Data system processing method, device, equipment and program product

Abstract

The invention discloses a data system processing method and device, equipment and a program product, and belongs to the field of computers. The target application system is disassembled into a plurality of sub-tasks based on the disassembly requirement of the target application system, a model combination matched with the sub-tasks is selected from a model pool according to the types of the sub-tasks to be analyzed to obtain formatted data, and the formatted data comprises at least one of a table association relation, code logic and a rule chain; calling the knowledge graph engine, and constructing a self-updating knowledge graph of the target application system based on the table structure, the business rule and the business event flow of the target application system; and performing semantic alignment on the knowledge graph and the formatted data, and constructing a multi-modal reasoning chain and an executable strategy based on the aligned data so as to drive code development of an alternative system of the target application system.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a data system processing method, apparatus, equipment and program product. Background Technology

[0002] Currently, the development of alternative systems heavily relies on the manual dismantling of existing systems, resulting in extremely low development efficiency. Taking Enterprise Resource Planning (ERP) systems as an example, their localization design heavily depends on the manual dismantling and code development of mature foreign systems such as Oracle ERP. The core bottlenecks can be summarized as low efficiency of manual dismantling and rigid static business rules. Therefore, there is an urgent need for an efficient and high-quality alternative system development methodology. Summary of the Invention

[0003] The purpose of the embodiments in this specification is to provide a data system processing method, apparatus, device, and program product to solve the problem of low development efficiency when replacing existing business systems in the prior art.

[0004] To achieve the above objectives, the embodiments in this specification adopt the following technical solutions: In a first aspect, an embodiment of the present invention provides a data system processing method, the method comprising: The target application system is decomposed into multiple sub-tasks based on its decomposition requirements. A model combination matching the multiple sub-tasks is selected from the model pool according to the type of the sub-tasks and parsed to obtain formatted data. The formatted data includes at least one of table relationships, code logic, and rule chains. The knowledge graph engine is invoked to construct a self-updating knowledge graph of the target application system based on the table structure, business rules, and business event flow of the target application system. The knowledge graph and the formatted data are semantically aligned, and a multimodal reasoning chain and an executable strategy are constructed based on the aligned data to drive the code development of an alternative system to implement the target application system.

[0005] Optionally, before the decomposition of the target application system's requirements into multiple sub-tasks, and the selection of a model combination matching the multiple sub-tasks from the model pool according to the type of the sub-tasks for parsing and processing to obtain formatted data, the method further includes: constructing a model pool, pre-training each model in the model pool, labeling the capability matrix, and generating an index table based on the capability matrix; wherein, the model pool includes code parsing models, financial and tax rule models, and multimodal models.

[0006] Optionally, the step of selecting a combination of models matching the multiple subtasks from the model pool based on the type of the subtask for parsing and processing to obtain formatted data includes: Based on the type of the subtask, a combination of models matching the multiple subtasks is selected from the model pool for parsing and processing; Conflict detection is performed on the first processing result output by the parsing process, and if a conflict is determined to exist, conflict resolution is performed on the first processing result output by the parsing process based on the confidence level corresponding to the first processing result output by the parsing process and / or the authority weight of the target data corresponding to the subtask. The second processing result output after the conflict resolution process is formatted and converted to obtain the formatted data.

[0007] Optionally, the semantic alignment of the knowledge graph and the formatted data includes: The knowledge graph is vectorized to obtain a vector of the knowledge graph; The vectors of the knowledge graph and the formatted data are mapped to the same vector space according to the condition of minimizing cosine similarity and contrast loss, so as to perform semantic alignment.

[0008] Optionally, after semantically aligning the knowledge graph and the formatted data, and constructing a multimodal inference chain and an executable strategy based on the aligned data to drive the code development of an alternative system to implement the target application system, the method further includes: After the alternative system is started, user behavior data and expert behavior data in the alternative system are collected, and the corresponding models and knowledge graphs in the model pool are dynamically updated based on the user behavior data and expert behavior data.

[0009] Optionally, the process of collecting user behavior data and expert behavior data from the alternative system, and dynamically updating the corresponding models and knowledge graphs in the model pool based on the user behavior data and expert behavior data, includes: The knowledge graph is updated based on the expert behavior data; A structural equation model was constructed based on the user behavior data to perform root cause analysis. Based on the results of the root cause analysis and the updated knowledge graph, the corresponding models in the model pool are updated.

