Financial information management method based on knowledge graph

Abstract

The invention discloses a financial information management method based on a knowledge graph, and the method comprises the following steps: S1, collecting multi-source heterogeneous financial data, and generating an initial financial data set; s2, extracting a financial entity and semantic relationship, constructing a triple set, and forming an initial financial knowledge graph; s3, performing semantic enhancement processing, and constructing a semantic enhanced financial graph; s4, setting an atlas updating scheduling mechanism, executing a dynamic updating operation, and generating a dynamic atlas structure; s5, inputting into an improved NodeFormer model, and outputting a risk path identification result; s6, constructing a map semantic question and answer structure, and outputting a structured query response result; and S7, constructing an atlas visual display module, and generating an atlas analysis view and an audit suggestion result. According to the method, the identification precision and the tracking capability of the risk path in the financial data are improved, and the transparency, the response efficiency and the risk prevention and control capability of the financial auditing process are effectively enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data intelligence and financial information management, and in particular to a financial information management method based on a knowledge graph. BACKGROUND

[0002] With the integrated development of financial information systems and the improvement of intelligent risk control needs, multi-source financial data management and analysis technologies around contract execution, budget control, invoice compliance and fund flow have attracted widespread attention. Existing financial management platforms mainly rely on structured databases and rule engines to interface financial systems, budget systems, invoice platforms and other modules, build field-level data correlation models, and perform operations such as voucher verification, budget review and risk prompt. However, the following problems exist in actual business:

[0003] Multi-source financial data has great heterogeneity in structure definition, field standard and time format. Traditional static data integration methods cannot adapt to frequently updated contract status, budget adjustments and invoice changes, resulting in delayed patterns, data gaps and lagging risk nodes. Risk identification mechanisms based on rule retrieval cannot effectively establish multi-hop logical paths between contracts, invoices and accounts, lack modeling capabilities for risk propagation chains, and are difficult to handle hidden financial abnormalities across cycles, modules and entities. Risk query functions for business personnel are mostly field filtering or form searching, lack natural language support and semantic understanding capabilities, have coarse response granularity and poor interpretability, and are difficult to meet the requirements of risk chain panoramic tracing in financial audit and supervision scenarios.

[0004] Therefore, how to provide a financial information management method based on a knowledge graph is a problem that those skilled in the art need to solve. SUMMARY

[0005] One object of the present application is to provide a financial information management method based on a knowledge graph. The present application makes full use of initial financial data set construction, semantic enhanced financial graph generation, graph update scheduling mechanism setting, improved NodeFormer model risk path identification, graph semantic question and answer structure construction and graph visualization display module output. The present application describes in detail the whole process of collecting multi-source heterogeneous financial data, constructing an initial financial knowledge graph, embedding time labels, business labels, risk labels, responsibility labels and source labels, executing dynamic graph structure update, identifying risk paths, generating structured query response results, and outputting graph analysis views and audit suggestion results. The present application has the advantages of unified data processing structure, rich graph expression semantics, high risk identification accuracy and strong audit assistance intelligence.

[0006] According to the financial information management method based on the knowledge graph, the following steps are included: S1, collect multi-source heterogeneous financial data from the financial system, the budget system, the contract management system, the bank interface and the invoice platform, perform field standardization processing, format normalization processing, timestamp unification processing and data desensitization processing, and generate an initial financial data set; S2, perform entity recognition and relationship extraction operations on the initial financial data set, extract financial entities and semantic relationships, construct a triple set, and form an initial financial knowledge graph; S3, perform semantic enhancement processing on the initial financial knowledge graph, embed time labels, business labels, risk labels, responsibility labels and source labels, and construct a semantic enhanced financial graph; S4, set up a graph update scheduling mechanism, perform dynamic update operations on the semantic enhanced financial graph, including regular updates based on time windows and local structure updates based on event triggers, and generate a dynamic graph structure; S5, input the dynamic graph structure into the improved NodeFormer model, perform risk path identification operations, and output risk path identification results; S6, construct a graph semantic question and answer structure based on the risk path identification results, and output a structured query response result; the graph semantic question and answer structure includes natural language question analysis, graph node matching and path retrieval operations; S7, construct a graph visualization display module, perform node highlighting, path tracking, label filtering and risk prompting operations, and generate a graph analysis view and an audit suggestion result.

[0007] Preferably, the S1 is specifically: A unified data structure template is set, which specifies field format, numerical accuracy, time specification and character encoding rules; Accounting voucher data, accounting subject data and payment and receipt detail data are collected from the financial system, budget item data, budget control data and approval records are collected from the budget system, contract numbers, clause information, performance status and payment plans are collected from the contract management system, account transaction records and fund flow information are collected from the bank interface, and invoice registration records, image recognition results and verification results are collected from the invoice platform; Perform field mapping and format conversion processing on various data, unify field structure, unit standard and data type, fill in missing fields and remove invalid records; unify the time format in all data, and identify and distinguish business time and system time; Perform desensitization processing on data related to identity information, account number, contract sensitive fields, use mask marking or encryption method to process sensitive fields, and generate an initial financial data set.

[0008] Preferably, the S2 is specifically: The financial entity category set is constructed, and a financial entity type is set; the financial entity type includes a contract number, a budget subject, an invoice number, an account identifier, a department name, an approval node, and a fund flow identifier; A lexical rule and a context structure template are constructed according to the financial entity category, an entity recognition operation is performed based on field content, a syntax structure, and a context position, and field content with specific semantics and financial business meaning in the initial financial data set is recognized; A set of financial semantic relationships is defined, and a semantic relationship type is set; the semantic relationship type includes an inclusion relationship, a control relationship, a fund flow relationship, a contract constraint relationship, a budget attribution relationship, and an invoice correspondence relationship; A relationship extraction model is constructed, a relationship determination operation is performed on the recognized entity group in combination with syntax dependency analysis and entity context analysis, and a semantic connection mode between entities is determined; The recognized financial entities and semantic relationships are organized in a triple format, the triple includes a starting entity, a relationship type, and a target entity, and an initial financial knowledge graph is generated.

