Method for updating and compliance verification of financial knowledge graph based on multi-source heterogeneous data

Abstract

The application relates to the technical field of big data management in the financial industry, in particular to a financial knowledge graph updating and compliance verification method based on multi-source heterogeneous data. The method comprises the following steps: establishing a global perception network for the financial field, collecting multi-source heterogeneous financial data in real time, and performing normalization processing; using a pre-trained large language model to perform named entity recognition and relationship extraction on the normalized financial data, and constructing a financial knowledge graph; when a new regulatory rule is released, detecting whether there is a conflict between the new regulatory rule and the old regulatory rule, and if there is a conflict, performing lossless dynamic updating on the financial knowledge graph; when a financial business request is received, taking the business subject and the business type corresponding to the current financial business request as an anchor point, extracting a local subgraph in the financial knowledge graph, using a graph neural network to reason the local subgraph, and finally outputting an interpretable compliance audit report.

Description

TECHNICAL FIELD

[0001] Embodiments of the present application relate to the technical field of financial industry big data management, and in particular to a financial knowledge graph updating and compliance verification method based on multi-source heterogeneous data. BACKGROUND

[0002] With the deepening of the digital transformation of the financial industry, artificial intelligence technology (AI) has been widely applied to core business scenarios such as credit approval, risk control, intelligent investment, and compliance review. Financial institutions accumulate massive amounts of data in their daily operations, and market information and regulatory policies in the external environment are constantly changing. How to efficiently manage and utilize these data has become a key to improving the level of service intelligence.

[0003] However, in the current financial industry big data management and application practice, there are still serious data fusion and processing problems. On the one hand, financial data has typical multi-source heterogeneous characteristics. In addition to structured data generated by the bank's internal core transaction system, there are also a large amount of external data in unstructured form, such as policy and regulation documents in PDF (Portable Document Format) or picture format published by regulatory agencies, annual financial reports of listed companies, prospectuses, and real-time financial public opinion on the Internet. Existing data processing technologies, such as ETL (Extraction Transformation Loading) tools, often have difficulty in efficiently performing deep semantic understanding and structured extraction on these unstructured texts, which results in a large amount of high-value information sleeping in the documents, forming a serious data island, and making it difficult for downstream AI applications and services to effectively utilize them. On the other hand, the financial industry has a strong regulatory attribute, and related regulatory rules are in frequent dynamic adjustment. For example, revisions of anti-money laundering regulations, adjustments of credit interest rate ceilings, and changes in real estate financing policies will directly affect the compliance determination of financial businesses. Most current mainstream financial AI applications are based on static knowledge bases or pre-trained large language models. When external regulatory rules change, these systems often lack automated sensing and updating mechanisms, and retraining large models or manually maintaining rule bases not only have high costs and long cycles, but also during the updating window period, AI systems are prone to produce erroneous decision recommendations based on outdated knowledge. This "knowledge lag" phenomenon is likely to cause "AI hallucination", i.e., the model outputs business recommendations that violate regulations, which can bring huge compliance risks and legal risks to financial institutions.

[0004] In addition, the existing AI decision system also has obvious shortcomings in explainability. When facing audits by regulatory agencies, "black box" models based on deep learning often have difficulty in clearly tracing their decision basis. Especially in high-sensitive scenarios such as credit rejection and anti-money laundering early warning, the lack of clear, reasonable regulatory references and logical derivation paths restricts the application of AI decision systems.

[0005] Therefore, the financial industry urgently needs an updating and compliance verification method of a financial knowledge graph applicable to multi-source heterogeneous data to break through the limitations of traditional methods. SUMMARY

[0006] Therefore, the embodiments of the present application propose an updating and compliance verification method of a financial knowledge graph based on multi-source heterogeneous data, which can fuse multi-source heterogeneous data in real time, automatically perceive rules and resolve rule conflicts, realize dynamic updating of the financial knowledge graph, support precise, real-time and explainable compliance verification of the whole process of financial business, and improve the user experience.

[0007] To achieve the above-mentioned purpose, the embodiments of the present application propose an updating and compliance verification method of a financial knowledge graph based on multi-source heterogeneous data, which comprises the following steps: establishing a global perception network for the financial field, collecting financial data including unstructured regulatory documents, semi-structured market disclosure files and streaming public opinion data in real time, and performing normalization processing; wherein the financial data collected by the global perception network in real time is multi-source heterogeneous data; using a pre-trained large language model to perform named entity recognition and relationship extraction on the normalized financial data, establishing knowledge triples, and constructing a financial knowledge graph with time evolution capability; when a new regulatory rule is released, starting a conflict detection engine to detect whether there is a conflict between the new regulatory rule and the old regulatory rule, if there is no conflict, keeping the financial knowledge graph unchanged, if there is a conflict, performing lossless dynamic updating of the financial knowledge graph; when a specific financial business request is received, taking the business subject and business type corresponding to the current financial business request as an anchor point, extracting a local subgraph in the financial knowledge graph, using a graph neural network to perform risk transmission analysis and compliance verification reasoning on the local subgraph, and finally outputting an explainable compliance audit report.

[0008] To achieve the above object, the embodiment of the present application also proposes a financial knowledge graph updating and compliance verification system based on multi-source heterogeneous data, comprising a data acquisition module, a financial knowledge graph construction module, a financial knowledge graph updating module and a risk transmission analysis and compliance verification module; the data acquisition module is used to establish a global perception network for the financial field, collect financial data including unstructured supervision documents, semi-structured market disclosure files and streaming public opinion data in real time, and perform normalization processing; wherein the financial data collected by the global perception network in real time is multi-source heterogeneous data; the financial knowledge graph construction module is used to perform named entity recognition and relationship extraction on the normalized financial data by using a pre-trained large language model, establish knowledge triples, and construct a financial knowledge graph with time sequence evolution capability; the financial knowledge graph updating module is used to start a conflict detection engine when there is a new supervision rule released, detect whether there is a conflict between the new supervision rule and the old supervision rule, keep the financial knowledge graph unchanged if there is no conflict, and perform lossless dynamic updating on the financial knowledge graph if there is a conflict; the risk transmission analysis and compliance verification module is used to extract a local subgraph in the financial knowledge graph when a specific financial business request is received, taking the business subject and the business type corresponding to the current financial business request as anchor points, perform risk transmission analysis and compliance verification reasoning on the local subgraph by using a graph neural network, and finally output an interpretable compliance audit report.

