Enterprise financial digital intelligent management system based on knowledge graph

Abstract

This invention discloses a knowledge graph-based intelligent digital management system for enterprise finance, comprising: a data access and preprocessing module for collecting multi-source enterprise financial data and generating a unified dataset; a knowledge graph construction module for constructing an enterprise financial knowledge graph and writing financial business element entities and their relationships; an improved RoFormer extraction module for outputting extraction results and evidence pointers; a graph fusion link generation module for forming business link instances; a key-value cache compression module for implementing bit-width split storage and constant bit mapping compression to generate a compressed key-value cache; and an incremental reasoning and early warning module for updating the knowledge graph and performing incremental reasoning, penetration verification, and early warning generation. This invention combines the improved RoFormer model with efficient reasoning through improved key-value cache compression to achieve automatic, penetrating verification and explainable early warnings for financial links.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph technology, and in particular to a knowledge graph-based intelligent digital management system for enterprise finance. Background Technology

[0002] Current enterprise financial digitization is typically carried out by multiple business systems, resulting in data being stored in a scattered manner across different systems for a long time. Differences exist among these systems in terms of business object identification systems, field definitions, master data rules, data granularity, and lifecycle management, making it difficult to establish a unified identifier and consistent relationships when financial business elements flow across systems. Due to the lack of unified data standards and cross-system indexes, the business links of contracts, invoices, payments, cash flow, vouchers, and accounting entries are difficult to automatically connect. Link construction and verification often rely on manual retrieval, comparison, and information supplementation across multiple systems, leading to insufficient link penetration, unstable traceability paths, and low efficiency in reconciliation and verification. Furthermore, unstructured or semi-structured data and attachments are difficult to reliably parse and structure for storage, resulting in scattered key evidence in the business links and difficulty in forming reusable structured representations, further impacting data consistency and traceability.

[0003] Traditional financial internal control and risk control processes still rely primarily on manual review and rule verification. Risk identification, risk rating, and decision-making typically depend on personnel summarizing, verifying, attributing, and recording various types of business data. This makes the anomaly detection and handling closed loop largely limited by personnel experience, job division, and processing load, making it difficult to meet the timeliness and consistency requirements of enterprises under conditions of high-concurrency business, multiple rounds of verification, and large-scale document processing. Existing early warning methods often focus on outputting anomaly conclusions or risk markers, lacking the ability to structurally bind anomaly conclusions to business chain relationships, original voucher sources, and textual evidence location, leading to... The early warning results are difficult to form a verifiable chain of evidence and an interpretable path. During the review, it is still necessary to manually trace back the original documents and rebuild the correlation, resulting in high review costs, low location efficiency, and long processing cycles. Under the condition of inconsistent data standards and scattered data across systems, the anomaly judgment is prone to repeated triggering, missed triggering, or inconsistent interpretation, which affects the stability and traceability of internal control execution. Moreover, in scenarios that require continuous tracking of rectification, review and write-back, and audit traceability, the lack of a unified closed-loop recording and status management method can easily lead to breakpoints in the handling process and difficulties in tracing responsibility, thereby reducing the closed-loop efficiency of risk handling and management reliability.

[0004] Therefore, how to provide a knowledge graph-based intelligent digital management system for enterprise finance is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a knowledge graph-based intelligent digital management system for enterprise finance. This invention comprehensively utilizes multi-source data preprocessing and caliber alignment, financial knowledge graph modeling and link instance generation, improved RoFormer model element and relationship extraction, and key-value caching compression inference cache management technology to form a complete process from multi-source data access, unified dataset generation, knowledge graph construction, text sequence extraction and evidence pointer output, graph fusion to complete the link, incremental inference penetration verification, to early warning generation and handling write-back. The improved RoFormer model achieves unified output of extraction results and evidence pointers through decoupled spatial flow, task-state structure mapping, and sparse state isolation evolution. Key-value caching compression achieves stable incremental inference and reduced resource consumption in long text and multi-round verification scenarios through bit-width split storage and constant structure bit mapping. Compared with existing technologies, this invention can achieve automatic association penetration and traceable verification of contract, invoice, payment, and accounting links, and improves the timeliness of anomaly identification, the completeness of interpretable evidence chains, and the efficiency of closed-loop handling, facilitating enterprise-level deployment.

[0006] According to an embodiment of the present invention, a knowledge graph-based intelligent management system for enterprise financial digitalization includes the following modules: The data access preprocessing module is used to collect multi-source financial data from enterprises and complete preprocessing to generate a unified dataset; The knowledge graph construction module is used to build an enterprise financial knowledge graph based on a unified dataset and write financial business element entities and their relationships. Improve the RoFormer extraction module to parse the unified dataset into a text sequence, complete the extraction of features and relationships, and output the extraction results and evidence pointers; The knowledge graph fusion link generation module is used to write the extracted results into the knowledge graph, perform entity alignment and relationship completion, and form business link instances. The key-value cache compression module is used to perform bit-width split storage and constant bit mapping compression on the inference cache to generate a compressed key-value cache; The incremental reasoning and early warning module is used to update the knowledge graph based on business link instances and extraction results, and to perform incremental reasoning, penetration verification and early warning generation.

[0007] Optionally, modules can be integrated using the following methods: Collect multi-source financial data from enterprises, perform preprocessing on the multi-source financial data, and generate a unified dataset; A corporate financial knowledge graph is constructed based on a unified dataset, and financial business element entities are written in, along with the relationships between these entities. An improved RoFormer model is constructed, and a unified dataset is parsed to obtain text sequences. Based on decoupled spatial flow, semantic information domain and evidence information domain encoding is performed on the text sequences. Task state structure mapping is used to select task states and map field types and relation types. Sparse state isolation evolution is introduced to extract retained states and perform non-retained state isolation to obtain extraction results and evidence pointers. The extraction results and evidence pointers are written into the enterprise financial knowledge graph, and entity alignment and relationship completion are performed based on the association index to form a link instance; Based on key-value cache compression, the inference cache of the RoFormer model is processed. Bit-width split storage is used to store high-bit segments and low-bit segments. Constant-bit compression is introduced to perform constant mapping allocation, resulting in a compressed key-value cache. The enterprise financial knowledge graph is updated by updating business link instances and extraction results, the compressed key-value cache is called for incremental reasoning calculation, penetrating verification is performed, and early warning information is generated.

