Method for generating an automated compliance review workflow based on a large model intelligent agent

Abstract

This invention discloses an automated compliance review workflow generation method based on large-scale model intelligent agent orchestration. It extracts text feature vectors from input documents using a document classification model and obtains a list of key clauses based on these vectors. Semantic intent analysis is performed on the key clause list to determine the clause type of each clause and identify potential risk points. If a clause type is classified as high-risk, a rule engine is used to match it with legal provisions in a knowledge base, obtaining comparison results and identifying non-compliant items, assessing the number and severity of these non-compliant items. When the number of non-compliant items exceeds a preset threshold, a sequence generation model is used to generate a dynamic step sequence based on the review context, and the input data for each step in the dynamic step sequence is obtained. This invention achieves fully automated processing from risk identification to compliance assessment, significantly improving review efficiency and accuracy, and providing efficient technical support for the compliance management of complex documents.

Description

Technical Field

[0001] This invention relates to the field of enterprise management technology, and in particular discloses an automated compliance review workflow generation method based on large model intelligent agent orchestration. Background Technology

[0002] In the field of corporate management, compliance review is a crucial step in ensuring that a company's operations comply with laws, regulations, and internal norms. Its accuracy and efficiency directly affect the company's risk control and sound development. Currently, many companies rely on manual review or software based on fixed rules for review. These methods prove inadequate when faced with complex and ever-changing business scenarios and constantly updated regulatory provisions.

[0003] Existing methods typically treat the review process as a series of isolated steps, lacking the ability to dynamically coordinate and break down the overall task. This leads to rigid processes that struggle to flexibly adapt to the changing needs of different review scenarios, such as contracts and invoices. Each adjustment requires significant manual redesign and reconfiguration, resulting in time and effort wasted. The root of this rigidity lies in the difficulty of effectively constraining and integrating the powerful semantic understanding capabilities of large models into deterministic business logic chains. On the one hand, large models can understand text, but their output is uncertain, and directly using them for serious compliance judgments may lead to unreliable results. On the other hand, deterministic business logic requires clear steps and traceable results. There is a contradiction between these two, and neither simply using large models nor fixed rules can achieve both simultaneously.

[0004] This leads to a core technical challenge: how to achieve deep collaboration between "intelligent, flexible understanding" and "determined business logic." Without this collaboration, it's impossible to build a reliable automated process that can both automatically parse document intent and strictly adhere to review steps. Specifically, the lack of collaboration between intelligent understanding and business logic creates a unique business problem in automated review. For example, when a system attempts to automatically review a contract, it might identify a key clause but fail to accurately determine which "review stage" should handle it, or how to precisely compare the identified information with specific legal provisions in the knowledge base. The entire process cannot be automatically connected and advanced.

[0005] Therefore, the key issue in achieving efficient, reliable, and automated compliance review is how to enable machines to not only understand the content but also autonomously and systematically complete the entire review process from analysis and comparison to judgment, just like experienced human beings. Summary of the Invention

[0006] This invention provides an automated compliance review workflow generation method based on large model intelligent agent orchestration, aiming to solve at least one of the defects existing in the above-mentioned prior art.

[0007] This invention relates to an automated compliance review workflow generation method based on large-scale intelligent agent orchestration, comprising the following steps:

[0008] S100. Extract the text feature vector of the input document through the document classification model, and obtain the list of key terms based on the text feature vector;

[0009] S200. Conduct semantic intent analysis on the list of key terms to determine the term type of each term and identify potential risk points;

[0010] S300 If the clause type is a high-risk category, the rule engine will match the legal provisions in the knowledge base to obtain the comparison results and identify non-compliance items, and assess the number and severity of non-compliance items.

[0011] S400 When the number of non-compliant items exceeds a preset threshold, a sequence generation model is used to generate a dynamic step sequence based on the review context, and the input data for each step in the dynamic step sequence is obtained.

[0012] S500. Determine the execution order and dependencies of steps in the dynamic step sequence, and determine the triggering conditions of subsequent steps based on the dependencies to form a complete review chain.

[0013] S600: Extract judgment criteria from the complete review chain and generate a compliance assessment report.

[0014] Further, step S100 includes:

[0015] S110. Obtain the input document, perform word segmentation and part-of-speech tagging on the input document to obtain a basic vocabulary set, and calculate the word frequency distribution matrix based on the basic vocabulary set;

[0016] S120. Input the word frequency distribution matrix into the document classification model for semantic parsing to determine the text feature vector;

[0017] S130. Calculate the set of similarity values ​​between the text feature vector and the benchmark clause vector library;

[0018] S140. If the value in the similarity value set is greater than the preset threshold, then filter the terms in the benchmark term vector library to obtain a list of key terms.

[0019] Further, step S200 includes:

[0020] S210. Use a bidirectional long short-term memory network to perform sequence modeling on the list of key terms to output an intent vector;

[0021] S220. Input the intent vector into the multilayer perceptron classifier to determine the clause type of each clause. The clause type is selected based on the probability distribution of the predefined semantic intent space.

[0022] S230. Extract the responsible parties and the boundaries of obligations according to the clause type, and map the responsible parties and the boundaries of obligations to the risk rule base to obtain risk characteristic data;

[0023] S240. If the deviation value in the risk characteristic data is higher than the preset threshold, the output will include the potential risk points with the clause type and violation mark.

[0024] Further, step S300 includes:

[0025] S310. If the clause type is a high-risk category, the named entity recognition model is used to extract the text feature vector of the clause corresponding to the high-risk category.

[0026] S320. Input the feature vector of the clause text into the preset rule engine to match the set of legal provisions in the pre-built knowledge base. The set of legal provisions is obtained through semantic similarity calculation.

[0027] S330. Use the textual implication model to perform logical conflict analysis on the feature vectors of the legal provisions and clause texts to obtain the comparison results, which include conflicting segments.

[0028] S340. Identify non-compliant items based on the comparison results and obtain the corresponding penalty amount. Assess the number and severity of non-compliant items based on the penalty amount.

[0029] Further, step S400 includes:

[0030] S410. If the cumulative number of non-compliance items exceeds a preset threshold, extract the review context containing historical review trajectories.

[0031] S420. Convert the review context into a model input vector;

[0032] S430. Generate a model by inputting the model input vector into the sequence to generate a dynamic step sequence;

[0033] S440. Parse the dynamic step sequence to obtain data index information, and retrieve the step execution parameters based on the data index information to obtain the input data for each step in the dynamic step sequence.

[0034] Further, step S500 includes:

[0035] S510. Parse the dynamic step sequence to extract the dependency identification data that defines the relationship between nodes, and map the dependency identification data to the node topology model to calculate the execution time sequence weight value.

[0036] S520. Construct a linear execution path containing serial and parallel logic based on the execution timing weight values, identify the blocking factors of nodes in the linear execution path and convert them into pre-start state constraints.

[0037] S530. If the pre-start state constraints meet the verification requirements, a trigger control instruction is generated and embedded into the linear execution path to form a complete review chain.

[0038] Further, step S600 includes:

[0039] S610. Traverse the complete review chain to capture node execution logs and status snapshots, and identify operational behavior data to form a set of original evidence fragments;

[0040] S620. Map the original set of evidence fragments to the pre-stored compliance rule base, and analyze the logical correspondence between the operational behavior and the clause to determine the set of judgment basis;

[0041] S630. Employ a multi-dimensional assessment matrix to quantify the set of judgment criteria, calculate the compliance matching degree, and aggregate to generate a risk distribution map;

[0042] S640. Based on the anomaly markers in the risk distribution map, trace back the associated logical path and fill the report structure with the compliance matching degree to generate a compliance assessment report.

