Enterprise risk assessment method and system based on large language model and high-frequency graph convolution

Abstract

The invention relates to the field of enterprise risk assessment, and discloses an enterprise risk assessment method and system based on a large language model and high-frequency graph convolution, and the method comprises the steps: S1, building an enterprise risk association graph: generating a structured joint weight between enterprises, and building the enterprise risk association graph; fusing the semantic weight of the large language model and the structured joint weight into a final edge weight; jointly inputting the fused risk map and the multi-modal feature into two-stage attention to generate a multi-modal attention feature; and S2, enterprise risk assessment: on the enterprise risk association graph endowed with the edge weight, carrying out information propagation and aggregation by adopting high-frequency graph convolution. According to the method, edge-level gating and fusion edge weight guide risk information to be preferentially diffused along a more credible and more relevant path, high-frequency signals such as financial mutation, negative news and major events can be quickly captured, and blind diffusion of the information in a network is reduced. And the amplification effect of market systematic impact is reduced.

Description

Technical Field

[0001] This application belongs to the field of enterprise risk assessment, and specifically relates to enterprise risk assessment methods and systems based on large language models and high-frequency graph convolution. Background Technology

[0002] In the context of increasingly complex corporate economic activities, corporate risk is influenced by a combination of factors, including financial condition, equity structure, news, and policy changes. To improve the foresight and accuracy of corporate risk identification, emerging technologies such as graph neural networks and large language models are gradually being introduced into corporate risk assessment tasks. However, corporate multimodal data exhibits significant differences in structure and semantics, and individuals have varying degrees of dependence on different information modalities. Furthermore, corporate risk signals often erupt in the form of localized mutations, posing challenges to existing methods in terms of feature representation and sensitivity response.

[0003] Current enterprise risk assessment methods still face several technical bottlenecks. Existing enterprise risk propagation mechanisms generally rely on static structural rules or simplified similarity matching, making it difficult to model semantically assisted potential risk transmission paths in complex enterprise networks, and lacking intelligent reasoning and precise quantification of enterprise risk weights. Secondly, graph neural networks are prone to oversmoothing during feature propagation, making it difficult to effectively perceive high-frequency risk signals such as sudden and drastic fluctuations preceding enterprise risks, resulting in model lag and insufficient sensitivity. These problems collectively limit the model's ability to characterize local anomalies, individual characteristics, and semantically guided propagation mechanisms, making it difficult to meet the actual needs for accurate identification, interpretation, and assessment of enterprise risks in complex financial environments. Summary of the Invention

[0004] To address or at least alleviate one or more of the above problems, a method and system for enterprise risk assessment based on large language models and high-frequency graph convolution is provided. This method overcomes the shortcomings of relying solely on structural rules or single static indicators, which cannot characterize semantically driven risk transmission, and effectively solves the homogenization (oversmoothing) problem that easily occurs in multi-layer propagation of graph neural networks.

[0005] To achieve the above objectives, in accordance with the first aspect of this application, a method for enterprise risk assessment based on large language models and high-frequency graph convolution is provided, comprising: S1. Construct a risk correlation map for enterprises; Based on multidimensional relationship data between enterprises, a risk relationship graph of enterprises is constructed with enterprises as nodes; the comprehensive risk score of enterprise nodes in the risk relationship graph is calculated; based on the comprehensive risk score of enterprise nodes, high-risk related enterprises are screened according to a predefined risk score threshold, and structured weights are assigned to different types of enterprise relationships based on the statistical distribution of high-risk related enterprises; for enterprise pairs with multiple relationships, the structured weights of each type of enterprise relationship are integrated to generate a structured joint weight between enterprises and construct an enterprise risk relationship graph. Based on a three-hop association subgraph and multimodal features, the multimodal features include at least corporate news, economic indicators, and event descriptions. A large language model is used to infer semantic weights. The semantic weights of the large language model are mapped to quantiles with the structured joint weights and then weighted and fused into final edge weights. The fused risk graph and the multimodal features are input into a two-level attention system. Intramodal feature extraction is performed on the multimodal features, and cross-modal attention fusion is performed between modalities. The multimodal features of each enterprise are weighted and combined according to their contribution to risk, thereby generating multimodal attention features with individual enterprise preference characteristics. S2. Enterprise Risk Assessment: On the enterprise risk association graph with the edge weights, high-frequency graph convolution is used for information propagation and aggregation. During message transmission, the high-frequency graph convolution dynamically generates edge-level gating coefficients based on the current multimodal attention features of the nodes at both ends of the edge. These edge-level gating coefficients and the edge weights are then used to collaboratively modulate the propagated message, outputting the enterprise's risk assessment result.

[0006] To achieve the above objectives, a second aspect of this application provides an enterprise risk assessment system based on a large language model and high-frequency graph convolution, the enterprise risk assessment system comprising: Enterprise Risk Association Graph Construction Unit: Used to construct enterprise risk association graphs. Based on multidimensional relationship data between enterprises, a risk relationship graph of enterprises is constructed with enterprises as nodes; the comprehensive risk score of enterprise nodes in the risk relationship graph is calculated; based on the comprehensive risk score of enterprise nodes, high-risk related enterprises are screened according to a predefined risk score threshold, and structured weights are assigned to different types of enterprise relationships based on the statistical distribution of high-risk related enterprises; for enterprise pairs with multiple relationships, the structured weights of each type of enterprise relationship are integrated to generate a structured joint weight between enterprises and construct an enterprise risk relationship graph. Based on the three-hop association subgraph and multimodal information, the semantic weights of the large language model are obtained through large language model reasoning. The semantic weights of the large language model are mapped to the structured joint weights using quantiles and then fused into the final edge weights. The fused risk graph and multimodal features, including corporate news, economic indicators, and event descriptions, are input into the risk model. Intramodal feature extraction is performed on each multimodal feature, and cross-modal attention fusion is performed between modalities. The economic status of the enterprise, news information, and event semantics are weighted and combined according to their contribution to risk, thereby generating multimodal attention features with individual enterprise preference characteristics. Enterprise risk assessment unit, used for enterprise risk assessment: On the enterprise risk association graph with the edge weights, high-frequency graph convolution is used for information propagation and aggregation. During message transmission, high-frequency graph convolution dynamically generates edge-level gating coefficients based on the current multimodal attention features of the nodes at both ends of the edge, and uses the edge-level gating coefficients and the edge weights to coordinately modulate the propagated message and output the enterprise's risk assessment results.

[0007] To achieve the above objectives, a third aspect of this application provides a computer-readable storage medium storing a computer program, which, when executed by a processor, is used to implement the enterprise risk assessment method based on a large language model and high-frequency graph convolution as described above.