[0009] To achieve the above object, the embodiment of the present application also proposes an electronic device, comprising a processor and a memory, wherein the memory stores instructions executable by the processor, and the processor is configured to execute the instructions so that the electronic device can implement the financial knowledge graph updating and compliance verification method based on multi-source heterogeneous data as described above.

[0010] To achieve the above object, the embodiment of the present application also proposes a computer readable storage medium storing a computer program, wherein the computer program is executed by a processor to implement the financial knowledge graph updating and compliance verification method based on multi-source heterogeneous data as described above.

[0011] In some optional embodiments, the unstructured regulatory documents at least include legal regulations in PDF format and red head documents in image format, the semi-structured market disclosure documents at least include financial reports, prospectuses and ESG (Environment Social Responsibility Governance) reports, and the streaming public opinion data at least includes financial news and social media trends; for the collected multi-source and heterogeneous financial data, multi-modal analysis is performed first, OCR (Optical Character Recognition) technology is used to extract text information in PDF format files and images, NLP (Natural Language Processing) technology is used to perform paragraph segmentation and cleaning on all long texts, and ASR (Automatic Speech Recognition) technology is used to recognize and transcribe audio data including conference recordings, and finally, the data is mapped to a unified intermediate data representation format to obtain normalized financial data.

[0012] In some optional embodiments, the normalized financial data is subjected to named entity recognition and relation extraction using a pre-trained large language model, knowledge triples are established, and a financial knowledge graph with time evolution capability is constructed, including: selecting a pre-trained BiLSTM (Bidirectional Long Short Term Memory) as the pre-trained large language model, using the pre-trained BiLSTM to perform named entity recognition and relation extraction on the normalized financial data, the recognized entities at least including enterprise entities, business entities, rule entities and time entities, and the extracted relations being factual relations between entities; establishing knowledge triples based on two entities and the factual relations therebetween, and labeling each knowledge triple with a timestamp attribute including an effective time and an invalid time based on a time slicing mechanism; based on all knowledge triples labeled with timestamp attributes, a financial knowledge graph with time evolution capability is constructed according to the logic of the objective world.

[0013] In some optional embodiments, when a new regulatory rule is released, the conflict detection engine is started to detect whether there is a conflict between the new regulatory rule and the old regulatory rule, if there is no conflict, the financial knowledge graph is kept unchanged, if there is a conflict, the financial knowledge graph is losslessly and dynamically updated, including: when a new regulatory rule is released, the conflict detection engine is started, the new regulatory rule is converted into a high-dimensional semantic vector, and semantic retrieval is performed in the financial knowledge graph, the regulatory rule with the highest semantic similarity with the new regulatory rule is taken as the old regulatory rule; compare the numerical constraints and logical constraints between the new regulatory rule and the old regulatory rule; if the numerical constraints and logical constraints between the new regulatory rule and the old regulatory rule are consistent, it is confirmed that there is no conflict between the new regulatory rule and the old regulatory rule, and the financial knowledge graph can be kept unchanged; if at least one of the numerical constraints and the logical constraints between the new regulatory rule and the old regulatory rule is inconsistent, it is confirmed that there is a conflict between the new regulatory rule and the old regulatory rule, and the financial knowledge graph needs to be losslessly and dynamically updated.

[0014] In some optional embodiments, the lossless and dynamic updating of the financial knowledge graph comprises: marking the attribute of the node corresponding to the old regulatory rule in the financial knowledge graph as invalid, and recording the invalidation time as the day before the new regulatory rule takes effect; inserting a node corresponding to the new regulatory rule in the financial knowledge graph, marking the attribute of the node corresponding to the new regulatory rule as effective, and recording the effective time as the first day of the new regulatory rule taking effect; in the financial knowledge graph, a relationship edge is established from the node corresponding to the new regulatory rule to the node corresponding to the old regulatory rule, forming a traceable regulatory rule evolution chain.

[0015] In some optional embodiments, when a specific financial service request is received, a local subgraph is extracted in the financial knowledge graph with the service subject and the service type corresponding to the current financial service request as the anchor point, a graph neural network is used to perform risk transmission analysis and compliance verification reasoning on the local subgraph, and finally an interpretable compliance audit report is output, including: when a specific financial service request is received, a local subgraph containing associated enterprises, guarantee chains, risk events and currently effective regulatory rules is extracted in the financial knowledge graph with the service subject and the service type corresponding to the current financial service request as the anchor point; using a graph neural network to perform risk transmission analysis and compliance verification reasoning on the local subgraph, analyzing whether the risk events of the associated enterprises can be transmitted to the service subject through the graph relationship, and verifying whether the attributes of the service subject meet the currently effective regulatory rules; based on the risk transmission analysis result and the compliance verification reasoning result, a comprehensive compliance score corresponding to the current financial service request is calculated; if the comprehensive compliance score meets the preset compliance threshold, the current financial service request is approved, and an interpretable compliance audit report representing the approval of the current financial service request is output; if the comprehensive compliance score does not meet the preset compliance threshold, the current financial service request is rejected, and an interpretable compliance audit report representing the rejection of the current financial service request is output, and the basis for the rejection is given in the interpretable compliance audit report.