[0043] The beneficial effects achieved by this invention are as follows:

[0044] This invention provides an automated compliance review workflow generation method based on large-scale intelligent agent orchestration. Addressing the challenges of multi-dimensional risk identification and dynamic processing in document compliance review, this invention proposes a solution integrating text feature extraction, semantic analysis, and dynamic review chain generation. By extracting feature vectors from document text, this invention accurately obtains a list of key clauses and, combined with semantic intent analysis, identifies clause types and potential risk points. Particularly for high-risk clauses, a rule engine is used to match regulatory provisions and assess the number and severity of non-compliance items. When non-compliance items exceed the limit, a sequence generation model is used to generate a dynamic sequence of steps. Based on the dependencies between steps, a complete review chain is formed, ultimately extracting the judgment criteria and generating a compliance assessment report. This invention, through an intelligent and dynamic review process, achieves fully automated processing from risk identification to compliance assessment, significantly improving review efficiency and accuracy, and providing efficient technical support for the compliance management of complex documents. Attached Figure Description

[0045] Figure 1 This is a flowchart illustrating an embodiment of the automated compliance review workflow generation method based on large model intelligent agent orchestration of the present invention. Detailed Implementation

[0046] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0047] like Figure 1 As shown, the first embodiment of the present invention proposes an automated compliance review workflow generation method based on large model intelligent agent orchestration, including the following steps:

[0048] Step S100: Extract the text feature vector of the input document through the document classification model, and obtain the list of key terms based on the text feature vector.

[0049] This step involves input document processing and key clause extraction. The compliance-related documents to be reviewed (such as contracts, agreements, business process documents, rules and regulations, etc.) are input into a pre-defined document classification model (such as a deep learning classification model based on BERT (Bidirectional Encoder Representations from Transformers) or RoBERTa (Robustly Optimized BERT Pretraining Approach)). The model extracts text feature vectors (768-1024 dimensions, adjustable according to model complexity) from the documents through text segmentation, feature encoding, and semantic understanding. Based on these text feature vectors, keyword matching, semantic clustering, and importance ranking algorithms are used to filter out core content directly related to compliance requirements, forming a list of key clauses (5-50 clauses, depending on document length and complexity). Each key clause includes the original text, its location in the document, and related business scenarios, providing a target for subsequent semantic analysis and risk identification.

[0050] The large model intelligent agent is an intelligent entity built on a large language model (LLM) that has the ability to make autonomous decisions, understand semantics, and orchestrate tasks. It can automatically complete operations such as text analysis, risk identification, and step generation. In this invention, it is used for clause semantic intent analysis, dynamic step sequence parsing, and review context processing. Its core is to realize the intelligentization and automation of compliance review work.

[0051] Document classification models are deep learning-based text processing models that focus on classifying and extracting features from input documents. They can accurately extract core features related to compliance review from unstructured documents, providing technical support for extracting key clauses. Common models include BERT and RoBERTa.

[0052] The key clause list is a collection of core clauses extracted from the input document that are directly related to the compliance review. It includes the original text of the clauses, their location, and related scenarios, and serves as the core foundation for subsequent semantic analysis, risk identification, and compliance comparison.

[0053] Step S200: Perform semantic intent analysis on the list of key terms to determine the term type of each term and identify potential risk points.

[0054] This step involves semantic analysis and risk identification of the clauses. Large-scale intelligent models (such as GPT-4 (Generative Pre-trained Transformer 4) and Tongyi Qianwen, etc.) are used to perform semantic intent analysis on each clause in the key clause list. Combined with the terminology and scenario databases in the compliance review knowledge base, the clause type of each clause is clarified. Specific classifications include: qualification, authority, process, responsibility, data security, privacy protection, etc. (which can be expanded according to industry compliance needs). Simultaneously, through semantic matching and risk rule comparison, potential compliance risks in each clause are identified (such as ambiguous clause wording, conflicts with regulations, unclear responsible parties, missing essential elements, etc.). These risk points are initially marked with a marking accuracy of ≥90% (preset accuracy threshold) to ensure the accuracy and comprehensiveness of risk point identification.

[0055] Step S300: If the clause type is a high-risk category, the rule engine is used to match the legal provisions in the knowledge base to obtain the comparison results and identify non-compliant items, and assess the number and severity of non-compliant items.

[0056] This step involves compliance comparison and non-compliance assessment of high-risk clauses. A pre-defined risk classification standard categorizes clauses into high-risk, medium-risk, and low-risk categories: high-risk categories include data security, privacy protection, and qualification compliance; medium-risk categories include process and authority-related clauses; and low-risk categories include supplementary clauses related to liability. If a clause is determined to be high-risk, the rules engine is triggered. The rules engine retrieves legal provisions from a pre-defined knowledge base, comparing the high-risk clause with the corresponding legal provisions sentence by sentence, outputting the comparison result (compliant / non-compliant / questionable), and accurately identifying non-compliant items. Simultaneously, non-compliant items are quantitatively assessed: the number of non-compliant items is based on actual statistics, and the severity of non-compliant items is divided into three levels (Level 1: minor non-compliance, requiring only textual correction; Level 2: general non-compliance, requiring adjustment of clause logic; Level 3: serious non-compliance, violating core regulations, requiring comprehensive rectification), with an assessment accuracy rate ≥85%.

[0057] A rules engine is an intelligent reasoning component that pre-defines compliance review rules and logic. It can call upon legal provisions in a knowledge base and automatically compare them with key clauses to quickly identify non-compliant items and output comparison results. Its core function is to transform legal rules into executable reasoning logic.

[0058] Step S400: When the number of non-compliant items exceeds a preset threshold, a sequence generation model is used to generate a dynamic step sequence based on the review context, and the input data for each step in the dynamic step sequence is obtained.

[0059] This step is the dynamic step sequence generation stage for compliance review. A preset threshold for the number of non-compliant items is established (the threshold ranges from 3 to 10 items, adjustable according to the review scenario; for example, 3 items for ordinary contracts and 5 items for core business contracts). When the number of identified non-compliant items exceeds this preset threshold, the sequence generation model (such as a Transformer-based sequence generation model) is activated. The model is input with the review context (including key clauses, non-compliant items, regulatory comparison results, and risk assessment conclusions) to generate a dynamic step sequence (4 to 15 steps, depending on the number and severity of non-compliant items) for rectifying non-compliant items and completing the compliance review. Simultaneously, a large-scale model agent analyzes the execution requirements of each step, clarifying the required input data for each step (such as rectification instructions, supplementary materials, qualification certificates, approval documents, etc.) to ensure that each step is implementable and executable, with input data completeness ≥95%.

[0060] The sequence generation model is a generative model built on the Transformer architecture. It can automatically generate an ordered and executable dynamic sequence of steps based on the input review context (key clauses, non-compliance items, etc.). It is used to guide the entire process of non-compliance rectification and compliance review, and has context awareness and logical coherence.

[0061] Step S500: Determine the execution order and dependencies of the steps in the dynamic step sequence, and determine the triggering conditions of subsequent steps based on the dependencies to form a complete review chain.

[0062] This step involves constructing the review chain and determining trigger conditions. Based on a dynamic step sequence, a directed graph analysis method is used to analyze the execution order of each step (e.g., "non-compliance item verification → rectification plan formulation → plan review → rectification implementation → re-review"), clarifying the dependencies between steps (e.g., "rectification plan review" depends on the completion of "rectification plan formulation," and "re-review" depends on the completion of "rectification implementation"). Based on these dependencies, trigger conditions are set for each subsequent step (e.g., trigger condition 1: previous steps are completed and the results are satisfactory; trigger condition 2: supplementary input data is complete and meets requirements; trigger condition 3: review comments are approved). The dynamic step sequence, execution order, dependencies, and trigger conditions are integrated to form a logically closed-loop, automatically executable complete review chain with a success rate of ≥90%.

[0063] The complete review chain is a logical closed loop consisting of a dynamic sequence of steps, the order of step execution, the dependencies between steps, and the triggering conditions. It can automate the entire process of compliance review from "clause extraction → risk identification → non-compliance comparison → rectification → review", ensuring the standardization and completeness of the review work.

[0064] Step S600: Extract the basis for judgment from the complete review chain and generate a compliance assessment report.

[0065] This step is the compliance assessment report generation stage. From the complete review chain, core information such as the execution results of each step, the rectification status of non-compliance items, the basis for regulatory comparison, and the risk assessment conclusions are extracted as the basis for the compliance assessment. Following a standardized template (including basic document information, a summary of key clauses, risk point identification results, details of non-compliance items and rectification status, review conclusions, and rectification recommendations), a compliance assessment report is generated. The report clearly indicates the compliance level (compliant, basically compliant, non-compliant), where: compliant (0 non-compliance items), basically compliant (number of non-compliance items ≤ preset threshold and all are of severity level 1), and non-compliant (number of non-compliance items exceeds the preset threshold or there are non-compliance items of severity level 2 or above). The report generation efficiency is ≤30 minutes per report (single document ≤50 pages).

