A financial task allocation method and system based on accounting knowledge graph updating

Abstract

This application provides a method and system for financial task allocation based on accounting knowledge graph updates. It acquires business form data and uses a pre-trained information extraction model to parse the data, identifying financial entities, financial relationships, and financial attribute information. The identified financial entities and relationships are converted into feature vectors, and the similarity between these feature vectors and existing node feature vectors in a pre-constructed accounting knowledge graph is calculated as a first matching degree. Based on this first matching degree, the accounting knowledge graph is dynamically updated. According to the updated knowledge graph, a second matching degree is recalculated between the financial tasks to be allocated and the nodes of each task execution entity. A dynamically adjusted financial task allocation scheme is then generated based on this second matching degree. Using this scheme, the dynamic collaborative evolution of the knowledge graph and task allocation can be achieved, thereby improving the system's ability to autonomously adapt to changes in financial business.

Description

Technical Field

[0001] This application relates to the field of knowledge graph technology, and more specifically, to a method and system for allocating financial tasks based on updating an accounting knowledge graph. Background Technology

[0002] In the fields of financial management and enterprise resource scheduling, knowledge graph technology is widely used in task allocation. By structuring financial entities, financial relationships and attribute information into graph nodes and edges, and combining graph embedding or feature vector representation, intelligent matching between tasks and execution entities can be achieved. This approach can not only express the semantic relationships of business data, but also assist task allocation decisions through similarity calculation between nodes, improve the rationality and traceability of allocation, and provide technical support for the automation and intelligence of financial business processes.

[0003] However, most existing knowledge graph-based task allocation methods rely on static graphs or fixed matching rules, which cannot cope with the challenges of frequent changes in business data and diverse financial task requirements. In practical applications, newly emerging financial entities or relationships often cannot be mapped to existing graph nodes in a timely manner, leading to a decrease in matching accuracy. In addition, node attribute updates are lagging, and the weights of associated edges cannot reflect the latest business status, further affecting the scientific nature of task allocation and execution effectiveness. This static and lagging allocation mechanism limits the system's adaptability to dynamic business scenarios. Therefore, how to achieve dynamic collaborative evolution between the graph and task allocation, thereby improving the system's autonomous adaptability to changes in financial business, has become a challenge for the industry. Summary of the Invention

[0004] This application provides a financial task allocation method and system based on accounting knowledge graph updates, which can realize the dynamic collaborative evolution of the graph and task allocation, thereby improving the system's ability to autonomously adapt to changes in financial business.

[0005] Firstly, this application provides a method for allocating financial tasks based on updating an accounting knowledge graph, comprising the following steps: Obtain business form data, and use a pre-trained information extraction model to parse the business form data in order to identify the financial entities, financial relationships and financial attribute information therein; The identified financial entities and financial relationship information are transformed into feature vectors, and the similarity between these feature vectors and the feature vectors of existing nodes in the pre-constructed accounting knowledge graph is calculated as the first matching degree. Based on the first matching degree, the accounting knowledge graph is dynamically updated: When the first matching degree is lower than the first preset threshold, a new corresponding node is created for the financial entity or financial relationship in the accounting knowledge graph, and an association edge is established between it and the relevant node. When the first matching degree is not lower than the first preset threshold but lower than the second preset threshold, the attribute information of the corresponding existing nodes in the accounting knowledge graph is supplemented, and the weight values ​​of the relevant associated edges are recalculated based on the supplemented graph structure. Based on the updated accounting knowledge graph, the second matching degree between the financial tasks to be assigned and the nodes of each task execution entity is recalculated, and a dynamically adjusted financial task allocation scheme is generated based on the second matching degree.

[0006] Preferably, the process of using a pre-trained information extraction model to parse the business form data to identify the financial entities, financial relationships, and financial attribute information specifically includes: Convert the acquired business form data into a unified text data format; Text data is input into a pre-trained information extraction model, which sequentially performs named entity recognition to locate financial entities, performs relation extraction to determine the financial relationships between the financial entities, and performs attribute extraction to obtain the financial attribute information of the financial entities. The identified financial entities, financial relationships, and financial attribute information will be output in a structured form.

[0007] Preferably, converting the identified financial entity and financial relationship information into feature vectors specifically includes: The identified financial entities are mapped into entity semantic vectors using a pre-trained word vector model; The identified financial relationships are mapped into relational semantic vectors through a pre-trained relational semantic model. The entity semantic vectors corresponding to the two financial entities connected by the financial relationship are combined to generate relational feature vectors. The feature vector is generated by concatenating all entity semantic vectors with the relation feature vectors.

[0008] Preferably, the similarity between the feature vector and the existing node feature vectors in the pre-constructed accounting knowledge graph is calculated as the first matching degree, specifically including: Identify the existing node subgraphs related to the current business type, and traverse each existing node in the subgraph; Calculate the cosine similarity between the feature vector and the feature vectors of each existing node traversed; Extract the top K values ​​with the highest cosine similarity from the calculated values, calculate a weighted average of the top K similarity values, and then determine the weighted average as the first matching degree, where K is a preset integer greater than 2.

