Enterprise exhaustion report generation method and device based on deep learning model

Abstract

The invention discloses an enterprise exhaustion report generation method and device based on a deep learning model, relates to the technical field of data processing, can be applied to the field of finance, and mainly aims to solve the problem that the existing enterprise exhaustion report generation efficiency is relatively low. The method mainly comprises the steps that multi-dimensional data of a target exhaustion note subject is acquired, and the multi-dimensional data comprises industry data and public opinion data of an industry where the target exhaustion note subject is located, an associated subject of the target exhaustion note subject and associated subject basic data; identifying the industry data and the public opinion data according to a pre-constructed risk assessment model to obtain a first risk parameter; constructing a multi-subject association relationship graph of the target exhaustion note subject according to the association subject and the association subject basic data, and extracting according to the multi-subject association relationship graph to obtain a second risk parameter; and according to the first risk parameter and the second risk parameter, generating an exhaustion report of the target exhaustion subject according to the report template configuration information. The method is mainly used for generating enterprise exhaustion reports.

Description

Technical Field

[0001] This invention relates to the field of data processing and can be applied to the financial field, particularly to a method and apparatus for generating corporate due diligence reports based on a deep learning model. Background Technology

[0002] In today's deeply globalized and digitally integrated business environment, enterprises face an increasingly complex and volatile operating environment. With the continuous development and innovation of financial markets, the relationships between enterprises are becoming increasingly close and complex, and activities such as investment, mergers and acquisitions, and cooperation are becoming more frequent. Against this backdrop, due diligence, as a crucial means of assessing a company's risk, value, and development potential, is becoming increasingly important. Whether financial institutions are making credit decisions, investment institutions are making equity investments, or companies themselves are selecting strategic partners, due diligence is essential to comprehensively understand the true situation of the target company. At the same time, the rapid development of emerging technologies such as big data and artificial intelligence has brought new opportunities and tools to due diligence work, making it possible to extract valuable information from massive amounts of data and driving continuous innovation in due diligence methods and technologies.

[0003] Current methods for generating due diligence reports primarily rely on manual data collection and analysis, which is not only time-consuming and labor-intensive but also susceptible to subjective influences, leading to incomplete information gathering and inaccurate analysis. In terms of data processing, existing technologies have limited ability to integrate multi-source heterogeneous data, failing to comprehensively identify the various risks faced by enterprises. In the risk assessment phase, the identification and quantification of risks are not accurate or detailed enough, making it difficult to meet the due diligence needs of complex business environments. From an efficiency perspective, the entire due diligence report generation process is cumbersome, requiring a considerable amount of time from data collection to report writing, resulting in low efficiency in report generation and an inability to provide timely support for enterprise decision-making. Summary of the Invention

[0004] In view of this, the present invention provides a method and apparatus for generating enterprise due diligence reports based on a deep learning model, the main purpose of which is to solve the problem of low generation efficiency of existing enterprise due diligence reports.

[0005] According to one aspect of the present invention, a method for generating enterprise due diligence reports based on a deep learning model is provided, comprising:

[0006] Obtain multidimensional data of the target due diligence subject, wherein the multidimensional data includes industry data and public opinion data of the industry in which the target due diligence subject is located, as well as the related subjects and basic data of the related subjects of the target due diligence subject;

[0007] The industry data and public opinion data are identified based on a pre-built risk assessment model to obtain the first risk parameter;

[0008] A multi-entity relationship graph of the target due diligence entity is constructed based on the associated entity and the basic data of the associated entity, and a second risk parameter is extracted based on the multi-entity relationship graph;

[0009] Based on the first risk parameter and the second risk parameter, a due diligence report for the target due diligence subject is generated according to the report template configuration information.

[0010] Furthermore, the risk assessment model includes a large model fine-tuned based on a financial knowledge base, a causal inference network, and a prediction network. The first risk parameter is obtained by identifying industry data and public opinion data based on the pre-constructed risk assessment model, including:

[0011] The large model, fine-tuned based on the financial knowledge base, extracts multiple risk factors from the industry data and public opinion data, as well as the semantic relationships between different risk factors;

[0012] By using the causal reasoning network, causal reasoning is performed based on the industry data, the public opinion data, and the semantic relationships to obtain the risk transmission chain;

[0013] The first risk parameter is obtained by classifying and predicting the semantic association and the risk transmission chain through a prediction network.

[0014] Furthermore, before identifying the industry data and public opinion data based on a pre-built risk assessment model to obtain the first risk parameter, the method further includes:

[0015] The process involves retrieving historical financial case data from a pre-built financial knowledge base, constructing fine-tuning training samples based on the historical financial case data, and then fine-tuning the large model by fine-tuning the low-rank matrix based on the fine-tuning training samples to obtain a large model fine-tuned based on the financial knowledge base.

[0016] Based on the causal discovery algorithm, risk factors and causal relationships between risk factors are extracted from the historical financial case data. An initial causal network structure is constructed with each risk factor as a node and the causal relationships between each risk factor as edges. Prior knowledge of causal relationships is added as a causal constraint to the initial causal network structure to obtain a causal inference network.

[0017] A multilayer perceptron is constructed. Prediction training samples are built based on the output of the large model fine-tuned based on the financial knowledge base and the output of the causal inference network. The multilayer perceptron is then trained based on the prediction training samples to obtain the prediction network.

[0018] Furthermore, the step of constructing a multi-subject relationship graph of the target due diligence subject based on the associated subject and the basic data of the associated subject includes:

[0019] Based on the basic data of the associated entities, the association relationship between the associated entities and the target due diligence entity is extracted, and an initial multi-entity association relationship graph is constructed with each associated entity as a node and the association relationship between each associated entity and the target due diligence entity as an edge;

[0020] The edge attributes matching different associations are identified from the preset weight mapping relationship set, and the edge attributes are embedded into their corresponding edges to obtain a multi-subject association relationship graph.

