Financial service processing method, apparatus and device

Abstract

This application provides a financial transaction processing method, apparatus, and device, relating to the fields of fintech or big data. It receives business request information input by a user, which is a financial transaction processing request submitted by the user through natural language, images, and / or voice. The method parses the business request information into a task tree, which is a hierarchical task structure. The task tree is processed based on a hierarchical collaboration framework to complete the execution of the business request. This method achieves intelligent, efficient, and automated processing of financial transaction requests, improving business response efficiency and execution accuracy.

Description

Technical Field

[0001] This application relates to the fields of financial technology or big data, and in particular to a financial business processing method, apparatus and equipment. Background Technology

[0002] As the digital transformation of the financial industry continues to accelerate, the channels for reaching financial services, business scenarios, and customer needs are all becoming more diversified and complex. As the core link between financial institutions and their clients, account managers not only need to handle basic tasks such as account management and risk assessment, but also address comprehensive financial service needs across various scenarios.

[0003] Traditional financial business processing relies primarily on multiple heterogeneous business systems, including customer relationship management systems, core business systems, and risk management systems. In actual operations, account managers must manually log into different systems to query, input, and verify data, completing business processes through offline aggregation and cross-system switching.

[0004] However, account managers spend a lot of time on data verification and format conversion when operating across systems, which not only increases the probability of human error but also significantly prolongs business response time, resulting in low efficiency and low accuracy in financial business processing. Summary of the Invention

[0005] This application provides a financial business processing method, apparatus, and equipment to solve the technical problems of low efficiency and low accuracy in financial business processing.

[0006] Firstly, this application provides a financial transaction processing method, applied to a transaction processing terminal, the method comprising:

[0007] Receive business requirement information input by the user, wherein the business requirement information is a financial business processing request submitted by the user through natural language, image and / or voice;

[0008] The business requirement information is parsed into a task tree, which is a hierarchical task structure;

[0009] The task tree is processed based on a hierarchical collaboration framework to complete the execution of business requirements. The hierarchical collaboration framework includes a strategy planning layer, a domain resolution layer, and a tool execution layer. The strategy planning layer is used to allocate resources in task processing. The domain resolution layer is used to decompose the task tree into atomic operation sequences. The tool execution layer is used to call multiple business processing models and existing services to process the atomic operation sequences. The atomic operation sequences are indivisible sequences of the smallest execution units. The existing services include the business system interfaces of financial institutions.

[0010] Secondly, this application provides a financial transaction processing apparatus, comprising:

[0011] The receiving module is used to receive business requirement information input by the user, which is a financial business processing request submitted by the user through natural language, images and / or voice.

[0012] The parsing module is used to parse the business requirement information into a task tree, wherein the task tree is a hierarchical task structure;

[0013] The processing module is used to process the task tree based on a hierarchical collaboration framework to complete the execution of business requirements. The hierarchical collaboration framework includes a strategy planning layer, a domain resolution layer, and a tool execution layer. The strategy planning layer is used to allocate resources in task processing. The domain resolution layer is used to decompose the task tree into atomic operation sequences. The tool execution layer is used to call multiple business processing models and existing services to process the atomic operation sequences. The atomic operation sequences are indivisible sequences of the smallest execution units. The existing services include the business system interfaces of financial institutions.

[0014] Thirdly, embodiments of this application provide a financial business processing device, including: a memory and a processor;

[0015] The memory stores computer-executed instructions;

[0016] The processor executes computer execution instructions stored in the memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0017] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.

[0018] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0019] The financial business processing method provided in this application, through a hierarchical task tree parsing mechanism that supports multimodal inputs such as natural language, images, and voice for receiving financial business requirements, combined with a layered collaborative framework that integrates resource allocation at the strategy planning layer, atomic operation sequence decomposition at the domain parsing layer, multi-business processing models at the tool execution layer, and collaborative scheduling of interfaces with existing financial institution business systems, not only breaks the dependence of traditional financial business processing on structured input and manual intervention, realizing the intelligent and automated end-to-end process of receiving, parsing, and executing business requirements, but also improves business response efficiency and execution accuracy through the minimization of atomic operation sequences and multi-entity collaborative scheduling. Attached Figure Description

[0020] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0021] Figure 1 Flowchart of the financial transaction processing method provided in this application Figure 1 ;

[0022] Figure 2 Flowchart of the financial transaction processing method provided in this application Figure 2 ;

[0023] Figure 3 A schematic diagram of the financial transaction processing device provided in this application;

[0024] Figure 4 A schematic diagram of the financial business processing equipment provided in this application.

[0025] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0026] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0027] It should be noted that the financial business processing methods, apparatus and equipment provided in this application can be used in the fields of financial technology or big data, or in any field other than financial technology or big data. This application does not limit the application fields of the financial business processing methods, apparatus and equipment.

[0028] As the digital transformation of the financial industry continues to accelerate, the channels for reaching financial services, business scenarios, and customer needs are all becoming more diversified and complex. As the core link between financial institutions and their clients, account managers are facing increasingly broader functional boundaries. They must not only implement basic tasks such as account management and risk assessment, but also respond to comprehensive financial service needs across various scenarios. Their work efficiency and service quality directly impact the customer experience of financial institutions.

[0029] Traditional financial transactions are primarily processed using multiple heterogeneous business systems, including key platforms such as customer relationship management systems, core business systems, and risk management systems. In actual business operations, account managers need to manually log into each system to complete data queries, information entry, and content verification in sequence. Then, through offline aggregation and sorting, and repeated switching between systems, the entire business process is finally completed.