[0009] Preferably, creating a new corresponding node for the financial entity or financial relationship in the accounting knowledge graph and establishing its association edges with related nodes specifically includes: Identify at least one existing node that is most relevant to the current business data as the anchor node; Create a new node for the financial entity or financial relationship; Based on the identified financial relationship information, determine the type and direction of the associated edge between the new node and the anchor node; Add the new node and the associated edge to the accounting knowledge graph.

[0010] Preferably, supplementing the attribute information of existing nodes in the accounting knowledge graph and recalculating the weight values ​​of related edges based on the supplemented graph structure specifically includes: The existing nodes in the accounting knowledge graph are identified as nodes to be added. The identified financial attribute information is merged with the existing attribute information of the node to be supplemented, and the attribute information of the node is updated. Based on the co-occurrence of the updated attribute information of the node to be supplemented with the attribute information of its neighboring nodes, the weight values ​​of the associated edges between the node to be supplemented and its neighboring nodes are recalculated and updated.

[0011] Preferably, the second preset threshold is greater than the first preset threshold.

[0012] Secondly, this application provides a financial task allocation system based on accounting knowledge graph updates, including: The acquisition module is used to acquire business form data and parse the business form data using a pre-trained information extraction model to identify financial entities, financial relationships and financial attribute information therein; The processing module is used to convert the identified financial entity and financial relationship information into feature vectors, calculate the similarity between the feature vector and the feature vector of existing nodes in the pre-built accounting knowledge graph, and use it as the first matching degree. The processing module is further configured to dynamically update the accounting knowledge graph based on the first matching degree: when the first matching degree is lower than a first preset threshold, a new node corresponding to the financial entity or financial relationship is created in the accounting knowledge graph, and an association edge is established between it and the relevant node; when the first matching degree is not lower than the first preset threshold but lower than a second preset threshold, the attribute information of the corresponding existing node in the accounting knowledge graph is supplemented, and the weight value of the relevant association edge is recalculated based on the supplemented graph structure. The execution module is used to recalculate the second matching degree between the financial tasks to be assigned and each task execution entity node based on the updated accounting knowledge graph, and generate a dynamically adjusted financial task allocation scheme based on the second matching degree.

[0013] Thirdly, this application provides a computer device including a memory and a processor, the memory storing code, and the processor being configured to acquire the code and execute the above-described financial task allocation method based on accounting knowledge graph updates.

[0014] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described financial task allocation method based on accounting knowledge graph updates.

[0015] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects: In this embodiment, based on the updated accounting knowledge graph, the second matching degree between the financial tasks to be assigned and the nodes of each task execution entity is recalculated, and a dynamically adjusted financial task allocation scheme is generated based on the second matching degree. First, the identified financial entities and financial relationship information are converted into feature vectors, and the similarity between these feature vectors and the feature vectors of existing nodes in the pre-constructed accounting knowledge graph is calculated as the first matching degree. The first matching degree can quantitatively characterize the degree of semantic overlap between a single business data and the existing knowledge base, thereby realizing novelty detection, confidence assessment, and priority ranking of candidate nodes, reducing mismatch and manual screening costs. Second, based on the first matching degree, the accounting knowledge graph is dynamically updated. Dynamic updates can expand the semantic coverage of the graph, correct outdated or incomplete node representations, and make the edge weights reflect the latest business. Co-occurrence and confidence levels are used to suppress knowledge drift and improve the timeliness and accuracy of the graph representation. Then, based on the updated accounting knowledge graph, the second matching degree between the financial tasks to be assigned and the nodes of each task execution entity is recalculated. Through the second matching degree, the task requirements and entity capabilities can be finely aligned in the vector space. Taking into account topological neighborhood, attribute similarity, and relationship strength, the candidate entities are ranked and fitted more reliably. Finally, a dynamically adjusted financial task allocation scheme is generated based on the second matching degree, which can realize automatic selection based on fit and business constraints, and can combine load balancing, confidence gating, and policy rules to perform multi-objective trade-offs and auditable decision output. In summary, the proposed scheme can realize the dynamic collaborative evolution of the graph and task allocation, thereby improving the system's ability to autonomously adapt to changes in financial business. Attached Figure Description

[0016] Figure 1This is a schematic diagram illustrating an application scenario of a financial task allocation method based on accounting knowledge graph updates, according to some embodiments of this application. Figure 2 This is an exemplary flowchart of a financial task allocation method based on accounting knowledge graph updates, as shown in some embodiments of this application. Figure 3 This is a schematic flowchart illustrating the process of determining a first matching degree according to some embodiments of this application; Figure 4 This is a schematic diagram of the structure of a financial task allocation system based on accounting knowledge graph updates, according to some embodiments of this application; Figure 5 This is a schematic diagram of the structure of a computer device that implements a financial task allocation method based on accounting knowledge graph updates, according to some embodiments of this application. Detailed Implementation

[0017] To better understand the technical solution of this application, the technical solution of this application will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] refer to Figure 1 This figure is a schematic diagram of an application scenario of a financial task allocation method based on accounting knowledge graph updates, according to some embodiments of this application. The figure mainly includes a business data source, a server, a communication network, and a task management terminal. The business data source is connected to the server through the communication network. The server obtains various business form data from the business data source, parses the financial entities, relationships, and attributes in the data using a pre-trained information extraction model, and then converts them into feature vectors. These vectors are then matched and calculated with a pre-constructed accounting knowledge graph. The graph structure is dynamically updated based on the matching results, and finally, a financial task allocation scheme is generated based on the updated graph. When a financial task processing request is received through the task management terminal, the server feeds back the dynamically adjusted allocation scheme to the terminal for execution by financial personnel or the system.

