Financial risk early warning method and device based on knowledge graph, equipment and medium

Abstract

The application provides a financial risk early warning method and device based on a knowledge graph, equipment and a medium, wherein the method comprises: performing structured semantic analysis on acquired multi-source heterogeneous data to obtain semantic units; based on the semantic units, a cross-modal semantic alignment technology is used to establish a risk correlation relationship between entities and generate a risk relationship triple; based on the semantic units and the risk relationship triple, an enterprise risk knowledge graph is constructed; a graph neural network analysis is performed on the enterprise risk knowledge graph to extract a risk feature vector; risk signal discrimination and confidence evaluation are performed according to the risk feature vector, and risk early warning information is output. The method can improve the accuracy of financial risk early warning and provide real-time and accurate decision support for enterprises in complex environments.

Description

Technical Field

[0001] This invention belongs to the field of deep learning technology, and in particular relates to a financial risk early warning method, device, equipment and medium based on knowledge graph. Background Technology

[0002] As financial risk early warning technology evolves from traditional statistical analysis to intelligent methods, early methods primarily relied on univariate financial ratios for basic risk assessment, which had limitations due to the reliance on single indicators and susceptibility to manipulation. With advancements in statistical methods, multivariate models have improved predictive capabilities by integrating multiple financial indicators, but these models impose stringent assumptions on data distribution and have failed to address the issue of collaborative analysis between financial and non-financial data. In recent years, big data and artificial intelligence technologies have propelled this field into an intelligent phase: achieving real-time risk monitoring by integrating multi-source heterogeneous data and combining it with machine learning models, while leveraging cloud computing to improve computing efficiency.

[0003] However, existing technologies still face key bottlenecks. The lack of semantic alignment of multi-source heterogeneous data, especially the semantic gap between the structured features of financial data and unstructured information such as audit texts and market sentiment, makes it difficult to dynamically associate cross-modal entities (such as contract terms and financial statement items), weakens the integrity of risk transmission path analysis, makes it difficult to achieve accurate financial risk early warning, and thus affects the reliability of information for company decision-making. Summary of the Invention

[0004] Therefore, it is necessary to provide a knowledge graph-based financial risk early warning method, device, equipment, and medium to address the aforementioned technical issues. This method can achieve deep semantic association analysis of cross-modal data, accurately distinguish between real risk signals and interference information, improve the accuracy of financial risk early warning, and provide enterprises with real-time and accurate decision support in complex environments.

[0005] Firstly, this application provides a knowledge graph-based financial risk early warning method, including:

[0006] The acquired multi-source heterogeneous data is subjected to structured semantic parsing to obtain semantic units;

[0007] Based on semantic units, risk association relationships between entities are established through cross-modal semantic alignment technology, generating risk relationship triples;

[0008] Construct an enterprise risk knowledge graph based on semantic units and risk relationship triples;

[0009] Graph neural network analysis is performed on the enterprise risk knowledge graph to extract risk feature vectors;

[0010] Risk signals are identified and confidence levels are assessed based on risk feature vectors, and risk warning information is output.

[0011] In one embodiment, the acquired multi-source heterogeneous data is subjected to structured semantic parsing to obtain semantic units, including:

[0012] Acquire structured financial statement data, listed company announcement data, and industry public opinion texts, and perform data preprocessing to generate multi-source heterogeneous data;

[0013] Parse the fields of structured financial statement data to generate semantic units for financial statements;

[0014] Text segmentation is performed on the data of listed company announcements to extract the text of audit report chapters. Then, named entity recognition is performed on the text of audit report chapters using a financial pre-trained language model to generate a set of audit entities. The financial pre-trained language model is a FinBERT model trained based on historical financial corpus.

[0015] Perform relation extraction processing on the audit entity set to generate audit semantic units;

[0016] Detect event-triggered words in industry public opinion texts to generate an initial event set;

[0017] Fine-grained sentiment analysis is performed on the initial event set to generate public opinion semantic units;

[0018] Semantic units are obtained by encapsulating the semantic units of financial statements, audits, and public opinion.

[0019] In one embodiment, risk association relationships between entities are established based on semantic units using cross-modal semantic alignment technology, generating risk relationship triples, including:

[0020] Obtain the predefined enterprise financial risk domain ontology, which includes entity type definitions and legal relationship constraint rules between entity types; wherein, the legal relationship constraint rules are used to limit the allowed combination of relationship types between entity types;

[0021] Based on the ontology of enterprise financial risk, entity type mapping is performed on semantic units to generate a set of mapped semantic units;

[0022] Based on a pre-trained domain-adaptive semantic embedding model, the numerical descriptions in semantic units are vectorized to generate numerical feature vectors, and the textual descriptions in semantic units are vectorized to generate textual feature vectors.

[0023] Numerical feature vectors and text feature vectors are projected into a unified semantic space to generate a cross-modal semantic vector set, and the cosine similarity of each pair of vectors from different sources in the cross-modal semantic vector set is calculated.

[0024] When the cosine similarity is greater than the dynamic domain threshold and the combination of ontology entity types of the semantic unit conforms to the legal relation constraint rules, the relation predicate is determined according to the combination of ontology entity types, and a risk relation triplet is generated with the semantic unit as the subject and the relation predicate as the relation.

[0025] In one embodiment, when the cosine similarity exceeds the dynamic domain threshold and conforms to the legal relation constraint rules, a triplet is instantiated to generate a risk relation triplet, including:

[0026] The dynamic neighborhood threshold is calculated using the following formula:

[0027]

[0028]

[0029] in, For dynamic domain threshold, This indicates finding the minimum value. Let be the probability density function of the normal distribution. This represents the mean of the similarity distribution of real risk samples. The standard deviation of the similarity distribution of real risk samples. The mean of the similarity distribution of the interference samples. The standard deviation of the similarity distribution of the interference samples is given.

[0030] In one embodiment, an enterprise risk knowledge graph is constructed based on semantic units and risk relationship triples, including:

[0031] Using semantic units as knowledge graph nodes and risk relationship triples as knowledge graph edges, graph structure initialization processing is performed to generate an initial knowledge graph;

[0032] Semantic similarity scores are extracted from risk relationship triples to generate basic association strength values;

[0033] Determine the reliability weight coefficient by matching the preset weight coefficient table according to the data source type;

[0034] Obtain the timestamp of the semantic unit and calculate the time difference with the current time;

[0035] Based on the time difference, the freshness decay factor is calculated using a preset exponential decay function, where the exponential decay function is expressed as follows:

[0036]

[0037] in, As a freshness decay factor, The decay rate constant is This is the time difference;

[0038] The edge dynamic weights are generated by multi-dimensional fusion of the basic association strength value, reliability weight coefficient, and freshness decay factor, as shown in the following formula:

[0039]

[0040] in, For edge dynamic weights, For reliability weighting coefficients, It is the freshness decay factor;

[0041] The dynamic weights of the edges are mapped to the corresponding edges of the initial knowledge graph to generate an enterprise risk knowledge graph.

[0042] In one embodiment, graph neural network analysis is performed on the enterprise risk knowledge graph to extract risk feature vectors, including:

[0043] The node feature vectors and edge dynamic weights of the enterprise risk knowledge graph are input into a temporal graph attention network. The neighbor node information is aggregated through a multi-layer time-aware attention mechanism to generate a temporal embedding representation of the node.

[0044] High-risk nodes are identified from temporal embedded representations, and risk transmission paths within a preset number of hops are extracted from the enterprise risk knowledge graph starting from these high-risk nodes.

[0045] Path encoding is performed on the node embeddings and edge weights of each risk transmission path to generate path feature vectors;

[0046] The temporal embedding representation of the node is concatenated and fused with the path feature vector to generate a risk feature vector.

[0047] In one embodiment, risk signal discrimination and confidence assessment are performed based on risk feature vectors, and risk warning information is output, including:

[0048] The risk feature vector is input into a pre-trained risk classifier, and the risk probability score of each node is calculated through a multilayer perceptron and a self-attention mechanism.

[0049] Nodes whose risk probability scores exceed the preset alarm threshold are identified as risk signal nodes;

[0050] Based on the path features in the risk feature vector, the significance index of the transmission path corresponding to the risk signal node is calculated.

[0051] Based on the reliability weight coefficient and freshness decay factor of the corresponding edge in the enterprise risk knowledge graph, calculate the data confidence score of the risk signal node;

[0052] The risk probability score, significance index, and data confidence score are input into a pre-trained multidimensional risk classification model for classification processing to obtain the risk level classification results.

[0053] Based on the risk level classification results, corresponding risk warning information is generated, which includes the risk type, scope of impact, and confidence level label.