[0010] In a second aspect, embodiments of the present invention provide a data system processing apparatus, the apparatus comprising: The hybrid model pool scheduling module is used to decompose the decomposition requirements of the target application system into multiple sub-tasks, and select model combinations that match the multiple sub-tasks from the model pool according to the type of the sub-tasks for parsing and processing to obtain formatted data. The formatted data includes at least one of table association, code logic, and rule chain. The knowledge graph construction module is used to call the knowledge graph engine to construct a self-updating knowledge graph of the target application system based on the table structure, business rules and business event flow of the target application system. The execution strategy generation module is used to semantically align the knowledge graph and the formatted data, and construct a multimodal inference chain and an executable strategy based on the aligned data to drive the code development of an alternative system to implement the target application system.

[0011] Thirdly, embodiments of the present invention provide an electronic device, including: a processor; and a memory arranged to store computer-executable instructions, which, when executed, cause the processor to perform the steps of the above-described method.

[0012] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, can implement the steps of the above-described method.

[0013] Fifthly, embodiments of the present invention provide a computer program product, the computer program product including a computer-readable storage medium storing a computer program operable to cause a computer to perform the steps of the method described above.

[0014] The above-described technical solutions adopted in the embodiments of this application can achieve the following beneficial effects: This specification provides a data system processing method, apparatus, device, and program product. The method significantly improves decomposition efficiency through the dynamic scheduling and combination of multiple heterogeneous models, standardizes the mapping of various types of data, and triggers attribute updates of the graph through real-time monitoring of event streams. The alignment of cross-modal semantics improves the accuracy of strategy generation, greatly enhancing the development efficiency of alternative systems. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 This is a flowchart illustrating a data system processing method according to this application; Figure 2 This is a flowchart illustrating the overall process of a data system processing method according to this application; Figure 3 This is a schematic diagram of the structure of a data system processing device according to this application; Figure 4 This is a schematic diagram of the structure of an electronic device according to this application. Detailed Implementation

[0016] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0017] To address the problems existing in the prior art, embodiments of this application provide a data system processing method. The executing entity of this method can be a terminal device or a server. The terminal device can be a mobile terminal device such as a mobile phone, tablet computer, or smartwatch, or a terminal device such as a computer. The server can be an independent server or a server cluster composed of multiple servers. Figure 1 As shown, the method specifically includes steps 101 to 103.

[0018] S101, based on the decomposition requirements of the target application system, the system is decomposed into multiple sub-tasks, and a model combination matching the multiple sub-tasks is selected from the model pool according to the type of the sub-tasks for parsing and processing to obtain formatted data. The formatted data includes at least one of table association, code logic, and rule chain.

[0019] In this application embodiment, the target application system is not limited and can be any kind of application system. For example, it can be a certain industrial system software, such as an electronic design automation system, a computer-aided design system, etc. Another example is an Enterprise Resource Planning (ERP) system, etc. Taking an ERP system as an example, an ERP system is a large-scale integrated business system containing numerous sub-modules. In the context of domestic substitution of foreign business systems, this application's technical solution is described using the domestic substitution of relevant sub-modules in an ERP system as an example, such as the substitution of the finance and tax module in an ERP system.

[0020] In one embodiment of this application, a domain-specific hybrid model pool can be pre-constructed before step 101. Employing a domain-specific hybrid model pool enables a dynamic collaborative scheduling mechanism, achieving task allocation and result fusion. The dynamic scheduling of the domain-specific hybrid model pool significantly improves decomposition efficiency.

[0021] Optionally, the model pool contains three types of specialized models: (1) Code parsing model (CodeLlama-34B): It can be finely tuned based on the LLaMA architecture, and optimizes parameters for Oracle PL / SQL syntax features such as foreign key constraints and nested logic of stored procedures. It supports reverse engineering and code logic extraction of languages ​​such as Java and Python, and outputs code logic.

[0022] (2) Tax and Finance Rules Model (CaiWuBERT): The RoBERTa framework can be used to parse policy documents and output rule chains. Secondary pre-training is performed on tax and finance corpora, State Taxation Administration announcements, and listed company financial reports. Domain dictionaries such as "input tax credit transfer" and "additional deduction" are embedded to accurately identify policy conditions such as "small-scale taxpayers with monthly sales of ≤100,000 yuan are exempt from value-added tax" with a recall rate of ≥90%.