[0019] The business data source can be an Enterprise Resource Planning (ERP) system, a financial software database, or an electronic document storage system; the task management terminal can be a personal computer, a dedicated financial workstation, or a mobile approval device; the server can be an independent physical server, a cluster or distributed system composed of multiple servers, or a cloud server that provides computing power services.

[0020] refer to Figure 2 The figure is an exemplary flowchart of a financial task allocation method based on accounting knowledge graph updates, according to some embodiments of this application. This financial task allocation method based on accounting knowledge graph updates mainly includes the following steps: In step 101, business form data is acquired, and the business form data is parsed using a pre-trained information extraction model to identify the financial entities, financial relationships, and financial attribute information therein.

[0021] It should be noted that the business form data refers to structured or semi-structured business voucher data submitted by various business departments and entering the financial process during corporate financial business processing. This includes, but is not limited to, expense reports, invoices, payment and receipt applications, purchase applications, contract approval forms, expense allocation forms, and their accompanying explanatory texts. These documents are used to record specific financial business transactions, amount information, responsible parties, and business process status. Business form data is typically acquired through the company's internal financial information system, business management system, or image archiving system. Specifically, the data access module automatically retrieves the business forms to be processed from these systems and, using character recognition, form structure parsing, and field extraction technologies, converts data from different sources, such as scanned documents, electronic forms, and online data, into a standardized input format that can be parsed by the model.

[0022] In some embodiments, the following steps can be used to parse the business form data using a pre-trained information extraction model to identify financial entities, financial relationships, and financial attribute information: Convert the acquired business form data into a unified text data format; Text data is input into a pre-trained information extraction model, which sequentially performs named entity recognition to locate financial entities, performs relation extraction to determine the financial relationships between the financial entities, and performs attribute extraction to obtain the financial attribute information of the financial entities. The identified financial entities, financial relationships, and financial attribute information will be output in a structured form.

[0023] It should be noted that the pre-trained information extraction model is a semantic parsing model used to automatically identify financial entities, financial relationships, and financial attribute information in text.

[0024] In specific implementation, firstly, through text cleaning and normalization procedures, irrelevant format markers and garbled characters are removed, and all characters are uniformly encoded. Finally, all business form data from different sources are output as plain text files with consistent encoding and no format markers, which serve as the unified text data format. Then, the pre-trained information extraction model is a neural network architecture based on a multi-layer bidirectional transformer. Its training process is divided into two stages: self-supervised pre-training on a mixed corpus containing massive amounts of general-domain text and financial-domain professional text to learn basic language representations; and further supervised fine-tuning on a manually annotated financial document dataset. In this dataset, each sentence is annotated with the boundaries and types of financial entities (such as accounts receivable, supplier B), relationships between entities (such as payment, attribution), and entity attributes. Model parameter settings include: a maximum sequence length of 512 characters, 12 attention heads, a feedforward network dimension of 3072, a training batch size of 16, and optimization using an adaptive moment estimation algorithm. When processing the input plain text, its named entity recognition module uses sequence tags... The annotation technology locates and classifies financial entities in the text; the relationship extraction module classifies and judges what predefined financial relationships exist between identified entity pairs based on the contextual semantics between entities; the attribute extraction module identifies and extracts attribute key-value pairs associated with specific financial entities, and the model outputs the original identification results containing entity type, relationship type, and attribute key-value pairs as financial attribute information; finally, after receiving the original identification results output by the information extraction model, a structured management program is called. This program maps and instantiates each item in the original results according to a predefined financial domain ontology data pattern. For example, an entity identified as "accounts payable", a "payment" relationship and its associated "amount" and "due date" attributes are instantiated into an "accounts payable" node with a unique identifier, a "payment" relationship edge pointing to a "supplier" node, and "amount" and "due date" attribute fields attached to the node. The program organizes and encapsulates all instantiated nodes, edges and attribute fields according to the specifications of graph data structures, and finally outputs the encapsulated standard graph structure data object as a structured form.

[0025] In step 102, the identified financial entities and financial relationship information are converted into feature vectors, and the similarity between the feature vectors and the feature vectors of existing nodes in the pre-constructed accounting knowledge graph is calculated as the first matching degree.

[0026] In some embodiments, converting the identified financial entity and financial relationship information into a feature vector can be achieved through the following steps: The identified financial entities are mapped into entity semantic vectors using a pre-trained word vector model; The identified financial relationships are mapped into relational semantic vectors through a pre-trained relational semantic model. The entity semantic vectors corresponding to the two financial entities connected by the financial relationship are combined to generate relational feature vectors. The feature vector is generated by concatenating all entity semantic vectors with the relation feature vectors.

[0027] It should be noted that the entity semantic vector is a vector used to represent the feature representation of a single financial entity in the semantic space; the relation semantic vector is a vector used to represent the semantic features of a specific financial relationship between financial entities; the relation feature vector is a vector used to comprehensively express the semantic information of financial relationships and their associated entities; and the word vector model is a model used to map words in text into continuous numerical vectors to express their semantic features.

