A method for constructing an audit model for related party transactions of an enterprise by using a knowledge graph

Abstract

This invention discloses a method for constructing an audit model for related-party transactions of enterprises using knowledge graphs, belonging to the field of related-party transaction auditing technology. The invention includes: collecting and preprocessing multi-source heterogeneous data; constructing a multi-layered knowledge graph for the film and television industry based on the preprocessed data; mining implicit relationships between enterprises, automatically discovering indirect relationships formed by equity holding on behalf of others and personnel crossover through a multi-layered path aggregation attention network; extracting transaction information from contract texts using natural language processing technology and aligning it with entities in the knowledge graph; establishing a dynamic cost reasonableness benchmark model; identifying anomalies and scoring risks in transactions for each film and television project; and generating a structured audit report. This invention, by constructing a multi-layered knowledge graph for the film and television industry and introducing a path aggregation attention network, can automatically mine implicit relationships between enterprises formed through indirect paths such as multi-layered equity nesting, personnel crossover, and historical cooperation.

Description

Technical Field

[0001] This invention belongs to the field of auditing related-party transactions of enterprises, and in particular relates to a method for constructing an auditing model of related-party transactions of enterprises using knowledge graphs. Background Technology

[0002] Related-party transactions refer to the transfer of resources, services, or obligations between an enterprise and its related parties. In auditing, the authenticity and fairness of related-party transactions have always been key areas of focus, as related parties may use non-market pricing to transfer benefits or manipulate profits. The identification of related parties is typically based on the definitions in the *Accounting Standards for Business Enterprises*, including direct or indirect control, joint control, significant influence relationships, and key management personnel and their close family members. In traditional auditing, the identification of related parties mainly relies on related-party relationship statements provided by the enterprise and auditors' inquiries into business registration information, manually identifying explicit related relationships through equity structure diagrams.

[0003] Co-investment is a common production model in the film and television industry, referring to a project operation method in which multiple investors jointly contribute capital, share risks, and share profits. Due to the large investment amounts and long production cycles of film and television projects, multiple companies and key creative personnel often participate. Under this model, the transaction structure is relatively complex, with potentially multi-layered nested equity relationships among investors, and key creative personnel may have overlapping positions or historical collaborations with investors. These indirectly formed connections may lead to implicit relationships, meaning that while there is no direct equity control in form, there is a potential for coordinating interests in substance.

[0004] Existing technologies for auditing joint investments in the film and television industry primarily face the challenge of uncovering implicit related-party relationships. Auditors typically rely on business registration information to identify direct shareholding or employment relationships. However, in practice, complex situations such as nominee shareholding, multi-layered nesting, and overlapping personnel make implicit relationships difficult to detect using conventional methods. For example, two companies may have a conflict of interest through joint investments in third-party projects, or implicit connections may exist through kinship among their respective senior executives—issues difficult to identify using traditional equity structure analysis. Due to the limitations in identifying related-party relationships, subsequent judgments on the fairness of transaction prices may lack a crucial foundation, leading to discrepancies between audit conclusions and actual circumstances. Therefore, the following solutions are proposed to address these issues. Summary of the Invention

[0005] The purpose of this invention is to provide a method for constructing an audit model of related-party transactions of enterprises using knowledge graphs. By constructing a multi-layered knowledge graph of the film and television industry and introducing a path aggregation attention network, it can automatically discover implicit relationships between enterprises formed through indirect paths such as multi-layered equity nesting, cross-appointment of personnel, and historical cooperation. This solves the problem that existing technologies rely on querying explicit relationships through business registration information and are unable to discover complex stakeholders.

[0006] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution:

[0007] This invention discloses a method for constructing an audit model for related-party transactions of an enterprise using knowledge graphs, comprising the following steps:

[0008] Step S1: Collect multi-source heterogeneous data and preprocess it. The multi-source heterogeneous data includes business registration data, film and television project filing data, main creator association data, contract texts and industry benchmark data.

[0009] Step S2: Construct a multi-layered knowledge graph for the film and television industry based on the preprocessed data. The knowledge graph includes enterprise entities, personnel entities, film and television project entities, contract entities, and various relationships between them.