[0023] (3) Multimodal Model (Pix2Struct): Integrates visual Transformer and attention mechanism to parse unstructured data such as Oracle interface prototype diagrams and ER diagrams, and generate standardized JSON descriptions, such as field layout and interaction logic. Of course, the model pool may also include other models besides the three types of dedicated models mentioned above, and this application embodiment does not limit this.

[0024] Optionally, in some embodiments of this application, a combination of models matching the plurality of subtasks is selected from the model pool according to the type of the subtask for parsing and processing to obtain formatted data, including: Based on the type of the subtask, a combination of models matching the multiple subtasks is selected from the model pool for parsing and processing; Conflict detection is performed on the first processing result output by the parsing process, and if a conflict is determined to exist, conflict resolution is performed on the first processing result output by the parsing process based on the confidence level corresponding to the first processing result output by the parsing process and / or the authority weight of the target data corresponding to the subtask. The second processing result output after the conflict resolution process is formatted and converted to obtain the formatted data.

[0025] For ease of understanding, this application example uses the decomposition of the financial module in an Enterprise Resource Planning (ERP) system into multiple sub-tasks as an example.

[0026] For example, the finance module can be divided into three subtasks: T1, T2, and T3. Subtask T1 parses the relationship between AP_INVOICES_ALL (Accounts Payable Invoices) and GL_JE_LINES (General Ledger Voucher Lines), relying on PL / SQL foreign key constraints and ER diagram semantics. Subtask T2 extracts the business process rules for "Prepayment Reconciliation," relying on user manual text analysis and interface operation logs. Subtask T3 matches Chinese local tax and financial compliance requirements, such as the "Prepayment - Invoice Reconciliation" rules, relying on policy document parsing and a historical case library. Of course, in real-world applications, there may be more complex business processes; the examples listed here are merely illustrative.

[0027] Based on the types of the three sub-tasks, a matching model can be selected from the model pool for processing to obtain the first processing result. For example, a code parsing model (CodeLlama-34B) and / or a multimodal model (Pix2Struct) can be matched for sub-task T1 to obtain code logic and / or structured data; a tax and finance rule model (CaiWuBERT) can be matched for sub-task T2 to obtain a rule chain; a tax and finance rule model (CaiWuBERT) and / or a multimodal model (Pix2Struct) can be matched for sub-task T3 to obtain a rule chain and / or structured data, and so on.

[0028] Furthermore, after obtaining the first processing result, it can be further aligned. Optionally, when aligning the data results, each model independently and in parallel processes its corresponding subtasks, and the output of each first processing result is converted into a unified format by an intermediate representation layer, such as using a JSON structure for code logic and a Drools rule chain for policy rules; then, conflict resolution processing can be performed on the unified formatted first processing results. Optionally, when conflicts occur between the first processing results output by the aforementioned models, a majority voting mechanism can be used to resolve the conflicts. For example, when the foreign key relationship parsed by CodeLLama-34B is inconsistent with the semantics of the ER graph extracted by Pix2Struct, the result with higher confidence score, such as model output probability ≥ 85%, or higher authority weight, and ER graph > code annotation, is given priority. The final result is as follows: Final result = argmax ; in, The confidence weight of the i-th model, such as when the CodeLlama parsing accuracy is ≥95%. =0.95; : The output of the i-th model; (·): Indicator function, returns 1 if the result matches, otherwise returns 0.

[0029] To provide a more intuitive understanding of the specific implementation of step 101, an illustrative processing flow can be shown below: (1) Model pool initialization: Load the pre-trained model parameters, label the capability matrix of each model, such as "CodeLlama code parsing accuracy ≥ 95%", "CaiWuBERT policy recall rate ≥ 90%", etc., and build an index table.

[0030] (2) Dynamic task allocation: After inputting the decomposition requirements, the optimal model combination can be matched from the capability matrix index based on the type of atomic subtask. For example, the aforementioned subtask T1 can be assigned to CodeLlama and Pix2Struct, subtask T2 to CaiWuBERT, subtask T3 to Llama-3, and so on.