[0028] In specific implementation, firstly, a phrase-level word vector model pre-trained on a large financial domain corpus, such as Phrase2Vec, is called. This corpus includes financial statements, audit reports, accounting academic literature, and financial news texts, and has already performed phrase annotation preprocessing for multi-word entities in the financial domain. The model uses a continuous bag-of-words architecture extended phrase embedding training method, with parameters including: 300-dimensional word / phrase vectors, a training window size of 5 words / phrases, 5 negative samples, and an initial learning rate of 0.025. For the identified financial entities: if it is a single word, such as "supplier," it is directly input into the model and outputs a 300-dimensional vector; if it is a multi-word entity, such as "accounts payable - supplier B," after inputting into the model, the model outputs a 300-dimensional vector of the phrase through a pre-trained phrase mapping table; if the model does not cover the phrase, the word vectors after entity splitting are fused into a 300-dimensional entity semantic vector using an average pooling formula. Then, a relational semantic model trained for relational semantics is called. The model is trained on a corpus containing numerous financial relationship statements (e.g., Company A_Payment_Company B). The training objective is to learn a distributed representation for relational vocabulary. The identified financial relationships are then input into the model to obtain a relational semantic vector. Simultaneously, the entity semantic vectors generated in the first step for the two financial entities connected by the relationship are retrieved, and then combined with the relational semantic vector through vector addition. Finally, all entity semantic vectors and relational feature vectors obtained from the first two steps of processing the current business form data are collected. These vectors are then concatenated in a predefined fixed order, such as the order in which entities appear in the text, followed by the relational feature vector. If a single vector has a dimension of 300, and a total of 3 entities and 1 relationship are identified, the concatenated vector will generate a 1200-dimensional long vector. Optionally, this long vector is normalized to a magnitude of 1 to improve the stability of subsequent similarity calculations. This concatenated and normalized long vector is used as the feature vector for the first matching degree calculation.

[0029] In some embodiments, reference Figure 3 As shown in the figure, this is a flowchart illustrating the process of determining the first matching degree in some embodiments of this application. In this embodiment, the similarity between the feature vector and the feature vector of an existing node in the pre-constructed accounting knowledge graph is calculated as the first matching degree, which can be achieved by the following steps: In step 1021, an existing node subgraph related to the current business type is determined, and each existing node in the subgraph is traversed. In step 1022, the cosine similarity between the feature vector and the feature vectors of each existing node traversed is calculated; In step 1023, the top K values ​​with the highest cosine similarity scores are extracted, and the weighted average of these top K similarity scores is calculated. The weighted average is then determined as the first matching degree, where K is a preset integer greater than 2.

[0030] It should be noted that the first matching degree in this application is an indicator used to quantify the similarity between the current business feature vector and the existing node feature vectors in the accounting knowledge graph.

[0031] In practice, firstly, based on the key category information extracted from the current business form data, such as procurement and expense reimbursement, this information is used as the query keyword and retrieved from the node type index of the pre-built accounting knowledge graph. All graph nodes labeled with this business type are located according to the index, and these nodes and their directly connected edges form a local subgraph. Further, each node is visited sequentially according to its storage order in the subgraph to prepare for subsequent calculations. Each visited node is used as an existing node for comparison. This is done because filtering nodes by business type to form a subgraph significantly reduces the number of nodes requiring calculation, avoiding inefficient full traversal across the entire graph, thus significantly improving computational efficiency. Secondly, for each existing node visited during the traversal, its pre-stored feature vector representation in the knowledge graph is read. Then, according to the cosine similarity calculation formula, the cosine similarity between the feature vector generated from the current business data and the node's feature vector is calculated. Specifically, this calculation is the ratio of the dot product of the two vectors to the product of their respective moduli. The calculated ratio is used as the cosine similarity between the existing node and the current business data. This is because cosine similarity effectively measures the directional similarity between two high-dimensional vectors, without being affected by their absolute length, making it particularly suitable for comparing feature vectors generated based on semantic information. After calculating the cosine similarity of all nodes in the subgraph, all the results are stored in a list, sorted in descending order, and the top K similarity values ​​are selected. Then, according to a preset weighting scheme (for example, the first-ranked similarity value is given the highest weight, and the weight decreases as the ranking decreases), these K values ​​are weighted and summed, and divided by the total weight to obtain a weighted average. Finally, this weighted average is determined as the first matching degree of this calculation. This is because taking only the single highest similarity value is easily affected by noise or random matching, while taking a weighted average based on the information of the top K most similar nodes makes the first matching degree more robust and reliable, and better reflects the degree of consistency between the current business and the overall relevant part of the knowledge graph.

[0032] In step 103, the accounting knowledge graph is dynamically updated based on the first matching degree.