[0030] However, existing technologies are limited by the heterogeneous technical architecture of multiple systems and the human-dominated operation mode. When handling cross-system business, account managers need to spend a lot of time on data verification and format conversion. This process not only significantly increases the probability of human error and reduces the accuracy of business data, but also seriously prolongs the business response cycle, directly resulting in low efficiency in financial business processing and making it difficult to meet customers' demands for timeliness and accuracy in the digital age.

[0031] To address the aforementioned issues, the financial business processing method provided in this application first parses unstructured business requirement information submitted by users in natural language, images, and / or voice into a hierarchical task tree. Then, it relies on a layered collaborative framework including a strategy planning layer, a domain analysis layer, and a tool execution layer to carry out task processing. The strategy planning layer is responsible for resource allocation for task processing, the domain analysis layer decomposes the task tree into an indivisible sequence of atomic operations, and the tool execution layer calls multiple business processing models and interfaces of existing financial institution business systems to execute the atomic operation sequence. Through the coordinated scheduling of multiple business processing models and existing services in the atomic operation sequence, the method ultimately achieves intelligent, efficient, and automated processing of financial business requirements, improving business response efficiency and execution accuracy.

[0032] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0033] Figure 1 Flowchart of the financial transaction processing method provided in this application Figure 1 In this embodiment, the executing entity is, for example, a financial business processing system. Figure 1 As shown, the method includes:

[0034] S101: Receive business request information input by the user. The business request information is a financial business processing request submitted by the user through natural language, image and / or voice.

[0035] Specifically, a multimodal information receiving module is built, equipped with functions such as natural language text recognition, image information extraction, speech-to-text conversion, and semantic recognition. This module can connect to financial institutions' online business channels (such as mobile banking apps, online banking, and official accounts) and offline channels (such as smart counter terminals and customer service voice systems). After a user submits a request, the module first standardizes the information format (such as speech-to-text conversion and key image information extraction), then performs preliminary validity checks. Once the request is confirmed to be within the scope of financial business, the standardized request information is passed to the subsequent parsing stage.

[0036] For example, an account manager might submit a business request through the online banking portal, typing "Analyze customer A's credit risk and generate a report" in natural language; or the customer might upload a photo (in image format) of a paper application form containing a credit risk assessment request via their mobile banking app; or the customer might call customer service and state in voice that "a credit risk analysis and report for customer A is needed," and the system's voice receiving module would then convert the speech to text to complete the user's input of the business request information.

[0037] S102: Parse business requirement information into a task tree, which is a hierarchical task structure.

[0038] The task tree is a hierarchical task structure generated by parsing business requirements information. It consists of a main task and several sub-tasks. The main task corresponds to the user's core business needs, and the sub-tasks are the various sub-steps required to complete the main task. There are clear logical connections between the sub-tasks.

[0039] Specifically, semantic parsing is performed on the standardized business requirement information. First, the main task is located using an intent recognition engine. Then, based on a financial business knowledge graph and combined with key elements in the requirements (such as customer identity, business type, and core demands), the main task is broken down into several sub-tasks with logical progression or parallel relationships. After decomposition, a task tree is constructed according to the hierarchical relationship of the main task leading the sub-tasks. Each task node is labeled with a clear task objective and required basic information, ultimately outputting structured task tree data.

[0040] For example, in response to the received request to "analyze customer A's credit risk and generate a report," the system first identifies the main task as "customer A's credit risk analysis and report generation." Based on the knowledge graph of financial credit risk assessment, the main task is broken down into four sub-tasks: "customer A's basic information query," "customer A's public opinion information analysis," "customer A's transaction behavior data collection and analysis," and "credit risk assessment result integration and report writing." A task tree is then constructed, with the main task as the root node and the four sub-tasks as first-level child nodes, forming a clearly hierarchical task tree structure.

[0041] S103: The task tree is processed based on a layered collaboration framework to complete the execution of business requirements. The layered collaboration framework includes a strategy planning layer, a domain resolution layer, and a tool execution layer. The strategy planning layer is used to allocate resources in task processing. The domain resolution layer is used to decompose the task tree into atomic operation sequences. The tool execution layer is used to call multiple business processing models and existing services to process atomic operation sequences. The atomic operation sequence is the smallest indivisible execution unit sequence. Existing services include the existing business system interfaces of financial institutions.

[0042] Specifically, first, the strategy planning layer receives the task tree and, based on a reinforcement learning model, allocates corresponding computing resources, data resources, and preset business processing model resource quotas to each task node, taking into account the current system resource usage, the priority and complexity of each subtask. Next, the domain resolution layer connects the task tree and the resource allocation results from the strategy planning layer. Based on semantic decoupling technology, it further decomposes each subtask into a sequence of atomic operations, with each atomic operation specifying its concrete execution action and the type of resources required. Finally, the tool execution layer loads the atomic operation sequence output by the domain resolution layer, initializes the business processing model call interface and existing service integration channels, and completes the preparation work for subsequent execution.

