing risk, the core of this process is to set and apply a dynamic, per-theme preset threshold, which is not a fixed global value, but is dynamically calculated according to the statistical characteristics of each theme in the standard contract text library. The specific method is that the system statistically analyzes the weight of each theme of hundreds of contracts of the same type in the standard contract text library in advance, calculates the mean and standard deviation of the weight of each theme, and the threshold setting refers to the "3-sigma" principle in statistical process control, that is, fluctuations within the range of mean plus or minus three times the standard deviation are considered normal, and fluctuations beyond this range are considered abnormal. Therefore, for each theme in the contract theme weight deviation list, the corresponding preset threshold is set to 2.5 times the weight standard deviation of the theme in the standard library. The reason for choosing 2.5 times instead of 3 times is based on backtesting analysis of historical data, which found that in the contract review scenario, 2.5 times the standard deviation can more sensitively capture significant content missing with potential legal risks while maintaining a low false positive rate. For example, if the weight standard deviation of the "breach of contract" theme in the standard library is 0.03, the preset threshold is 0.03*2.5, i.e. 0.075. The system will traverse each theme in the contract theme weight deviation list, compare the absolute value of the weight difference of each theme with the calculated preset threshold of the corresponding theme, and if the difference is greater than the threshold, for example, the absolute value of the weight difference of the "breach of contract" theme of a certain contract is 0.08, which is greater than 0.075, the system will mark this theme as a missing risk theme, and extract the name of the theme and the absolute value of the weight difference. Summarize all the marked themes to form a missing risk theme list.

[0136] Based on the contract intelligibility risk index, the system begins to make the final determination of ambiguity risk, which requires setting an explicit ambiguity risk threshold. The threshold is based on the empirical analysis of historical contract data. The system maintains a database of thousands of completed contracts, each associated with its final legal status, such as "smooth execution", "minor dispute", "major litigation". The system calculates the contract intelligibility risk index for each contract in the historical database, obtaining the distribution of contract risk index under different legal status. By drawing the receiver operating characteristic (ROC) curve, the true positive rate (correctly identifying risky contracts) and false positive rate (misjudging non-risk contracts as risky) under different thresholds are analyzed. The risk index value corresponding to the point with the largest Youden index (true positive rate - false positive rate) is selected as the optimal ambiguity risk threshold. For example, after analyzing a historical database of 2000 software procurement contracts, it is found that when the threshold is set to 0.95, the system can identify contracts that have caused major litigation with an accuracy of 85%, while the false positive rate is controlled within 10%. Therefore, 0.95 is set as the ambiguity risk threshold for this type of contract. When reviewing a new contract, the system compares the previously calculated contract intelligibility risk index (e.g. 1.1522) with this 0.95 threshold. Since 1.1522 is greater than 0.95, the system determines that the contract has ambiguity risk. Finally, the system integrates all analysis results and combines the ambiguity risk determination conclusion (yes or no) with the missing risk theme list generated in the previous step to form a comprehensive risk result report.

[0137] The above is only a preferred embodiment of the present application, and does not limit the form of the present application. Any skilled person in the art can use the disclosed technology to make changes or modifications to equivalent embodiments applied to other fields. However, any simple modification, equivalent change and modification made in accordance with the technical essence of the present application to the above embodiments, without departing from the technical solution content of the present application, still falls within the protection scope of the present application.

Claims

1. A method of risk review of contract terms, characterized by, The method comprises the following steps: obtaining a to-be-reviewed contract text, deconstructing the sentences of each contract clause to obtain a clause analysis feature set; based on the clause analysis feature set, assigning a preset weight to the syntax tree depth, the reference distance, the logical nesting layer number and the professional term density of each recorded clause, generating a single clause analysis complexity score of each clause, and calculating the single clause analysis complexity scores of all clauses in the contract to obtain a contract understandability risk index; based on the to-be-reviewed contract text and a standard contract text library of the same type, calculating the word co-occurrence frequency of the to-be-reviewed contract text and the standard contract text library of the same type and clustering into topics to generate an actual topic distribution of the to-be-reviewed contract and a standard topic distribution, comparing the weight of each topic in the actual topic distribution of the to-be-reviewed contract with the corresponding item in the standard topic distribution one by one to obtain a contract topic weight deviation list; screening topics in the contract topic weight deviation list that exceed a preset threshold to determine missing risks, and determining that the contract understandability risk index exceeds a threshold to determine ambiguity risks, and generating a risk result; the step of obtaining the single clause analysis complexity score is: based on the clause analysis feature set, extracting the syntax tree depth value, the reference distance value, the logical nesting layer number value and the professional term density value of each contract clause, respectively performing maximum and minimum difference scaling processing, and sequentially mapping the original values to the closed interval 0 to 1 to record the standardized syntax tree depth value, the standardized reference distance value, the standardized logical nesting layer number value and the standardized professional term density value; according to the standardized syntax tree depth value, the standardized reference distance value, the standardized logical nesting layer number value and the standardized professional term density value, calculating the single clause analysis complexity value of each contract clause, and the calculation formula is: ； wherein, represents the number of clauses in the i-th contract clause, represents the clause complexity value of the i-th contract clause, represents the normalized syntax tree depth value in the i-th contract clause, represents the normalized anaphora distance value in the i-th contract clause, represents the normalized logical nesting level value in the i-th contract clause, represents the normalized professional terminology density value in the i-th contract clause, represents the normalized syntax tree depth value, the normalized anaphora distance value, the normalized logical nesting level value and the normalized professional terminology density value, respectively, represents the normalized syntax tree depth value, the normalized anaphora distance value, the normalized logical nesting level value and the normalized professional terminology density value, respectively, represents the normalized syntax tree depth value, the normalized anaphora distance value, the normalized logical nesting level value and the normalized professional terminology density value, respectively, represents the normalized syntax tree depth value, the normalized anaphora distance value, the normalized logical nesting level value and the normalized professional terminology density value, respectively, represents the normalized syntax tree depth value, the normalized anaphora distance value, the normalized logical nesting level value and the normalized professional terminology density value, respectively, represents the normalized syntax tree depth value, the normalized anaphora distance value, the normalized logical nesting level value and the normalized professional terminology density value, respectively, represents the normalized syntax tree depth value, the normalized anaphora distance value, the normalized logical nesting level value and the normalized professional terminology density value, respectively, the step of obtaining the clause analysis feature set is: obtaining a to-be-reviewed contract text, performing sentence processing on the to-be-reviewed contract text, splitting each contract clause into independent sentences, marking each syntactic component in each sentence, analyzing the direct dependency relationship between each syntactic component, determining the syntax tree structure of each sentence according to the dependency relationship link, counting the longest path node number from the syntax tree root node to the leaf node, and generating a syntax tree depth value; based on the syntax tree depth value, traversing the word sequence in each contract clause sentence by sentence, identifying all referring words and referring entities in the sentence, associating and pairing the referring words with the corresponding referring entities, recording the actual word distance between each association and pairing, taking the number of intermediate interval words in the pairing as the word distance value, and forming a word distance value set of all referring word-referring entity pairs in the clause; Based on the set of lexical distance values, the conditional logical operators and temporal logical operators of the contract clauses are identified sentence by sentence, the identified conditional logical operators and temporal logical operators are expanded one by one and recursively in turn, the number of nested layers of each operator in the sentence is counted, and the frequency of each professional term in the preset list in the contract clause sentence is counted, so as to generate a clause analysis feature set with the syntax tree depth value, the set of lexical distance values, the number of nested layers of the operator, and the frequency of the professional term as the clause characteristic dimensions.