[0033] It should be noted that, based on the first matching degree, the dynamic updating of the accounting knowledge graph includes: when the first matching degree is lower than a first preset threshold, creating a corresponding new node for the financial entity or financial relationship in the accounting knowledge graph and establishing its association edge with the relevant node; when the first matching degree is not lower than the first preset threshold but lower than a second preset threshold, supplementing the attribute information of the corresponding existing node in the accounting knowledge graph, and recalculating the weight value of the relevant association edge based on the supplemented graph structure; it should also be noted that the second preset threshold is greater than the first preset threshold, and the method for setting the first and second preset thresholds is: based on historical new business The similarity statistical distribution of feature vectors with existing node feature vectors in the pre-built accounting knowledge graph (rather than the similarity distribution between existing nodes) is set. Specifically, feature vectors generated from all new business data in the past 6-12 months are first collected, and their cosine similarity with existing node feature vectors in the corresponding business type subgraph is used to form a historical similarity sample set, i.e., the new-old matching similarity set. Then, the lower quantile value (e.g., 25th percentile, such as 0.4) in this sample set is selected as the first preset threshold, which determines that it is a brand new business and a new node needs to be created. The higher quantile value (e.g., 75th percentile, such as 0.8) is selected as the second preset threshold, which determines that the business is highly matched and no node attributes need to be updated.

[0034] It should also be noted that when the first matching degree is higher than the second preset threshold, it indicates that the semantic consistency between the current business data and the existing nodes in the accounting knowledge graph is high, and it can be fully expressed in the existing graph structure. At this time, there is no need to add new nodes or supplement attribute information. Instead, the target node with the highest matching degree is directly reused as the graph mapping object of the business data, and the existing relationship structure of the target node is used to participate in the subsequent task matching and allocation calculation, thereby avoiding unnecessary graph modification, maintaining the stability and consistency of the graph structure, and ensuring that task allocation is processed based on the most reliable and mature node semantic representation.

[0035] In some embodiments, creating a corresponding new node for the financial entity or financial relationship in the accounting knowledge graph and establishing its association with related nodes can be achieved through the following steps: Identify at least one existing node that is most relevant to the current business data as the anchor node; Create a new node for the financial entity or financial relationship; Based on the identified financial relationship information, determine the type and direction of the associated edge between the new node and the anchor node; Add the new node and the associated edge to the accounting knowledge graph.

[0036] It should be noted that, in this application, anchor nodes are reference nodes used to connect new nodes in the accounting knowledge graph to establish relationships.

[0037] In specific implementation, firstly, the intermediate data generated during the calculation of the first matching degree is queried. This data records a list of similarities between the current business feature vector and the feature vectors of each existing node in the knowledge graph subgraph. Then, the existing node with the highest cosine similarity value is selected from this list. If the highest value is still lower than the first preset threshold, it indicates that there are no highly relevant nodes. At this time, the system searches for root nodes or high-frequency nodes with central hub status and type relevance in the graph according to the business type. Finally, the selected one or more existing nodes are used as anchor nodes connecting new knowledge. Secondly, a unique node identifier is assigned to the financial entity or financial relationship to be added in the global scope of the knowledge graph. Then, the structured information obtained from the information extraction step is assigned to the new node: the identified name or description is used as the node label, the feature vector output by the information extraction model is used as the initial feature vector of the node, and the identified financial attribute information is stored as the attribute key-value pairs of the node. After the above assignment is completed, the data object with the identifier, label, feature vector, and attributes is used as a new node to be added to the knowledge graph. Then, the financial relationship information identified from the business form is read. This information is in the form of triples. The relationships are stored formally, and based on the relational predicates in the triples, a match is searched in a predefined dictionary of financial relationship types to determine the formal type names of the associated edges. Simultaneously, the directionality of the relationship is determined based on the roles of the subject and object in the triples: from the node representing the subject entity to the node representing the object entity. If the new node represents an entity, the direction is from the anchor node to the new node, or vice versa. If the new node itself represents a relationship, multiple edges need to be created pointing from that relationship node to the relevant entity nodes. The determined relationship type names and their directionality rules are used as the class for constructing the associated edges. The process involves determining the type and direction of the new node data. Finally, it accesses the accounting knowledge graph data currently stored in memory or a graph database. First, it inserts the newly created node data into the graph's node set according to the graph's node data format. Then, based on the determined type and direction of the associated edges, it creates one or more edges between the anchor node and the new node, or between the new relation node and the related entity node. Each edge is assigned a corresponding type label and direction attribute. The updated graph data is then saved to persistent storage, completing this addition operation. The updated graph structure data, with the nodes and edges inserted, serves as the latest accounting knowledge graph.

[0038] In some embodiments, supplementing the attribute information of existing nodes in the accounting knowledge graph and recalculating the weight values ​​of related edges based on the supplemented graph structure can be achieved through the following steps: The existing nodes in the accounting knowledge graph are identified as nodes to be added. The identified financial attribute information is merged with the existing attribute information of the node to be supplemented, and the attribute information of the node is updated. Based on the co-occurrence of the updated attribute information of the node to be supplemented with the attribute information of its neighboring nodes, the weight values ​​of the associated edges between the node to be supplemented and its neighboring nodes are recalculated and updated.

[0039] It should be noted that the co-occurrence rate in this application is an indicator used to quantify the degree to which the attribute information of the node to be supplemented and the attribute information of its neighboring nodes appear simultaneously or are related; the weight value of the associated edge is a value used to reflect the strength and importance of the relationship between nodes in the knowledge graph.