[0021] Furthermore, the extraction of the second risk parameter based on the multi-subject relationship graph includes:

[0022] Search for subgraphs that match the preset risk pattern from the multi-entity relationship graph;

[0023] Based on the paths between different entities in the multi-entity relationship graph, risk transmission paths under different risk types are extracted;

[0024] Clustering of each entity based on the distance of risk characteristics yields risk entity clusters;

[0025] The second risk parameter is obtained by weighting the number of subgraphs corresponding to each preset risk mode, the number of risk transmission paths corresponding to each risk type, and the number of risk entities in each risk entity cluster.

[0026] Furthermore, the acquisition of multidimensional data of the target due diligence subject includes:

[0027] Unstructured data is crawled through a distributed crawler network, following the collection and execution mechanism of a distributed task queue, and structured data is crawled through multi-source data interfaces.

[0028] The structured data is normalized and entity disambiguated sequentially by a pre-built real-time data pipeline to obtain the first structured data.

[0029] The unstructured data is processed by extracting key information using a pre-built natural language model, and the extracted key information is then structurally transformed to obtain the second structured data.

[0030] The first structured data and the second structured data are spatiotemporally aligned, and the spatiotemporally aligned data is classified and integrated according to preset data dimensions to obtain industry data and public opinion data, related entities and basic data of related entities.

[0031] Furthermore, the multidimensional data also includes the target entity's basic data, and the generation of the due diligence report for the target entity based on the first risk parameter and the second risk parameter, according to the report template configuration information, includes:

[0032] The report template configuration information is parsed to identify multiple user configuration components, wherein the user configuration components are determined based on the user's selection operation on at least one component of the report template component configuration tree;

[0033] A report template is generated based on the multiple user configuration components, and the target entity's basic data, the first risk parameter, and the second risk parameter are associated and added to their respective corresponding components to obtain an initial due diligence report;

[0034] The built-in compliance verifier filters out abnormal data in the basic data of the target entity, marks the abnormal data, and generates a due diligence report for the target entity based on the initial due diligence report with the special marking.

[0035] According to another aspect of the present invention, an apparatus for generating enterprise due diligence reports based on a deep learning model is provided, comprising:

[0036] The acquisition module is used to acquire multidimensional data of the target due diligence subject, wherein the multidimensional data includes industry data and public opinion data of the industry in which the target due diligence subject is located, as well as the related subjects and basic data of the related subjects of the target due diligence subject;

[0037] The identification module is used to identify the industry data and public opinion data based on a pre-built risk assessment model to obtain the first risk parameter;

[0038] The extraction module is used to construct a multi-subject relationship graph of the target due diligence subject based on the associated subject and the basic data of the associated subject, and to extract the second risk parameter based on the multi-subject relationship graph;

[0039] The generation module is used to generate a due diligence report for the target due diligence subject based on the first risk parameter and the second risk parameter, according to the report template configuration information.

[0040] According to another aspect of the present invention, a storage medium is provided, wherein at least one executable instruction is stored therein, the executable instruction causing a processor to perform operations corresponding to the above-described method for generating enterprise due diligence reports based on deep learning models.

[0041] According to another aspect of the present invention, a computer device is provided, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other via the communication bus;

[0042] The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the above-described deep learning model-based enterprise due diligence report generation method.

[0043] By employing the above-described technical solutions, the technical solutions provided by the embodiments of the present invention have at least the following advantages:

[0044] This invention provides a method and apparatus for generating enterprise due diligence reports based on a deep learning model. First, multi-dimensional data of the target due diligence subject is acquired. This multi-dimensional data includes industry data and public opinion data of the target due diligence subject's industry, as well as related subjects and their basic data. The industry data and public opinion data are identified using a pre-constructed risk assessment model to obtain a first risk parameter. A multi-subject relationship graph of the target due diligence subject is constructed based on the related subjects and their basic data. A second risk parameter is extracted from this graph. Based on the first and second risk parameters, a due diligence report for the target due diligence subject is generated according to the report template configuration information. Compared with existing technologies, this invention comprehensively assesses risks by integrating multi-dimensional information such as industry data, public opinion data, and multi-subject relationships. This allows for more accurate capture of potential risk points of the target due diligence subject, effectively improving the accuracy and comprehensiveness of the due diligence report and providing a more reliable and detailed basis for enterprise decision-making.

[0045] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0046] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0047] Figure 1 This invention provides a flowchart of a method for generating enterprise due diligence reports based on a deep learning model, according to an embodiment of the present invention.

[0048] Figure 2 This invention provides a flowchart of another method for generating enterprise due diligence reports based on a deep learning model, according to an embodiment of the present invention.

[0049] Figure 3This invention provides a block diagram of another enterprise due diligence report generation device based on a deep learning model, according to an embodiment of the present invention.

[0050] Figure 4 A schematic diagram of the structure of a computer device provided in an embodiment of the present invention is shown. Detailed Implementation

[0051] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0052] This invention provides a method for generating enterprise due diligence reports based on a deep learning model, such as... Figure 1 As shown, the method includes:

[0053] 101. Obtain multi-dimensional data of the target due diligence subject.

[0054] In this embodiment of the invention, in the field of financial investment, when an investment institution considers investing in a company, it needs to fully understand the company. Therefore, it needs to acquire multi-dimensional data on the company to conduct due diligence and generate a due diligence report. The company to be understood in this process is the target due diligence subject. The multi-dimensional data includes industry data and public opinion data of the target due diligence subject's industry, as well as related entities and their basic data. Industry data can be obtained by collecting information such as the market size, growth rate, and competitive landscape of the entire industry over the past few years. Public opinion data can be obtained by collecting positive and negative evaluations and public attention regarding the company and its industry through social media, news websites, and other channels. Simultaneously, to gain a more comprehensive understanding of the company, it is also necessary to acquire the target due diligence subject's related entities, such as its parent company, subsidiaries, major partners, suppliers, and shareholders, and collect their basic data, such as the financial status, operating performance, and market reputation of a related company.

