Natural language based zero-code worksheet dynamic generation method and related device

Abstract

The application provides a natural language-based zero-code worksheet dynamic generation method and related equipment, semantic recognition is performed on table content setting information and a name input by a user to determine an industry to which a worksheet to be generated belongs, a corresponding worksheet structure template is extracted from a pre-constructed industry knowledge base, and the worksheet structure template is optimized and adjusted in combination with the table content setting information input by the user, so that a worksheet structure meeting user requirements can be generated, high-quality worksheets can be generated through AI technologies such as deep learning and natural language processing, diversified user requirements can be met, worksheets meeting user requirements can be automatically generated, time and effort of manual definition of fields and table structures by the user are reduced, an error rate of manual operation is reduced, worksheet accuracy and consistency are improved, and work efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of information processing, and in particular to a zero-code worksheet dynamic generation method based on natural language and related equipment. BACKGROUND

[0002] With the development of low-code or zero-code technology, existing zero-code platforms generally have visual operation capabilities. By providing a graphical user interface, users can create applications through drag-and-drop components and parameter configuration without writing underlying code, significantly reducing the technical threshold for application development, and the technology has been widely applied in enterprise office, data management and other scenarios. In the development of data management applications, the worksheet is the core data carrier, and the construction of the table structure is a key link. However, the current zero-code platform still relies on a manual configuration dominated technical solution in the worksheet generation link, that is, the user needs to complete the table structure design, field attribute setting and table interrelation logic definition one by one through the configuration interface provided by the platform. This technical solution has some significant technical bottlenecks, which limits the ease of use, adaptability and intelligence level of the platform, and has high configuration complexity and technical threshold. The existing solution requires users to have basic knowledge of table structure design, and manually complete field naming, data type selection, mandatory item setting, uniqueness constraint configuration and multi-table interrelation definition operations. For non-technical background users, such configuration involves the basic logic of database design, which has high learning cost and operation threshold, and is easy to cause subsequent data management confusion due to improper configuration. Moreover, the existing technology lacks flexibility in industry adaptation, has limited template reusability, lacks intelligent generation capability, and is insufficient in professional scenario adaptation. The current technical solution only supports basic field generation based on simple text instructions, and cannot generate complex and professional field combinations and interrelation logic in combination with industry characteristics and specific use scenarios. It fails to truly realize intelligent cost reduction and efficiency improvement. SUMMARY

[0003] The present application aims to at least solve one of the above technical defects, and in view of this, the present application provides a zero-code worksheet dynamic generation method based on natural language and related equipment to solve the technical defects of the limited function of the zero-code platform in dynamically generating tables in the prior art.

[0004] A natural language-based zero-code worksheet dynamic generation method, the method comprising: determining attribute parameters of a target table to be processed, wherein the attribute parameters of the target table include a name of a target worksheet to be generated, content setting information, and image data; performing multi-dimensional cross-modal semantic fusion analysis on the attribute parameters of the target table to determine target worksheet structured demand information corresponding to user demand; performing similarity matching analysis on the target worksheet structured demand information and a preset dynamic industry knowledge graph, and calculating a matching degree of the target worksheet structured demand information and each subgraph in the preset dynamic industry knowledge graph; determining at least one candidate industry subgraph corresponding to the target worksheet structured demand information according to the matching degree of the target worksheet structured demand information and each subgraph in the preset dynamic industry knowledge graph; instantiating an initial worksheet structure template based on each of the determined candidate industry subgraphs to obtain a target worksheet structure template, wherein the target worksheet structure template includes field definitions of the worksheet, connection logic of the worksheet and an external data source, and cross-table reference formulas based on relationships of the preset dynamic industry knowledge graph; presenting the target worksheet structure template to the user and capturing all operations of the user on the target worksheet structure template; dynamically analyzing all operations of the user on the target worksheet structure template, dynamically optimizing the target worksheet structure template, until a target worksheet structure finally determined by the user is generated; after the user confirms the final target worksheet structure, presenting the final target worksheet structure to the user again, automatically configuring an external data source corresponding to the target worksheet structure, generating a dynamic worksheet corresponding to the user's demand, and displaying the dynamic worksheet to the user.

[0005] Preferably, the multi-dimensional cross-modal semantic fusion analysis on the attribute parameters of the target table to determine target worksheet structured demand information corresponding to user demand comprises: performing multi-dimensional cross-modal semantic fusion analysis on the attribute parameters of the target table to obtain semantic analysis results of the target table; determining target worksheet structured demand information corresponding to user demand based on the semantic analysis results of the target table.

[0006] Preferably, the multi-dimensional cross-modal semantic fusion analysis on the attribute parameters of the target table is performed to obtain a semantic analysis result of the target table, including: performing deep semantic analysis on the name and content setting information of the to-be-generated worksheet included in the attribute parameters of the target table, extracting the business entity, intent and logical constraint information between fields of the to-be-generated worksheet as the text semantic analysis result of the target table; using a preset worksheet visual analysis model to analyze the image data included in the attribute parameters of the target table, and extracting the header structure, data area and table style features included in the attribute parameters of the target table as the text visual analysis result of the target table, wherein the preset worksheet visual analysis model is trained by taking the image data included in the attribute parameters of the target table as a training sample and taking the header structure, data area and table style features contained in the image data included in the attribute parameters of the target table as a sample label; and the semantic analysis result of the target table is generated by aligning and fusing the text semantic analysis result of the target table and the text visual analysis result of the target table.

[0007] Preferably, the process of extracting the header structure, data area and table style features included in the attribute parameters of the target table by the preset worksheet visual analysis model includes: performing target detection and OCR recognition processing on the image data included in the attribute parameters of the target table; jointly encoding the recognized text information and the control position information thereof in the image data included in the attribute parameters of the target table to determine the table physical structure and logical hierarchy included in the attribute parameters of the target table; and extracting the header structure, data area and table style features included in the attribute parameters of the target table based on the table physical structure and logical hierarchy included in the attribute parameters of the target table.

[0008] Preferably, the method further includes: crawling real-time authoritative information of each industry; using a preset training language analysis model to extract industry entity relationship information from the crawled real-time authoritative information of each industry and construct an industry entity triple; and incorporating the industry entity triple into an original dynamic industry knowledge graph to update the preset dynamic industry knowledge graph in real time.

[0009] Preferably, the incorporating of the industry entity triple into the original dynamic industry knowledge graph includes: mapping the entities and relationships in the industry entity triple to a corresponding continuous vector space of the dynamic industry knowledge graph; and finding a subgraph closest to the structured demand representation corresponding to the industry entity triple in the vector space and embedding the entities and relationships in the industry entity triple.

[0010] Preferably, the method further comprises: generating a data provenance report corresponding to the target worksheet structure while generating the target worksheet structure determined by the user finally, wherein the data provenance report comprises the source of each field of the target worksheet structure, the calculation logic and the basis node thereof in the dynamic industry knowledge graph.

[0011] A natural language-based zero-code worksheet dynamic generation system applied to the natural language-based zero-code worksheet dynamic generation method described in any of the preceding embodiments, the system comprising: a multi-modal input interface module, a cross-modal semantic fusion analysis module, a dynamic industry knowledge graph module, a graph neural network matching processing module, a dynamic template generation module, an interaction optimization module, and a deployment module; wherein the cross-modal semantic fusion analysis module comprises a text analysis submodule, a visual analysis submodule, and a multi-modal fusion submodule; wherein the multi-modal input interface module receives and processes user input data, determines the attribute parameters of the target table to be processed, and transmits them to the cross-modal semantic fusion analysis module, wherein the attribute parameters of the target table include the name of the target worksheet to be generated, content setting information, image data, and document data; the user's input methods include natural language text input, image input, document upload input, and voice input; the text analysis submodule uses a pre-trained language model based on Transformer to perform deep semantic analysis on the name of the target worksheet to be generated and the content setting information included in the attribute parameters of the target table, including named entity recognition, relation extraction, intent classification, and other functions; uses a multi-head attention mechanism to capture long-distance dependency relationships; extracts the business entities, intents, and logical constraint information between fields of the target worksheet to be generated as the text semantic analysis results of the target table, and transmits them to the multi-modal fusion submodule; the visual analysis submodule uses a preset worksheet visual analysis model to analyze the image data included in the attribute parameters of the target table, extracts the header structure, data area, and table style features included in the attribute parameters of the target table as the text visual analysis results of the target table, and transmits them to the multi-modal fusion submodule, wherein the preset worksheet visual analysis model is trained using the image data included in the attribute parameters of the target table as training samples and the header structure, data area, and table style features contained in the image data included in the attribute parameters of the target table as sample labels; the multi-modal fusion submodule uses a cross-attention-based multi-modal fusion mechanism to align and fuse the text semantic analysis results of the target table with the text visual analysis results of the target table through a dynamic weight distribution algorithm, generates target worksheet structured demand information corresponding to the user's demand, and transmits it to the graph neural network matching processing module; the graph neural network matching processing module is responsible for similarity matching analysis of the target worksheet structured demand information and the preset dynamic industry knowledge graph, calculates the matching degree of the target worksheet structured demand information and each subgraph in the preset dynamic industry knowledge graph, and transmits it to the dynamic industry knowledge graph module; the dynamic industry knowledge graph module determines at least one candidate industry subgraph corresponding to the target worksheet structured demand information according to the matching degree of the target worksheet structured demand information and each subgraph in the preset dynamic industry knowledge graph, and transmits it to the dynamic template generation module;The dynamic industry knowledge graph module further includes a knowledge acquisition submodule, a knowledge storage submodule, and a knowledge update submodule. The knowledge acquisition submodule is responsible for establishing an automated knowledge extraction pipeline to continuously extract knowledge from industry websites, standard document libraries, and API documents. A relation joint extraction model based on BERT is used to extract industry entity triples. The knowledge storage submodule is responsible for storing various industry knowledge using a graph database to support complex graph query and reasoning. The schema layer of the knowledge graph includes multiple dimensions such as industry ontology, data objects, business rules, and calculation formulas. The knowledge update submodule is responsible for implementing a knowledge freshness evaluation mechanism based on time decay to automatically update outdated knowledge on a regular basis, ensuring the timeliness of the dynamic industry knowledge graph. The dynamic template generation module is responsible for instantiating an initial worksheet structure template based on the determined candidate industry subgraphs, obtaining a target worksheet structure template, and transmitting it to the interactive optimization module. The target worksheet structure template includes field definitions of the worksheet, connection logic between the worksheet and external data sources, and cross-table reference formulas based on the preset dynamic industry knowledge graph relationships. The interactive optimization module presents the target worksheet structure template to the user and captures all user operations on the target worksheet structure template. The dynamic analysis of all user operations on the target worksheet structure template dynamically optimizes the target worksheet structure template until the user finally determines the target worksheet structure and transmits it to the deployment module. The deployment module presents the final target worksheet structure to the user again after the user confirms the final target worksheet structure, automatically configures the external data sources corresponding to the target worksheet structure, generates a dynamic worksheet corresponding to the user's demand, and displays it to the user.

[0012] A natural language-based zero-code worksheet dynamic generation device includes one or more processors and a memory. The memory stores computer readable instructions that, when executed by the one or more processors, implement the steps of the natural language-based zero-code worksheet dynamic generation method described in any of the preceding embodiments.

[0013] A readable storage medium stores computer readable instructions that, when executed by one or more processors, cause the one or more processors to implement the steps of the natural language-based zero-code worksheet dynamic generation method described in any of the preceding embodiments.

[0014] It can be seen from the above introduction that the application can determine the attribute parameters of the target table to be processed, wherein the attribute parameters of the target table include the name, content setting information and image data of the target worksheet to be generated; so that the attribute parameters of the target table can be subjected to multi-dimensional cross-modal semantic fusion analysis to determine target worksheet structured demand information corresponding to the user demand; and the target worksheet structured demand information is subjected to similarity matching analysis with the preset dynamic industry knowledge graph, and the matching degree of the target worksheet structured demand information and each subgraph in the preset dynamic industry knowledge graph is calculated; and at least one candidate industry subgraph corresponding to the target worksheet structured demand information is determined according to the matching degree of the target worksheet structured demand information and each subgraph in the preset dynamic industry knowledge graph; so that an initial worksheet structure template can be instantiated based on each determined candidate industry subgraph, so that a target worksheet structure template can be obtained, wherein the target worksheet structure template can include each field definition of the worksheet, the connection logic of the worksheet and the external data source, and the cross-table reference formula based on the preset dynamic industry knowledge graph relationship; since the obtained target worksheet structure template needs to meet the user demand, the target worksheet structure template can be presented to the user, and all operations of the user on the target worksheet structure template are captured; and all operations of the user on the target worksheet structure template are dynamically analyzed, and the target worksheet structure template is dynamically optimized until the target worksheet structure finally determined by the user is generated; after the user confirms the final target worksheet structure, the final target worksheet structure is presented to the user again, and the external data source corresponding to the target worksheet structure is automatically configured, a dynamic worksheet corresponding to the user's demand is generated and displayed to the user.