[0008] By adopting the above technical solution, this application has the following beneficial effects compared with the prior art: In this application, a high-frequency graph convolution and edge-gated propagation mechanism are adopted to avoid the coarse-grained averaging that treats all neighbors equally in traditional graph convolution. This allows for more precise screening and differentiated weighting of risk signal transmission paths in multimodal heterogeneous graphs, thereby effectively alleviating the problem of excessive feature smoothing and significantly improving the model's ability to model sudden anomalies and implicit transmission chains.

[0009] In this application, multi-source enterprise information is uniformly modeled on an enterprise risk association graph, and combined with semantic reasoning and graph propagation of a large language model, enterprise risks can be comprehensively and timely identified and forward-looking warnings can be generated, thereby improving the accuracy and timeliness of enterprise risk assessment.

[0010] In this application, the source of risk edge weights is decomposable, meaning that both structural relationships and semantic reasoning are traceable, and the propagation path and node modal contributions are clear and auditable, improving interpretability and robustness; it also has stronger robustness to data missingness, noise, and annual distribution changes.

[0011] The specific embodiments of this application will be described in further detail below with reference to the accompanying drawings. Attached Figure Description

[0012] The accompanying drawings, which form part of this application, are used to provide a further understanding of the application. The illustrative embodiments and descriptions of the application are used to explain the application, but do not constitute an undue limitation of the application. Obviously, the drawings described below are merely some embodiments, and those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0013] In the attached diagram: Figure 1 This is a flowchart illustrating the enterprise risk assessment method based on a large language model and high-frequency graph convolution in this specific implementation. Figure 2 This is a schematic diagram of the system flow of the enterprise risk assessment method based on large language model and high-frequency graph convolution in this specific embodiment; Figure 3 This is a schematic diagram of the three-hop association subgraph of the enterprise risk assessment method based on a large language model and high-frequency graph convolution in this specific embodiment; Figure 4 This is a schematic diagram of the system architecture of the enterprise risk assessment system based on a large language model and high-frequency graph convolution in this specific embodiment. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions in the embodiments will be clearly and completely described below with reference to the accompanying drawings. The following embodiments are used to illustrate this application, but are not intended to limit the scope of this application.

[0015] Please see Figure 1 , Figure 2 , Figure 3 This application provides a method for enterprise risk assessment based on large language models and high-frequency graph convolution, including: S1. Construct a risk correlation map for enterprises; Based on multidimensional relationship data between enterprises, a risk relationship graph of enterprises is constructed with enterprises as nodes; the comprehensive risk score of enterprise nodes in the risk relationship graph is calculated; based on the comprehensive risk score of enterprise nodes, high-risk related enterprises are screened according to a predefined risk score threshold, and structured weights are assigned to different types of enterprise relationships based on the statistical distribution of high-risk related enterprises; for enterprise pairs with multiple relationships, the structured weights of each type of enterprise relationship are integrated to generate a structured joint weight between enterprises and construct an enterprise risk relationship graph. Based on a three-hop association subgraph and multimodal features, the multimodal features include at least corporate news, economic indicators, and event descriptions. A large language model is used to infer semantic weights. The semantic weights of the large language model are mapped to quantiles with the structured joint weights and then weighted and fused into final edge weights. The fused risk graph and the multimodal features are input into a two-level attention system. Intramodal feature extraction is performed on the multimodal features, and cross-modal attention fusion is performed between modalities. The multimodal features of each enterprise are weighted and combined according to their contribution to risk, thereby generating multimodal attention features with individual enterprise preference characteristics. S2. Enterprise Risk Assessment: On the enterprise risk association graph with the edge weights, high-frequency graph convolution is used for information propagation and aggregation. During message transmission, high-frequency graph convolution dynamically generates edge-level gating coefficients based on the current multimodal attention features of the nodes at both ends of the edge, and uses the edge-level gating coefficients and the edge weights to coordinately modulate the propagated message and output the enterprise's risk assessment results.

[0016] It should be noted that the execution entity of the enterprise risk assessment method based on large language models and high-frequency graph convolution in this embodiment is an enterprise risk assessment system based on large language models and high-frequency graph convolution. This system can be an electronic device, a component in an electronic device, an integrated circuit, or a chip. The electronic device can be a mobile electronic device or a non-mobile electronic device. For example, mobile electronic devices can be mobile phones, tablets, laptops, PDAs, in-vehicle electronic devices, wearable devices, etc., and non-mobile electronic devices can be servers and personal computers, etc., which are not specifically limited in this application. The following description uses a server as the execution entity to illustrate the enterprise risk assessment method based on large language models and high-frequency graph convolution in this embodiment.

[0017] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0018] For example, this embodiment uses the CSMAR financial database to collect business relationship information such as equity structure, controlling relationships, and investment relationships of A-share listed companies in China from 2019 to 2023. Based on this business relationship information, five enterprise risk relationship maps for different years are constructed to simulate the relationship network between enterprises. All nodes are uniquely identified by securities codes or personnel IDs to ensure identifiability and cross-year traceability in the map. On this basis, structured edges reflecting economic interests, control relationships, and governance relationships are constructed according to the disclosed investment, shareholding, controlling, and employment relationships between enterprises. All structured information is uniformly converted into a "entity-relationship-entity" triple format and constructed as independent map slices by year to ensure consistency over time.

[0019] It should be noted that the graph data used in this embodiment comes from various relationships between enterprises, such as holding, shareholding, and investment, aiming to capture potential risk propagation paths and structural dependencies between enterprises. The graph is constructed using years as time slices, and is based on multi-entity, multi-relationship data of enterprises and their shareholders, executives, controllers, and investors, to build a graph of enterprise risk association objects in the Neo4j graph database.