[0016] In some optional embodiments, based on the risk transmission analysis result and the compliance verification reasoning result, a comprehensive compliance score corresponding to the current financial service request is calculated, which is realized by the following formula:

[0017] ;

[0018] ;

[0019] ;

[0020] ;

[0021] ;

[0022] wherein, represents the comprehensive compliance score, represents the regulatory rule satisfaction identifier, represents that the attributes of the service subject do not meet the currently effective regulatory rules, represents that the attributes of the service subject meet the currently effective regulatory rules, represents the risk transmission impact score, represents the compliance matching degree score, and are respectively and corresponding weight coefficient, , , , The value range of each of the above is , represents the impact degree of the first risk event, represents the loss scale of the first risk event, that is, the ratio of the loss amount of the first risk event to the net assets of the business subject, represents the violation degree of the first risk event, represents the range of the first risk event, , , corresponding weight coefficient, , , corresponding weight coefficient, , represents the probability of the first risk event being transmitted to the business subject, represents the relationship strength between the first risk event and the business subject, represents the dependence degree of the business subject on the first risk event, that is, the ratio of the annual resource amount obtained by the business subject from the first risk event to the total annual resource amount of the business subject, represents the length of the transmission path of the first risk event, is a preset path attenuation coefficient, represents the total number of risk events, represents the total number of index items of the currently effective regulatory rules, represents the actual value of the first index item of the business subject, represents the compliance value of the first index item.

[0023] The embodiment of the present application proposes a financial knowledge graph updating and compliance verification method based on multi-source heterogeneous data. Compared with the traditional AI decision model, at least the following beneficial effects are brought.

[0024] Firstly, the traditional technology can only process structured data, and the embodiment of the present application can effectively mine high-value information in unstructured supervision documents and streaming public opinions, and construct a more complete and three-dimensional financial knowledge base.

[0025] Second, the "time slicing mechanism" and "conflict detection engine" are introduced innovatively, solving the problem that the static knowledge base cannot cope with the frequent changes of new financial rules. The application can automatically complete knowledge iteration at the first time when new regulatory rules are released, eliminating the illusion risk of traditional AI decision models due to knowledge lag.

[0026] Third, unlike end-to-end black box models, the application based on graph path reasoning can provide accurate regulatory references and factual basis for each decision result, perfectly meeting the strong regulatory requirements of the financial industry for auditability and traceability, and improving the time sensitivity and compliance guarantee of financial decision-making.

[0027] Fourth, through risk transmission analysis of correlation, the application can discover implicit risks (such as complex guarantee circle risks or related transaction risks) that are difficult to detect by a single data source, improving the penetration of risk control. BRIEF DESCRIPTION OF DRAWINGS

[0028] In order to more clearly illustrate the technical solutions in the embodiments of the application or the related art, the drawings needed to be used in the description of the embodiments of the application or the related art will be briefly introduced. Obviously, the following drawings are only some embodiments of the application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of these drawings. The drawings described herein are only used to explain the application and do not limit the application.

[0029] Figure 1 is a flowchart of a financial knowledge graph updating and compliance verification method based on multi-source heterogeneous data provided in an embodiment of the application;

[0030] Figure 2 is a schematic diagram of a multi-source heterogeneous data acquisition process provided in an embodiment of the application;

[0031] Figure 3 is a schematic diagram of a financial knowledge graph construction process provided in an embodiment of the application;

[0032] Figure 4 is a schematic diagram of a financial knowledge graph updating process provided in an embodiment of the application;

[0033] Figure 5 is a schematic diagram of a compliance verification process provided in an embodiment of the application;

[0034] Figure 6 is a schematic diagram of a multi-source heterogeneous data sensing, intelligent analysis engine and normalized data processing architecture provided in an embodiment of the application;

[0035] Figure 7is a schematic diagram of time-slicing based rule conflict detection and graph evolution logic provided in an embodiment of the present application;

[0036] Figure 8 is a schematic diagram of local subgraph retrieval based risk transmission analysis and compliance verification inference process provided in an embodiment of the present application;

[0037] Figure 9 is a structural schematic diagram of a multi-source heterogeneous data based financial knowledge graph updating and compliance verification system provided in another embodiment of the present application;

[0038] Figure 10 is a structural schematic diagram of an electronic device provided in another embodiment of the present application. DETAILED DESCRIPTION

[0039] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the embodiments of the present application will be described in detail below with reference to the drawings. Those skilled in the art can understand that in the embodiments of the present application, many technical details are proposed in order to make the readers better understand. However, the technical solutions claimed by the present application can be implemented even without these technical details and various changes and modifications based on the following embodiments. The following embodiments are divided for the convenience of description, and should not constitute any limitation on the specific implementation of the present application. The following embodiments can be combined and referenced to each other without contradiction.

[0040] An embodiment of the present application proposes a multi-source heterogeneous data based financial knowledge graph updating and compliance verification method. The implementation details of the multi-source heterogeneous data based financial knowledge graph updating and compliance verification method proposed in the embodiment will be described in detail below. The following content only provides implementation details for the convenience of understanding, and is not necessary for implementing the present solution.

[0041] The specific process of the multi-source heterogeneous data based financial knowledge graph updating and compliance verification method proposed in the embodiment can be as shown in Figure 1 , including:

[0042] Step 11, a global perception network for the financial field is established, real-time collection of financial data including unstructured regulatory documents, semi-structured market disclosure files, and streaming public opinion data is performed, and normalization processing is performed.

[0043] In a specific implementation, the premise of building a financial knowledge graph is to collect sufficient financial data. In this embodiment, a global perception network for the financial field is established. The network can collect financial data in real time, including unstructured regulatory documents, semi-structured market disclosure files, and streaming public opinion data. These financial data have different sources and formats, belong to multi-source heterogeneous data, and cannot be directly used. Therefore, the collected financial data also need to be normalized, and the normalized financial data can be used as the basic data for building the financial knowledge graph.

[0044] In one example, the multi-source heterogeneous data collection process can be as shown in Figure 2 According to the different functions, it can be divided into multi-source heterogeneous data perception, intelligent analysis engine and normalized data processing as shown in Figure 6

[0045] The unstructured regulatory documents at least include laws and regulations, red head files in PDF format and picture format, the semi-structured market disclosure files at least include financial reports, prospectuses, ESG reports, and the streaming public opinion data at least include financial news and social media dynamics.