[0015] Therefore, the application can generate a high-quality worksheet by performing semantic recognition on the table content setting information and name input by the user to determine the industry to which the generated worksheet belongs, extracting a corresponding worksheet structure template from a pre-constructed industry knowledge base, and optimizing and adjusting the worksheet structure template in combination with the table content setting information input by the user, thereby meeting the user's diversified demand, automatically generating a worksheet meeting the user's demand, reducing the time and effort of the user to manually define fields and table structures, reducing the error rate of manual operation, improving the accuracy and consistency of the worksheet, and improving work efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed to be used in the description of the embodiments or the prior art will be briefly introduced. Obviously, the accompanying drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor. Figure 1 A flow chart of a natural language-based zero-code worksheet dynamic generation method provided by the present application is shown in the following. Figure 2 A natural language-based zero-code worksheet dynamic generation system architecture schematic diagram is shown in the following. Figure 3 A hardware structure block diagram of a natural language-based zero-code worksheet dynamic generation device disclosed by the present application is shown in the following. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments only constitute some embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0018] In view of the fact that most of the current natural language-based zero-code worksheet dynamic generation schemes are difficult to adapt to complex and changeable business requirements, the present applicant has researched a natural language-based zero-code worksheet dynamic generation scheme. The natural language-based zero-code worksheet dynamic generation method judges the industry to which the generated worksheet belongs by performing semantic recognition on the table content setting information and name input by the user, extracts the corresponding worksheet structure template from the pre-constructed industry knowledge base, and optimizes and adjusts the worksheet structure template in combination with the table content setting information input by the user, so as to generate a worksheet structure meeting the user's requirements. Through deep learning and natural language processing AI technology, the system can generate high-quality worksheets to meet the user's diversified needs, automatically generate worksheets meeting the user's requirements, reduce the time and effort of the user to manually define fields and table structures, reduce the error rate of manual operation, improve the accuracy and consistency of the worksheet, and improve the work efficiency.

[0019] The method provided by the embodiments of the present application can be used in a plurality of general-purpose or special-purpose computing device environments or configurations. For example, a personal computer, a server computer, a handheld device or a portable device, a tablet device, a multi-processor device, a distributed computing environment including any of the above devices or equipment, and the like. The present application provides a zero-code worksheet dynamic generation method based on natural language, which can be applied to various management systems, and can also be applied to various computer terminals or intelligent terminals, and the execution subject can be a processor or a server of a computer terminal or an intelligent terminal.

[0020] The flow of the zero-code worksheet dynamic generation method based on natural language provided by the embodiments of the present application will be introduced below in combination with Figure 1 , as shown in FIG. 1, the flow can include the following steps: Figure 1

[0021] In step S101, the attribute parameters of the target table to be processed are determined.

[0022] Specifically, in order to ensure that the table is "available, accurate, and adapted to the needs", when automatically generating a worksheet (such as Excel, Google Sheets, or a structured table generated by code), the attribute parameters of the target table to be processed can be determined. The attribute parameters of the target table include the name of the target worksheet to be generated, content setting information, and image data. These attribute parameters are the "design blueprint" of the table, which directly determines whether the table can correctly carry data, support subsequent processing (such as calculation, analysis, and visualization), and whether it meets the use scenarios of users or systems. If this step is skipped and the table is directly generated, problems such as "data cannot be filled in", "format is chaotic", and "function is invalid" (for example, a cell that should store a date is set to a text format, causing the date sorting to be incorrect; a cell that should store a numerical value is set to a character format, causing the sum formula to be incorrect) are likely to occur. The core problem of avoiding "table cannot carry data" is to clearly define the "storage rules" of data. The essence of a table is a "structured data container", and the attribute parameters directly define the "capacity, shape, and material" of this "container", that is, how data can be stored in what form and what rules. Without clear parameters, data and the table will be incompatible and cannot be used normally. The core attribute parameters that need to be determined and the corresponding effects are shown in Table 1:

[0023]

[0024] Defining the "function rules" of the table is a key capability to ensure that the table supports subsequent processing. The automatically generated worksheet is not usually "static storage", but needs to support subsequent operations (such as formula calculation, data filtering, sorting, and linkage with other tables). The attribute parameters directly determine whether these functions can work normally, avoiding the embarrassment of "table generated but cannot be used". ​

[0025] Typical scenarios and dependent attribute parameters are as follows:

[0026] 1. Formula calculation requirements: If the table needs to automatically calculate "Employee monthly salary = hourly wage x working hours", it needs to be determined in advance that the data types of "hourly wage" and "working hours" columns are "numeric" (not text), otherwise the formula cannot be recognized. If you need to keep 2 decimal places, set the cell format of the "monthly salary" column to "numeric (2 decimal places)" to avoid the calculation result showing "5200.3333".

[0027] 2. Data filtering and sorting requirements: If you need to filter "2024 new employees" by "date of employment", you need to determine in advance that the data type of the "date of employment" column is "date" (not text), otherwise you cannot accurately filter by "year" and "month" (text format "2024-05-01" will be sorted as a string, not in chronological order).

[0028] 3. Cross-table linkage requirements: If the "employee performance table" needs to link "employee ID" to "employee information table" to obtain "department" information (such as VLOOKUP function), you need to determine in advance that the "employee ID" columns of the two tables have the same data type (both text or both numeric), and "employee ID" is a unique value (no duplication), otherwise the linkage will return an error value (#N / A).

[0029] Defining the "adaptation rules" of the table can adapt to the individual needs of "different use scenarios". In different scenarios, the "user" and "purpose" of the table are different, and the requirements for the table are completely different. Attribute parameters need to match scenario requirements to avoid "general tables that cannot meet specific scenarios". Common scenarios and corresponding attribute parameters can be shown in Table 2:

[0030]

[0031] Attribute parameters are the "design blueprint" of "automatically generated worksheets". The process of determining attribute parameters is to "translate vague requirements into clear technical rules". From "I need an employee table", it is refined to "the employee table needs 6 columns, employee ID (unique), name (non-empty), date of employment, department, salary (2 decimal places), and whether to be regularized". Only when these rules are clear can the automatically generated table "accurately match the requirements", neither data storage problems nor subsequent calculations, analysis, and linkage can be supported, and the ultimate goal of "generate and use" can be achieved. Conversely, if this step is skipped, the generated table is most likely a "semi-finished product" that needs to be manually adjusted for use, losing the efficiency of "automatic generation".

[0032] In step S102, multi-dimensional cross-modal semantic fusion analysis is performed on the attribute parameters of the target table to determine the target worksheet structured demand information corresponding to the user demand.

[0033] Specifically, the user's demand is often not directly expressed by "explicit instructions" (such as "I want a 3-column, numerical type table"), but is hidden in multi-modal information such as natural language description, reference document, data example, and use scenario. In order to solve the problem of "basic parameters cannot accurately match the user's deep demand", after determining the basic attribute parameters (such as the number of columns, data type, and format constraint) of the target table, multi-dimensional cross-modal semantic fusion analysis is further performed on the attribute parameters of the target table to determine the target worksheet structured demand information corresponding to the user demand. Only by "fusion interpretation" of these scattered and ambiguous information, can the basic attribute parameters be converted into "structured demand fully matching the user's actual use", avoiding the deviation of "correct table parameters but cannot be used".

[0034] In practice, users rarely express their demands in "table design language", but more often express their demands through natural language (such as "help me make a table to record customer follow-up, which can show the time and result of each communication, and distinguish whether the customer has an intention"), reference pictures (such as "similar to the Excel table I sent, but add a contract amount column"), and data examples (such as "I have several rows of customer data, make a table according to this format"). These information itself does not directly correspond to attribute parameters such as "column definition" and "data constraint". The basic parameters (such as "column name = communication time, data type = date") only define the "technical rules of the table", but cannot answer the questions "what is this column used for? Whether it needs to be linked with other columns? Whether it needs to support a specific analysis scenario?". These "use level" questions are the key to determine whether the table is "useful".

[0035] Therefore, multi-dimensional cross-modal semantic fusion analysis is to complete the conversion from "technical parameters" to "business use" by "integrating multi-source and multi-type information", and finally determine the accurate structured demand. It can solve the "ambiguity and ambiguity of demand", from "literal understanding" to "semantic accurate interpretation". The demand expressed by the user through natural language or a single modality often has ambiguity or information missing, which cannot be covered by basic parameter definition. Multi-dimensional cross-modal fusion can eliminate ambiguity and accurately locate the user's real intention through "multi-information mutual verification".

[0036] User requirements are usually in natural language, such as "Help me create a table to record product sales, and calculate how many of each product have been sold and how much money has been earned." There are limitations if only looking at basic parameters. For example, if only literal information is extracted, the basic parameters may be defined as follows: column name = product name (text), sales (numeric), profit (numeric). However, this parameter design has obvious ambiguities: "profit" refers to "single-item profit" or "total profit"? If the user actually needs "total profit = single-item profit x sales," defining only the "profit" column will result in the need for manual calculation, which violates the "automatic recording" requirement; does "product name" need to distinguish between "model"? If the user needs to filter by "product model" later, only the "product name" column may not be sufficient (e.g., "iPhone 15" and "iPhone 15 Pro" need to be counted separately). If both the "data sample" provided by the user (e.g., "product model: iPhone 15 Pro, unit price: 8999, cost: 6000, sales: 10") and the "usage scenario description" (e.g., "I need to calculate the total profit of each model every week to facilitate commission calculation") are considered, the following information can be determined through fusion analysis:

[0037] 1. The "unit price (numeric), cost (numeric)" columns need to be added, and the "total profit" column needs to have a preset formula (= (unit price - cost) x sales) instead of being manually filled in.

[0038] 2. "Product name" needs to be split into "product category (text) + product model (text)" two columns to support filtering and statistics by "model."

[0039] 3. The "statistical week (text / date)" column needs to be added to adapt to the "weekly statistics" scenario.

[0040] Through the fusion of "natural language requirements + data samples + scenario descriptions," the ambiguity of basic parameters is eliminated, and the table design fully meets the user's core purpose of "statistical commission."

[0041] User's demand information is often scattered in multiple modalities (such as "natural language description + reference table picture + target system requirements"), and information from different modalities may be complementary, conflicting, or repetitive, requiring semantic alignment through fusion analysis. Map information from different sources to a unified "table structure dimension" (such as column definition, association rule, and functional constraint) to avoid information omission or contradiction. Key fusion dimensions and their roles are shown in Table 3:

[0042]

[0043] It can be seen that the information of a single mode can only support part of the attribute parameters, while through the cross-modal fusion analysis of "natural language + reference + data + system constraints", the scattered information can be "pieced together into a complete structured demand", ensuring that the table is not only "correct in parameters", but also adapts to the user's operation habits, data characteristics and external system requirements.

[0044] The user's demand exists "explicit demand" and "implicit demand". The explicit demand can be met by the basic parameters, but if the implicit demand is not excavated, the table will need to be modified significantly after a period of use due to "insufficient functions". Multi-dimensional cross-modal fusion analysis can predict these potential demands through "scene correlation" and "semantic reasoning", making the structured design more "forward-looking".

[0045] For example, the user's explicit demand (natural language + reference table): "Help me create an employee attendance table to record whether they are present or absent each day, and also mark if they are late or absent." Basic parameter design: column name = employee name (text), date (date), attendance status (drop-down options: present / late / absent / absent). Problem: only meets the "recording" requirement, but does not consider the user's subsequent "attendance statistics" requirement, which would require manual filtering and statistics if the user needs to calculate "number of times late" and "attendance rate" every week, which is extremely inefficient.

[0046] Further cross-modal fusion to excavate implicit demands can be as follows:

[0047] 1. From "user identity" (assuming HR), infer that HR not only records attendance, but also calculates "monthly attendance rate" and "late deduction", which is an implicit demand;

[0048] 2. From "notes in the reference table" (such as a handwritten note in the reference table that "late 1 time deducts 50 yuan"), verify that the user indeed has a "deduction calculation" requirement;

[0049] 3. From "industry conventions" (enterprise attendance often needs to be associated with "work calendar" to exclude holidays), supplement that it is necessary to avoid counting holidays as "absent".

