Multi-agent driven refinancing information analysis method and device, equipment and medium

Abstract

This application discloses a multi-agent-driven method, apparatus, device, and medium for refinancing information parsing, relating to the technical field of interdisciplinary research in fintech and artificial intelligence. Based on a closed-loop iteration of multi-agent-driven refinancing information parsing, it achieves reliable fusion of contradictory multi-source data, accurate inference of event causality, and reliable extraction of financing information. The method includes: calling a data acquisition agent to obtain multi-source heterogeneous financial data verified through evidence causal graph fusion; based on the multi-source heterogeneous financial data, calling a document parsing agent to parse financial documents and extract enterprise information, abstracting the enterprise information into event tuples and arranging them chronologically to form an event chain; based on the context constraints of the event chain, calling an information verification agent to semantically extract refinancing fields from the financial documents; generating reverse verification questions based on the refinancing fields to verify the logical consistency between the refinancing fields; and outputting verified structured refinancing information.

Description

Technical Field

[0001] This application relates to the technical field of the intersection of financial technology and artificial intelligence, and in particular to a multi-agent driven method, apparatus, device and medium for refinancing information analysis. Background Technology

[0002] With the deep integration of artificial intelligence technology and the financial sector, the automatic parsing of financing information has become a key technological direction for improving the data processing efficiency of financial institutions. Corporate financing-related announcements, financial reports, and regulatory documents are scattered across exchange websites and multiple information platforms, encompassing various unstructured data types such as portable documents, word processing documents, and hypertext markup language. Furthermore, the daily update volume is large, requiring analysts to manually search and read each document individually. Processing a single document is time-consuming, making it difficult to achieve timely and comprehensive parsing of all financing information.

[0003] Existing technologies mainly employ optical character recognition, natural language processing, and general multi-agent data analysis to extract text, match keywords, and classify financing-related documents. Based on a preset workflow, these steps are linked together into a linear processing pipeline, outputting structured fields and classification labels for financing information. The compliance review process relies on manual comparison of regulatory rules, manually verifying the extracted fields against regulatory thresholds.

[0004] However, existing technologies in financing information analysis rely on shallow feature matching and linear processing. When multiple data sources provide contradictory information about the same financing event, they only rely on a single credibility index for selection and cannot utilize the supporting and contradictory relationships between data sources for fusion and judgment. Furthermore, for the correlation between adjacent events in a financing event chain, they only rely on surface feature markers such as keyword overlap, failing to distinguish between true causality and pseudo-correlation, leading to downstream analysis and verification being misled by incorrect context. In summary, existing technologies lack the ability to fuse and reason about multi-source contradictory data using graph structures and the ability to infer counterfactual causal relationships between events, making it difficult to achieve accurate and reliable analysis of financing information. Summary of the Invention

[0005] In view of this, this application provides a multi-agent driven method, apparatus, device and medium for refinancing information analysis. The main purpose is to solve the problem that the existing technology lacks the graph structure fusion reasoning ability of multi-source contradictory data and the counterfactual inference ability of event causality, making it difficult to achieve accurate and reliable analysis of financing information.

[0006] The first aspect provides a multi-agent-driven method for parsing refinancing information, which includes: In response to a financing event, a data collection agent is invoked to obtain multi-source heterogeneous financial data obtained through evidence causal graph fusion verification. The evidence causal graph fusion verification includes: constructing an evidence causal graph with data sources as nodes and supporting and contradictory relationships between data sources as edges; deriving the posterior distribution of the credibility of each data source based on the evidence causal graph, and outputting the fusion conclusion with the highest posterior probability as multi-source heterogeneous financial data. Based on the multi-source heterogeneous financial data, a document parsing agent is invoked to parse the financial documents and extract enterprise information. The enterprise information is abstracted into event tuples and arranged in chronological order to form an event chain. The event chain is labeled with causal relationships. The causal relationships are determined by removing preceding events to construct counterfactual assumptions and inferring whether subsequent events will necessarily occur. Based on this, true causality and pseudo-correlation can be distinguished. Based on the context constraints of the event chain, the information verification agent is invoked to perform semantic extraction on the financial document to obtain the refinancing field. Based on the refinancing field, a reverse verification question is generated to verify the logical consistency between the refinancing fields. If a contradiction is found, a second extraction is triggered to form a closed-loop iteration, and the verified structured refinancing information is output.

[0007] Secondly, a multi-agent driven refinancing information analysis device is provided, the device comprising: The acquisition unit is used to respond to a financing event by calling a data acquisition agent to acquire multi-source heterogeneous financial data obtained through evidence causal graph fusion verification. The evidence causal graph fusion verification includes: constructing an evidence causal graph with data sources as nodes and supporting and contradictory relationships between data sources as edges; deriving the posterior distribution of the credibility of each data source based on the evidence causal graph; and outputting the fusion conclusion with the highest posterior probability as multi-source heterogeneous financial data. The abstraction unit is used to call the document parsing agent to parse the financial documents and extract enterprise information based on the multi-source heterogeneous financial data. The enterprise information is abstracted into event tuples and arranged in chronological order to form an event chain. The event chain is labeled with causal relationships. The causal relationships are determined by removing the preceding events to construct counterfactual assumptions and inferring whether subsequent events will necessarily occur. Based on this, true causality and pseudo-correlation are distinguished. The verification unit is used to call the information verification agent to perform semantic extraction on the financial document to obtain the refinancing field based on the context constraints of the event chain, generate a reverse verification question based on the refinancing field to verify the logical consistency between the refinancing fields, and trigger a second extraction to form a closed loop iteration if a contradiction is found, and output the verified structured refinancing information.

[0008] Thirdly, a multi-agent driven refinancing information parsing device is provided, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor. When the processor executes the program, it implements the aforementioned multi-agent driven refinancing information parsing method.

[0009] Fourthly, a storage medium is provided on which a computer program is stored, which, when executed by a processor, implements the aforementioned multi-agent-driven refinancing information parsing method.

[0010] By employing the aforementioned technical solutions, this application provides a multi-agent-driven method, apparatus, device, and medium for refinancing information analysis. Compared to existing technologies that rely on shallow feature matching and linear processing for financing information analysis, this application constructs an evidence causal graph through a data acquisition agent to perform fusion reasoning on multi-source contradictory data. It derives the posterior distribution of credibility by utilizing the supporting and contradictory relationships between data sources. Compared to a single credibility scoring approach, this method can uncover implicit connections between data sources at the graph structure level, outputting the most probable fusion conclusion and improving the credibility of multi-source heterogeneous financial data. Furthermore, by constructing a causal event chain through a document parsing agent to remove preceding events and constructing counterfactual hypotheses to infer whether subsequent events will necessarily occur, this method distinguishes between true causality and pseudo-correlation from a causal perspective. Compared to shallow feature labeling methods such as keyword overlap, this method avoids the misleading influence of pseudo-correlation on downstream analysis and verification, providing accurate causal context support. Finally, by generating a reverse verification problem through an information verification agent to verify the logical consistency between refinancing fields and triggering secondary extraction to form a closed-loop iteration when contradictions are discovered, this method can automatically eliminate extraction illusions and improve the accuracy and reliability of structured refinancing information.

[0011] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0012] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating a multi-agent-driven refinancing information parsing method in one embodiment of this application; Figure 2 yes Figure 1 A flowchart illustrating a specific implementation method of step 101; Figure 3 yes Figure 2A flowchart illustrating a specific implementation method for step 204; Figure 4 yes Figure 1 A flowchart illustrating a specific implementation method for step 102; Figure 5 yes Figure 1 A schematic diagram of a specific implementation method for step 103; Figure 6 This is a schematic flowchart illustrating a specific implementation method of this application after outputting verified structured refinancing information; Figure 7 This is a flowchart illustrating another specific implementation of an embodiment of this application after outputting verified structured refinancing information; Figure 8 This is a schematic diagram of the structure of a multi-agent driven refinancing information parsing system in one embodiment of this application; Figure 9 This is a schematic diagram of the structure of a multi-agent driven refinancing information parsing device in one embodiment of this application; Figure 10 This is a schematic diagram of the device structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0013] The present application will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of the present application can be combined with each other.