[0010] Step S3: Employ graph neural network models to uncover implicit relationships between enterprises, and automatically discover indirect relationships formed by nominee shareholding and personnel overlap through multi-layer path aggregation attention networks;

[0011] Step S4: Extract transaction information from the contract text using natural language processing technology and align it with entities in the knowledge graph;

[0012] Step S5: Establish a dynamic cost rationality benchmark model for various costs in film and television production to assess whether the transaction price is fair;

[0013] Step S6: Combining the implicit relationships discovered in Step S3 and the cost benchmark in Step S5, perform anomaly identification and risk scoring for transactions in each film and television project.

[0014] Step S7: Generate a structured audit report based on the analysis results of step S6.

[0015] Furthermore, the preprocessing described in step S1 includes: standardizing the company names in the business registration data, disambiguating the names of personnel, performing OCR recognition and text cleaning on the contract text, and storing all the data uniformly and establishing an index.

[0016] Furthermore, the entity types of the knowledge graph described in step S2 include enterprise entities, personnel entities, film and television project entities, and contract entities, and the relationship types include equity investment relationships, employment relationships, investment relationships, production relationships, participation relationships, and contractual relationships. A unique ID is generated for each entity, and a timestamp index is established.

[0017] Furthermore, the graph neural network model described in step S3 is a multi-layer path aggregation attention network, which is specifically implemented as follows: extracting a subgraph for each film and television project and initializing a feature vector for each node; learning the implicit representation of nodes through multi-layer convolution using the graph attention network; introducing a path-level attention mechanism to calculate the weights of multiple paths between any two enterprise nodes and comprehensively considering the time decay factor to obtain the association strength score; and marking enterprise pairs with association strength scores higher than a preset threshold as implicit related parties.

[0018] Furthermore, the path-level attention mechanism enumerates all paths between nodes, inputs the representation vector of each path into the attention layer to calculate the weight, and uses a multilayer perceptron to output the path score, finally aggregating to obtain the association strength score between node pairs, where the time decay factor is calculated based on the time difference between the most recent relationship on the path and the current time.

[0019] Furthermore, the extraction of transaction information in step S4 includes: using a named entity recognition model to identify the company name, personnel name, amount, date, and project name in the contract text; using a relation extraction model to determine the relationship type between entities and extracting transaction triples; matching and aligning the extracted entities with nodes in the knowledge graph, creating temporary nodes for those that cannot be matched, and storing the transaction information as attributes of the contract entities.

[0020] Furthermore, the dynamic cost rationality benchmark model described in step S5 includes: extracting features affecting costs from a knowledge graph, collecting real cost data from historical contracts as a training set, training a machine learning regression model to obtain predicted values ​​and prediction intervals, and periodically updating the model based on new data.

[0021] Furthermore, the anomaly identification and risk scoring in step S6 specifically includes: for each transaction, determining whether there is an explicit or implicit relationship between the two parties; when there is a relationship, calculating the deviation between the actual transaction price and the predicted value of the benchmark model; calculating the transaction risk score by combining the relationship strength score and the deviation score; and aggregating the risks of all transactions in a film and television project to obtain a project-level risk score.

[0022] Furthermore, the deviation is calculated by dividing the difference between the actual transaction price and the benchmark forecast value by the benchmark forecast value, taking into account whether it exceeds the forecast range; the risk score is a weighted sum of the correlation strength score and the portion of the deviation that exceeds the allowable threshold.

[0023] Furthermore, the audit report described in step S7 includes basic project information, a list of hidden related parties, details of abnormal transactions, risk levels, and visualization charts, and supports exporting to various file formats for auditors to review.

[0024] The present invention has the following beneficial effects:

[0025] 1. This invention constructs a multi-layered knowledge graph and introduces a path-aggregating attention network, which can automatically uncover implicit relationships between enterprises formed through indirect paths such as multi-layered equity nesting, cross-appointment of personnel, and historical cooperation. This method can discover more potential stakeholders, thereby broadening the coverage of related-party transaction audits and enabling auditors to focus on more transaction entities that may have risks of transferring benefits, providing more comprehensive clues for subsequent substantive testing.

[0026] 2. This invention utilizes industry benchmark data to construct a dynamic cost rationality model and learns patterns in historical transaction data through machine learning algorithms, enabling it to automatically update the benchmark range based on market changes. This method avoids the judgment lag problem that may be caused by relying on fixed proportions or static industry standards in traditional auditing, allowing the fair assessment of transaction prices to be in line with the current market situation and improving the synchronization between audit judgment and industry practice.