[0031] (3) Parallel processing and result fusion: Each subtask's corresponding model processes its own subtask independently and in parallel, then outputs the processing result (i.e., the first processing result mentioned above), which is then converted into a unified format by an intermediate representation layer. For example, code logic is converted into a JSON structure, policy rules are converted into Drools rule chains, and so on. Then, for conflicting results in the unified formatted processing results, a majority voting mechanism is used to determine the final result, i.e., the second processing result mentioned above. Preferably, the models in the model pool can also be lightweight optimized, for example, by using model pruning and quantization techniques to compress the number of model parameters by 30%-50%, reducing GPU computing power requirements and supporting edge deployment.

[0032] S102, invoke the knowledge graph engine to construct a self-updating knowledge graph of the target application system based on the table structure, business rules and business event flow of the target application system.

[0033] Table structure is a crucial source of the self-updating knowledge graph of the target application system, belonging to structured data. Optionally, database connections can be made via JDBC or other database interfaces to automatically extract table structures (field names, data types, primary and foreign key relationships) and import them into a graph database according to the "entity-relationship-attribute" schema. During the extraction of one table structure, ETL tools, such as Apache NiFi, can be used to import the data into Neo4j in RDF triple format, preserving the original semantics, such as field constraints and indexing strategies. Let's take the finance module in an ERP system as an example. For instance, the AP_INVOICES_ALL table can be mapped to the "Accounts Payable Invoices" node, with attributes including INVOICE_ID (primary key), AMOUNT (amount), and VENDOR_ID (supplier ID), and a "issued by _" relationship can be established with the "Supplier" node.

[0034] Business rules are another important source of self-updating knowledge graphs for target application systems, and are generally unstructured data. Optionally, business rules in the user manual can be converted into rule chains such as Drools rule files and then embedded into the "approval policy" node of the knowledge graph to support dynamic policy execution.

[0035] The sources of business rules mainly include policy text parsing and log behavior mining.

[0036] For policy text parsing, a predefined Prompt template can be used to batch process the data using CaiWuBERT to generate structured rule chains. For example, a predefined Prompt template (role: tax expert; task: extract policy name, effective date, applicable conditions, and execution action) can be used to process the State Taxation Administration's PDF file using the CaiWuBERT tool, resulting in a rule chain such as "New policy for additional deduction of R&D expenses → Effective date 2024-01-01 → Applicable industry = manufacturing → Execution action: additional deduction ratio = 120%).

[0037] For log behavior mining, analysis tools can be used to analyze user operation logs, extract high-frequency business patterns, and abstract them into "best practice" nodes. For example, LDA topic model analysis tools can be used to analyze logs to extract the business pattern of "month-end closing process," and then abstract it into a rule chain of "automatic verification of voucher integrity before closing."

[0038] Business event streams are also an important source of self-updating knowledge graphs for target application systems. Optionally, business event streams can be monitored, and the monitored event data can be converted into graph update instructions to update graph attributes, ensuring data real-time performance.

[0039] For example, Apache Kafka can be used to listen to ERP business event streams (such as "invoice approval completed" and "general ledger voucher posting") and convert the listened event data into graph update instructions such as Cypher instructions. For instance, the "invoice-approval status=approved" attribute in the graph can be updated via the "invoice approval completed" event. In some application scenarios, the latency is less than 1 second, or even less.

[0040] Of course, it should be understood that knowledge conflicts may occur when constructing the knowledge graph. A multi-level strategy can be chosen to resolve knowledge conflicts. Generally, this multi-level strategy needs to consider time priority and authority weight. For time priority, the latest effective rules need to be retained; for example, the 2024 VAT rate covers the 2023 version. For authority weight, policies issued by more authoritative departments have higher weight; for example, documents from the Ministry of Finance have higher priority than local policies.

[0041] Alternatively, one priority formula is as follows: Rule priority = α * time decay factor + β * authority weight; Where α and β are adjustment coefficients, and the time decay factor is... λ represents the undecayed rate. This is the current time.

[0042] In one optional parameter configuration, α=0.7, β=0.3, the weight of the Ministry of Finance document is 1.0, and the weight of local policy is 0.6.

[0043] Furthermore, manual review can be triggered in situations where priorities cannot be determined, priorities are the same, or core processes have changed. For example, changes to core processes (such as adjustments to the approval chain) can trigger a manual confirmation process. Introducing a manual review mechanism, in addition to time priority and authority weighting, can further ensure system stability.

[0044] Of course, version management of the knowledge graph can also be implemented, with timestamps and version tags attached to each graph node, such as "Policy Version v2.1.3", supporting historical version tracing and one-click rollback.