[0050] Final structured demand optimization: add "workday flag (Boolean value)" column: automatically associate holiday data (non-workdays are marked as "no" by default and are not counted in attendance statistics); add "monthly number of times late (numeric value)" column: preset COUNTIF formula (=COUNTIF (current month attendance status column, "late")), automatically calculate; add "late deduction (numeric value)" column: preset formula (=monthly number of times late x 50), automatically calculate; add "monthly attendance rate (percentage)" column: preset formula (= (monthly attendance days - absent days) / monthly attendance days), automatically generate.

[0051] At this point, the table not only satisfies the explicit need of "recording attendance," but also anticipates the implicit need of "statistics and calculation," truly achieving "meeting what users think of and covering what users haven't mentioned."

[0052] Therefore, cross-modal semantic fusion serves as a precise bridge between requirements and forms, transforming users' vague multimodal needs into clear, structured rules; integrating fragmented information into complete form designs; and extending surface-level functional requirements to potential application scenarios. This analysis ensures that the generated worksheets not only have correct technical parameters but also accurately match users' actual business needs, operational habits, and future requirements, truly achieving the goal of "automatically generated, usable, and easy to use." Conversely, skipping this step and generating forms solely based on basic parameters will likely result in users receiving forms that are "not what they wanted," requiring repeated modifications and ultimately reducing efficiency.

[0053] Step S103: Perform similarity matching analysis between the structured requirement information of the target worksheet and the preset dynamic industry knowledge graph, and calculate the matching degree between the structured requirement information of the target worksheet and each subgraph in the preset dynamic industry knowledge graph.

[0054] Specifically, to ensure that the generated worksheets accurately meet user needs while deeply conforming to industry standards and practical logic, and to avoid ineffective designs that "meet superficial user requirements but are detached from industry realities," the automatic worksheet generation process performs similarity matching analysis between the target worksheet's structured requirements and a pre-defined dynamic industry knowledge graph. It also calculates the matching degree between the target worksheet's structured requirements and each subgraph in the pre-defined dynamic industry knowledge graph, thus calibrating "vague user requirements" into "precise, industry-usable requirements." Generally, user-proposed worksheet requirements often suffer from "superficiality and ambiguity." For example, a user might only say "I need an e-commerce inventory table," but not specify that it should include "SKU codes, batch numbers, near-expiration warning thresholds, and inventory turnover rate fields," while these details are precisely the core elements of e-commerce inventory management. The dynamic industry knowledge graph stores structured knowledge validated through industry practice (e.g., the "inventory management subgraph" in e-commerce includes the relationships between "products-inventory-supply chain," required fields, and data format specifications; the "expense reimbursement subgraph" in finance includes the constraint logic of "reimbursement category-approval process-invoice type"). Through similarity matching, the user's "fuzzy requirements" can be aligned with the corresponding "industry standard subgraph" in the knowledge graph: if the user's requirements match the "e-commerce inventory subgraph" by 90%, then the required industry fields not mentioned by the user (such as "near-expiration warning days") can be automatically filled in, calibrating the "table the user wants" into a "table usable in the industry," avoiding omissions in requirements or logical deviations.

[0055] It is also helpful to ensure the "industry compliance" and "practice adaptability" of the worksheet. Different industries have strict specifications or default practices for the structure, fields, and logic of the worksheet. If only user demand is generated, it is easy to have "non-compliant industry rule" errors, such as the "salary table" in the financial field, which needs to include fields such as "pre-tax salary, special additional deductions, social security fund base", and the calculation logic needs to comply with relevant regulations; the "production plan table" in the manufacturing industry needs to be associated with "work order number, material BOM list, equipment capacity data", otherwise it cannot be connected to the subsequent production execution system.

[0056] The preset dynamic industry knowledge graph is a structured deposit of public industry rules, business logic, and data standards (each "subgraph" corresponds to a subdivided scene). Through matching degree calculation, the "industry subgraph" most relevant to user demand can be filtered out; based on the constraint logic of the subgraph, it can be verified whether the user demand is compliant (such as automatically supplementing if the user does not mention "special additional deductions" to ensure compliance with tax laws); it can allow the generated worksheet to directly interface with existing industry processes (such as data format compatibility with ERP systems and financial software), avoiding the problem of "table can be generated but cannot be used". If it does not rely on the industry knowledge graph, each time a worksheet is generated, it needs to "interpret user demand from zero + manually design structure", which is extremely inefficient; while the knowledge graph can provide "predefined industry templates and logic", through matching analysis to achieve "automated completion and optimization", it can reduce manual intervention. For example, if the user demand matches the "retail industry sales report subgraph" with a matching degree of 85%, the system can directly generate the basic structure based on the "field template" of this subgraph without the user having to list the fields one by one. It can also intelligently optimize the logic, and the subgraph of the knowledge graph not only contains "fields", but also contains "associations between fields". After matching, these industry-wide calculation logic can be automatically embedded without the user having to manually set formulas; it can also quickly locate the core demand. When the user demand is complex, it can be matched with the "cross-border e-commerce inventory subgraph" and the "cross-border logistics subgraph" to preferentially integrate the logic of the subgraph with a higher matching degree, quickly locate the core scene, and avoid demand interpretation deviation.

[0057] In general, industry rules and business scenarios are not fixed, and the core feature of the "dynamic industry knowledge graph" is "real-time update". Through "demand and dynamic subgraph matching analysis", it can ensure that the generated worksheet can keep up with the latest changes in the industry. If the user demand matches the "updated subgraph", it will automatically integrate the new industry requirements; if the user demand has a high matching degree with the old subgraph and a low matching degree with the new subgraph, it can prompt that "the current demand may not comply with the latest industry specifications" and give optimization suggestions to avoid generating "outdated and ineffective" worksheets.

[0058] User demand is the "starting point", industry knowledge graph is the "scale", and similarity matching and matching degree calculation are the "bridge", which can ensure that the automatically generated worksheet not only "meets the user's spoken needs", but also "covers the user's unspoken but industry-required needs", while taking into account compliance, practicality, timeliness, and finally achieving the intelligent goal of "generate and use immediately".

[0059] Step S104, according to the matching degree of the target worksheet structured demand information and each subgraph in the preset dynamic industry knowledge graph, at least one candidate industry subgraph corresponding to the target worksheet structured demand information is determined.

[0060] Specifically, the preset dynamic industry knowledge graph is a "knowledge collection" covering multiple scenarios and multiple subfields, for example, a "retail industry knowledge graph" may include "store sales subgraph", "online e-commerce subgraph", "member management subgraph", "inventory turnover subgraph", and multiple subgraphs. Each subgraph corresponds to the knowledge (fields, logic, specifications) of a specific business scenario. User demand is often a single or explicit sub-scenario. If all subgraphs of the entire industry knowledge graph are directly called, it will lead to "knowledge overload", and it is impossible to determine whether to use "sales fields" or "inventory fields", or to determine whether to embed "sales same period comparison calculation logic" or "member points rules". By "matching degree screening candidate subgraph", "precise support can be locked from generalized knowledge", if the user demand and "store sales subgraph" match degree reaches 92% (much higher than 30% with "inventory turnover subgraph"), the "candidate subgraph is store sales subgraph" can be determined, and the subsequent design core is focused on the "sales date, store number, category, sales, same period / period comparison calculation logic" and other key knowledge of the subgraph, avoiding the confusion of knowledge call. Therefore, in the process of automatically generating the worksheet, at least one candidate industry subgraph corresponding to the target worksheet structured demand information can be further determined according to the matching degree of the target worksheet structured demand information and each subgraph in the preset dynamic industry knowledge graph.

[0061] The most suitable industry knowledge carrier (sub-graph) is selected through matching degree, providing a "precise, compliant, and feasible" basis for the subsequent structured design of the work table, avoiding directionless knowledge calling or incorrect scenario adaptation. The "generalized industry knowledge" can be focused on "precise demand support", ensuring the "high relevance" of the "candidate sub-graph" to the demand and avoiding directional errors. One of the core risks of automatically generating a work table is "scenario mismatch". For example, the user wants a "financial expense reimbursement table", but a "human resource attendance table" is generated, the root cause is that the sub-graph matching the demand is not filtered out. When the demand and a sub-graph have a significantly higher matching degree than other sub-graphs, it can be directly determined as the "core candidate sub-graph", ensuring that all subsequent designs (field selection, formula logic, format specification) are based on the business scenario corresponding to the sub-graph, avoiding directional deviation from the source. For example, the user's demand is to "generate a cross-border e-commerce logistics cost accounting table". Through matching degree calculation, if the "cross-border logistics cost sub-graph" has a matching degree of 88% and the "domestic e-commerce sales sub-graph" has a matching degree of 25%, the candidate sub-graph is locked to the former, and the subsequent process will automatically introduce "international shipping, customs, customs clearance fees, and logistics time delay costs" specific to cross-border e-commerce, rather than irrelevant fields such as "delivery fees and delivery areas" for domestic e-commerce.

[0062] Not all user demands are "single scenarios". Some demands may involve "cross-sub-scenario integration", such as the user wanting to "generate an 'inventory-sales linked e-commerce operation table'". This demand involves both "inventory management" (requiring inventory quantity and safety stock threshold) and "sales analysis" (requiring sales and hot-selling categories), and a single sub-graph cannot fully support it. At this time, "determining multiple candidate sub-graphs based on matching degree" becomes crucial: the system can calculate the matching degree of the demand and each sub-graph, filter out "multiple sub-graphs with high matching degrees" (such as "e-commerce inventory sub-graph" with a matching degree of 80% and "e-commerce sales sub-graph" with a matching degree of 75%), and determine both as "candidate sub-graphs". When generating the work table, the core knowledge of both sub-graphs can be integrated (such as taking "SKU, current inventory, and safety stock" from the inventory sub-graph and "monthly sales, sales trend" from the sales sub-graph, and embedding "inventory warning = current inventory - safety stock" and "restocking suggestion = monthly sales - current inventory" and other linkage logic). Without filtering candidate sub-graphs through matching degree, it is impossible to determine which two sub-graphs to integrate (it may incorrectly integrate "inventory sub-graph" and "logistics sub-graph"), resulting in a generated table that cannot meet the core demand of "inventory-sales linkage".

[0063] The determination of the candidate industry subgraph can also provide a "landable basis" for subsequent "worktable structured design". After determining the candidate industry subgraph, all subsequent design actions such as "field selection, formula setting, format specification, compliance verification" need to be based on the knowledge of the candidate subgraph. The candidate subgraph is an "intermediate bridge" for "transforming requirements into specific table structure". The determination of the candidate subgraph based on the matching degree provides "precise positioning" for this bridge. The candidate subgraph contains "industry required fields", for example, the "financial reimbursement subgraph" contains "reimbursement person, department, reimbursement category, amount, invoice number, approval status". If the user's demand does not mention "invoice number", it can be automatically completed based on the candidate subgraph. The candidate subgraph contains "business logic between fields", for example, the "retail sales subgraph" contains "average unit price = sales amount / transaction number". The calculation formula can be automatically embedded based on the candidate subgraph, without the need for manual setting by the user. The candidate subgraph contains "industry compliance requirements", for example, the "salary subgraph" contains "individual income tax deduction items and social security accumulation fund fields". The system can check whether the user's demand has omitted the compliance fields based on the candidate subgraph, to ensure that the table meets the industry rules. If the candidate subgraph is not determined, the subsequent design will have "no basis to rely on", neither can it determine which fields to fill in, nor can it determine what logic to embed. The final generated table may only be "a pile of scattered fields mentioned by the user", rather than a "structured worktable with logic, compliance and usability". Only by determining the candidate subgraph can "abstract user requirements" and "generalized industry knowledge" be transformed into "specific table design scheme", to finally ensure that the generated worktable not only meets the user's demand, but also conforms to the industry practice, avoiding blind trial and error or incorrect scene adaptation.

[0064] In step S105, based on the determined each candidate industry subgraph, an initial worktable structure template is instantiated to obtain a target worktable structure template.