[0027] 3. This invention integrates heterogeneous data from multiple sources, such as business registration data, project filing data, and contract texts, into a single knowledge graph, enabling information that was originally scattered across different systems to be interconnected and verified. For example, transaction information in a contract can be compared with the business registration equity structure, and personnel cooperation records can be cross-verified with project investment relationships. This multi-dimensional data association capability helps to discover hard-to-detect logical contradictions and provides richer factual evidence for audit conclusions.

[0028] 4. This invention automates multiple steps that originally required manual querying, comparison, and judgment by integrating data collection, graph construction, relationship mining, price comparison, and report generation. Auditors no longer need to spend a lot of time on information retrieval and preliminary screening, but can directly obtain the risk list and related paths output by the system, thereby focusing their energy on in-depth verification of high-risk areas and improving the overall efficiency of auditing work.

[0029] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0030] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a flowchart illustrating a method for constructing an audit model for related-party transactions of an enterprise using knowledge graphs, according to the present invention. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] Please see Figure 1 As shown, this invention provides a method for constructing an audit model for related-party transactions of an enterprise using knowledge graphs. The method includes the following steps:

[0034] Step S1: Collect and preprocess multi-source heterogeneous data, which includes business registration data, film and television project filing data, data related to key creative personnel, contract texts, and industry benchmark data.

[0035] Step S2: Construct a multi-layered knowledge graph for the film and television industry based on the preprocessed data. The knowledge graph includes enterprise entities, personnel entities, film and television project entities, contract entities, and various relationships between them.

[0036] Step S3: Employ graph neural network models to uncover implicit relationships between enterprises, and automatically discover indirect relationships formed by nominee shareholding and personnel overlap through multi-layer path aggregation attention networks;

[0037] Step S4: Extract transaction information from the contract text using natural language processing technology and align it with entities in the knowledge graph;

[0038] Step S5: Establish a dynamic cost rationality benchmark model for various costs in film and television production to assess whether the transaction price is fair;

[0039] Step S6: Combining the implicit relationships discovered in Step S3 and the cost benchmark in Step S5, perform anomaly identification and risk scoring for transactions in each film and television project.

[0040] Step S7: Generate a structured audit report based on the analysis results of step S6.

[0041] The preprocessing in step S1 includes: standardizing the company names in the business registration data, disambiguating the names of personnel, performing OCR recognition and text cleaning on the contract text, and storing all the data uniformly and creating an index.

[0042] In step S2, the entity types of the knowledge graph include enterprise entities, personnel entities, film and television project entities, and contract entities. The relationship types include equity investment relationships, employment relationships, investment relationships, production relationships, participation relationships, and contractual relationships. A unique ID is generated for each entity, and a timestamp index is established.

[0043] In step S3, the graph neural network model is a multi-layer path aggregation attention network. Its specific implementation includes: extracting a subgraph for each film and television project and initializing a feature vector for each node; learning the implicit representation of nodes through multi-layer convolution using the graph attention network; introducing a path-level attention mechanism to calculate the weights of multiple paths between any two enterprise nodes and comprehensively considering the time decay factor to obtain the association strength score; and marking enterprise pairs with association strength scores higher than a preset threshold as implicit related parties.

[0044] The path-level attention mechanism enumerates all paths between nodes, inputs the representation vector of each path into the attention layer to calculate the weight, and uses a multilayer perceptron to output the path score. Finally, it aggregates to obtain the association strength score between node pairs, where the time decay factor is calculated based on the time difference between the most recent relationship on the path and the current time.

[0045] Step S4 involves extracting transaction information as follows: using a named entity recognition model to identify the company name, personnel name, amount, date, and project name in the contract text; using a relation extraction model to determine the relationship type between entities and extracting transaction triples; matching and aligning the extracted entities with nodes in the knowledge graph, creating temporary nodes for those that cannot be matched, and storing the transaction information as attributes of the contract entities.

[0046] The dynamic cost rationality benchmark model in step S5 includes: extracting features affecting costs from the knowledge graph, collecting real cost data from historical contracts as a training set, using a machine learning regression model to train and obtain predicted values ​​and prediction intervals, and updating the model regularly based on new data.