[0043] For example, regarding the generated task tree of "Customer A's Credit Risk Analysis and Report Generation," the strategy planning layer first assesses the priority of each subtask, assigning a higher priority to "Customer A's Basic Information Query," allocating access resources to the customer information database and a quota for one basic information query business processing model; and assigning a medium priority to "Transaction Behavior Data Collection and Analysis," allocating computing resources to the transaction data platform and a quota for one data analysis business processing model. The domain parsing layer further breaks down the four subtasks into atomic operation sequences: "Query Customer A's Basic Information → Collect Customer A's Public Opinion Data → Analyze Customer A's Public Opinion Risk Level → Collect Customer A's Transaction Data for the Past 3 Years → Analyze the Rationality of Customer A's Transaction Behavior → Integrate Risk Assessment Indicators → Write the Report." The tool execution layer, corresponding to each atomic operation, connects existing service channels such as the customer information query interface and the public opinion data collection interface, and prepares the calling environment for business processing models such as the data analysis business processing model and the report writing model.

[0044] Optionally, the multiple business processing models may include at least:

[0045] A knowledge-based question-answering business processing model is used to answer business knowledge questions;

[0046] Data analytics business processing model, used to analyze customer profiles and transaction behavior;

[0047] A process automation business processing model is used to execute business process operations.

[0048] Specifically, the knowledge-based question-answering business processing model is deeply integrated with the financial business knowledge base through dedicated functional modules configured on the multi-business processing model management platform. The entire process relies on the platform architecture to ensure efficient output, enabling accurate reception, retrieval, and response to business knowledge query requests. The system allocates dedicated functional modules within this management platform, clearly defining the tasks of "knowledge retrieval, semantic matching, and answer output." It also configures a dedicated data interface, limiting its use solely to data interaction with the financial business knowledge base to ensure targeted and secure access. Through this dedicated interface, the model establishes a fixed connection with the knowledge base, encompassing a full range of financial business knowledge, including regulatory policies, business specifications, product descriptions, and operational guidelines. This allows for real-time reading of the knowledge base content and supports dynamic updates (e.g., after regulatory policy updates, the model can update the knowledge base through the interface). (Quickly obtain the latest content); The model receives various business knowledge query requests initiated by users through the system's collaborative scheduling module (such as "requirements of a certain regulatory policy for the sale of wealth management products" or "the handling regulations for personal loan business"). After receiving the request, the semantic parsing function of the dedicated functional module is automatically triggered to perform word segmentation and semantic recognition on the query request to clarify the user's core needs. After the semantic parsing is completed, the retrieval function of the financial business knowledge base is called through the dedicated data interface. Based on the parsed keywords and semantic logic, the knowledge base is accurately searched and matched to select the most relevant knowledge content. After the functional module organizes and optimizes the information (removing redundant information and sorting out the logical structure) to form a clear and accurate answer, it is fed back to the user through the collaborative scheduling module, completing the entire process of business knowledge Q&A.

[0049] The data analysis business processing model is used to analyze customer profiles and transaction behavior. This model operates on a core logic of "data acquisition - algorithm analysis - structured output," relying on interfaces configured on a multi-business processing model management platform to connect to multiple data sources. Through built-in algorithms, it achieves accurate analysis of customer profiles and transaction behavior. The system configures dedicated functional modules for this model within the management platform. These modules include sub-functions such as data collection, data cleaning, algorithm analysis, and result structuring. Multiple dedicated data interfaces are also configured to connect to core data sources such as customer information databases, transaction data platforms, and product holding databases. Data access permissions and data transmission specifications for each interface are clearly defined. The model establishes real-time connections with each data source through these dedicated interfaces. When it receives user-initiated commands for customer profile analysis or transaction behavior analysis (such as "analyze the financial preferences of a certain customer group" or "statistically analyze the frequency and amount distribution of customer transactions over the past three months"), it automatically triggers the data collection interface. According to the scope and dimensions required by the command, it collects basic customer information (age, occupation, asset size) and transaction data (transactions) from the corresponding data sources. Data such as time, amount, product type, and transaction channel are collected. The raw data is then fed into the data cleaning sub-function of the model's dedicated module to perform data deduplication, outlier removal, and format standardization to ensure data accuracy. The cleaned data is then fed into the algorithm analysis sub-function. This module has built-in algorithms for customer profile building, transaction behavior clustering, and trend analysis. For customer profile analysis, the algorithm integrates multi-dimensional customer data and extracts customer tags (such as risk preference, financial needs, and consumption habits) to build a complete customer profile. For transaction behavior analysis, the algorithm statistically analyzes transaction characteristics, mines transaction patterns, and identifies abnormal transactions, completing the core analysis process. After the analysis is completed, the model uses the result structuring sub-function of the dedicated module to convert the analysis results into structured forms such as tables, charts, and text summaries (e.g., customer profile tag tables, transaction behavior trend charts, and analysis conclusion summaries). The structured results are then fed back to the user through the system's collaborative scheduling module. The system also supports the export and archiving of results to meet the needs of business decision-making, customer management, and other scenarios.