[0054] Secondly, this application also provides a knowledge graph-based financial risk early warning device, including:

[0055] The semantic parsing module is used to perform structured semantic parsing on the acquired multi-source heterogeneous data to obtain semantic units;

[0056] The cross-modal alignment module is used to establish risk association relationships between entities based on semantic units and through cross-modal semantic alignment technology, and generate risk relationship triples.

[0057] The graph construction module is used to build an enterprise risk knowledge graph based on semantic units and risk relationship triples;

[0058] The risk analysis module is used to perform graph neural network analysis on the enterprise risk knowledge graph and extract risk feature vectors;

[0059] The risk warning module is used to identify risk signals and assess confidence levels based on risk feature vectors, and output risk warning information.

[0060] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-mentioned knowledge graph-based financial risk early warning method.

[0061] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described knowledge graph-based financial risk early warning method.

[0062] The aforementioned knowledge graph-based financial risk early warning method, device, equipment, and medium obtain semantic units by performing structured semantic parsing on acquired multi-source heterogeneous data. Based on these semantic units, risk relationships between entities are established using cross-modal semantic alignment technology, generating risk relationship triples. This achieves dynamic cross-modal entity association between structured and unstructured data in the financial domain. A corporate risk knowledge graph is constructed based on the semantic units and risk relationship triples, integrating dispersed risk information in a structured graph form to present a corporate risk relationship network, providing a structured knowledge carrier for precise risk analysis. Graph neural network analysis is performed on the corporate risk knowledge graph to extract risk feature vectors. Utilizing the graph neural network's ability to model complex relationships, key features that comprehensively reflect the corporate risk status are extracted from the graph, providing precise feature input for risk signal discrimination. Risk signal discrimination and confidence assessment are performed based on the risk feature vectors, and risk early warning information is output. This technical solution effectively distinguishes between real risk signals and interference information through precise analysis of risk features, reducing misjudgments and omissions, improving the accuracy of financial risk early warning, and providing real-time and accurate decision support for enterprises in complex environments. Attached Figure Description

[0063] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0064] Figure 1 This is a flowchart illustrating a knowledge graph-based financial risk early warning method provided in an embodiment of the present invention.

[0065] Figure 2 This is a schematic diagram of a knowledge graph-based financial risk early warning device provided in an embodiment of the present invention. Detailed Implementation

[0066] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0067] First, a brief introduction to the terms used in the embodiments of this application will be given.

[0068] A semantic unit refers to the smallest unit of information with independent business meaning extracted from multi-source heterogeneous data. In multimodal data fusion scenarios, semantic units carry the core semantics of the original data through standardized representations (such as entities, attributes, and events), providing a basic data structure for cross-modal alignment. Essentially, it transforms unstructured text (such as audit report clauses) and semi-structured data (such as financial statement fields) into machine-processable semantic carriers, solving the information silo problem in traditional financial analysis.

[0069] Enterprise risk knowledge graph is a dynamic topological network constructed with semantic units as nodes and risk relationship triples as edges. This graph integrates financial entities (such as companies and suppliers), risk elements (such as debt ratios and litigation events) and their relationships to form a visualized and analyzable risk transmission network. Its core feature that distinguishes it from traditional databases is that it supports graph structure reasoning and can reveal hidden risk links through multi-hop relationships between nodes.

[0070] Graph Neural Network Analysis (GNN) refers to the process of extracting risk feature vectors from knowledge graphs using graph deep learning models (such as GCN and GAT). Its technical essence is to capture the risk transmission strength between nodes through neighborhood aggregation and dynamically weight key risk paths using graph attention. This technology transforms discrete triples into machine learning-readable features with quantifiable risk strength to support subsequent signal discrimination.

[0071] Based on the above definitions, the implementation environment of a knowledge graph-based financial risk early warning method provided in this application embodiment will be described. Indicatively, the implementation environment includes: a terminal, sensing devices, a processor, and a memory. The terminal is connected to the processor, sensing devices, and memory via a network. Sensing devices include, but are not limited to, financial data interface devices, text data extraction devices, and network data acquisition interfaces that interface with public opinion APIs. The processor can be a central processing unit, a graphics processing unit, an artificial intelligence chip, etc. The memory can be a distributed cloud storage system or a local server cluster, and is not limited here.

[0072] Based on the above explanations of terms and implementation environments, the application scenarios of the embodiments of this application are described. The knowledge graph-based financial risk early warning method provided in the embodiments of this application can be applied to scenarios including but not limited to the following:

[0073] In corporate lending, financial institutions need to accurately assess the financial risks of borrowers to control bad debt rates. Both pre-loan approval and post-loan monitoring stages face the challenge of collaborative analysis of multi-source data. This technical solution can acquire borrowers' financial statements, tax returns, credit reports, industry sentiment, and supply chain partner evaluations. After generating semantic units, it establishes a correlation between "debt repayment ability indicators - credit delinquency records - industry risk events," constructing a corporate risk knowledge graph. Before lending, graph neural networks are used to analyze the graph and extract risk feature vectors to determine the company's repayment ability and potential risks. After lending, the graph data is updated in real time to dynamically monitor changes in corporate financial indicators and the evolution of risk correlation paths, promptly identifying early warning signals such as "declining revenue - rising debt - guarantee chain risk," helping financial institutions optimize credit approval decisions and reduce post-loan risk losses.

[0074] As an illustration, the knowledge graph-based financial risk early warning method provided in this application embodiment can also be applied to other application scenarios. This is only an example and does not limit the specific application scenarios.

[0075] In one exemplary embodiment, such as Figure 1 As shown, a knowledge graph-based financial risk early warning method is provided. This embodiment illustrates the application of this method to a terminal in the aforementioned implementation environment. It is understood that this method can also be applied to a server, or to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps 101 to 105:

[0076] Step 101: Perform structured semantic parsing on the acquired multi-source heterogeneous data to obtain semantic units.

[0077] Specifically, this method can acquire multi-source heterogeneous data from enterprise ERP systems, financial data platforms, and internet public opinion monitoring systems, including structured financial statements, unstructured audit reports, and market dynamic texts. Specifically, it uses natural language processing technology to perform dependency parsing on unstructured text to identify risk entities (such as "accounts receivable") and their attributes (such as "overdue for 90 days"). Simultaneously, it parses key fields in tabular data (such as "debt-to-equity ratio") using a predefined ontology rule base to generate standardized semantic units. For example, it can use a BERT model combined with a BiLSTM-CRF architecture to extract entity relationships from the text, and use regular expressions to match numerical units in the tables to ensure the computability of the financial attributes of the semantic units. This step transforms fragmented data into associative semantic primitives, solving the problem of data modality fragmentation in traditional methods.

[0078] Step 102: Based on semantic units, establish risk association relationships between entities through cross-modal semantic alignment technology and generate risk relationship triples.

[0079] Specifically, multimodal embedding technology is used to map semantic units describing the same entity from different data sources to a unified vector space. For example, the semantic similarity of the "abnormal cash flow" item in financial statements and the textual description of "customer performance risk" in audit opinions is calculated using an entity linking algorithm. When the similarity exceeds a preset threshold, a risk relationship triple "Company A - Cash Flow Risk Exists - Customer B" is established. For instance, a dual-tower neural network model based on knowledge distillation can be used to encode financial values ​​and textual descriptions separately, align cross-modal entities using cosine similarity, and then generate dynamic triples through a relation extraction model. This process overcomes the limitations of single-data-source analysis, achieving a deep correlation between financial events and market behavior.

[0080] Step 103: Construct an enterprise risk knowledge graph based on semantic units and risk relationship triples.

[0081] Specifically, this method can define an ontology framework to classify entities into categories such as enterprises and guarantors, and limit relationship types to risk transmission types such as guarantees and holding companies, storing the triple structure in a Neo4j graph database. For example, a dynamic graph update mechanism is introduced. When newly acquired judicial documents show that an enterprise has defaulted on its debts, a triple "Enterprise C - Involved in Debt Default - Court D" is automatically created and associated with existing entities to form a guarantee chain topology. This graph supports multi-hop risk path queries (such as Enterprise A → Guarantee → Enterprise B → Holding Company → Enterprise C), providing a structured semantic foundation for subsequent graph analysis.

[0082] Step 104: Perform graph neural network analysis on the enterprise risk knowledge graph to extract risk feature vectors.

[0083] Specifically, this method inputs node attributes (such as registered capital) and edge weights (such as the proportion of guarantee amount) from the graph into a graph convolutional network (GCN), capturing the intensity of risk transmission through multi-layer neighborhood aggregation operations. Specifically, it iteratively aggregates the risk features of the first-order neighbors for each enterprise node, dynamically allocating path weights using a graph attention mechanism (GAT). For example, when a multi-layer equity pledge ring is detected in an enterprise, the network strengthens the features of high-leverage nodes in the ring structure, outputting a 128-dimensional risk feature vector representing the probability of debt default. This step transforms discrete relationships into quantifiable and assessable machine learning features.