[0047] The anomaly identification and risk scoring in step S6 specifically includes: for each transaction, determining whether there is an explicit or implicit relationship between the two parties; when a relationship exists, calculating the deviation between the actual transaction price and the predicted value of the benchmark model; calculating the transaction risk score by combining the relationship strength score and the deviation score; and aggregating the risks of all transactions in a film and television project to obtain a project-level risk score.

[0048] The deviation is calculated by dividing the difference between the actual transaction price and the benchmark forecast by the benchmark forecast, taking into account whether it exceeds the forecast range; the risk score is the weighted sum of the correlation strength score and the portion of the deviation that exceeds the allowable threshold.

[0049] The audit report in step S7 includes basic project information, a list of hidden related parties, details of abnormal transactions, risk levels, and visualization charts, and supports exporting to various file formats for auditors to review.

[0050] One specific application of this embodiment is:

[0051] Step S1: Multi-source heterogeneous data acquisition and preprocessing

[0052] Collect the following four types of data and perform structured processing:

[0053] Business registration data: Business registration information of film and television industry-related companies is obtained from the National Enterprise Credit Information Publicity System and third-party commercial databases, including company name, unified social credit code, shareholders and shareholding ratio, senior executives, foreign investment records, change records, etc.; the data is stored in tabular form, and each record includes fields such as company ID, shareholder ID, shareholding ratio, and start date of employment.

[0054] Film and television project filing data: Obtain film and television project information from the filing and publicity system, including project name, filing unit, list of investors, list of production companies, list of main creative personnel (director, screenwriter, main actors), project status (pre-production, filming, post-production, release), and estimated investment amount; the data is stored in JSON format and the entity relationships need to be parsed out;

[0055] Key creative personnel association data: Collect historical work collaboration records of key creative personnel from film and television industry databases such as Maoyan Professional Edition and Lighthouse Professional Edition, including personnel names, projects they have participated in, roles they have played, and collaborating companies, in order to build a personnel collaboration network;

[0056] Contract texts and industry benchmark data: Collect anonymized texts of investment contracts, production contracts, actor remuneration contracts, etc. related to film and television projects, as well as various production cost benchmark data obtained from industry reports, such as the remuneration range of actors of different levels, the average cost per minute of special effects production, and the proportion of costume, prop and set expenses; industry benchmark data should be stored by project type (film / TV series), genre (historical / modern), and year.

[0057] Preprocessing steps: Standardize company names in business registration data to eliminate abbreviation ambiguity; disambiguate personnel names (identifying different people with the same name) and associate them using ID numbers or unique identifiers; perform OCR recognition (if scanned copies are available) and text cleaning on contract texts to remove irrelevant information; store all data in a unified Hadoop distributed file system and create an index for subsequent steps.

[0058] Step S2: Construction of a multi-layered knowledge graph for the film and television industry

[0059] Based on the preprocessed data, a knowledge graph containing entities and relationships is constructed and stored in a graph database (such as Neo4j); entity types include:

[0060] Corporate entities: Attributes include company ID, name, registered capital, and date of establishment;

[0061] Personnel entity: Attributes include personnel ID, name, date of birth, occupation, etc.;

[0062] Film and television project entities: Attributes include project ID, name, filing number, type, filing time, etc.;

[0063] Contract entity: Attributes include contract ID, signing date, amount, subject matter, etc.; Relationship types include:

[0064] Equity investment relationship: The shareholding relationship between Company A and Company B, including the shareholding ratio and the start date of shareholding;

[0065] Employment Relationship: The employment relationship between an individual and an enterprise, including attributes such as job title and start date of employment;

[0066] Investment relationship: The investment relationship between a company and a film and television project, including the investment amount and investment ratio;

[0067] Production Relationship: The production relationship between a company and a film or television project, including attributes such as production costs;

[0068] Participation relationship: The participation relationship between an individual and a film or television project, including attributes such as role and remuneration;

[0069] Contractual relationship: The contractual relationship between the contract and the contracting party (company or individual), including attributes such as the signing date and amount;

[0070] Construction process: Using the batch import tool of the graph database, entities and relationships are imported in CSV format, and a unique ID is generated for each entity; at the same time, a timestamp index is created to take time decay into account later.