[0009] Preferably, the S3 is specifically: A set of label types is set, and the set of label types includes a time label, a business label, a risk label, a responsibility label, and a source label; Based on the attribute information and the relationship structure of the financial entities in the initial financial knowledge graph, corresponding time dimensions, business classifications, risk identifiers, organizational attributions, and data source field contents are extracted; the extraction results are respectively converted into structured label data, wherein the time label includes annual, quarterly, and monthly information, the business label includes expense category, expenditure item, and project type information, the risk label includes an over-budget mark, a frequent transaction mark, and a repeated reimbursement mark, the responsibility label includes the department to which the label belongs and the responsible personnel information, and the source label includes a system source identifier and a data generation mode identifier; The structured label data is embedded into the node attribute set and the edge attribute set of the initial financial knowledge graph, corresponding label information is added to each node and edge in the graph, and a semantic-enhanced financial graph is constructed.

[0010] Preferably, the S4 is specifically: The time window update scheduling unit sets a fixed update period parameter, performs a graph structure update operation at a set daily, weekly, or monthly cycle, parses and processes new financial data in each cycle, compares the parsing results with the nodes and edges in the semantic-enhanced financial graph, and if new entities or new relationships are recognized, performs a node addition, edge connection, or attribute supplement operation; The event trigger update scheduling unit sets event listening rules and trigger conditions, including budget execution exceptions, contract state changes, invoice cancellation records, approval chain changes, and fund flow exceptions. If an event that meets the event listening rules and trigger conditions is detected, the corresponding local graph update operation of the region is immediately performed. The update content includes node state marker update, relationship connection reconstruction, and label content revision; The time window update scheduling unit and the event trigger update scheduling unit output update operation logs synchronously, record the timestamp, modification type, modification target, and modification source of the semantic enhancement type financial graph modification, and generate a dynamic graph structure.

[0011] Preferably, the improved NodeFormer model includes a label attention guide module, a state evolution memory module, a global path comparison module, and a fusion output module, specifically: The structure guide connection module constructs a node adjacency candidate matrix based on the dynamic graph structure, calculates the structural similarity weight between nodes, forms a structure attention distribution, and maps the attention distribution to a structure guide connection strength matrix; The label attention guide module extracts the time label, business label, risk label, responsibility label, and source label attached to the node and edge, constructs a multi-dimensional label embedding matrix, calculates the label similarity weight between node pairs based on the label embedding matrix, and performs element-wise fusion operation on the label similarity weight and the structure guide connection strength matrix to generate a label-aware connection matrix; The state evolution memory module constructs a node state time series, constructs the connection changes, attribute updates, and label changes of the node in multiple time windows into a state input sequence, introduces a gated recurrent structure and a memory unit to construct a node state evolution trajectory, and encodes the state input sequence to generate a node evolution representation vector; The global path comparison module constructs a candidate path set, extracts the node evolution representation and label-aware connection strength in the path, constructs a path feature sequence, selects a low-risk financial path as a comparison sample, performs comparison learning operation based on the path feature sequence, and calculates the path deviation score; The fusion output module performs weighted fusion on the node evolution representation, label-aware connection strength, and path deviation score to generate a path confidence distribution. Paths with confidence greater than a pre-set confidence threshold are marked as risk paths, and the risk path recognition result is output.

[0012] Preferably, S6 specifically includes: Perform lexical analysis, dependency syntax analysis, and semantic role labeling on the input question text, extract target entities, target relationships, and inquiry intentions, construct a graph semantic question and answer structure, and generate structured response data; The structured response data is converted into a question and answer output format, and a query response result containing a target node, a path node sequence, a path type, a key label and risk prompt information is output.

[0013] Preferably, the graph semantic question and answer structure includes natural language question analysis, graph node matching and path retrieval operations, specifically: Based on the target entity name in the inquiry intent, the name similarity score of the target entity and the node name in the semantic enhanced financial graph is calculated, and the corresponding time label, business label, risk label, responsibility label and source label are extracted, and the label consistency score is calculated. The name similarity score and the label consistency score are normalized and weighted fused to generate a comprehensive matching score. Set the comprehensive matching threshold, filter the nodes with a comprehensive matching score greater than the comprehensive matching threshold as the candidate node set; Based on the candidate node set and the target relationship in the inquiry intent, the associated path containing the candidate node is retrieved from the dynamic graph structure, and the pattern matching operation is performed on the node sequence, relationship type and label content in the path and the inquiry intent to generate a path candidate set; Call the risk path identification result, mark the path type attribute of the risk path and the non-risk path in the path candidate set respectively, and organize the structured response data according to the set format according to the node state evolution information, label content and connection strength in the path.

[0014] Preferably, the S7 is specifically: Based on the node set and edge set in the dynamic graph structure, render the graph structure graph view, display the relationship connection, attribute information and label content between nodes; Perform node highlighting, path tracking, label filtering and risk prompt operations to generate graph analysis view and audit suggestion results; the graph analysis view includes node graph, path connection, label information and highlight identification; the audit suggestion result includes risk source, risk node, associated path and suggestion content list.