[0065] Specifically, in the process of automatically generating a worksheet, in order to transform the "abstract industry knowledge" (such as field definition, logical relationship, compliance rule) in the candidate subgraph into "concrete table structure" (such as column name, data type, formula position, format specification), a "skeleton" is built for the final generation of a directly usable worksheet. Based on the determined each candidate industry subgraph, an initial worksheet structure template can be instantiated to obtain a target worksheet structure template. The target worksheet structure template can include each field definition of the worksheet, the connection logic of the worksheet and the external data source, and the cross-table reference formula based on the preset dynamic industry knowledge graph relationship. The candidate industry subgraph is a "structured knowledge carrier", not a "directly usable table template". For example, in the "cross-border e-commerce logistics cost subgraph", the "knowledge description" stored is "field: international freight (definition: basic fee generated by cross-border transportation, data type: numerical value), customs (definition: import tax collected by customs, calculation logic: cargo value x customs tax rate), compliance rule: include 'customs declaration number' field for traceability" and the like. These information is "abstract, non-visual", and cannot be directly used by users for data entry or analysis. The "instantiation of the initial template" can map the abstract knowledge elements in the subgraph one by one to the specific structure components of the table. For example: the "field definition" of the subgraph can correspond to the "column name + data type" of the table (such as "international freight" can correspond to the column name "international freight (yuan)" and the data type "numerical value (retain 2 decimal places)" in the table); the "calculation logic" of the subgraph can correspond to the "formula cell" of the table (such as "customs = cargo value x customs tax rate" corresponds to embedding the formula "=C2*D2" in the corresponding cell of the "customs" column, where C column is "cargo value" and D column is "customs tax rate"); the "compliance rule" of the subgraph can correspond to the "mandatory column label" of the table (such as "include customs declaration number" corresponds to labeling "mandatory" beside the "customs declaration number" column, and setting data validity check). If no transformation is performed, the knowledge of the candidate subgraph will always remain at the "theoretical level" and cannot be transformed into a table form that users can perceive and use.

[0066] Based on the determined individual candidate industry subgraphs, an initial worksheet structure template is instantiated, which ensures that the "table structure" is 100% aligned with the "candidate subgraph knowledge", avoiding knowledge omissions. The candidate subgraph is a "precise demand support" screened by matching degree, and each knowledge element (field, logic, rule) it contains is strongly related to user demand. For example, in the "financial expense reimbursement subgraph", "reimbursement category (limited options: travel expenses / office expenses / entertainment expenses), approval status (pending approval / passed / rejected), and amount verification (total amount = sum of each detailed amount)" are all core elements that the "expense reimbursement table" must have. The process of "instantiating the initial template" is the process of "landing knowledge by element", which will traverse all knowledge nodes of the candidate subgraph to ensure that each node corresponds to the table structure, avoiding key information omissions. For example: if the subgraph contains "reimbursement category limited options", the "drop-down menu" will be automatically set for the "reimbursement category" column during instantiation, rather than allowing users to input freely; if the subgraph contains "total amount verification logic", the "sum formula" will be automatically embedded in the last row of the "total amount" column during instantiation, and "data validity" will be set (if the total amount ≠ the sum of details, a prompt will be popped up). Conversely, if the table is generated directly without instantiation, problems such as "missing key fields" and "logic not landed" may occur, resulting in a table that does not meet industry knowledge requirements and cannot meet user actual use requirements.

[0067] As mentioned earlier, some user requirements require "multiple candidate subgraphs to support together" (for example, the "inventory-sales linked e-commerce operation table" requires "e-commerce inventory subgraph" and "e-commerce sales subgraph" to support together). At this time, "instantiating the initial template" can integrate the knowledge elements of multiple subgraphs into the same table template, avoiding the problem of "subgraph knowledge fragmentation". The specific fusion logic includes the following:

[0068] 1. Field de-duplication and priority sorting: If two subgraphs contain the same field (such as "SKU number" in both the inventory subgraph and the sales subgraph), the instantiation will automatically de-duplicate and retain one "SKU number" column, and use the field definition of the "subgraph with a higher match degree" as the standard (such as the inventory subgraph's definition of "SKU" containing "inventory location" attributes, and the sales subgraph only containing "product name", then use the inventory subgraph's definition as the standard, and add a "inventory location" note next to the "SKU number" column).

[0069] 2. Logic association and linkage design: If there is a business association between two subgraphs (such as "current inventory" in the inventory subgraph and "monthly sales" in the sales subgraph need to be linked to calculate "inventory sellable days"), the instantiation will automatically add an associated field (such as the "inventory sellable days" column) and embed a linkage formula ("=B2 / C2", where B column is "current inventory" and C column is "monthly sales").

[0070] 3. Format uniformity and partition design: To avoid confusion caused by the fusion of multiple subgraphs, the fields are partitioned according to the "business module" during instantiation (e.g., the left side of the table is the "inventory module" field: SKU, current inventory, safety stock; the right side is the "sales module" field: monthly sales, sales, hot-selling level, etc.), and the readability is improved through borders and background colors.

[0071] If fusion instantiation is not performed, the knowledge of multiple subgraphs will become "two independent table fragments", which cannot form a complete demand support for "inventory-sales linkage", and the user still needs to manually splice the table, losing the significance of "automatic generation".

[0072] Based on the determined candidate industry subgraphs, an initial worksheet structure template is instantiated to generate an "adjustable initial template" to leave enough space for subsequent "personalized adaptation". The "instantiated initial template" is not a "final unchangeable table", but a "structured initial skeleton" (i.e., the target worksheet structure template), which contains the core knowledge of the candidate subgraph (ensuring compliance and usability) and leaves room for "personalized adjustment" (meeting the user's special needs).

[0073] For example, a user requirement is to "generate a monthly sales table for retail stores", and the candidate subgraph is "store sales subgraph". The initial template obtained after instantiation contains core columns such as "store number", "date", "category", "sales", and "year-on-year growth rate". If the user additionally requires "adding a'store manager name' column" (a personalized requirement not included in the subgraph), since the initial template has already built the "field column + data type" skeleton, it only needs to insert a column "store manager name" (data type "text") in the template, without the need to redesign the entire table structure. If the user requires "changing the number of decimal places in the 'year-on-year growth rate' column from 2 to 1", it only needs to adjust the "data format setting" in the template, without affecting other core logic. This "first skeleton, then adjustment" mode not only avoids the inefficiency of "designing a table from scratch", but also flexibly adapts to the user's personalized requirements, balancing "standardization (industry knowledge)" and "personalization (user's special requirements)".

[0074] Based on the determined individual candidate industry subgraphs, instantiating an initial worksheet structure template can also provide an "executable structural basis" for "final worksheet generation". In subsequent processes, data needs to be filled based on the "target worksheet structure template" (such as extracting data from user data sources and filling in the corresponding columns), format rendering (such as setting table header background color and adjusting column width), and generating final files (such as Excel, Google Sheets), and the "target worksheet structure template" is the "only basis" for these operations. For example, when filling data, it needs to know that the "transportation fee" field in the data source corresponds to the "international transportation fee (yuan)" column in the template, otherwise the problem of "data filling the wrong column" will occur; when rendering the format, the system needs to know that the "approval status" column needs to be marked with color (pending approval = yellow, passed = green), otherwise it cannot achieve automatic format beautification; when generating files, the system needs to know that the table contains 10 columns, the default column width is 15 characters, and the table header font is bold 12, in order to generate a file that meets the user's reading habits. Without this template, subsequent "data filling, format rendering, and file generation" will lose direction and cannot form a final usable worksheet.

[0075] In step S106, the target worksheet structure template is presented to the user, and all user operations on the target worksheet structure template are captured.

[0076] Specifically, after generating the target worksheet structure template, in order to better meet the user's needs, the target worksheet structure template can be presented to the user, and all user operations on the target worksheet structure template are captured, which helps to solve the problem of "system self-judgment may deviate from the user's actual use habits, implicit needs". The "system designed template" can be converted into a "template that the user really recognizes", ensuring that the final generated worksheet not only "meets industry rules", but also "fits the user's operation habits, scene details, and unexpressed needs".

[0077] Although through "cross-modal semantic fusion", "industry knowledge graph matching", "instantiated template", and other steps, the template may have been as close to the user's needs as possible, but there are still two core deviation risks, for example, based on algorithm to interpret the user's multi-modal needs (such as natural language, reference examples), but it may miss the user's "unexpressed but default habits" (such as the user's habit of placing the "customer name" column on the far left of the table, rather than the system default "customer ID" column); Secondly, the subgraph of the industry knowledge graph is "general specification", but the user's specific scene may have "special adjustments" (such as the industry subgraph default "reimbursement amount" retains 2 decimal places, but the user's company requires 1 decimal place, or needs to be labeled "whether to expedite reimbursement").

[0078] The template is presented to the user and the operation is captured, which is "let the user be the final decision maker", through the user's manual adjustment (such as dragging the column order, adding or deleting fields, modifying the format), correcting the system's understanding deviation, and calibrating the "general template" to the "user-specific available template". For example: the "customer follow-up table" generated by the system contains "follow-up date, customer name, communication content" by default, and the user adds the "next follow-up time" column through operation, supplementing the implicit needs that the system did not predict, and the final template is truly consistent with the user's actual use scenario. Secondly, the self-generated template may have the problem of "seemingly meeting the needs, but actually cannot be used", for example, the user wants to "record the overtime table of employees", and the template generated based on the "human resource attendance chart" contains "employee ID, overtime date, overtime length", but the user actually needs to count the overtime length by "department", and the template does not have the "department" column. If the final table is directly generated, the user will abandon it due to "lack of key fields", losing the significance of automatic generation.

[0079] The template is presented to the user, who can intuitively check the "field integrity" (whether there are missing key information), "logical rationality" (whether the formula conforms to the actual calculation habit, such as "total overtime" whether it is calculated by "hour" or "minute"), "format adaptability" (whether the date format is the user's common "YYYY / MM / DD", instead of the system default "YYYY-MM-DD"); By capturing the user's "modification operation" (such as adding a "department" column, adjusting the formula unit, modifying the date format), directly verify and correct the deviation of the template, these operations are the "direct feedback" of the user on "whether the template meets the needs", which is more accurate than algorithm prediction. Different users have different operation habits and scene details, even if they have the same needs (such as "e-commerce inventory table"), different users may have different requirements, for example, e-commerce operator A is used to putting "SKU code" in the first column for quick search; E-commerce operator B is used to putting "inventory warning status" (red / yellow / green) in the first column to pay attention to out-of-stock goods first; The financial position user may need to add a "stock amount" column (= stock quantity * unit price) in the "inventory table", while the operation position user may not need this calculation field. These "personalized needs" cannot be fully predicted by industry knowledge graph or algorithm, and industry sub-graphs can only provide "general fields", but cannot cover "column order preferences, personalized fields, and format habits". By capturing user operations (such as dragging columns to adjust the order, inserting new columns, deleting unnecessary columns, and modifying cell formats), you can retain the user's "personalized settings" (such as remembering the user's preferred column order, and automatically following it when generating similar tables in the future); You can also supplement "personalized fields not covered by industry templates" (such as the user-added "stock amount" column, which the system can automatically embed matching formulas for); You can adjust "general format to user habit format" (such as changing the date format to the user's common style). The final generated template is no longer a "unified industry version", but a "version tailored to the user's personal habits", greatly improving the user's experience and efficiency. If you skip the "template presentation and operation capture" step and directly generate the final worksheet, the user will need to modify it after discovering the deviation, which will result in higher costs. If you need to add a field, you need to manually insert a column and adjust the position of the existing data, which may cause data misplacement; If you need to modify the formula, you need to adjust the formula for all data rows one by one to avoid partial cell calculation errors; If you need to adjust the column order, it may damage the linkage relationship with other tables. Capturing user operations in the "template stage" has very low modification cost. In the template stage, there is no actual data, so adding / deleting columns, adjusting the order, and modifying formulas will not affect the data; You can "optimize the template structure once based on user operations", and directly follow the optimized structure when filling in data in the future, avoiding the hassle of "data + structure" double modification.For example, the user discovers that the "Supplier" column is missing in the template stage, and inserts a new column. If the insertion is made after the final table is generated, the user needs to manually move the "Inventory Quantity", "Unit Price", and other columns to the right, which is prone to errors. The modification in the template stage can completely avoid such risks.

[0080] Furthermore, capturing the user's operations on the template not only optimizes the current template, but also provides key data for the "subsequent iterations" of the system. It can count the "common modification operations of users with similar needs" and integrate these "high-frequency personalized needs" into the subgraph of the industry knowledge graph in the future to reduce the modification cost of future users. It can record the "user's preferences for format and logic" and automatically adapt these preferences when generating similar templates in the future to improve the accuracy of algorithm prediction. It can enable the ability to automatically generate worksheets to evolve from "meeting general needs" to "precisely matching the habit needs of specific user groups", forming a "generation-feedback-optimization" closed loop.

[0081] The generated target worksheet structure template is a "semi-finished product based on algorithms and industry knowledge", and "presenting it to the user and capturing the operation" is a key step to let the user polish it into a "finished product that meets individual needs". It can verify and correct the system's understanding of the demand deviation; it can supplement the individualized needs not covered by the industry template; it can reduce the subsequent modification cost and improve the usability of the final table. To ensure that the automatically generated worksheet is not only "technically compliant and logically correct", but also "in line with the user's operation habits and actual scenarios", it truly achieves the goal of "generation can be used, and used in order", avoiding the awkward situation of "the system thinks it meets the needs, but the user can't use it".