[0014] Existing technologies for analyzing financing information rely on shallow feature matching and linear processing. When multiple data sources provide contradictory information about the same financing event, they only rely on a single credibility index for selection and cannot utilize the supporting and contradictory relationships between data sources for fusion and judgment. Furthermore, for the correlation between adjacent events in a financing event chain, they only rely on surface-level feature markers such as keyword overlap, failing to distinguish between true causality and pseudo-correlation, leading to downstream analysis and verification being misled by incorrect context. In summary, existing technologies lack the ability to fuse and reason about multi-source contradictory data using graph structures and the ability to infer counterfactual causal relationships between events, making it difficult to achieve accurate and reliable analysis of financing information.

[0015] To address this problem, this embodiment provides a multi-agent-driven method for parsing refinancing information, such as... Figure 1 As shown, it includes the following steps: 101. In response to a financing event, the data collection agent is invoked to obtain multi-source heterogeneous financial data verified by evidence causal graph fusion.

[0016] In this embodiment, the trigger signal for the financing event comes from the exchange's announcement system. When a financing event is detected, the data collection agent is invoked to collect relevant data from multiple financial information platforms. Different platforms may provide inconsistent information for the same field of the financing event. For example, platform A shows that the issuance amount is 1 billion yuan, while platform B shows that the issuance amount is 1.2 billion yuan. At this time, the evidence causal graph fusion verification is triggered.

[0017] The specific evidence causal graph fusion verification includes: constructing an evidence causal graph with data sources as nodes and supporting and contradictory relationships between data sources as edges; deriving the posterior distribution of the credibility of each data source based on the evidence causal graph, and outputting the fusion conclusion with the highest posterior probability as multi-source heterogeneous financial data.

[0018] In the causal evidence graph, each node represents an independent data provider, such as a stock exchange announcement platform, a corporate credit information disclosure system, or a third-party financial information terminal. Edges between nodes are constructed based on the descriptive relationships of the various data sources regarding the same financing event. If two data sources provide consistent and mutually corroborating descriptions of key fields such as financing amount and counterparty, a supporting edge is established; if field values ​​conflict or contradict each other, a contradictory edge is established. This causal evidence graph comprehensively depicts the mutual verification structure among multi-source data, providing a topological foundation for subsequent credibility reasoning.

[0019] After constructing the evidence causal graph, the prior probability of each node is initialized based on a pre-defined credibility assessment model. This prior probability comprehensively considers factors such as the historical accuracy, update frequency, and authority level of the data source. Then, an iterative message passing algorithm is executed along the edges of the evidence causal graph, allowing each node to dynamically adjust its credibility estimate based on supporting or contradictory feedback from neighboring nodes. After multiple iterations until the posterior distribution converges, the field value corresponding to the data source with the highest posterior probability is selected as the final fusion conclusion. For example, when three data sources record the amount of a certain equity financing as A1, A2, and A3, and the first two have a strong supporting relationship while the third has a contradictory relationship, the output with the higher posterior probability between A1 and A2 will be used as the accurate amount after fusion, thus effectively eliminating noise interference from a single data source.

[0020] It should be noted that the multi-source heterogeneous financial data obtained through fusion verification not only includes structured fields but also retains the contextual information of the original documents. The fusion process records the posterior probability changes of each node and the basis for determining the relationships between edges throughout the entire process, forming a traceable verification log to ensure the audit compliance of the fusion conclusions. By embedding the data conflict resolution mechanism into the causal reasoning framework, this embodiment achieves a shift from passively receiving data to actively verifying data, improving the accuracy of financing information analysis and providing a reliable data foundation for downstream clue identification.

[0021] 102. Based on the multi-source heterogeneous financial data, call the document parsing intelligent agent to parse the financial documents and extract enterprise information, and abstract the enterprise information into event tuples and arrange them in chronological order to form an event chain.

[0022] In this embodiment, financial documents are various unstructured or semi-structured text materials directly related to financing events. These documents typically exist in portable document formats, Hypertext Markup Language web pages, or scanned copies, but due to their diverse formats and complex layouts, they are difficult to parse directly by traditional rule engines. Therefore, it is necessary to rely on multi-source heterogeneous financial data as external anchors and invoke a document parsing intelligent agent to accurately parse and extract information from the financial documents.

[0023] In practice, the document parsing agent first uses a multimodal document layout analysis model to structurally model financial documents, constructing a multi-layered tree structure that includes heading levels, paragraph hierarchies, table rows and columns, and merged cell relationships, while filtering out redundant content such as headers, footers, watermarks, and duplicate declarations. Then, based on a pre-trained entity recognition model, named entities are extracted from this tree structure. The extracted results are then matched with a standard entity library using vector similarity to generate a candidate entity list. Contextual semantic rules, such as equity penetration relationships, consistency of registered location, and industry classification matching, are then used to verify the candidate entities, ultimately determining the target entities and their corresponding company information. On this basis, the relevant information of each target entity is abstracted into event tuples containing event subject, event type, event time, and event attributes, and arranged in ascending order of event time to form an event chain. Throughout the parsing process, the fusion conclusions provided by multi-source heterogeneous financial data are used as a reference to ensure that the extracted company information is consistent with verified facts.

[0024] After forming an event chain, causal relationships are labeled for adjacent or cross-time period event tuples. Correspondingly, the event chain is labeled with causal relationships. These causal relationships are determined by removing preceding events to construct counterfactual hypotheses and inferring whether subsequent events will necessarily occur, thus distinguishing between true causality and spurious correlations. Specifically, for any pair of successive events, the preceding event is removed, and a counterfactual hypothesis is constructed to infer whether the subsequent event will still necessarily occur in the absence of the preceding event. For example, when it is detected that a company completes a new round of equity financing within three months of receiving government industry subsidies, the subsidy event is removed, and a scenario where the company did not receive subsidies is simulated. If the inference shows that the financing event is still highly likely to occur, then the two are determined to be only related by time sequence rather than true causality; conversely, if the probability of the financing event significantly decreases, then the subsidy event is confirmed as the true cause of the financing event. This counterfactual inference process effectively filters out spurious correlations caused by confounding factors such as industry cycles and market sentiment, ensuring that the causal relationships in the event chain have substantial explanatory power.

[0025] The event chain labeled with causal relationships not only retains the original temporal information, but also embeds a verified causal logic structure. By deeply integrating the counterfactual reasoning mechanism into the event chain construction process, this embodiment achieves a leap from simple temporal arrangement to causal semantic understanding, thereby improving the business interpretability of financing information analysis.

[0026] 103. Based on the context constraints of the event chain, the information verification agent is invoked to perform semantic extraction on the financial document to obtain the refinancing field. Based on the refinancing field, a reverse verification question is generated to verify the logical consistency between the refinancing fields. If a contradiction is found, a second extraction is triggered to form a closed loop iteration, and the verified structured refinancing information is output.

[0027] In this embodiment, the event chain annotated with causal relationships accurately depicts the identities of the entities involved in each financing event, the sequence of occurrence, and the true driving logic, providing an irreplaceable anchoring framework for semantic extraction. This framework enables the system to distinguish which text fragments belong to the substantive content of the financing event to be parsed, and which are merely background information or distracting information, thereby ensuring the business relevance and factual accuracy of the obtained fields from the source.

[0028] In the specific implementation process, the information verification agent first determines the list of fields to be extracted and their semantic roles based on the attribute definitions of the current node in the event chain. Then, using the preceding and following causal events of this node in the event chain as query conditions, it retrieves paragraphs or table areas in the financial document that semantically match these conditions. Subsequently, a semantic understanding model for the financial domain is used to extract fine-grained information from the candidate areas, and the extraction results are aligned and verified with the confirmed entities and time-series information in the event chain. Only field values ​​consistent with the context constraints are retained to form the refinancing field.

[0029] Furthermore, since semantic extraction is a single output, large language models may produce illusions, leading to discrepancies between the extracted refinancing fields and the original document, or logical contradictions between the fields. For example, the lock-up period field value may not match the type of the issuing object. A single extraction cannot automatically detect and correct such contradictions. Therefore, it is necessary to generate a reverse verification problem to verify the logical consistency between the refinancing fields.