[0084] Step 105: Based on the risk feature vector, perform risk signal discrimination and confidence assessment, and output risk warning information.

[0085] For example, a logistic regression model can be used to classify feature vectors and output risk level labels, while an ensemble learning framework is used to calculate confidence probabilities. Specifically, risk feature vectors can be input into a voting model composed of XGBoost and random forest. When the model unanimously determines that there is a risk of "inflated income" with a probability exceeding 85%, a high-confidence warning report is generated. The output results include a visual traceability of the risk transmission path (such as an abnormal fund flow chain) to help users locate core risk nodes.

[0086] The aforementioned knowledge graph-based financial risk early warning method, device, equipment, and medium obtain semantic units by performing structured semantic parsing on acquired multi-source heterogeneous data. Based on these semantic units, risk relationships between entities are established using cross-modal semantic alignment technology, generating risk relationship triples. This achieves dynamic cross-modal entity association between structured and unstructured data in the financial domain. A corporate risk knowledge graph is constructed based on the semantic units and risk relationship triples, integrating dispersed risk information in a structured graph form to present a corporate risk relationship network, providing a structured knowledge carrier for precise risk analysis. Graph neural network analysis is performed on the corporate risk knowledge graph to extract risk feature vectors. Utilizing the graph neural network's ability to model complex relationships, key features that comprehensively reflect the corporate risk status are extracted from the graph, providing precise feature input for risk signal discrimination. Risk signal discrimination and confidence assessment are performed based on the risk feature vectors, and risk early warning information is output. This technical solution effectively distinguishes between real risk signals and interference information through precise analysis of risk features, reducing misjudgments and omissions, improving the accuracy of financial risk early warning, and providing real-time and accurate decision support for enterprises in complex environments.

[0087] In one embodiment, the acquired multi-source heterogeneous data is subjected to structured semantic parsing to obtain semantic units, including:

[0088] Acquire structured financial statement data, listed company announcement data, and industry sentiment text, and perform data preprocessing to generate multi-source heterogeneous data.

[0089] Specifically, the structured financial statement data originates from XBRL format files exported from the enterprise's ERP system, listed company announcement data comes from stock exchange disclosure platforms, and industry sentiment text is collected from news media and social media content via web crawlers. This method performs standardized preprocessing on the raw data, including missing value imputation, timestamp alignment, and encoding format conversion, generating a standardized data input stream. For example, to address the issue of inconsistent currency units in financial statements, regular expressions are used for matching and uniformly converting the currency units to the base currency unit, ensuring data comparability.

[0090] The structured financial statement data is parsed to generate semantic units for the financial statements.

[0091] Specifically, the structured financial statement data is parsed to generate semantic units for the financial statements. For example, a mapping rule base for accounting items and business terms is established, mapping "1122_Accounts Receivable" in the balance sheet to the standard entity "Accounts Receivable"; numerical fields are parsed through financial indicator calculation formulas, such as automatically recognizing "(Operating Revenue - Operating Costs) / Operating Revenue" as the "Gross Profit Margin" indicator, generating structured semantic units carrying numerical attributes and timestamps.

[0092] Text segmentation is performed on the data of listed company announcements to extract the text of audit report chapters. Then, named entity recognition is performed on the audit report chapter text using a financial pre-trained language model to generate a set of audit entities. The financial pre-trained language model is a FinBERT model trained based on historical financial corpus.

[0093] For example, this method uses a segmentation model based on chapter titles (such as BiLSTM-CRF) to identify chapters specific to the "Audit Report," isolating irrelevant textual interference. Specifically, the segmented audit report text is input into the pre-trained financial language model FinBERT, which has been micro-trained on historical financial corpora (such as SEC documents and audit opinions) and possesses the ability to extract sensitive features of financial terminology. For example, FinBERT identifies specialized entities such as "management fraud risk" and "going concern uncertainty," generating a set of audit entities that includes entity types and context.

[0094] Relational extraction is performed on the audit entity set to generate audit semantic units.

[0095] Specifically, a joint model of dependency parsing and semantic role labeling is used to analyze the logical relationships between entities. For example, the triple <raw material price increase, triggering, inventory impairment> is extracted from "inventory impairment due to rising raw material prices" and associated with negative statements in the audit opinion as risk labels to form audit semantic units with causal logic.

[0096] Event trigger words are detected in industry public opinion texts to generate an initial event set.

[0097] Specifically, a dictionary of financial event trigger words such as "merger and acquisition", "default" and "litigation" is constructed. The BiLSTM-CRF model is used to identify the event subject and time elements. For example, the "supply chain disruption" event type is extracted from "a new energy vehicle company was cut off by its supplier", with the event subject being "vehicle company A", and a basic event framework is generated.

[0098] Fine-grained sentiment analysis is performed on the initial event set to generate public opinion semantic units.

[0099] Specifically, a hierarchical sentiment analysis model based on domain dictionaries is adopted. First, the sentiment polarity of the event is determined, and then the specific risk dimension is identified through a convolutional neural network. For example, the word "surge" in "soaring battery raw material prices cause industry anxiety" is identified as negative sentiment in the dimension of "raw material costs". The sentiment intensity is quantified as 0.92, forming a public opinion semantic unit with sentiment weight.

[0100] Semantic units are obtained by encapsulating the semantic units of financial statements, audits, and public opinion.

[0101] Specifically, the data source, timestamp, entity type, attribute value, and confidence label are encapsulated in JSON-LD format. For example, the triple "raw material price increase, triggering inventory impairment" in the audit semantic unit is associated with the "decline in inventory turnover rate" indicator in the financial statements, forming a unified semantic expression that is readable by machines. This embodiment can achieve deep semantic parsing of multi-source heterogeneous data, effectively extracting key information from multiple aspects such as finance, auditing, and public opinion, and transforming it into a unified semantic unit. This not only makes up for the shortcomings of data processing in existing technologies, but also provides high-quality data support for subsequent risk association analysis and knowledge graph construction.

[0102] In one embodiment, risk association relationships between entities are established based on semantic units using cross-modal semantic alignment technology, generating risk relationship triples, including:

[0103] Obtain the predefined enterprise financial risk domain ontology, which includes entity type definitions and legal relationship constraint rules between entity types; wherein, the legal relationship constraint rules are used to limit the allowed combination of relationship types between entity types.

[0104] In this context, a domain ontology is a structured knowledge model of entity types (concepts), attributes, relationships, and logical constraints within a specific professional domain (such as financial risk). For example, this corporate financial risk domain ontology can be constructed using the OWL language. Entity type definitions cover categories such as corporate entities, guarantors, and debt instruments, while legal relationship constraints are explicitly defined using the SWRL rule set. For instance, it restricts the relationship between a "corporate entity" and a "guarantor" to only allow "guarantee" or "holding" relationships, prohibiting direct non-risk-transfer relationships such as "supply." This ontology is initialized through an expert knowledge base and continuously expands its entity types and constraint combinations as cases accumulate, ensuring that aligned behaviors conform to the logical norms of the financial risk domain.

[0105] Based on the ontology of enterprise financial risk, entity type mapping is performed on semantic units to generate a set of mapped semantic units.

[0106] Specifically, the "inventory turnover rate" in the financial statement semantic unit is mapped to the "operating efficiency" entity type in the ontology; the "bad debt provision ratio" in the audit semantic unit is mapped to the "credit risk" entity type; and the "raw material shortage event" in the public opinion semantic unit is mapped to the "supply chain disruption" entity type. The mapping results are verified through the SPARQL query engine to ensure that they conform to ontology constraints. For example, the text event "employee satisfaction" is rejected from being mapped to the financial risk entity, thus ensuring the logical consistency of the entity type system.

[0107] Based on a pre-trained domain-adaptive semantic embedding model, the numerical descriptions in semantic units are vectorized to generate numerical feature vectors, and the textual descriptions in semantic units are vectorized to generate textual feature vectors.

[0108] For example, the domain-adaptive semantic embedding model can be a BERT model trained on financial domain corpora, capable of transforming text and numerical descriptions into high-dimensional feature vectors. Furthermore, a dual-channel neural network architecture can be employed, where: the numerical channel encodes financial indicator values ​​and statistical features through a fully connected layer; the text channel loads a FinBERT model to extract deep semantic representations. By introducing a domain adversarial training mechanism, a gradient inversion layer is added during the pre-training phase, enabling the model to learn the numerical-text association patterns specific to the financial domain, such as aligning the numerical vector of "gross profit margin decreased by 15% year-on-year" with the text vector of "product price war intensified" in the latent space.

[0109] Numerical feature vectors and text feature vectors are projected into a unified semantic space to generate a cross-modal semantic vector set, and the cosine similarity of each pair of vectors from different sources in the cross-modal semantic vector set is calculated.