[0071] Step S3: Implicit Relationship Mining Based on Graph Neural Networks

[0072] The core of this step lies in proposing a multi-layered path aggregation attention network to automatically discover potential implicit relationships between enterprises, especially those formed through indirect paths such as nominee shareholding, personnel overlap, and historical cooperation. The algorithm flow is as follows:

[0073] Step S31, Subgraph Extraction: For each film and television project, extract the subgraph within a certain range around it, including all investors, production companies and their related companies (connected by equity and employment relationships), as well as relevant key creative personnel and their collaborating personnel and companies; the subgraph range is limited by the number of hops (e.g., 3 hops) to reduce the amount of computation;

[0074] Step S32, Node Feature Initialization: Generate an initial feature vector for each node; enterprise node features include registered capital, years of establishment, one-hot encoding of industry classification, number of historical investment projects, etc.; personnel node features include years of service, number of projects participated in, awards received, etc.; film and television project node features include investment scale, type, filing time, etc.; all features are normalized and concatenated into an initial vector. ;

[0075] Step S33, Graph Convolutional Layer: Multiple convolutional layers are performed using a Graph Attention Network (GAT) to learn the implicit representations of nodes; in the... Layers, nodes The representation is updated as follows:

[0076]

[0077] In the formula, For nodes In the The feature vector representation of a layer; For activation functions; For nodes The neighbor node index; For nodes The set of neighboring nodes; For the first Layer nodes For neighboring nodes The attention coefficient represents right Importance weights; For the first The trainable weight matrix of the layer; For neighboring nodes In the Feature vector representation of a layer; attention coefficient The calculation is as follows:

[0078]

[0079] In the formula, It is an exponential function; For a leaky linear rectifier unit, an activation function that allows small gradients for negative values; For the first The transpose of the attention parameter vector of the layer; This is a vector concatenation operation; For nodes The vector obtained by linear transformation of the upper-level features; For neighboring nodes The vector obtained by linear transformation of the upper-level features; For any neighboring node The vector obtained by linear transformation of the upper-level features;

[0080] Step S34, Path Aggregation Layer: To capture the semantics of multi-hop paths, path-level attention is introduced; for any two enterprise nodes... and Enumerate all paths between them (length not exceeding L), for each path The representation is obtained by concatenating or averaging the representations of nodes on the path; then the weight of each path is calculated:

[0081]

[0082] In the formula, For path Attention weights; The transpose of the trainable parameter vector is used to map the hidden representation to a scalar; It is the hyperbolic tangent activation function; is a trainable weight matrix; For path The representation vector; It is a trainable bias vector; For nodes and nodes The set of all paths between; For a path index in the set; final node pair The association strength score is:

[0083]

[0084] In the formula, For nodes and nodes The score of the strength of the implicit association between them; For path Attention weights; For path representation of multilayer perceptron The mapping result outputs a scalar representing the contribution of the path to the association score; MLP is a multilayer perceptron; simultaneously, considering a time decay factor, more weights are given to recent equity changes or collaborations: the weight of each path is multiplied by... ,in, The time difference between the most recent relationship on the path and the current time. The attenuation coefficient;

[0085] Step S35, Model Training and Inference: Using known explicit associations (such as direct shareholding) as positive samples, randomly sample negative samples to train the MPAAN model. The optimization objective is to maximize the difference between the scores of positive and negative samples. After training, calculate the association strength score for any pair of companies and set a threshold. (e.g., 0.8), scores higher than For companies marked as implicit related parties, output a list.

[0086] Step S4: Extracting contract text information and aligning it with transaction records

[0087] Natural language processing techniques are used to extract key transaction information from contract texts and associate it with entities in a knowledge graph; specifically, this includes:

[0088] Step S41, Named Entity Recognition: The BERT-BiLSTM-CRF model is used to recognize entities such as company names, personnel names, amount figures, dates, and project names in the contract text; the model is fine-tuned on the annotated contract corpus;

[0089] Step S42, Relationship Extraction: For the identified entity pairs, use the pre-trained BERT model to determine the relationship type between the entities, such as "contracting party", "transaction amount", "transaction target", etc.; for example, for "Party A, Company A, pays Party B, Actor B, a fee of 10 million yuan", extract the triples (Company A, payment, Actor B) and (Actor B, fee, 10 million yuan).