[0020] For example, please refer to Figure 2 First, a risk association graph of enterprises is obtained. This graph is constructed based on entities such as enterprises, shareholders, actual controllers, and senior executives, and edge relationships are built according to structured relationships such as investment, shareholding, controlling stake, and employment to depict the association structure of enterprises within a year. Next, the comprehensive risk score of enterprise nodes in the risk association graph is calculated. Based on the comprehensive risk score, high-risk associated enterprises are screened according to predefined risk score thresholds. For enterprise pairs with multiple association relationships, the structured weights of each association relationship type are integrated to generate a structured joint weight between enterprises and construct the enterprise risk association graph. Figure 2 In the enterprise risk association graph on the left, the numerical values ​​(0.3, 0.27, 0.22, 0.15, 0.32) labeled on the sides represent the structured joint risk weights between enterprises. Based on the three-hop association subgraph and multimodal features, the semantic weights of the large language model are obtained using large language model inference. Figure 2 In this context, 0.22, 0.28, and 0.14 represent the risk weights obtained by the large language model based on dimensions such as graph relationships, financial data, and external events. The large language model integrates these risk weights to obtain a semantic weight of 0.28 for companies A and B, which is used for subsequent weight mapping and fusion. The semantic weights of the large language model are then mapped to quantiles with the structured joint weights and weighted and fused to form the final edge weights. Figure 2The final weight between companies A and B is determined by mapping the semantic weight (0.28) of the large language model to the structured joint weight (0.32) using quantiles, and then weighted and fused to obtain the final edge weight (0.27). The risk graph with the final edge weight is then input into the risk model along with multimodal features, including company news, economic indicators, and event descriptions. The model extracts intra-modal features from the news text, financial indicators, and event records, identifying the key information that best represents the company's operating status or changes in the external environment within each modality, forming a stable modal representation. The model further performs cross-modal attention fusion, weighting and combining the company's economic status, news information, and event semantics according to their contribution to risk, thereby generating multimodal attention features with individual company preferences. High-frequency graph convolution dynamically generates gating coefficients based on the current multimodal features of the companies at both ends of the edge. Figure 2 The edges connecting neighboring companies B, C, and D to company A are exemplarily labeled with "high" and "low" values ​​to represent the two possible outcomes of the edge during actual gating modulation. This coefficient, along with the fused edge weights, is applied to the neighborhood information to emphasize paths with stronger risk relevance, while suppressing weakly correlated or noisy paths. For example... Figure 2 The circular structure on the left illustrates example edge weights between target company A and its neighboring companies (B, C, D) in a risk graph with final edge weights, such as 0.30, 0.27, and 0.15, representing the strength of risk transmission from different neighbors to company A. Finally, the high-frequency graph convolution outputs a company representation that fuses structural relationships, semantic evidence, and high-frequency risk signals for subsequent risk prediction.

[0021] For example, the comprehensive risk score of enterprise nodes in the enterprise risk association object graph is first calculated. To construct a risk guidance mechanism for enterprise structural relationships, this paper designs a one-dimensional comprehensive risk score to quantify the structural risk level of enterprise nodes. The score is constructed based on seven structured risk factors, including the total number of litigation events, the number of arbitrations, audit opinions, the number of violations, litigation labels, arbitration labels, and major violation labels. Considering the collinearity and uneven distribution among the original variables, this paper adopts a linear modeling strategy with regularization terms to learn the combined weights of risk factors from historical samples, ultimately forming a fusion score to reflect the comprehensive risk level of enterprises.

[0022] Based on the aforementioned comprehensive risk score, the 80th percentile of the annual score is used as the risk threshold to define a set of high-risk enterprises, which is then used to construct structural edge weights to quantify the potential risk propagation intensity. This quantile threshold strategy can expand the identification range of potential high-risk nodes and improve the coverage of structural risk estimation. Furthermore, it has been widely validated in tasks such as credit scoring, risk control modeling, and graph anomaly detection, demonstrating good practical stability and empirical support.

[0023] In some embodiments, based on multiple risk characteristics of enterprises, a linear regression model is used to calculate the comprehensive risk score of each enterprise node; a risk score threshold is set to screen out high-risk enterprises; for each type of association, the edge in the association type where both ends are high-risk enterprises is marked as a high-risk edge; the proportion of high-risk edges in the association type is defined as the structured weight of the association type; for the same pair of enterprises with multiple associations, a hybrid strategy combining geometric mean and maximum value is used to fuse the structured weights of each association type to generate a structured joint weight that characterizes the strength of the enterprise's association with the overall risk.

[0024] For example, the calculation of structured weights is completed in two steps: First, the seven risk factors (number of litigation events, number of arbitrations, audit opinions, number of major violations, litigation labels, arbitration labels, and violation event labels) are integrated into a comprehensive risk score for the enterprise using linear regression; a quantile threshold q (0.8) is set for the year to identify the set of "high-risk enterprises". Subsequently, the proportion of edges with "high risk at both ends" in the edge set of the type is statistically analyzed according to the relationship type, and this proportion is used as the constant weight of the type for the year and assigned to all candidate edges of the type. Second, for multiple records of the same enterprise under multiple relationships, a hybrid strategy of "geometric mean + maximum value" is used to synthesize the structured joint weight of the enterprise pair, which is an important indicator reflecting the risk transmission capability under multiple structural relationships. The range of the structured joint weight is [0,1].

[0025] For example, the proportion of high-risk edges among high-risk firms in various structural relationships within the enterprise graph is statistically analyzed to construct structural risk weights at the relationship type level. When constructing structural risk at the relationship type level, let... For any relation type The structural risk edge of this type of relationship is defined as follows: ; in, Represents an enterprise node. This indicates that all elements in the graph belong to the relation type. The set of edges. This ratio measures the type of relation. The proportion of high-risk edges between high-risk nodes can be regarded as the potential risk transmission capacity of this relationship in the graph.

[0026] To unify the understanding of enterprises To address the risk weights of various structural relationships, this embodiment introduces a hybrid fusion mechanism, combining the combined effects of a single dominant risk relationship and multiple risk relationships to construct a robust edge-level structural risk representation.

[0027] Specifically, for the set of relation types that exist between node pairs The structured joint weights are defined as follows: ; in, This is the type with the highest structure weight among all relationship types, emphasizing the most dominant single structural risk path. [0,1] is the fusion coefficient, used to adjust the weight balance between the dominant risk relationship and the cumulative risk under the combined effect of multiple relationships. When When =1, fusion relies entirely on the strongest relationship channel, suitable for focusing on high-risk edge situations; when When =0, fusion relies entirely on the overall level of multiple relations, placing greater emphasis on consistency risks within the structure. This embodiment defaults to 0.5. In this mechanism, the geometric mean term GeoMean is defined as follows: ; in, Indicates the first Structured weights for class-based relationships. For node pairs The total number of relation types. A larger GeoMean value indicates that multiple relations carry strong risks, reflecting the synergistic accumulation effect of structural risk factors. Compared to the arithmetic mean, the geometric mean is more sensitive to extreme values, only showing a significant increase when multiple relation types are all high-risk. Therefore, it is more suitable for characterizing the accumulation mechanism of systemic risk in networks with overlapping multi-relationship structures. To further enhance the expressive power of graph structures for the connection characteristics of high-risk enterprises, this paper maps various heterogeneous structural relations into edge-level risk weights between enterprise node pairs based on the integration of risk scoring and structural statistics. Furthermore, it constructs an enterprise risk association graph based on this.

[0028] In some embodiments, a set of core enterprises is first identified on the enterprise risk association object map, and then the structural data and multimodal data of enterprise pairs are organized using a three-hop association subgraph as a carrier. The data are then input into a large language model for semantic reasoning, and the resulting risk propagation probability is the semantic weight of the large language model.