[0030] Among them, the reverse verification problem is a verification problem automatically generated based on the extracted refinancing fields, used to verify whether the logical relationship between the fields is valid. For example, based on the two refinancing fields that the financing type is private placement and the issuance target includes the controlling shareholder, a verification problem is generated: whether the lock-up period of the controlling shareholder's subscription to the private placement complies with the regulatory rules. Specifically, in the process of verifying the logical consistency between refinancing fields, a semantic understanding model is deployed in the information verification agent. The generated reverse verification questions, along with the original refinancing fields and corresponding document fragments, are input into the semantic understanding model, which then determines whether the pre-defined logical relationship of the questions holds true. If the semantic understanding model returns a negative or uncertain answer, the relevant fields are determined to be contradictory, and the fields are immediately marked as needing correction. The secondary extraction strategy is adjusted according to the type of contradiction: for numerical contradictions, all expressions involving that value in the document are relocated and the correct value is comprehensively determined; for semantic contradictions, the search scope is expanded to related chapters or supplementary cross-evidence from multi-source heterogeneous data is consulted. The corrected fields undergo another round of reverse verification until all questions receive positive answers. The structured refinancing information that passes this closed-loop iterative verification not only contains accurate field values ​​but also includes complete verification paths and causal traceability records. It can be directly used in scenarios with high credibility requirements such as clue identification, risk warning, and compliance auditing, truly achieving a reliable transformation from original documents to business-usable knowledge.

[0031] The structured refinancing information that passes closed-loop iterative verification not only contains accurate field values ​​but also includes a complete verification trajectory and causal path of the event chain, ensuring that every piece of output data can be traced back to the specific location in the original document and verified through logical consistency. By deeply embedding the causal semantics of the event chain into the entire extraction and verification process, this embodiment realizes the transformation from isolated field extraction to context-aware information parsing, enabling structured refinancing information to possess both data-level accuracy and business-level interpretability and reliability, providing a high-quality and highly reliable data foundation for financial intelligent analysis.

[0032] The multi-agent-driven refinancing information parsing method provided in this application, compared with the existing technology that relies on shallow feature matching and linear processing for financing information parsing, constructs an evidence causal graph through a data acquisition agent to perform fusion reasoning on multi-source contradictory data. It derives the posterior distribution of credibility by utilizing the supporting and contradictory relationships between data sources. Compared with the single credibility scoring method, it can mine the implicit correlations between data sources at the graph structure level, output the fusion conclusion with the optimal probability, and improve the credibility of multi-source heterogeneous financial data. The document parsing agent constructs an event chain of causal relationships to remove antecedent events and constructs counterfactual hypotheses to infer whether subsequent events will necessarily occur. It distinguishes between true causality and pseudo-correlation from the perspective of causality. Compared with shallow feature labeling methods such as keyword overlap, it can avoid the misleading effect of pseudo-correlation on downstream parsing and verification, and provide accurate causal context support. The information verification agent generates a reverse verification problem to verify the logical consistency between refinancing fields and triggers secondary extraction to form a closed-loop iteration when contradictions are found. It can automatically eliminate extraction illusions and improve the accuracy and reliability of structured refinancing information.

[0033] In practical applications, financing-related financial data is typically sourced from diverse locations, updated at varying paces, and is highly time-sensitive. Relying solely on periodic full-scale data collection is not only resource-intensive but also makes it difficult to capture sudden announcements or regulatory updates during critical windows. Conversely, relying entirely on passive monitoring may miss underlying data changes that, while not triggering explicit events, actually impact financing decisions. Therefore, a dual-mode data collection mechanism with both proactive sensing and comprehensive coverage capabilities must be constructed. This mechanism should be able to respond quickly when financing events occur while ensuring the continuity and integrity of basic data through routine inspections. Specifically, such as... Figure 2 As shown, step 101 above includes the following steps: 201. The data acquisition agent adopts a dual-mode architecture that combines event-driven directional acquisition with direct interface connection for full-volume inspection to access multi-source data.

[0034] 202. Based on a preset event system, listen for trigger signals of financing events, dynamically adjust the data collection frequency according to the importance weight of financing events, and execute financial data collection tasks through distributed deployment of various types of collectors.

[0035] 203. Verify the credibility of the collected financial data based on at least one of the following dimensions: source authority, data timeliness, and content completeness.

[0036] 204. Perform evidence causal graph fusion verification on the financial data that has passed the credibility verification to obtain multi-source heterogeneous financial data.

[0037] In terms of implementation, the system monitors the trigger signals of financing events in real time based on a preset event system. This financing event system covers company lifecycle events such as the date of the board of directors' preliminary announcement, the date of the shareholders' meeting review, the date of the financial report disclosure, the date of the response to regulatory inquiries, and the date of the previous financing lock-up period lifting; regulatory policy events such as the revision of refinancing rules, the update of industry regulatory policies and the dynamic update of exchange review; and market events such as abnormal fluctuations in industry indices, financing announcements of leading companies, and the implementation of major projects in the industrial chain.

[0038] Specifically, the data acquisition agent is equipped with a dynamic acquisition scheduling algorithm, which sets a basic acquisition frequency when no financing events are triggered. Once a trigger signal for any financing event is detected, the data collection frequency will be dynamically adjusted. See the following formula:

[0039] in, To pre-allocate weights for different financing events, important financing events such as board proposal announcements are given the highest weight, and the collection frequency immediately increases to the minute level and continues for several trading days after being triggered, before gradually decreasing back to the basic frequency; less important financing events are given a lower weight and are increased to the hour level. This serves as an upper limit for the collection frequency, acting as a protection threshold to prevent an unlimited surge in collection frequency due to an excessive number of financing events and the accumulation of weights.

[0040] The data collection task is executed collaboratively by a variety of distributed collectors, including direct-connection collectors, webpage rendering collectors, and document parsing collectors. Each collector supports breakpoint resumption and failure retry mechanisms to ensure a high success rate of data capture within a short period after the announcement is released, thereby efficiently obtaining financial data related to the financing event.

[0041] Because the data quality varies across different data sources, directly using unverified data may lead to unreliable downstream parsing results. Therefore, it is necessary to verify the credibility of the collected financial data. The credibility verification process calculates a credibility score based on three dimensions: source authority, data timeliness, and content completeness. The values ​​for each dimension range from 0 to 1. See the following formula:

[0042] Among them, the weight of source authority Evaluation based on data source; timeliness weighting Judging based on publication time; completeness weighting is based on... Field completeness is linearly reduced. Financial data with a credibility score greater than a preset threshold passes verification; financial data that fails verification is moved to the review queue.

[0043] It should be noted that credibility verification can only exclude low-quality data. However, when multiple data sources provide contradictory information about the same financing event, a single credibility score cannot derive the optimal conclusion by utilizing the correlation between data sources. Therefore, it is necessary to perform evidence causal graph fusion verification on the financial data that has passed credibility verification.

[0044] Specifically, such as Figure 3 As shown, step 204 above includes the following steps: 301. When different data sources provide contradictory information about the same financing event, the credibility scores of each data source are used as prior probabilities to initialize the credibility of each node in the causal graph of evidence. The support and contradiction relationships between data sources are determined by comparing the field values ​​of different data sources for the same financing event.

[0045] 302. Construct the edges of the evidence causal graph based on the supporting and contradictory relationships, and perform iterative message passing along the edges to update the posterior distribution of the credibility of each node until convergence. Select the field value corresponding to the data source with the highest posterior probability as the fusion conclusion to obtain multi-source heterogeneous financial data.

[0046] In this embodiment, when different data sources provide contradictory information about the same financing event, relying solely on a single credibility score for selection ignores potential mutual support or contradictory relationships between the data sources. For example, if two low-credibility data sources have the same field value for the same financing event, the credibility of that field value should be enhanced. However, a single score cannot reflect this correlation effect. Therefore, it is necessary to construct an evidence causal graph, using the credibility scores of each data source as prior probabilities to initialize the credibility of each node in the evidence causal graph.

[0047] Specifically, three different data sources are used as nodes in the causal graph of evidence. The credibility score of the first data source is 0.95, the second data source is 0.85, and the third data source is 0.7, which are used as the prior probabilities of the credibility of each node. Then, the supporting and contradictory relationships between data sources are determined by comparing the field values ​​of the same financing event from different data sources. For example, for the issuance amount field, the first data source shows 1 billion yuan, and the second data source also shows 1 billion yuan, so their field values ​​are consistent and constitute a supporting relationship. The second data source shows 1.2 billion yuan, which is inconsistent with the first two and constitutes a contradictory relationship. For the financing type field, all three data sources show private placement, and there are supporting relationships between each pair of them.