[0082] Step S107, dynamically analyze all user operations on the target worksheet structure template, dynamically optimize the target worksheet structure template, and generate the target worksheet structure that the user finally determines.

[0083] Specifically, to solve the deviation between "machine pre-judgment" and "user real demand", ensure that the final output fully meets the user's personalized and scenario-based demands, in the whole process from "demand analysis to generating the final worksheet structure", dynamically analyze all user operations on the target worksheet structure template, dynamically optimize the target worksheet structure template, and generate the user's final target worksheet structure, which can make up for the natural deviation between "machine cognition" and "user real demand". Generally, the initial template generated by the machine through "structured demand information + industry knowledge graph matching" is a "pre-judgment output" based on standardized data and general industry rules, but the user's demand often has "implicit details" or "personalized preferences" that the machine cannot capture. These details cannot be completely covered by the early semantic analysis and graph matching, for example: in the e-commerce industry, the initial template generated by the machine based on the "order management" sub-graph may contain "order number, product ID, payment amount" and other general fields, but a user actually needs to add an additional "customer membership level" field or delete a "logistics number" field. These personalized needs cannot be exhausted by machine pre-judgment. For another example: user preferences for table format, field linkage logic, etc. all belong to "implicit demand" that cannot be covered by early structured analysis.

[0084] Dynamic analysis of user operations (such as user adding / deleting fields, adjusting format, modifying formula) can make the machine "passively learn" the user's real demands, and modify the "standardized pre-judgment" to "personalized adaptation" through optimized templates. The initial template only solves the problem of "meeting the user's basic demand framework", but does not solve the core problem of "whether it is easy to use or not". The user ultimately needs is not "a logical structure", but "a tool that can be directly used", if not dynamically optimized: assuming that the "customer information" column in the initial template contains "name, phone, address", but the user actually needs to filter customers by "region" in the actual business, but does not set a separate "region" field in the template. At this time, the template may "meet the industry general structure", but it cannot meet the user's actual operation needs (such as filtering and statistics), and ultimately becomes "invalid structure". By capturing user operations (such as manually inserting the "region" column and adding a sum formula for the "amount" column), the template structure can be adjusted in real time (such as fixing the position of the "region" column and defaulting to keep the sum formula), so that the final structure not only "meets the demand", but also "directly supports the user's business operations" (such as filtering regional customers and automatically calculating the total amount), and truly realizes "take and use".

[0085] The previous "structured demand analysis → atlas matching → initial template generation" is a "one-way derivation" process, and the "real verification" of the user is not passed. If the final structure is directly determined by skipping "dynamic optimization", the small deviation in the early stage may be amplified, and the final output may be completely out of line with the user's demand. For example, the user's demand is "generate a table for'monthly sales analysis'", and the machine matches the "sales daily report" subgraph based on the industry atlas (because there is a small deviation between "daily report" and "monthly analysis" in some field overlap, the initial template contains "daily sales, daily sales" and other fields); If not dynamically optimized, the user needs to manually change the "daily" field to "monthly" and add the "monthly same period / period" field after receiving it, which is very high in operation cost; Through dynamic analysis, the machine can capture the user's modification operation and optimize the template to "monthly sales, monthly sales, same period growth rate" and other fields, correct the deviation of the early subgraph matching, and form a closed loop. Dynamic optimization is "to involve the user in the structure design", through real-time verification and adjustment, the deviation in the early derivation process is corrected in time, the demand mispositioning caused by "one-way derivation" is avoided, and the consistency between the final structure and the user's demand is ensured.

[0086] Further, not all users have "structured design ability", most users only know "what function they need", but do not know "how to achieve the function through the table structure", which can be improved by analyzing the user's operation. For example, when the user selects the "sales" column multiple times to try to calculate the proportion, the machine can automatically add the "sales proportion" column in the template and preset the formula "(sales of a certain product / total sales) x 100%", reducing the user's operation threshold (no need to manually input the formula). The user's operation is "the most direct signal of demand". For example, the user repeatedly adjusts the format of the "date" column, indicating that the format setting of the initial template does not meet the user's habits. The machine can optimize the "date format" to default, so that the final structure is more in line with the user's use habits, and the "precision" of demand satisfaction is improved.

[0087] Step S108, after the user confirms the final target worksheet structure, the final target worksheet structure is presented to the user again, and the external data source corresponding to the target worksheet structure is automatically configured, a dynamic worksheet corresponding to the user's demand is generated and displayed to the user.

[0088] Specifically, to solve the problem of "structure only staying at the 'empty template' stage, unable to directly carry data and support business use", while ensuring "structure without deviation, data linkage, and use effective" through "secondary presentation + automatic configuration", the final deliverable is a truly usable dynamic tool for users. After the user confirms the final target worksheet structure, the final target worksheet structure is presented to the user again, and the corresponding external data source of the target worksheet structure is automatically configured, generating a dynamic worksheet corresponding to the user's needs and displaying it to the user, which is also a closed loop of "structure confirmation -> data activation", turning the template from an "empty shelf" to a "usable tool". The "final worksheet structure" confirmed by the user is a "static template without data", and the core requirement of the user is to "manage / analyze actual data with a table". If it only stays at the "structure confirmation" stage, the table cannot carry any business value, losing the core significance of "automatic generation". Secondly, by automatically configuring the external data source, the "static structure" can be filled with "dynamic worksheet containing real data", while ensuring that the data and the fields of the structure are accurately aligned. For example, the structure confirmed by the user is "e-commerce order analysis table" (columns: order number, product name, order time, payment amount, logistics status), which can automatically connect the user's "e-commerce backend database" (external data source), extract "order data from September 1, 2024 to September 30", match and fill according to the structure columns, and generate a dynamic table containing real order data. At this time, the table has changed from an "empty template" to a "tool that can be used to analyze the order trend in September", truly meeting the user's business needs.

[0089] In addition, even if the user has confirmed the "final structure" of the worksheet, there may still be two implicit risks: the confirmed structure of the worksheet does not match the "fields of the data source": for example, the user confirmed structure contains a "customer mobile number" column, but the field name in the external data source is "contact phone", and if it is directly filled, it will result in "data empty" or "filling the wrong column". The user has "memory bias" for the confirmed structure of the worksheet, for example, after confirming the structure, the user may forget the "design purpose of a certain column", and later find that it does not match the expectation. Presenting the structure again and automatically configuring the data source can allow the user to confirm again that the "column name, data type, and constraint rules" match the expectation before data filling (such as "customer mobile number" column exists and data type is "text"), avoiding the need to modify the structure after data filling due to "memory bias"; and automatically detecting the matching relationship between "structure column" and "data source field", if there is a mismatch, prompting the user to "please confirm whether the data source field 'contact phone' corresponds to the 'customer mobile number' column", and after the user confirms, it is filled, avoiding the invalidity of the table caused by "data mismatch".

[0090] For example, the user confirms the structure of the "product category" column, and the corresponding field in the data source is "category name". Through the mapping prompt during secondary presentation, the user confirms the corresponding relationship of "category name → product category", ensuring that the "product category" column data is accurate after filling, rather than empty.

[0091] In the scenario of users using tables, it is often not "one-time viewing of static data", but "continuous tracking of data changes" (such as e-commerce operations needing to view order data every day, HR needing to update employee attendance data every week). If a "static table" is generated (data is filled and no longer updated), the user needs to manually repeat the operation of "extracting data → copying and pasting → adjusting format", which is extremely inefficient.

[0092] The core value of "dynamic worksheet" is "data linkage and real-time update". When automatically configuring the data source, the system will establish a "dynamic association" between the "worksheet structure" and the "external data source" (such as through API interface, database connection, file timing synchronization, etc.), rather than "one-time data copying". After the user confirms the final target worksheet structure, the final target worksheet structure is presented to the user again, and the external data source corresponding to the target worksheet structure is automatically configured. The dynamic worksheet corresponding to the user's needs is generated and displayed to the user. When the user opens the worksheet again, the latest data will be automatically pulled from the external data source, without the need for manual operation.

[0093] For example, the user generates a "store sales daily report table" and configures a dynamic association with the "POS system database". Every morning, the sales data of the previous day is automatically extracted and updated to the table. The user can view the latest daily report by opening the table the next day without manually importing data, which truly realizes the goal of "table serving business efficiency".

[0094] "Generate and display", reduce the user threshold, to achieve "demand to tool seamless delivery", the user's final appeal is "quickly get the table", rather than "participate in the complex configuration process". If the user manually operates "connect data source → select data range → fill data → set update rule" after "structure confirmation", it will greatly increase the use threshold, which is contrary to the efficiency of "automatic generation". "Automatic configuration of data source and direct display after generation" is "seamless delivery", and the whole process of "data source connection, field mapping, data filling, dynamic association setting" is completed automatically, and the user does not need to master any technical operation (such as database connection, API configuration); Directly display the dynamic worksheet after generation, the user can immediately perform "filtering, sorting, calculation, visualization" and other operations, realize "demand confirmation → tool available" zero transition. For example: after the HR user confirms the "employee salary table" structure, the system automatically connects the "salary management system", extracts all employees' "basic salary, performance, deduction" data, fills it into the table and calculates "actual salary", and directly displays the complete salary table after generation. HR does not need to manually import data or calculate, open it to be used for salary check, which greatly reduces the operation cost.

[0095] After the user confirms the final target worksheet structure, the final target worksheet structure is presented to the user again, and the external data source corresponding to the target worksheet structure is automatically configured. The dynamic worksheet corresponding to the user's demand is generated and displayed to the user, which can provide "data basis" for subsequent "table function expansion". The user may need to "expand the function" of the table (such as adding data visualization charts, setting data early warning, and linking other tables), and the realization of these functions depends on "the table has carried real data" and "dynamic association with data source". For example, the user wants to add "monthly sales trend chart" in "order analysis table". If the table has been dynamically associated with the data source, the chart can be automatically generated based on real-time data, and the chart will change synchronously with the data update. If the user wants to set "inventory early warning" (when "current inventory" is lower than "safety inventory", it will be marked red), the system can automatically trigger the early warning rule based on the dynamic data in the table, without the user manually monitoring. If the data source is not automatically configured and only static structure is generated, the user needs to manually fill in the data before setting the function, which is very inefficient. "Dynamic worksheet" provides convenience for subsequent function expansion by establishing "data basis" in advance, so that the table has "continuous iterative value".

[0096] After the user confirms the final target worksheet structure, the final target worksheet structure is presented to the user again, and the corresponding external data source of the target worksheet structure is automatically configured, a dynamic worksheet corresponding to the user's demand is generated and displayed to the user, which can ensure that the "structure matches the data source" and avoid data mismatch; the table is changed from a "static template" to a "dynamically updated business tool" to adapt to continuous use scenarios; the available table is seamlessly delivered, the user threshold is reduced, and the goal of "automatically generating and being usable" is truly realized. Starting from the user's vague demand, through semantic analysis, graph matching, template instantiation, and user adjustment, the dynamic worksheet is finally landed as "precise structure, real-time data, and directly used for business", which fully meets the user's core demand.

[0097] As can be known from the above introduction, the table content setting information and the name input by the user are subjected to semantic recognition to determine the industry to which the generated worksheet belongs, and the corresponding worksheet structure template is extracted from the pre-constructed industry knowledge base, and then the worksheet structure template is optimized and adjusted in combination with the table content setting information input by the user, so that a worksheet structure meeting the user's demand is generated. Through deep learning and natural language processing AI technology, the system can generate high-quality worksheets to meet the user's diversified needs, automatically generate worksheets meeting the user's demand, reduce the user's time and effort for manually defining fields and table structures, reduce the error rate of manual operation, improve the accuracy and consistency of the worksheet, and improve the work efficiency.

[0098] As can be known from the above introduction, the application can perform multi-dimensional cross-modal semantic fusion analysis on the attribute parameters of the target table to determine the target worksheet structured demand information corresponding to the user's demand. Next, the process is introduced, which can include the following:

[0099] In step S201, the name and content setting information of the generated worksheet included in the attribute parameters of the target table are subjected to deep semantic analysis, and the business entity, intent and field logical constraint information of the generated worksheet are extracted as the text semantic analysis result of the target table.

[0100] Specifically, in the multi-dimensional cross-modal semantic fusion analysis, the vague and unstructured user demand is converted into precise and landable structured demand information. The name and content setting information of the generated worksheet included in the attribute parameters of the target table are subjected to deep semantic analysis, and the business entity, intent and field logical constraint information of the generated worksheet are extracted as the text semantic analysis result of the target table, so as to lay a "precise alignment" foundation for subsequent matching of industry subgraphs and generation of worksheet structures.