[0090] Step S43, Transaction Record Alignment: Match the extracted transaction entities with nodes in the knowledge graph; align company names with standard names in business registration data using fuzzy matching algorithms (such as edit distance); align personnel names with ID numbers or unique identifiers; align project names with project names in the filing data; if a match cannot be found, create a temporary node for subsequent manual confirmation; after alignment, store the transaction information as attributes of the contract entity and establish the relationship between the contract and related companies, personnel, and projects.

[0091] Step S5: Construct a dynamic cost rationality benchmark model

[0092] A dynamic benchmark model is established to assess the fairness of transaction prices for various costs in film and television production (actor salaries, special effects costs, costume and prop costs, etc.); taking actor salaries as an example:

[0093] Step S51, Feature Engineering: Extract features affecting actor remuneration from the knowledge graph, including: actor's years of service, total box office revenue of past works, number of awards (Golden Rooster Award, Hundred Flowers Award, etc.), current popularity (number of Weibo followers), project type (movie / TV series), project theme (historical / modern), investment scale, etc.; standardize all features;

[0094] Step S52, Baseline Model Training: Collect real salary data from historical contracts as the training set, and train an XGBoost regression model to predict a reasonable salary range; the model outputs predicted values. and prediction interval For example, upper and lower limits can be obtained through quantile regression; similarly, for special effects production costs, a similar model can be established based on features such as special effects duration and complexity (e.g., CGI proportion).

[0095] Step S53, Dynamic Update: The model is retrained periodically (e.g., monthly) with new data to adapt to market changes.

[0096] Step S6: Identification and Risk Scoring of Abnormal Related Transactions

[0097] Based on the implicit relationships discovered in step S3 and the cost benchmark in step S5, the transactions in each film and television project are audited; specific process:

[0098] Step S61, Relationship Determination: For each transaction (such as an actor's salary contract), first obtain the node IDs of the two parties (such as the actor and the production company) in the knowledge graph; if the two parties have a direct equity relationship or employment relationship, it is an explicit relationship; otherwise, query the list of implicit related parties obtained in Step S3, and if the relationship strength score between the two parties is... If so, it is considered a implicit association;

[0099] Step S62, Price Deviation Calculation: If there is a relationship (explicit or implicit) between the two parties to the transaction, the actual transaction price is calculated. Compared with benchmark price Deviation:

[0100]

[0101] Meanwhile, if the actual price exceeds the predicted range If so, it is marked as an abnormal price;

[0102] Step S63, Risk Scoring: Calculate the transaction risk score by combining the correlation strength score and the deviation score.

[0103]

[0104] In the formula, , The weighting coefficients are set through expert experience or learned from historical data. This represents the permissible deviation threshold; if the actual price is lower than the benchmark, it may involve the transfer of benefits, but overstating costs is usually more common, so only positive deviations are considered; cases below the benchmark can be marked separately.

[0105] Step S64, Project-level Risk Aggregation: For all transactions in a film and television project, calculate the average risk score and list all high-risk transactions; at the same time, consider indicators such as the overall return on investment of the project to assist in the judgment;

[0106] Step S7: Generate Audit Report

[0107] The above analysis results are summarized to generate a structured audit report; the report content includes:

[0108] Basic project information: Name, registration number, investors, production company;

[0109] List of hidden related parties: Lists potential hidden related companies and individuals discovered through MPAAN, along with examples of related paths;

[0110] Abnormal Transaction Details: For each marked transaction, the following information is displayed: the parties involved in the transaction, the transaction amount, the benchmark price range, the deviation, the risk score, and the type of relationship.

[0111] Risk Level: Based on the project-level risk score, three risk levels are given: low, medium, and high.

[0112] Visual charts: Provide knowledge graph subgraphs and highlight abnormal transaction paths;

[0113] The report is output in PDF format and can also be exported to Excel for further review by auditors.

[0114] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0115] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for constructing an audit model for related-party transactions of an enterprise using knowledge graphs, characterized in that, The method includes the following steps: Step S1: Collect multi-source heterogeneous data and preprocess it. The multi-source heterogeneous data includes business registration data, film and television project filing data, main creator association data, contract texts and industry benchmark data. Step S2: Construct a multi-layered knowledge graph for the film and television industry based on the preprocessed data. The knowledge graph includes enterprise entities, personnel entities, film and television project entities, contract entities, and various relationships between them. Step S3: Employ graph neural network models to uncover implicit relationships between enterprises, and automatically discover indirect relationships formed by nominee shareholding and personnel overlap through multi-layer path aggregation attention networks; Step S4: Extract transaction information from the contract text using natural language processing technology and align it with entities in the knowledge graph; Step S5: Establish a dynamic cost rationality benchmark model for various costs in film and television production to assess whether the transaction price is fair; Step S6: Combining the implicit relationships discovered in Step S3 and the cost benchmark in Step S5, perform anomaly identification and risk scoring for transactions in each film and television project. Step S7: Generate a structured audit report based on the analysis results of step S6.