[0015] Preferably, the execution of node highlighting, path tracking, label filtering and risk prompt operation is specifically: According to the risk path identification result, mark the risk node and the risk path, perform the node highlighting operation on the risk node, superimpose the color identification and the label information prompt, perform the path highlighting operation on the risk path, render the path connection style and the connection weight; Set the path tracking function, receive the start node and the target node input, retrieve the path sequence in the dynamic graph structure, display the intermediate nodes, relationship types and connection order in the path, and mark the style of the risk nodes in the path; Set the label filtering condition set, including time label, business label, risk label, responsibility label and source label, keep the node and path meeting any label filtering condition in visible state, execute the hiding processing on the node and path not meeting the label filtering condition; Based on the risk label level of the node and the path confidence distribution, the risk prompt operation is performed on the node and path greater than the preset risk threshold, the warning mark is attached in the dynamic graph structure, and the risk label type, offset score and associated path information are displayed.

[0016] The beneficial effects of the present application are: The present application constructs an initial financial knowledge graph and performs semantic enhancement processing, generates an initial financial data set with a unified structure by field standardization, format regularization and timestamp unification, constructs a triple set through entity recognition and relation extraction, further constructs a semantic enhancement type financial graph, realizes accurate modeling of financial information at the semantic level and multi-angle expression of semantic dimension, sets a graph update scheduling mechanism in the graph dynamic update link, fuses regular update based on time window and local structure update based on event trigger, realizes high-frequency and efficient update of dynamic graph structure, guarantees the freshness of knowledge graph content and the timeliness of structure, inputs the dynamic graph structure into the improved NodeFormer model in the risk identification link, the model constructs path confidence distribution and outputs risk path identification result by introducing structure guided connection, label attention fusion, state evolution memory and global path comparison mechanism, improves the robustness and precision of financial risk identification, constructs graph semantic question and answer structure based on risk path identification result in the auxiliary query link, realizes automatic execution of natural language problem analysis, graph node matching and path retrieval, generates structured query response result, improves the intelligent level of financial query, finally executes node highlighting, path tracking, label filtering and risk prompt operation through the graph visualization display module, outputs graph analysis view and audit suggestion result, enhances the visibility and operability of audit result, significantly improves the accuracy, real-time performance and intelligent level of financial information management. BRIEF DESCRIPTION OF DRAWINGS

[0017] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, illustrate the application together with the embodiments thereof, and explain the application without limiting the application. In the drawings:

[0018] Fig. 1 A flowchart of a financial information management method based on a knowledge graph according to the present application is provided; Fig. 2 A schematic diagram of an improved NodeFormer model according to the present application is provided; Fig. 3 A data flow diagram of a financial information management method based on a knowledge graph is proposed for the present application. DETAILED DESCRIPTION

[0019] The present application will now be further described in detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams and only illustrate the basic structure of the present application in a schematic manner, and therefore only show the components related to the present application.

[0020] REFERENCE Figs. 1-3 A financial information management method based on a knowledge graph includes the following steps: S1, collect multi-source heterogeneous financial data from a financial system, a budget system, a contract management system, a bank interface and an invoice platform, perform field standardization processing, format normalization processing, timestamp unification processing and data desensitization processing, and generate an initial financial data set; S2, perform entity recognition and relationship extraction operations on the initial financial data set, extract financial entities and semantic relationships, construct a triple set, and form an initial financial knowledge graph; S3, perform semantic enhancement processing on the initial financial knowledge graph, embed time labels, business labels, risk labels, responsibility labels and source labels, and construct a semantic enhanced financial graph; S4, set up a graph update scheduling mechanism, perform dynamic update operation on the semantic enhanced financial graph, the dynamic update operation includes regular update based on time window and local structure update based on event trigger, and generate dynamic graph structure; S5, input the dynamic graph structure into the improved NodeFormer model, perform risk path identification operation, and output risk path identification result; S6, construct a graph semantic question and answer structure based on the risk path identification result, and output a structured query response result; the graph semantic question and answer structure includes natural language question analysis, graph node matching and path retrieval operation; S7, construct a graph visualization display module, perform node highlighting, path tracking, label filtering and risk prompting operation, and generate a graph analysis view and an audit suggestion result.

[0021] The present embodiment collects multi-source heterogeneous financial data from a financial system, a budget system, a contract management system, a bank interface and an invoice platform, performs field standardization processing, format normalization processing, timestamp unification processing and data desensitization processing, and generates an initial financial data set with uniform structure, clear semantics and strong comparability, realizing the centralized fusion and cleaning of multi-system scattered data; by performing entity recognition and relationship extraction operations on the initial financial data set, a triple set is constructed and an initial financial knowledge graph is formed, improving the structured degree and semantic expression ability of the financial data.

[0022] Further, by embedding time labels, business labels, risk labels, responsibility labels, and source labels into the initial financial knowledge graph, a semantic enhanced financial graph is constructed, enhancing the expression ability of the graph structure in terms of time, responsibility attribution, business background, and risk dimensions; a graph update scheduling mechanism is set, and regular updates based on time windows and local structure updates based on event triggers are performed to generate dynamic graph structures, making the graph content have timeliness, dynamics, and change response ability; the dynamic graph structure is input into the improved NodeFormer model, combined with structure guided connection, label attention mechanism, state evolution memory, and path comparison to calculate path confidence distribution, effectively identifying high-risk paths, and realizing deep mining and dynamic early warning of financial risks.

[0023] Further, based on the risk path identification results, a graph semantic question and answer structure is constructed to support natural language question analysis, graph node matching, and path retrieval operations, improving the queryability and interactivity of complex graph data; at the same time, a graph visualization display module is constructed to perform node highlighting, path tracking, label filtering, and risk prompt operations to generate graph analysis views and audit suggestion results, realizing the visualization of key risk information and audit assistance, thereby improving the intelligent level, risk identification ability, and decision support effect of the financial information management process.