[0045] This invention achieves dynamic rule updates and version management by constructing a knowledge graph and combining a three-tiered conflict resolution strategy of time priority, authority weight, and manual review.

[0046] S103, semantically align the knowledge graph and the formatted data, and construct a multimodal reasoning chain and an executable strategy based on the aligned data to drive the code development of an alternative system to implement the target application system.

[0047] It should be understood that the knowledge graph and the formatted data can be semantically aligned in various ways. Optionally, one way to semantically align the knowledge graph and the formatted data is as follows: The knowledge graph is vectorized to obtain a vector of the knowledge graph; The vectors of the knowledge graph and the formatted data are mapped to the same vector space according to the condition of minimizing cosine similarity and contrast loss, so as to perform semantic alignment.

[0048] It should be understood that vectorized knowledge graph vectors can represent the topological relationships between knowledge graph nodes. Optionally, various algorithms can be used to encode the knowledge graph during vectorization. For example, the GraphSAGE algorithm can be used to encode the knowledge graph, generating a 256-dimensional vector representation, and so on. This vector can represent the topological relationships between nodes; taking the financial module in an ERP system as an example again, it can represent the semantic association of "accounts payable invoice → supplier," and so on.

[0049] It should be understood that generative adversarial networks (GANs) can be used to achieve semantic alignment. Optionally, a GAN can be designed to map knowledge graph vectors to formatted data embeddings represented by 4096-dimensional text in Llama-3 to the same vector space, minimizing cosine similarity and contrastive loss to ensure semantic consistency. For example, the node “Accounts Payable Invoices” is highly correlated with the text description “AP_INVOICES_ALL Table”, and semantic alignment can be performed.

[0050] For ease of understanding, the loss function for training a generative adversarial model applicable to embodiments of this application is as follows: ; The formula for cosine similarity loss is: =1- ; The formula for contrast loss is: =-log ; in, Knowledge graph vectors generated for GraphSAGE; The text embedding vectors generated for LLama-3, i.e. the formatted data mentioned earlier; is the temperature coefficient, defaulting to 0.05; K is the number of negative samples. By training the generative adversarial model above to align the knowledge graph with the text semantic space, cross-modal reasoning consistency can be ensured.

[0051] Furthermore, constructing multimodal inference chains and executable policies based on aligned data can include: Based on the aligned data, multiple models are invoked to output intermediate results, which include at least one of code logic, policy rules, and interface description. Based on the intermediate results, related nodes in the knowledge graph are retrieved to create a multimodal reasoning chain and generate an executable policy with constraints.

[0052] In one implementation, during specific strategy generation, based on the aforementioned aligned data, a multi-level supplier collaboration module is designed to trigger collaboration among relevant models in the hybrid model pool. For example, CodeLLama is used to parse the iSupplier module code of the finance module in the aforementioned ERP system to extract the logic of "supplier qualification review" and "order status synchronization"; CaiWuBERT is used to match the "supply chain risk assessment" and "domestic production ratio ≥ 60%" rules in the State-owned Assets Supervision and Administration Commission's policies; and LLama-3 is used to generate a design summary, such as "introducing blockchain to achieve qualification traceability," etc. Then, the best practice node for "supplier collaboration" in the knowledge graph is retrieved, associated with the "blockchain" technology node, and policy compliance is verified, such as marking the "domestic production ratio ≥ 60%" constraint, to generate a strategy with constraints.

[0053] Of course, furthermore, strategies can be converted into compilable code based on template engines such as Freemarker and deployed to an independent and controllable ERP system.

[0054] This invention employs cross-modal semantic alignment and joint reasoning techniques, which can improve the accuracy of policy generation.

[0055] Of course, it should be understood that after S103, the following steps are also included: after the alternative system is running, collecting user behavior data and expert behavior data in the alternative system, and dynamically updating the corresponding models and knowledge graphs in the model pool based on the user behavior data and expert behavior data.

[0056] It should be understood that user behavior data can be obtained through user behavior tracking. An SDK can be embedded in the interface of a self-controllable ERP system to record click flows, form submission duration, and abnormal events (such as data validation failures). It should be understood that expert behavior data can be obtained through an expert annotation platform. A visual interface should be designed for experts to annotate policy deficiencies (such as "rules for overseas subsidiaries not covered") and link them to specific business nodes.