[0029] For example, eigenvector centrality is used to identify core enterprises. Starting with pairing core enterprises, the three-hop association subgraphs of each pair of core enterprises in the annual graph are retrieved, and multimodal information such as economic indicators, litigation, major violations, and related news are collected. The large language model performs semantic reasoning by analyzing the structural paths of investment, shareholding, controlling stake, or employment between enterprises extracted through the three-hop subgraphs based on preset expert prompts. It also combines multi-source text data related to the enterprises (including financial descriptions, news events, and public announcements) to infer potential risk transmission relationships between enterprises. The model outputs the probability value of its risk transmission and generates risk association interpretations based on structural paths and textual evidence. The probability is defined as the semantic weight of the large language model for the directed pairs of enterprises.

[0030] Specifically, the generation of core risk enterprises and candidate pairs involves the following steps: When constructing the risk semantic reasoning task, the systemic core nodes are first identified in the annual enterprise graph based on the feature vector centrality index, and the top 5% of enterprises are selected as core risk enterprises. This strategy is based on the theoretical assumption of "concentrated risk flow" in complex networks, meaning that a few structurally central enterprises often bear a higher potential risk propagation capacity. Therefore, prioritizing the generation of enterprise pairs between these nodes can cover a larger range of systemic impact with lower semantic computation costs. Then, based on pairing core enterprises, the corresponding subgraphs within three hops are retrieved from the annual graph. The scope mainly covers typical risk channels between enterprises such as "holding, shareholding, investment, and employment." Each enterprise pair is associated with its three-hop subgraph, serving as an important input background for semantic reasoning in the large language model.

[0031] For example, in the process of screening core risk enterprises and extracting three-hop association subgraphs, based on feature vector centrality, nodes with prominent structural influence are identified from the annual enterprise graph, and a compact, task-related local subgraph structure is constructed. This provides a semantically clear structural context for subsequent Large Language Model (LLM) inference, thereby improving the quantitative accuracy and explanatory power of risk propagation relationships.

[0032] Let the annual corporate chart be in This refers to all listed companies. This includes four types of directed structural relationships: investment, shareholding, controlling stake, and employment. To assess the structural importance of enterprise nodes in the graph, this paper employs the eigenvector centrality method. This method not only measures the number of direct connections to a node but also integrates the centrality scores of its neighboring nodes, enabling more effective identification of key nodes with spillover influence in the system. Let... The corresponding centrality vector satisfy: ;

[0033] in, The largest eigenvalue, Indicates enterprise The structural influence score. After sorting, select the top-ranked items. The nodes form a core risk enterprise set : ;

[0034] in, =5%. To avoid introducing data illusions, a type of intermediary node with extremely high connectivity but lacking actual business activity was removed during the construction process. After completing the screening of core risk enterprises, this paper focuses on each pair of core enterprises. The structural subgraph within three hops of the original graph is retrieved to construct the semantic context of risk propagation. ; in, This represents a three-hop association subgraph between core enterprise u and core enterprise v in the risk association object graph. A three-hop association subgraph refers to a structured subgraph in the risk association object graph that takes all connection paths between the two core enterprises with a length not exceeding three and consists of all nodes covered by these paths and their corresponding edges. Let represent the set of nodes and the set of edges in the subgraph, respectively.

[0035] This subgraph retains all nodes and edges along the paths, covering various relationship types such as holding, investment, shareholding, and employment, to enrich structural semantic information. To improve the accuracy of semantic reasoning, redundant nodes, multiple paths, and structurally ambiguous edges are filtered and regularized to ensure that the subgraph structure is compact and semantically clear.

[0036] It should be noted that, to further aid in understanding the risk core enterprise screening and three-hop context subgraph extraction mechanism proposed in this paper, Figure 3 This is a structural diagram of a three-hop relational subgraph sample, where nodes represent different types of entities, and edges represent the relationships between entities. In this example, red nodes (Company A, Company B) are core enterprises with high structural importance; blue nodes (Company C, Company D, Company E, Company F, and Person 1) represent non-core entities that have direct or indirect relationships with the core enterprises, including different types of nodes such as enterprises, shareholders, or senior executives.

[0037] The different edges in the diagram indicate the specific types of relationships, including "investment relationship," "shareholding relationship," and "employment relationship." "Investment relationship" represents one company investing in another; "shareholding relationship" represents a shareholder or enterprise holding shares in an invested company; and "employment relationship" represents a connection of positions between individuals within companies. The direction of the arrows indicates the directional nature of the relationship, such as from the investor to the invested company, or from the employee to their employer.

[0038] Specifically, the system packages multimodal data and designs prompts: When constructing semantic reasoning input for enterprise pairs, the system organizes and packages four types of core data information for each enterprise pair, forming a unified input template. First, there is graph data, directly presented as images. Second, there is economic data, referencing each enterprise's key financial indicators for the year, using mean-filling to fill in missing values, ensuring data consistency and standardized content. Finally, there is event and news evidence, including statistical characteristics of lawsuits and major violations involved in the year, as well as text content such as news headlines and summaries, which are cleaned and deduplicated to extract representative content.

[0039] For example, to improve the stability and accuracy of large language model inference, this method constructs expert role-oriented prompts for systemic financial risk analysis scenarios when calling the large language model: The prompt explicitly specifies that the model plays the role of a "systemic financial risk expert," possessing knowledge graph reasoning and multimodal financial data analysis capabilities, and explains that it will simultaneously provide information such as the three-hop relationship subgraph between company A and company B, annual economic indicators, and recent events and news. The prompt details the tasks the model needs to complete, including: determining whether company A may transmit risk to company B based on the structural relationships of the three-hop relationship subgraph, financial indicators, and external semantic information; inferring potential transmission paths and impact mechanisms; and providing refined risk probabilities in the 0-1 range based on a pre-set ten-level risk rating table. To facilitate subsequent program parsing, a fixed structured output format is pre-set at the end of the prompt, requiring the model to provide four fields: "Does a risk transmission relationship exist?", "Explanation of risk transmission", "Risk correlation weight (retaining two decimal places)", and "Risk level (A-J)". This achieves automatic extraction and standardized expression of the semantic weights of companies to the large language model.

[0040] In some embodiments, KDE-CDF distribution alignment is used for same-scale calibration: Considering that the numerical range and distribution shape of the large language model output and the structured joint weights are often inconsistent, this embodiment uses KDE-CDF distribution alignment (quantile mapping) for same-scale calibration: Kernel density estimation is performed on the semantic weights of the large language model and the structured joint weights to obtain a smooth distribution, and then their respective cumulative distribution functions (CDF) are calculated; each semantic weight of the large language model is first mapped to its quantile position, and then the structural scale value at the corresponding quantile is retrieved through the inverse CDF on the structural side. This method makes the weights on both sides of the same order of magnitude and more continuous and additive without destroying the monotonicity of the ranking, which is convenient for subsequent fusion; boundary pruning can be performed on extreme values ​​to improve robustness. After obtaining the semantic weights of the large language model at the same scale, they are convexly weighted and fused with the structured joint weights to generate the final edge weights for model training and inference.