2. The method for constructing an audit model for related-party transactions of an enterprise using knowledge graphs according to claim 1, characterized in that, The preprocessing described in step S1 includes: standardizing the company names in the business registration data, disambiguating the names of personnel, performing OCR recognition and text cleaning on the contract text, and storing all the data uniformly and creating an index.

3. The method for constructing an audit model for related-party transactions of an enterprise using knowledge graphs according to claim 1, characterized in that, The entity types of the knowledge graph mentioned in step S2 include enterprise entities, personnel entities, film and television project entities, and contract entities. The relationship types include equity investment relationships, employment relationships, investment relationships, production relationships, participation relationships, and contractual relationships. A unique ID is generated for each entity, and a timestamp index is established.

4. The method for constructing an audit model for related-party transactions of an enterprise using knowledge graphs according to claim 1, characterized in that, The graph neural network model described in step S3 is a multi-layer path aggregation attention network. Its specific implementation includes: extracting a subgraph for each film and television project and initializing a feature vector for each node; learning the implicit representation of nodes through multi-layer convolution using the graph attention network; introducing a path-level attention mechanism to calculate the weights of multiple paths between any two enterprise nodes and comprehensively considering the time decay factor to obtain the association strength score; and marking enterprise pairs with association strength scores higher than a preset threshold as implicit related parties.

5. The method for constructing an audit model for related-party transactions of an enterprise using knowledge graphs according to claim 4, characterized in that, The path-level attention mechanism enumerates all paths between nodes, inputs the representation vector of each path into the attention layer to calculate the weight, and uses a multilayer perceptron to output the path score. Finally, it aggregates the scores to obtain the association strength score between node pairs. The time decay factor is calculated based on the time difference between the most recent relationship on the path and the current time.

6. The method for constructing an audit model for related-party transactions of an enterprise using knowledge graphs according to claim 1, characterized in that, Step S4, which involves extracting transaction information, includes: using a named entity recognition model to identify the company name, personnel name, amount, date, and project name in the contract text; using a relation extraction model to determine the relationship type between entities and extracting transaction triples; matching and aligning the extracted entities with nodes in the knowledge graph, creating temporary nodes for those that cannot be matched, and storing the transaction information as attributes of the contract entities.

7. The method for constructing an audit model for related-party transactions of an enterprise using knowledge graphs according to claim 1, characterized in that, The dynamic cost rationality benchmark model described in step S5 includes: extracting features affecting costs from a knowledge graph, collecting real cost data from historical contracts as a training set, training a machine learning regression model to obtain predicted values ​​and prediction intervals, and periodically updating the model based on new data.

8. The method for constructing an audit model for related-party transactions of an enterprise using knowledge graphs according to claim 1, characterized in that, The anomaly identification and risk scoring in step S6 specifically includes: for each transaction, determining whether there is an explicit or implicit relationship between the two parties; when there is a relationship, calculating the deviation between the actual transaction price and the predicted value of the benchmark model; calculating the transaction risk score by combining the relationship strength score and the deviation score; and aggregating the risks of all transactions in a film and television project to obtain a project-level risk score.

9. A method for constructing an audit model for related-party transactions of an enterprise using knowledge graphs according to claim 8, characterized in that, The deviation is calculated by dividing the difference between the actual transaction price and the benchmark forecast value by the benchmark forecast value, taking into account whether it exceeds the forecast range. The risk score is a weighted sum of the association strength score and the deviation score exceeding the allowable threshold.

10. The method for constructing an audit model for related-party transactions of an enterprise using knowledge graphs according to claim 1, characterized in that, The audit report mentioned in step S7 includes basic project information, a list of hidden related parties, details of abnormal transactions, risk levels, and visualization charts, and supports exporting to various file formats for auditors to review.

Images