[0048] Furthermore, based on the established support and contradiction relationships, an evidence causal graph is constructed with edges. Support relationship edges connect data source nodes with consistent field values, while contradiction relationship edges connect data source nodes with inconsistent field values. Iterative message passing occurs along the edges: support relationship edges transmit strengthening signals, and receiving nodes update their own credibility based on the credibility of the sending node and the strength of support; contradiction relationship edges transmit weakening signals, and receiving nodes reduce their own credibility based on the credibility of the sending node and the strength of contradiction. After multiple iterations until the posterior distribution of the credibility of each node converges, the first data source, due to the support enhancement from the second data source and the lack of contradiction weakening, has the highest posterior probability. The third data source, due to contradiction with most data sources and a low prior probability, has the lowest posterior probability. The field value corresponding to the first data source with the highest posterior probability is selected as the fusion conclusion, with the issuance amount set at one billion yuan and the financing type at private placement, resulting in multi-source heterogeneous financial data.

[0049] In practical applications, financial documents are typically in portable or word processing document formats. Directly processing these documents makes it difficult to accurately locate structured data within tables. Therefore, a multimodal document layout analysis model is needed to construct a multi-layered tree structure for financial documents. Specifically, for example... Figure 4 As shown, step 102 above includes the following steps: 401. Call the document parsing intelligent agent and use a multimodal document layout analysis model to construct a multi-layered tree structure for financial documents.

[0050] 402. Based on the pre-trained entity recognition model, named entities are extracted from the multi-layer tree structure. The extracted named entities are matched with the standard entity library by vector similarity to obtain candidate entities. The candidate entities are then checked for rules based on context information to determine the target entity.

[0051] 403. Abstract the enterprise information corresponding to the target entity into event tuples containing event subject, event type, event time and event attributes, and arrange them in ascending order of time to form an event chain.

[0052] In this embodiment, a multimodal document layout analysis model is deployed in the document parsing agent. The multimodal document layout analysis model is a pre-trained document understanding model that can simultaneously process three modalities of input: document page image, optical character recognition text, and text coordinate information. It obtains the layout information by extracting visual features from the page image, and locates the spatial position of each content area by combining the text coordinate information. The output includes layout labels covering cover title, table of contents, chapter titles, body paragraphs, list items, tables, headers and footers, footnotes, disclaimers, and watermark redundancy.

[0053] Based on this, the construction of a multi-level tree structure includes at least the following: constructing the parent-child relationship of chapters based on the font size, indentation, and numbering rules of the heading level, for example, the first-level heading is the parent node and its subordinate second-level headings are child nodes; using a table structure recognition algorithm to restore the row and column relationships of cells in the table area and outputting a structured two-dimensional table; and filtering redundant content, including headers, footers, and disclaimers.

[0054] Correspondingly, the output is a tree-structured data containing node identifier, parent node identifier, type, content, and confidence level.

[0055] Considering that the same entity may be expressed differently in different documents within a financial document, a pre-trained entity recognition model is deployed in the document parsing agent. This pre-trained model is then used to extract named entities from a multi-layered tree structure and perform cross-document alignment. Specifically, an entity recognition model based on a pre-trained language model is used to identify various entities in the text, such as company names, personal names, stock codes, amounts, and dates. The extracted named entities are matched with a standard entity library using vector similarity to obtain candidate entities. After encoding the identified entities and standard library entities into vectors, the cosine similarity of these vectors is calculated. Based on the cosine similarity, the top-ranked candidate entities are selected. Combined with contextual stock codes, industry classifications, and related entities, rule verification is performed to determine the target entity from the candidate entities. A unique matching result is output, completing entity normalization.

[0056] After entity normalization, the enterprise information corresponding to the target entity is abstracted into tuples containing event subject, event type, event time, and event attributes, and arranged in ascending chronological order to form a vertical event chain. This abstraction process is not simply a listing of facts, but rather calculates the correlation strength between adjacent events based on dimensions such as event type, involved subjects, and keyword overlap, and then labels them as causal, progressive, or unrelated relationships. The resulting event chain not only faithfully records the timeline of the enterprise's financing activities but also embeds a logical dependency structure between events.

[0057] In practical applications, financing information in financial documents is often fragmented, implicit, and prone to multi-source conflicts. Simply relying on model extraction or rule matching is insufficient to guarantee the consistency of business logic and factual accuracy between fields. Therefore, a dual verification mechanism for financing information based on event chains is needed to enable the information to support highly reliable clue identification and compliance judgment. Specifically, such as... Figure 5 As shown, step 103 above includes the following steps: 501. Using the historical financing event time sequence distribution and causal relationship annotation of the target enterprise in the event chain as context constraints, the information verification agent is invoked to perform rule engine screening to locate candidate paragraphs in the financial document, and intent disambiguation is performed on the candidate paragraphs to extract the refinancing field from the paragraphs after intent disambiguation.

[0058] 502. Decompose the regulatory rules into structured rules that include at least applicable scenarios, constraints, quantification thresholds, and special cases. Use the causal relationship between the refinancing field and the preceding events in the event chain as a joint retrieval condition to match the target rules. Use the result of quantification threshold comparison, scenario judgment, and exception exclusion of the target rules, together with the logical consistency verification result between the refinancing field, as the basis for judging the reverse verification problem.

[0059] 503. When there is a contradiction in the determination criteria, a correction instruction is generated based on the document fragment associated with the preceding event in the event chain that has a causal relationship with the contradiction point, triggering a second extraction of the document fragment until the verification is passed or the preset number of iterations is reached.

[0060] 504. Calculate the historical risk score based on the event chain superimposed with negative risk factors. When the historical risk score is greater than a preset threshold, mark the historical risk label. Combine the verified refinancing field with the historical risk label as verified structured refinancing information.

[0061] In its implementation, the information verification agent deploys a rule engine. The rule engine initially screens candidate paragraphs based on a keyword dictionary for refinancing built from an automaton, and uses multiple regular expressions to extract directly structured fields such as issuance amount and number of subscribers. Two levels of recall thresholds are set to ensure a high initial recall rate, preventing the omission of crucial information. Subsequently, intent disambiguation is performed on the candidate paragraphs. Based on a general large language model, a large-scale financial refinancing model is constructed using low-rank adaptation and instruction fine-tuning. The training set includes manually annotated historical refinancing announcements and semi-structured financial statements. Fine-tuning tasks include three categories: refinancing field extraction, implicit intent recognition, and clause semantic understanding. Refinancing field extraction uses structured output instructions to force the model to output formatted data, covering the target company, refinancing type, issuance amount, number of subscribers, pricing benchmark date, lock-up period, conversion price, and use of proceeds. To identify the implicit intent of ambiguous statements, a contrastive learning framework is used to train an intent classification model. Positive samples are paragraphs that explicitly disclose refinancing plans, while negative samples are paragraphs that mention funding needs but do not have refinancing plans. The loss function uses information noise contrast estimation to estimate the loss. The model output is the probability of refinancing intent. If the probability is greater than a set threshold, it is judged as a potential target. The refinancing field is extracted from the paragraphs after intent disambiguation.

[0062] The specific loss function used is the InfoNCE loss, as shown in the following formula:

[0063] in, Cosine similarity; The temperature coefficient is used; the model output is the probability of refinancing intention. If the probability is greater than a set threshold, it is judged as a potential clue, thus realizing preliminary discovery.

[0064] It should be noted that the extraction of the refinancing field is a single output. The large language model may produce illusions, resulting in the extracted field values ​​not matching the financial documents or logical contradictions between fields. This leads to different regulations imposing different constraints on the same situation. Relying solely on logical consistency checks between fields cannot cover the detection of contradictions in the compliance dimension. Therefore, regulatory rules need to be broken down into structured rules and used together with the logical consistency check results as the basis for judging reverse verification problems.

[0065] Specifically, regulatory rules are broken down into structured rules that include at least applicable scenarios, constraints, quantitative thresholds, and special cases. A structured rule index and a semantic rule vector library are constructed simultaneously. Compliance verification employs a three-tiered progressive process: First, a structured coarse-grained recall of rules is conducted based on structured fields such as refinancing type and issuing entity segment to quickly filter the rule set; second, for non-quantitative principle clauses such as reasonable financing scale, semantically relevant rules are matched using vector retrieval to complete the structured rules for easily overlooked scenarios; finally, quantitative threshold comparison, scenario judgment, and exception exclusion are performed on each recalled rule; the causal relationship between the refinancing field and preceding events in the event chain is used as a joint retrieval condition to obtain the target rule. The results of the target rule's quantitative threshold comparison, scenario judgment, and exception exclusion, along with the logical consistency verification results between the refinancing field, serve as the basis for judging reverse verification issues.