[0046] For multi-source heterogeneous financial data, intelligent analysis engine is needed for multi-modal analysis, including using OCR technology (OCR text recognition engine) to extract text information from PDF format files and images, using TableNet to restore the table structure of semi-structured data, using NLP technology (NLP text cleaning and segmentation) to segment and clean all long texts, using ASR technology (ASR speech / semantic analysis) to recognize and transcribe audio data including conference recordings, and finally mapping to a unified intermediate data representation format, such as JSON format (JavaScript Object Notation, JS key-value pair data) or XML format (Extensible Markup Language, Extensible Markup Language), to obtain normalized financial data.

[0047] In one example, the collection of multi-source heterogeneous financial data can also be divided by dimension. In the regulatory dimension (unstructured), PDF format and picture format laws and regulations and red head files published by the People's Bank of China and the China Banking and Insurance Regulatory Commission can be captured regularly. In the entity dimension (semi-structured), WIND and Futu can be accessed to obtain the annual financial statements and prospectus texts of target enterprises. In the public opinion dimension (streaming data), financial news push of Caifengwang and Caixin can be subscribed, and negative public opinion discussions about target enterprises on social media can be tried.

[0048] ​In one example, for the collected PDF format regulatory documents, OCR technology can be used to recognize the text, and NLP technology can be used to remove the header, footer and watermark noise. For mixed layout financial reports, table parsing algorithm can be used to restore the table structure, and key indicators such as "net asset-liability ratio" and "cash short debt ratio" can be extracted as structured key-value pairs.

[0049] Step 12, using a pre-trained large language model to perform named entity recognition and relationship extraction on the normalized financial data, establish knowledge triples, and construct a financial knowledge graph with time evolution capability.

[0050] In a specific implementation, after obtaining the normalized financial data, a pre-trained large language model can be used to perform named entity recognition and relationship extraction on the normalized financial data, establish knowledge triples, and construct a financial knowledge graph with time evolution capability.

[0051] In one example, the construction process of the financial knowledge graph can be as shown in Figure 3

[0052] This embodiment selects BiLSTM (BERT-BiLSTM-CRF) as a pre-trained large language model, and uses the pre-trained BiLSTM to perform named entity recognition and relationship extraction on the normalized financial data. The recognized entities include at least enterprise entities (such as A Group), business entities (such as loans), rule entities (such as the three red lines), and time entities (such as 2025), and the extracted relationships are factual relationships between entities.

[0053] Next, based on two entities and the factual relationship between them, "entity-relation-entity" knowledge triples are established, and based on the "time slicing mechanism", each knowledge triple is labeled with a timestamp attribute including "effective time" and "invalid time". The timestamp attribute format is year-month-day.

[0054] Finally, based on all the knowledge triples with timestamp attributes labeled, a financial knowledge graph with time evolution capability is constructed according to the logic of the objective world.

[0055] Step 13, when a new regulatory rule is released, a conflict detection engine is started to detect whether there is a conflict between the new regulatory rule and the old regulatory rule. If there is no conflict, the financial knowledge graph remains unchanged, and if there is a conflict, the financial knowledge graph is updated dynamically without loss.

[0056] ​In a specific implementation, when a new regulatory rule is released, a conflict detection engine is started immediately to detect whether there is a conflict between the new regulatory rule and the old regulatory rule. If there is no conflict, the financial knowledge graph remains unchanged. If there is a conflict, the financial knowledge graph is updated dynamically without loss.

[0057] In one example, the updating process of the financial knowledge graph can be as shown in Figure 4 The rule conflict detection and graph evolution logic based on time slicing can be as shown in Figure 7

[0058] When a new regulatory rule is released, the conflict detection engine is started, the new regulatory rule is converted into a high-dimensional semantic vector (this process is also called semantic alignment), and semantic retrieval is performed in the financial knowledge graph. The regulatory rule with the highest semantic similarity to the new regulatory rule is taken as the old regulatory rule.

[0059] After the old regulatory rule is determined, it can be compared whether the numerical constraints and logical constraints between the new regulatory rule and the old regulatory rule are consistent. The change of the numerical constraint is, for example, the interest rate cap changes from 36% to 24%. The change of the logical constraint is, for example, the access condition is changed.

[0060] If the numerical constraints and logical constraints between the new regulatory rule and the old regulatory rule are consistent, it is confirmed that there is no conflict between the new regulatory rule and the old regulatory rule, and the financial knowledge graph remains unchanged.

[0061] If at least one of the numerical constraints and logical constraints between the new regulatory rule and the old regulatory rule is inconsistent, it is confirmed that there is a conflict between the new regulatory rule and the old regulatory rule, and the financial knowledge graph needs to be updated dynamically without loss.

[0062] In one example, when the financial knowledge graph is updated dynamically without loss, the attribute of the node corresponding to the old regulatory rule in the financial knowledge graph is marked as “invalidated”, and the invalidation time is recorded as the day before the new regulatory rule takes effect. Then, a node corresponding to the new regulatory rule is inserted in the financial knowledge graph, and the attribute of the node corresponding to the new regulatory rule is marked as “in effect”, and the effective time is recorded as the first day of the new regulatory rule taking effect. Finally, in the financial knowledge graph, a relationship edge is established from the node corresponding to the new regulatory rule to the node corresponding to the old regulatory rule, forming a traceable regulatory rule evolution chain.

[0063] In one example, as shown in Figure 7 ​As shown, the new regulatory rule released in 2025 requires that the asset-liability ratio should not be greater than 70%, and the old regulatory rule released in 2021 requires that the asset-liability ratio should not be greater than 80%, which conflicts in numerical constraints. The new threshold 70% is more stringent than the old threshold 80%, so the new regulatory rule actually belongs to rule tightening. The old regulatory rule needs to be invalidated, and the invalidation time is recorded as December 31, 2024. The new regulatory rule needs to be validated, and the validation time is recorded as January 1, 2025.