[0008] Optionally, the enterprise's multi-source financial data includes enterprise contract data, invoice platform data, online banking fund data, expense reimbursement data, and ERP system data.

[0009] Optionally, generating a unified dataset includes: Structured and semi-structured data are obtained from various business sources corresponding to the enterprise's multi-source financial data. Data source identifiers, business object identifiers, and time identifiers are generated for the obtained data to form a set of data to be processed. Preprocessing is performed on the dataset to be processed. Preprocessing includes standardizing field formats, units, currencies, dates and times, encoding rules, and handling null and outlier values ​​to generate a standardized dataset. The standardized dataset is processed by master data mapping, caliber alignment, and deduplication verification. The deduplication verification process includes merging records based on unique identifiers and removing duplicate records based on business object identifiers and time identifiers, thereby generating a unified dataset.

[0010] Optionally, the step of constructing an enterprise financial knowledge graph based on a unified dataset includes: A corporate financial knowledge graph is built based on a unified dataset, including a set of entity types of financial business elements, a set of entity attributes, a set of entity identifier fields, a set of entity relationship types, and a set of relationship constraint fields. Corresponding data field mapping rules are configured for each entity type and relationship type. The unified dataset is instantiated using the enterprise financial knowledge graph, which converts the records in the unified dataset into entity instances and relation instances of financial business elements. A unique entity identifier is generated for each entity instance, a unique relation identifier is generated for each relation instance, and entity attribute values ​​and relation attribute values ​​are written at the same time. Consistency checks and incremental updates are performed on entity instances and relationship instances. Consistency checks include entity duplication checks and relationship conflict checks. Incremental updates include new instance writing, modified instance updating, and invalid instance marking.

[0011] Optionally, obtaining the extraction results and evidence pointers includes: An improved RoFormer model is constructed and a task state set is configured. The task state set includes contract terms extraction task state, payment receipt extraction task state, and accounting element extraction task state. The text data in the unified dataset is parsed and processed to obtain a text sequence. By performing dual-domain parallel encoding processing on the text sequence through decoupled spatial flow, a semantic representation sequence is generated from the text sequence in the semantic information domain, and an evidence representation sequence is generated from the text sequence in the evidence information domain. The output of the evidence information domain includes the text fragment location field. Based on the task state structure mapping, structural mapping processing is performed. According to the target task state, a structural mapping path corresponding to the target task state is selected. Field type mapping processing and relation type mapping processing are performed on the semantic representation sequence according to the selected structural mapping path to obtain the field type sequence and relation type sequence. Sparse state isolation evolution is introduced to perform inter-layer state processing. In each layer, the retained state is extracted from the intermediate state and passed to the next layer. The non-retained state is confined to the current layer and blocked from being passed to the next layer, resulting in a sequence of retained states after cross-layer evolution. Extraction results and evidence pointers are generated based on the field type sequence, relation type sequence, reserved state sequence, and evidence representation sequence. The evidence pointers include the text fragment start position identifier, text fragment end position identifier, page number identifier, and line number identifier.

[0012] Optionally, forming a link instance includes: The extraction results and evidence pointers are converted into graph writing data. The graph writing data includes entity writing records, relationship writing records and evidence writing records. The evidence writing records include text fragment start position identifier, text fragment end position identifier, page number identifier and line number identifier. Based on the association index, entity alignment processing is performed on the graph data. Existing entity instances are matched according to the unique identifier, entity merging is performed, and new entity instances are created when no matching entity instance is found, and entity attribute values ​​are updated and written. The entity alignment results are used to complete the relationship in the graph data. Based on the business object identifier and time identifier, contract-invoice relationship, invoice-payment relationship, payment-accounting relationship and fund flow-payment relationship are generated. Relationship attribute values ​​and evidence writing records are written to form business link instances.

[0013] Optionally, obtaining the compressed key-value cache includes: In the process of improving the RoFormer model inference, a key-value cache is established. The key-value cache includes a set of key vectors and a set of value vectors stored in layers. A token identifier and a cache index are generated for each token in the text sequence. Bit-width split storage is used to perform split storage on the key-value cache. The numerical representation of each key vector and each value vector is divided into high-bit segments according to the bit width and written to the high-bit segment cache area, and the low-bit segments are written to the low-bit segment cache area. The high-bit segment cache area and the low-bit segment cache area are respectively associated with the token identifier. A constant bit compression structure is introduced to perform constant mapping allocation on the key-value cache after split storage. A fixed cache address range is allocated for each token identifier according to the constant mapping table, and the write addresses of the high-bit segment cache area and the low-bit segment cache area are mapped to the fixed cache address range. Cache access is performed based on a constant mapping table. The high-order segments of the key vector and the high-order segments of the value vector are read from a fixed cache address range according to the token identifier. When the preset precision trigger condition is met, the low-order segments of the key vector and the low-order segments of the value vector are read from the fixed cache address range to generate the access result of the compressed key-value cache.

[0014] Optionally, generating the early warning information includes: The enterprise financial knowledge graph is updated based on business link instances and extraction results. Updates are written to entity attributes, and additions and changes are written to entity relationships. Evidence pointers are recorded and corresponding to the extraction results. Incremental inference is performed on the text sequence by calling the compressed key-value cache. New key vectors and new value vectors are generated for new text fragments and written to the compressed key-value cache. Based on the compressed key-value cache, the field type and relation type corresponding to the new text fragments are obtained and the incremental extraction results are output. By performing penetrating verification on business link instances using the updated enterprise financial knowledge graph, early warning information is generated. The penetrating verification includes consistency verification of the executing entity, consistency verification of amount and tax, and consistency verification of duplicate documents. Verification results and link paths are generated, and early warning information includes the anomaly type, the set of entities involved, the link path, and the original evidence fragments located by the evidence pointer.