[0110] Specifically, heterogeneous vectors are transformed into semantic vectors of the same dimension through a shared weighted projection matrix, and the projection process is optimized based on optimal transport theory to minimize modal difference loss. For example, the cosine similarity between the numerical vector of "extended accounts payable turnover days" and the text vector of "surge in supplier demand letters" is calculated to generate an index that can quantify the strength of cross-modal association. Furthermore, a bilinear alignment layer can be designed to transform numerical and text feature vectors into the same vector space through a learnable matrix. For example, the cosine similarity between the numerical vector of "200 million yuan provision for inventory impairment" in the audit report and the text vector of the news text "risk of unsold products of a certain supplier" can be calculated.

[0111] When the cosine similarity is greater than the dynamic domain threshold and the combination of ontology entity types of the semantic unit conforms to the legal relation constraint rules, the relation predicate is determined according to the combination of ontology entity types, and a risk relation triplet is generated with the semantic unit as the subject and the relation predicate as the relation.

[0112] Specifically, the dynamic domain threshold is adjusted based on industry risk sensitivity: a lower threshold of 0.7 is set for high-volatility industries (such as real estate), while a threshold of 0.8 is used for stable industries (such as utilities). Relationship establishment is triggered only when the similarity meets the standard and the combination of ontology types is legal (such as allowing "debt risk transmission" relationship between "enterprise-supplier"). This embodiment can achieve deep semantic alignment of cross-modal data, effectively establish risk association relationships between entities, and generate risk relationship triples.

[0113] In one embodiment, when the cosine similarity exceeds the dynamic domain threshold and conforms to the legal relation constraint rules, a triplet is instantiated to generate a risk relation triplet, including:

[0114] The dynamic neighborhood threshold is calculated using the following formula:

[0115]

[0116]

[0117] in, For dynamic domain threshold, This indicates finding the minimum value. Let be the probability density function of the normal distribution. This represents the mean of the similarity distribution of real risk samples. The standard deviation of the similarity distribution of real risk samples. The mean of the similarity distribution of the interference samples. The standard deviation of the similarity distribution of the interference samples is given.

[0118] Specifically, when calculating the dynamic domain threshold, this method first constructs a similarity distribution model between real risk samples and interfering samples. For example, verified cross-modal association pairs are extracted from a historical risk event database as a positive sample set, such as the audited and confirmed relationship pair "raw material price increase leads to a decrease in gross profit margin"; simultaneously, association pairs judged by experts as false alarms are collected as a negative sample set, such as the falsely associated relationship pair "employee turnover rate affects fixed asset turnover". The cosine similarity of the cross-modal semantic vectors of the two sets of samples is calculated separately, forming two independent data distributions. Further, the maximum likelihood estimation method is used to calculate the distribution parameters: a normal distribution is fitted to the similarity data of the positive sample set to obtain the real risk distribution parameters. and Fit a normal distribution to the similarity data of the negative sample set to obtain the distribution parameters of the interference samples. and For example, in manufacturing risk scenarios, typical parameters might present... (High-risk association intensity concentration area) and The distribution characteristics of the (noise dispersion region) are analyzed. Based on the obtained distribution parameters, this method solves for the threshold T using a numerical optimization algorithm. Specifically, within the defined domain... Within the interval, the golden section search algorithm is used for iterative calculation. The minimum point. For example, when , and , At that time, the optimal solution can be converged in five iterations. This threshold will serve as the dynamic discrimination criterion for subsequent triple generation. The technical effectiveness of this dynamic neighborhood threshold is reflected in three aspects: first, it enables adaptive adjustment of the threshold, unlike the traditional fixed threshold method; The value is dynamically updated according to the distribution of real risk samples and interference samples, ensuring accurate differentiation under different risk scenarios; secondly, it improves the accuracy of risk association determination by calculating the optimal separation point based on probability distribution, reducing misjudgments caused by unreasonable single threshold settings (such as avoiding misjudging highly similar interference samples as real risks, or omitting low-similarity samples that are indeed real risks); thirdly, it enhances the robustness of the method by using the smooth description of sample distribution characteristics by normal distribution, reducing the impact of outliers on threshold calculation, making the threshold determination results more stable and reliable. The dynamic domain threshold calculated by the above formula, together with the legal relationship constraint rules, constitutes the dual judgment criteria for risk relationship triple generation. When the cosine similarity of cross-modal semantic unit pairs exceeds and conforms to the legal relationship between entity types, it can be confirmed that there is a real risk association between the two, providing a scientific and quantitative basis for subsequent triple instantiation, effectively improving the accuracy and reliability of risk relationship extraction.

[0119] In one embodiment, an enterprise risk knowledge graph is constructed based on semantic units and risk relationship triples, including:

[0120] Using semantic units as knowledge graph nodes and risk relationship triples as knowledge graph edges, a graph structure initialization process is performed to generate an initial knowledge graph.

[0121] Specifically, semantic units, as the basic semantic carriers of multi-source heterogeneous data after structured parsing, encompass multi-dimensional information such as financial statements, audits, and public opinion. Each semantic unit has a unique identifier, entity type, and attribute characteristics. Directly using them as nodes in the knowledge graph can preserve the original semantic information of each risk element. Risk relationship triples contain structured association information of "subject semantic unit - relation predicate - object semantic unit." Using them as edges connecting nodes can clarify the risk association logic between nodes. In the graph structure initialization phase, a unique graph node ID needs to be assigned to each node, establishing a mapping relationship with the unique identifier of the semantic unit; a unique graph edge ID needs to be assigned to each edge, associating it with the corresponding risk relationship triple, and recording the direction of the edge (from the subject semantic unit node to the object semantic unit node) and the type of relation predicate. The initial knowledge graph is stored using a graph database (such as Neo4j or JanusGraph). The node table stores information such as the ID, entity type, and core attributes of semantic units, while the edge table stores information such as the edge ID, starting node ID, ending node ID, and relational predicates. Through the association between nodes and edges, a preliminary network graph structure is formed, which intuitively presents the prototype of enterprise risk elements and their relationships.

[0122] Semantic similarity scores are extracted from risk relationship triples to generate basic association strength values.

[0123] Specifically, the semantic similarity score can be the cosine similarity value calculated during cross-modal semantic alignment. It reflects the degree of semantic association between the subject and object semantic units in the risk relationship triple. A higher similarity score indicates a stronger semantic association and a more reliable risk relationship. During extraction, this method uses the unique identifier of the risk relationship triple to associate it with the corresponding vector pair in the cross-modal semantic vector set, reads its calculated cosine similarity value, and directly uses this value as the basic association strength value. For example, for the triple <"Accounts receivable turnover decreased", "Affected by", "Customer default rate increased">, if its corresponding cosine similarity is 0.82, then the basic association strength value is... Set to 0.82; for the triple <"Audit Adjustment for Allowance for Doubtful Accounts Receivable", "Adjustment", "Overstated Accounts Receivable Balance">, if the cosine similarity is 0.76, then Set to 0.76. The range of the basic association strength value is [0,1], which provides an initial association strength benchmark for the subsequent calculation of edge dynamic weights and quantifies the basic reliability of the risk relationship between nodes.

[0124] The reliability weight coefficient is determined by matching the data source type with a preset weight coefficient table.

[0125] Specifically, semantic units and risk relation triples from different sources have varying information reliability, which needs to be assessed using reliability weighting coefficients. The method adjusts the weighting coefficient based on the source: the more authoritative the source and the higher the data quality, the larger the corresponding weighting coefficient, and the stronger its contribution to the dynamic weighting of the edge. The preset weighting coefficient table is formulated based on the authoritativeness, accuracy, and timeliness of the data source and has been reviewed and confirmed by domain experts. For example: structured financial data generated by an enterprise's internal financial system is reliable and can be weighted at 0.9; audit report data audited by an accounting firm is weighted at 0.85 because it has undergone professional audit verification; publicly available listed company announcements from stock exchanges have official authority and are weighted at 0.8; industry sentiment texts released by third-party financial information platforms may contain subjective biases or information deviations and are weighted at 0.6; unofficial enterprise-related discussions on social media have lower information credibility and are weighted at 0.4. When determining the reliability weighting coefficient, this method extracts the data source type of the subject semantic unit and the object semantic unit in the risk relationship triple and takes the lower of the two weighting coefficients as the corresponding weighting coefficient for that triple. (Adopting a conservative approach to avoid the reliability issues of a single high-weight source masking the reliability of another low-weight source). For example, if the subject semantic unit comes from the company's internal financial system (weight 0.9) and the object semantic unit comes from social media discussions (weight 0.4), then this triple... The weight is set at 0.4; if both are from the audit report (weight 0.85), then... The value was set at 0.85. By introducing a reliability weight coefficient, the weights of the edges can reflect the quality differences of the data sources, thereby improving the knowledge graph's ability to distinguish the reliability of risk associations.