[0057] It should be understood that dynamically updating the corresponding models and knowledge graphs in the model pool based on the user behavior data and expert behavior data may include: The knowledge graph is updated based on the expert behavior data; A structural equation model was constructed based on the user behavior data to perform root cause analysis. Based on the root cause analysis results and the updated knowledge graph, the corresponding models in the model pool are updated.

[0058] For user behavior data, user operation logs (clickstream, abnormal events) can be stored in Elasticsearch in real time for root cause analysis; for expert behavior data, expert-annotated data (such as policy defect markers) can be associated with knowledge graph nodes to form training samples.

[0059] Then, a structural equation model (SEM) can be built to analyze the path coefficients of variables and output a root cause report; the CodeLlama model can be incrementally trained and new rule data (such as "overseas subsidiary approval process") can be injected to improve the accuracy of strategy generation.

[0060] Of course, the adjustments to the knowledge graph and model mentioned above can be implemented using a hot update method, and the corresponding effects can be verified.

[0061] For example, update commands (such as Drools rule files and API routing policies) can be issued through the Nacos dynamic configuration center; some users can be assigned to the new policy group through A / B testing, key indicators (such as API success rate and inventory turnover rate) can be monitored, and full deployment can be carried out after the statistical significance reaches the target (p<0.05).

[0062] If the system malfunctions after the update (such as service unavailability), it will automatically revert to a previous version, and the fault point will be located through log analysis.

[0063] In addition, such as Figure 2 As shown in the embodiments of this specification, the overall flow of the data system processing technology solution based on hybrid model pools and knowledge graphs is also provided, as follows: (1) The user submits a decomposition requirement, and a hybrid large model scheduling module, i.e., a model pool, is constructed. Based on the user's submitted decomposition requirement, the task is decomposed into multiple sub-tasks. According to the type of the sub-task, the model combination matching the multiple sub-tasks is called from the model pool for parsing and processing. Then, the processing results are subjected to conflict detection and elimination. Finally, the output results of each model are aligned to obtain formatted data. The model pool applicable to the embodiments of this application may include CodeLlama parsing code model, CaiWuBERT parsing policy model, and Pix2Struct parsing chart model, corresponding to the specific manifestations of the above-mentioned code parsing model, financial and tax rule model, and multimodal model. (2) Call the dynamic knowledge graph engine to construct the knowledge graph based on structured data injection, unstructured data parsing, and real-time event updates. Structured data injection corresponds to the injection of the table structure mentioned above, unstructured data parsing corresponds to the parsing of the business rules mentioned above, and real-time event updates correspond to the updating of the business event flow mentioned above. (3) Based on the collaborative reasoning module, it can also be an execution strategy generation module, which semantically aligns the knowledge graph and the formatted data, and constructs a multimodal reasoning chain and an executable strategy based on the aligned data, generating cross-modal strategies to drive the code development and interface generation of the domestic ERP system; (4) In order to continuously update the system through closed-loop feedback, the feedback optimization module, also known as the feedback update module, collects user behavior data and expert behavior data from the domestic ERP system, updates the knowledge graph based on the expert behavior data, constructs a structural equation model based on the user behavior data to perform root cause analysis, and finally updates the corresponding model in the model pool based on the root cause analysis results and the updated knowledge graph.

[0064] The above describes the data system processing method based on a hybrid model pool and knowledge graph provided in the embodiments of this specification. Based on the same idea, the embodiments of this specification also provide a data system processing apparatus based on a hybrid model pool and knowledge graph, such as... Figure 3 As shown.

[0065] In a second aspect, embodiments of the present invention provide a data system processing apparatus, the apparatus comprising: The data system processing method apparatus based on hybrid model pool and knowledge graph includes: a hybrid model pool scheduling module 202, a knowledge graph construction module 204, an execution strategy generation module 206, and a feedback update module 208.

[0066] The hybrid model pool scheduling module 202 is used to decompose the decomposition requirements of the target application system into multiple sub-tasks, and select model combinations that match the multiple sub-tasks from the model pool according to the type of the sub-tasks for parsing and processing to obtain formatted data that can be recognized by the knowledge graph engine. The formatted data that can be recognized includes at least one of table association, code logic, and rule chain. The knowledge graph construction module 204 is used to call the knowledge graph engine to construct a self-updating knowledge graph of the target application system based on the table structure, business rules and business event flow of the target application system. The execution strategy generation module 206 is used to semantically align the knowledge graph and the formatted data, and construct a multimodal inference chain and an executable strategy based on the aligned data to drive the code development of an alternative system to implement the target application system. The feedback update module 208 is used to collect user behavior data and expert behavior data, and dynamically update the corresponding models and knowledge graphs in the model pool based on the user behavior data and expert behavior data.