[0015] The beneficial effects of this invention are: This invention preprocesses and aligns multi-source financial data for enterprises, using a unified enterprise financial knowledge graph to represent financial business elements and their relationships. This enables the automatic construction and penetrating verification of the business chain, including contracts, invoices, payments, and accounting entries. Compared to existing methods relying on manual comparison and decentralized system queries, this invention uses an improved RoFormer model to extract textual elements and relationship clues, simultaneously outputting evidence pointers. This allows the extracted results to be stably bound to graph relationship instances, thereby improving the consistency of cross-system associations, the traceability of verification, and the efficiency of chain continuity, while reducing verification costs caused by chain breakpoints and inconsistencies.

[0016] This invention implements bit-width split storage and structure-constant bit mapping compression for the inference cache through key-value caching compression, supporting incremental inference computation in scenarios involving long texts and multi-round verification. After knowledge graph updates, it performs penetrating verification and early warning generation, achieving an interpretable early warning and handling closed loop. Compared with existing early warning results that lack evidence chains and have low handling efficiency, this invention can output early warning information containing link paths and original text evidence fragment location information, improving the timeliness of anomaly identification, interpretation consistency, and review efficiency. At the same time, it reduces inference resource consumption and improves system processing throughput and deployment stability, facilitating the implementation of enterprise-level internal control and risk control processes. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a knowledge graph-based intelligent management system for enterprise finance. Figure 2 This is a structural block diagram of a knowledge graph-based intelligent digital management method for enterprise finance proposed in this invention. Figure 3 This is a functional diagram of the improved RoFormer model, which is a knowledge graph-based intelligent digital management method for enterprise finance proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figure 1 A knowledge graph-based intelligent management system for enterprise finance digitalization includes the following modules: The data access preprocessing module is used to collect multi-source financial data from enterprises and complete preprocessing to generate a unified dataset; The knowledge graph construction module is used to build an enterprise financial knowledge graph based on a unified dataset and write financial business element entities and their relationships. Improve the RoFormer extraction module to parse the unified dataset into a text sequence, complete the extraction of features and relationships, and output the extraction results and evidence pointers; The knowledge graph fusion link generation module is used to write the extracted results into the knowledge graph, perform entity alignment and relationship completion, and form business link instances. The key-value cache compression module is used to perform bit-width split storage and constant bit mapping compression on the inference cache to generate a compressed key-value cache; The incremental reasoning and early warning module is used to update the knowledge graph based on business link instances and extraction results, and to perform incremental reasoning, penetration verification and early warning generation.

[0020] refer to Figure 2 and Figure 3 A knowledge graph-based digital intelligent management method for enterprise finance includes: Collect multi-source financial data from enterprises, perform preprocessing on the multi-source financial data, and generate a unified dataset; A corporate financial knowledge graph is constructed based on a unified dataset, and financial business element entities are written in, along with the relationships between these entities. An improved RoFormer model is constructed, and a unified dataset is parsed to obtain text sequences. Based on decoupled spatial flow, semantic information domain and evidence information domain encoding is performed on the text sequences. Task state structure mapping is used to select task states and map field types and relation types. Sparse state isolation evolution is introduced to extract retained states and perform non-retained state isolation to obtain extraction results and evidence pointers. The extraction results and evidence pointers are written into the enterprise financial knowledge graph, and entity alignment and relationship completion are performed based on the association index to form a link instance; Based on key-value cache compression, the inference cache of the RoFormer model is processed. Bit-width split storage is used to store high-bit segments and low-bit segments. Constant-bit compression is introduced to perform constant mapping allocation, resulting in a compressed key-value cache. The enterprise financial knowledge graph is updated by updating business link instances and extraction results, the compressed key-value cache is called for incremental reasoning calculation, penetrating verification is performed, and early warning information is generated.

[0021] In this embodiment, the enterprise's multi-source financial data includes enterprise contract data, invoice platform data, online banking fund data, expense reimbursement data, and ERP system data.

[0022] In this embodiment, generating a unified dataset includes: Structured and semi-structured data are obtained from various business sources corresponding to the enterprise's multi-source financial data. Data source identifiers, business object identifiers, and time identifiers are generated for the obtained data to form a set of data to be processed. Preprocessing is performed on the dataset to be processed. Preprocessing includes standardizing field formats, units, currencies, dates and times, encoding rules, and handling null and outlier values ​​to generate a standardized dataset. The standardized dataset is processed by master data mapping, caliber alignment, and deduplication verification. The deduplication verification process includes merging records based on unique identifiers and removing duplicate records based on business object identifiers and time identifiers, thereby generating a unified dataset.