[0126] Obtain the timestamp of the semantic unit and calculate the time difference with the current time.

[0127] Specifically, the timestamp of a semantic unit refers to the time when the original data corresponding to that unit was generated or published (such as the preparation time of financial statements, the issuance time of audit reports, and the publication time of public opinion texts), which is recorded and stored when the semantic unit is generated. The current time is the system time for constructing the enterprise risk knowledge graph, and the time difference ( The timestamp of the semantic unit is calculated by subtracting the timestamp of the semantic unit from the current time, with the unit being uniformly days (e.g., if the timestamp of a semantic unit is January 1, 2025, and the current time is June 20, 2025, then...). (Days). For the time difference corresponding to the risk relationship triples, the average of the timestamps of the subject semantic unit and the object semantic unit is calculated with the current time to ensure that the time difference can comprehensively reflect the overall timeliness of the information involved in the relationship. The calculation of the time difference provides a basic parameter for the subsequent calculation of the freshness decay factor, enabling the knowledge graph to distinguish the degree of influence of different time information on the current risk analysis.

[0128] Based on the time difference, the freshness decay factor is calculated using a preset exponential decay function, where the exponential decay function is expressed as follows:

[0129]

[0130] in, As a freshness decay factor, The decay rate constant is This represents the time difference.

[0131] Specifically, financial risk information is highly time-sensitive; recently generated information more accurately reflects the current risk status, while outdated information may lose its reference value due to changes in the company's operating conditions. Therefore, it is necessary to use a freshness decay factor to correct for time decay in the weights of the edges. The exponential decay function can simulate the non-linear decay process of information value over time: when... When the information is generated at the current time, it indicates that the information is completely fresh and the weight does not decay; as... Increase The decay rate gradually decreases, with a rapid initial decrease followed by a slower decline, consistent with the actual characteristics of risk information value decay (e.g., information decays more slowly over the past month, and significantly over the past six months). Decay Rate Constant Dynamic settings based on the type of risk information: For information with extremely high time sensitivity (such as public opinion events, sudden risk announcements). Larger values ​​cause the information value to decay rapidly; this is especially true for information with a longer time frame (such as annual financial statements and long-term audit adjustments). The smaller value ensures that the information value maintains a certain weight over a longer period of time. By calculating the freshness decay factor, the weights of the knowledge graph edges can dynamically reflect the time value of the information, avoiding interference from outdated information in the current risk analysis.

[0132] The edge dynamic weights are generated by multi-dimensional fusion of the basic association strength value, reliability weight coefficient, and freshness decay factor, as shown in the following formula:

[0133]

[0134] in, For edge dynamic weights, For reliability weighting coefficients, It is the freshness decay factor;

[0135] The dynamic weights of the edges are mapped to the corresponding edges of the initial knowledge graph to generate an enterprise risk knowledge graph.

[0136] For example, this method uses a multi-dimensional fusion of basic association strength value, reliability weight coefficient, and freshness decay factor to generate dynamic edge weights. The range of dynamic edge weights is [0,1], and the higher the value, the better the comprehensive performance of the corresponding risk relationship in the three dimensions of semantic association, source reliability, and information freshness. The dynamic edge weights are mapped to the corresponding edges of the initial knowledge graph to generate an enterprise risk knowledge graph. During the mapping process, the calculated dynamic edge weights are written into the edge table of the graph database using the edge's unique identifier, and stored in association with other attributes of that edge. The generated enterprise risk knowledge graph not only contains the topological structure of nodes (semantic units) and edges (risk relationship triples), but also assigns a quantified dynamic weight to each edge, which can intuitively distinguish the importance of different risk relationships. Edges with higher weights correspond to closer, more reliable, and timely risk associations, and should be given priority in subsequent risk analysis; edges with lower weights correspond to relatively weaker risk associations and can be used as auxiliary references. For example, when visualizing the graph, the weight differences can be intuitively presented through the thickness or color depth of the edges, helping analysts quickly identify core risk transmission paths. In summary, this embodiment constructs an enterprise risk knowledge graph through graph structure initialization, multi-dimensional weight fusion, and dynamic weight mapping. This not only fully preserves the semantic relationships of multi-source heterogeneous data, but also achieves quantitative differentiation of risk relationships through dynamic edge weights. This solves the problem of single edge attributes and difficulty in reflecting differences in the strength of associations in traditional knowledge graphs. It provides high-quality graph data with rich semantic and quantitative features for subsequent graph neural network analysis, improving the accuracy and efficiency of financial risk analysis.

[0137] In one embodiment, graph neural network analysis is performed on the enterprise risk knowledge graph to extract risk feature vectors, including:

[0138] The node feature vectors and edge dynamic weights of the enterprise risk knowledge graph are input into the temporal graph attention network. The information of neighboring nodes is aggregated through a multi-layer time-aware attention mechanism to generate the temporal embedding representation of the nodes.

[0139] Specifically, the node feature vectors of the enterprise risk knowledge graph originate from the vectorization results of the semantic units in the early stages. These vectors cover the core semantic features of the corresponding semantic units in dimensions such as finance, auditing, and public opinion. Each node feature vector maintains a uniform dimensionality to adapt to network input requirements. The edge dynamic weights are quantitative indicators generated during the initial construction of the enterprise risk knowledge graph, reflecting the semantic tightness of risk associations between nodes, the reliability of data sources, and the comprehensive characteristics of information timeliness. The temporal graph attention network is an improved model that introduces time-dimensional perception capabilities on top of traditional graph attention networks. Its core lies in a multi-layered time-aware attention mechanism. This mechanism not only measures the association strength between neighboring nodes and the current node based on edge dynamic weights but also adjusts the attention allocation ratio based on the timestamp characteristics of node information. For example, for two neighboring nodes of the current node, if their edge dynamic weights are similar, but one neighboring node's information timestamp is closer to the current analysis time, the time-aware attention mechanism will allocate a higher attention weight to that neighboring node, making the aggregation process more focused on recent and effective information. During network computation, each attention mechanism first performs a weighted summation of the features of all neighboring nodes of the current node. The weighting coefficients are determined by the dynamic weights of the edges and the time-aware weights. Then, the weighted neighbor features are fused with the current node's own features, and after processing by an activation function, the node feature representation of that layer is obtained. Through multiple layers of such iterative computation, the neighbor information of the current node under different time dimensions and different association strengths is gradually integrated, and finally a node temporal embedding representation that can comprehensively reflect the temporal change features of the node and the network topology association features is generated. This representation not only preserves the semantic attributes of the node itself, but also contains its dynamic association information with neighboring nodes in the knowledge graph.

[0140] High-risk nodes are identified from the temporal embedding representation, and risk transmission paths within a preset number of hops are extracted from the enterprise risk knowledge graph starting from the high-risk nodes.

[0141] Specifically, the temporal embedding representation of a node contains its comprehensive characteristics within the risk association network. High-risk nodes can be identified by assessing the risk level of this embedding representation. In the specific assessment process, this method inputs the node's temporal embedding representation into a preset risk classifier. This classifier is a model trained based on historical risk node labeled data, capable of outputting the node's risk level based on the feature patterns in the embedding representation (such as similarity to historical high-risk node embeddings and the magnitude of risk-related dimensions in the embedding vector). When a node's risk level reaches a preset high-risk threshold, it is determined to be a high-risk node. For example, in financial indicator nodes, the "debt-to-asset ratio temporal embedding representation" has a similarity exceeding a set value to historical high-debt default nodes, or in public opinion nodes, the risk dimension feature value of the "negative event temporal embedding representation" is significantly higher than that of normal nodes; these are all identified as high-risk nodes. After identifying a high-risk node, a risk transmission path is extracted starting from that node. The preset number of hops is set according to the actual patterns of enterprise risk transmission, typically 2-3 hops, ensuring coverage of the main risk transmission range while avoiding path redundancy due to excessive hops. The path extraction process employs a depth-first search algorithm, starting from a high-risk node and sequentially traversing its directly related neighbor nodes (1-hop path) and the related nodes of those neighbor nodes (2-hop path), recording the nodes and connecting edges at each step. Simultaneously, it combines edge dynamic weights for path filtering, retaining only node connections where the edge dynamic weight exceeds the path filtering threshold, and eliminating weakly related paths with excessively low weights. This ensures that the extracted risk transmission paths are all effective paths with actual risk transmission significance. For example, starting from the high-risk node of "customer default," it extracts a 2-hop path of "customer default → overdue accounts receivable → cash flow shortage," or a 3-hop path of "supplier cut-off → production stagnation → revenue decline → profit decline."

[0142] Path encoding is performed on the node embeddings and edge weights of each risk transmission path to generate path feature vectors.