[0066] Furthermore, the automated compliance review workflow generation method based on large model intelligent agent orchestration provided in this embodiment includes step S100 as follows:

[0067] S110. Obtain the input document, perform word segmentation and part-of-speech tagging on the input document to obtain a basic vocabulary set, and calculate the word frequency distribution matrix based on the basic vocabulary set.

[0068] The word frequency of the input document is obtained using the following formula:

[0069]

[0070] In formula (1), Indicates the first The document in the first The word frequency of each word Indicates the first The document in the first The number of times each word appears, Indicates the first All in each document The total number of occurrences of each word. The control logic of formula (1) is the standardized implementation of text word frequency statistics. Its core value is: 1. Normalization: It unifies the word frequency statistics of documents of different lengths to the same scale, avoiding the high word frequency of long documents due to the large total number of words; 2. Feature extraction basis: The word frequency distribution matrix is ​​the input of subsequent text analysis algorithms such as TF-IDF and topic model, providing quantitative text features for compliance review; 3. Engineering scalability: It is compatible with various input documents (such as regulations, contracts, and technical documents) and can be directly embedded into the automated workflow of large model intelligent agent orchestration.

[0071] The word frequency distribution matrix is ​​obtained using the following formula:

[0072] (2)

[0073] In formula (2), Represents the word frequency distribution matrix. Indicates the total number of documents. This represents the total number of words in the basic vocabulary set. The control logic of formula (2) is the matrix encapsulation of text word frequency features. Its core value is: 1. Structured expression: It transforms the scattered word frequency statistics into a structured matrix, which is convenient for efficient computer storage and calculation; 2. Dimension alignment: It clarifies the correspondence between matrix dimensions and the number of documents and words, ensuring the consistency of feature extraction; 3. Workflow adaptation: It can be directly used as the input feature of large models and machine learning models to support the subsequent process of automated compliance review.

[0074] The automated compliance review system first acquires the input document, such as a contract text file, and then performs word segmentation and part-of-speech tagging using natural language processing tools. Specifically, word segmentation breaks the text down into word units; for example, "contract terms" is divided into "contract" and "terms." Part-of-speech tagging identifies the type of each word, such as noun or verb, thus obtaining a basic vocabulary set, which may contain high-frequency words such as "rights" and "obligations." Based on this basic vocabulary set, the process of calculating the word frequency distribution matrix involves counting the frequency of each word in the document and constructing a matrix where rows represent words, columns represent document paragraphs, and matrix elements are frequency values. For example, in a lease contract, the word "rent" appears in multiple paragraphs, and its frequency can be obtained by counting, thus forming a matrix to capture word distribution patterns. This word frequency distribution matrix helps quantify the text structure and provides a data foundation for subsequent semantic analysis.

[0075] S120. Input the word frequency distribution matrix into the document classification model for semantic parsing to determine the text feature vector.

[0076] The text feature vector is determined using the following formula:

[0077] (3)

[0078] In formula (3), Represents the text feature vector. This represents the activation function of the document classification model. Represents the word embedding matrix, The bias term is represented by the model bias term. The control logic of formula (3) is the core of the forward propagation of the document classification model. Its core value is: 1. Frequency-semantic conversion: mapping discrete frequency features to continuous semantic embeddings, preserving text semantic information; 2. Nonlinear modeling: introducing nonlinearity through activation functions to improve the document classification model's ability to model complex text semantics; 3. Standardized feature output: generating fixed-dimensional text feature vectors to provide a unified input for subsequent compliance review tasks.

[0079] The word frequency distribution matrix is ​​input into a document classification model for semantic parsing to determine text feature vectors. Specifically, the document classification model can employ a pre-trained model such as BERT. By inputting the word frequency distribution matrix, the model learns the semantic relationships between words. For example, in contract analysis, the model parses the association between "breach of contract" and "compensation," outputting high-dimensional feature vectors, such as a 128-dimensional vector representing the core semantics of the document, thereby achieving a deep understanding of the text. This parsing process extracts features step by step through a multi-layer neural network. First, a shallow representation is obtained from the word frequency distribution matrix, then an attention mechanism is used to capture contextual dependencies, and finally, text feature vectors are generated to quantify the document's topic.

[0080] S130. Calculate the set of similarity values ​​between the text feature vector and the benchmark clause vector library.

[0081] The similarity score set is obtained using the following formula:

[0082] (4)

[0083] In formula (4), Represents a set of similarity scores. Represents the cosine similarity function. Represents the text feature vector. This represents the feature vector of a specific clause in the benchmark clause vector library. Indicating the reference clause vector library The feature vector of each clause. The control logic of formula (4) is the standardized implementation of semantic matching between text features and benchmark clauses. Its core value is: 1. Semantic similarity quantification: transforming abstract text semantic similarity into calculable values ​​to achieve automated matching; 2. Batch efficient calculation: obtaining the similarity between the document and all benchmark clauses in one traversal, adapting to large-scale compliance review scenarios; 3. Standardized output: generating a set of similarity values. It can be directly used for subsequent sorting, threshold judgment and risk classification, supporting the automated workflow of large model intelligent agent orchestration.

[0084] The cosine similarity function between the text feature vector and the benchmark clause vector library is derived using the following formula:

[0085] (5)

[0086] In formula (5), The cosine similarity function between the text feature vector and the benchmark clause vector library. This represents the transpose of the feature vector of a specific clause in the benchmark clause vector library. The L2 norm of the text feature vector is represented. This represents the L2 norm of the feature vector of a certain clause in the benchmark clause vector library. The control logic of formula (5) is implemented in matrix form of cosine similarity. Its core values ​​are: 1. Direction priority: It only focuses on the difference in vector direction, eliminates the interference of vector length (word frequency, document length, etc.), and measures semantic similarity more accurately; 2. Numerical standardization: It compresses the similarity to [ 1,1] interval, which makes it easy to set thresholds for automatic matching; 3. Computationally efficient: matrix form is adapted to GPU / distributed computing, and can efficiently process large-scale documents and clause libraries.

[0087] When calculating the similarity scores between the input text feature vector and the benchmark clause vector library, a cosine similarity formula is used. The benchmark clause vector library is a pre-built database containing vector representations of standard contract clauses; for example, the library stores the text feature vector of a "confidentiality clause." By comparing the angle between the input text feature vector and each text feature vector in the benchmark clause vector library, a set of similarity values, such as 0.85 or 0.92, is obtained to assess the degree of matching.

[0088] S140. If the value in the similarity value set is greater than the preset threshold, then filter the terms in the benchmark term vector library to obtain a list of key terms.

[0089] The list of key terms is derived using the following formula:

[0090] (6)

[0091] In formula (6), This indicates a list of key terms. This indicates a preset threshold for similarity. The control logic of formula (6) is the key clause screening logic for automated compliance review. Its core value is: 1. Precise matching: By filtering through the threshold, only clauses with high similarity are retained, avoiding interference from irrelevant clauses; 2. Efficiency improvement: The large-scale benchmark clause library is reduced to a small number of key clauses, significantly reducing the workload of subsequent analysis; 3. Standardized output: The generated list of key clauses. It can be directly connected to large-scale intelligent models for subsequent processes such as risk identification and report generation.

[0092] If the value in the similarity set is greater than a preset threshold, for example, a preset threshold of 0.8, the corresponding clauses in the benchmark clause vector library are filtered to obtain a list of key clauses, thereby extracting clauses that are highly relevant to the input document, such as "termination conditions". This helps to automate contract review and improve efficiency.

[0093] Preferably, the automated compliance review workflow generation method based on large model intelligent agent orchestration provided in this embodiment includes step S200 as follows:

[0094] S210. Use a bidirectional long short-term memory network to perform sequence modeling on the list of key terms to output an intent vector.