[0055] It is important to note that acquiring multidimensional data provides comprehensive and rich foundational information for subsequent risk assessments. Industry data helps understand the macro-environment and development trends of the target due diligence entity; public opinion data promptly reflects market and public perceptions and attitudes towards the entity and its industry; and related entity and basic data can indirectly reveal the potential risks and advantages of the target due diligence entity. Integrating this data allows for a more accurate assessment of the overall situation of the target due diligence entity, avoiding assessment biases caused by insufficient information and improving the accuracy and reliability of due diligence.

[0056] 102. Identify the industry data and public opinion data based on the pre-constructed risk assessment model to obtain the first risk parameter.

[0057] In this embodiment of the invention, a pre-built risk assessment model is constructed based on the data characteristics of multi-dimensional data and trained using a large amount of historical data and industry experience. Collected industry data, such as fluctuations in industry growth rates and the intensity of market competition, as well as public opinion data, such as the number of negative news stories and public sentiment, are input into this risk assessment model. The model learns the enterprise risk represented by the industry data and public opinion data, i.e., the first risk parameter. For example, if the industry growth rate continues to decline and there are numerous negative reports in the public opinion regarding the target company's product quality issues, the model will determine that the company faces high market and reputational risks and provide corresponding risk values.

[0058] It should be noted that using a pre-built risk assessment model to identify industry and public opinion data can quickly and objectively quantify the risks faced by the target due diligence entity in terms of the industry environment and public opinion. Through the primary risk parameter, investment institutions can clearly understand the degree of risk faced by the target due diligence entity in the external environment, improving the accuracy of data analysis and providing important reference for subsequent investment decisions.

[0059] 103. Construct a multi-entity relationship graph of the target due diligence entity based on the associated entity and the basic data of the associated entity, and extract the second risk parameter based on the multi-entity relationship graph.

[0060] In this embodiment of the invention, a multi-entity relationship graph is constructed using graph database technology based on the collected related entities and their basic data. In this graph, each related entity is treated as a node, and the relationships between them (such as controlling stakes, partnerships, etc.) are connected by edges. By analyzing this graph, potential risk points can be identified. For example, if the parent company of the target technology company is in poor financial condition, faces significant debt default risks, and has financial dealings with the target company, then the risk may be transferred to the target company. This potential risk information is extracted from the graph and quantified as a second risk parameter. Extracting the second risk parameter through graph analysis allows for a more comprehensive assessment of the risk status of the target due diligence entity, especially risks arising from related-party transactions, internal group risk transmission, etc., improving the accuracy of data analysis and thus providing a more comprehensive assessment of the overall risk profile of the target due diligence entity.

[0061] 104. Based on the first risk parameter and the second risk parameter, generate a due diligence report for the target due diligence subject according to the report template configuration information.

[0062] In this embodiment of the invention, after obtaining the first risk parameter (reflecting the target due diligence entity's risks in terms of industry and public opinion) and the second risk parameter (reflecting the risks brought by related entities), these parameters are integrated according to the pre-set report template configuration information. The report template includes various parts of the due diligence report, such as company overview, industry analysis, risk assessment, conclusions and recommendations, etc. Based on the magnitude and nature of the parameter values, detailed descriptions and analyses are provided in the corresponding sections. For example, if the first risk parameter shows that the target company is at a disadvantage in industry competition, and the second risk parameter indicates that related companies have significant financial risks, then the risk assessment section will focus on explaining the impact of these risks on the target company and provide corresponding risk ratings.

[0063] It should be noted that generating due diligence reports according to the report template configuration information ensures the standardization and consistency of the reports, making reports from different due diligence projects comparable. Integrating the primary and secondary risk parameters into the report provides investment institutions with comprehensive and accurate risk assessment results and decision-making basis.

[0064] In one embodiment of the present invention, for further illustration and limitation, such as Figure 2 As shown, the step described above involves identifying the industry data and public opinion data based on a pre-built risk assessment model to obtain a first risk parameter, including:

[0065] 201. Extract multiple risk factors from the industry data and public opinion data, as well as the semantic relationships between different risk factors, through the large model based on the financial knowledge base and fine-tuning.

[0066] 202. Through the causal reasoning network, causal reasoning is performed based on the industry data, the public opinion data, and the semantic relationships to obtain the risk transmission chain;

[0067] 203. The semantic association and the risk transmission chain are classified and predicted by the prediction network to obtain the first risk parameter.

[0068] In this embodiment of the invention, the risk assessment model includes a large model fine-tuned based on a financial knowledge base, a causal inference network, and a prediction network. The large model fine-tuned based on the financial knowledge base can identify risk factors from industry data, such as policy changes potentially leading to reduced subsidies for enterprises, and rapid technological updates causing existing equipment to become obsolete. It can also extract risk factors from public opinion data, such as consumer dissatisfaction with product quality potentially leading to a decline in market share, and negative media reports impacting a company's brand image. Furthermore, the large model can uncover semantic relationships between different risk factors. For example, policy change risks may affect a company's capital investment, thereby affecting technological research and development progress, ultimately leading to a decline in product quality, thus establishing logical connections between these risk factors at the semantic level.

[0069] Causal reasoning networks can perform causal reasoning based on industry data, public opinion data, and semantic relationships extracted from large models to analyze the causal logic between various risk factors and construct a complete risk transmission chain. For example, it may be found that tightening policies (a risk factor in industry data) reduce financial support for enterprises, preventing them from updating production equipment in a timely manner (the intermediate link reflected in semantic relationships), which in turn leads to unstable product quality, increased consumer complaints (a risk factor in public opinion data), and ultimately a decline in the enterprise's market share. This series of causal relationships constitutes a complete risk transmission chain.

[0070] The predictive network takes semantic relationships and risk transmission chains as input information and uses machine learning algorithms for classification and prediction. Since the predictive network is trained on similar past cases, it can determine the degree of risk faced by the current due diligence subject and quantify these risks to derive the primary risk parameter. For example, based on historical data of similar companies facing the same risk transmission chain, it can predict the current company's potential market share loss and the degree of deterioration in financial indicators, presenting the primary risk parameter in numerical form.