[0143] Specifically, each risk transmission path contains an ordered sequence of nodes and a corresponding sequence of edge weights. The core of path encoding is to convert this sequenced topological and weight information into a fixed-dimensional feature vector. During the encoding process, the temporal embedding representation of each node on the path is weighted. The weighting coefficient is determined by the node's position in the path and its corresponding edge weight. High-risk nodes at the path's starting point and nodes with higher edge weights are assigned higher weighting coefficients to highlight the feature contribution of core risk nodes. Mean pooling is then performed on all weighted node embedding representations to obtain the comprehensive node features of the path. Simultaneously, statistical features are extracted from the edge weight sequence on the path, calculating the mean, maximum, minimum, and variance of the edge weights. These statistical features reflect the overall correlation strength and volatility of the path. The comprehensive node features and the statistical features of the edge weights are concatenated in a preset order and mapped to a unified dimension through a fully connected layer to generate the path feature vector. This vector not only contains the semantic and temporal features of the nodes on the path, but also incorporates the statistical information of the edge weights, which can fully characterize the overall features of a risk transmission path. For example, for the path "customer default → overdue accounts receivable → cash flow shortage", its path feature vector will contain the high-risk features of the "customer default" node, the correlation features of the "overdue accounts receivable" and "cash flow shortage" nodes, as well as the mean and maximum values ​​of the two edge weights, etc., which comprehensively reflect the risk transmission characteristics of the path.

[0144] The temporal embedding representation of the node is concatenated and fused with the path feature vector to generate a risk feature vector.

[0145] Specifically, the temporal embedding representation of a node focuses on characterizing the semantic, temporal, and local correlation features of a single node, while the path feature vector focuses on characterizing the overall features of the risk transmission path starting from that node. The fusion of the two can achieve a dimensional expansion from "individual node features" to "node-path global features". During the splicing and fusion process, it is necessary to ensure that the dimensionality of the node temporal embedding representation and the path feature vector are compatible. If the two dimensions are different, the path feature vector is mapped to the same dimension as the node temporal embedding representation through a fully connected layer. The two are spliced ​​in the order of feature categories. For example, the financial dimension features, audit dimension features, and public opinion dimension features in the node temporal embedding are spliced ​​first, and then the node comprehensive features and edge weight statistical features in the path feature vector are spliced. The spliced ​​feature vector is batch normalized to eliminate the numerical scale difference between different feature dimensions and avoid the dominance of a certain dimension feature due to excessively large values ​​in the subsequent risk judgment process. The generated risk feature vector retains the core risk attributes of the node itself and contains the path association information of the node in the risk transmission network. It can provide comprehensive and multi-dimensional feature support for subsequent risk signal judgment. For example, the risk feature vector of a "cash flow shortage" node contains not only the cash flow temporal change features of the node itself, but also the feature information of multiple paths such as "customer default → overdue accounts receivable → cash flow shortage" and "supplier arrears → delayed cash collection → cash flow shortage", which fully reflects the risk source and transmission impact of the node.

[0146] In one embodiment, risk signal discrimination and confidence assessment are performed based on risk feature vectors, and risk warning information is output, including:

[0147] The risk feature vector is input into a pre-trained risk classifier, and the risk probability score of each node is calculated through a multilayer perceptron and a self-attention mechanism.

[0148] Specifically, the risk feature vector integrates the temporal embedding representation of nodes with path feature vectors, including the node's own risk attributes and the transmission characteristics of associated paths, providing a comprehensive feature foundation for risk classification. The pre-trained risk classifier is a deep learning model trained on historical risk event data. Its core structure consists of a multilayer perceptron and a self-attention mechanism. The multilayer perceptron performs nonlinear transformations on the risk feature vector through multiple fully connected layers, progressively extracting higher-order risk features. Each layer uses the ReLU activation function to introduce nonlinear factors, and dropout is used to suppress overfitting. The self-attention mechanism is used to capture the correlation importance of features in different dimensions of the risk feature vector. For example, it automatically identifies the strong correlation between the "cash flow shortage" feature and the "supplier default" path feature, assigning higher attention weights to these key feature pairs, making the classifier more focused on feature combinations that play a decisive role in risk judgment. When calculating the risk probability score, the output layer of the classifier uses the sigmoid activation function, mapping the output value to the [0,1] interval. This value is the node's risk probability score, reflecting the likelihood of the node having actual risk. The closer the score is to 1, the more significant the risk signal corresponding to the node.

[0149] Nodes whose risk probability scores exceed the preset alarm threshold are identified as risk signal nodes.

[0150] Specifically, the preset alarm threshold is a value set based on the balance analysis of recall and precision of historical risk cases. It is usually determined by the inflection point of the Receiver Operating Characteristic (ROC) curve to ensure that while capturing as many true risk signals as possible, the false alarm rate is reduced. For example, when the alarm threshold is set to 0.7, all nodes with a risk probability score ≥ 0.7 are marked as risk signal nodes. The semantic units corresponding to these nodes may involve risk-related information such as abnormal financial indicators, audit adjustments, or negative public opinion events. After identifying the risk signal nodes, their entity type and attribute information in the enterprise risk knowledge graph are associated. For example, if the risk signal node is a financial indicator, its specific indicator name and numerical trend need to be recorded; if it is a risk event, information such as the subject involved and the area of ​​impact needs to be recorded to provide basic data for subsequent risk analysis.

[0151] Based on the path features in the risk feature vector, the significance index of the transmission path corresponding to the risk signal node is calculated.

[0152] Specifically, the path features in the risk feature vector include the comprehensive node features and edge weight statistics of each transmission path, which are the core basis for measuring the importance of the path. The calculation of the significance index focuses on the degree of influence of the path on the risk signal node. The specific process is as follows: extract the edge weight sequence of all associated transmission paths of the risk signal node, calculate the sum of edge weights of each path to reflect the overall association strength of the path; count the number of high-risk nodes in each path. The higher the proportion of high-risk nodes, the greater the significance contribution of the path; finally, adjust the path based on the path length (number of hops), appropriately reducing the significance weight of paths that are too short (incomplete information) or too long (high noise interference). The features of the above three dimensions are weighted and summed to obtain the significance score of each path. Then, the scores of all associated paths are normalized to generate a significance index in the range of 0-1. The higher the index value, the more critical the influence of the corresponding transmission path on the risk signal node, and the more critical the core path that should be given priority in risk analysis.

[0153] Based on the reliability weight coefficient and freshness decay factor of the corresponding edge in the enterprise risk knowledge graph, the data confidence score of the risk signal node is calculated.

[0154] Specifically, the data confidence score is used to assess the overall reliability of the data upon which the risk signal node is based. Its calculation takes all edges associated with the risk signal node as the analysis object. For each associated edge, its reliability weight coefficient and freshness decay factor are extracted, and the two are multiplied to obtain the data quality score of a single edge. This score comprehensively reflects the reliability of the source and the freshness of the information connected by the edge. A weighted average is calculated on the data quality scores of all associated edges of the risk signal node, with the weighting coefficient being the proportion of the edge's dynamic weight in the total weight of all associated edges. This ensures that the edge with the greater impact on the node contributes more to the final result in terms of its data quality score. The calculated result, after normalization, becomes the data confidence score of the risk signal node, ranging from 0 to 1. A higher score indicates a more reliable data source and more timely information supporting the risk signal, thus increasing the credibility of the risk signal.

[0155] The risk probability score, significance index, and data confidence score are input into a pre-trained multidimensional risk classification model for classification processing to obtain the risk level classification results.

[0156] Specifically, the multidimensional risk classification model is an ensemble model built on the gradient boosting tree algorithm, capable of comprehensively determining risk levels based on three dimensions of quantitative indicators. During the model training phase, labeled data from historical risk events is used to classify risk levels into four categories: low, medium, high, and extremely high. Each category corresponds to a different range of risk probability scores, significance indicators, and data confidence scores. For example, an "extremely high" risk level corresponds to a combination of a risk probability score ≥ 0.85, a significance indicator ≥ 0.8, and a data confidence score ≥ 0.75, indicating a clear risk signal, a crucial transmission path, and highly reliable data. During classification, the model performs feature cross-processing on the three input indicators to capture the interaction relationships between them (e.g., a combination of high risk probability score and low data confidence score requires a lower risk level). Through ensemble voting of multiple decision trees, the final risk level classification result of the risk signal node is output, ensuring the stability and accuracy of the classification results.

[0157] Based on the risk level classification results, corresponding risk warning information is generated, which includes the risk type, scope of impact, and confidence level label.