[0064] In the financial knowledge graph, a relationship edge from the node corresponding to the new regulatory rule to the node corresponding to the old regulatory rule needs to be established, indicating “amended from”, to form a traceable regulatory rule evolution chain.

[0065] Step 14, when receiving a specific financial business request, taking the business subject and business type corresponding to the current financial business request as anchor points, extracting a local subgraph in the financial knowledge graph, and using a graph neural network to conduct risk transmission analysis and compliance verification reasoning on the local subgraph, finally outputting an interpretable compliance audit report.

[0066] In a specific implementation, the present embodiment needs to receive a specific financial business request in real time. When receiving a financial business request, taking the business subject and business type corresponding to the current financial business request as anchor points, extracting a local subgraph in the financial knowledge graph, and using a graph neural network to conduct risk transmission analysis and compliance verification reasoning on the local subgraph, finally outputting an interpretable compliance audit report.

[0067] In one example, the compliance verification process can be as shown in Figure 5 The risk transmission analysis and compliance verification reasoning process based on local subgraph retrieval can be as shown in Figure 8

[0068] When receiving a specific financial business request, the current financial business request needs to be parsed to identify the business subject (such as A Group) and the business type (such as initiating real estate loans), and then taking the business subject and business type corresponding to the current financial business request as anchor points, extracting a local subgraph containing associated enterprises, guarantee chains, risk events, and currently effective regulatory rules in the financial knowledge graph.

[0069] ​Next, we use graph neural networks to perform risk transmission analysis and compliance verification reasoning on local subgraphs. We analyze whether risk events of related enterprises can be transmitted to the business entity through graph relationships (e.g., if a "commercial bill overdue" event is detected in subsidiary B of Group A yesterday, this risk can be transmitted from B to Group A through equity relationships). We also verify whether the attributes of the business entity meet the currently effective regulatory rules (e.g., if Group A's asset-liability ratio is detected to be 75%, exceeding the requirement of no more than 70% under the currently effective regulatory rules).

[0070] Finally, based on the risk transmission analysis results and compliance verification reasoning results, the comprehensive compliance score corresponding to the current financial business request is calculated. If the comprehensive compliance score meets the preset compliance threshold, the current financial business request is approved, and an interpretable compliance audit report indicating approval of the current financial business request is output. If the comprehensive compliance score does not meet the preset compliance threshold, the current financial business request is rejected, and an interpretable compliance audit report indicating rejection of the current financial business request is output. The interpretable compliance audit report provides the basis for rejection (e.g., a recommendation to reject, reason one being that the asset-liability ratio is 75%, exceeding the 70% upper limit, and reason two being the existence of high-risk public opinion association).

[0071] In one example, based on the results of risk transmission analysis and compliance verification reasoning, the comprehensive compliance score corresponding to the current financial business request can be calculated using the following formula:

[0072] ;

[0073] ;

[0074] ;

[0075] ;

[0076] ;

[0077] in, This indicates the overall compliance score. This indicates that the regulatory rules are met. This indicates that the attributes of the business entity do not meet the currently effective regulatory rules. This indicates that the attributes of the business entity comply with the currently effective regulatory rules. This indicates the impact of risk transmission on the score. This indicates the compliance matching score. and They are respectively and The corresponding weighting coefficients, , , , The range of values ​​is , Indicates the first The extent of the impact of each risk event Indicates the first The scale of loss from the first risk event, i.e., the first The ratio of the loss amount of each risk event to the net assets of the business entity. Indicates the first The degree of violation in each risk event Indicates the first The scope of impact of each risk event, , , They are respectively , , The corresponding weighting coefficients, , Indicates the first The probability that a risk event will be transmitted to the business entity. Indicates the first The strength of the relationship between the initiator of each risk event and the business entity. Indicates the business entity's position on the first The degree of interdependence between the initiators of each risk event, i.e., the business entity from the first... The ratio of the annual resource amount received by the initiator of each risk event to the total annual resource amount of the business entity. Indicates the first The length of the transmission path of a risk event. The preset path attenuation coefficient, This represents the total number of risk events. This represents the total number of indicators for currently effective regulatory rules. The first one representing the business entity The actual value of each indicator item Indicates the first The compliance value of each indicator item.

[0078] In one example and The values ​​are 0.6 and 0.4 respectively.

[0079] In one example , , The values ​​are 0.4, 0.4, and 0.2 respectively.

[0080] In one example, the violation degree of the risk event is divided into no violation and no negative impact, no violation but negative impact, general violation, and serious violation, and the values are 0, 0.3, 0.6, and 1.0, respectively.

[0081] In one example, the scope of the risk event is related to the initiator of the risk event, and only affects the initiator and the business subject as a small range, the value is 0.2, the associated enterprises except the initiator and the business subject are 1-2 as a medium range, the value is 0.5, and the associated enterprises except the initiator and the business subject are more than 2 as a large range, the value is 1.0.

[0082] In one example, the relationship strength between the initiator of the risk event and the business subject depends on the holding relationship of the business subject relative to the initiator, no direct holding is 0, ordinary cooperation party is 0.2, strategic cooperation party or important supplier is 0.5, holding company is 0.8, and wholly-owned subsidiary is 1.0.

[0083] In one example, the length of the transmission path of the risk event depends on the number of nodes, if the risk event is directly transmitted from the initiator to the business subject, the length of the transmission path of the risk event is 1, and if the risk event needs to pass through a node to be transmitted to the business subject from the initiator, the length of the transmission path of the risk event is 2.

[0084] In one example, the preset path attenuation coefficient is fixed at 2.

[0085] In one example, the preset compliance threshold is 85.

[0086] The financial knowledge graph updating and compliance verification method based on multi-source heterogeneous data proposed in this embodiment at least has the following beneficial effects compared with the traditional AI decision model.