[0066] When there are contradictions in the judgment criteria, for example, if the verification finds that the issuer includes a controlling shareholder but the lock-up period is only 18 months, while the regulatory rules require that the controlling shareholder's subscription lock-up period be no less than 36 months, it is necessary to locate the source of the contradiction based on the causal context of the event chain and trigger a correction. Specifically, a correction instruction is generated based on the document fragment associated with the preceding event in the event chain that has a causal relationship with the contradiction point, triggering a second extraction of the document fragment, re-identifying the composition information of the issuer, correcting the lock-up period to 36 months, and generating a reverse verification question for verification again, until the verification passes or the preset number of iterations is reached, forming a closed-loop iteration of extraction and verification.

[0067] Since the target company's historical negative events may affect the feasibility and risk level of the current financing plan, relying solely on the extraction and verification results of the current document is insufficient to assess the cumulative effect of historical risks. Therefore, it is necessary to calculate the historical risk score based on the event chain and by overlaying negative risk factors. Specifically, the historical risk score is calculated by calling the target company's event chain and overlaying negative risk factors, as shown in the following formula:

[0068] in, Branches are set for the score of individual negative events, including negative events such as regulatory inquiries, administrative penalties, and pre-fundraising violations; The time decay factor is used. Each negative event score is assigned a branch weight based on its type and multiplied by the time decay factor to reflect the timeliness of the risk, resulting in a historical risk score. Finally, targets with historical risk scores exceeding a preset threshold are automatically marked with a historical risk tag and placed in a priority review queue. The verified refinancing field is combined with this risk tag to output verified structured refinancing information, containing both accurate and complete transaction elements and a traceable risk profile, fully supporting subsequent intelligent decision-making and manual review.

[0069] In practical applications, verified structured refinancing information only contains attribute fields of the financing event itself. Directly using it for lead ranking lacks horizontal comparability. Companies in different industries and with different market capitalizations exhibit significant differences in financial indicators and industry prosperity. Fixed-weight scoring cannot accommodate these differences. Therefore, it is necessary to construct a multi-dimensional quantitative feature system using verified structured refinancing information as input, and achieve fair horizontal comparison based on dynamic benchmarks and conditional variable weighting mechanisms. Furthermore, such as... Figure 6 As shown, after outputting the verified structured refinancing information, the method further includes: invoking the clue evaluation agent to perform the following steps: 601. Construct a multi-dimensional quantitative feature system using the verified structured refinancing information as input. Divide the indicators into groups according to industry classification and market capitalization size, calculate the dynamic benchmark quantile value of each indicator, dynamically calculate the scoring weight based on the grouping status of the multi-dimensional quantitative features, and obtain the clue score by weighted summation of the multi-dimensional quantitative features.

[0070] 602. Based on the knowledge graph of the industrial chain, a graph attention network is used to transmit clue scores.

[0071] 603. Using the business end feedback results as labels, the scoring weights are iteratively updated using an online gradient descent algorithm. After each update, single-dimensional weight fluctuation truncation and normalization processing are performed to ensure that the total weight is constant and the single-dimensional fluctuation does not exceed the preset range.

[0072] In the specific implementation process, verified structured refinancing information is used as input to construct a multi-dimensional quantitative feature system covering four dimensions: policy adaptation, enterprise qualifications, industry prosperity, and business value. All raw features are first standardized to eliminate the influence of dimensions, then outliers are truncated using standard deviation and linearly scaled to a preset range. Subsequently, the system is grouped according to secondary industry classification and market capitalization. Within each group, the quantile values ​​of each indicator are calculated as a dynamic benchmark. This benchmark is automatically updated after each quarterly financial report disclosure and incorporates a macroeconomic cycle adjustment factor to reflect industry growth deviations and changes in prosperity. The macroeconomic cycle adjustment factor is shown in the following formula:

[0073] in, This indicates a deviation in growth rate; This addresses the deviation in industry prosperity growth. Individual company metrics are mapped to their percentile position within their respective groups, with full marks awarded to those above the 75th percentile and low marks to those below the 25th percentile, thus ensuring fair horizontal comparison.

[0074] Based on this, the scoring weights are dynamically adjusted according to the grouping status of the multidimensional quantitative features. For example, the weights of policy adaptation and industry prosperity dimensions are appropriately increased for enterprises in the group with strong policy support and high prosperity. Finally, the clue scores are obtained by weighted summation of the features of each dimension, ensuring that the scoring results are both industry-comparable and reflect individual characteristics.

[0075] The specific clue value score uses a weighted summation model, with initial weights determined by expert scoring. The output ranges from 0 to 100, with higher scores indicating greater value. The clue value is calculated using the following formula:

[0076] in, To adapt to policy dimensions and characteristics; Characteristics of enterprise qualifications; Characteristics of industry prosperity; Features representing the business value dimension.

[0077] It should be noted that relying solely on clue scores for ranking cannot distinguish between high-value, high-risk clues and high-value, low-risk clues. Therefore, a risk coefficient needs to be calculated and combined with the clue score to form a two-dimensional assessment. The risk coefficient integrates three categories: compliance risk, issuance failure risk, and market volatility risk. Compliance risk is derived from the risk level output by compliance verification; issuance failure risk is calculated based on the pass rate and withdrawal rate of similar financing plans in historical issuance data; and market volatility risk is calculated based on industry index volatility and market liquidity indicators. The three types of risk are integrated using a multiplicative model. A high value in any risk dimension will significantly amplify the overall risk coefficient; the higher the value, the greater the risk. For the specific multiplicative model, please refer to the following formula:

[0078] The higher the risk coefficient, the greater the risk. A four-quadrant classification is used, with a lead score of 70 and a risk coefficient of 0.4 as thresholds. Leads with a score greater than 70 and a risk coefficient less than 0.4 are classified as high-value, low-risk, and should be prioritized for immediate follow-up. Leads with a score greater than 70 and a risk coefficient greater than 0.4 are classified as high-value, high-risk, requiring a thorough assessment of risk controllability before making a decision. Leads with a score less than 70 and a risk coefficient less than 0.4 are classified as low-value, low-risk, and can be tracked routinely. Leads with a score less than 70 and a risk coefficient greater than 0.4 are classified as low-value, high-risk, and should be temporarily suspended or abandoned. This dual-dimensional assessment enables differentiated ranking and hierarchical management of leads.

[0079] In this embodiment, the industry chain knowledge graph includes multiple heterogeneous nodes and heterogeneous edges carrying different semantics. Upstream and downstream nodes whose transmission scores reach a preset association threshold are marked as potential association clues, and semantic explanations representing the transmission path are output simultaneously. When a node triggers a refinancing event and obtains an initial clue score, it propagates upstream and downstream along the industry chain edges through a graph attention network. During the transmission process, the attention weight is jointly determined by the business relevance and the proportion of historical transactions, and adaptively aggregates neighbor node information in conjunction with a feature transformation matrix. Only when the transmission score reaches the preset association threshold are the corresponding upstream and downstream nodes marked as potential association clues, and semantic explanations representing the transmission path are output simultaneously, enabling analysts to understand the logical basis of the clue association.

[0080] For the specific conduction process, please refer to the following formula:

[0081] in, For conduction score; The attention weight is determined by both business relevance and transaction proportion. The characteristic transformation matrix; For nodes The neighboring nodes. If the transmission score is ≥0.6, its upstream and downstream nodes are marked as potential correlation clues and automatically included in the tracking range.

[0082] Furthermore, an online gradient descent algorithm is used to dynamically optimize the weights, eliminating the need for full retraining. The weights are updated quarterly using the feedback results from the business side as labels. For the specific dynamic weight optimization process, please refer to the following formula:

[0083] Among them, the loss function The mean squared error loss and the learning rate are... The value is set to 0.01, and regularization is added to prevent overfitting. After each gradient update, single-dimensional weight fluctuation truncation and normalization are performed to ensure that the total weight sum is constant and the single-dimensional fluctuation amplitude is small, thus ensuring model stability.

[0084] In practical applications, if verified structured refinancing information is not tracked and managed after output, the entire process from lead identification to implementation lacks status control. Business personnel cannot be informed of lead progress in a timely manner, and leads that are not processed within the time limit may miss the project window. Therefore, it is necessary to convert verified structured refinancing information into lead data and drive its migration between different tracking states based on preset status flow rules. Specifically, such as... Figure 7 As shown, after outputting the verified structured refinancing information, the method further includes: invoking the information verification agent to perform the following steps: 701. Convert the verified structured refinancing information into clue data, drive the clue data to migrate between different tracking states based on preset state transition rules, set timeout supervision trigger conditions for each tracking state, and record the time and operation identifier of each tracking state migration to form an audit log.