[0023] In this embodiment, the step of constructing an enterprise financial knowledge graph based on a unified dataset includes: A corporate financial knowledge graph is built based on a unified dataset, including a set of entity types for financial business elements, a set of entity attributes, a set of entity identifier fields, a set of entity relationship types, and a set of relationship constraint fields. Corresponding data field mapping rules are configured for each entity type and relationship type. Specifically, the construction of the corporate financial knowledge graph based on the unified dataset involves: Based on the unified dataset, fields are sorted, master data is merged and aligned with definitions. Stable business objects are identified to form a set of financial business element entity types. For each entity type, a set of entity attributes and a set of entity identifier fields are defined. Based on the association key field, number reference field and time link field in the unified dataset, the connection method between business objects is determined to form a set of entity relationship types. For each relationship type, a set of relationship attribute fields and a set of relationship constraint fields are defined. Data field mapping rules define the correspondence between source fields in the unified dataset and target fields in the knowledge graph. The rules include the mapping between source field names and target field names, the encoding conversion between source field values ​​and target field values, the conversion method between units and currencies, the unification method of date and time formats, the handling method of null values ​​and outliers, the generation method of primary keys and foreign keys, and the value selection method when there are multiple source conflicts. Based on the data field mapping rules, the unified dataset is written into entity attribute fields, entity identifier fields and relationship attribute fields to form an enterprise financial knowledge graph. The unified dataset is instantiated using the enterprise financial knowledge graph, which converts the records in the unified dataset into entity instances and relation instances of financial business elements. A unique entity identifier is generated for each entity instance, a unique relation identifier is generated for each relation instance, and entity attribute values ​​and relation attribute values ​​are written at the same time. Consistency checks and incremental updates are performed on entity instances and relationship instances. Consistency checks include entity duplication checks and relationship conflict checks. Incremental updates include new instance writes, modified instance updates, and invalid instance marking. Specifically, the consistency checks and incremental updates for entity instances and relationship instances are performed as follows: When performing entity duplication validation on an entity instance, a unique entity identifier key is generated based on the entity identifier field set. Entity instances with the same unique entity identifier key in the knowledge graph are retrieved and identified as duplicate entities. Duplicate entities are merged into the same entity instance. During merging, attribute consistency is compared on the entity attributes. If the attribute values ​​are the same, the values ​​are retained. If the attribute values ​​are different, the written value is determined according to the conflict value selection method in the data field mapping rules, and the source identifier and time identifier are recorded. When performing relation conflict verification on a relation instance, a unique relation key is generated based on the relation type identifier, the unique identifier key of the relation starting entity, the unique identifier key of the relation ending entity, and the relation constraint fields. Relation instances with the same unique relation key in the knowledge graph are retrieved and determined to be conflicting relations. The relation attributes of the conflicting relations are compared item by item according to the relation constraint fields. If the values ​​of the relation constraint fields are consistent, the relation instance is retained. If the values ​​of the relation constraint fields are inconsistent, the conflict resolution method in the field mapping rules determines whether to retain the relation instance and records the conflict fields. In incremental update processing, new instances are written for entity unique identifier keys and relation unique keys that do not exist in the knowledge graph in the unified dataset. For instances that exist but whose attribute or relation attributes are inconsistent with the written values, change instance updates are performed and the change time identifier is recorded. For instances that have not appeared in the continuous update cycle or whose business status is marked as invalid, invalid instance marking is performed and historical versions are retained.

[0024] In this embodiment, obtaining the extraction result and the evidence pointer includes: An improved RoFormer model is constructed and a task state set is configured. This set includes task states for extracting contract terms, payment receipts, and accounting elements. Text data from a unified dataset is parsed to obtain a text sequence, where: An improved RoFormer model is constructed, specifically by setting up two parallel representation paths—a semantic information domain and an evidence information domain—on the original RoFormer encoding framework. The semantic information domain is used to generate semantic representations of fields, and the evidence information domain is used to generate locatable evidence representations. Locating fields are retained at the output. An internally pre-built task-state driven structure mapping table maps task state identifiers to a fixed label space and output field set, and maps the output of field types and relation types to entity attribute fields and relation types of the knowledge graph. A sparse state isolation evolution structure is adopted, defining a set of retained states and propagating only retained states between layers. Non-retained states participate in the calculation in the current layer but do not enter the next layer, forming a key state sequence for cross-layer evolution. Configure the task state set, specifically: determine the task boundaries of the task state set based on the entity type set, relation type set, and field mapping rules of the enterprise financial knowledge graph; configure a task state identifier, input text range identifier, field type set, relation type set, and evidence pointer output field set for each task state; configure the correspondence between task state and structure mapping path; extract the clause attribute fields and clause relationship types of the contract entity bound to the task state from the contract clauses; extract the association fields and relationship types of the payment receipt and the payment entity and the fund flow entity bound to the task state from the payment receipt and the fund flow entity; extract the account, period, and amount fields and relationship types of the voucher entity and the accounting element bound to the task state from the accounting element. The text data in the unified dataset is parsed to obtain a text sequence. Specifically, the text data in the unified dataset is restored according to the data source identifier, and page number identifiers, line number identifiers and paragraph boundary identifiers are generated. The text is cleaned and normalized, control characters are removed and the formats of numbers, amounts and dates are unified. The text is segmented into text segments according to the paragraph boundary identifiers and line number identifiers and a token sequence is generated. Each token is appended with a page number identifier, line number identifier and segment start and end position identifier to form an input sequence with positioning fields, thus obtaining the text sequence. A decoupled spatial flow approach is used to perform dual-domain parallel encoding on the text sequence. A semantic representation sequence is generated in the semantic information domain, and an evidence representation sequence is generated in the evidence information domain. The output of the evidence information domain includes a text fragment location field, where: In the semantic information domain, a semantic representation sequence is generated from the text sequence. Specifically, the tokens in the text sequence are mapped to semantic embedding vectors and fed into the encoding structure of the improved RoFormer model. The semantic embedding vector of each token is linearly mapped to obtain the query vector, key vector, and value vector. The attention score is calculated based on the query vector and key vector. The value vectors are weighted and converged to obtain the context representation. The context representation is transformed by the feedforward network and connected to the residual to output the semantic state of the current layer. The semantic state is iteratively updated in the encoding. The output layer is arranged in the order of the tokens to form a semantic representation sequence. In the evidence information domain, an evidence representation sequence is generated from the text sequence. Specifically, an evidence embedding vector is generated for each token in the text sequence. The evidence embedding vector is encoded by page number identifier, line number identifier, paragraph identifier, text segment start and end position identifier and token sequence identifier. It serves as the input representation of the evidence information domain. The evidence information domain updates the evidence embedding vector between layers along independent encoding paths. The update process keeps the correspondence between token and positioning field unchanged. The evidence embedding vector is aggregated in context to obtain the evidence state. The output layer arranges the evidence states of each token in token order and carries the page number identifier, line number identifier and text segment start and end position identifier to form an evidence representation sequence. Based on the task-state structure mapping, structure mapping processing is performed. A structure mapping path corresponding to the target task state is selected. Field type mapping and relation type mapping processing are then performed on the semantic representation sequence according to the selected structure mapping path, resulting in a field type sequence and a relation type sequence, where: The structure mapping path corresponding to the target task state is selected based on the target task state. Specifically, the structure mapping parameters corresponding to the target task state identifier are read. The structure mapping parameters include the input text range identifier, field type set identifier, relation type set identifier, output label space identifier, and evidence pointer output field set identifier. Based on the structure mapping parameters, the output channel matching the target task state is activated in the output layer. The output channel is bound to the field type set and relation type set corresponding to the target task state to form the structure mapping path corresponding to the target task state and the output range and label set of the path are limited. The semantic representation sequence is processed by field type mapping and relation type mapping according to the selected structure mapping path. Specifically, the semantic representation vector of each token in the semantic representation sequence is input into the field type mapping layer of the structure mapping path. The semantic representation vector is mapped to the label index corresponding to the field type set and the field type identifier is output. The field type sequence is obtained by arranging the tokens in order. Paired tokens in the semantic representation sequence are input into the relation type mapping layer of the structure mapping path. The paired representations are mapped to the label index corresponding to the relation type set and the relation type identifier is output. At the same time, the relation start token identifier and relation end token identifier are recorded. The relation type sequence is obtained by arranging the relation instances in order. Sparse state isolation evolution is introduced to perform inter-layer state processing. At each layer, retained states are extracted from intermediate states and passed to the next layer. Non-retained states are confined to the current layer and their transmission to the next layer is blocked, resulting in a sequence of retained states after cross-layer evolution. The inter-layer state processing specifically involves: Using the intermediate state of each layer's output as input, the intermediate states are filtered and marked according to the definition of retained states. The token state corresponding to the retained state is written into the inter-layer transfer buffer and used as the source of the input state for the next layer. The token state corresponding to the non-retained state is retained in the current layer and participates in the current layer's calculation but is not written into the inter-layer transfer buffer. When calculating the next layer, the retained states from the inter-layer transfer buffer and the basic input of the layer represent the input state of the layer. The filtering, marking and transfer process is repeated until the last layer. The retained states obtained from the final transfer of each layer are collected in token order to form a sequence of retained states after cross-layer evolution. Extraction results and evidence pointers are generated based on the field type sequence, relation type sequence, reserved state sequence, and evidence representation sequence. The evidence pointers include the text fragment start position identifier, text fragment end position identifier, page number identifier, and line number identifier.