[0024] In the embodiment, S1 is specifically: A unified data structure template is set, specifying field format, numerical precision, time specification, and character encoding rules; Accounting voucher data, accounting subject data, and payment and receipt detail data are collected from the financial system, budget item data, budget control data, and approval records are collected from the budget system, contract numbers, clause information, performance status, and payment plans are collected from the contract management system, account transaction records and fund flow information are collected from the bank interface, and invoice registration records, image recognition results, and verification results are collected from the invoice platform; Field mapping and format conversion processing is performed on various data to unify field structure, unit standard, and data type, fill in missing fields, and remove invalid records; the time format in all data is unified, and business time and system time are identified and distinguished; Desensitization processing is performed on data related to identity information, account number, and contract sensitive fields, and sensitive fields are processed using mask marking or encryption to generate an initial financial data set.

[0025] In the embodiment, S2 is specifically: A set of financial entity categories is constructed, and financial entity types are set; the financial entity types include contract number, budget subject, invoice number, account identifier, department name, approval node, and fund flow identifier; constructing a lexical rule and a context structure template according to the financial entity category, performing an entity recognition operation based on field content, syntax structure and context position, and identifying field content with specific semantics and financial business meaning in the initial financial data set; defining a set of financial semantic relationships and setting semantic relationship types; the semantic relationship types include inclusion relationship, control relationship, fund flow relationship, contract constraint relationship, budget attribution relationship and invoice correspondence relationship; constructing a relationship extraction model, combining syntax dependency analysis and entity context analysis, performing relationship determination operation on the identified entity group, and determining the semantic connection mode between entities; organize the identified financial entities and semantic relationships in a triple format, the triple includes the starting entity, the relationship type and the target entity, and generate an initial financial knowledge graph.

[0026] In the embodiment, the S3 is specifically: Based on the attribute field information of each node and edge in the initial financial knowledge graph, a label extraction parameter table is constructed, which records the field name, label type, legal value range and abnormal value filtering rule in the label extraction parameter table; Based on the label extraction parameter table, the initial financial knowledge graph is traversed node by node and edge by edge, the time field, business field, risk information field, organization attribution field and source field are read, the date analysis and dimension decomposition are performed on the time field, the annual, quarterly and monthly discrete values are extracted, the keyword mapping and field merging are performed on the business field, and the expense subject, expenditure item and project type content are extracted; Based on the risk information field, it is judged whether there is a risk mark such as budget overrun, fund flow anomaly, frequent reimbursement, invoice cancellation, etc., a risk label is written for the node or edge containing the risk behavior, and a default security label is written for the node or edge without risk record; Based on the organization attribution field, the department name, responsible person name and post identifier are extracted, the data source system type, interface identifier and generation mode identifier are extracted from the source field, and the above extraction results are constructed into a label description vector; Construct a label standardization mapping rule to convert time values, text labels and risk marks into a unified encoding form, the encoding format includes numerical encoding, Boolean encoding and enumeration encoding, and form a node label attribute set and an edge label attribute set through vector splicing; Construct a label conflict detection rule to detect data with repeated, missing or contradictory labels, if the same node has inconsistent label content in multiple data sources, perform label priority judgment according to timestamp and source field, and retain the latest or more reliable label content; The label attribute set is embedded into the knowledge graph storage structure, a label attribute key-value pair is added to the node attribute field, a time label, a risk label and the like are added to the edge attribute field, and an empty label placeholder is automatically filled for data without labels, so as to ensure the integrity of the graph structure and the uniformity of the label field; A label integrity checking rule is constructed, all nodes and edges in the graph are traversed, it is verified whether the five label types of time label, business label, risk label, responsibility label and source label are contained, the number of missing labels and the coverage rate index are recorded, if the coverage rate is lower than the set threshold, an early warning information is output, and an un-covered element list is written in the update log; Further, an index table associated with the label and the structure is constructed, the index table records the mapping relationship between the label value and the node number, the edge number and the relationship type, and supports subsequent fast retrieval, statistical analysis and risk screening operations based on the label.

[0027] In the embodiment, the S4 is specifically: A graph update scheduling mechanism is set, the graph update scheduling mechanism includes a time window update scheduling unit and an event triggered update scheduling unit; The time window update scheduling unit sets the update cycle type and the update timestamp range, supports the timing scanning logic of the day cycle, the week cycle and the month cycle; at each cycle arrival time, the newly added financial data in the cycle start and end range is obtained, the field standardization processing, the format regularization processing, the timestamp unification processing and the data desensitization processing procedures are called, and the standardized financial data in the update cycle is generated; The standardized financial data is subjected to entity recognition and relationship extraction operations, a triple candidate set is extracted, and similarity comparison is performed with the node set and the edge set in the semantic enhanced financial graph, if the starting entity, the target entity or the semantic relationship not contained in the graph is found, it is marked as a graph difference unit; The structure supplement operation is performed on the graph difference unit, including node addition, edge connection and attribute expansion, the node addition operation writes the entity identifier and the label set in the graph structure, the edge connection operation constructs the relationship type, the connection strength and the connection timestamp, and the attribute expansion operation writes the business field and the semantic label field; The event trigger update scheduling unit configures an event listening list, the event listening list including a budget execution exception event, a contract state change event, an invoice cancellation event, an approval chain change event and a fund flow exception event; a trigger field, a trigger logic and a weight factor of each type of event are set, if an event record meeting the event listening rule and the corresponding trigger condition is listened to in the new data, the trigger state is marked as valid; the event listening rule is a listening range rule set based on data source, entity type and semantic label, the listening range including a change behavior marked by a specified system type, a specific entity category and a specific risk label; the trigger condition is a rule set based on value range change, state transition or logic conflict determination, the trigger situation including a budget execution progress exceeding a threshold, a contract performance state changing from execution to termination, an invoice state changing from true to cancellation, an approval process adding a node or replacing an approver, a fund flow direction having an abnormal ring jump or repeated circulation; For the event record marked as valid, its corresponding graph sub-region is extracted, its position in the semantic enhanced financial graph is tracked through entity identification, and local structure update operation is performed on the corresponding region; the local structure update operation includes node state field modification, edge connection state reconstruction and label field revision, wherein the state field modification includes risk level marking, freezing state identification and responsibility person transfer marking, the edge connection state reconstruction includes chain breaking, reconnection and connection direction adjustment, and the label field revision includes time label updating, business label refinement and risk label upgrading; The time window update scheduling unit and the event trigger update scheduling unit synchronously write a graph update log, the update log including an operation timestamp, an operation type, an operation target number and an operation source system, the log content and the dynamic graph structure being bound to form a structured update index; Further, a graph structure snapshot recording mechanism is constructed, a structure snapshot of the dynamic graph structure being generated after each update, the structure snapshot including a current node set, an edge set and a label state mapping table; each structure snapshot is provided with a unique number and a timestamp, which can be used for subsequent graph evolution comparison, update rollback and version recovery operations.