[0050] The automated business process model is used to execute business process operations. This model centers on "standard templates - instruction triggering - automatic execution - progress feedback," relying on the collaborative mechanism of a multi-business processing model management platform to achieve automated triggering, execution, and control of financial business processes. The system configures a dedicated functional module within this management platform. The core functions of this module include process template management, interface calls, status synchronization, result archiving, and progress feedback. It also configures existing service interface call permissions to connect with the service interfaces of various business systems. The model also presets various standard financial business process templates, covering common processes such as loan approval, wealth management product redemption, customer account opening, and business archiving. Each template clearly defines the process nodes, the execution order of each node, triggering conditions, corresponding service interfaces, and operating specifications, and supports custom modification and addition of templates. The model receives process execution instructions initiated by users through the system's collaborative scheduling module (such as "execute a customer's loan approval process" or "process a wealth management product redemption process"). After receiving the instruction, its dedicated functional module automatically matches the corresponding standard business process template, confirms the process nodes, execution order, and required service interfaces, and then triggers the automatic execution of the process. After the process is triggered... The model automatically calls the corresponding existing service interfaces according to the preset node order in the template to complete the operations of each node (e.g., in the loan approval process, it automatically calls the customer credit inquiry interface and the approval opinion entry interface; in the wealth management product redemption process, it automatically calls the product holding inquiry interface and the fund transfer interface). At the same time, the module will synchronize the execution status of each task node in real time (e.g., "pending execution", "in execution", "completed", "abnormal interruption"). For abnormal nodes, an automatic reminder mechanism will be triggered and synchronized to the relevant operators. During the process execution, the model will also automatically complete the tasks of synchronizing the status, storing intermediate results, and archiving the final results without manual intervention. During the process execution, the model will collect the execution progress of each node in real time through a dedicated functional module and feed back the progress information (e.g., "process completed 30%, currently at the credit inquiry node") to the user through the collaborative scheduling module to ensure that the user can keep track of the process dynamics in real time. When all process nodes are completed, the model will automatically complete the final result archiving, generate a process execution report and feed it back to the user, and mark the process as completed, forming a closed loop of "instruction triggering - automatic execution - progress feedback - result archiving", realizing the automation and standardization of business process operations.

[0051] The financial business processing method provided in this embodiment receives business requirement information input by the user. This business requirement information is a financial business processing request submitted by the user through natural language, images, and / or voice. The method parses the business requirement information into a task tree, which is a hierarchical task structure. The task tree is processed based on a layered collaboration framework to complete the execution of the business requirement. The layered collaboration framework includes a strategy planning layer, a domain resolution layer, and a tool execution layer. The strategy planning layer allocates resources in task processing; the domain resolution layer decomposes the task tree into atomic operation sequences; and the tool execution layer calls multiple business processing models and existing services to process the atomic operation sequences. The atomic operation sequences are indivisible sequences of the smallest execution units. Existing services include existing business system interfaces of financial institutions. This method achieves intelligent, efficient, and automated processing of financial business requirements, improving business response efficiency and execution accuracy.

[0052] Figure 2 Flowchart of the financial transaction processing method provided in this application Figure 2 ,like Figure 2 As shown, in this embodiment... Figure 1 Based on the examples, the financial business processing method is described in detail, which includes:

[0053] S201: Receive business request information input by the user. The business request information is a financial business processing request submitted by the user through natural language, images and / or voice.

[0054] Step S201 is similar to step S101, and will not be described again here.

[0055] S202: The business requirement information is parsed using a multimodal fusion algorithm to generate a task tree; the multimodal fusion algorithm is a parsing algorithm that combines text, image and speech.

[0056] Among them, the multimodal fusion algorithm is a comprehensive algorithm that integrates the parsing capabilities of text, image and speech. Its core function is to break the limitations of single-modal parsing, realize the complementarity and synergy of multi-source information, and accurately extract the core intent of business needs.

[0057] The task tree is a hierarchical task structure formed by breaking down core requirements. It includes one main task (corresponding to the core requirements) and several sub-tasks (corresponding to the detailed steps to complete the main task).

[0058] Specifically, the system calls a multimodal fusion algorithm to first perform modal parsing on the requirement information. The text modality extracts keywords and semantic intent, the image modality recognizes text content and application elements, and the speech-to-text conversion completes semantic understanding. Then, through the algorithm's feature fusion and intent alignment modules, the parsing results of each modality are integrated to eliminate information ambiguity and accurately locate the core business intent. Finally, based on the financial business logic, the core intent is transformed into a main task, which is then broken down into multiple logically related sub-tasks. A task tree is constructed and output according to the main-sub hierarchy.

[0059] For example, in response to the received request to "analyze customer A's credit risk and generate a report" (taking text modality as an example), the multimodal fusion algorithm first extracts the keywords "customer A," "credit risk analysis," and "report," determining the core intent as "to complete customer A's credit risk assessment and issue a report." Subsequently, this core intent is defined as the core main task, which is then broken down into five sub-tasks: "customer A's basic information query," "customer A's public opinion information collection and analysis," "customer A's transaction behavior data statistics and analysis," "credit risk level assessment," and "report writing and integration." With the main task as the root node and the sub-tasks as first-level nodes, a corresponding task tree is generated.

[0060] S203: At the policy planning layer, resources are allocated through a reinforcement learning model to generate a global task decomposition path.

[0061] Among them, the strategy planning layer is the top-level decision-making layer of the hierarchical collaboration framework, which is used to coordinate the resources and paths for task execution globally.

[0062] Reinforcement learning models are algorithmic models with autonomous learning and decision optimization capabilities, continuously optimizing decision strategies through interaction and feedback with the system environment.

[0063] Specifically, after receiving the task tree, the strategy planning layer first obtains the current system resource usage and the response status of each financial business interface through a reinforcement learning model; then, based on the complexity, urgency, and dependencies of each task in the task tree, it formulates a resource allocation plan through the model's decision module, matching suitable computing resources, data permissions, and preset business processing models for each task node; finally, combining the resource allocation plan and task dependencies, it plans the execution order of each task, generates a global task decomposition path, and outputs it to the domain parsing layer.