[0158] Specifically, the risk type is determined based on the entity type and associated path characteristics of the risk signal nodes. For example, financial indicator nodes correspond to types such as "debt repayment ability risk" and "profitability risk," while risk event nodes correspond to types such as "supply chain risk" and "credit risk." The type description is further refined by combining the core relational predicates in the path (e.g., "cash flow risk caused by customer default"). The scope of impact is determined by analyzing the extended nodes of the risk transmission path. If the path involves nodes in multiple areas such as production, sales, and finance, it is marked as "global impact"; if it is limited to a single business link, it is marked as "local impact," and the key departments or business modules affected are listed. Confidence level labels are generated based on a combination of data confidence scores and risk levels, and are divided into four levels: "extremely high confidence," "high confidence," "medium confidence," and "low confidence." This directly reflects the reliability of the risk warning information. The generated risk warning information is presented in structured text format, including a detailed description of the risk type, a clear definition of the scope of impact, and confidence level labels, facilitating corporate decision-makers to quickly understand the core elements of the risk and formulate targeted response strategies.

[0159] In summary, the knowledge graph-based financial risk early warning method provided in this application breaks through the data barriers of traditional financial analysis by generating standardized semantic units through structured semantic parsing of multi-source heterogeneous data; based on the enterprise financial risk domain ontology and adaptive semantic embedding model, it establishes risk association relationships between entities through cross-modal semantic alignment technology, generating risk relationship triples that conform to domain constraints; on this basis, it integrates semantic units and risk relationship triples to construct a dynamic enterprise risk knowledge graph, and innovatively introduces time decay functions and data source reliability weights to realize dynamic edge weight calculation; by performing time-series graph neural network analysis on the graph, it extracts multi-dimensional risk feature vectors by combining node states and risk transmission path features; and it generates hierarchical early warning decisions by combining a multi-dimensional classification model of risk probability scores, transmission path significance, and data confidence. The above technical solution eliminates the semantic gap between structured and unstructured data through cross-modal deep semantic association, accurately captures the risk transmission logic by using dynamic graph modeling and path analysis, and effectively distinguishes between real risk signals and market noise interference by using multi-dimensional confidence assessment. Thus, it significantly improves the accuracy of financial risk early warning in complex business environments and provides enterprises with decision support that is real-time, accurate and interpretable.

[0160] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0161] Based on the same inventive concept, this application also provides a knowledge graph-based financial risk early warning device 10 for implementing the knowledge graph-based financial risk early warning method described above. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the knowledge graph-based financial risk early warning device 10 provided below can be found in the limitations of the knowledge graph-based financial risk early warning method described above, and will not be repeated here.

[0162] In one exemplary embodiment, such as Figure 2 As shown, a knowledge graph-based financial risk early warning device 10 is provided, comprising:

[0163] Semantic parsing module 11 is used to perform structured semantic parsing on the acquired multi-source heterogeneous data to obtain semantic units;

[0164] The cross-modal alignment module 12 is used to establish risk association relationships between entities based on semantic units and through cross-modal semantic alignment technology, and generate risk relationship triples.

[0165] Graph construction module 13 is used to construct an enterprise risk knowledge graph based on semantic units and risk relationship triples;

[0166] Risk analysis module 14 is used to perform graph neural network analysis on the enterprise risk knowledge graph and extract risk feature vectors;

[0167] The risk warning module 15 is used to identify risk signals and assess confidence based on risk feature vectors, and output risk warning information.

[0168] In one embodiment, the semantic parsing module 11 includes:

[0169] The data preprocessing unit is used to acquire structured financial statement data, listed company announcement data, and industry public opinion text, and perform data preprocessing to generate multi-source heterogeneous data;

[0170] The financial statement parsing unit is used to parse fields in structured financial statement data and generate financial statement semantic units.

[0171] The audit entity recognition unit is used to segment the data of listed company announcements, extract the text of audit report chapters, and perform named entity recognition on the audit report chapter text using a financial pre-trained language model to generate a set of audit entities. The financial pre-trained language model is a FinBERT model trained based on historical financial corpus.

[0172] The audit relation extraction unit is used to extract relations from the audit entity set and generate audit semantic units.

[0173] The public opinion event detection unit is used to detect event trigger words in industry public opinion texts and generate an initial event set.

[0174] The public opinion sentiment analysis unit is used to perform fine-grained sentiment analysis on the initial event set and generate public opinion semantic units;

[0175] The semantic unit encapsulation unit is used to encapsulate financial statement semantic units, audit semantic units, and public opinion semantic units to obtain semantic units.

[0176] In one embodiment, the cross-modal alignment module 12 includes:

[0177] The domain ontology unit is used to obtain a predefined enterprise financial risk domain ontology, which includes entity type definitions and legal relationship constraint rules between entity types; wherein, the legal relationship constraint rules are used to limit the allowed combination of relationship types between entity types;

[0178] The entity type mapping unit is used to map semantic units to entity types based on the enterprise financial risk domain ontology, and generate a set of mapped semantic units;

[0179] Semantic vectorization units are used to vectorize numerical descriptions in semantic units based on pre-trained domain-adaptive semantic embedding models, generating numerical feature vectors, and to vectorize textual descriptions in semantic units, generating textual feature vectors.

[0180] The semantic similarity unit is used to project numerical feature vectors and text feature vectors onto a unified semantic space to generate a cross-modal semantic vector set and to calculate the cosine similarity of each pair of vectors from different sources in the cross-modal semantic vector set.

[0181] The triplet generation unit is used to determine the relation predicate based on the combination of ontology entity types when the cosine similarity is greater than the dynamic domain threshold and the combination of ontology entity types of the semantic unit conforms to the legal relation constraint rules, and to generate a risk relation triplet with the semantic unit as the subject and the relation predicate as the relation.

[0182] In one embodiment, the triplet generation unit is used to perform the following steps:

[0183] The dynamic neighborhood threshold is calculated using the following formula:

[0184]

[0185]

[0186] in, For dynamic domain threshold, This indicates finding the minimum value. Let be the probability density function of the normal distribution. This represents the mean of the similarity distribution of real risk samples. The standard deviation of the similarity distribution of real risk samples. The mean of the similarity distribution of the interference samples. The standard deviation of the similarity distribution of the interference samples is given.

[0187] In one embodiment, the map construction module 13 includes:

[0188] The graph initialization unit is used to perform graph structure initialization processing with semantic units as knowledge graph nodes and risk relationship triples as knowledge graph edges to generate an initial knowledge graph.

[0189] The basic association strength unit is used to extract semantic similarity scores from risk relationship triples and generate basic association strength values.

[0190] The reliability weighting unit is used to determine the reliability weighting coefficient by matching a preset weighting coefficient table according to the data source type.

[0191] The time difference calculation unit is used to obtain the timestamp of the semantic unit and calculate the time difference with the current time.

[0192] The freshness decay factor calculation unit is used to calculate the freshness decay factor based on the time difference using a preset exponential decay function, where the exponential decay function is expressed as follows:

[0193]

[0194] in, As a freshness decay factor, The decay rate constant is This is the time difference;

[0195] The edge dynamic weight fusion unit is used to perform multi-dimensional fusion of the basic association strength value, reliability weight coefficient, and freshness decay factor to generate edge dynamic weights, as shown in the following formula:

[0196]

[0197] in, For edge dynamic weights, For reliability weighting coefficients, It is the freshness decay factor;

[0198] The weight mapping unit is used to dynamically map the edge weights to the corresponding edges in the initial knowledge graph, thereby generating an enterprise risk knowledge graph.

[0199] In one embodiment, the risk analysis module 14 includes:

[0200] The temporal embedding generation unit is used to input the node feature vectors and edge dynamic weights of the enterprise risk knowledge graph into the temporal graph attention network, and aggregate the neighbor node information through a multi-layer time-aware attention mechanism to generate the temporal embedding representation of the node;

[0201] The risk path extraction unit is used to identify high-risk nodes from the temporal embedded representation, and extract risk transmission paths within a preset number of hops from the enterprise risk knowledge graph, starting from the high-risk nodes.

[0202] The path encoding unit is used to perform path encoding operations on the node embeddings and edge weights of each risk transmission path to generate path feature vectors;

[0203] The risk feature fusion unit is used to concatenate and fuse the temporal embedded representation of a node with the path feature vector to generate a risk feature vector.

[0204] In one embodiment, the risk warning module 15 includes:

[0205] The risk probability calculation unit is used to input the risk feature vector into the pre-trained risk classifier and calculate the risk probability score of each node through a multilayer perceptron and a self-attention mechanism.

[0206] The risk signal node determination unit is used to determine nodes whose risk probability scores exceed a preset alarm threshold as risk signal nodes.

[0207] The path salience calculation unit is used to calculate the salience index of the transmission path corresponding to the risk signal node based on the path features in the risk feature vector.

[0208] The data confidence calculation unit is used to calculate the data confidence score of risk signal nodes based on the reliability weight coefficient and freshness decay factor of the corresponding edge in the enterprise risk knowledge graph.