[0028] In the embodiment, the improved NodeFormer model includes a label attention guide module, a state evolution memory module, a global path comparison module and a fusion output module, specifically: The node set and the edge set in the dynamic graph structure are used to construct an initial adjacency candidate matrix in the structure-guided connection module. The structure correlation is calculated according to the connection times, the connection direction and the connection timestamp for any two nodes. The structure similarity weight between the nodes is calculated by a structure similarity function. The structure similarity weight is mapped into a structure attention distribution by a normalization method. A structure-guided connection strength matrix is generated by a weighted mapping method. Each element in the structure-guided connection strength matrix represents the potential connection strength and the influence degree of two nodes in the graph structure. The time label, the business label, the risk label, the responsibility label and the source label of each node and edge in the graph are encoded and processed in the label attention guiding module to generate a multi-dimensional label embedding matrix. The label similarity weight between the node pairs is calculated by the vector similarity. The label similarity weight and the structure-guided connection strength matrix are fused by the element-by-element multiplication and addition method. A label-aware connection matrix is generated by the weight normalization method. The label-aware connection matrix retains the joint weight information of the structure strength and the label semantics, and can describe the business similarity and the structure correlation between the nodes. In the state evolution memory module, a state input sequence is constructed for each node. The state input sequence is composed of the structure connection change value, the attribute field update record and the label change record of the node in multiple time windows. The state input sequence is received by a gated recurrent structure. The long-term dependence information is saved by a memory unit to form a node state evolution trajectory. The node evolution representation vector is generated by an encoding mapping method. The node evolution representation vector contains the node historical behavior pattern, the risk change trend and the label evolution path information. If the prediction error needs to be calculated, the node state offset trend is obtained by the least mean square error fitting. In the global path comparison module, a candidate path set is generated based on the dynamic graph structure. The path length can be fixed or adaptive. The node evolution representation of each node in the path and the label-aware connection strength of each connection in the path are extracted to construct a path feature sequence. From the historical data, the paths that meet the budget normal execution, the contract compliance and the no abnormal transaction record are selected as low-risk financial paths. The path feature sequence is subjected to a comparative learning operation. The path offset score is obtained by calculating the difference between the high-risk path and the low-risk path in the node evolution representation and the connection weight. The risk discrimination boundary can be generated by the logistic regression fitting of the path offset score. The node evolution representation vector, the label-aware connection matrix and the path offset score are fused by the fusion output module to generate a path confidence distribution. Each element in the path confidence distribution represents the probability value of the path being a risk path. The paths with a confidence greater than a threshold are selected by setting a preset confidence threshold. The risk path recognition result is output. The risk node, the risk label information and the offset score are retained as traceable records. By jointly modeling the node structure features, semantic label features, and time evolution information, hidden risk paths can be accurately identified in complex financial dynamic environments, enabling high-precision risk tracing analysis and path-level risk positioning. Compared to models relying solely on static graphs or single structure features, this approach has higher accuracy, interpretability, and real-time performance.

[0029] In this embodiment, S6 is specifically: Performing lexical analysis, dependency syntax analysis, and semantic role labeling on the input question text, extracting target entities, target relationships, and query intentions; constructing a graph semantic question answering structure and generating structured response data; Converting the structured response data into a question and answer output format, and outputting the query response results containing target nodes, path node sequences, path types, key labels, and risk prompt information.