[0064] For example, for the task tree of "Credit Risk Analysis and Report Generation for Customer A", the reinforcement learning model in the strategy planning layer first detects that the current customer information database has a low load and the data analysis business processing model is idle. Combining the task logic, it sets "Customer A Basic Information Query" as the highest priority, allocating dedicated access permissions to the customer information database, basic computing resources, and one basic information query business processing model. It sets "Transaction Behavior Data Statistics and Analysis" as the second priority, allocating access permissions to the transaction data platform, high-performance computing resources, and one data analysis business processing model. Finally, it generates a global task decomposition path of "Basic Information Query → Public Opinion Information Analysis → Transaction Behavior Analysis → Risk Level Assessment → Report Writing", clarifying the resource matching rules for each link.

[0065] Optionally, the reinforcement learning model dynamically updates its parameters through a reward function. The input features of the reinforcement learning model include the system resource state and the task execution state, which are used as input features of the model.

[0066] Among them, the reward function is the core basis for updating model parameters. It is used to quantitatively evaluate the merits of model decisions (such as resource allocation schemes) and output reward values ​​to guide the model to iterate towards the optimal decision direction.

[0067] System resource status refers to the current occupancy, idleness, and response status of various resources in the business processing system, including computing resource load, data resource availability, and the idle status of business processing models and existing service interfaces.

[0068] Task execution status refers to the progress of each task in the task tree, including the number of completed tasks, the progress of the currently executing task, the execution time of the task, and the number of task execution failures / retries.

[0069] Specifically, firstly, the system collects and organizes system resource status and task execution status data in real time, performs standardized preprocessing on the data (such as normalizing resource load values ​​and quantifying task progress), and forms feature vectors that meet the model input requirements. Then, the feature vectors are input into the reinforcement learning model, which outputs resource allocation decisions based on the current feature status (such as allocating computing resources to specific tasks or scheduling designated business processing models). After the decision is executed, the reward function, combined with preset evaluation indicators (such as task execution efficiency, resource utilization, and task failure rate), quantifies and scores the effect of this decision and generates a reward value. The better the decision effect, the higher the reward value, and vice versa. Finally, guided by the reward value, the model dynamically updates its internal parameters and adjusts its decision-making strategy through optimization algorithms such as gradient descent, realizing a closed-loop iteration of "input features → decision output → reward feedback → parameter update", continuously improving the rationality and efficiency of the decision.

[0070] For example, in the business scenario of "analyzing customer A's credit risk and generating a due diligence report," the specific input features of the reinforcement learning model are: system resource status (the customer information query interface is currently idle, one data analysis business processing model is idle, high-performance computing resources are at 30% load, and access to the public opinion data resource library is available), and task execution status (currently in the initial stage of the "customer A basic information query" task execution, with 0 completed tasks and no failure / retry records). Based on these input features, the model outputs the decision to "allocate idle customer information query interfaces and basic computing resources to the 'customer A basic information query' task." After the decision is executed, the task is completed efficiently (taking only 2 seconds, below the preset threshold of 5 seconds). The reward function determines that this decision has optimized execution efficiency and improved resource utilization, and outputs a positive high reward value. After receiving this reward value, the model dynamically updates its parameters, reinforcing the decision logic of "prioritizing idle interfaces to match high-priority basic information query tasks."

[0071] S204: In the domain resolution layer, the task tree is decomposed into atomic operation sequences in conjunction with the knowledge graph. The domain resolution layer has a pre-set knowledge graph.

[0072] The domain resolution layer is a task refinement module of the hierarchical collaboration framework, which further breaks down the subtasks of the task tree into the smallest directly executable units.

[0073] A knowledge graph is a pre-built structured knowledge network in the financial field, including core knowledge such as financial business processes, task decomposition rules, and data relationships.

[0074] An atomic operation sequence is a sequence of indivisible smallest execution units arranged in logical order. Each unit corresponds to a specific execution action and has no subdivision space.

[0075] Specifically, the domain parsing layer loads a pre-built financial domain knowledge graph and receives the task tree and global task decomposition path. Based on the task decomposition rules in the knowledge graph, each subtask is decomposed into several minimum execution actions (atomic operations). Then, in combination with the execution order of the global task decomposition path, all atomic operations are logically sorted to ensure that the results of the preceding atomic operations can support the execution of the following operations. Finally, an ordered sequence of atomic operations is output.

[0076] For example, the domain analysis layer, combined with the financial credit risk assessment knowledge graph, breaks down the subtasks of the "Customer A's Credit Risk Analysis and Report Generation" task tree: "Customer A's basic information query" is broken down into "calling the customer information query interface → extracting name / qualification / credit information"; "public opinion information collection and analysis" is broken down into "calling the public opinion data interface → filtering negative public opinion → assessing the public opinion risk level"... Finally, combined with the global path, an atomic operation sequence is generated: "calling the customer information query interface → extracting customer A's basic information → calling the public opinion data interface → filtering negative public opinion about customer A → assessing customer A's public opinion risk level → calling the transaction data interface → statistically analyzing customer A's transaction data for the past 3 years → analyzing the rationality of customer A's transactions → assessing the credit risk level → integrating information and writing a report."

[0077] Optionally, the domain semantic relationships in the knowledge graph can be dynamically updated by integrating the business rule base with historical service call logs;

[0078] Adjust the service matching weights in the knowledge graph based on user feedback data.