[0087] Firstly, it breaks through the limitation of traditional technology that can only process structured data, can effectively mine high-value information in unstructured supervision documents and streaming public opinions, and constructs a more complete and three-dimensional financial knowledge base.

[0088] Secondly, the "time slicing mechanism" and "conflict detection engine" are innovatively introduced, which solves the problem that the static knowledge base cannot cope with the frequent changes of new financial regulations. The embodiment can automatically complete knowledge iteration at the first time when new supervision rules are released, eliminating the illusion risk of traditional AI decision model caused by knowledge lag.

[0089] Thirdly, unlike the end-to-end black box model, the embodiment based on graph path reasoning can provide accurate regulation references and factual basis for each decision result, perfectly meeting the strong supervision requirements of the financial industry for auditable and traceable, and improving the time sensitivity and compliance guarantee of financial decision-making.

[0090] Fourthly, the embodiment can find implicit risks (such as complex guarantee circle risks or associated transaction risks) that are difficult to be perceived by a single data source through risk transmission analysis of the association relationship, and the penetration of the risk control is improved.

[0091] The division of steps in the above methods is only for the purpose of clearly describing the technical solutions, and in specific implementation, one step can be combined or some steps can be divided into multiple steps, as long as the same logical relationship is included, which is within the protection scope of the present application. Irrelevant modifications or irrelevant designs added in the algorithm or the process are within the protection scope of the present application as long as the core of the algorithm and the process is not changed.

[0092] Another embodiment of the present application provides a financial knowledge graph updating and compliance verification system based on multi-source heterogeneous data. Details of the financial knowledge graph updating and compliance verification system based on multi-source heterogeneous data provided by the embodiment are described below. The following details are provided for easy understanding and are not necessary for implementing the present application.

[0093] The structure of the financial knowledge graph updating and compliance verification system based on multi-source heterogeneous data provided by the embodiment is shown in Figure 9 The financial knowledge graph updating and compliance verification system based on multi-source heterogeneous data provided by the embodiment includes a data acquisition module 21, a financial knowledge graph construction module 22, a financial knowledge graph updating module 23, and a risk transmission analysis and compliance verification module 24.

[0094] The data acquisition module 21 is configured to establish a global perception network for the financial field, acquire financial data including unstructured regulatory documents, semi-structured market disclosure files, and streaming public opinion data in real time, and perform normalization processing. The financial data acquired by the global perception network in real time is multi-source heterogeneous data.

[0095] The financial knowledge graph construction module 22 is configured to perform named entity recognition and relationship extraction on the normalized financial data by using a pre-trained large language model, establish knowledge triples, and construct a financial knowledge graph with time evolution capability.

[0096] The financial knowledge graph updating module 23 is configured to start a conflict detection engine to detect whether there is a conflict between the new regulatory rules and the old regulatory rules in the case of publishing new regulatory rules. If there is no conflict, the financial knowledge graph remains unchanged. If there is a conflict, the financial knowledge graph is dynamically updated without loss.

[0097] The risk transmission analysis and compliance verification module 24 is configured to, when a specific financial service request is received, take the service subject and the service type corresponding to the current financial service request as an anchor point, extract a local subgraph in the financial knowledge graph, perform risk transmission analysis and compliance verification reasoning on the local subgraph by using a graph neural network, and finally output an interpretable compliance audit report.

[0098] It is worth mentioning that each module involved in the embodiment is a logical module. In actual application, one logical module can be one physical module, or a part of one physical module, or an organic combination of multiple physical modules. In addition, in order to highlight the innovative part of the present application, modules not closely related to solving the technical problems proposed in the present application are not introduced in the embodiment, but this does not mean that there are no other modules in the embodiment.

[0099] It can be found that the embodiment is a system embodiment corresponding to the above-mentioned method embodiment, and the embodiment can be implemented in cooperation with the above-mentioned method embodiment. The related technical details and technical effects mentioned in the above-mentioned method embodiment are still valid in the embodiment. In order to reduce repetition, they will not be described here. Correspondingly, the related technical details mentioned in the embodiment can also be applied to the above-mentioned method embodiment.

[0100] Another embodiment of the present application provides an electronic device, as shown in the figure, comprising a processor 31 and a memory 32, wherein the memory 32 stores instructions executable by the processor 31, and the processor 31 is configured to execute the instructions, so that the electronic device can implement a method for updating and compliance verification of a financial knowledge graph based on multi-source heterogeneous data as described in the above-mentioned method embodiment. Figure 10

[0101] The memory and the processor are connected in a bus mode, the bus includes any number of interconnected buses and bridges, and the bus connects various circuits of one or more processors and memories together. The bus can also connect various other circuits such as peripheral devices, voltage stabilizers, and power management circuits together, which are well known in the art, and therefore will not be further described herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be one element or multiple elements such as multiple receivers and transmitters, which provide a unit for communicating with various other devices on the transmission medium. The data processed by the processor is transmitted on the wireless medium through the antenna, and further, the antenna also receives data and transmits the data to the processor.

[0102] The processor is responsible for managing the bus and general processing, and can also provide various functions, including timing, peripheral interface, voltage regulation, power management, and other control functions. The memory can be used to store data used by the processor during operation.​

[0103] Another embodiment of the present application provides a computer readable storage medium, wherein a computer program is stored in the computer readable storage medium, and the computer program, when executed by a processor, enables a method for updating and compliance checking of a financial knowledge graph based on multi-source heterogeneous data to be implemented.

[0104] That is, a person skilled in the art can understand that all or part of the steps in the above method embodiments can be completed by programs instructing relevant hardware, and the programs are stored in a storage medium and include a plurality of instructions for causing a device (such as a single-chip microcomputer, a chip, etc.) or a processor to execute all or part of the steps of the method described in the method embodiments of the present application. The foregoing storage medium includes a U disk, a mobile hard disk, a read-only memory, a random access memory, a magnetic disk or an optical disk, and various media that can store program codes.