[0095] The intent vector is generated using the following formula:

[0096] (7)

[0097] In formula (7), Represents the intent vector. Indicates the sequence length of key clauses. This indicates that BiLSTM is at time step The hidden layer output. The control logic of formula (7) is the global intent extraction logic of BiLSTM sequence modeling. Its core value is: 1. Context semantic aggregation: By mean pooling, the bidirectional context information captured by BiLSTM is compressed into global intent, avoiding the influence of sequence length differences; 2. Intent standardization: Generate a fixed-dimensional intent vector to provide a unified input for the workflow orchestration of subsequent large model agents; 3. Strong robustness: Mean pooling is not sensitive to local noise and can stably extract the core business intent of key terms.

[0098] BiLSTM at time step Hidden layer output This can be derived from the following formula:

[0099]

[0100] In formula (8), This indicates that BiLSTM is at time step The hidden layer output, This represents the Sigmoid activation function. This represents the hidden layer weight matrix. Indicates that the forward LSTM is in Step output, Indicates that the inverse LSTM is in Step output, Indicates time step Input, The hidden layer bias term is represented. The control logic of formula (8) is the core computational unit of bidirectional LSTM. Its core value is: 1. Bidirectional context modeling: It utilizes both past and future contexts to capture the semantic dependencies of the clause sequence more accurately than unidirectional LSTM; 2. Information fusion: It efficiently concatenates the three types of information, namely historical, future and current input, to provide comprehensive semantic features for intent recognition; 3. Nonlinear expression: Through Sigmoid activation, the model can learn complex semantic patterns and improve the accuracy of intent recognition.

[0101] An automated compliance review system processes a list of key terms in a lease agreement, such as a sequence containing terms like "rent payment" and "liability for breach of contract." It first employs a bidirectional long short-term memory (BSSM) network for sequence modeling. Specifically, the BSSM network is a recurrent neural network architecture capable of capturing sequence dependencies from both directions. For example, when processing the "rent payment" term, the forward layer of the BSSM network analyzes the impact of preceding context such as "contract term," while the backward layer considers the context of following context such as "termination conditions." This allows for the regulation of information flow through forget gates, input gates, and output gates, calculating the hidden state sequence and ultimately outputting an intent vector, such as a 64-dimensional vector representing the overall payment intent.

[0102] S220. Input the intent vector into the multilayer perceptron classifier to determine the clause type of each clause. The clause type is selected based on the probability distribution of the predefined semantic intent space.

[0103] The probability distribution of the predefined semantic intent space is derived using the following formula:

[0104] (9)

[0105] In formula (9), Represents the probability distribution of a predefined semantic intent space. Represents the normalized exponential function, This represents the weight matrix of the MLP classifier. This represents the bias term of the MLP classifier. The control logic of formula (9) is the standardized transformation from intent vector to clause type probability distribution. Its core value is: 1. Intent classification: transforming the abstract intent vector into an interpretable clause type probability to achieve automated identification of compliance intent; 2. Probability normalization: the softmax function ensures that the output is a legal probability distribution, which is convenient for subsequent threshold judgment and confidence assessment; 3. Workflow-driven: the classification result is directly used to orchestrate the corresponding compliance review sub-processes of the large model agent to achieve automated mapping of "intent → workflow".

[0106] The type of clause is determined by the following formula:

[0107] (10)

[0108] In formula (10), Indicates the type of terms. The maximum index function is represented. The control logic of formula (10) is a standardized mapping from probability distribution to discrete clause type. Its core value is: 1. Deterministic classification: the probability distribution output by softmax is transformed into a unique category label to realize the final output of intent recognition; 2. Efficient decision-making: by taking the maximum probability index, the classification result is guaranteed to be consistent with the confidence of the model prediction; 3. Workflow adaptation: the classification result can be directly used as the trigger condition for intelligent agent orchestration to realize the automatic closed loop of "intent → type → workflow".

[0109] The intent vector is input into a multilayer perceptron classifier to determine the clause type of each clause. The multilayer perceptron classifier is a feedforward neural network comprising an input layer, multiple hidden layers, and an output layer, which handles non-linear relationships using activation functions such as ReLU. Specifically, for the input intent vector, the multilayer perceptron classifier first performs weighted summation and activation in the hidden layers to progressively extract high-level features. Then, in the output layer, it calculates a probability distribution based on a softmax function and selects the type with the highest probability in a predefined semantic intent space. For example, "rent payment" is classified as an "obligation-type" clause because its probability distribution shows that the proportion of obligation intent (0.75) is higher than that of other intents such as rights (0.2), thus achieving accurate type determination.

[0110] S230. Extract the responsible parties and the boundaries of obligations according to the clause type, and map the responsible parties and the boundaries of obligations to the risk rule base to obtain risk characteristic data.

[0111] Risk characteristic data is derived using the following formula:

[0112] (11)

[0113] In formula (11), Data representing risk characteristics, This represents the feature vector indicating the responsible party and the boundary of obligations under the current clause. This represents the baseline risk feature vector in the risk rule base. The L2 norm of the vector is represented. The control logic of formula (11) is a method for quantifying the difference between compliance clauses and risk rules. Its core values ​​are: 1. Difference quantification: transforming unstructured compliance elements such as the responsible party and the boundary of obligations into calculable vector differences, thereby achieving numerical measurement of risk; 2. Normalization: eliminating the influence of the benchmark vector scale, making the risk values ​​of different types of clauses comparable; 3. Risk interpretability: It intuitively reflects the degree to which the terms deviate from the standards, facilitating subsequent risk classification and handling.

[0114] Based on the clause type, the responsible party and the boundary of obligations are extracted. For example, in an "obligation-type" clause, the automated compliance review system uses named entity recognition technology to parse the text, identify the responsible party such as "tenant," and define the boundary of obligations such as "the amount and deadline for monthly rent payment." These elements are then mapped to the risk rule base. Specifically, the risk rule base is a pre-built database that stores standard risk patterns, such as deviation rules corresponding to "payment delay." The mapping process involves comparing the degree of matching between the boundary of obligations and the rules in the database, and calculating the deviation value. For example, if the actual boundary of obligations is "payment before the 1st of each month" while the rule requires "strict punctuality," the deviation is quantified using the Euclidean distance formula to obtain risk characteristic data such as a set of deviation values ​​of 0.9, thereby revealing potential non-compliance points.

[0115] S240. If the deviation value in the risk characteristic data is higher than the preset threshold, the output will include the potential risk points with the clause type and violation mark.

[0116] Potential risk points are derived using the following formula:

[0117] (12)

[0118] In formula (12), Indicate potential risk points, Indicates a violation sign. The deviation is preset threshold. The control logic of formula (12) is the standardized generation logic of compliance risk points. Its core values ​​are: 1. Accurate risk marking: only high-risk items above the threshold are retained to avoid interference from low-risk items; 2. Information structuring: the clause type is bound to the violation mark to facilitate subsequent risk tracing and handling; 3. Automated closed loop: directly connects to the workflow of the large model intelligent agent to realize the automated closed loop of "risk identification → intelligent handling".

[0119] If the deviation value in the risk characteristic data is higher than a preset threshold such as 0.8, then the potential risk point containing the term type and violation identifier will be output, such as "obligation type - payment delay violation", so as to facilitate timely intervention by the user.

[0120] Furthermore, the automated compliance review workflow generation method based on large model intelligent agent orchestration provided in this embodiment includes step S300 as follows:

[0121] S310. If the clause type is a high-risk category, then the named entity recognition model is used to extract the text feature vector of the clause corresponding to the high-risk category.

[0122] The feature vector of the clause text corresponding to the high-risk category is obtained by the following formula:

[0123] (13)

[0124] In formula (13), Represents the feature vector of the clause text. This represents the sequence of entities output by the named entity recognition model. Represents the entity embedding matrix. express Model bias terms This represents a list of high-risk category clauses. The control logic of formula (13) is the refined feature extraction logic for high-risk clauses. Its core value is: 1. Entity-driven feature extraction: Unlike the general BiLSTM intent extraction, it focuses on key compliance entities in the clauses (such as amount, subject, authority, etc.), providing more granular features than global intent; 2. Precise high-risk positioning: through The model identifies the core risk elements in high-risk clauses, providing a data foundation for subsequent accurate alerts and rectification; 3. Structured model output: generates fixed-dimensional feature vectors of clause text. This facilitates accurate matching and risk classification in the subsequent risk rule base.