[0079] Specifically, the system periodically reads new / revised rules from the business rule base (such as adding public opinion data verification rules for credit risk analysis), and extracts valid information from historical service call logs (such as the call association between frequently occurring "transaction behavior analysis" and "credit calculation interface"). Through a semantic alignment algorithm, the system transforms the rule content and log association information into standardized domain semantic relationships, replacing or supplementing them in the knowledge graph to achieve dynamic updates of domain semantic relationships. It also collects user feedback data (such as ratings / comments on service matching accuracy and execution efficiency). Through data cleaning and quantitative analysis, it determines the service matching scenario corresponding to the feedback. If the feedback shows a large matching error between a service and a specific task, the matching weight of that service is lowered; if the feedback is good, the weight is raised, thus completing the dynamic adjustment of service matching weights in the knowledge graph.

[0080] For example, in the business scenario of "analyzing customer A's credit risk and generating a due diligence report," the business rule base adds a rule "credit risk analysis requires supplementing the customer's supply chain transaction data verification for the past 6 months." Historical service call logs show multiple associated call records for "supply chain transaction data verification" and "supply chain finance system interface." The system uses semantic alignment to supplement the knowledge graph with the semantic relationship between "credit risk analysis—supply chain transaction data verification—supply chain finance system interface." After the update, the task tree decomposition will automatically add a sub-task of "customer A's supply chain transaction data query" and associate it with the corresponding interface. If multiple users report that "the public opinion conclusions for customer A generated by the ordinary public opinion analysis business processing model have significant deviations, while the professional financial public opinion business processing model is more accurate," the system quantitatively analyzes this feedback and increases the matching weight between the "public opinion information analysis" task and the "professional financial public opinion business processing model" in the knowledge graph from 0.6 to 0.9, while simultaneously decreasing the matching weight of the ordinary public opinion analysis business processing model from 0.4 to 0.1.

[0081] S205: At the tool execution layer, the atomic operation sequence of multi-service processing models and existing service processing is scheduled through a semantic-level service discovery mechanism; the semantic-level service discovery mechanism refers to dynamically matching and calling the required services through vectorized representation of service description text and intent matching.

[0082] The tool execution layer is the execution implementation module of the layered collaboration framework, responsible for the scheduling and execution of specific operations.

[0083] The semantic-level service discovery mechanism is a service matching and invocation framework based on semantic parsing. It transforms the intent of atomic operations into vector representations, which are then matched with the descriptive text vectors of multi-business processing models and existing services to achieve dynamic and accurate invocation.

[0084] Specifically, after receiving the atomic operation sequence, the tool execution layer performs semantic parsing on each atomic operation to extract the core execution intent and the required resource type. Through a semantic-level service discovery mechanism, the operation intent is vectorized and matched with the functional description vectors of multi-business processing models (such as data analysis business processing models and report writing business processing models) and the service description vectors of existing service interfaces (such as customer information interfaces and public opinion interfaces) in the system. After a successful match, a scheduling instruction is generated, which triggers the corresponding multi-business processing models or existing services to start in sequence according to the order of the atomic operation sequence, executes the specific operation, and obtains the execution result.

[0085] For example, for a sequence of atomic operations, the tool execution layer first parses the semantic intent of the first operation, "call the customer information query interface," as "obtain basic information about customer A." Through the semantic-level service discovery mechanism, it matches this intent vector with the service description vector of the "customer information query interface" (matching degree 98%), triggering the call to this existing service. It then parses the intent of "analyze the rationality of customer A's transaction" as "transaction data modeling and analysis," matches it with the functional description vector of the "data analysis business processing model," and schedules the startup of this business processing model. Finally, it completes the scheduling of all atomic operations in sequence, obtaining the execution results of each operation (such as basic information, public opinion risk level, and transaction analysis conclusions).

[0086] Optionally, a semantic-level service discovery mechanism can be used to schedule atomic operation sequences for multi-service processing models and existing service processing. Specific implementation methods include:

[0087] The service description text of the atomic operation sequence is vectorized using a model for extracting semantic features from the service description text.

[0088] Based on the knowledge graph matching service, the matching results are obtained by describing the text and the intent of the task requirements.

[0089] The matching results are converted into requests adapted to heterogeneous interface protocols and the corresponding existing services are invoked.

[0090] Specifically, firstly, the system invokes a service description text semantic feature extraction model to perform semantic parsing on the service requirement description text corresponding to each operation in the atomic operation sequence, extracting core semantic features and completing vectorized representation to obtain semantic vectors that can be used to calculate the matching degree. Next, based on a pre-built financial domain knowledge graph, the generated semantic vectors are compared with the semantic description vectors of multi-business processing models and existing services stored in the knowledge graph. By calculating vector similarity, the service description text and task requirement intent are accurately matched, and the matching result, including the matched service type, matching degree, and call priority, is output. Finally, the system invokes a protocol adaptation module to convert the matching result in a unified format into a call request that conforms to the protocol specification, based on the heterogeneous interface protocol type of the target service in the matching result. The call is initiated through a standardized interface channel, while the call status and return results are monitored to ensure the smooth execution of the atomic operations.