[0030] In this embodiment, the graph semantic question answering structure includes natural language question analysis, graph node matching, and path retrieval operations, specifically: Setting the node name similarity calculation method, using an edit distance-based matching algorithm and a vector space cosine similarity calculation method to construct a node name matching function, and extracting the edit distance features and vector semantic features between the target entity name and each node name in the semantic enhanced financial graph; performing normalization processing on the edit distance features and vector semantic features, and calling a weighted fusion function to generate a name similarity score; the name similarity score is fitted through the node name matching function and the name similarity fusion rule; Extracting the time labels, business labels, risk labels, responsibility labels, and source labels corresponding to the target entity and the candidate nodes, and constructing a label nested set; performing item-by-item comparison operations on the target entity label set and the candidate node label set according to the corresponding order of the label fields, recording the total number of consistent items and the proportion of the total number of fields, and generating a label consistency score; Performing normalization processing on the name similarity score and the label consistency score, and calling a linear weight function to calculate the fusion result according to the set proportion weight, and generating a comprehensive matching score; setting a comprehensive matching threshold, and selecting the graph nodes with a comprehensive matching score greater than the comprehensive matching threshold to form a candidate node set; Setting a target relationship mode set, performing path traversal operations on the candidate nodes in the dynamic graph structure according to the target relationship field in the query intention, limiting the maximum path length and relationship jump number range, and performing structure mode matching and semantic mode matching operations on the node sequence, relationship type, and label content in the path and the target relationship mode set, outputting the path sequence set that meets the conditions, and generating a path candidate set; The risk path identification result is called to call a risk path marking module for each path in the path candidate set, read the path confidence distribution, and mark the path greater than the risk path threshold as a risk path and mark other paths as non-risk paths, and respectively assign path type attributes; For each path in the path candidate set, read the node state evolution time sequence, node label set and node connection weight, collect and process each type of information according to the preset structured response format, organize the structured response data, and output the query response result containing the target node, path node sequence, path type, key label and risk prompt information.

[0031] In the embodiment, S7 is specifically: Based on the node set and edge set in the dynamic graph structure, a graph structure graph view is rendered to display the relationship connection, attribute information and label content between nodes; Perform node highlighting, path tracking, label filtering and risk prompt operations to generate a graph analysis view and an audit suggestion result; the graph analysis view includes node graphics, path connections, label information and highlight identifiers; the audit suggestion result includes risk sources, risk nodes, associated paths and suggestion content lists.

[0032] In the embodiment, the execution of the node highlighting, path tracking, label filtering and risk prompt operations is specifically: According to the path confidence distribution and risk path label set output in the risk path identification result, a marked index table is established for the nodes and edges contained in all paths marked as risk paths; a node graphics rendering engine is called to map the nodes in the marked index table to a color identification scheme according to the risk level, and to superimpose highlight colors, border styles and label floating window content in the graph visualization interface to perform node highlighting operations; a path connection rendering engine is called to adjust the line width, color and solid-void style of the path style according to the connection relationship corresponding to the risk path, and to control the path gradient effect according to the connection weight value to perform path highlighting operations; A path tracking controller is constructed to receive user input of starting node and target node identifiers, call a graph search function from the dynamic graph structure to retrieve path sequences that meet the connection conditions; perform format standardization processing on all intermediate nodes, relationship types and connection direction information in the path sequence to render a continuous path navigation view; identify whether there are marked risk nodes in the path sequence, and if so, add highlight markers and risk label prompt information to the risk nodes in the path navigation view; Set a set of label filtering conditions, including five structured filtering dimensions of time label, business label, risk label, responsibility label and source label; extract the embedded label content of each node and edge in the graph structure, perform label matching operation, filter the graph elements meeting any filtering condition and add them to the visible set, add the graph elements not meeting the condition to the hidden set, call the graph rendering engine to refresh the current visualization area and update the display content after filtering; Extract the risk label level field content of the nodes in the identified risk path, and extract the path confidence score of the path; set a preset risk threshold, filter out the nodes and paths whose risk label level is greater than the set level threshold or whose path confidence score is greater than the preset confidence threshold; perform a risk prompt operation on the nodes and paths, and add warning icons, offset score prompts and risk label type explanations in the graph structure.

[0033] Embodiment 1 In order to verify the feasibility of the application in implementation, the application is applied to the financial information management practice of a large manufacturing enterprise group. The group has multiple subsidiaries under its flag, and needs to process a large amount of heterogeneous financial data from the financial system, budget system, contract management system, bank interface and invoice platform on a daily basis. In the traditional information management process, there are problems such as serious data redundancy, difficulty in identifying risk transactions, and lagging audit response, which makes it difficult to support the complex financial review and risk control needs within the group.

[0034] In this scenario, the enterprise introduces a financial information management method based on knowledge graph proposed by the application into its financial data governance platform, and realizes the core functions of unified modeling, semantic enhancement, risk identification and intelligent question answering of massive financial data by deploying the graph engine, improved NodeFormer model, semantic question answering interface and graph visualization components.

[0035] In the application process, the enterprise first completes the collection and processing of the initial financial data set based on the application, and collects more than 4.7 million structured and semi-structured data from five systems, including 570,000 contract records, 1.04 million budget items, 1.22 million invoice data, 980,000 bank flow data and 890,000 voucher data, and uniformly performs field normalization, timestamp synchronization and desensitization processing. Then, according to S2, entity recognition and relationship extraction are performed, more than 3.7 million triples are constructed, and five types of labels of time, business, risk, responsibility and source are embedded in the graph through S3, realizing multi-dimensional semantic enhancement of the graph.

[0036] In a three-month deployment run, the system of the present application performs a time window-based periodic update every 24 hours and listens to abnormal events such as invoice cancellation, approval chain changes, etc. to trigger local updates in real time, effectively ensuring the real-time nature of the dynamic graph structure. For the key business department's fund flow and contract performance data, the improved NodeFormer model performs risk path identification operations and automatically generates graph semantic question and answer responses and graph visualization outputs, supporting the audit team to quickly locate potential risk links.

[0037] The following is the running result of the system of the present application in three months and the comparison with the traditional method, see Table 1 Risk Path Identification Comparison Data Statistics Table: Table 1 Risk Path Identification Comparison Data Statistics Table

[0038] As can be seen in the analysis of Table 1, the present application method has a significant advantage in risk path identification. The traditional method mainly relies on rule matching and keyword filtering, and has the problems of serious missed detection and weak correlation, while the present application can effectively identify more potential risk paths by introducing an improved NodeFormer model, combining structure-guided connection and state evolution representation mechanism, especially in the contract-fund flow-invoice correlation chain. The average number of risk paths identified in three months is increased by about 142.9% compared with the traditional method, and the number of paths marked as high risk is steadily increasing, reflecting the significant improvement in the screening ability of path confidence distribution, and the higher value of decision support for auditors.