[0025] In this embodiment, forming a link instance includes: The extraction results and evidence pointers are converted into graph writing data. This graph writing data includes entity writing records, relationship writing records, and evidence writing records. Each evidence writing record contains a text fragment start position identifier, a text fragment end position identifier, a page number identifier, and a line number identifier. Specifically, the conversion of the extraction results and evidence pointers into graph writing data involves: The extraction results are categorized according to the field type and relation type corresponding to the task state. The field value, field type identifier, and corresponding token range are used as candidate entity attributes. Entity write records are generated according to the field mapping rules. The entity write record contains the entity type identifier, entity unique identifier key, attribute field name, and attribute field value. Relationship threads are categorized according to the relation type identifier and relation endpoint token range. Relationship write records are generated according to the relation mapping rules. The relation write record contains the relation type identifier, relation start entity unique identifier key, relation end entity unique identifier key, and relation attribute field value. The start position identifier, end position identifier, page number identifier, and line number identifier of the text fragment in the evidence pointer are bound to the corresponding entity writing record or relationship writing record to generate an evidence writing record. The evidence writing record contains the evidence identifier, the bound object identifier, the page number identifier, the line number identifier, and the fragment start and end position identifier, forming the graph writing data. Based on the association index, entity alignment processing is performed on the graph data. Existing entity instances are matched according to the unique identifier, entity merging is performed, and new entity instances are created when no matching entity instance is found, and entity attribute values ​​are updated and written. The entity alignment results are used to complete the relationship in the graph data. Based on the business object identifier and time identifier, contract-invoice relationship, invoice-payment relationship, payment-accounting relationship and fund flow-payment relationship are generated. Relationship attribute values ​​and evidence writing records are written to form business link instances.