[0040] In practice, firstly, the similarity list generated during the calculation of the first matching degree is queried, and the existing node with the highest cosine similarity value is selected. It is determined that this highest similarity value falls between a first preset threshold and a second preset threshold, meeting the condition for triggering attribute supplementation and updating. Then, using the node's unique identifier, the complete data of the node is located and loaded into the accounting knowledge graph. This located and loaded existing node is designated as the node to be supplemented in this knowledge enhancement operation. Secondly, the financial attribute information identified and structured from the current business form is read. This information is organized in the form of attribute key-value pairs. These new attribute key-value pairs are traversed. For each attribute key, it is checked whether it exists in the existing attribute set of the node to be supplemented. If it exists, the new attribute value is merged or replaced with the original attribute value of the node. For example, for the "recent transaction time" attribute, the new time replaces the old time; for the "number of transactions" attribute, the new and old values ​​are added together. If it does not exist... The new attribute key-value pair is directly added to the node's attribute set. After the fusion of all attribute pairs is completed, the fused attribute key-value pair set with updated content is used as the latest attribute information of the node to be supplemented. Then, all directly adjacent nodes of the node to be supplemented in the graph are obtained. For each adjacent node, the updated attribute information of the node to be supplemented and the attribute information of the adjacent node are extracted respectively. The co-occurrence of the two attributes is quantified by statistically analyzing the frequency of the same or related attribute values ​​in the two types of information in historical business data, or by applying predefined association rules to calculate their confidence. The calculated co-occurrence value is then normalized and mapped to a standardized weight value. This weight value is then used to update the weight field stored on the association edge between the node to be supplemented and the adjacent node. After traversing all adjacent nodes and completing the update, the standardized values ​​on all updated association edges are used as weight values ​​reflecting the latest association strength between nodes.

[0041] In step 104, based on the updated accounting knowledge graph, the second matching degree between the financial tasks to be assigned and each task execution entity node is recalculated, and a dynamically adjusted financial task allocation scheme is generated based on the second matching degree.

[0042] In some embodiments, recalculating the second matching degree between the financial tasks to be assigned and each task execution entity node based on the updated accounting knowledge graph can be achieved using the following steps: Based on the updated accounting knowledge graph, the semantic feature vectors of the financial tasks to be assigned and the capability feature vectors of each task execution entity node are obtained through graph embedding algorithm. Calculate the similarity between the semantic feature vector of the financial task to be assigned and the capability feature vector of each task execution entity node; The calculated similarity is used as the second matching degree of the corresponding task execution subject node.

[0043] It should be noted that the capability feature vector in this application is a vector used to represent the various capability features of the task execution subject node when executing financial tasks; the second matching degree is an indicator used to quantify the degree of matching between the financial tasks to be assigned and the capability features of each task execution subject node.

[0044] In practical implementation, firstly, the dynamically updated accounting knowledge graph is used as input, and a graph embedding algorithm is invoked. This algorithm generates a large number of node sequences originating from different nodes through a simulated random walk approach. These sequences are treated as "sentences" and input into a shallow neural network for training. The training objective is to make neighboring nodes in the original graph have similar vector representations in the vector space. Through this process, the algorithm learns and generates a fixed-dimensional low-dimensional vector for each node in the graph. This low-dimensional vector representing the task execution entity node is used as its capability feature vector, and the low-dimensional vector representing the financial task to be assigned is used as its semantic feature vector. The reason for this is that the graph embedding algorithm can compress and encode the complex graph topology, node attributes, and relationships into a unified numerical vector, making the matching between tasks and entities a highly efficient vector computation. Secondly, the semantic feature vectors of the financial tasks to be assigned are obtained, and each task execution entity node is read sequentially. For each task execution entity node, the cosine similarity between its capability feature vector and the task semantic feature vector is calculated. This calculation is performed by dividing the dot product of the two vectors by the product of their respective magnitudes, resulting in a value between -1 and 1. This calculated cosine similarity value is used as the similarity between the task execution entity node and the current task. This is done because cosine similarity is a standard method for measuring the consistency of the directions of two vectors, is simple and efficient to calculate, and can effectively assess the alignment of task requirements and entity capabilities in terms of semantic and structural features. Then, a temporary record is created for each task execution entity node participating in the calculation, and the calculated similarity between the node and the task is stored as the score field of the record. After the similarity of all nodes has been calculated, the system organizes these record items into a list, and the similarity score stored in each record item in the list is used as the second matching degree of the corresponding task execution entity node.

[0045] In some embodiments, generating a dynamically adjusted financial task allocation scheme based on a second matching degree can be achieved through the following steps: Based on the calculated second matching degree of each task execution entity node, all candidate task execution entity nodes are sorted in descending order. Based on the preset allocation strategy, at least one task execution subject node is selected from the sorted list; Generate a list of allocation schemes containing the correspondence between the financial tasks to be assigned and the selected task execution entity nodes.

[0046] It should be noted that the allocation scheme list in this application is a structured data set used to record the correspondence between the financial tasks to be allocated and the selected task execution entity nodes.