[0125] For key clauses in lease agreements, if they are determined to be in a high-risk category such as "liability for breach of contract," the automated compliance review system first uses a named entity recognition (NAME) model to extract textual features of the clauses. Specifically, a NAME model is a natural language processing technology based on deep learning that can identify specific entities in text. For example, in a clause stating "the lessee must compensate for losses," the NAME model uses a trained neural network, such as the BERT architecture, to scan the text sequence and mark entities such as "lessee" as the liable party and "compensate for losses" as the description of the obligation. This outputs a feature set, such as a list of entity labels and location information. These features capture the core semantic elements of the clause, providing basic data for subsequent matching.

[0126] S320. Input the feature vector of the clause text into the preset rule engine to match the set of legal provisions in the pre-built knowledge base. The set of legal provisions is obtained through semantic similarity calculation.

[0127] The highest matching degree between the feature vector of the clause text and the feature vector of the legal provisions in the knowledge base is obtained by the following formula:

[0128]

[0129] In formula (14), Indicates the highest matching degree. This represents the feature vector of a specific legal provision in the knowledge base. This represents the function that takes the maximum value. Represents the feature vector of the clause text With the feature vector of a certain legal provision in the knowledge base The cosine similarity. The control logic of formula (14) is the precise matching logic between high-risk clauses and the regulatory knowledge base. Its core value is: 1. Precise matching: semantic association is quantified by cosine similarity, avoiding the limitations of keyword matching, and can identify synonyms and variant clauses; 2. Highest similarity screening: only the most relevant regulatory clauses are retained, avoiding interference from irrelevant information and improving the efficiency of compliance review; 3. Strong interpretability: the matching degree value intuitively reflects the degree of association between the clause and the regulation, which is convenient for subsequent risk rating and manual review.

[0130] The set of matched legal provisions is obtained using the following formula:

[0131]

[0132] In formula (15), Represents a collection of legal provisions. This represents the threshold for regulatory matching similarity. The control logic of formula (15) is the batch matching logic between high-risk clauses and the regulatory knowledge base. Its core value is: 1. Precise screening: Through threshold filtering, only regulatory clauses that are highly semantically related to high-risk clauses are retained, avoiding interference from irrelevant information; 2. Collective output: The matching results are encapsulated into a structured set, which is convenient for subsequent automated processing (such as risk rating, generation of rectification suggestions, and writing of review reports); 3. Strong scalability: It can be adjusted It balances recall and precision to adapt to compliance scenarios in different industries.

[0133] The textual features of the clauses are input into a preset rule engine to match legal provisions in a pre-built knowledge base. The rule engine is an automated decision-making system that processes the input through logical rules and pattern matching. For example, the engine loads laws stored in the knowledge base, such as Article 114 of the Contract Law regarding breach of contract compensation, and then calculates semantic similarity. The specific process involves converting the feature vector into an embedded representation and using the cosine similarity formula to evaluate the similarity. If the angle between the clause feature vector and the law vector is less than a preset threshold of 0.2, a match is determined, thereby obtaining a list of relevant legal provisions to ensure the accuracy and comprehensiveness of the analysis.

[0134] S330. Use the textual implication model to perform logical conflict analysis on the set of legal provisions and the feature vectors of the clause texts to obtain the comparison results, which include conflicting segments.

[0135] The output of the text entailment model is obtained using the following formula:

[0136]

[0137] In formula (16), This represents the output of the text entailment model. Represents a natural language reasoning model. This represents the implied probability. Indicates a neutral probability. The formula (16) represents the probability of conflict. The control logic of the formula (16) is the logic of logical conflict analysis between high-risk clauses and legal provisions. Its core value is: 1. Logical compliance verification: It upgrades from semantic matching to logical reasoning, which can identify complex scenarios such as "partial compliance" and "implicit conflict", and is more accurate than simple similarity matching; 2. Probabilistic output: The three types of probabilities intuitively reflect the degree of compliance, which is convenient for risk classification and confidence assessment; 3. Conflict location: It directly outputs conflict fragments, which greatly reduces the cost of manual investigation and improves the efficiency of rectification.

[0138] The set of conflicting fragments is derived using the following formula:

[0139] (17)

[0140] In formula (17), Represents a set of conflicting fragments. Indicates a single legal provision. Representing fragments The probability of conflict, The threshold for conflict determination is indicated. The control logic of formula (17) is the precise screening and aggregated output logic for compliance conflicts. Its core value is: 1. High confidence screening: Through threshold filtering, only conflict fragments judged by the model as having high confidence are retained, avoiding interference from low confidence noise; 2. Structured output: The conflict content is encapsulated into a set form, which is convenient for subsequent automated processing (such as generating rectification suggestions, risk rating, and writing review reports); 3. Strong interpretability: Directly locates conflict fragments, making the compliance review results transparent and traceable, and greatly reducing the cost of manual investigation.

[0141] A textual implication model is used to perform logical conflict analysis on the features of the legal provisions and clauses to obtain comparison results. The textual implication model is a natural language reasoning tool, such as a Transformer-based model. When inputting a pair of legal provisions and clauses, the textual implication model uses an attention mechanism to determine the implication relationship. For example, if the legal provisions require "immediate compensation" while the clauses state "in installment compensation", the textual implication model outputs conflict labels and fragments such as "inconsistent compensation time limit". The results include specific conflict fragments to highlight the inconsistencies.

[0142] S340. Identify non-compliant items based on the comparison results and obtain the corresponding penalty amount. Assess the number and severity of non-compliant items based on the penalty amount.

[0143] The severity of non-compliance is determined by the following formula:

[0144]

[0145] In formula (18), Indicates the severity of non-compliance items. Indicates non-compliant items The corresponding amount of the penalty for violation, The weighting indicates the severity of the risk. The control logic of formula (18) is the quantitative assessment logic of the severity of compliance risk. Its core values ​​are: 1. Quantifying risk: transforming the abstract "compliance conflict" into a calculable value, realizing the quantification and comparability of risk; 2. Punishment-oriented: using the penalty for violations as the core indicator, directly reflecting the economic consequences of violations, and aligning with the actual risk concerns of enterprises; 3. Adjustable weighting: through 4. Aggregated calculation: Supports the aggregation of risks from multiple conflicting segments to comprehensively assess the overall risk level of a single clause.

[0146] The number of non-compliant items is calculated using the following formula:

[0147] (19)

[0148] In formula (19), Indicates the number of non-compliant items. This represents the number of elements in the set of conflict fragments. The control logic of formula (19) is the statistical logic of the number of compliance conflicts. Its core value is: 1. Intuitive quantification: The number of non-compliant items is directly counted using the set cardinality, realizing the intuitive quantification of compliance risk; 2. Simple and efficient: The calculation logic is extremely simple, requiring no complex calculations, and the results can be obtained quickly; 3. Risk auxiliary assessment: Similar to the severity of formula (18) In coordination, a comprehensive assessment of compliance risks should be conducted from two dimensions: "quantity" and "severity".

[0149] Based on the comparison results, non-compliant items are identified and corresponding penalty amounts are obtained. For example, conflicting segments are extracted from the results, mapped to a penalty database to obtain the amount, such as 5000 yuan. Then, the number of non-compliant items is assessed, such as 3 items, and the severity of each non-compliant item is evaluated, such as Level 1: minor non-compliance requiring only text correction. A weighted score is then used, such as multiplying the number by a severity coefficient of 0.8, to obtain the total risk value, thus facilitating contract optimization for users. This method achieves efficient risk assessment.

[0150] Preferably, the automated compliance review workflow generation method based on large model intelligent agent orchestration provided in this embodiment includes step S400 as follows:

[0151] S410. If the cumulative number of non-compliance items exceeds a preset threshold, extract the review context containing historical review trajectories.