[0026] In this embodiment, obtaining the compressed key-value cache includes: In the process of improving the RoFormer model inference, a key-value cache is established. The key-value cache includes a set of key vectors and a set of value vectors stored in layers. A token identifier and a cache index are generated for each token in the text sequence. Bit-width split storage is used to perform split storage on the key-value cache. The numerical representation of each key vector and each value vector is divided into high-order segments according to the bit width and written to the high-order segment cache area, while the low-order segments are written to the low-order segment cache area. The high-order segment cache area and the low-order segment cache area are respectively associated with the token identifier. Specifically, the bit-width division into high-order segments and writing to the high-order segment cache area, and the low-order segment writing to the low-order segment cache area are as follows: For each token, the key vector and value vector are first determined to have a numerical representation format and bit width allocation rules. The bit width allocation rules include the total bit width, sign bit position, exponent range, and mantissa range. Each numerical representation is then split into a high-order segment and a low-order segment according to the bit width allocation rules. The high-order segment consists of the sign bit, exponent, and the high-order part of the mantissa, while the low-order segment consists of the low-order part of the mantissa. The high-order segment is packaged into a continuous bit sequence according to a preset high-order segment storage format and written to the corresponding address in the high-order segment cache. Similarly, the low-order segment is packaged into a continuous bit sequence according to a preset low-order segment storage format and written to the corresponding address in the low-order segment cache. Both the high-order segment cache write address and the low-order segment cache write address are indexed to the token, enabling the token to locate the corresponding high-order segment data and low-order segment data respectively. The default high-order segment storage format is used to store the high-order segment bit sequence in the high-order segment cache area. The format definition includes the high-order segment field arrangement order, field length and alignment method. The high-order segment field arrangement order specifies the high-order part of the sign field, exponent field and mantissa field. The high-order segment field length is determined by the bit width allocation rule. The alignment method specifies that the high-order segment is aligned to words or blocks with a fixed word length as the basic storage unit. A fixed number of high-order segment storage units are allocated to each token for random access. The default low-order segment storage format is used to store the low-order segment bit sequence in the low-order segment buffer. The format definition includes the low-order segment field arrangement order, field length and packing method. The low-order segment field length is determined by the bit width allocation rule. The packing method specifies that the low-order segment is compactly packed and stored according to a fixed block size and the offset within the block is recorded, so that each token can locate the corresponding low-order segment data through the block address and the offset within the block. A constant-bit compression structure is introduced to perform constant mapping allocation on the key-value cache after split storage. A fixed cache address range is allocated to each token identifier according to the constant mapping table, mapping the write addresses of the high-bit and low-bit cache areas to the fixed cache address range. Specifically, the constant mapping allocation for the key-value cache after split storage is as follows: Constant mapping allocation is based on a constant mapping table, which establishes a correspondence between token identifiers and cache address ranges. The cache address range consists of a high-order segment address range and a low-order segment address range, and the length of the address range is fixed. For each token identifier, the constant mapping table is queried to obtain the corresponding fixed cache address range. The high-order segment of the key vector and the high-order segment of the value vector corresponding to the token identifier are written into the high-order segment address range of the high-order segment cache area, and the low-order segment of the key vector and the low-order segment of the value vector corresponding to the token identifier are written into the low-order segment address range of the low-order segment cache area. When a new token identifier is added to the text sequence, a new fixed cache address range is allocated according to the constant mapping table and the new high-order segment and the new low-order segment are written. When a token identifier appears repeatedly, the fixed cache address range already allocated to the token identifier is reused and overwritten. Cache access is performed based on a constant mapping table. The high-order segments of the key vector and value vector are read from a fixed cache address range according to the token identifier. When a preset precision trigger condition is met, the low-order segments of the key vector and value vector are read from the fixed cache address range, generating the access result of the compressed key-value cache. The constant mapping table is specifically an address allocation and retrieval table. The table entries use the token identifier as the index key. The table entry content includes the starting address of the high-order address range, the length of the high-order address range, the starting address of the low-order address range, the length of the low-order address range, and a layer identifier field. The constant mapping table is generated during the initialization of the key-value cache according to the maximum context length and the fixed storage unit length of each token. The token in the text sequence occupies a unique table entry in the constant mapping table according to the token identifier. During cache access, the corresponding high-order address range and low-order address range are directly located through the token identifier and the reading is completed. The preset precision trigger conditions are as follows: the precision trigger conditions include the judgment condition that the difference of the amount field falls into the range, the judgment condition that the ticket matching verification result is in an uncertain state, the judgment condition that the subject identifier matching score falls into the range, the judgment condition that the same token has multiple candidate field types in the field type mapping output, and the judgment condition that multiple candidate relation types appear in the relation type mapping output. When any judgment condition is met, the low-order segment address range is read and combined with the already read high-order segment to obtain the complete key vector and value vector. When the judgment condition is not met, only the high-order segment is used to generate the access result of the compressed key-value cache.

[0027] In this embodiment, generating the early warning information includes: The enterprise financial knowledge graph is updated based on business link instances and extraction results. Updates are written to entity attributes, and additions and changes are written to entity relationships. Evidence pointers are recorded and corresponding to the extraction results. Incremental inference is performed on the text sequence by calling the compressed key-value cache. New key vectors and new value vectors are generated for new text fragments and written to the compressed key-value cache. Based on the compressed key-value cache, the field type and relation type corresponding to the new text fragments are obtained and the incremental extraction results are output. By performing penetrating verification on business link instances using the updated enterprise financial knowledge graph, early warning information is generated. This penetrating verification includes consistency checks on the executing entity, amount and tax, and duplicate document consistency. Verification results and link paths are generated, and early warning information includes the anomaly type, the set of entities involved, the link path, and original evidence fragments located by evidence pointers. Specifically, the penetrating verification involves: The penetrating verification process extracts the main field and amount field along the business link instances to form a verification field set, and compares them item by item. The main consistency verification is based on the equality of the main unique identifier key. When the main unique identifier key is missing, it is based on the consistency of any two of the three standardized results: the main name, the unified social credit code, and the payment account. The amount and tax amount consistency verification is based on the difference in amount not exceeding the threshold. The threshold for absolute difference in amount is five, and the threshold for relative difference in amount is 0.5%. The threshold for absolute difference in tax amount is one, and the threshold for relative difference in tax amount is 0.5%. The duplicate document consistency verification uses the combination of invoice code and invoice number, payment receipt number, fund flow number, or voucher number as the deduplication key. If multiple link instances corresponding to the same deduplication key are found and the amount difference does not exceed the absolute difference threshold, it is determined to be duplicate. If any verification fails, an early warning message is generated. The early warning message is written with the original evidence fragments that include the exception type, the set of entities involved, the link path, and the evidence pointer location.

[0028] Example 1: To verify the feasibility of this invention in practice, it was applied to a financial shared service scenario in a group enterprise. Contract management, invoice management, fund payment, expense reimbursement, and ERP operated independently. The business object codes and field definitions were inconsistent for a long time, resulting in frequent breaks in the contract, invoice, payment, and accounting links. Financial personnel had to repeatedly search and verify across multiple systems to confirm the upstream and downstream relationships of the same business transaction. Internal control and risk control relied on manual sampling and experience-based judgment. Anomaly warnings often only provided conclusions without verifiable evidence. During verification, it was still necessary to go back to the original contract terms, invoice elements, receipt summaries, and attachment content for segment-by-segment comparison. Anomaly identification was lagging, the closed-loop processing cycle was long, the audit and evidence collection costs were high, and the consistency of the results was difficult to guarantee.