2. The method of risk review of contract terms of claim 1, wherein, The obtaining step of the contract intelligibility risk index is: Based on the single clause analysis complexity values corresponding to all clauses, the contract intelligibility risk index is calculated.

3. The method of risk review of contract terms of claim 1, wherein, The obtaining steps of the actual subject distribution of the to-be-reviewed contract and the standard subject distribution are: The to-be-reviewed contract text and the standard contract text library are respectively subjected to paragraph segmentation, each paragraph is subjected to word-by-word segmentation, and each word element is converted into a word vector by calling a pre-trained semantic model, then the word vectors of all word elements in each paragraph are subjected to weighted average to generate a paragraph vector, and a fixed theme keyword group is set according to the business type, and a corresponding theme center vector is generated by averaging the keyword vectors of each group, so as to obtain a paragraph vector set and a theme center vector set; According to the paragraph vector set and the theme center vector set, the semantic weight value of each theme in the to-be-reviewed contract is calculated; Based on the semantic weight value of each theme, the semantic weight value of the to-be-reviewed contract and the standard contract text library is calculated respectively, and a corresponding weight composition vector is output in ascending order of theme number, and the actual subject distribution of the to-be-reviewed contract and the standard subject distribution are constructed respectively.

4. The method of claim 1, wherein, The obtaining step of the contract theme weight deviation list is: Based on the actual subject distribution of the to-be-reviewed contract and the standard subject distribution, the weight values of the same themes in the theme distributions are compared one by one, the absolute value of the difference between the weight value of each theme in the actual subject distribution of the to-be-reviewed contract and the weight value in the standard subject distribution is calculated, the absolute values of the differences of the themes are arranged in descending order, and a contract theme weight deviation list is formed.

5. The method of claim 1, wherein, The obtaining step of the risk result is: Based on the contract theme weight deviation list, a preset threshold is set, each theme in the contract theme weight deviation list is traversed, if the weight difference absolute value corresponding to a theme is greater than the preset threshold, the theme is marked as a missing risk theme, and all marked missing risk themes and weight difference absolute values are extracted to form a missing risk theme list.

6. The method of risk review of contract terms of claim 5, wherein, The obtaining step of the risk result further includes: based on the contract intelligibility risk index, setting an ambiguity risk threshold, judging the value size relationship between the contract intelligibility risk index and the ambiguity risk threshold, if the value of the contract intelligibility risk index is greater than the ambiguity risk threshold, it is determined that there is an ambiguity risk, and the risk result is formed in combination with the missing risk theme list.

7. The risk review system of the contract clause risk review method according to any one of claims 1 to 6, characterized in that, It includes: A contract analysis module acquires a to-be-reviewed contract text, deconstructs the sentences of each contract clause, and obtains a clause analysis feature set; a complexity evaluation module, which assigns preset weights to the syntax tree depth, the distance of reference, the number of logical nesting layers, and the density of professional terms of each clause of the record based on the clause parsing feature set, generates a single-clause parsing complexity score of each clause, and obtains a contract understandability risk index by aggregating and calculating the single-clause parsing complexity scores of all clauses in the contract; a theme deviation analysis module, which calculates the word co-occurrence frequency of the to-be-reviewed contract text and the same-type standard contract text library and clusters the words into themes based on the to-be-reviewed contract text and the same-type standard contract text library, generates an actual theme distribution of the to-be-reviewed contract and a standard theme distribution, compares each theme weight in the actual theme distribution of the to-be-reviewed contract with a corresponding item in the standard theme distribution one by one, and obtains a contract theme weight deviation list; a risk identification module, which screens themes with a theme weight deviation list exceeding a preset threshold, determines that the themes are missing risks, determines that the contract understandability risk index exceeding a threshold is an ambiguity risk, and generates a risk result.

Images