[0152] Extract the review context using the following formula:

[0153] (20)

[0154] In formula (20), Represents the review context vector. Represents the feature vector of the historical review trajectory. The control logic of formula (20) is the review context aggregation logic in high-risk scenarios. Its core values ​​are: 1. Full-dimensional information aggregation: integrate the current review results with the historical trajectory to form a complete context including "intent-clause-conflict-severity-history"; 2. Temporal risk perception: introduce historical review trajectories to support the identification of recurring compliance issues and risk evolution trends; 3. Structured output: encapsulate context information in vector form to facilitate subsequent automated analysis and model input.

[0155] When the number of non-compliance items accumulated in the automated compliance review system exceeds a preset threshold, such as 5 items, the automated compliance review system automatically extracts the review context containing historical review trajectories. For example, it pulls the review records of previous lease contracts from the database, including the history of clause modification and risk labeling logs, thereby providing a complete background for subsequent processing.

[0156] S420. Transform the review context into a model input vector.

[0157] The model input vector is obtained using the following formula:

[0158]

[0159] In formula (21), Represents the model input vector. Indicates a context-embedded function, The context embedding weight matrix represents the weight matrix. This represents the context embedding bias term. The control logic of formula (21) is the standardized transformation logic from the review context to the model input. Its core value is: 1. Standardized adaptation: It maps heterogeneous review contexts (text, numerical values, historical trajectories) into dense vectors to adapt to various downstream models; 2. Feature learning: It uses a trainable context embedding weight matrix. With context embedding bias 3. Highly scalable: Automatically learns the optimal representation of contextual features, improving the performance of downstream tasks; It can be flexibly replaced (such as using BERT, LSTM, linear layers, etc.) to adapt to different scenario requirements.

[0160] Transforming the review context into a model input vector involves natural language processing techniques. The review context, such as the review trajectory in text form, is first split into word sequences using a word segmentation tool. Then, a word embedding model, such as Word2Vec, is used to map each word into a high-dimensional vector. For example, when processing the trajectory of the "liability for breach of contract" clause, the automated compliance review system identifies keywords such as "compensation" and "time limit" and generates vector representations. These vectors capture semantic relationships and are aggregated into a single input vector through average pooling operations, ensuring that the model can understand the overall meaning of the context. This transformation process helps to preserve the structured features of historical data, providing a reliable foundation for dynamic generation.

[0161] S430. Input the model input vector into the sequence to generate a model and generate a dynamic step sequence.

[0162] The dynamic step sequence is generated using the following formula:

[0163] (twenty two)

[0164] In formula (22), Represents a dynamic sequence of steps. Represents a sequence generation model. Indicates the first The step vector of the step, Indicates the first The step vector of the steps. The control logic of formula (22) is the sequence generation logic from compliance review results to automated handling process. Its core value is: 1. Dynamic adaptation: Based on the input risk context, it generates handling steps with appropriate length and content to avoid the rigidity of fixed processes; 2. End-to-end generation: It generates executable steps directly from the review context without human intervention, improving the efficiency of compliance handling; 3. Structured output: The sequence form is easy for machine parsing and execution, and also easy for human understanding and review.

[0165] No. Step output The conditional probability is derived using the following formula:

[0166]

[0167] In formula (23), Indicates the first Step output The conditional probability, This represents the Softmax activation function. This represents the output of the hidden layer of the generative model. Indicates the weights of the generative model. This represents the bias term. The control logic of formula (23) is the autoregressive probability prediction logic of the sequence generation model. Its core value is: 1. Dynamic generation constraints: By modeling conditional probability, the generated step sequence is guaranteed. 1. Semantically coherent and logically feasible, avoiding the generation of meaningless steps; 2. Context-aware: incorporating review context. 1. Based on historical steps, the generated actions can accurately respond to the current compliance risk status; 2. Probabilistic output: Softmax provides the confidence level of the generated actions, which can be used to control the rigor of the generation (such as selecting the highest probability in high-risk scenarios and random sampling to increase diversity in low-risk scenarios).

[0168] The input vector of the model is input into the sequence generation model to generate a dynamic sequence of steps. The sequence generation model is an architecture based on recurrent neural networks such as LSTM, which can predict the next action sequence based on the input vector. For example, after inputting the review vector of the lease contract, the sequence generation model calculates the hidden state layer by layer and outputs a sequence of steps such as "check the rent terms - verify compliance - generate a report", thereby adapting to the review needs of different contract scenarios.

[0169] S440. Parse the dynamic step sequence to obtain data index information, and retrieve the step execution parameters based on the data index information to obtain the input data for each step in the dynamic step sequence.

[0170] The input data for each step in the dynamic step sequence is obtained using the following formula:

[0171] (twenty four)

[0172] In formula (24), Indicates the first Step input data, Indicate steps Data index information, This represents a data retrieval function. This indicates the review of the data knowledge base. The control logic of formula (24) is the standardized acquisition logic from generation steps to execution data. Its core value is: 1. Data-driven execution: It transforms the abstract generation steps into specific executable tasks and provides the complete input required for each step; 2. Precise data supply: It accurately pulls relevant data from the knowledge base through index matching, avoiding interference from invalid information; 3. Process automation: It supports the fact that every action output by the sequence generation model can be actually executed by the AI ​​agent, forming a complete closed loop.

[0173] The dynamic step sequence is parsed to obtain data index information. Based on this data index information, step execution parameters are retrieved to obtain input data for each step in the dynamic step sequence. For example, the index ID 001 corresponds to "check rental terms" in the dynamic step sequence. The automated compliance review system then retrieves specific data from the parameter library, such as rental calculation formulas and historical default cases, thereby providing precise input for each step and achieving automated optimization of the review process. This approach improves the efficiency of contract risk management.

[0174] Furthermore, the automated compliance review workflow generation method based on large model intelligent agent orchestration provided in this embodiment includes step S500 as follows:

[0175] S510. Parse the dynamic step sequence to extract the dependency identification data that defines the association between nodes, and map the dependency identification data to the node topology model to calculate the execution time sequence weight value.

[0176] The execution time weight values ​​are obtained using the following formula:

[0177]

[0178] In formula (25), This indicates the execution timing weight value. Indicate steps For steps Dependence, The total dependency is represented by the control logic of formula (25), which is the quantification and sorting logic of the logical relationship between compliance processing steps. Its core value is: 1. Quantification of temporal logic: Transforming the semantic dependency relationship between steps into calculable numerical weights to realize the automatic arrangement of the execution order of steps; 2. Critical path identification: Through the weight size, the key steps with high weight and high importance can be accurately located to ensure that the core compliance review tasks are executed first; 3. Dynamic adaptation: The weight calculation is based on the dynamically generated step sequence, which can be adjusted in real time as the content of the steps changes to ensure the effectiveness of the temporal strategy.

[0179] The node topology model is derived using the following formula:

[0180] (26)

[0181] In formula (26), Represents a node topology model. The graph construction function is represented by formula (26). The control logic of formula (26) is the structured mapping logic from dynamic step sequence to executable topology process. Its core value is: 1. Process structuring: transforming linear step sequence into graph structure, clearly expressing the complex dependencies between steps; 2. Weight-driven scheduling: embedding temporal weights into graph edges, providing a quantitative basis for process scheduling and resource allocation; 3. Strong executability: the topology graph can be directly parsed by the scheduling engine to realize the automated execution of the compliance review process.

[0182] For the lease contract review process, the dynamic step sequence is first analyzed. For example, the dynamic step sequence includes "terms review - risk assessment - report generation". Dependency identification data is extracted from this sequence, such as "risk assessment" depending on the output of "terms review", thus defining the relationship between nodes. The dependency identification data is then mapped to a node topology model, which is a graph structure representation where nodes represent review steps and edges represent dependencies. For example, in processing housing lease contracts, the node topology model connects the "rent terms review" node with the "breach of contract liability assessment" node. By calculating edge weights, such as dependency strength based on historical data statistics, the execution sequence weight value is obtained. For example, if "rent terms" review usually needs to be performed first, its weight value is 0.8, indicating high priority. This mapping process involves traversing the topology graph and applying a weighted algorithm to evaluate the sequence. For example, starting from the starting node, the path weights are accumulated to calculate the quantitative value of the overall execution order, thus providing a basis for subsequent path construction.