[0029] After deploying the system of this invention, multi-source financial business data enters the data access preprocessing module to complete field standardization, master data mapping, caliber alignment, and deduplication verification, generating a unified dataset and writing it into the knowledge graph construction module to form an enterprise financial knowledge graph. Contract text, invoice text, receipt text, and attachment text enter the improved RoFormer extraction module to be parsed into text sequences, outputting elements and relationship clues in the semantic information domain, outputting evidence pointers in the evidence information domain, and writing them into the graph. The graph fusion link generation module completes entity alignment and relationship completion, forming a penetrable business link instance. In the reasoning stage, the key-value cache compression uses bit-width split storage and structure constant bit mapping compression to support incremental reasoning of long texts and multi-round verification. The incremental reasoning early warning module performs penetrating verification and generates an early warning after the graph is updated. The early warning also carries the link path and the location information of the original text evidence fragment. After the processing is completed, the graph is written back to form a closed loop trace.

[0030] During trial runs across multiple business cycles, the system integrated data from various financial business systems and operated continuously, processing a large volume of contract texts, invoice data, payment receipts and fund flows, as well as vouchers and accounting records. Simultaneously, it formed a large-scale entity and relationship network within the knowledge graph. The level of automated business process connectivity significantly improved, with a marked reduction in verification time for individual business transactions. Daily automated verification processing capabilities improved and remained stable. The system continuously generated risk warnings and provided evidence pointers that could directly pinpoint the original text location for most warnings, enabling reviewers to quickly trace back to the original text fragments corresponding to contract terms, invoice elements, and receipt summaries. Warning review time decreased significantly, anomaly identification became more timely, and the closed-loop processing efficiency was higher. In batch processing scenarios involving long texts and multiple rounds of verification, the resource consumption on the inference side was effectively controlled, and overall throughput remained stable, meeting enterprise-level deployment and operational requirements.

[0031] Table 1. Comparison of the overall performance of financial link penetration and risk control early warning

[0032] As can be seen from Table 1, in terms of link connectivity and verification efficiency, the link connectivity rate of this invention reaches 94.7%, which is significantly better than rule penetration (62.4%) and baseline extraction (86.8%), indicating that cross-system links are easier to connect automatically. The verification time is 115 seconds / transaction, which is significantly reduced compared to rule penetration (980 seconds / transaction) and baseline extraction (310 seconds / transaction), demonstrating that the verification speed is faster and the processing is more efficient after the link instance is formed.

[0033] In terms of early warning effectiveness, the early warning hit rate of this invention is 91.5%, which is higher than the rule penetration rate of 71.2% and the baseline extraction rate of 84.1%, indicating more accurate anomaly identification; the false alarm rate is 6.8%, which is lower than the rule penetration rate of 18.7% and the baseline extraction rate of 11.2%, indicating less interference in early warning; the evidence traceability rate is 93.2%, which is significantly higher than the map without evidence rate of 55.4%, indicating that the early warning results can stably bind the original evidence fragments and link paths.

[0034] In terms of review, closure, and resource consumption, the review time of this invention is 170 seconds per item, which is better than the 360 ​​seconds per item for no evidence in the graph; the closure time is 9.4 hours per item, which is lower than the baseline extraction time of 14.6 hours per item; the peak inference memory consumption is 8.3GB, which is lower than the baseline extraction time of 14.2GB. This shows that while improving the efficiency of review and processing, it maintains low resource consumption and is suitable for batch verification and continuous operation scenarios.

[0035] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A knowledge graph-based enterprise financial digital intelligent management system, characterized in that, The method comprises the following steps: A data access preprocessing module is used to collect enterprise financial multi-source data, complete preprocessing to generate a unified data set; A knowledge graph construction module is used to construct an enterprise financial knowledge graph according to the unified data set, and write in financial business element entities and associated relationships; An improved RoFormer extraction module is used to parse the unified data set into a text sequence, complete element and relationship extraction, and output extraction results and evidence pointers; A graph fusion link generation module is used to write the extraction results into the knowledge graph, perform entity alignment and relationship completion, and form business link instances; A key-value cache compression module is used to implement bit width split storage and constant bit mapping compression on the reasoning cache, and generate compressed key-value cache; An incremental reasoning early warning module is used to update the knowledge graph based on the business link instances and the extraction results, perform incremental reasoning, penetration verification and early warning generation. 2.The knowledge graph-based enterprise financial digital intelligent management method according to claim 1, applied to the knowledge graph-based enterprise financial digital intelligent management system of claim 1, characterized in that, It comprises: Collecting enterprise financial multi-source data, preprocessing the enterprise financial multi-source data, and generating a unified data set; According to the unified data set, an enterprise financial knowledge graph is constructed, and financial business element entities are written in, and the associated relationships between the financial business element entities are written in; An improved RoFormer model is constructed, the unified data set is parsed to obtain a text sequence, semantic information domain and evidence information domain coding are performed on the text sequence based on decoupled space flow, task state selection, field type and relationship type mapping are performed by adopting task state structure mapping, reserved state extraction is performed by introducing sparse state isolation evolution, non-reserved state isolation is performed, and extraction results and evidence pointers are obtained; The extraction results and the evidence pointers are written into the enterprise financial knowledge graph, entity alignment and relationship completion are performed based on the associated index, and link instances are formed; Based on key-value cache compression, the reasoning cache of the RoFormer model is processed, high bit segment and low bit segment storage are performed by adopting bit width split storage, constant mapping allocation is performed by introducing structure constant compression, and compressed key-value cache is obtained; The enterprise financial knowledge graph is updated through the business link instances and the extraction results, the compressed key-value cache is called to perform incremental reasoning calculation, penetration verification is performed, and early warning information is generated. 3.The knowledge graph-based enterprise financial digital intelligent management method of claim 2, characterized in that, The enterprise financial multi-source data comprises enterprise contract data, invoice platform data, online banking fund data, expense reimbursement data and ERP system data.