[0041] For example, in this embodiment, KDE fitting is performed on the semantic weight set of the large language model output by the Large Language Model (LLM) and the structured joint weights to obtain continuous probability density curves. KDE can effectively smooth the discrete sample distribution into a continuous function, avoiding the step-like problem commonly found in histograms, while improving the interpolationability of numerical distributions, making the subsequent mapping process more stable and reliable. The cumulative distribution functions corresponding to the two distributions are then calculated. and CDF stands for "the percentage of samples less than or equal to a certain value," which is the quantile position of a given value within its distribution. Using this property, the following mapping function can be constructed: ; in, Represents the semantic weights of a large language model. This represents the semantic enhancement edge weights, obtained by mapping the semantic weights of a large language model through quantiles, and scaled to the same as the structured joint weights. Represents the same-scale quantile mapping function; The mapping function first converts semantic edge weights into their quantile positions in the LLM distribution, and then finds the corresponding weight values ​​at the quantiles in the structural edge weight distribution, thus aligning edge weights from different sources in terms of distribution position. This process maintains the monotonicity (order remains unchanged) and continuity (smooth numerical distribution) of the values, ensuring that the mapped semantic weights and structural weights are on the same scale and are additive. After achieving scale unification, this paper employs a linear fusion strategy to construct the final propagation edge weights: ; in, This represents the structured joint weights and fusion coefficients of nodes u and v. [0,1] represents the relative weights of the control structure and semantic edge weights. This embodiment uses the default weights. =0.5, meaning that the structured joint weights and the semantic weights of the large language model after quantile mapping are equally weighted, avoiding bias towards either source and achieving a balance between structural pattern stability and semantic perception capability. The final generated edge weights Assigning structural connections between all enterprises in the graph as inputs to risk weights during the propagation stage of the graph neural network provides more accurate and interpretable edge weight modeling support for downstream risk assessment tasks.

[0042] In some embodiments, the enterprise node multimodal features are calculated: in the multimodal fusion method, two-level attention is used to obtain the final input features of the enterprise node, that is, firstly, lightweight alignment and filtering are performed within the modality, and then weight allocation is performed between the modalities to form a modality preference attention vector; finally, the obtained enterprise node multimodal features are used as the node features of the enterprise.

[0043] For example, each enterprise node is connected to three types of information sources: economic indicators, litigation and violation events, and news texts. For economic indicators, missing values ​​are filled with the mean, without further scaling, standardization, or extreme value removal, to maintain consistency between financial statements and business interpretations. The news modality uses the year's news headlines as text input, performs basic deduplication and empty / duplicate removal, and directly generates fixed-length embeddings through a text encoder. Economic, event, and news modalities are strictly aligned temporally with the annual graph slices. If a modality is missing, a placeholder vector is used in conjunction with a mask to participate in subsequent fusion, ensuring stable training. The economic modality uses an autoencoder (64-dimensional), and news / event modalities use BERT (768-dimensional each), which is then linearly mapped to a unified latent space before entering a two-level attention process.

[0044] For example, please refer to Figure 2 The two-level attention mechanism consists of intra-modal feature extraction and cross-modal attention fusion. For intra-modal feature extraction, for target company A, the system constructs its internal risk representation from three types of information sources: news-internal modality, event-internal modality, and economic-internal modality. First, in the news-internal modality, the system maps multiple news texts related to company A, such as "News 1, News 2, News 3," into news feature vectors using a text encoding network. Second, in the event-internal modality, the system extracts key events occurring in company A, such as "Event 1, Event 2," and encodes them into event feature representations using a text encoding network. Third, in the economic-internal modality, the system extracts inter-company economic relationships related to company A, such as the financial indicators of companies B, C, and D, and constructs a multi-dimensional economic feature vector reflecting the economic connections between companies. Subsequently, the model performs cross-modal attention fusion on the extracted news features, event features, and economic features. Cross-modal attention fusion dynamically learns the weights of different modalities based on their contribution to the target company's risk and merges the three types of modal features into a unified multi-modal attention feature.

[0045] For example, please refer to Figure 2 The two-level attention is divided into intra-modal feature extraction and cross-modal attention fusion: for intra-modal feature extraction, enterprise nodes... Information input with three modalities: structured economic indicators News text vector Event text vector The economic modality also includes the economic characteristics of its structural neighbors after aggregation. To account for the differences in perception methods between different modalities, intra-modal feature extraction is first performed on features of each modality, and a differentiated strategy is adopted for different modalities:

[0046] For news and event modalities, they primarily reflect the company's own news sentiment and risk events. In corporate bankruptcy prediction, risk is endogenous; bankruptcy triggers often originate directly from the company itself rather than related companies. Introducing news information from related companies can easily introduce noise and affect the identification of the true risk source. Therefore, a self-feature modeling strategy is adopted for these two modalities. (For news and event modalities...) , ; in, This is the original vector of the mode of node u; Represents the Sigmoid function; Linearly map it to the unified latent space ; The learnable intramodal attention vector is used to compress the mapped representation into a single score; the score is then nonlinearly introduced via LeakyReLU and then... Pressing down to (0,1) yields node-wise and mode-wise scalar gating. Finally, this gating method is used to scale the mapped vector proportionally to obtain the modal embedding. The calculation formula is as follows: ; This process relies solely on the modal input of the node itself, without introducing cross-modal or neighbor information, and is therefore suitable for reflecting the direct semantic contribution of news / events to the enterprise ontology.

[0047] In economic models, a company's financial condition depends not only on its own operations but also on the influence of its related companies. This modeling strategy is particularly suitable for revealing structural financial risks such as "financial deterioration of holding companies" and "losses of invested companies" when modeling risk transmission paths.

[0048] For economic modes, a graph-based economic mode modeling method was used: for enterprise nodes... First, collect its structural neighbor set. Each neighbor node economic characteristics and the economic characteristics of the nodes themselves. Append it to the end of the list to get the concatenated combined input: ; The symbol ⊕ indicates vector concatenation.

[0049] Let the shared linear mapping matrix be... The neighbor and self features are projected onto a unified semantic space, and the concatenated features are then input into an attention scoring network to calculate the attention weights. : ; in, This is a learnable attention direction vector. This refers to the Sigmoid function, used to compress attention scores to the [0,1] interval. This represents the economic modal characteristics of enterprise node u.