[0183] S520. Construct a linear execution path containing serial and parallel logic based on the execution timing weight values, identify the blocking factors of nodes in the linear execution path and convert them into pre-start state constraints.

[0184] The linear execution path is derived using the following formula:

[0185] (27)

[0186] In formula (27), Indicates a linear execution path. This represents a weight-based ranking function. This represents the index of the sorted steps. The control logic of formula (27) is the quantitative sorting logic of dynamic step sequence to executable linear path. Its core value is: 1. Weight-driven priority: It transforms abstract dependency relationship into executable order, with high-weight steps executed first, ensuring the advancement of core compliance tasks; 2. Linearization and parallel compatibility: While maintaining the linear execution order, it can identify parallel executable steps through pre-state constraints, improving process efficiency; 3. Strong executability: The generated linear path can be directly parsed and executed by the workflow engine without additional manual intervention.

[0187] The pre-start state constraints are derived using the following formula:

[0188]

[0189] In formula (28), Indicates the first Pre-start state constraints. Indicate steps The blocking factor; Indicates the first Step by step; Indicates the first Step by step; This indicates an index constraint, limiting consideration to only those items in the execution path that occur earlier than... Execution steps ; This indicates a dependency decision condition, when Depends on At that time, Inclusion of blocking factors Preconditions. The control logic of formula (28) is the constraint expression logic of the inter-step dependencies in the linear execution path. Its core value is: 1. Precise constraint generation: Only retain the preconditions that actually exist as dependencies, avoid redundant constraints, and improve execution efficiency; 2. Temporal logic guarantee: Through 1. Ensure that the preceding steps are executed before the current steps to avoid process deadlock; 2. Strong executability: The constraint set can be directly parsed by the scheduling engine to achieve automated process control.

[0190] A linear execution path is constructed based on the execution sequence weight values. For example, steps with a weight higher than 0.5 are set as serial logic, such as "risk assessment" immediately after "terminal review," while steps with similar weights, such as "compliance check" and "legal consultation," are processed in parallel, forming a path such as "serial review - parallel check - serial report." Blocking factors for nodes in the linear execution path are identified. For example, if the "risk assessment" node lacks input data, it is considered a blockage and converted into a pre-start state constraint, such as requiring "terminal review" to complete its output.

[0191] S530. If the pre-start state constraints meet the verification requirements, a trigger control instruction is generated and embedded into the linear execution path to form a complete review chain.

[0192] The complete review chain is formed using the following formula:

[0193]

[0194] In formula (29), This indicates the complete review chain. Indicate steps Trigger control commands, The constraint verification function is represented. The control logic of formula (29) is the final encapsulation logic from the linear execution path to the automated execution review chain. Its core values ​​are: 1. Closed-loop control: The constraint verification is bound to the trigger command, and only the steps that meet the conditions can be executed, ensuring the correctness of the process logic; 2. Executability: The complete review chain can be directly parsed by the scheduling engine, realizing a closed loop from "step sequence" to "automated execution"; 3. Traceability: The trigger command of each step corresponds one-to-one with the constraint verification result, which is convenient for compliance audit and problem investigation.

[0195] If the pre-start state constraints meet the verification requirements, a trigger control instruction is generated. For example, after the constraint verification passes, the trigger control instruction, such as "start report generation", is embedded in the path to form a complete review chain, thereby ensuring a smooth process.

[0196] Preferably, the automated compliance review workflow generation method based on large model intelligent agent orchestration provided in this embodiment includes step S600 as follows:

[0197] S610. Traverse the complete review chain to capture node execution logs and status snapshots, and identify operational behavior data to form a set of original evidence fragments.

[0198] The original set of evidence fragments is derived using the following formula:

[0199] (30)

[0200] In formula (30), Represents the set of original evidence fragments. Indicate steps Execution log, Indicate steps The state snapshot. The control logic of formula (30) is the collection logic from the complete review chain to the traceable evidence library. Its core value is: 1. Full process traceability: covering the behavior and status of every step in the review chain, realizing "operation traceability and status traceability"; 2. Evidence integrity: collecting behavior logs and status snapshots at the same time, restoring the execution process from the two dimensions of "what was done" and "what the status became"; 3. Compliance guarantee: the original evidence fragments can be directly used for compliance audits and regulatory inspections, meeting the requirements of data retention and traceability.

[0201] The system traverses the entire review chain, capturing execution logs and status snapshots of each node. Specifically, in the accounts receivable review process of supply chain finance, the automated compliance review system automatically collects the operation timestamps, reviewer actions, and snapshots of the "Contract Authenticity Confirmation" document generated upon completion of the "Trade Background Verification" node. These data collectively constitute fragments of original evidence, such as recording that "Operator A called a third-party business registration data interface at a specific time and marked the contract seal as valid."

[0202] S620. Map the original set of evidence fragments to the pre-stored compliance rule base, and analyze the logical correspondence between the operation behavior and the clause to determine the set of judgment basis.

[0203] The set of judgment criteria is determined using the following formula:

[0204] (31)

[0205] In formula (31), This represents the set of criteria for judgment. Represents the evidence-rule mapping function. This indicates a pre-stored compliance rule base. The control logic of formula (31) is a structured mapping logic from original evidence to compliance judgment basis. Its core values ​​are: 1. Evidence ruleization: transforming scattered execution logs and status snapshots into judgment basis bound to compliance rules, making the review conclusion "based on evidence"; 2. Judgment standardization: unifying judgment standards through the pre-stored rule base, avoiding the subjectivity of manual review, and improving the consistency and accuracy of review results; 3. Traceability: the judgment basis is directly related to the original evidence and compliance rules, forming a complete traceable chain of "evidence → rule → basis → conclusion", which meets the requirements of compliance audit.

[0206] The logical correspondence between operational actions and terms is analyzed using the following formula:

[0207] (32)

[0208] In formula (32), This indicates the logical correspondence between operational actions and clauses. This represents the association mapping function. The control logic of formula (32) is the semantic association logic from operational behavior data to compliance clauses. Its core values ​​are: 1. Interpretability: It binds abstract operational behaviors with specific compliance clauses, making the review results "based on evidence"; 2. Compliance accuracy: It accurately locates the compliance requirements corresponding to each operation, avoiding generalities; 3. Audit friendliness: The logical correspondence can be directly used as the core evidence for compliance audit, proving the compliance or violation of the operation.

[0209] The original set of evidence fragments is mapped to a pre-stored compliance rule base, which stores clauses such as "the verification of the authenticity of trade background must include verification of the business registration information of the contract signatories." The automated compliance review system parses the content of the "Contract Authenticity Confirmation Letter" using natural language processing, extracts key fields such as "signatory name" and "verification result," and logically matches them with the clauses in the rule base to determine whether the operation constitutes a set of criteria that meet the requirements of the clauses.

[0210] S630. Employ a multi-dimensional assessment matrix to quantify the set of judgment criteria, calculate the compliance matching degree, and aggregate to generate a risk distribution map.

[0211] The compliance matching degree is calculated using the following formula:

[0212]

[0213] In formula (33), Indicates compliance matching degree. Indicate the basis for judgment Compliance score, Indicate the basis for judgment The weight of the formula (33) is the control logic of the set of judgment criteria to the quantitative assessment logic of compliance matching degree. The core value is: 1. Quantitative assessment: transforming qualitative compliance judgment into quantitative matching degree value, which is convenient for intuitive comparison and decision-making; 2. Weighted fairness: distinguishing the importance of different criteria by weight, avoiding "one-size-fits-all", and more in line with actual business scenarios; 3. Normalized comparability: the normalized matching degree can be compared horizontally / vertically between different review objects and different time periods, improving the universality and comparability of the assessment.

[0214] The risk distribution map is generated by aggregating data using the following formula:

[0215] (34)

[0216] In formula (34), Represents the risk distribution map. This represents the heatmap generation function. The control logic of formula (34) is the aggregation generation logic from compliance quantitative data to risk visualization map. Its core value is: 1. Risk visualization: It transforms abstract compliance matching data into an intuitive heatmap, making the complex risk distribution clear at a glance; 2. Multi-dimensional aggregation: It supports multi-dimensional aggregation by business line, department, clause, etc., to meet the risk control perspectives of different levels; 3. Decision support: It provides intuitive data support for compliance resource allocation and rectification priority formulation.