[0091] For example, in the business scenario of "analyzing customer A's credit risk and generating a due diligence report," for the "customer A's transaction behavior data analysis" operation in the atomic operation sequence: First, the service description text semantic feature extraction model parses the service requirement description of this operation, "obtain customer A's transaction data for the past 3 years and analyze its rationality," extracting core semantic features such as "customer A," "transaction data for the past 3 years," and "rationality analysis," generating corresponding semantic vectors; then, based on a financial domain knowledge graph, this semantic vector is compared with the semantic description vector of the "data analysis business processing model" in the knowledge graph (including related semantics such as "transaction data processing" and "risk-related feature analysis") and the "transaction data platform"... The semantic description vector of the "Transaction Data Platform Interface" is compared to obtain the matching result of "Data Analysis Business Processing Model (matching degree 95%, priority 1) + Transaction Data Platform Interface (matching degree 92%, priority 2)". Finally, the system recognizes that the "Transaction Data Platform Interface" uses the HTTP protocol and the "Data Analysis Business Processing Model" uses the internal RPC protocol. The protocol adaptation module converts the matching result into a transaction data query request that conforms to the HTTP protocol and a data analysis instruction that conforms to the RPC protocol, respectively. The calls are initiated in sequence. First, the transaction data of customer A for the past 3 years is obtained through the transaction data platform interface. Then, the data analysis business processing model completes the data rationality analysis and outputs the analysis results.

[0092] The financial business processing method provided in this embodiment receives business requirement information input by the user. This information is a financial business processing request submitted by the user via natural language, image, and / or voice. A multimodal fusion algorithm is used to parse the business requirement information and generate a task tree. The multimodal fusion algorithm combines text, image, and voice parsing algorithms. At the strategy planning layer, a reinforcement learning model allocates resources and generates a global task decomposition path. At the domain parsing layer, a knowledge graph is pre-set to decompose the task tree into atomic operation sequences. At the tool execution layer, a semantic-level service discovery mechanism schedules multiple business processing models and existing service processing atomic operation sequences. The semantic-level service discovery mechanism refers to dynamically matching and invoking the required services through vectorized representations of service description text and intent matching. This method achieves intelligent, efficient, and automated processing of financial business requirements, improving business response efficiency and execution accuracy.

[0093] Figure 3 A schematic diagram of the financial transaction processing device provided in this application is shown below. Figure 3 As shown, the financial transaction processing device 300 provided in this embodiment includes:

[0094] The receiving module 301 is used to receive business requirement information input by the user, which is a financial business processing request submitted by the user through natural language, image and / or voice.

[0095] Parsing module 302 is used to parse business requirement information into a task tree, which is a hierarchical task structure;

[0096] Processing module 303 is used to process the task tree based on a hierarchical collaboration framework to complete the execution of business requirements. The hierarchical collaboration framework includes a strategy planning layer, a domain resolution layer, and a tool execution layer. The strategy planning layer is used to allocate resources in task processing. The domain resolution layer is used to decompose the task tree into atomic operation sequences. The tool execution layer is used to call multiple business processing models and existing services to process atomic operation sequences. The atomic operation sequence is the smallest indivisible execution unit sequence. Existing services include the existing business system interfaces of financial institutions.

[0097] In one possible implementation, the financial business processing device 300 further includes: a generation module 305 and a disassembly module 306;

[0098] The generation module 305 is used to allocate resources and generate global task decomposition paths through a reinforcement learning model at the policy planning layer.

[0099] The decomposition module 306 is used to decompose the task tree into atomic operation sequences in the domain parsing layer in conjunction with the knowledge graph. The domain parsing layer is pre-set with a knowledge graph.

[0100] The processing module 303 is also used to schedule the atomic operation sequence of multi-service processing models and existing service processing through a semantic-level service discovery mechanism at the tool execution layer. The semantic-level service discovery mechanism refers to dynamically matching and calling the required services through vectorized representation of service description text and intent matching.

[0101] In one possible implementation, the financial business processing device 300 further includes: a determination module 307 and a calling module 308;

[0102] Processing module 303 is used to vectorize the service description text of the atomic operation sequence using a model for extracting semantic features of the service description text;

[0103] The determination module 307 is used to obtain the matching result by matching the description text with the intent of the task requirements based on the knowledge graph matching service;

[0104] Module 308 is invoked to convert the matching results into requests adapted to heterogeneous interface protocols and invoke the corresponding existing services.

[0105] In one possible implementation, the determining module 307 is further configured to dynamically update the parameters of the reinforcement learning model through a reward function. The input features of the reinforcement learning model include system resource status and task execution status, which serve as model input features.

[0106] In one possible implementation, the financial transaction processing device 300 further includes: an update module 309 and an adjustment module 310;

[0107] Update module 309 is used to dynamically update the domain semantic relationships in the knowledge graph by integrating the business rule base and historical service call logs;

[0108] Adjustment module 310 is used to adjust the service matching weights in the knowledge graph based on user feedback data.

[0109] In one possible implementation, the generation module 305 is further configured to parse the business requirement information using a multimodal fusion algorithm to generate a task tree; the multimodal fusion algorithm is a parsing algorithm that combines text, image, and speech.

[0110] In one possible implementation, the determining module 307 is also used in the knowledge question-answering business processing model to answer business knowledge questions;

[0111] The determination module 307 is also used for data analysis business processing models, for analyzing customer profiles and transaction behavior;

[0112] The determination module 307 is also used in the process automation business processing model to execute business process operations.

[0113] The financial business processing device provided in this embodiment can execute the methods provided in the above method embodiments. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0114] Figure 4 A schematic diagram of the financial transaction processing equipment provided in this application. Figure 4 As shown, the electronic device of this embodiment may include: at least one processor 401; and a memory 402 communicatively connected to the at least one processor; wherein the memory 402 stores instructions that can be executed by the at least one processor 401, and the instructions are executed by the at least one processor 401 to cause the electronic device to perform the method as described in any of the above embodiments.