[0050] Attention weight Applied to neighbor representations, a structure-aware representation of the economic mode is obtained: ; in, Represents nodes The stack of economic features of all neighboring nodes with structural connections, after linear mapping, represents the fused representation of neighboring economic semantics in the semantic space. This is the attention gating coefficient calculated jointly from information about the node itself and its neighbors. This output not only preserves the node's own economic semantics but also incorporates the structural economic characteristics of its associated enterprises, thus capturing the influence of local structures while achieving structural perception modeling of the node's economic state.

[0051] For example, please refer to Figure 2 For cross-modal attention fusion, in order to achieve adaptive integration of multi-source risk information, this embodiment introduces a cross-modal attention mechanism in the modal dimension, learns the degree of preference of nodes for different information sources, and generates a unified node representation in the fusion space.

[0052] Specifically, set up enterprise nodes The three modal characteristics are We first project it onto a common semantic space using a shared linear mapping matrix to ensure comparability: ; in, This is used to unify the differences in dimensionality and distribution among different modalities, so that subsequent attention scores can be calculated in a unified space; Describe the space of a real matrix. This represents the dimension after mapping the original features of different modalities to a unified semantic space. Input feature dimension; The symbol "~" represents the modal features projected onto a unified semantic space after linear mapping. "~" is used to represent features after linear transformation or semantic space projection to distinguish them from the original features.

[0053] Next, through a shared attention direction vector All modes are scored and normalized along the mode dimension to obtain the nodes. Attention weights for each modality: ; Among them, symbols This represents the summation index in softmax normalization, used for traversing the set. All modes in the equation, distinguished from the m phase in the molecule; exp represents the natural exponential function. Weights This reflects the degree of node dependence on different modes and satisfies... It represents the attention weight of node u to mode m. If a node is more sensitive to the economic mode, then its attention weight on that mode is... The value will be larger; conversely, if more reliance is placed on news or breaking events, the weight of the news or event modality will be dominant.

[0054] After obtaining the attention weights, an independent linear transformation matrix is ​​configured for each modality. , This represents the dimension of the output feature after the linear transformation. This is used to preserve the individual differences of modalities and to fuse the attention-weighted modal representations into the final node representation: ; Final output Let u be the weighted fusion representation of node u, where the importance of each modality is dynamically adjusted through an attention mechanism.

[0055] In some embodiments, a three-hop subgraph is first extracted from the core enterprise pairs on the annual graph, and multimodal data is fused. The semantic weights of the large language model are then output by the large language model. This probability is then projected onto the structural scale through kernel density estimation-cumulative distribution function quantile mapping, and convexly added with the structured joint weights by a coefficient α to obtain the fused edge weights. Thus, each edge simultaneously possesses objective structural strength and current semantic tendency; these edge weights are multiplied by the edge-level gating in HighConv. The model prioritizes amplifying high-frequency risk signals on reliable channels, effectively mitigating over-smoothing, and outputting enterprise risk probabilities with stronger interpretability and robustness.

[0056] For example, for each edge connecting the source node and the target node, the current representations of the source node and the target node are concatenated; the concatenated representation is passed through a nonlinear transformation function to generate an edge-level gating coefficient with a numerical range of [-1,1]; the edge-level gating coefficient is multiplied by the edge weight to obtain the final message modulation factor of the edge; the final message modulation factor is used to scale the message strength transmitted from the source node to the target node, thereby achieving context-dependent and global weight-dependent collaborative modulation.

[0057] It should be noted that high-frequency graph convolution is introduced during the graph propagation stage to address the homogenization (oversmoothing) problem that easily occurs in GNNs during multi-layer propagation. The core idea of ​​high-frequency graph convolution is to simultaneously utilize the fused edge weights and semantic gating to control the intensity of information passing through each edge, so that "neighbor signals with greater differences" are preserved and amplified, thereby making it more sensitive to sudden, local, and heterogeneous risk signals.

[0058] During message transmission, this embodiment adaptively generates edge-level gating coefficients based on the current representations of the nodes at both ends of the edge. The coefficients range from suppression to amplification. These coefficients are multiplied by the fusion risk weights of the corresponding edges to finely modulate the message strength from adjacent nodes. At the target node, summation and aggregation are used to achieve information aggregation. Edge-level modulation is accomplished collaboratively by dynamic gating and fusion risk weights. For each edge, the gating is obtained by nonlinear transformation of the concatenation of the representations of the current nodes at both ends of the edge. The numerical domain is [-1, 1], and it is a context-sensitive continuous amplification / suppression coefficient: when the representations at both ends are significantly different, it can generate large positive / negative amplitudes, thereby strengthening or canceling neighbor signals during aggregation, reflecting the responsiveness to high-frequency patterns. Complementing this, the edge weights, resulting from the fusion of structured joint weights and LLM weights, characterize the global capacity and semantic consistency of the channel in the dimensions of "structural strength + semantic transmission tendency." This mechanism not only avoids the coarse-grained averaging that treats all neighbors equally in traditional graph convolution, but also enables precise screening and differentiated weighting of risk signal transmission paths in multimodal heterogeneous graphs. This effectively alleviates the problem of excessive feature smoothing and significantly improves the model's ability to model sudden anomalies and hidden transmission chains.

[0059] For example, to more effectively model the localized sudden risk characteristics before corporate bankruptcy, for each edge (𝑢,𝑣), the target node... Its neighboring nodes The features are concatenated and input into the gating network to obtain the edge-level semantic weights: ; in, These are learnable parameters. The weights reflect the contribution of neighboring nodes to the target node's high-frequency propagation, and can dynamically adjust the importance of different neighbors; the symbol " "" indicates a vector concatenation operation, which connects two vectors along their dimensions; This is a bias term used to improve the flexibility of mapping. The semantic gating coefficients, i.e., edge-level semantic weights, are used to dynamically adjust the intensity of high-frequency risk transmission from neighbor node 𝑣 to target node 𝑢. Based on this, the final edge weights from the risk-weighted graph are introduced. This is used to depict the strength of risk relationships between enterprises. Therefore, the messages transmitted along each edge... Defined as: ; Finally, the representation of the target node is obtained by aggregating the weighted messages of all its neighbors: ; This design effectively incorporates risk information from neighboring nodes while preserving the characteristics of the target node itself. Semantic gating. Used to highlight the high-frequency signals represented by local differences, and the final edge weights This ensures the consistency between propagation intensity and risk relationship. By combining these two methods, HighConv (High-Frequency Graph Convolution) can improve the model's accuracy and interpretability in depicting real risk propagation paths while suppressing excessive smoothing. To balance the robust semantic information obtained from cross-modal fusion with the high-frequency information captured by HighConv, residual concatenation and linear projection are performed after high-frequency graph convolution. Cross-modal fusion is represented as... HighConv output is First, we need to assemble the parts: ; in, The concatenated and fused features of node u are then passed through a linear mapping layer and ReLU activation is applied to obtain the final node representation for classification: ; in, The learnable parameter matrix representing the linear projection. Indicates the bias term; The concatenation of the cross-modal fusion representation and the HighConv output is linearly mapped to achieve dimensional alignment and feature weighting integration, thereby enhancing the sensitivity to risk differences while maintaining semantic robustness.