[0047] In practice, the process begins with two steps. First, a sorting algorithm is invoked. This algorithm compares the second matching degree values ​​of any two nodes, placing the node with the larger value at the beginning of the sequence and the node with the smaller value at the end, thus rearranging all nodes. The resulting ordered sequence, from the node with the highest second matching degree to the node with the lowest second matching degree, serves as the input queue for subsequent allocation strategy processing. This is done because sorting all candidates in descending order of matching degree clearly demonstrates their suitability priority for the task requirements, providing a clear and orderly decision-making basis for subsequent automated or semi-automated decisions. This is the first step towards achieving efficient and reasonable allocation. Second, the allocation strategy rules pre-configured in the strategy management module are read. These rules may include options such as "directly select the first-ranked node," "randomly select from the top three nodes," or "select the node with the lightest current workload from the top-ranked nodes." The sorted list, along with relevant real-time data, is then input into the strategy logic. The process involves performing calculations. Based on the results of the strategy logic, one or more nodes are selected from the sorted list. The node(s) pointed to by the strategy calculation results are chosen as the final task execution entity nodes. This is because relying solely on matching degree sorting may not meet complex real-world business constraints, such as load balancing and fairness. The preset allocation strategy introduces business rules and real-time status, ensuring that the final selection result is both reasonable in terms of capability matching and meets the needs of actual management and scheduling. Then, a structured data list is created. Each record in the list contains at least two core fields. The first field is filled with the unique task identifier of the financial task to be allocated, and the second field is filled with the unique node identifier of the task execution entity node selected in the second step. If a task needs to be allocated to multiple entities, records containing multiple corresponding relationships are generated. Finally, the data list encapsulating the task-entity correspondence records is used as the final output of this dynamic task allocation process, namely the allocation scheme list.

[0048] On the other hand, in some embodiments, this application provides a financial task allocation system based on accounting knowledge graph updates, referencing Figure 4 The figure is a schematic diagram of the structure of a financial task allocation system based on accounting knowledge graph updates, according to some embodiments of this application. The financial task allocation system 400 based on accounting knowledge graph updates includes: an acquisition module 401, a processing module 402, and an execution module 403, which are described below: The acquisition module 401 in this application is mainly used to acquire business form data and use a pre-trained information extraction model to parse the business form data in order to identify the financial entities, financial relationships and financial attribute information therein. Processing module 402, in this application, is used to convert the identified financial entity and financial relationship information into feature vectors, and calculate the similarity between the feature vectors and the existing node feature vectors in the pre-constructed accounting knowledge graph as the first matching degree; In this application, the processing module 402 is further configured to dynamically update the accounting knowledge graph based on the first matching degree: when the first matching degree is lower than a first preset threshold, a new node corresponding to the financial entity or financial relationship is created in the accounting knowledge graph, and an association edge is established between it and the relevant node; when the first matching degree is not lower than the first preset threshold but lower than a second preset threshold, the attribute information of the corresponding existing node in the accounting knowledge graph is supplemented, and the weight value of the relevant association edge is recalculated based on the supplemented graph structure. The execution module 403 in this application is mainly used to recalculate the second matching degree between the financial tasks to be assigned and each task execution entity node based on the updated accounting knowledge graph, and generate a dynamically adjusted financial task allocation scheme based on the second matching degree.

[0049] In addition, this application also provides a computer device, the computer device including a memory and a processor, the memory storing code, the processor being configured to acquire the code and execute the above-described financial task allocation method based on accounting knowledge graph updates.

[0050] In some embodiments, reference Figure 5 The figure is a schematic diagram of the structure of a computer device implementing a financial task allocation method based on accounting knowledge graph updates, according to some embodiments of this application. The financial task allocation method based on accounting knowledge graph updates in the above embodiments can be implemented through... Figure 5 The computer device shown is used to implement this, and the computer device 500 includes at least one processor 501, a communication bus 502, a memory 503, and at least one communication interface 504.

[0051] Processor 501 can be a general-purpose central processing unit (CPU) or an application-specific integrated circuit (ASIC).

[0052] The communication bus 502 can be used to transmit information between the aforementioned components.

[0053] Memory 503 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CDROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital versatile optical discs, Blu-ray discs, etc.), magnetic disks or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. Memory 503 may exist independently and be connected to processor 501 via communication bus 502. Memory 503 may also be integrated with processor 501.

[0054] The memory 503 stores program code for executing the scheme of this application, and its execution is controlled by the processor 501. The processor 501 executes the program code stored in the memory 503. The program code may include one or more software modules. The financial task allocation method based on accounting knowledge graph updates in the above embodiments can be implemented by the processor 501 and one or more software modules in the program code in the memory 503.

[0055] Communication interface 504 uses any transceiver-like device to communicate with other devices or communication networks, such as Ethernet, radio access network (RAN), wireless local area networks (WLAN), etc.

[0056] In a specific implementation, as one example, a computer device may include multiple processors, each of which may be a single-core (single CPU) processor or a multi-core (multi CPU) processor. Here, a processor may refer to one or more devices, circuits, and / or processing cores used to process data (e.g., computer program instructions).

[0057] The aforementioned computer device can be a general-purpose computer device or a special-purpose computer device. In specific implementations, the computer device can be a desktop computer, a portable computer, a network server, a handheld digital assistant (PDA), a mobile phone, a tablet computer, a wireless terminal device, a communication device, or an embedded device. This application does not limit the type of computer device.