[0071] It should be noted that through fine-tuning, the large model can better adapt to specific due diligence tasks and data characteristics, accurately extracting risk factors and semantic relationships, providing a comprehensive and detailed foundation for subsequent risk analysis. By configuring a causal inference network, the inherent causal logic between risk factors can be deeply analyzed, constructing a complete risk transmission chain. The prediction network can quantify complex risk situations, deriving the primary risk parameter. The entire risk assessment model, through the synergy of these three parts, achieves a comprehensive and accurate identification of risks to the target due diligence subject in terms of industry data and public opinion data, providing reliable risk parameter support for the due diligence report.

[0072] In one embodiment of the present invention, for further explanation and limitation, before the step of identifying the industry data and public opinion data based on a pre-built risk assessment model to obtain the first risk parameter, the method further includes:

[0073] The process involves retrieving historical financial case data from a pre-built financial knowledge base, constructing fine-tuning training samples based on the historical financial case data, and then fine-tuning the large model by fine-tuning the low-rank matrix based on the fine-tuning training samples to obtain a large model fine-tuned based on the financial knowledge base.

[0074] Based on the causal discovery algorithm, risk factors and causal relationships between risk factors are extracted from the historical financial case data. An initial causal network structure is constructed with each risk factor as a node and the causal relationships between each risk factor as edges. Prior knowledge of causal relationships is added as a causal constraint to the initial causal network structure to obtain a causal inference network.

[0075] A multilayer perceptron is constructed. Prediction training samples are built based on the output of the large model fine-tuned based on the financial knowledge base and the output of the causal inference network. The multilayer perceptron is then trained based on the prediction training samples to obtain the prediction network.

[0076] In this embodiment of the invention, the financial knowledge base stores financial cases of different types and periods, covering various aspects such as stock market fluctuations, corporate mergers and acquisitions, and credit defaults. When constructing training samples, information such as text descriptions, key events, and risk indicators from historical financial cases can be labeled and organized to form a structured data format. For example, in a case where a company's stock price plummeted due to financial fraud, the risk factor of financial fraud, the result of the stock price plummeting, and the correlation between them are labeled. After obtaining the training samples, the large model is fine-tuned using a low-rank matrix fine-tuning method. Specifically, this involves using the LoRA (Low-Rank Adaptation) method to insert low-rank decomposition adapter modules into each layer of the original large model. During fine-tuning, only the parameters of these adapter modules are updated, while the parameters of the original model remain unchanged. This allows the large model to better learn knowledge and features in the financial field with fewer computational resources, resulting in a large model fine-tuned based on the financial knowledge base. This fine-tuned large model can more accurately understand the semantic information in financial texts, identify potential risk factors, and provide a solid foundation for subsequent risk assessment.

[0077] The causal inference network is constructed based on a causal discovery algorithm, specifically the PC algorithm (Peter-Clark Algorithm). This algorithm extracts causal relationships between multiple risk factors, such as policy changes, market demand fluctuations, and corporate financial conditions. For example, policy tightening leads to a decrease in market demand, thus affecting a company's sales and profits. After extracting the risk factors, an initial causal network structure is constructed, with each risk factor as a node and the causal relationships between them as edges. This initial structure includes the interactions and transmission paths between different risk factors. To further improve the accuracy of the initial causal network structure, prior knowledge of causal relationships is incorporated as a constraint. This prior knowledge consists of rules and principles summarized by financial experts based on long-term experience. For example, during economic recessions, corporate credit default risk typically increases—a widely accepted causal relationship. Adding this prior knowledge to the initial causal network structure yields a more accurate and reliable causal inference network. This network can simulate the causal transmission process between risk factors, providing a logically rigorous reasoning basis for risk assessment.

[0078] The prediction network is built upon a Multilayer Perceptron (MLP). A MLP is a feedforward artificial neural network with powerful nonlinear fitting capabilities. For example, assuming the fine-tuned large model outputs risk factors and semantic associations from industry and public opinion data, and the causal inference network outputs risk transmission chains and causal relationship strengths, this information is combined to construct prediction training samples. During training, the MLP continuously adjusts its weights and bias parameters to learn the mapping relationship between input information and a first risk parameter. After training with a large number of samples, the prediction network is obtained, which can accurately predict the degree of risk faced by the target due diligence subject based on the input industry and public opinion data, presenting it in the form of a first risk parameter.

[0079] In one embodiment of the present invention, for further explanation and limitation, the step of constructing a multi-subject relationship graph of the target due diligence subject based on the associated subject and the basic data of the associated subject includes:

[0080] Based on the basic data of the associated entities, the association relationship between the associated entities and the target due diligence entity is extracted, and an initial multi-entity association relationship graph is constructed with each associated entity as a node and the association relationship between each associated entity and the target due diligence entity as an edge;

[0081] The edge attributes matching different associations are identified from the preset weight mapping relationship set, and the edge attributes are embedded into their corresponding edges to obtain a multi-subject association relationship graph.

[0082] In this embodiment of the invention, the basic data of related entities may include their equity structure, transaction records, cooperation agreements, etc. From the equity structure data, the controlling or participating relationship between the related entity and the target due diligence entity can be determined by analyzing the shareholder list and shareholding ratio. For example, if related entity A holds 51% of the shares of the target due diligence entity, then it is clear that A has a controlling relationship with the target due diligence entity. Transaction record data reflects the business dealings between the related entity and the target due diligence entity, such as procurement, sales, and lending. By analyzing indicators such as transaction frequency and transaction amount, the closeness of the business dealings can be determined. Cooperation agreement data directly reflects the cooperative relationship between the two parties in specific projects or business areas. Each related entity is treated as a node, and both the related entity and the target due diligence entity are considered independent individuals in the graph. The relationships between each related entity and the target due diligence entity are used as edges, connecting the corresponding nodes. For example, if related entity B has a long-term procurement cooperation relationship with the target due diligence entity, then there is an edge in the graph connecting node B and the target due diligence entity node, with the edge type labeled "procurement cooperation". In this way, all related entities and their relationships with the target due diligence entity are integrated to construct an initial multi-entity relationship map.