[0101] The table requirements proposed by users are often abstract, colloquial or fragmented, and the "worksheet name" and "content setting" are the most direct "text carriers" of user requirements. Extracting "business entities" through deep semantic analysis is to find concrete business anchors for ambiguous requirements. These entities directly correspond to core objects in actual business scenarios, avoiding the deviation of the generated worksheet from the user's real business scenario.

[0102] Users create tables to achieve specific goals, and "intent" is a precise extraction of the core goal. If the intent analysis is skipped, the table may have a "structure but no function". For example: the same "customer table", if the user's implicit intent is "to filter high-value customers for follow-up", the table needs to include "consumption amount" and "repeat purchase rate" fields that can be filtered; if the intent is "to record the basic contact information of customers", only "name", "phone" and "address" basic fields are needed. By analyzing the "content setting information" (such as the user's supplement "need to distinguish between new and old customers, and can see the consumption in the last 3 months"), the core intent of "filtering and classification" and "time range statistics" can be extracted, ensuring that the structure of the generated table can directly support the user's operation goal, rather than simply stacking fields.

[0103] The core value of a worksheet is not only "contains the required fields", but also the logical association and constraints between fields. Without logical constraints, the table may have data contradictions, calculation errors and other problems, losing its practical value.

[0104] For example, the "content setting" in the user's requirement includes "the order table needs to automatically calculate the total price, and the unpaid order cannot fill in the delivery information". Through semantic analysis, two key logical constraints can be extracted:

[0105] 1. Calculation constraint: the "total price" field is derived from the "unit price" and "quantity" fields (not manually input);

[0106] 2. State constraint: the filling of the "delivery date" field depends on the "order status" being "paid" (precondition constraint).

[0107] These constraint information will be directly translated into "logical rules" of the subsequent worksheet structure (such as Excel formulas and data validation rules), ensuring the accuracy of the table data and the standardization of the operation, avoiding the user's subsequent redundant work of manually adjusting the logic.

[0108] The multi-dimensional cross-modal semantic fusion analysis usually combines multi-modal information such as text (such as user text requirements), voice (such as user voice description), and even visual sketches (such as user hand-drawn table frames). The text semantic analysis result (entity, intent, constraint) is the common semantic language of all modal information. The business entity (complaint, processing progress), intent (track complaint processing status), and constraint (unhandled state marked in red) obtained by analyzing the text attribute parameters can be used as a translator to convert the "marked in red" in the voice into a conditional formatting rule and align the fields in the hand-drawn sketch with the entity, finally realizing semantic unification of multi-modal information, avoiding ambiguity or conflict between different modal information, and ensuring that the target worksheet structured demand information obtained after fusion is complete and consistent.

[0109] In step S202, the preset worksheet visual analysis model is used to analyze the image data included in the attribute parameters of the target table, extract the table header structure, data area, and table style features included in the attribute parameters of the target table as the text visual analysis result of the target table.

[0110] Specifically, in order to supplement the "visual and spatial requirements" that cannot be covered by text semantic analysis, the finally generated worksheet not only "matches the function of the business", but also "matches the user's usage habits", and realizes the complete fusion of "text + visual" multi-modal information. In the process of determining the structured demand information of the target worksheet, the preset worksheet visual analysis model can be used to analyze the image data included in the attribute parameters of the target table, and extract the table header structure, data area, and table style features included in the attribute parameters of the target table as the text visual analysis result of the target table. The preset worksheet visual analysis model uses the image data included in the attribute parameters of the training target table as the training sample, and uses the table header structure, data area, and table style features contained in the image data included in the attribute parameters of the training target table as the sample label, and is trained.

[0111] As can be known from the above introduction, the attribute parameters of the target table can include image data. For example, it can also include user input files (the files can be document or image data or video data), and the user's demand for the table is often not only expressed in words, but also implicitly structured expectations are transmitted in the form of images, such as hand-drawn table sketches, table screenshots for reference, or even diagrams marked with “divide into two columns here” and “merge cells in this part”. These requirements cannot be obtained through text semantic analysis and must rely on visual analysis. For example, the user provides a “competitor sales report” screenshot as a reference, and the text only states “make a similar table”. At this time, visual analysis can accurately extract the table header structure: “product category (merge 2 rows) → monthly sales (divide into 1-12 month columns) → annual total (right-aligned)”; data area: “the sales data area is below the table header, each row corresponds to one product, and the numerical value is rounded to one decimal place”; style characteristics: “table header background color is blue, data row changes color every other row, and total row is bold”. If visual analysis is skipped and only text “similar to competitor table” is used to generate a table, the table header level may be chaotic (such as not merging cells), the data format may not match (such as rounding to an integer), and the deviation from the user's expectations is large.

[0112] The “usability” of a table depends not only on whether the fields are complete, but also on the spatial arrangement logic of the fields, that is, “which fields go together, how many levels the table header is, and how the data area and summary area are divided”. This spatial logic directly affects the user's browsing efficiency, and text semantic analysis (such as extracting “customer, order, and note” entities) cannot determine “how the entities should be visually sorted”. Visual analysis can solve this problem by extracting the spatial characteristics of “table header structure” and “data area”: table header level: identify “multi-level table headers” (such as “annual sales” under “Q1-Q4”, and “Q1” under “1-3 months”), avoid generating “flat table headers” that cause information confusion; area division: identify the spatial boundaries of “data area” and “summary area” to ensure the visual hierarchy of the table; field sorting: identify the visual habits of “high-frequency browsing fields on the left” (such as customer name and order number) and “auxiliary fields on the right” (such as notes), to avoid generating “important fields on the right, requiring horizontal scrolling to find” inefficient structure. For example: visual analysis finds that in the reference image provided by the user, “customer ID” and “customer name” are always on the far left, and are separated from the middle “order information” by a vertical line, which can directly convert this spatial logic into the layout rules of the target table, improving the user's browsing efficiency when using it later.

[0113] The "style features" of a table (such as color, font, cell format, border) seem to be "aesthetic needs", but in fact they are strongly related to the use scenario and user habits, and even affect the readability and professionalism of the data. Text semantic analysis cannot obtain these style information, which must be extracted through visual analysis and landed. Secondly, different business scenarios have default specifications for table styles, for example, financial statements usually require "table header in dark color, data area in light color, amount field right-aligned and with thousand separator"; project progress table may need "completed task row marked in green, uncompleted in yellow". Visual analysis can extract these scenario-based styles from the image data provided by the user, ensuring that the generated table meets the industry or enterprise specifications without the need for users to manually adjust the format later. Some users have fixed style habits. Visual analysis can identify these personalized features, for example, from the user's hand-drawn sketch, it is found that "all table headers are marked with a circle 'bold' ", so in the generated template, the "table header bold" style is automatically applied, improving the user's comfort and acceptance.

[0114] As mentioned earlier, "multi-dimensional cross-modal semantic fusion" needs to integrate multi-modal information such as text (such as name, content setting) and vision (such as image data). And "text visual analysis results" (table header, data area, style) are the key bridge connecting text semantics and visual needs, which can avoid conflicts or omissions between the two modalities of information. If the text description is "to contain 10 fields", but the visual image only shows 8 fields with clear 2-column merging, visual analysis can prompt "text field number does not match image layout", and prefer to follow the user's visual expectation (8 columns, including merged cells), or further confirm the requirements; text analysis extracts the intent of "statistical monthly sales" and the logical constraints of "order amount = unit price x quantity", while visual analysis supplements the visual rules of "monthly sales column in table header 3-14 columns" and "order amount column right-aligned and marked red". After the fusion of the two, the generated target worksheet structured demand information contains both "functional logic" (text) and "visual form" (vision), ensuring that the subsequent generated template is a complete solution that "matches both function and form".

[0115] Among them, the process of extracting the table header structure, data area, and table style features included in the attribute parameters of the target table includes the following: first, performing target detection and OCR recognition processing on the image data included in the attribute parameters of the target table; then, jointly encoding the recognized text information and its control position information in the image data included in the attribute parameters of the target table to determine the table physical structure and logical hierarchy included in the attribute parameters of the target table; finally, based on the table physical structure and logical hierarchy included in the attribute parameters of the target table, extracting the table header structure, data area, and table style features included in the attribute parameters of the target table.

[0116] Specifically, in order to accurately analyze the "physical form" and "logical relationship" of the table from unstructured image data, ensure that the extracted table header structure, data area and style features conform to both visual presentation and business logic of the table, when processing image data by using a preset worksheet visual analysis model, the "target detection and OCR recognition → joint coding → feature extraction" technology is adopted. Image data (such as table screenshots, hand-drawn sketches) is a collection of pixels, containing a large amount of irrelevant information, while the system needs is "the effective content of the table" (such as text, lines, cell position). Target detection and OCR recognition are the basis for "stripping off irrelevant information from the image and extracting core content":

[0117] Target detection locates the physical boundaries of the "table area" in the image through the model, such as the outer frame of the table and the cell separation line, which can exclude interference elements such as background and watermark, and clearly "which pixels belong to the table itself". OCR recognition can convert the "imaged text" (such as "customer name" and "amount" in the screenshot) in the table area into "editable text information", solving the problem of "unable to directly read image text".

[0118] The core feature of the table is the unity of "physical form" (such as the position, size and arrangement of cells) and "logical hierarchy" (such as the "dependence of table headers and data" and "parent-child relationship of multi-level table headers"). "Joint coding" (text information + control position information) can establish the mapping of "physical position" and "logical relationship", the text information (OCR result) provides "content semantics", and the control position information (cell coordinates, size and relative position obtained by target detection) provides "spatial constraints". Through joint coding, the logical relationship can be inferred based on spatial constraints, and if only relying on text information, the logical structure of the table may be damaged. The extraction of table header structure, data area and style features depends on the accurate understanding of the "physical structure" and "logical hierarchy" of the table. And "extracting features based on physical structure and logical hierarchy" is to avoid misjudgment caused by "only looking at visual appearance" and ensure that the extracted results not only conform to the visual presentation but also fit the business logic. For table header structure extraction, it can be based on the logical hierarchy (determined by joint coding) to distinguish between "multi-level table headers" (such as "yearly, quarterly and monthly") and "ordinary rows" and avoid misjudging the "first row of the data area" as a table header. For example, if the position of a certain row of text (such as "total") is at the end of the data area and the font is bold (style feature), it can be judged as a "summary row" rather than a table header in combination with the logical hierarchy. For data area extraction, it can be based on the physical structure (cell position and range) to accurately divide the "start and end rows / columns of the data area" (such as "the first 50 rows below the table header are the data area") and confirm the "corresponding relationship between the data area and the table header" (such as "the data in the third column corresponds to the 'amount' table header") in combination with the logical hierarchy. If only relying on position information, it may misjudge the "blank area of the merged cell" as the data area, resulting in range errors. For style feature extraction, it can combine the physical structure (such as the color pixel value and border line thickness of the cell) with the logical hierarchy (such as the semantics of "table header row" and "summary row") to accurately extract "scenario-based styles" (such as "the background color of the table header row is blue" and "the font of the summary row is bold") and avoid misjudging "accidental stains" as "background color" or confusing "table header styles" with "data styles".

[0119] In step S203, the text semantic analysis result of the target table is aligned and fused with the text visual analysis result of the target table to generate a semantic analysis result of the target table.

[0120] Specifically, to solve the problem of "text demand" and "visual demand" being separated, convert the scattered single-modal information into a unified semantic scheme that is "functionally complete, structurally reasonable and visually matched", and ensure that the final generated target worksheet can accurately meet the user's business logic and fit their visual usage habits. The text semantic analysis result of the target table can be aligned and fused with the text visual analysis result of the target table to generate a semantic analysis result of the target table.

[0121] Text semantic analysis and visual analysis extract information based on different dimensions. If they are not aligned and fused, there may be conflicts between "business logic" and "visual form", resulting in generated tables that cannot be used. One of the core roles of alignment and fusion is to identify and resolve such conflicts, ensuring that the logic of the two is consistent. Text semantic analysis extracts "two fields, 'quarterly sales' and'monthly sales', must be included", but visual analysis finds that in the reference graph provided by the user, the "monthly sales" column is merged under "quarterly sales" (the visual form is "quarterly as the first-level table header, monthly as the second-level table header"). If not fused, two errors may occur: generating two independent columns according to the text (destroying the visual hierarchy); and generating a multi-level table header according to the visual, but missing the "quarterly and monthly summary logic" (destroying the business logic). During fusion processing, the "text field association" and "visual hierarchy" are aligned to determine that "quarterly sales" is the first-level table header, "1-3 months" is the second-level table header, and "quarterly sales = the sum of 1-3 month sales" (the logical constraints of the text are embedded in the visual hierarchy), solving the conflict.