[0085] 702. Based on user profiles, a collaborative filtering algorithm and a rule matching model are used to calculate the push priority, and the clue data is pushed to the target users in a differentiated manner according to the push priority.

[0086] 703. Receive landing result data returned from the business end, update the tracking status of the corresponding clue data based on the landing result data, and use the landing result data as a feedback tag for subsequent scoring weight iteration.

[0087] In this embodiment, different tracking states include, but are not limited to, pending identification, identified, pushed, following up, due diligence, landed / lost, and archived. Each tracking state is set with a timeout supervision trigger condition.

[0088] For example, if a status update has been pushed but not followed up within three working days, a supervisory reminder will be triggered. The time and operation identifier of each status update will be recorded to form an audit log, supporting full-process audit traceability.

[0089] Because personnel across different business lines have varying interests in lead types and industry preferences, pushing the same content to all personnel would be inefficient and lack targeting. Therefore, it is necessary to use collaborative filtering algorithms and rule-matching models based on user profiles to calculate push priorities and implement differentiated pushes. Specifically, user profiles are constructed based on business line, industry, region, and historical preferences. Collaborative filtering algorithms are used to predict target users' preference scores for leads based on the historical behavior of similar users. Rule-matching models are used to perform hard matching and filtering based on the user's business line and the industry attributes of the leads. The push priority is calculated by combining both methods, and lead data is pushed to target users in a differentiated manner according to the push priority, supporting three push channels: terminal to-do list, instant messaging, and email.

[0090] If the follow-up and implementation results from the business side after lead generation are not fed back to the system, the scoring model cannot be optimized based on actual results, and the system's capabilities will gradually become rigid. Therefore, it is necessary to receive implementation result data fed back from the business side to form a closed loop. Specifically, standardized interfaces are provided, including a forward push interface and a reverse implementation result feedback interface. This interface receives implementation result data fed back from the business side and updates the tracking status of the corresponding lead data based on the implementation result data. For example, if the business side reports that due diligence is complete, the tracking status is updated to "implemented"; if the business side reports that the customer has abandoned the business, the tracking status is updated to "churned". The implementation result data is used as a feedback label for subsequent scoring weight iterations, enabling the scoring model to be continuously optimized based on actual implementation results.

[0091] Furthermore, as a specific implementation of the above method, embodiments of this application provide a multi-agent driven refinancing information parsing system, such as... Figure 8 As shown, this system demonstrates the overall topology and collaborative relationships between agents at each level of the multi-agent system for analyzing refinancing clues of listed companies based on artificial intelligence. The diagram clearly presents a star-shaped three-tier architecture consisting of a central control scheduling layer, an execution layer, and an iterative optimization layer. The central control scheduling agent is located at the center, connecting six domain-specific execution agents through a directed acyclic graph dynamically orchestrating instructions, while simultaneously forming a bidirectional feedback path with the iterative optimization agent. The central control layer embeds event-driven logic from the financial domain, enabling it to generate differentiated collaborative links in real time based on different event types such as board proposals and regulatory inquiries, rather than relying on fixed preset workflows.

[0092] In terms of specific components, the central control and scheduling intelligent agent, as the core of global scheduling, integrates a capital market event calendar engine and a regulatory rule dynamic perception module. Its responsibility is not simply task distribution, but rather to dynamically generate and adjust the directed acyclic graph scheduling sequence of the executing intelligent agents based on the urgency of financial events and real-time changes in regulatory rules. The execution layer consists of six types of dedicated functional agents: The data acquisition agent includes an event-triggered acquisition unit, a multi-source alignment unit, and a hierarchical labeling unit, responsible for the targeted acquisition and standardized access of multi-source financial data; the document preprocessing agent includes a structure tree parsing unit, an entity disambiguation unit, and an event chain construction unit, completing the deep structured preprocessing of financial documents; the information parsing agent includes a rule initial screening unit, a large model extraction unit, and an intent recognition unit, achieving accurate extraction of core information and implicit intent in refinancing; the compliance verification agent includes a rule semantic encoding unit, a clause comparison unit, and a risk tracing unit, independently performing clause-level compliance verification and risk labeling, and supporting hot rule updates; the business opportunity assessment agent includes a dynamic benchmark scoring unit, a two-dimensional matrix unit, and an event transmission analysis unit, completing the quantitative assessment of business opportunity value and risk; and the business output agent includes a visualization output unit, a differentiated push unit, and a lifecycle management unit, responsible for result display, business integration, and full-process status management.

[0093] The iterative optimization agent is responsible for the continuous evolution of system capabilities. Its internal data feedback processing unit is responsible for cleaning and verifying the follow-up and implementation of data fed back from the business side; the model incremental training unit performs incremental fine-tuning on entity recognition, intent recognition, and scoring models based on newly labeled data, avoiding the resource consumption and performance fluctuations caused by full retraining; the regulatory rule update unit receives publicly available regulatory rule data uniformly collected by the data collection agent, completes rule parsing, structured decomposition, and knowledge base updates, and synchronously updates the rule quadruple library and terminology dictionary after manual review. The entire iterative optimization layer and the central control scheduling layer form a closed-loop feedback mechanism, enabling the system to continuously receive business experience without interrupting service, ensuring that the multi-agent collaborative system maintains high accuracy throughout long-term operation.

[0094] Furthermore, as a specific implementation of the above method, embodiments of this application provide a multi-agent driven refinancing information parsing device, such as... Figure 9 As shown, the device includes: an acquisition unit 81, an abstraction unit 82, and a verification unit 83.

[0095] The acquisition unit 81 is used to respond to a financing event by calling the data acquisition agent to acquire multi-source heterogeneous financial data obtained through evidence causal graph fusion verification. The evidence causal graph fusion verification includes: constructing an evidence causal graph with data sources as nodes and supporting and contradictory relationships between data sources as edges; deriving the posterior distribution of the credibility of each data source based on the evidence causal graph; and outputting the fusion conclusion with the highest posterior probability as multi-source heterogeneous financial data. Abstraction unit 82 is used to call document parsing agent to parse financial documents and extract enterprise information based on the multi-source heterogeneous financial data, and abstract the enterprise information into event tuples and arrange them in chronological order to form an event chain. The event chain is marked with causal relationship. The causal relationship is determined by removing the preceding events to construct counterfactual assumptions and inferring whether the subsequent events will necessarily occur. Based on this, true causality and pseudo-correlation are distinguished. The verification unit 83 is used to call the information verification agent to perform semantic extraction on the financial document to obtain the refinancing field based on the context constraints of the event chain, generate a reverse verification question based on the refinancing field to verify the logical consistency between the refinancing fields, and trigger a second extraction to form a closed loop iteration if a contradiction is found, and output the verified structured refinancing information.

[0096] The multi-agent-driven refinancing information parsing device provided in this invention, compared with the existing technology that relies on shallow feature matching and linear processing for financing information parsing, constructs an evidence causal graph through a data acquisition agent to perform fusion reasoning on multi-source contradictory data. It derives the posterior distribution of credibility by utilizing the supporting and contradictory relationships between data sources. Compared with a single credibility scoring approach, this method can uncover implicit connections between data sources at the graph structure level, outputting the most probable fusion conclusion and improving the credibility of multi-source heterogeneous financial data. Furthermore, the document parsing agent constructs a causal event chain to remove antecedent events and constructs counterfactual hypotheses to infer whether subsequent events will necessarily occur. This distinguishes between true causality and pseudo-correlation from a causal perspective. Compared with shallow feature labeling methods such as keyword overlap, this avoids the misleading influence of pseudo-correlation on downstream parsing and verification, providing accurate causal context support. Finally, the information verification agent generates a reverse verification problem to verify the logical consistency between refinancing fields and triggers secondary extraction to form a closed-loop iteration when contradictions are found. This automatically eliminates extraction illusions and improves the accuracy and reliability of structured refinancing information.

[0097] In specific application scenarios, the acquisition unit is specifically used for: The data acquisition agent uses a dual-mode architecture that combines event-driven directional acquisition with direct interface connection for full-volume inspection to access multi-source data. Based on a preset event system, the system monitors the trigger signals of financing events, dynamically adjusts the data collection frequency according to the importance weight of the financing events, and executes financial data collection tasks through distributed deployment of various types of collectors. Based on at least one of the dimensions of source authority, data timeliness, and content completeness, the collected financial data is verified for credibility. Financial data with a credibility score greater than a preset threshold passes the verification, while financial data that fails the verification is transferred to the review queue. Evidence causal graph fusion verification is performed on financial data that has passed the credibility verification to obtain multi-source heterogeneous financial data.