[0083] To enrich the quantitative information in each edge, a pre-constructed set of weighted mapping relationships is used to find matching edge attributes based on the edge type in the initial multi-entity relationship graph. This set includes edge attributes corresponding to different relationship types. Edge attributes can include weight, trust level, and risk transmission coefficient. Weight reflects the closeness or importance of the relationship; for example, a controlling stake typically has a higher weight than a typical partnership because it has a greater impact on the decision-making and operations of the target due diligence entity. Trust level can be determined based on factors such as the related entity's historical credit record and the stability of the partnership; a high trust level implies more reliable business cooperation. The risk transmission coefficient measures the likelihood of risk transmission within the relationship; for example, a lending relationship with a financially weak related entity may have a higher risk transmission coefficient. Taking the graph of automobile manufacturing company C as an example, assuming the preset weight mapping relationship sets the weight of the controlling relationship at 0.8 and the risk transmission coefficient at 0.3; in the procurement and supply relationship, for a supplier with a credit rating of A, the trust level is 0.9 and the weight is 0.6; in the joint venture and cooperation relationship, if the partner is strong, the weight is 0.7 and the risk transmission coefficient is 0.4. Then, in C's graph, the edge embedding weight of the controlling relationship between D and C is 0.8 and the risk transmission coefficient is 0.3; the edge embedding trust level of the procurement and supply relationship between E and C is 0.9 and the weight is 0.6; and the edge embedding weight of the joint venture and cooperation relationship between F and C is 0.7 and the risk transmission coefficient is 0.4.

[0084] In one embodiment of the present invention, for further explanation and limitation, the step of extracting the second risk parameter based on the multi-subject relationship graph includes:

[0085] Search for subgraphs that match the preset risk pattern from the multi-entity relationship graph;

[0086] Based on the paths between different entities in the multi-entity relationship graph, risk transmission paths under different risk types are extracted;

[0087] Clustering of each entity based on the distance of risk characteristics yields risk entity clusters;

[0088] The second risk parameter is obtained by weighting the number of subgraphs corresponding to each preset risk mode, the number of risk transmission paths corresponding to each risk type, and the number of risk entities in each risk entity cluster.

[0089] In this embodiment of the invention, a graph matching algorithm is used to search for subgraphs in a multi-agent association graph that match a preset risk pattern. Specifically, a subgraph isomorphism algorithm can be used, which involves traversing all possible subgraphs in the multi-agent association graph and checking whether the nodes and edges of each subgraph match the preset risk pattern. Figure 1 One-to-one correspondence. If a matching subgraph is found, it indicates that a structure consistent with the preset risk pattern exists within the multi-entity relationship, potentially concealing corresponding risks. The preset risk pattern is a pre-constructed structure of relationships between nodes that may trigger risks, defined in the form of a subgraph. It clarifies the type and number of nodes (representing different entities) and edges (representing relationships), as well as the connections between them. For example, the preset risk pattern might be a triangular structure of "holding company - subsidiary - related party," where the holding company controls the subsidiary, and the subsidiary has large, unusual transactions with the related party. In the multi-entity relationship graph, a graph matching algorithm is used to search. If it finds that Company A controls Company B, and Company B has large, unusual financial transactions with Company C, then a subgraph matching the preset risk pattern has been found, indicating a potential risk of asset transfer through related-party transactions.

[0090] The multi-entity relationship graph includes various risk types, each with its own pre-defined transmission method and relationship characteristics. For example, credit risk is typically transmitted through debt-creditor relationships and guarantee relationships; market risk may affect related entities through price fluctuations in the upstream and downstream of the industrial chain. Risk types can include credit risk, market risk, and operational risk, and can also be customized according to the application scenario; this embodiment of the invention does not impose limitations. Based on the paths between entities in the multi-entity relationship graph, and combined with the characteristics of different risk types, risk transmission paths are extracted. First, the initial risk source entity is determined; for example, for credit risk, it might be a company with a low credit rating; for market risk, it might be a raw material supplier that is greatly affected by market price fluctuations. Then, along the edges of the graph, according to the transmission rules of risk types, paths that may transmit risk to other entities are sought. For example, in credit risk transmission, it is transmitted from the debtor to the creditor along the edges of the debt-creditor relationship; in guarantee relationships, it is transmitted from the guaranteed party to the guarantor.

[0091] Risk characteristics are extracted from each entity in a multi-entity relationship graph. These characteristics may include the entity's financial indicators (such as debt-to-equity ratio, current ratio, etc.), industry attributes, relationship complexity, and historical risk event records. Distance metrics are used to calculate the distance between different entities' risk characteristics; these metrics can be Euclidean distance, Manhattan distance, cosine similarity, etc. Based on the calculated risk characteristic distances, clustering algorithms (such as K-means clustering, hierarchical clustering, etc.) are used to cluster the entities. Clustering algorithms group entities with similar risk characteristics together, forming risk entity clusters.

[0092] Let N1 be the number of subgraphs corresponding to the preset risk model, and w1 be their weight; let N2 be the number of risk transmission paths corresponding to each risk type, and w2 be their weight; let N3 be the number of risk entities in each risk entity cluster, and w3 be their weight. Then the formula for calculating the second risk parameter R is:

[0093] R = w1×N1 + w2×N2 + w3×N3;

[0094] The weights of each item can be determined based on the correlation between various factors and the final risk loss in a large number of past risk events. For example, if historical data shows that the occurrence of a pre-defined risk pattern subplot is highly correlated with the occurrence of major risk events, then a higher weight should be assigned to the pre-defined risk pattern subplot.

[0095] In one embodiment of the present invention, for further explanation and limitation, the step of obtaining multidimensional data of the target due diligence subject includes:

[0096] Unstructured data is crawled through a distributed crawler network, following the collection and execution mechanism of a distributed task queue, and structured data is crawled through multi-source data interfaces.

[0097] The structured data is normalized and entity disambiguated sequentially by a pre-built real-time data pipeline to obtain the first structured data.

[0098] The unstructured data is processed by extracting key information using a pre-built natural language model, and the extracted key information is then structurally transformed to obtain the second structured data.