[0122] Secondly, there are "information blind spots" in text semantic analysis and visual analysis. Text can clearly specify "what fields are needed, what logic is between fields", but it cannot clearly specify "how to arrange the fields, what they look like"; visual can clearly specify "how to arrange the fields, what they look like", but it cannot clearly specify "why the fields exist, the logical relationship between the fields". Alignment and fusion use the advantages of one modality to fill in the gaps of the other modality, forming a complete demand portrait. Visual supplements the "layout missing" of text: text semantic analysis extracts "customer, order number, amount, date" as four business entities, but does not specify the "column order of the four fields". At this time, the "column order of customer -> order number -> date -> amount in the reference graph" extracted by visual analysis can supplement this information, avoiding the generation of "amount on the leftmost side and customer on the rightmost side" which is against common sense, and improving user convenience. Visual analysis identifies from the reference graph that there are "subtotal and total rows, but the calculation logic is not labeled". At this time, the logical constraints extracted by text semantic analysis, "subtotal = sum of the same customer order amount, total = sum of all subtotals", can supplement the "calculation rules" that visual cannot convey, ensuring that the generated "subtotal / total" rows are not "empty rows", but functional rows that can automatically calculate. For example, the user only provides a "inventory table" screenshot (visual information: there are "product name, inventory quantity, warning value" three columns, and the row where "inventory quantity < warning value" is marked red), without using text to explain the "red marking rule". During fusion, if the text semantic analysis does not extract the logical constraints, it can combine the "red marking feature" of visual to deduce "add the condition format logic of 'highlight when inventory quantity < warning value'", supplementing the rule information missing in the text.

[0123] The finally generated "target table semantic analysis result" needs to meet both "machine understandable" and "user perceptible". It needs to clearly define the structured information such as "fields, logic, data sources", and the user needs to clearly define the visualized information such as "what does the table look like, how to use it". Alignment and fusion can convert scattered text and visual information into a unified semantic scheme that is "machine executable and user recognizable".

[0124] For the system, the fusion result will clearly define the mapping relationship between "business entity and visual location" (such as "customer name" corresponding to "1st column, left alignment, table header bold") and the association rule between "logical constraints and visual representation" (such as "order amount > 1000" corresponding to "cell background color is light blue"). These structured semantic information is the core basis for subsequent "automatic configuration of data source" and "generation of dynamic worksheet" - the machine can accurately know "which field corresponds to which data source field" and "which logic needs to be implemented through format or formula", avoiding execution deviation. For the user, the fusion result will present the "visual effect after the landing of business requirements" (such as "including customer and order information, customer column on the left, amount column right-aligned, over 1000 yuan marked blue"). The user can perceive "whether the final table meets the expectations" in advance through the semantic description after fusion, reducing the subsequent adjustment cost. For example: after fusion, the user is informed that "the table will contain all the order fields you need, and the columns are arranged in the order of the reference diagram, and the amount over 1000 yuan will be marked blue", the user can directly confirm "whether it meets the requirements", without the need to modify after generating the template. If the alignment and fusion step is skipped and the template is generated based on text or visual single modality, subsequent problems such as "fields are complete but layout is chaotic" and "layout is good but logic is missing" may occur, requiring repeated iteration and adjustment. Alignment and fusion can greatly reduce the subsequent optimization cost and improve the one-time matching accuracy by integrating the dual modality information in advance. For example, without fusion, if the template is generated based on text first, it may contain all the fields but the layout is random, the user may feedback "column order is wrong, format is not good-looking", and the visual needs to be adjusted again. If the template is generated based on visual first, it may have a good layout but lack key logic, the user may feedback "no subtotal row, amount will not be calculated automatically", and the logic needs to be supplemented again. The template generated after fusion contains "correct fields, reasonable layout, complete logic, and matching style", the user does not need to repeatedly propose "adjust the layout" and "supplement the logic" requirements, reducing the iteration times and improving the efficiency.

[0125] In step S204, based on the semantic analysis result of the target table, the structured demand information corresponding to the user demand of the target worksheet is determined.

[0126] Specifically, the problem that "semantic analysis results cannot directly guide the generation of workbooks" is that the semantic analysis results are "a complete interpretation of user requirements", while the structured requirement information is "a conversion of the interpretation into machine-executable and standardized table design rules". Therefore, after completing the alignment and fusion of the text semantics and visual semantics of the target table, the structured requirement information of the target worktable corresponding to the user requirements can be determined based on the semantic analysis results of the target table, so as to convert "unstructured semantic interpretation" into "structured design rules". The semantic analysis results of the target table (after text + visual fusion) are "a natural language / descriptive interpretation of user requirements", containing information such as business entities, intentions, logical constraints, and visual features, but these information is "dispersed and non-standardized", and cannot be directly used by machines to generate workbooks (machines need clear "field definition, format parameter, logical formula" rules, etc.). For example, the semantic analysis results may be described as: "record customer order information, including customer name, purchased goods, unit price, quantity, automatically calculate total amount, mark blue where total amount exceeds 1000 yuan, customer name on the left, font in Songti", which is easy for people to understand, but "has no clear execution standard" for machines. The "determination of the structured requirement information of the target worktable" is to convert these descriptive interpretations into machine-executable standardized rules.

[0127] Only through this "structured conversion", can the machine clearly "generate how many columns, what type of each column, how to write the formula, and how to set the format", laying a "executable" foundation for subsequent generation of worktable templates. The structured requirement information is equivalent to the "landing instruction manual" of the requirement interpretation, which converts each information point in the semantic analysis results (whether it is business logic or visual preference) into "unique and clear design rules", ensuring that the table generated by the machine "neither misses any requirement points in the semantics, nor adds designs that are not in the semantics", achieving the precise alignment of "interpretation as design".

[0128] As mentioned earlier, the subsequent process needs to "match structured requirement information with dynamic industry knowledge graph to determine candidate sub-graphs" and "realize initial templates based on candidate sub-graphs". Both steps require "standardized structured requirement information" as "matching reference", otherwise it is impossible to achieve efficient and accurate docking. For "industry knowledge graph matching", the sub-graphs in the knowledge graph store "standardized field definitions and logical rules". Only when the structured requirement information also adopts the standardized format of "field name + data type + logical constraint" can it be "accurately compared" with the knowledge nodes in the sub-graphs, avoiding "matching failure" due to inconsistent formats (such as the semantic description "can automatically calculate total amount" cannot be compared with the rule "order amount = unit price x quantity" in the sub-graph). For "template instantiation", the instantiation process needs "clear field order, format parameters, and formula logic" to generate specific table structures (such as Excel templates). If only based on semantic analysis results (such as "total amount is on the right side"), it is impossible to determine "whether the right side is the 5th column or the 6th column" and "column width" during instantiation. However, structured requirement information clearly states "order amount column is in the 5th column with a column width of 15 characters", so instantiation can be directly generated according to the rules without additional judgment. Structured requirement information is not only "the execution standard of the machine" but also "the clear basis for users to confirm their requirements". Compared with abstract semantic analysis results (such as "contains order-related fields and format meets expectations"), structured requirement information can be presented to users in "tabular and standardized" form (such as "field list, formula rule, and format parameter"), allowing users to intuitively determine "whether it fully meets their own requirements" and avoid subsequent generation and significant adjustment due to "understanding bias".

[0129] For example, users can confirm through structured requirement information whether "all required fields are included" (such as confirming that there is a "customer type" column and it is a drop-down option); "whether the logical rules are correct" (such as confirming that the formula for "order amount" is "unit price x quantity" instead of "unit price + quantity"); and "whether the visual format meets expectations" (such as confirming that "amount over 1000 yuan is blue background, not blue font"). If users find deviations (such as "customer type should include 'potential customer' option"), they can also make "precise adjustments" (change "drop-down options" from "new customer / old customer" to "new customer / old customer / potential customer") based on structured requirement information, without the need to re-analyze semantic analysis results, significantly reducing communication and adjustment costs.

[0130] In practical application, the application can also crawl authoritative information of each industry in real time; and use a preset training language analysis model to extract industry entity relationship information from the crawled authoritative information of each industry and construct industry entity triples; then map the entities and relationships in the industry entity triples to the corresponding continuous vector space of the dynamic industry knowledge graph; so as to find the nearest subgraph of the structured demand corresponding to the industry entity triples in the vector space and embed the entities and relationships in the industry entity triples, so as to update the preset dynamic industry knowledge graph in real time.

[0131] In practical application, the application can also generate a data bloodline report corresponding to the target worksheet structure while generating the target worksheet structure finally determined by the user, wherein the data bloodline report includes the source, calculation logic and basis node in the dynamic industry knowledge graph of each field of the target worksheet structure.

[0132] Specifically, in order to upgrade from "structure delivery" to "full-link traceable value delivery", the corresponding data bloodline report is generated synchronously when the target worksheet structure is generated, which not only solves the problem of "how to use the table", but also answers the key questions of "where does the data come from, where does it go, and how does it change", providing core support for subsequent data use, maintenance and risk control. The "synchronous delivery" of the data bloodline report (recording the flow path, processing rules and correlation of data from the "source" to the "target worksheet") and the target worksheet structure essentially gives the worksheet "explainability" and "manageability", which helps to reduce the data use threshold: let "table users" quickly understand the meaning of data. The target worksheet structure only presents "what fields are there and what are the field types" (such as the "order amount" field is a numerical type), but cannot answer "is this 'order amount' a direct reference to the 'amount' of the original order table, or a calculated result after adding 'coupon deduction'?". The data bloodline report can directly supplement such key information: for example, it clearly states "target table. Order amount = source table A. Original amount - source table B. Coupon amount", and marks the storage location of the source table (such as "database DB_01. Table order_raw"). This can help users (especially non-technical personnel) quickly grasp the "business meaning" of the field, and avoid incorrect analysis caused by "misunderstanding the data source" (such as mistakenly using "calculated amount" as "original amount").

[0133] When the target worksheet has data anomalies, the worksheet structure alone cannot determine whether the problem is due to errors in the source data itself or errors in the processing rules. The data lineage report can provide a "full link trace path" to help technical personnel quickly narrow down the scope of investigation and avoid the inefficient problem of "blindly checking the full link". In addition, the current data compliance requirements clearly require "data processing activities to be traceable", especially for fields involving personal information. The synchronously generated lineage report can directly record the source and processing process of each field in the target worksheet. It can quickly respond to compliance audits and also identify "sensitive data flow risks" in a timely manner. The data lineage report can serve as a "maintenance manual" to simplify later maintenance and iteration, reduce table modification costs, and only need to expand based on existing lineage relationships when modifying, while also checking "whether the modification affects downstream" to avoid the chain reaction caused by "modifying a single hair and affecting the whole body".

[0134] The generation of the data lineage report needs to focus on the "fields of the target worksheet". The process of generating a data lineage report corresponding to the target worksheet structure can be as follows:

[0135] 1. The "full link data element information" of the target worksheet can be collected first

[0136] The prerequisite for generating a lineage report is to obtain the "association information between the target table and the upstream data". Two types of core element information need to be collected to ensure that there is no omission. Extract the target worksheet structure element information from the final determined worksheet structure, including "field name, field type, field description, primary key / foreign key". Also extract the upstream data link element information: trace the "data source" of the target table field, and collect 3 types of information:

[0137] (1) Source data table information: the field corresponding to the upstream table name, storage location, and source field name;

[0138] (2) Data processing rule information: record the "processing logic" of the field from the source to the target table, including calculation rules, conversion rules, and filtering rules;

[0139] (3) Data flow link information: if there is an "intermediate processing table", the complete link needs to be recorded, and the processing tool at each step needs to be marked.

[0140] In practice, table structure element information can be obtained through a data dictionary (such as Hive Metastore, MySQL Information Schema); processing rules and flow links can be obtained through ETL tool logs (such as DataWorks, Flink task logs); if unstructured data (such as Excel source files) is involved, "data source specification documents" provided by business personnel need to be manually supplemented.

[0141] 2. Analyze "field-level lineage relationships" and establish a mapping model

[0142] Based on the collected meta-information, the one-to-one / one-to-many blood relationship between the target table field and the upstream data can be determined through the method of "technical analysis + business verification", avoiding analysis errors, for example, using "blood analysis tools" to automatically identify field mapping relationships and performing reverse analysis based on "data processing scripts / task configurations". Open source tools (such as Apache Atlas, DataHub) can automatically analyze the blood of SQL scripts and ETL tasks; commercial tools (such as Informatica DataGovernance, IBM InfoSphere) support the integration of blood from multiple data sources. Secondly, technical analysis may have "semantic bias", which needs to be confirmed by business personnel. The business rationality of processing rules can be checked, and "business-level blood explanation" can also be supplemented.