4. The knowledge graph-based enterprise financial digital intelligent management method according to claim 2, characterized in that, The generation of the unified data set comprises: Obtaining structured data and semi-structured data from each business corresponding to the enterprise financial multi-source data, generating data source identifiers, business object identifiers and time identifiers for the obtained data, and forming a to-be-processed data set; The to-be-processed data set is preprocessed, the preprocessing comprises field format uniform processing, unit uniform processing, currency uniform processing, date and time uniform processing, coding rule uniform processing and null value and abnormal value processing, and a standardized data set is generated; The standardized data set is subjected to main data mapping processing, caliber alignment processing and de-duplication verification processing, the de-duplication verification processing comprises record merging processing based on unique identifiers and repeated record elimination processing based on business object identifiers and time identifiers, and a unified data set is generated. 5.The knowledge graph-based enterprise financial digital intelligent management method according to claim 2, characterized in that, The construction of the enterprise financial knowledge graph according to the unified data set comprises: According to the unified data set, an enterprise financial knowledge graph is established, including a set of financial business element entity types, a set of entity attributes, a set of entity identification fields, a set of entity relationship types, and a set of relationship constraint fields, and a corresponding data field mapping rule is configured for each entity type and relationship type; Through the enterprise financial knowledge graph, instantiation processing is performed on the unified data set, and records in the unified data set are converted into financial business element entity instances and relationship instances, an entity unique identifier is generated for each entity instance, a relationship unique identifier is generated for each relationship instance, and entity attribute values and relationship attribute values are written simultaneously; Consistency checking processing and incremental updating processing are performed on the entity instances and relationship instances, the consistency checking processing includes entity duplication checking and relationship conflict checking, and the incremental updating processing includes new instance writing processing, changed instance updating processing, and invalid instance marking processing. 6.The knowledge graph-based enterprise financial digital intelligent management method according to claim 2, characterized in that, The obtained extraction result and evidence pointer include: An improved RoFormer model is constructed and a task state set is configured, the task state set includes a contract clause extraction task state, a payment receipt extraction task state, and an entry element extraction task state, and text data in the unified data set is parsed to obtain a text sequence; Through decoupled space flow, double-domain parallel coding processing is performed on the text sequence, semantic representation sequences are generated for the text sequence in the semantic information domain, and evidence representation sequences are generated for the text sequence in the evidence information domain, the evidence information domain outputs a text segment positioning field; Based on the structure mapping processing of the task state structure, a structure mapping path corresponding to the target task state is selected according to the target task state, field type mapping processing and relationship type mapping processing are performed on the semantic representation sequences according to the selected structure mapping path, and field type sequences and relationship type sequences are obtained; Interlayer state processing is performed by introducing sparse state isolation evolution, reserved states are extracted from the intermediate state at each layer and passed to the next layer, non-reserved states are limited in the current layer and blocked from passing to the next layer, and a reserved state sequence after cross-layer evolution is obtained; According to the field type sequence, the relationship type sequence, the reserved state sequence, and the evidence representation sequence, an extraction result and an evidence pointer are generated, the evidence pointer includes a text segment start position identifier, a text segment end position identifier, a page number identifier, and a line number identifier.

7. The knowledge graph-based enterprise financial digital intelligent management method according to claim 2, characterized in that, The link instance is formed, including: The extraction result and the evidence pointer are converted into graph writing data, the graph writing data includes entity writing records, relationship writing records, and evidence writing records, and the evidence writing records include text segment start position identifiers, text segment end position identifiers, page number identifiers, and line number identifiers; Based on the association index, entity alignment processing is performed on the graph writing data, existing entity instances are matched according to the unique identifier, entity merging is performed, and when there is no matching entity instance, a new entity instance is created, and entity attribute values are updated and written; Through the entity alignment result, relationship completion is performed on the graph writing data, contract-invoice relationships, invoice-payment relationships, payment-accounting relationships, and fund flow-payment relationships are generated according to business object identifiers and time identifiers, relationship attribute values and evidence writing records are written, and business link instances are formed. 8.The knowledge graph-based enterprise financial digital intelligent management method of claim 2, wherein, The obtained compressed key-value cache includes: In the process of improving the RoFormer model inference, a key-value cache is established. The key-value cache includes a set of key vectors and a set of value vectors stored in layers. A token identifier and a cache index are generated for each token in the text sequence. Bit-width split storage is used to perform split storage on the key-value cache. The numerical representation of each key vector and each value vector is divided into high-bit segments according to the bit width and written to the high-bit segment cache area, and the low-bit segments are written to the low-bit segment cache area. The high-bit segment cache area and the low-bit segment cache area are respectively associated with the token identifier. A constant bit compression structure is introduced to perform constant mapping allocation on the key-value cache after split storage. A fixed cache address range is allocated for each token identifier according to the constant mapping table, and the write addresses of the high-bit segment cache area and the low-bit segment cache area are mapped to the fixed cache address range. Cache access is performed based on a constant mapping table. The high-order segments of the key vector and the high-order segments of the value vector are read from a fixed cache address range according to the token identifier. When the preset precision trigger condition is met, the low-order segments of the key vector and the low-order segments of the value vector are read from the fixed cache address range to generate the access result of the compressed key-value cache. 9.The knowledge graph-based enterprise financial digital intelligent management method of claim 2, wherein, The generation of early warning information includes: The enterprise financial knowledge graph is updated based on business link instances and extraction results. Updates are written to entity attributes, and additions and changes are written to entity relationships. Evidence pointers are recorded and corresponding to the extraction results. Incremental inference is performed on the text sequence by calling the compressed key-value cache. New key vectors and new value vectors are generated for new text fragments and written to the compressed key-value cache. Based on the compressed key-value cache, the field type and relation type corresponding to the new text fragments are obtained and the incremental extraction results are output. By performing penetrating verification on business link instances using the updated enterprise financial knowledge graph, early warning information is generated. The penetrating verification includes consistency verification of the executing entity, consistency verification of amount and tax, and consistency verification of duplicate documents. Verification results and link paths are generated, and early warning information includes the anomaly type, the set of entities involved, the link path, and the original evidence fragments located by the evidence pointer.

Images