[0099] The first structured data and the second structured data are spatiotemporally aligned, and the spatiotemporally aligned data is classified and integrated according to preset data dimensions to obtain industry data and public opinion data, related entities and basic data of related entities.

[0100] In this embodiment of the invention, during the acquisition of multidimensional data of the target due diligence subject, a distributed crawler network and a distributed task queue are used to rapidly crawl unstructured data from multiple data sources simultaneously. This distributed approach greatly improves the efficiency and stability of data crawling, can handle massive data crawling tasks, and avoids data acquisition interruptions due to single points of failure. Simultaneously, structured data is crawled through multi-source data interfaces, ensuring the diversity and comprehensiveness of data sources. These multi-source data interfaces can include interfaces from public query platforms such as those for industry and commerce, taxation, customs, and the judiciary. After data collection, a pre-built real-time data pipeline is used to process the structured data. Normalization standardizes data of different formats and dimensions, eliminating differences between data and providing a consistent data foundation for subsequent analysis. Entity disambiguation accurately identifies and processes different representations of the same entity in the data, avoiding data confusion. These two steps result in accurate and standardized first-order structured data. For unstructured data, a pre-built natural language model is used to accurately extract key information and convert this key information into a structured format, i.e., second-order structured data, achieving effective utilization of unstructured data. To ensure data consistency and relevance across time and space, the first and second structured data were spatiotemporally aligned and then categorized and integrated according to preset data dimensions to obtain multi-dimensional data, including industry data, public opinion data, related entities, and basic data of related entities. This provides rich and high-quality data support for subsequent comprehensive and in-depth risk assessments and due diligence of the target entities, effectively improving the accuracy and comprehensiveness of the due diligence work.

[0101] In one embodiment of the present invention, for further explanation and limitation, the step of generating a due diligence report for the target due diligence subject according to the report template configuration information based on the first risk parameter and the second risk parameter includes:

[0102] The report template configuration information is parsed to identify multiple user configuration components, wherein the user configuration components are determined based on the user's selection operation on at least one component of the report template component configuration tree;

[0103] A report template is generated based on the multiple user configuration components, and the target entity's basic data, the first risk parameter, and the second risk parameter are associated and added to their respective corresponding components to obtain an initial due diligence report;

[0104] The built-in compliance verifier filters out abnormal data in the basic data of the target entity, marks the abnormal data, and generates a due diligence report for the target entity based on the initial due diligence report with the special marking.

[0105] In this embodiment of the invention, when generating a due diligence report for the target due diligence subject, the report template configuration information is first parsed to identify multiple user configuration components. These components are determined by the user based on the report template component configuration tree, according to their own needs for report content and structure, through selection operations, which can accurately meet the diverse and personalized report presentation requirements of different users. Next, a report template is generated based on the parsed user configuration components, and the target subject's basic data, first risk parameter, and second risk parameter are accurately associated and added to the corresponding components to obtain an initial due diligence report, ensuring that all kinds of key information are presented in an organized and logical manner in the report. Finally, a built-in compliance validator is used to comprehensively screen the target subject's basic data, quickly and accurately identify abnormal data and mark it specially, enabling report users to quickly identify potential problems. Based on the initial due diligence report with special markings, a final target due diligence report is generated, effectively improving the accuracy, compliance, and practicality of the report, and providing a high-quality and reliable decision-making basis for due diligence work.

[0106] This invention provides a method for generating enterprise due diligence reports based on a deep learning model. First, multi-dimensional data of the target due diligence subject is acquired. This multi-dimensional data includes industry data and public opinion data of the target due diligence subject's industry, as well as related subjects and their basic data. The industry data and public opinion data are identified using a pre-constructed risk assessment model to obtain a first risk parameter. A multi-subject relationship graph of the target due diligence subject is constructed based on the related subjects and their basic data, and a second risk parameter is extracted from this graph. Based on the first and second risk parameters, a due diligence report for the target due diligence subject is generated according to the report template configuration information. Compared with existing technologies, this invention comprehensively assesses risks by integrating multi-dimensional information such as industry data, public opinion data, and multi-subject relationships. This allows for more accurate capture of potential risk points of the target due diligence subject, effectively improving the accuracy and comprehensiveness of the due diligence report and providing a more reliable and detailed basis for enterprise decision-making.

[0107] Furthermore, as a response to the above Figure 1 The implementation of the method shown in this embodiment of the invention provides an enterprise due diligence report generation device based on a deep learning model, such as... Figure 3 As shown, the device includes:

[0108] The acquisition module 31 is used to acquire multidimensional data of the target due diligence subject, wherein the multidimensional data includes industry data and public opinion data of the industry in which the target due diligence subject is located, as well as the related subjects and basic data of the related subjects of the target due diligence subject;

[0109] The identification module 32 is used to identify the industry data and public opinion data based on a pre-built risk assessment model to obtain the first risk parameter;

[0110] Extraction module 33 is used to construct a multi-subject relationship graph of the target due diligence subject based on the associated subject and the basic data of the associated subject, and extract the second risk parameter based on the multi-subject relationship graph;

[0111] The generation module 34 is used to generate a due diligence report for the target due diligence subject based on the first risk parameter and the second risk parameter, according to the report template configuration information.

[0112] Furthermore, the identification module 32 includes:

[0113] The semantic extraction unit is used to extract multiple risk factors from the industry data and the public opinion data, as well as the semantic relationships between different risk factors, through the large model fine-tuned based on the financial knowledge base.

[0114] The causal reasoning unit is used to perform causal reasoning based on the industry data, the public opinion data, and the semantic relationships through the causal reasoning network to obtain the risk transmission chain.

[0115] The classification prediction unit is used to classify and predict the semantic association and the risk transmission chain through the prediction network to obtain the first risk parameter.

[0116] Furthermore, the device also includes:

[0117] The first construction module is used to retrieve historical financial case data from a pre-built financial knowledge base, construct fine-tuning training samples based on the historical financial case data, and fine-tune the large model by fine-tuning the low-rank matrix based on the fine-tuning training samples to obtain a large model fine-tuned based on the financial knowledge base.