[0067] The above are the data system processing devices provided in the embodiments of this specification. Based on the same idea, the embodiments of this specification also provide an electronic device, such as... Figure 4 As shown.

[0068] The electronic device can provide a terminal device or server, etc., for the above embodiments.

[0069] Electronic devices can vary considerably due to differences in configuration or performance, and may include one or more processors 301 and memory 302. Memory 302 may store one or more application programs or data. Memory 302 may be temporary or persistent storage. The application programs stored in memory 302 may include one or more modules (not shown in the figures), each module including a series of computer-executable instructions for the electronic device. Furthermore, processor 301 may be configured to communicate with memory 302 and execute the series of computer-executable instructions stored in memory 302 on the electronic device. The electronic device may also include one or more power supplies 303, one or more wired or wireless network interfaces 304, one or more input / output interfaces 305, and one or more keyboards 306.

[0070] Specifically, in this embodiment, the electronic device includes a memory and one or more programs, wherein one or more programs are stored in the memory, and one or more programs may include one or more modules, and each module may include a series of computer-executable instructions for use in the electronic device, and is configured to be executed by one or more processors. The one or more programs include computer-executable instructions for performing the following: The target application system is decomposed into multiple sub-tasks based on its decomposition requirements. A model combination matching the multiple sub-tasks is selected from the model pool according to the type of the sub-tasks and parsed to obtain formatted data. The formatted data includes at least one of table relationships, code logic, and rule chains. The knowledge graph engine is invoked to construct a self-updating knowledge graph of the target application system based on the table structure, business rules, and business event flow of the target application system. The knowledge graph and the formatted data are semantically aligned, and a multimodal reasoning chain and an executable strategy are constructed based on the aligned data to drive the code development of an alternative system to implement the target application system.

[0071] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the electronic device embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0072] Furthermore, based on the above Figures 1 to 4 The method shown in this specification, along with one or more embodiments, also provides a computer-readable storage medium for storing computer-executable instruction information. In one specific embodiment, the storage medium can be a USB flash drive, optical disc, hard disk, etc. When the computer-executable instruction information stored in the storage medium is executed by a processor, it can achieve the following process: The target application system is decomposed into multiple sub-tasks based on its decomposition requirements. A model combination matching the multiple sub-tasks is selected from the model pool according to the type of the sub-tasks and parsed to obtain formatted data. The formatted data includes at least one of table relationships, code logic, and rule chains. The knowledge graph engine is invoked to construct a self-updating knowledge graph of the target application system based on the table structure, business rules, and business event flow of the target application system. The knowledge graph and the formatted data are semantically aligned, and a multimodal reasoning chain and an executable strategy are constructed based on the aligned data to drive the code development of an alternative system to implement the target application system.

[0073] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the above-described storage medium embodiment is basically similar to the method embodiment, so the description is relatively simple; relevant parts can be referred to the description of the method embodiment.

[0074] Furthermore, based on the above Figures 1 to 4 The method shown in this specification, along with one or more embodiments, also provides a computer program product including a computer program that, when executed by a processor, performs the following process: The target application system is decomposed into multiple sub-tasks based on its decomposition requirements. A model combination matching the multiple sub-tasks is selected from the model pool according to the type of the sub-tasks and parsed to obtain formatted data. The formatted data includes at least one of table relationships, code logic, and rule chains. The knowledge graph engine is invoked to construct a self-updating knowledge graph of the target application system based on the table structure, business rules, and business event flow of the target application system. The knowledge graph and the formatted data are semantically aligned, and a multimodal reasoning chain and an executable strategy are constructed based on the aligned data to drive the code development of an alternative system to implement the target application system.

[0075] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the above-described embodiment of a computer program product is relatively simple in description because it is fundamentally similar to the method embodiment; relevant parts can be referred to the description of the method embodiment.