[0060] In some embodiments, the fused node representation is input into a linear classifier, the probability distribution of enterprise risk is output through softmax, and weighted cross-entropy is used as the loss function to address the class imbalance problem.

[0061] For example, the node-level semantic features obtained by multimodal attention fusion are concatenated with the structure-sensitive features enhanced by high-frequency graph convolution at the feature level. Then, the concatenated vector is compressed and reconstructed through a fully connected mapping with non-linear activation to obtain the fused node representation. The fused node representation is input into a linear classifier, and the softmax function outputs a probability distribution of "bankruptcy / non-bankruptcy". During the training phase, weighted cross-entropy is used as the loss function to alleviate the bias caused by class imbalance and improve the overall recognition performance.

[0062] Specifically, for each node The final representation is obtained after the aforementioned modality modeling and convolution. And input it into a linear classifier to obtain the predicted probability. , The final linear classification layer, coupled with a Sigmoid activation function, represents the probability that company u is deemed to have a risk of bankruptcy. Considering the imbalance between positive and negative samples, a weighted cross-entropy loss function is used during training, assigning different weights to positive and negative classes. Let the true label be... The category weights are respectively , The loss function is then defined as: ; in, This represents the set of training nodes. Weights Used to enhance the influence of a minority group (bankrupt companies), The model assigns weights corresponding to the majority class (normal enterprises) to mitigate bias caused by uneven sample distribution. During training, the model parameters are updated via gradient descent to make the predicted distribution closer to the true distribution, ultimately enabling the identification of enterprise bankruptcy risk.

[0063] It should be noted that this embodiment performs feature-level concatenation between the node-level semantic features generated by multimodal attention fusion and the structure-sensitive features enhanced by high-frequency graph convolution. The concatenated vector is compressed and reconstructed through a fully connected mapping layer with non-linear activation. This preserves the global semantics of multimodal collaborative expression while enhancing the responsiveness to high-frequency mutations and local anomalous structures in neighboring enterprise nodes, providing a more discriminative enterprise representation for downstream enterprise risk prediction tasks.

[0064] Based on the same inventive concept, please see Figure 4 This application also provides an enterprise risk assessment system based on a large language model and high-frequency graph convolution, the enterprise risk assessment system comprising: Enterprise Risk Association Graph Construction Unit: Used to construct enterprise risk association graphs. Based on multidimensional relationship data between enterprises, a risk relationship graph of enterprises is constructed with enterprises as nodes; the comprehensive risk score of enterprise nodes in the risk relationship graph is calculated; based on the comprehensive risk score of enterprise nodes, high-risk related enterprises are screened according to a predefined risk score threshold, and structured weights are assigned to different types of enterprise relationships based on the statistical distribution of high-risk related enterprises; for enterprise pairs with multiple relationships, the structured weights of each type of enterprise relationship are integrated to generate a structured joint weight between enterprises and construct an enterprise risk relationship graph. Based on the three-hop association subgraph and multimodal information, the semantic weights of the large language model are obtained through large language model reasoning. The semantic weights of the large language model are mapped to the structured joint weights using quantiles and then fused into the final edge weights. The fused risk graph and multimodal features, including corporate news, economic indicators, and event descriptions, are input into the risk model. Intramodal feature extraction is performed on each multimodal feature, and cross-modal attention fusion is performed between modalities. The economic status of the enterprise, news information, and event semantics are weighted and combined according to their contribution to risk, thereby generating multimodal attention features with individual enterprise preference characteristics. Enterprise risk assessment unit, used for enterprise risk assessment: On the enterprise risk association graph with the edge weights, high-frequency graph convolution is used for information propagation and aggregation. During message transmission, high-frequency graph convolution dynamically generates edge-level gating coefficients based on the current multimodal attention features of the nodes at both ends of the edge, and uses the edge-level gating coefficients and the edge weights to coordinately modulate the propagated message and output the enterprise's risk assessment results.

[0065] Based on the same inventive concept, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the enterprise risk assessment method based on a large language model and high-frequency graph convolution as described above.

[0066] The program product of this application for implementing the above method may employ a portable compact disk read-only memory and include program code, and may run on a terminal device, such as a personal computer. However, the program product of this application is not limited thereto. In this application, the readable storage medium may be any tangible medium containing or storing a program that may be used by or in conjunction with an instruction execution system, system, or device.

[0067] It should be noted that a computer-readable storage medium may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable storage medium may also be any readable medium other than a readable storage medium, capable of sending, propagating, or transmitting programs for use by or in conjunction with an instruction execution system, system, or device. The program code contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0068] The above are merely preferred embodiments of this application and are not intended to limit this application in any way. Although this application has disclosed preferred embodiments as described above, it is not intended to limit this application. Any person skilled in the art can make some modifications or alterations to the above-mentioned technical content to create equivalent embodiments without departing from the scope of the technical solution of this application. The implementation schemes in the above embodiments can also be further combined or replaced. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of this application without departing from the content of the technical solution of this application shall still fall within the scope of this application.

Claims

1. A method for enterprise risk assessment based on large language models and high-frequency graph convolution, characterized in that, include: S1. Construct a risk correlation map for enterprises; Based on multidimensional relationship data among enterprises, a risk relationship map of enterprises with enterprises as nodes is constructed. Calculate the comprehensive risk score of the enterprise nodes in the enterprise risk association object graph; Based on the comprehensive risk score of the enterprise nodes, high-risk associated enterprises are screened according to the predefined risk score threshold, and structured weights are assigned to different types of enterprise associations based on the statistical distribution of high-risk associated enterprises. For enterprise pairs with multiple relationships, the structured weights of each relationship type are integrated to generate structured joint weights between enterprises and construct an enterprise risk relationship graph; Based on the three-hop association subgraph and multimodal features, the semantic weights of the large language model are obtained through large language model inference. The semantic weights of the large language model and the structured joint weights are mapped by quantiles and then fused into the final edge weights. The fused risk map and the multimodal features are input into two-level attention. Intramodal feature extraction is performed on the multimodal features, and cross-modal attention fusion is performed between modalities. The multimodal features of each enterprise are weighted and combined according to the contribution of each multimodal feature to the risk, thereby generating multimodal attention features with enterprise individual preference characteristics. S2. Enterprise Risk Assessment: On the enterprise risk association graph with the edge weights, high-frequency graph convolution is used for information propagation and aggregation. During message transmission, high-frequency graph convolution dynamically generates edge-level gating coefficients based on the current multimodal attention features of the nodes at both ends of the edge, and uses the edge-level gating coefficients and the edge weights to coordinately modulate the propagated message and output the enterprise's risk assessment results.