[0118] The second construction module is used to extract risk factors and causal relationships between risk factors from the historical financial case data based on the causal discovery algorithm. It constructs an initial causal network structure with each risk factor as a node and the causal relationships between each risk factor as edges. It also adds prior knowledge of causal relationships as causal constraints to the initial causal network structure to obtain a causal inference network.

[0119] The third construction module is used to construct a multilayer perceptron. It constructs prediction training samples based on the output of the large model fine-tuned based on the financial knowledge base and the output of the causal inference network, and trains the multilayer perceptron based on the prediction training samples to obtain the prediction network.

[0120] Furthermore, the extraction module 33 includes:

[0121] The construction unit is used to extract the relationship between the associated subject and the target due diligence subject based on the basic data of the associated subject, and construct an initial multi-subject relationship graph with each associated subject as a node and the relationship between each associated subject and the target due diligence subject as an edge;

[0122] The identification unit is used to identify the edge attributes that match different associations from a preset weight mapping relationship set, and embed the edge attributes into their respective corresponding edges to obtain a multi-subject association relationship graph.

[0123] Furthermore, the extraction module 33 also includes:

[0124] The search unit is used to search for a subgraph that matches a preset risk pattern from the multi-subject relationship graph;

[0125] The extraction unit is used to extract risk transmission paths under different risk types based on the paths between different entities in the multi-entity relationship graph;

[0126] Clustering units are used to cluster various entities based on the distance of risk characteristics, resulting in risk entity clusters.

[0127] The calculation unit is used to perform weighted calculations on the number of subgraphs corresponding to each preset risk mode, the number of risk transmission paths corresponding to each risk type, and the number of risk entities in each risk entity cluster to obtain the second risk parameter.

[0128] Furthermore, the acquisition module 31 includes:

[0129] The data collection unit is used to collect unstructured data through a distributed crawler network and according to the collection and execution mechanism of a distributed task queue, and to collect structured data through a multi-source data interface.

[0130] The preprocessing unit is used to perform normalization and entity disambiguation on the structured data sequentially through a pre-built real-time data pipeline to obtain the first structured data.

[0131] The key information extraction unit is used to extract key information from the unstructured data using a pre-built natural language model, and to perform structural transformation on the extracted key information to obtain the second structured data.

[0132] The integration unit is used to perform spatiotemporal alignment of the first structured data and the second structured data, and to classify and integrate the spatiotemporally aligned data according to preset data dimensions to obtain industry data and public opinion data, related entities and related entity basic data.

[0133] Furthermore, the generation module 34 includes:

[0134] A parsing unit is used to parse multiple user configuration components in the report template configuration information, wherein the user configuration components are determined based on the user's selection operation on at least one component of the report template component configuration tree;

[0135] The template generation unit is used to generate a report template based on the multiple user configuration components, and associate the target subject basic data, the first risk parameter and the second risk parameter with their respective corresponding components to obtain an initial due diligence report;

[0136] The generation unit is used to filter out abnormal data in the basic data of the target entity through a built-in compliance verifier, mark the abnormal data in a special way, and generate a due diligence report of the target entity based on the initial due diligence report with the special marking.

[0137] This invention provides a deep learning model-based enterprise due diligence report generation device. First, it acquires multi-dimensional data of the target due diligence subject, including industry data and public opinion data of the target due diligence subject's industry, as well as related subjects and their basic data. Then, it identifies the industry data and public opinion data using a pre-built risk assessment model to obtain a first risk parameter. Next, it constructs a multi-subject relationship graph of the target due diligence subject based on the related subjects and their basic data, and extracts a second risk parameter from this graph. Finally, it generates a due diligence report for the target due diligence subject according to the first and second risk parameters and the report template configuration information. Compared with existing technologies, this invention comprehensively assesses risks by integrating multi-dimensional information such as industry data, public opinion data, and multi-subject relationships, enabling more accurate capture of potential risk points of the target due diligence subject, effectively improving the accuracy and comprehensiveness of the due diligence report, and providing a more reliable and detailed basis for enterprise decision-making.

[0138] According to one embodiment of the present invention, a storage medium is provided, the storage medium storing at least one executable instruction, the computer-executable instruction being able to execute the enterprise due diligence report generation method based on a deep learning model in any of the above method embodiments.

[0139] Figure 4 The diagram illustrates a structural schematic of a computer device according to an embodiment of the present invention. The specific embodiments of the present invention do not limit the specific implementation of the computer device.

[0140] like Figure 4As shown, the computer device may include: a processor 402, a communications interface 404, a memory 406, and a communications bus 408.

[0141] The processor 402, communication interface 404, and memory 406 communicate with each other via communication bus 408.

[0142] Communication interface 404 is used to communicate with other network elements such as clients or other servers.

[0143] The processor 402 is used to execute program 410, specifically to execute the relevant steps in the above embodiment of the enterprise due diligence report generation method based on deep learning model.

[0144] Specifically, program 410 may include program code that includes computer operation instructions.

[0145] Processor 402 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention. The computer device includes one or more processors, which may be processors of the same type, such as one or more CPUs; or processors of different types, such as one or more CPUs and one or more ASICs.

[0146] Memory 406 is used to store program 410. Memory 406 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0147] Specifically, program 410 can be used to cause processor 402 to perform the following operations:

[0148] Obtain multidimensional data of the target due diligence subject, wherein the multidimensional data includes industry data and public opinion data of the industry in which the target due diligence subject is located, as well as the related subjects and basic data of the related subjects of the target due diligence subject;

[0149] The industry data and public opinion data are identified based on a pre-built risk assessment model to obtain the first risk parameter;

[0150] A multi-entity relationship graph of the target due diligence entity is constructed based on the associated entity and the basic data of the associated entity, and a second risk parameter is extracted based on the multi-entity relationship graph;

[0151] Based on the first risk parameter and the second risk parameter, a due diligence report for the target due diligence subject is generated according to the report template configuration information.