[0105] A person skilled in the art can understand that each of the above embodiments is a specific embodiment of the present application, and in actual application, various changes can be made in form and detail without departing from the spirit and scope of the present application. For those skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can also be made, which are also considered to be within the protection scope of the present application.

Claims

1. A method for updating and compliance verification of a financial knowledge graph based on multi-source heterogeneous data, characterized in that, The method comprises: A global perception network for the financial field is established, financial data including unstructured regulatory documents, semi-structured market disclosure files, and streaming public opinion data are collected in real time, and normalization processing is performed; wherein the financial data collected by the global perception network in real time is multi-source heterogeneous data; A pre-trained large language model is used to perform named entity recognition and relationship extraction on the normalized financial data, establish knowledge triples, and construct a financial knowledge graph with time evolution capability; When a new regulatory rule is released, a conflict detection engine is started to detect whether there is a conflict between the new regulatory rule and the old regulatory rule, if there is no conflict, the financial knowledge graph remains unchanged, if there is a conflict, the financial knowledge graph is updated dynamically without loss; When a specific financial business request is received, the business subject and the business type corresponding to the current financial business request are taken as anchor points, a local subgraph is extracted in the financial knowledge graph, a graph neural network is used to perform risk transmission analysis and compliance verification reasoning on the local subgraph, and finally an interpretable compliance audit report is outputted; When a specific financial business request is received, the business subject and the business type corresponding to the current financial business request are taken as anchor points, a local subgraph is extracted in the financial knowledge graph, a graph neural network is used to perform risk transmission analysis and compliance verification reasoning on the local subgraph, and finally an interpretable compliance audit report is outputted; When a specific financial business request is received, the business subject and the business type corresponding to the current financial business request are taken as anchor points, a local subgraph is extracted in the financial knowledge graph, a graph neural network is used to perform risk transmission analysis and compliance verification reasoning on the local subgraph, and finally an interpretable compliance audit report is outputted; The risk transmission analysis and compliance verification reasoning of the local subgraph are performed by using the graph neural network, whether the risk events of the associated enterprises can be transmitted to the business subject through the graph relationship is analyzed, and whether the attributes of the business subject meet the current effective regulatory rules are verified; Based on the risk transmission analysis result and the compliance verification reasoning result, a comprehensive compliance score corresponding to the current financial business request is calculated; If the comprehensive compliance score meets the preset compliance threshold, the current financial business request is approved, and an interpretable compliance audit report representing the approval of the current financial business request is outputted; If the comprehensive compliance score does not meet the preset compliance threshold, the current financial business request is rejected, an interpretable compliance audit report representing the rejection of the current financial business request is outputted, and the basis for the rejection is given in the interpretable compliance audit report; Based on the risk transmission analysis result and the compliance verification reasoning result, a comprehensive compliance score corresponding to the current financial business request is calculated, which is realized by the following formula: ； ； ； ； ； in, This indicates the overall compliance score. This indicates that the regulatory rules are met. This indicates that the attributes of the business entity do not meet the currently effective regulatory rules. This indicates that the attributes of the business entity comply with the currently effective regulatory rules. This indicates the impact of risk transmission on the score. This indicates the compliance matching score. and They are respectively and The corresponding weighting coefficients, , , , The range of values ​​is , Indicates the first The extent of the impact of each risk event Indicates the first The scale of loss from the first risk event, i.e., the first The ratio of the loss amount of each risk event to the net assets of the business entity. Indicates the first The degree of violation in each risk event Indicates the first The scope of impact of each risk event, , , They are respectively , , The corresponding weighting coefficients, , Indicates the first The probability that a risk event will be transmitted to the business entity. Indicates the first The strength of the relationship between the initiator of each risk event and the business entity. Indicates the business entity's position on the first The degree of interdependence between the initiators of each risk event, i.e., the business entity from the first... The ratio of the annual resource amount received by the initiator of each risk event to the total annual resource amount of the business entity. Indicates the first The length of the transmission path of a risk event. The preset path attenuation coefficient, This represents the total number of risk events. This represents the total number of indicators for currently effective regulatory rules. The first one representing the business entity The actual value of each indicator item Indicates the first compliance value for the indicator item.

2. The method of claim 1, wherein, The unstructured regulatory documents at least include laws and regulations, red head files in PDF format and picture format, the semi-structured market disclosure files at least include financial reports, prospectuses, and ESG reports, and the streaming public opinion data at least include financial news and social media dynamics; For the collected multi-source heterogeneous financial data, first multi-modal analysis is performed, OCR technology is used to extract text information in PDF format files and images, NLP technology is used for paragraph segmentation and cleaning of all long texts, ASR technology is used for recognition and transcription of audio data including conference recordings, and finally a unified intermediate data representation format is mapped to obtain normalized financial data.

3. The method of claim 2, wherein, The pre-trained large language model is used for named entity recognition and relation extraction of the normalized financial data, and a knowledge triple is established to construct a financial knowledge graph with time evolution capability, including: A pre-trained BiLSTM is selected as a pre-trained large language model, and the pre-trained BiLSTM is used for named entity recognition and relation extraction of the normalized financial data, and the recognized entities include at least enterprise entities, business entities, rule entities, and time entities, and the extracted relationships are factual relationships between entities; Based on two entities and the factual relationship between them, a knowledge triple is established, and based on a time slicing mechanism, a timestamp attribute including an effective time and an invalid time is labeled for each knowledge triple; Based on all knowledge triples with labeled timestamp attributes, a financial knowledge graph with time evolution capability is constructed according to the logic of the objective world.