[0098] In specific application scenarios, the acquisition unit is further used for: When different data sources provide contradictory information about the same financing event, the credibility scores of each data source are used as prior probabilities to initialize the credibility of each node in the evidence causal graph. The support and contradiction relationships between data sources are determined by comparing the field values ​​of different data sources for the same financing event. Based on the supporting and contradictory relationships, an evidence causal graph is constructed with edges. The credibility posterior distribution of each node is updated by iterative message passing along the edges until convergence. The field value corresponding to the data source with the highest posterior probability is selected as the fusion conclusion to obtain multi-source heterogeneous financial data.

[0099] In specific application scenarios, the abstract unit is specifically used for: The document parsing agent uses a multimodal document layout analysis model to construct a multi-layered tree structure for financial documents. The construction of the multi-layered tree structure includes at least: constructing parent-child relationships for chapters based on the title level; using a table structure recognition algorithm to restore the row and column relationships and merging relationships of cells in the table area; and filtering redundant content. The pre-trained entity recognition model extracts named entities from the multi-layer tree structure, performs vector similarity matching between the extracted named entities and the standard entity library to obtain candidate entities, and performs rule verification on the candidate entities in combination with context information to determine the target entity. The enterprise information corresponding to the target entity is abstracted into event tuples containing event subject, event type, event time and event attributes, and arranged in ascending order of time to form an event chain.

[0100] In specific application scenarios, the verification unit is specifically used for: Using the historical financing event time sequence distribution and causal relationship annotation of the target enterprise in the event chain as context constraints, the information verification intelligent agent is called to perform rule engine screening to locate candidate paragraphs in the financial document, and intent disambiguation is performed on the candidate paragraphs. The refinancing field is extracted from the paragraphs after intent disambiguation. The regulatory rules are broken down into structured rules that include at least applicable scenarios, constraints, quantitative thresholds, and special cases. The causal relationship between the refinancing field and the preceding events in the event chain is used as a joint retrieval condition to match the target rule. The result of quantitative threshold comparison, scenario judgment, and exception exclusion of the target rule, together with the logical consistency verification result between the refinancing field, is used as the basis for judging the reverse verification problem. When the judgment criteria indicate a contradiction, a correction instruction is generated based on the document fragment associated with the preceding event in the event chain that has a causal relationship with the contradiction point, triggering a second extraction of the document fragment until the verification is passed or the preset number of iterations is reached. Historical risk scores are calculated based on the event chain and superimposed negative risk factors. When the historical risk score is greater than a preset threshold, a historical risk label is marked. The verified refinancing field is merged with the historical risk label as verified structured refinancing information.

[0101] In specific application scenarios, the device further includes: a first execution unit; the first execution unit is specifically used for: After outputting the verified structured refinancing information, a multi-dimensional quantitative feature system is constructed using the verified structured refinancing information as input. The dynamic benchmark quantile value of each indicator is calculated by dividing the data into groups according to industry classification and market capitalization. The scoring weight is dynamically calculated based on the group to which the multi-dimensional quantitative feature belongs. The clue score is obtained by weighted summation of the multi-dimensional quantitative features. Based on the knowledge graph of the industrial chain, a graph attention network is used to transmit clue scores. The knowledge graph of the industrial chain includes multiple heterogeneous nodes and heterogeneous edges carrying different semantics. Upstream and downstream nodes whose transmission scores reach a preset association threshold are marked as potential association clues, and semantic interpretations representing the transmission path are output simultaneously. Using the feedback results from the business side as labels, the scoring weights are iteratively updated using an online gradient descent algorithm. After each update, single-dimensional weight fluctuation truncation and normalization are performed to ensure that the total weights are constant and the single-dimensional fluctuation does not exceed a preset range.

[0102] In specific application scenarios, the device further includes: a second execution unit; the second execution unit is specifically used for: After the verified structured refinancing information is output, the verified structured refinancing information is converted into clue data. The clue data is driven to migrate between different tracking states based on preset state transition rules. Each tracking state is set with an overdue supervision trigger condition. The time and operation identifier of each tracking state migration are recorded to form an audit log. Based on user profiles, collaborative filtering algorithms and rule matching models are used to calculate push priorities, and the clue data is pushed to target users in a differentiated manner according to the push priorities; Receive landing result data fed back from the business side, update the tracking status of the corresponding clue data based on the landing result data, and use the landing result data as a feedback tag for subsequent scoring weight iteration.

[0103] Based on the above-mentioned multi-agent driven refinancing information parsing method, this application embodiment also provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the above-mentioned multi-agent driven refinancing information parsing method.

[0104] Based on this understanding, the technical solution of this application can be embodied in the form of a software product. The software product can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, or portable hard drive), and includes several instructions to cause a computer device (such as a personal computer, server, or network device) to execute the methods described in the various implementation scenarios of this application.

[0105] Based on the above-described multi-agent-driven refinancing information parsing method and corresponding virtual device embodiments, in order to achieve the above objectives, this application embodiment also provides a physical device for multi-agent-driven refinancing information parsing, which may specifically be a computer, smartphone, tablet computer, smartwatch, server, or network device, etc. The physical device includes a storage medium and a processor; the storage medium is used to store computer programs; the processor is used to execute the computer programs to implement the above-described multi-agent-driven refinancing information parsing method.

[0106] Optionally, the physical device may also include a user interface, a network interface, a camera, radio frequency (RF) circuitry, sensors, audio circuitry, a Wi-Fi module, etc. The user interface may include a display screen, input units such as a keyboard, etc., and optional user interfaces may also include USB interfaces, card reader interfaces, etc. The network interface may optionally include standard wired interfaces, wireless interfaces (such as Wi-Fi interfaces), etc.

[0107] In an exemplary embodiment, see Figure 10 The aforementioned physical device includes a communication bus, a processor, a memory, and a communication interface. It may also include an input / output interface and a display device. The various functional units can communicate with each other via the bus. The memory stores a computer program, and the processor executes the program stored in the memory to perform the multi-agent-driven refinancing information parsing method described in the above embodiments.

[0108] Those skilled in the art will understand that the entity device structure for multi-agent driven refinancing information parsing provided in this embodiment does not constitute a limitation on the entity device, and may include more or fewer components, or combine certain components, or have different component arrangements.

[0109] The storage medium may also include an operating system and a network communication module. The operating system is a program that manages the hardware and software resources of the entity device involved in the multi-agent-driven refinancing information parsing, supporting the operation of the information processing program and other software and / or programs. The network communication module is used to enable communication between the various components within the storage medium, as well as communication with other hardware and software in the information processing entity device.

[0110] Through the above description of the implementation methods, those skilled in the art can clearly understand that this application can be implemented using software plus necessary general-purpose hardware platforms, or it can be implemented using hardware. By applying the technical solution of this application, compared with the existing methods, this application constructs an evidence causal graph through a data acquisition intelligent agent, uses the supporting and contradictory relationships between data sources to deduce the posterior distribution of credibility, and achieves the probabilistic optimal fusion of multi-source contradictory data from the graph structure level; based on the fused credible data, a document parsing intelligent agent constructs an event chain of causal relationships to remove antecedent events, constructs counterfactual hypotheses to infer whether subsequent events will necessarily occur, distinguishes between true causality and pseudo-correlation from the causal level, and provides accurate causal context for downstream parsing; based on causal context constraints, an information verification intelligent agent generates a reverse verification problem to verify the logical consistency between refinancing fields and triggers secondary extraction to form a closed-loop iteration when contradictions are found, automatically eliminating extraction illusions, and realizing progressive enhancement of credible data fusion, accurate causal inference, and reliable extraction verification.

[0111] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing this application. Those skilled in the art will understand that the modules in the apparatus of the embodiment can be distributed within the apparatus of the embodiment as described, or can be modified to be located in one or more apparatuses different from this embodiment. The modules of the above-described embodiment can be combined into one module, or further divided into multiple sub-modules.

[0112] The serial numbers in this application are for descriptive purposes only and do not represent the superiority or inferiority of any particular implementation scenario. The above disclosures are merely a few specific implementation scenarios of this application; however, this application is not limited thereto, and any variations conceived by those skilled in the art should fall within the protection scope of this application.