2. The method according to claim 1, characterized in that, Based on the comprehensive risk score, high-risk related enterprises are screened according to a predefined risk score threshold. Structured weights are assigned to different types of relationships based on the statistical distribution of these high-risk related enterprises. For enterprise pairs with multiple relationships, the structured weights of each relationship type are merged to generate a structured joint weight between the enterprises, including: Based on multiple risk factors of an enterprise, including at least the number of lawsuits, arbitrations, audit opinions, major violations, litigation tags, arbitration tags, and violation tags, a linear regression model is used to calculate the comprehensive risk score for each enterprise node. A risk score threshold is set to screen out high-risk enterprises. For each type of relationship, the edge in which both ends are high-risk enterprises is marked as a high-risk edge. The proportion of high-risk edges in the edges of the relationship type is defined as the structured weight of the relationship type. For the same pair of companies with multiple relationships, a hybrid strategy combining geometric mean and maximum value is adopted to fuse the structured weights of each relationship type and generate a structured joint weight that represents the strength of the company's association with the overall risk.

3. The method according to claim 2, characterized in that, For the same pair of enterprises with multiple relationships, a hybrid strategy combining geometric mean and maximum value is used to fuse the structured weights of each relationship type, including: For the set of relation types that exist between node pairs The structured joint weights are defined as follows: ； in, This is the type with the highest structure weight among all relationship types, emphasizing the most dominant single structural risk path; For enterprise nodes Structured joint weights, For relational types, for Structured weights of class relationships; [0,1] is the fusion coefficient, used to adjust the weight balance between the dominant risk relationship and the cumulative risk under the combined effect of multiple relationships; when When =1, fusion relies entirely on the strongest relationship channel, suitable for focusing on high-risk edge situations; when When =0, fusion depends entirely on the overall level of multiple relationships, emphasizing the consistency risk in the structure; The geometric mean term GeoMean is defined as follows: ； in, Indicates the first Structured weights for class-based relationships. For structured weights of structural relationships of types 1, 2, ..., m, For node pairs The total number of relationship types possessed; a larger GeoMean value indicates that multiple relationships carry greater risks, reflecting the synergistic accumulation effect of structural risk factors.

4. The method according to claim 2, characterized in that, Based on the three-hop association subgraph and multimodal features, the semantic weights of the large language model are obtained through large language model inference, including: The three-hop association subgraph is a structured subgraph formed by taking all connection paths between two core enterprises with a length not exceeding three from the enterprise risk association object graph, and consisting of all nodes covered by these paths and the corresponding edges of the nodes. The core enterprises are identified using eigenvector centrality. Starting with pairing core enterprises, the three-hop relational subgraphs of the core enterprises in the annual graph are retrieved, and the multimodal features of the enterprises to each other are collected. The multimodal features include at least enterprise news, economic indicators, and event descriptions, which serve as important inputs for semantic reasoning in the large language model. The semantic reasoning of the large language model includes at least judging the probability of risk transmission between the two enterprises based on preset expert prompt words, combined with structural paths and multimodal features. The probability is the semantic weight of the large language model and a brief explanation of the mechanism.

5. The method according to claim 4, characterized in that, The quantile mapping between the semantic weights of the large language model and the structured joint weights includes: KDE-CDF distribution alignment is used for same-scale calibration: kernel density estimation is performed on the semantic weights and structured joint weights of all edges of the large language model to fit the probability distribution, and the cumulative distribution function of each is calculated accordingly; each semantic weight of the large language model is mapped to a quantile value through the cumulative distribution function of the semantic weight side of the large language model, and then mapped to the numerical scale of the structured weights through the inverse cumulative distribution function of the structured joint weight side to obtain the calibrated same-scale semantic weights of the large language model.

6. The method according to claim 1, characterized in that, The edge-level gating coefficients are dynamically generated based on the current representations of the nodes at both ends of the edge, and the propagated message is co-modulated using the edge-level gating coefficients and the edge weights, including: For each edge connecting the source node and the target node, the current representations of the source node and the target node are concatenated; the concatenated representation is then transformed by a nonlinear transformation function to generate an edge-level gating coefficient with a numerical range of [-1, 1]; the edge-level gating coefficient is multiplied by the edge weight to obtain the final message modulation factor of the edge; the final message modulation factor is used to scale the message strength transmitted from the source node to the target node, thereby achieving context-dependent and global weight-dependent collaborative modulation.

7. An enterprise risk assessment system based on large language models and high-frequency graph convolution, characterized in that, The enterprise risk assessment system includes: Enterprise Risk Association Graph Construction Unit: Used to construct enterprise risk association graphs. Based on multidimensional relationship data between enterprises, a risk relationship graph of enterprises is constructed with enterprises as nodes; the comprehensive risk score of enterprise nodes in the risk relationship graph is calculated; based on the comprehensive risk score of enterprise nodes, high-risk related enterprises are screened according to a predefined risk score threshold, and structured weights are assigned to different types of enterprise relationships based on the statistical distribution of high-risk related enterprises; for enterprise pairs with multiple relationships, the structured weights of each type of enterprise relationship are integrated to generate a structured joint weight between enterprises and construct an enterprise risk relationship graph. Based on the three-hop association subgraph and multimodal information, the semantic weights of the large language model are obtained through large language model reasoning. The semantic weights of the large language model are mapped to the structured joint weights using quantiles and then fused into the final edge weights. The fused risk graph and multimodal features, including corporate news, economic indicators, and event descriptions, are input into the risk model. Intramodal feature extraction is performed on each multimodal feature, and cross-modal attention fusion is performed between modalities. The economic status of the enterprise, news information, and event semantics are weighted and combined according to their contribution to risk, thereby generating multimodal attention features with individual enterprise preference characteristics. Enterprise risk assessment unit, used for enterprise risk assessment: On the enterprise risk association graph with the edge weights, high-frequency graph convolution is used for information propagation and aggregation. During message transmission, high-frequency graph convolution dynamically generates edge-level gating coefficients based on the current multimodal attention features of the nodes at both ends of the edge, and uses the edge-level gating coefficients and the edge weights to coordinately modulate the propagated message and output the enterprise's risk assessment results.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it is used to implement the enterprise risk assessment method based on a large language model and high-frequency graph convolution as described in any one of claims 1-6.

Images