4. The method of claim 3, wherein, When a new regulatory rule is released, a conflict detection engine is started to detect whether there is a conflict between the new regulatory rule and the old regulatory rule, if there is no conflict, the financial knowledge graph remains unchanged, if there is a conflict, the financial knowledge graph is updated dynamically without loss, including: When a new regulatory rule is released, a conflict detection engine is started, the new regulatory rule is converted into a high-dimensional semantic vector, and semantic retrieval is performed in the financial knowledge graph, the regulatory rule with the highest semantic similarity to the new regulatory rule is selected as the old regulatory rule; Compare the numerical constraints and logical constraints between the new regulatory rule and the old regulatory rule; If the numerical constraints and logical constraints between the new regulatory rule and the old regulatory rule are consistent, it is confirmed that there is no conflict between the new regulatory rule and the old regulatory rule, and the financial knowledge graph remains unchanged; If at least one of the numerical constraints and logical constraints between the new regulatory rule and the old regulatory rule is inconsistent, it is confirmed that there is a conflict between the new regulatory rule and the old regulatory rule, and the financial knowledge graph needs to be updated dynamically without loss.

5. The method of claim 4, wherein, The dynamic updating of the financial knowledge graph without loss includes: Marking the attribute of the node corresponding to the old regulatory rule in the financial knowledge graph as invalid, and recording the invalid time as the day before the new regulatory rule takes effect; Insert a node corresponding to the new regulatory rule in the financial knowledge graph, mark the attribute of the node corresponding to the new regulatory rule as effective, and record the effective time as the first day of the new regulatory rule taking effect; In the financial knowledge graph, a relationship edge is established from the node corresponding to the new regulatory rule to the node corresponding to the old regulatory rule, forming a traceable regulatory rule evolution chain.

6. A system for updating and compliance verification of a financial knowledge graph based on multi-source heterogeneous data, characterized in that, The system comprises: The data collection module is configured to establish a global perception network for the financial field, collect financial data including unstructured regulatory documents, semi-structured market disclosure files, and streaming public opinion data in real time, and perform normalization processing. The global perception network collects financial data in real time, which is multi-source heterogeneous data. The financial knowledge graph construction module is configured to use a pre-trained large language model to perform named entity recognition and relationship extraction on the normalized financial data, establish knowledge triples, and construct a financial knowledge graph with time evolution capability. The financial knowledge graph updating module is configured to start a conflict detection engine when there is a new regulatory rule released, detect whether there is a conflict between the new regulatory rule and the old regulatory rule, keep the financial knowledge graph unchanged if there is no conflict, and perform lossless dynamic updating of the financial knowledge graph if there is a conflict. The risk transmission analysis and compliance verification module is configured to extract a local subgraph in the financial knowledge graph when a specific financial business request is received, use a graph neural network to perform risk transmission analysis and compliance verification reasoning on the local subgraph, and finally output an interpretable compliance audit report. When a specific financial business request is received, the business subject and business type corresponding to the current financial business request are used as anchor points to extract a local subgraph in the financial knowledge graph, and a graph neural network is used to perform risk transmission analysis and compliance verification reasoning on the local subgraph, and finally output an interpretable compliance audit report, including: When a specific financial business request is received, the business subject and business type corresponding to the current financial business request are used as anchor points to extract a local subgraph in the financial knowledge graph, and a graph neural network is used to perform risk transmission analysis and compliance verification reasoning on the local subgraph, and finally output an interpretable compliance audit report, including: The risk transmission analysis and compliance verification module is configured to extract a local subgraph in the financial knowledge graph when a specific financial business request is received, use a graph neural network to perform risk transmission analysis and compliance verification reasoning on the local subgraph, and finally output an interpretable compliance audit report. The risk transmission analysis and compliance verification module is configured to extract a local subgraph in the financial knowledge graph when a specific financial business request is received, use a graph neural network to perform risk transmission analysis and compliance verification reasoning on the local subgraph, and finally output an interpretable compliance audit report. Based on the risk transmission analysis result and the compliance verification reasoning result, the comprehensive compliance score corresponding to the current financial business request is calculated. If the comprehensive compliance score meets the preset compliance threshold, the current financial business request is approved, and an interpretable compliance audit report representing the approval of the current financial business request is output. If the comprehensive compliance score does not meet the preset compliance threshold, the current financial business request is rejected, and an interpretable compliance audit report representing the rejection of the current financial business request is output, and the basis for the rejection is given in the interpretable compliance audit report. ； ； ； ； ； wherein, denotes a comprehensive compliance score, denotes a regulatory rule satisfaction identifier, denotes that the attribute of the business subject does not satisfy the currently effective regulatory rule, denotes that the attribute of the business subject satisfies the currently effective regulatory rule, denotes a risk transmission influence score, denotes a compliance matching degree score, and are respectively and corresponding weight coefficients, , , , the value ranges of , denotes an influence degree of the th risk event, denotes a loss scale of the th risk event, i.e., a ratio of a loss amount of the th risk event to a net asset of the business subject, denotes a violation degree of the th risk event, denotes a range of influence of the th risk event, , , are respectively , , corresponding weight coefficients, , denotes a probability that the th risk event is transmitted to the business subject, denotes a relationship strength between an initiator of the th risk event and the business subject, denotes a dependency degree of the business subject on initiators of the th risk event, i.e., a ratio of an annual resource amount obtained by the business subject from the initiators of the th risk event to an annual total resource amount of the business subject, denotes a length of a transmission path of the th risk event, is a preset path attenuation coefficient, denotes a total number of risk events, denotes a total number of index items of the currently effective regulatory rule, denotes an actual value of the th index item of the business subject, denotes a compliance value for the indicator item.

7. An electronic device, comprising: Based on the risk transmission analysis result and the compliance verification reasoning result, the comprehensive compliance score corresponding to the current financial business request is calculated. including: A processor and a memory having stored therein instructions executable by the processor, the processor being configured to execute the instructions to enable the electronic device to implement the method for updating and compliance checking of a financial knowledge graph based on multi-source heterogeneous data according to any one of claims 1 to 5.

8. A computer readable storage medium storing a computer program, characterized in that, The computer program, when executed by a processor, can implement the method for updating and compliance checking of a financial knowledge graph based on multi-source heterogeneous data according to any one of claims 1 to 5.

Images