[0217] A multi-dimensional evaluation matrix is ​​used to quantify the set of judgment criteria. This matrix includes multiple dimensions such as "completeness of evidence," "operational standardization," and "timeliness compliance." For example, regarding the judgment criterion of "verification of business information," the automated compliance review system assigns quantitative scores from 0 to 1 based on factors such as the authority of the data source it calls, the completeness of the verification steps, and whether it is completed within the prescribed processing time limit. Finally, a comprehensive compliance matching score, such as 0.85, is obtained through weighted calculation. After aggregating the matching scores of all nodes in the process, the risks of each stage are marked on the topology map with different color intensities (e.g., green = low risk, yellow = medium risk, red = high risk), forming a risk distribution map.

[0218] S640. Based on the anomaly markers in the risk distribution map, trace back the associated logical path and fill the report structure with the compliance matching degree to generate a compliance assessment report.

[0219] The compliance assessment report is generated using the following formula:

[0220] (35)

[0221] In formula (35), This indicates a compliance assessment report. This indicates the report population function. This indicates a fixed report structure. This represents the risk backtracking function. The control logic of formula (35) is the automatic generation logic of compliance reports based on risk data and logical relationships. Its core values ​​are: 1. Full-link traceability: The report content is directly linked to the risk map and operational behavior-clause logic, realizing the traceability of "risk → behavior → basis → clause"; 2. Report automation: It replaces manual writing, automatically fills in structured content and visual charts, and greatly improves the efficiency and consistency of report generation; 3. Decision readability: The structured report clearly shows the risk distribution, the root cause of the problem and the direction of rectification, which makes it easy for management and regulatory agencies to quickly understand the compliance status.

[0222] Based on the anomaly markers in the risk distribution map, the system traces back the associated logical path. For example, if the risk distribution map shows that the "Logistics Document Review" node has a matching degree of only 0.3 and is marked as a red anomaly, the automated compliance review system automatically traces back to the preceding node, "Goods Inbound Record Verification," to check the completeness and accuracy of its output data. Combining the specific compliance matching degree values ​​of these two nodes, the automated compliance review system fills the corresponding sections of the standardized report template with specific issues such as "missing key receipt information in logistics documents" and "discontinuous timestamps in inbound records," automatically generating a compliance assessment report that points out the specific deficiencies.

[0223] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. An automatic compliance review workflow generation method based on large model agent arrangement, characterized in that, Includes the following steps: S100. Extract the text feature vector of the input document through a document classification model, and obtain a list of key terms based on the text feature vector; S200. Perform semantic intent analysis on the list of key terms to determine the term type of each term and identify potential risk points; S300. If the clause type is a high-risk category, the rule engine will match the legal provisions in the knowledge base to obtain the comparison results and identify non-compliant items, and assess the number and severity of the non-compliant items. S400. When the number of non-compliant items exceeds a preset threshold, a sequence generation model is used to generate a dynamic step sequence based on the review context, and the input data of each step in the dynamic step sequence is obtained. S500. Determine the execution order and dependencies of the steps in the dynamic step sequence, and determine the triggering conditions of subsequent steps based on the dependencies to form a complete review chain. S600. Extract the judgment criteria from the complete review chain and generate a compliance assessment report; Step S400 includes: S410. If the cumulative number of non-compliance items exceeds a preset threshold, extract the review context containing historical review trajectories. S420. Convert the review context into a model input vector; S430. The model input vector is input to the sequence to generate a model to generate a dynamic step sequence; S440. Parse the dynamic step sequence to obtain data index information, and retrieve the step execution parameters according to the data index information to obtain the input data for each step in the dynamic step sequence; Step S500 includes: S510. Parse the dynamic step sequence to extract the dependency identification data that defines the association between nodes, and map the dependency identification data to the node topology model to calculate the execution time sequence weight value. S520. Construct a linear execution path containing serial and parallel logic based on the execution timing weight values, identify the blocking factors of nodes in the linear execution path and convert them into pre-start state constraints. S530. If the pre-start state constraint meets the verification requirements, a trigger control instruction is generated and embedded into the linear execution path to form a complete review chain.

2. The method of claim 1, wherein, Step S100 includes: S110. Obtain the input document, perform word segmentation and part-of-speech tagging on the input document to obtain a basic vocabulary set, and calculate the word frequency distribution matrix based on the basic vocabulary set; S120. Input the word frequency distribution matrix into the document classification model for semantic parsing to determine the text feature vector; S130. Calculate the set of similarity values ​​between the text feature vector and the benchmark clause vector library; S140. If the value in the similarity value set is greater than a preset threshold, then filter the terms in the benchmark term vector library to obtain a list of key terms.

3. The method of claim 1, wherein, Step S200 includes: S210. Use a bidirectional long short-term memory network to perform sequence modeling on the list of key terms to output an intent vector; S220. Input the intent vector into a multilayer perceptron classifier to determine the clause type of each clause, wherein the clause type is selected based on the probability distribution of a predefined semantic intent space; S230. Extract the responsible parties and the boundaries of obligations according to the clause type, and map the responsible parties and the boundaries of obligations to the risk rule base to obtain risk characteristic data; S240. If the deviation value in the risk characteristic data is higher than a preset threshold, then output the potential risk point containing the clause type and violation identifier.

4. The automated compliance review workflow generation method based on large-scale intelligent agent orchestration according to claim 1, characterized in that, Step S300 includes: S310. If the clause type is a high-risk category, then the named entity recognition model is used to extract the clause text feature vector corresponding to the high-risk category. S320. Input the feature vector of the clause text into a preset rule engine to match the set of legal provisions in a pre-built knowledge base, wherein the set of legal provisions is obtained through semantic similarity calculation. S330. A textual entailment model is used to perform logical conflict analysis on the set of legal provisions and the feature vectors of the clause texts to obtain a comparison result, the comparison result containing conflicting segments; S340. Identify non-compliant items based on the comparison results and obtain the corresponding penalty amount for violations, so as to assess the number and severity of the non-compliant items based on the penalty amount for violations.

5. The method of claim 1, wherein, Step S600 includes: S610. Traverse the complete review chain to capture node execution logs and status snapshots, and identify operational behavior data to form a set of original evidence fragments; The original set of evidence fragments is derived using the following formula: in, Represents the set of original evidence fragments. Indicate steps Execution log, Indicate steps State snapshot, This indicates the complete review chain; S620. Map the set of original evidence fragments to a pre-stored compliance rule base, and parse the logical correspondence between the operation behavior and the clause to determine the set of judgment basis; S630. Use a multi-dimensional evaluation matrix to quantify the set of judgment criteria, calculate the compliance matching degree, and aggregate to generate a risk distribution map; S640. Based on the anomaly markers in the risk distribution map, trace back the associated logical path and fill the report structure with the compliance matching degree to generate a compliance assessment report.

6. The automated compliance review workflow generation method based on large-model intelligent agent orchestration according to claim 5, characterized in that, In step S620, the set of judgment criteria is determined using the following formula: in, This represents the set of criteria for judgment. Represents the evidence-rule mapping function. This indicates a pre-stored compliance rule base; The logical correspondence between operational actions and terms is analyzed using the following formula: in, This indicates the logical correspondence between operational actions and clauses. Represents the associative mapping function. This indicates a list of key terms.

7. The automated compliance review workflow generation method based on large-model intelligent agent orchestration according to claim 6, characterized in that, In step S630, the compliance matching degree is obtained using the following formula: in, Indicates compliance matching degree. Indicate the basis for judgment Compliance score, Indicate the basis for judgment The weights; The risk distribution map is generated by aggregating data using the following formula: in, Represents the risk distribution map. This represents the heatmap generation function. Indicates the severity of non-compliance items. Indicates the number of non-compliant items.

8. The method of claim 7, wherein, In step S640, a compliance assessment report is generated using the following formula: in, This indicates a compliance assessment report. This indicates the report population function. This indicates a fixed report structure. This represents the risk backtracking function. This indicates the logical correspondence between operational actions and clauses.

Images