[0152] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0153] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for generating enterprise due diligence reports based on a deep learning model, characterized in that, include: Obtain multidimensional data of the target due diligence subject, wherein the multidimensional data includes industry data and public opinion data of the industry in which the target due diligence subject is located, as well as the related subjects and basic data of the related subjects of the target due diligence subject; The industry data and public opinion data are identified based on a pre-built risk assessment model to obtain the first risk parameter; A multi-entity relationship graph of the target due diligence entity is constructed based on the associated entity and its basic data, and a second risk parameter is extracted based on the multi-entity relationship graph. Based on the first risk parameter and the second risk parameter, a due diligence report for the target due diligence subject is generated according to the report template configuration information.

2. The method according to claim 1, characterized in that, in, The risk assessment model includes a large model fine-tuned based on a financial knowledge base, a causal inference network, and a prediction network. The first risk parameter is obtained by identifying industry data and public opinion data based on the pre-constructed risk assessment model, including: The large model, fine-tuned based on the financial knowledge base, extracts multiple risk factors from the industry data and public opinion data, as well as the semantic relationships between different risk factors; By using the causal reasoning network, causal reasoning is performed based on the industry data, the public opinion data, and the semantic relationships to obtain the risk transmission chain; The first risk parameter is obtained by classifying and predicting the semantic association and the risk transmission chain through a prediction network.

3. The method according to claim 2, characterized in that, Before identifying the industry data and public opinion data based on a pre-built risk assessment model to obtain the first risk parameter, the method further includes: The process involves retrieving historical financial case data from a pre-built financial knowledge base, constructing fine-tuning training samples based on the historical financial case data, and then fine-tuning the large model by fine-tuning the low-rank matrix based on the fine-tuning training samples to obtain a large model fine-tuned based on the financial knowledge base. Based on the causal discovery algorithm, risk factors and causal relationships between risk factors are extracted from the historical financial case data. An initial causal network structure is constructed with each risk factor as a node and the causal relationships between each risk factor as edges. Prior knowledge of causal relationships is added as a causal constraint to the initial causal network structure to obtain a causal inference network. A multilayer perceptron is constructed. Prediction training samples are built based on the output of the large model fine-tuned based on the financial knowledge base and the output of the causal inference network. The multilayer perceptron is then trained based on the prediction training samples to obtain the prediction network.

4. The method according to claim 1, characterized in that, The construction of a multi-entity relationship graph of the target due diligence entity based on the associated entity and the basic data of the associated entity includes: Based on the basic data of the associated entities, the association relationship between the associated entities and the target due diligence entity is extracted, and an initial multi-entity association relationship graph is constructed with each associated entity as a node and the association relationship between each associated entity and the target due diligence entity as an edge; The edge attributes matching different associations are identified from the preset weight mapping relationship set, and the edge attributes are embedded into their corresponding edges to obtain a multi-subject association relationship graph.

5. The method according to claim 4, characterized in that, The second risk parameter extracted based on the multi-entity relationship graph includes: Search for subgraphs that match the preset risk pattern from the multi-entity relationship graph; Based on the paths between different entities in the multi-entity relationship graph, risk transmission paths under different risk types are extracted; Clustering of each entity based on the distance of risk characteristics yields risk entity clusters; The second risk parameter is obtained by weighting the number of subgraphs corresponding to each preset risk mode, the number of risk transmission paths corresponding to each risk type, and the number of risk entities in each risk entity cluster.

6. The method according to claim 1, characterized in that, The acquisition of multidimensional data of the target due diligence subject includes: Unstructured data is crawled through a distributed crawler network, following the collection and execution mechanism of a distributed task queue, and structured data is crawled through multi-source data interfaces. The structured data is normalized and entity disambiguated sequentially by a pre-built real-time data pipeline to obtain the first structured data. The unstructured data is processed by extracting key information using a pre-built natural language model, and the extracted key information is then structurally transformed to obtain the second structured data. The first structured data and the second structured data are spatiotemporally aligned, and the spatiotemporally aligned data is classified and integrated according to preset data dimensions to obtain industry data and public opinion data, related entities and basic data of related entities.

7. The method according to claim 1, characterized in that, The multidimensional data also includes the target entity's basic data. The process of generating a due diligence report for the target entity based on the first risk parameter and the second risk parameter, according to the report template configuration information, includes: The report template configuration information is parsed to identify multiple user configuration components, wherein the user configuration components are determined based on the user's selection operation on at least one component of the report template component configuration tree; A report template is generated based on the multiple user configuration components, and the target entity's basic data, the first risk parameter, and the second risk parameter are associated and added to their respective corresponding components to obtain an initial due diligence report; The built-in compliance verifier filters out abnormal data in the basic data of the target entity, marks the abnormal data, and generates a due diligence report for the target entity based on the initial due diligence report with the special marking.

8. A device for generating enterprise due diligence reports based on a deep learning model, characterized in that, include: The acquisition module is used to acquire multidimensional data of the target due diligence subject, wherein the multidimensional data includes industry data and public opinion data of the industry in which the target due diligence subject is located, as well as the related subjects and basic data of the related subjects of the target due diligence subject; The identification module is used to identify the industry data and public opinion data based on a pre-built risk assessment model to obtain the first risk parameter; The extraction module is used to construct a multi-subject relationship graph of the target due diligence subject based on the associated subject and the basic data of the associated subject, and to extract the second risk parameter based on the multi-subject relationship graph; The generation module is used to generate a due diligence report for the target due diligence subject based on the first risk parameter and the second risk parameter, according to the report template configuration information.

9. A storage medium, characterized in that, The storage medium stores at least one executable instruction that causes the processor to perform the operation corresponding to the enterprise due diligence report generation method based on a deep learning model as described in any one of claims 1-7.

10. A computer device, characterized in that, include: The processor, memory, communication interface, and communication bus are provided, wherein the processor, memory, and communication interface communicate with each other via the communication bus. The memory is used to store at least one executable instruction that causes the processor to perform the operation corresponding to the enterprise due diligence report generation method based on a deep learning model as described in any one of claims 1-7.

Images