[0115] Optionally, the memory 402 can be either standalone or integrated with the processor 401. When the memory 402 is set up independently, the device also includes a bus for connecting the memory 402 and the processor 401.

[0116] The implementation principle and technical effects of the electronic device provided in this embodiment can be found in the foregoing embodiments, and will not be repeated here.

[0117] This application also provides a computer-readable storage medium storing computer-executable instructions. When the computer-executable instructions are executed by a processor, the methods provided in any of the foregoing embodiments can be implemented.

[0118] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the method provided in any of the foregoing embodiments.

[0119] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0120] It should be further noted that although the steps in the flowchart are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0121] It should be understood that the above-described device embodiments are merely illustrative, and the device of this application can also be implemented in other ways. For example, the division of units / modules in the above embodiments is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple units, modules, or components may be combined, or integrated into another system, or some features may be ignored or not executed.

[0122] Furthermore, unless otherwise specified, the functional units / modules in the various embodiments of this application can be integrated into one unit / module, or each unit / module can exist physically separately, or two or more units / modules can be integrated together. The integrated units / modules described above can be implemented in hardware or as software program modules.

[0123] Unless otherwise specified, the processor can be any suitable hardware processor, such as a CPU, GPU, FPGA, DSP, and ASIC. Unless otherwise specified, the storage unit can be any suitable magnetic or magneto-optical storage medium, such as resistive random access memory (RRAM), dynamic random access memory (DRAM), static random access memory (SRAM), enhanced dynamic random access memory (EDRAM), high-bandwidth memory (HBM), hybrid memory cube (HMC), etc.

[0124] If the integrated unit / module is implemented as a software program module and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0125] In the above embodiments, the descriptions of each embodiment have their own emphasis. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0126] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0127] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method of processing financial transactions, characterized by, Applied to the business processing end, the method includes: Receive business requirement information input by the user, wherein the business requirement information is a financial business processing request submitted by the user through natural language, image and / or voice; The business requirement information is parsed into a task tree, which is a hierarchical task structure; The task tree is processed based on a hierarchical collaboration framework to complete the execution of business requirements. The hierarchical collaboration framework includes a strategy planning layer, a domain resolution layer, and a tool execution layer. The strategy planning layer is used to allocate resources in task processing. The domain resolution layer is used to decompose the task tree into atomic operation sequences. The tool execution layer is used to call multiple business processing models and existing services to process the atomic operation sequences. The atomic operation sequences are indivisible sequences of the smallest execution units. The existing services include the business system interfaces of financial institutions.

2. The method of claim 1, wherein, The processing of the task tree based on the hierarchical collaboration framework includes: In the strategy planning layer, resources are allocated through a reinforcement learning model to generate global task decomposition paths; In the domain parsing layer, the task tree is decomposed into the atomic operation sequence in conjunction with the knowledge graph, and the domain parsing layer is pre-set with the knowledge graph; At the tool execution layer, the atomic operation sequence is processed by multiple service processing models and existing services through a semantic-level service discovery mechanism. The semantic-level service discovery mechanism refers to dynamically matching and invoking the required service through vectorized representation of service description text and intent matching.

3. The method according to claim 2, characterized in that, The process of scheduling the atomic operation sequence through the semantic-level service discovery mechanism to handle multiple service processing models and existing services includes: The service description text of the atomic operation sequence is vectorized and represented using a model for extracting semantic features from the service description text. Based on the knowledge graph matching service, the matching results are obtained by describing the text and the intent of the task requirements. The matching result is converted into a request adapted to the heterogeneous interface protocol and the corresponding existing service is invoked.

4. The method according to claim 2 or 3, characterized in that, The reinforcement learning model dynamically updates its parameters through a reward function. The input features of the reinforcement learning model include system resource status and task execution status, which serve as model input features.

5. The method according to claim 2 or 3, characterized in that, The method further includes: By integrating the business rule base with historical service call logs, the domain semantic relationships in the knowledge graph are dynamically updated; The service matching weights in the knowledge graph are adjusted based on user feedback data.

6. The method according to any one of claims 1-3, characterized in that, The step of parsing the business requirement information into a task tree includes: The business requirement information is parsed using a multimodal fusion algorithm to generate the task tree; the multimodal fusion algorithm is a parsing algorithm that combines text, image, and speech.

7. The method according to any one of claims 1-3, characterized in that, The multiple business processing models include at least: A knowledge-based question-answering model is used to answer business knowledge questions. Data analytics models are used to analyze customer profiles and transaction behavior; A process automation model is used to execute business process operations.

8. A financial transaction processing device, characterized in that, include: The receiving module is used to receive business requirement information input by the user, which is a financial business processing request submitted by the user through natural language, images and / or voice. The parsing module is used to parse the business requirement information into a task tree, wherein the task tree is a hierarchical task structure; The processing module is used to process the task tree based on a hierarchical collaboration framework to complete the execution of business requirements. The hierarchical collaboration framework includes a strategy planning layer, a domain resolution layer, and a tool execution layer. The strategy planning layer is used to allocate resources in task processing. The domain resolution layer is used to decompose the task tree into atomic operation sequences. The tool execution layer is used to call multiple business processing models and existing services to process the atomic operation sequences. The atomic operation sequences are indivisible sequences of the smallest execution units. The existing services include the business system interfaces of financial institutions.

9. An electronic device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-7.

11. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method described in any one of claims 1-7.

Images