[0058] In addition, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described financial task allocation method based on accounting knowledge graph updates.

[0059] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0060] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A financial task allocation method based on accounting knowledge graph updates, characterized in that, include: Obtain business form data, and use a pre-trained information extraction model to parse the business form data in order to identify the financial entities, financial relationships and financial attribute information therein; The identified financial entities and financial relationship information are transformed into feature vectors, and the similarity between these feature vectors and the feature vectors of existing nodes in the pre-constructed accounting knowledge graph is calculated as the first matching degree. Based on the first matching degree, the accounting knowledge graph is dynamically updated: When the first matching degree is lower than the first preset threshold, a new corresponding node is created for the financial entity or financial relationship in the accounting knowledge graph, and an association edge is established between it and the relevant node. When the first matching degree is not lower than the first preset threshold but lower than the second preset threshold, the attribute information of the corresponding existing nodes in the accounting knowledge graph is supplemented, and the weight values ​​of the relevant associated edges are recalculated based on the supplemented graph structure. Based on the updated accounting knowledge graph, the second matching degree between the financial tasks to be assigned and the nodes of each task execution entity is recalculated, and a dynamically adjusted financial task allocation scheme is generated based on the second matching degree.

2. The method as described in claim 1, characterized in that, The business form data is parsed using a pre-trained information extraction model to identify financial entities, financial relationships, and financial attribute information, specifically including: Convert the acquired business form data into a unified text data format; Text data is input into a pre-trained information extraction model, which sequentially performs named entity recognition to locate financial entities, performs relation extraction to determine the financial relationships between the financial entities, and performs attribute extraction to obtain the financial attribute information of the financial entities. The identified financial entities, financial relationships, and financial attribute information will be output in a structured form.

3. The method as described in claim 1, characterized in that, The process of converting identified financial entities and financial relationship information into feature vectors specifically includes: The identified financial entities are mapped into entity semantic vectors using a pre-trained word vector model; The identified financial relationships are mapped into relational semantic vectors through a pre-trained relational semantic model. The entity semantic vectors corresponding to the two financial entities connected by the financial relationship are combined to generate relational feature vectors. The feature vector is generated by concatenating all entity semantic vectors with the relation feature vectors.

4. The method as described in claim 1, characterized in that, Calculating the similarity between this feature vector and the existing node feature vectors in the pre-constructed accounting knowledge graph, as the first matching degree, specifically includes: Identify the existing node subgraphs related to the current business type, and traverse each existing node in the subgraph; Calculate the cosine similarity between the feature vector and the feature vectors of each existing node traversed; Extract the top K values ​​with the highest cosine similarity from the calculated values, calculate a weighted average of the top K similarity values, and then determine the weighted average as the first matching degree, where K is a preset integer greater than 2.

5. The method as described in claim 1, characterized in that, Creating corresponding new nodes for the financial entities or financial relationships in the accounting knowledge graph, and establishing their association edges with related nodes, specifically includes: Identify at least one existing node that is most relevant to the current business data as the anchor node; Create a new node for the financial entity or financial relationship; Based on the identified financial relationship information, determine the type and direction of the associated edge between the new node and the anchor node; Add the new node and the associated edge to the accounting knowledge graph.

6. The method as described in claim 1, characterized in that, The process of supplementing the attribute information of existing nodes in the accounting knowledge graph and recalculating the weight values ​​of related edges based on the supplemented graph structure specifically includes: The existing nodes in the accounting knowledge graph are identified as nodes to be added. The identified financial attribute information is merged with the existing attribute information of the node to be supplemented, and the attribute information of the node is updated. Based on the co-occurrence of the updated attribute information of the node to be supplemented with the attribute information of its neighboring nodes, the weight values ​​of the associated edges between the node to be supplemented and its neighboring nodes are recalculated and updated.

7. The method as described in claim 1, characterized in that, The second preset threshold is greater than the first preset threshold.

8. A financial task allocation system based on accounting knowledge graph updates, characterized in that, include: The acquisition module is used to acquire business form data and parse the business form data using a pre-trained information extraction model to identify financial entities, financial relationships and financial attribute information therein; The processing module is used to convert the identified financial entity and financial relationship information into feature vectors, calculate the similarity between the feature vector and the feature vector of existing nodes in the pre-built accounting knowledge graph, and use it as the first matching degree. The processing module is further configured to dynamically update the accounting knowledge graph based on the first matching degree: when the first matching degree is lower than a first preset threshold, a new node corresponding to the financial entity or financial relationship is created in the accounting knowledge graph, and an association edge is established between it and the relevant node; when the first matching degree is not lower than the first preset threshold but lower than a second preset threshold, the attribute information of the corresponding existing node in the accounting knowledge graph is supplemented, and the weight value of the relevant association edge is recalculated based on the supplemented graph structure. The execution module is used to recalculate the second matching degree between the financial tasks to be assigned and each task execution entity node based on the updated accounting knowledge graph, and generate a dynamically adjusted financial task allocation scheme based on the second matching degree.

9. A computer device comprising a memory and a processor, the memory storing code, characterized in that, The processor is configured to acquire the code and execute the financial task allocation method based on accounting knowledge graph updates as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the financial task allocation method based on accounting knowledge graph updates as described in any one of claims 1 to 7.

Images