[0209] The risk level classification unit is used to input the risk probability score, significance index and data confidence score into the pre-trained multidimensional risk classification model for classification processing to obtain the risk level classification result;

[0210] The early warning information generation unit is used to generate corresponding risk early warning information based on the risk level classification results. The risk early warning information includes risk type, scope of impact, and confidence level label.

[0211] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the knowledge graph-based financial risk early warning method as described above.

[0212] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0213] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0214] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A financial risk early warning method based on knowledge graphs, characterized in that, The method includes: The acquired multi-source heterogeneous data is subjected to structured semantic parsing to obtain semantic units; Based on the semantic units, risk association relationships between entities are established through cross-modal semantic alignment technology, generating risk relationship triples; Based on the semantic units and the risk relationship triples, a corporate risk knowledge graph is constructed. Graph neural network analysis is performed on the enterprise risk knowledge graph to extract risk feature vectors; Based on the risk feature vector, risk signals are identified and confidence levels are assessed, and risk warning information is output. The step of establishing risk association relationships between entities based on the semantic units and generating risk relationship triples through cross-modal semantic alignment technology includes: Obtain a predefined enterprise financial risk domain ontology, wherein the enterprise financial risk domain ontology includes entity type definitions and legal relationship constraint rules between entity types; wherein the legal relationship constraint rules are used to limit the allowed combinations of relationship types between entity types; Based on the enterprise financial risk domain ontology, entity type mapping is performed on the semantic units to generate a set of mapped semantic units; Based on a pre-trained domain-adaptive semantic embedding model, the numerical descriptions in the semantic units are vectorized to generate numerical feature vectors, and the text descriptions in the semantic units are vectorized to generate text feature vectors. The numerical feature vector and the text feature vector are projected into a unified semantic space to generate a cross-modal semantic vector set, and the cosine similarity of each pair of vectors from different sources in the cross-modal semantic vector set is calculated. When the cosine similarity is greater than the dynamic domain threshold and the combination of ontology entity types of the semantic unit conforms to the legal relation constraint rules, the relation predicate is determined according to the combination of ontology entity types, and a risk relation triplet is generated with the semantic unit as the subject and the relation predicate as the relation. Specifically, when the cosine similarity exceeds the dynamic domain threshold and conforms to the legal relationship constraint rules, triple instantiation is performed to generate the risk relationship triple, including: The dynamic neighborhood threshold is calculated using the following formula: ； ； in, The dynamic neighborhood threshold, This indicates finding the minimum value. Let be the probability density function of the normal distribution. This represents the mean of the similarity distribution of real risk samples. The standard deviation of the similarity distribution of real risk samples. The mean of the similarity distribution of the interference samples. This represents the standard deviation of the similarity distribution of the interference samples.

2. The method according to claim 1, characterized in that, The process of performing structured semantic parsing on the acquired multi-source heterogeneous data to obtain semantic units includes: The system acquires structured financial statement data, listed company announcement data, and industry public opinion text, and performs data preprocessing to generate the multi-source heterogeneous data. The structured financial statement data is parsed to generate semantic units for the financial statements. The data of the listed company's announcements is segmented into text, the text of the audit report section is extracted, and the text of the audit report section is used to perform named entity recognition on the text of the audit report section using a financial pre-trained language model to generate an audit entity set. The financial pre-trained language model is a FinBERT model trained based on historical financial corpus. The audit entity set is subjected to relation extraction processing to generate audit semantic units; Event trigger words are detected in the industry public opinion text to generate an initial event set; Fine-grained sentiment analysis is performed on the initial event set to generate public opinion semantic units; The semantic unit is obtained by encapsulating the financial statement semantic unit, the audit semantic unit, and the public opinion semantic unit.

3. The method according to claim 1, characterized in that, The construction of an enterprise risk knowledge graph based on the semantic units and the risk relationship triples includes: Using the semantic units as knowledge graph nodes and the risk relationship triples as knowledge graph edges, a graph structure initialization process is performed to generate an initial knowledge graph. Semantic similarity scores are extracted from the risk relationship triples to generate basic association strength values; Determine the reliability weight coefficient by matching the preset weight coefficient table according to the data source type; Obtain the timestamp of the semantic unit and calculate the time difference with the current time; Based on the time difference, a freshness decay factor is calculated using a preset exponential decay function, wherein the exponential decay function is expressed as follows: ； in, The freshness decay factor is... The decay rate constant is The time difference value; The basic association strength value, the reliability weight coefficient, and the freshness decay factor are fused in multiple dimensions to generate dynamic edge weights, as shown in the following formula: ； in, The basic correlation strength value, The dynamic weight of the edge, This refers to the reliability weighting coefficient; The dynamic weights of the edges are mapped to the corresponding edges of the initial knowledge graph to generate the enterprise risk knowledge graph.

4. The method according to claim 3, characterized in that, The step of performing graph neural network analysis on the enterprise risk knowledge graph to extract risk feature vectors includes: The node feature vectors and edge dynamic weights of the enterprise risk knowledge graph are input into a temporal graph attention network. A multi-layer time-aware attention mechanism is used to aggregate neighbor node information, generating a temporal embedding representation of the nodes. High-risk nodes are identified from the temporal embedding representation, and risk transmission paths within a preset number of hops are extracted from the enterprise risk knowledge graph starting from the high-risk nodes. Perform path encoding operations on the node embeddings and edge weights of each risk transmission path to generate path feature vectors; The temporal embedding representation of the node is concatenated and fused with the path feature vector to generate the risk feature vector.

5. The method according to claim 3, characterized in that, The step of performing risk signal discrimination and confidence assessment based on the risk feature vector, and outputting risk warning information, includes: The risk feature vector is input into a pre-trained risk classifier, and the risk probability score of each node is calculated through a multilayer perceptron and a self-attention mechanism. Nodes whose risk probability scores exceed a preset alarm threshold are identified as risk signal nodes; Based on the path features in the risk feature vector, the significance index of the transmission path corresponding to the risk signal node is calculated; Based on the reliability weight coefficient and the freshness decay factor of the corresponding edge in the enterprise risk knowledge graph, calculate the data confidence score of the risk signal node; The risk probability score, the significance index, and the data confidence score are input into a pre-trained multidimensional risk classification model for classification processing to obtain the risk level classification result. Based on the risk level classification results, corresponding risk warning information is generated, which includes risk type, scope of impact, and confidence level label.

6. A knowledge graph-based financial risk early warning device, used to implement the method according to any one of claims 1 to 5, characterized in that, The device includes: The semantic parsing module is used to perform structured semantic parsing on the acquired multi-source heterogeneous data to obtain semantic units; The cross-modal alignment module is used to establish risk association relationships between entities based on the semantic units and generate risk relationship triples through cross-modal semantic alignment technology. The graph construction module is used to construct an enterprise risk knowledge graph based on the semantic units and the risk relationship triples; The risk analysis module is used to perform graph neural network analysis on the enterprise risk knowledge graph and extract risk feature vectors; The risk warning module is used to identify risk signals and assess confidence levels based on the risk feature vector, and output risk warning information. The cross-modal alignment module includes: Obtain a predefined enterprise financial risk domain ontology, wherein the enterprise financial risk domain ontology includes entity type definitions and legal relationship constraint rules between entity types; wherein the legal relationship constraint rules are used to limit the allowed combinations of relationship types between entity types; Based on the enterprise financial risk domain ontology, entity type mapping is performed on the semantic units to generate a set of mapped semantic units; Based on a pre-trained domain-adaptive semantic embedding model, the numerical descriptions in the semantic units are vectorized to generate numerical feature vectors, and the text descriptions in the semantic units are vectorized to generate text feature vectors. The numerical feature vector and the text feature vector are projected into a unified semantic space to generate a cross-modal semantic vector set, and the cosine similarity of each pair of vectors from different sources in the cross-modal semantic vector set is calculated. When the cosine similarity is greater than the dynamic domain threshold and the combination of ontology entity types of the semantic unit conforms to the legal relation constraint rules, the relation predicate is determined according to the combination of ontology entity types, and a risk relation triplet is generated with the semantic unit as the subject and the relation predicate as the relation. Specifically, when the cosine similarity exceeds the dynamic domain threshold and conforms to the legal relationship constraint rules, triple instantiation is performed to generate the risk relationship triple, including: The dynamic neighborhood threshold is calculated using the following formula: ； ； in, The dynamic neighborhood threshold, This indicates finding the minimum value. Let be the probability density function of the normal distribution. This represents the mean of the similarity distribution of real risk samples. The standard deviation of the similarity distribution of real risk samples. The mean of the similarity distribution of the interference samples. This represents the standard deviation of the similarity distribution of the interference samples.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 5.

Images