[0039] In order to further evaluate the effect of the present application in intelligent question and answer and visualization, the comparison of artificial response time and problem positioning efficiency in the audit process is counted, see Table 2 Intelligent Question and Answer and Graph Analysis Auxiliary Effect Comparison Table: Table 2 Intelligent Question and Answer and Graph Analysis Auxiliary Effect Comparison Table

[0040] From the results of Table 2, the traditional method needs to manually search contract documents, budget control details and other data sources when handling audit queries, which takes a long time and the path is not clear, while the system of the present application relies on the graph semantic question and answer structure and the dynamic graph structure, and can directly parse natural language questions into graph structured retrieval tasks, achieving second-level response. The average query time is compressed to less than 40 seconds, the problem positioning efficiency is improved by more than 75%, and the average adoption rate of the audit suggestion results automatically generated by the system in three months reaches 77.7%, reflecting the good practicality and reference value of the present application method in actual work.

[0041] The embodiment fully verifies the feasibility and superiority of the method in complex financial scenarios, not only significantly improves the risk identification ability and audit response speed, but also improves the reference of audit recommendations and the intelligent level of the whole system. By introducing a dynamic atlas structure and an improved NodeFormer model, the application realizes unified management, deep mining and efficient application of multi-source financial data.

[0042] The above describes only the preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art within the technical range disclosed by the present application, according to the technical solution and the inventive concept of the present application, equivalent replacement or change, should be covered within the protection scope of the present application.

Claims

1. A knowledge graph-based financial information management method, characterized in that, Includes the following steps: S1. Collect multi-source heterogeneous financial data from financial systems, budget systems, contract management systems, bank interfaces and invoice platforms, perform field standardization, format regularization, timestamp unification and data anonymization processing to generate an initial financial dataset; S2. Perform entity recognition and relation extraction operations on the initial financial dataset to extract financial entities and semantic relations, construct a set of triples, and form an initial financial knowledge graph. S3. Perform semantic enhancement processing on the initial financial knowledge graph, embed time tags, business tags, risk tags, responsibility tags and source tags to construct a semantically enhanced financial graph; S4. Set up a graph update scheduling mechanism to perform dynamic update operations on the semantically enhanced financial graph. The dynamic update operations include periodic updates based on time windows and local structure updates based on event triggers to generate a dynamic graph structure. S5. Input the dynamic graph structure into the improved NodeFormer model, perform the risk path identification operation, and output the risk path identification result; S6. Construct a graph semantic question-and-answer structure based on the risk path identification results, and output the structured query response results; The graph semantic question-answering structure includes natural language question parsing, graph node matching, and path retrieval operations; S7. Construct a graph visualization module, perform node highlighting, path tracing, tag filtering and risk warning operations, and generate graph analysis views and audit recommendation results.

2. The financial information management method based on knowledge graphs according to claim 1, characterized in that, Specifically, S1 is: Establish a unified data structure template, specifying field formats, numerical precision, time specifications, and character encoding rules; Data is collected from the financial system, including accounting voucher data, accounting subject data, and payment details; from the budget system, including budget item data, budget control data, and approval records; from the contract management system, including contract numbers, terms, performance status, and payment plans; from the bank interface, including account transaction records and fund flow information; and from the invoice platform, including invoice registration records, image recognition results, and verification results. Perform field mapping and format conversion processing on various types of data, unify field structure, unit standards and data types, fill in missing fields and remove invalid records; unify the time format in all data, and identify and distinguish between business time and system time; Data involving identity information, account numbers, and contract sensitive fields is anonymized, and sensitive fields are processed using masking or encryption methods to generate an initial financial dataset.

3. The financial information management method based on knowledge graphs according to claim 2, characterized in that, Specifically, S2 is: Construct a set of financial entity categories and define financial entity types; the financial entity types include contract number, budget item, invoice number, account identifier, department name, approval node, and fund flow identifier; Lexical rules and contextual structure templates are constructed based on financial entity categories. Entity recognition operations are performed based on field content, syntactic structure, and contextual position to identify field content in the initial financial dataset that has specific semantic and financial business meaning. Define a set of financial semantic relationships and set semantic relationship types; the semantic relationship types include inclusion relationship, control relationship, fund flow relationship, contractual constraint relationship, budget attribution relationship and invoice correspondence relationship; A relation extraction model is constructed, and by combining syntactic dependency analysis and entity pair context analysis, relation determination operations are performed on the identified entity groups to determine the semantic connection between entities. The identified financial entities and semantic relationships are organized in a triple format, which includes the starting entity, the relationship type, and the target entity, to generate an initial financial knowledge graph.

4. The financial information management method based on knowledge graphs according to claim 3, characterized in that, Specifically, S3 is: Define a set of tag types, which includes time tags, business tags, risk tags, responsibility tags, and source tags; Based on the attribute information and relationship structure of financial entities in the initial financial knowledge graph, the corresponding time dimension, business classification, risk identifier, organizational affiliation, and data source fields are extracted. The extracted results are then converted into structured tagged data, where the time tags include annual, quarterly, and monthly information; the business tags include expense categories, expenditure items, and project types; the risk tags include over-budget markers, frequent transaction markers, and duplicate reimbursement markers; the responsibility tags include department and responsible personnel information; and the source tags include system source identifiers and data generation method identifiers. Structured labeled data is embedded into the node attribute set and edge attribute set of the initial financial knowledge graph, and corresponding label information is attached to each node and edge in the graph to construct a semantically enhanced financial graph.