[0076] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0077] Those skilled in the art will understand that the embodiments of this specification can be provided as methods, systems, or computer program products. Therefore, this specification may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this specification may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0078] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0079] The above are merely embodiments of this specification and are not intended to limit the scope of this specification. Various modifications and variations can be made to this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this specification should be included within the scope of the claims of this specification. Furthermore, all other embodiments obtained by those skilled in the art without inventive effort should fall within the protection scope of this document.

Claims

1. A data system processing method, characterized in that, The method includes: The target application system is decomposed into multiple sub-tasks based on its decomposition requirements. A model combination matching the multiple sub-tasks is selected from the model pool according to the type of the sub-tasks and parsed to obtain formatted data. The formatted data includes at least one of table relationships, code logic, and rule chains. The knowledge graph engine is invoked to construct a self-updating knowledge graph of the target application system based on the table structure, business rules, and business event flow of the target application system. The knowledge graph and the formatted data are semantically aligned, and a multimodal reasoning chain and an executable strategy are constructed based on the aligned data to drive the code development of an alternative system to implement the target application system.

2. The method according to claim 1, characterized in that, Before the process of breaking down the target application system's requirements into multiple sub-tasks and selecting a model combination matching the multiple sub-tasks from the model pool according to the type of the sub-tasks for parsing and processing to obtain formatted data, the process further includes: constructing a model pool, pre-training each model in the model pool, labeling the capability matrix, and generating an index table based on the capability matrix; wherein, the model pool includes code parsing models, financial and tax rule models, and multimodal models.

3. The method according to claim 1, characterized in that, The step of selecting a combination of models matching the multiple subtasks from the model pool based on the type of the subtask, and then parsing and processing it to obtain formatted data includes: Based on the type of the subtask, a combination of models matching the multiple subtasks is selected from the model pool for parsing and processing; Conflict detection is performed on the first processing result output by the parsing process, and if a conflict is determined to exist, conflict resolution is performed on the first processing result output by the parsing process based on the confidence level corresponding to the first processing result output by the parsing process and / or the authority weight of the target data corresponding to the subtask. The second processing result output after the conflict resolution process is formatted and converted to obtain the formatted data.

4. The method according to claim 1, characterized in that, The semantic alignment of the knowledge graph and the formatted data includes: The knowledge graph is vectorized to obtain a vector of the knowledge graph; The vectors of the knowledge graph and the formatted data are mapped to the same vector space according to the condition of minimizing cosine similarity and contrast loss, so as to perform semantic alignment.

5. The method according to claim 1, characterized in that, After semantically aligning the knowledge graph and the formatted data, and constructing a multimodal inference chain and executable strategy based on the aligned data to drive the code development of an alternative system to implement the target application system, the method further includes: After the alternative system is started, user behavior data and expert behavior data in the alternative system are collected, and the corresponding models and knowledge graphs in the model pool are dynamically updated based on the user behavior data and expert behavior data.

6. The method according to claim 5, characterized in that, The system collects user behavior data and expert behavior data from the alternative system, and dynamically updates the corresponding models and knowledge graphs in the model pool based on the user behavior data and expert behavior data, including: The knowledge graph is updated based on the expert behavior data; A structural equation model was constructed based on the user behavior data to perform root cause analysis. Based on the results of the root cause analysis and the updated knowledge graph, the corresponding models in the model pool are updated.

7. A data system processing apparatus, characterized in that, The device includes: The hybrid model pool scheduling module is used to decompose the decomposition requirements of the target application system into multiple sub-tasks, and select model combinations that match the multiple sub-tasks from the model pool according to the type of the sub-tasks for parsing and processing to obtain formatted data. The formatted data includes at least one of table association, code logic, and rule chain. The knowledge graph construction module is used to call the knowledge graph engine to construct a self-updating knowledge graph of the target application system based on the table structure, business rules and business event flow of the target application system. The execution strategy generation module is used to semantically align the knowledge graph and the formatted data, and construct a multimodal inference chain and an executable strategy based on the aligned data to drive the code development of an alternative system to implement the target application system.

8. An electronic device, comprising: processor; And a memory arranged to store computer-executable instructions, characterized in that, when executed, the executable instructions cause the processor to perform the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing computer-executable instructions, characterized in that, When the computer-executable instructions are executed by a processor, they can implement the steps of the method described in any one of claims 1-6.

10. A computer program product, the computer program product comprising a computer-readable storage medium storing a computer program, characterized in that, The computer program is operable to cause the computer to perform the method as described in any one of claims 1 to 6.

Images