Claims

1. A multi-agent driven method for parsing refinancing information, characterized in that, include: In response to a financing event, the data collection agent is invoked to obtain multi-source heterogeneous financial data verified by evidence causal graph fusion. The evidence causal graph fusion verification includes: constructing an evidence causal graph with data sources as nodes and supporting and contradictory relationships between data sources as edges; deriving the posterior distribution of the credibility of each data source based on the evidence causal graph; and outputting the fusion conclusion with the highest posterior probability as multi-source heterogeneous financial data. Based on the multi-source heterogeneous financial data, a document parsing agent is invoked to parse the financial documents and extract enterprise information. The enterprise information is abstracted into event tuples and arranged in chronological order to form an event chain. The event chain is labeled with causal relationships. The causal relationships are determined by removing preceding events to construct counterfactual assumptions and inferring whether subsequent events will necessarily occur. Based on this, true causality and pseudo-correlation can be distinguished. Based on the context constraints of the event chain, the information verification agent is invoked to perform semantic extraction on the financial document to obtain the refinancing field. Based on the refinancing field, a reverse verification question is generated to verify the logical consistency between the refinancing fields. If a contradiction is found, a second extraction is triggered to form a closed-loop iteration, and the verified structured refinancing information is output.

2. The method according to claim 1, characterized in that, The process of calling the data collection agent to acquire multi-source heterogeneous financial data obtained through evidence causal graph fusion verification includes: The data acquisition agent uses a dual-mode architecture that combines event-driven directional acquisition with direct interface connection for full-volume inspection to access multi-source data. Based on a preset event system, the system monitors the trigger signals of financing events, dynamically adjusts the data collection frequency according to the importance weight of the financing events, and executes financial data collection tasks through distributed deployment of various types of collectors. Based on at least one of the dimensions of source authority, data timeliness, and content completeness, the collected financial data is verified for credibility. Financial data with a credibility score greater than a preset threshold passes the verification, while financial data that fails the verification is transferred to the review queue. Evidence causal graph fusion verification is performed on financial data that has passed the credibility verification to obtain multi-source heterogeneous financial data.

3. The method according to claim 2, characterized in that, The process of performing evidence causal graph fusion verification on the financial data that has passed the credibility check yields multi-source heterogeneous financial data, including: When different data sources provide contradictory information about the same financing event, the credibility scores of each data source are used as prior probabilities to initialize the credibility of each node in the evidence causal graph. The support and contradiction relationships between data sources are determined by comparing the field values ​​of different data sources for the same financing event. Based on the supporting and contradictory relationships, an evidence causal graph is constructed with edges. The credibility posterior distribution of each node is updated by iterative message passing along the edges until convergence. The field value corresponding to the data source with the highest posterior probability is selected as the fusion conclusion to obtain multi-source heterogeneous financial data.

4. The method according to claim 1, characterized in that, Based on the multi-source heterogeneous financial data, the document parsing intelligence is invoked to parse the financial documents and extract enterprise information. This enterprise information is then abstracted into event tuples and arranged chronologically to form an event chain, including: The document parsing agent is invoked, and a multi-modal document layout analysis model is used to construct a multi-layered tree structure for financial documents. The construction of the multi-layered tree structure includes at least: constructing the parent-child relationship of chapters based on the title level; using a table structure recognition algorithm to restore the row and column relationships and merging relationships of cells in the table area; and filtering redundant content. The pre-trained entity recognition model extracts named entities from the multi-layer tree structure, performs vector similarity matching between the extracted named entities and the standard entity library to obtain candidate entities, and performs rule verification on the candidate entities in combination with context information to determine the target entity. The enterprise information corresponding to the target entity is abstracted into event tuples containing event subject, event type, event time and event attributes, and arranged in ascending order of time to form an event chain.

5. The method according to claim 1, characterized in that, Based on the context constraints of the event chain, the information verification agent is invoked to semantically extract the refinancing field from the financial document. A reverse verification question is generated based on the refinancing field to verify the logical consistency between the refinancing fields. If a contradiction is found, a second extraction is triggered to form a closed-loop iteration, outputting verified structured refinancing information, including: Using the historical financing event time sequence distribution and causal relationship annotation of the target enterprise in the event chain as context constraints, the information verification intelligent agent is called to perform rule engine screening to locate candidate paragraphs in the financial document, and intent disambiguation is performed on the candidate paragraphs. The refinancing field is extracted from the paragraphs after intent disambiguation. The regulatory rules are broken down into structured rules that include at least applicable scenarios, constraints, quantitative thresholds and special cases. The causal relationship between the refinancing field and the preceding events in the event chain is used as a joint retrieval condition to match the target rule. The result of quantitative threshold comparison, scenario judgment and exception exclusion of the target rule, together with the logical consistency verification result between the refinancing field, is used as the basis for judging the reverse verification problem. When the judgment criteria indicate a contradiction, a correction instruction is generated based on the document fragment associated with the preceding event in the event chain that has a causal relationship with the contradiction point, triggering a second extraction of the document fragment until the verification is passed or the preset number of iterations is reached. Historical risk scores are calculated based on the event chain and superimposed negative risk factors. When the historical risk score is greater than a preset threshold, a historical risk label is marked. The verified refinancing field is merged with the historical risk label as verified structured refinancing information.

6. The method according to claim 1, characterized in that, After outputting the verified structured refinancing information, the method further includes: invoking the clue evaluation agent to perform the following steps: Using the verified structured refinancing information as input, a multi-dimensional quantitative feature system is constructed. The dynamic benchmark quantile value of each indicator is calculated by dividing the data into groups according to industry classification and market capitalization. The scoring weight is dynamically calculated based on the group to which the multi-dimensional quantitative feature belongs. The clue score is obtained by weighted summation of the multi-dimensional quantitative features. Based on the knowledge graph of the industrial chain, a graph attention network is used to transmit clue scores. The knowledge graph of the industrial chain includes multiple heterogeneous nodes and heterogeneous edges carrying different semantics. Upstream and downstream nodes whose transmission scores reach a preset association threshold are marked as potential association clues, and semantic interpretations representing the transmission path are output simultaneously. Using the feedback results from the business side as labels, the scoring weights are iteratively updated using an online gradient descent algorithm. After each update, single-dimensional weight fluctuation truncation and normalization are performed to ensure that the total weights are constant and the single-dimensional fluctuation does not exceed a preset range.

7. The method according to any one of claims 1 to 6, characterized in that, After outputting the verified structured refinancing information, the method further includes: invoking an information verification agent to perform the following steps: The verified structured refinancing information is converted into clue data. The clue data is driven to migrate between different tracking states based on preset state transition rules. Each tracking state is set with an overdue supervision trigger condition. The time and operation identifier of each tracking state migration are recorded to form an audit log. Based on user profiles, collaborative filtering algorithms and rule matching models are used to calculate push priorities, and the clue data is pushed to target users in a differentiated manner according to the push priorities; Receive landing result data fed back from the business side, update the tracking status of the corresponding clue data based on the landing result data, and use the landing result data as a feedback tag for subsequent scoring weight iteration.

8. A multi-agent driven refinancing information analysis device, characterized in that, include: The acquisition unit is used to respond to financing events by calling the data acquisition agent to acquire multi-source heterogeneous financial data obtained through evidence causal graph fusion verification; The evidence causal graph fusion verification includes: constructing an evidence causal graph with data sources as nodes and supporting and contradictory relationships between data sources as edges; deriving the posterior distribution of the credibility of each data source based on the evidence causal graph; and outputting the fusion conclusion with the highest posterior probability as multi-source heterogeneous financial data. The abstraction unit is used to call the document parsing agent to parse the financial documents and extract enterprise information based on the multi-source heterogeneous financial data. The enterprise information is abstracted into event tuples and arranged in chronological order to form an event chain. The event chain is labeled with causal relationships. The causal relationships are determined by removing the preceding events to construct counterfactual assumptions and inferring whether subsequent events will necessarily occur. Based on this, true causality and pseudo-correlation are distinguished. The verification unit is used to call the information verification agent to perform semantic extraction on the financial document to obtain the refinancing field based on the context constraints of the event chain, generate a reverse verification question based on the refinancing field to verify the logical consistency between the refinancing fields, and trigger a second extraction to form a closed loop iteration if a contradiction is found, and output the verified structured refinancing information.

9. 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 steps of the method according to any one of claims 1 to 7.

10. A computer storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images