5. The financial information management method based on knowledge graphs according to claim 4, characterized in that, Specifically, S4 is: The time window update scheduling unit sets a fixed update cycle parameter and performs graph structure update operations on a regular basis according to the set daily, weekly or monthly cycle. In each cycle, the newly added financial data is parsed and processed, and the parsing results are compared with the nodes and edges in the semantically enhanced financial graph. If a new entity or new relationship is identified, node addition, edge connection or attribute supplementation operations are performed. The event trigger update scheduling unit sets event listening rules and trigger conditions, including budget execution anomalies, contract status changes, invoice cancellation records, approval chain changes, and cash flow anomalies; If an event that meets the event listening rules and trigger conditions is detected, the corresponding local graph update operation will be executed immediately. The update content includes updating node status markers, reconstructing relationship connections, and revising label content. The time window update scheduling unit and the event-triggered update scheduling unit synchronously output update operation logs, recording the timestamp, modification type, modification target, and modification source of the semantically enhanced financial graph modification, and generating a dynamic graph structure.

6. The financial information management method based on knowledge graphs according to claim 5, characterized in that, The improved NodeFormer model includes a label attention guidance module, a state evolution memory module, a global path comparison module, and a fusion output module, specifically: The structure-guided connection module constructs a node adjacency candidate matrix based on the dynamic graph structure, calculates the structural similarity weights between nodes, forms a structural attention distribution, and maps the attention distribution to a structure-guided connection strength matrix. The tag-aware guidance module extracts time tags, business tags, risk tags, responsibility tags, and source tags attached to nodes and edges, constructs a multi-dimensional tag embedding matrix, calculates the tag similarity weights between node pairs based on the tag embedding matrix, and performs an element-wise fusion operation with the tag similarity weights and the structure-guided connection strength matrix to generate a tag-aware connection matrix. The state evolution memory module constructs a node state time series, and constructs the connection changes, attribute updates and label changes of the node in multiple time windows into a state input sequence. It introduces a gated loop structure and memory unit to construct the node state evolution trajectory, and encodes the state input sequence to generate a node evolution representation vector. The global path comparison module constructs a candidate path set, extracts the node evolution representation and label-aware connection strength in the path, constructs a path feature sequence, selects low-risk financial paths as comparison samples, performs comparison learning operations based on the path feature sequence, and calculates the path offset score. The fusion output module performs weighted fusion of node evolution representation, label-aware connection strength, and path offset score to generate a path confidence distribution. Paths with confidence values ​​greater than a preset confidence threshold in the confidence distribution are selected as risk paths, and the risk path identification results are output.

7. The financial information management method based on knowledge graphs according to claim 6, characterized in that, Specifically, S6 is: Lexical analysis, dependency parsing, and semantic role labeling are performed on the input question text to extract target entities, target relationships, and query intent; a graph semantic question-answering structure is constructed to generate structured response data; The structured response data is converted into a question-and-answer output format, and the output includes the target node, path node sequence, path type, key tags and risk warning information.

8. The financial information management method based on knowledge graphs according to claim 7, characterized in that, The graph semantic question-answering structure includes natural language question parsing, graph node matching, and path retrieval operations, specifically: Based on the target entity name in the query intent, calculate the name similarity score between the target entity and the node name in the semantically enhanced financial graph, and extract the corresponding time tag, business tag, risk tag, responsibility tag and source tag, and calculate the tag consistency score; perform normalization and weighted fusion on the name similarity score and tag consistency score to generate a comprehensive matching score; Set a comprehensive matching threshold and select nodes with a comprehensive matching score greater than the comprehensive matching threshold as a candidate node set; Based on the target relationship between the candidate node set and the query intent, the associated paths containing candidate nodes are retrieved from the dynamic graph structure. The node sequence, relationship type and label content in the path are matched with the query intent to generate a path candidate set. The risk path identification results are invoked, and risky paths and non-risk paths in the path candidate set are marked with path type attributes respectively. The structured response data is organized according to the set format for the node state evolution information, label content and connection strength in the path.

9. A knowledge graph-based financial information management method according to claim 8, characterized in that, Specifically, S7 is: Based on the set of nodes and edges in the dynamic graph structure, a graphical view of the graph structure is rendered, displaying the relationships between nodes, attribute information, and label content. Perform node highlighting, path tracing, tag filtering, and risk alert operations to generate a graph analysis view and audit recommendations. The graph analysis view includes node graphics, path connections, label information, and highlighted icons; The audit recommendations include risk sources, risk points, related paths, and a list of recommendations.

10. A knowledge graph-based financial information management method according to claim 9, characterized in that, The execution node highlighting, path tracing, tag filtering, and risk warning operations are specifically as follows: Based on the risk path identification results, risk nodes and risk paths are marked. Risk nodes are highlighted and color-coded with label information. Risk paths are highlighted and path connection styles and weights are rendered. Set up a path tracing function, receive input from the starting node and the target node, retrieve the path sequence in the dynamic graph structure, display the intermediate nodes, relationship types and connection order in the path, and perform style marking on the risk nodes in the path. Set a set of tag filtering conditions, including time tags, business tags, risk tags, responsibility tags, and source tags. Keep nodes and paths that meet any tag filtering conditions visible, and hide nodes and paths that do not meet the tag filtering conditions. Based on the risk label level and path confidence distribution of nodes, risk warning operations are performed on nodes and paths that exceed the preset risk threshold. Warning markers are added to the dynamic graph structure to display the risk label type, offset score and associated path information.

Images