[0143] 3. Structured output data blood report, clear presentation dimensions

[0144] The parsed blood relationship is converted into a structured report that is understandable by humans and recognizable by machines, which needs to include five core presentation dimensions to ensure complete and clear information. In practice, there are two versions of data blood reports, one is a "visual report" (such as HTML / PDF format, including flowchart) for business personnel to view, and the other is a "structured file" (such as JSON format, easy to access data governance platform later) for machine calling.

[0145] 4. Synchronize the consistency of the report with the target worksheet

[0146] After generating the report, the last step of verification is needed to ensure that the blood report and the target worksheet structure are "completely matched", avoiding "field omission, rule error", for example, checking whether each field of the target worksheet has a corresponding entry in the blood report, and if there is omission, it needs to be traced back and supplemented. It can also be confirmed through "sample data verification" whether the processing rules are correct, and whether the "data transfer link" in the report is consistent with the actual ETL task execution link. In practice, an automated script (such as Python) can be written to compare the "target table field list" with the "report field list"; the data quality tool (such as Great Expectations) can be used to verify the accuracy of the processing rules.

[0147] For example, the specific implementation process of the present application can be as follows: the name and table content setting information of the to-be-generated worksheet A can be received; then the name and table content setting information of the worksheet A are subjected to semantic analysis by the LLM large model to extract key entity information and semantic relationship information; the industry information of the worksheet A is generated according to the key entity information and the semantic relationship information; the industry information is matched with the industries in the pre-constructed industry knowledge base to obtain the industry to which the worksheet A belongs; wherein, the industry knowledge base covering multiple industries can be pre-constructed, and the worksheet structure template, field name, data type and association relationship corresponding to each industry are stored in the industry knowledge base; the key entity information and the semantic relationship information are matched with the corresponding industry in the industry knowledge base, and the industry with the highest matching degree is taken as the industry to which the worksheet A belongs. Then the worksheet A structure template is determined according to the industry to which the worksheet A belongs; the worksheet A structure template is adjusted according to the table content setting information to generate the worksheet structure A. For example, the field name, data type and association relationship are determined according to the table content setting information; the worksheet A structure template is adjusted according to the field name, data type and association relationship, and then the worksheet structure A can be generated. After the worksheet structure A is generated, the component information of each form component in the worksheet structure A can also be output one by one in a streaming output manner; then the worksheet structure is rendered according to the component information of all the form components to generate the interface of the worksheet A, wherein the component information of the worksheet structure A includes: component name, component type, verification logic, default value and prompt information. After the interface of the worksheet A is generated, feedback information of the user can also be received, and the feedback information is subjected to semantic analysis by the LLM large model to obtain modification intention information; so that the worksheet structure A can be adjusted according to the modification intention information, and the worksheet interface is re-rendered according to the adjusted worksheet structure A2.

[0148] The following will be described in combination with Figure 2 , a system architecture for realizing dynamic production of a zero-code platform is introduced, which can be applied to the natural language-based zero-code worksheet dynamic generation method described above, such as Figure 2As shown, the system architecture can include: a multi-modal input interface module, a cross-modal semantic fusion analysis module, a dynamic industry knowledge graph module, a graph neural network matching processing module, a dynamic template generation module, an interaction optimization module, a deployment module; wherein the cross-modal semantic fusion analysis module includes a text analysis submodule, a visual analysis submodule, a multi-modal fusion submodule. The multi-modal input interface module can receive and process the input data of the user, and determine the attribute parameters of the target table to be processed and transmit them to the cross-modal semantic fusion analysis module, wherein the attribute parameters of the target table include the name of the target worksheet to be generated, the content setting information, the image data, the document data; the input mode of the user includes natural language text input, image input, document upload input, voice input. The text analysis submodule can use a pre-trained language model based on Transformer to perform deep semantic analysis on the name of the target worksheet to be generated and the content setting information included in the attribute parameters of the target table, which can include named entity recognition, relationship extraction, intent classification, etc. of the target worksheet to be generated; the multi-head attention mechanism can be used to capture long-distance dependency relationships; the business entities, intents and logical constraint information between fields of the target worksheet to be generated are extracted as the text semantic analysis results of the target table, and all the analysis results are transmitted to the multi-modal fusion submodule. The visual analysis submodule can use a preset worksheet visual analysis model to analyze the image data included in the attribute parameters of the target table to extract the header structure, data area and table style features included in the attribute parameters of the target table as the text visual analysis results of the target table and transmit them to the multi-modal fusion submodule. In practice, the preset worksheet visual analysis model can be trained with the image data included in the attribute parameters of the target table as training samples, and the header structure, data area and table style features contained in the image data included in the attribute parameters of the target table as sample labels, and trained. The multi-modal fusion submodule can align and fuse the text semantic analysis results of the target table and the text visual analysis results of the target table based on the multi-modal fusion mechanism of cross attention, and generate target worksheet structured demand information corresponding to the user demand and transmit it to the graph neural network matching processing module. The target worksheet structured demand information corresponding to the user demand represents the relevant information of the worksheet that the user wants. Therefore, the graph neural network matching processing module can be responsible for similarity matching analysis of the target worksheet structured demand information and the preset dynamic industry knowledge graph, and calculating the matching degree of the target worksheet structured demand information and each subgraph in the preset dynamic industry knowledge graph and transmitting it to the dynamic industry knowledge graph module.The dynamic industry knowledge graph module can determine at least one candidate industry subgraph corresponding to the structured demand information of the target worksheet according to the matching degree between the structured demand information of the target worksheet and each subgraph in the preset dynamic industry knowledge graph and transmit the at least one candidate industry subgraph to the dynamic template generation module; in order to dynamically update the industry knowledge graph, the dynamic industry knowledge graph module can further include a knowledge acquisition submodule, a knowledge storage submodule and a knowledge updating submodule, wherein the knowledge acquisition submodule can be responsible for establishing an automated knowledge extraction pipeline to continuously extract knowledge from industry websites, standard document libraries and API documents; a BERT-based relation joint extraction model can be used to extract industry entity triples; the knowledge storage submodule is responsible for storing each industry knowledge using a graph database to support complex graph query and reasoning, wherein the mode layer of the knowledge graph includes multiple dimensions such as industry ontology, data object, business rule and calculation formula; the knowledge updating submodule is responsible for implementing a knowledge freshness evaluation mechanism based on time decay to automatically update outdated knowledge regularly and ensure the timeliness of the dynamic industry knowledge graph. In the field of information extraction in natural language processing (NLP), "relation extraction" is one of the core tasks, and the goal is to extract structured triples (such as `<Xiaoming, lives in, Beijing>` and `<Apple Inc., founder, Steve Jobs>`) of "entity pair + relationship type" from unstructured text. Traditional relation extraction follows a two-step pipeline mode, while "joint extraction" is born to solve the defects of the pipeline mode, as shown in Table 4.

[0149]

[0150] BERT (Bidirectional Encoder Representations from Transformers) is a pre-trained language model that can capture the contextual semantic information of text (such as the different meanings of "apple" in "eating an apple" and "Apple Inc.") through a bidirectional Transformer. In relation joint extraction, the role of BERT is "semantic encoder", which converts the original text into vector representation (i.e. "word embedding") containing rich contextual information, providing high-quality semantic input for subsequent "entity recognition" and "relationship judgment". Compared with traditional word embeddings (such as Word2Vec, GloVe, which can only provide static, context-free vectors), the advantage of BERT is that the vectors of the same word in different contexts are different, which is more consistent with the ambiguity of natural language; it can capture long-distance dependencies (such as "Xiaoming joined A Company in 2020, and he is now the product manager of the company", where "he" and "the company" are related to "Xiaoming" and "A Company").

[0151] The overall process of the BERT relation joint extraction model can be summarized as "BERT encoding → joint decoding (entity + relation)". The joint decoding is the technical core, and the mainstream idea is divided into two categories: "parameter sharing multi-task learning" and "unified label modeling". Parameter sharing multi-task learning (Multi-Task Learning) mainly regards "entity extraction" and "relation judgment" as two sub-tasks, shares the encoder parameters of BERT, and optimizes through two independent decoders (such as CRF, full connection layer) respectively, and finally gets the triplets through "entity pair matching". The specific process is as follows:

[0152] 1. BERT encoding: input text (such as "A founded B company"), output the context vector of each token (such as the vectors of "A", "create", "establish", "B", "public", "company").

[0153] 2. Entity extraction sub-task: a CRF (Conditional Random Field) layer is connected after the BERT output layer (a classic layer for processing sequence labeling, which can capture the boundary constraints of entities, such as "B-Person" can only be followed by "I-Person" or "O"), complete entity labeling (such as "A" is labeled as "B-Person → I-Person → I-Person", "B company" is labeled as "B-Organization → I-Organization → I-Organization → I-Organization").

[0154] 3. Relation judgment sub-task: a fully connected layer is connected after the BERT output layer, input "entity pair vector" (usually take the average or maximum value of entity token vector), judge the relation type of the entity pair (such as `<A, B company>` relation is "founder", `<B company, A>` is "co-founder", and no relation is marked as "NA").

[0155] 4. Triple generation: combine the extracted entities two by two, filter out the entity pairs with "NA" relation, and finally get the effective triplets (such as `<A, founder, B company>`).

[0156] Unified label modeling (Unified Labeling) does not separate "entity" and "relation" tasks, but designs a unified label system to bind "entity position" and "relation type", and directly outputs triplets through a single generator (such as sequence labeling, pointer network). This idea more thoroughly realizes "joint", and the typical representative is TPLinker (Token Pair Linking), whose core design is as follows: label definition: the "entity pair" is represented as "token pair" (such as the i-th token and the j-th token in the text), and the label contains two types of information:

[0157] 1. Entity Type: Mark whether a token pair is the "start-end of an entity" (e.g., `<1,3>` means that the 1st token to the 3rd token is a "person" class entity, corresponding to "Jobs");

[0158] 2. Relationship Type: Mark whether a token pair is the "head entity start-tail entity start" (e.g., `<1,7>` means that the 1st token is the head entity start of the "founder" relationship, and the 7th token is the tail entity start, corresponding to "A → B company").

[0159] After BERT encoding, a classifier is used to predict the labels of all token pairs, and then the "head entity-relation-tail entity" triplets are spliced according to the label rules. This completely avoids the redundant combination of entity pairs (such as the need to traverse all entity pairs in parameter sharing multi-task learning), directly locates the entity boundaries corresponding to the effective relationship, and is more efficient and accurate.

[0160] Take the commonly used "BERT + CRF + relationship classification" architecture as an example to intuitively understand the model structure:

[0161] Input text: [CLS] A founded B company [SEP] ( [CLS] is the sentence-level vector mark of BERT, and [SEP] is the separator); BERT encoder: output the context vector of each token (including the vector of [CLS]);

[0162] Branch 1: Entity extraction (CRF layer), entity labeling result: O B-Person I-Person I-Person O O O B-Organization I-Organization I-Organization I-Organization O; Extracted entities: Person (A), Organization (B company).

[0163] Branch 2: Relationship classification (full connection layer), input entity pair vector (A vector, B company vector); relationship prediction result: founder (non-NA).

[0164] Final triplet: <A, founder, B company>;

[0165] Compared with the traditional pipeline model or the joint model without BERT, the BERT joint extraction model has stronger semantic understanding. The bidirectional context encoding of BERT solves the problems of "semantic ambiguity" and "long-distance dependence" of the traditional model. For example, it can accurately identify that "C" is an "organization" entity in "C released a new mobile phone" and "Xiaoming works at C". There is less error propagation. The joint extraction mode avoids the chain reaction of "entity extraction error leading to relationship error" in the pipeline. For example, even if the entity boundary is slightly biased, the semantic information of the relationship judgment can correct the entity annotation in the opposite direction. The generalization ability is better. The pre-training-fine-tuning mode of BERT (pre-training on general corpus and fine-tuning on specific relationship dataset) enables it to perform well in small sample scenarios. For example, only a small amount of "partner" relationship data can accurately extract similar relationships.

[0166] The dynamic template generation module can be responsible for instantiating an initial worksheet structure template based on the determined various candidate industry subgraphs, obtaining a target worksheet structure template and transmitting it to the interactive optimization module, wherein the target worksheet structure template includes various field definitions of the worksheet, connection logic of the worksheet and external data sources, and cross-table reference formulas based on preset dynamic industry knowledge graph relationships; the interactive optimization module presents the target worksheet structure template to the user and captures all operations of the user on the target worksheet structure template; dynamically analyze all operations of the user on the target worksheet structure template, dynamically optimize the target worksheet structure template, until the user finally determines the target worksheet structure and transmits it to the deployment module; the deployment module can present the final target worksheet structure to the user again after the user confirms the final target worksheet structure, and automatically configure the external data sources corresponding to the target worksheet structure, generate a dynamic worksheet corresponding to the user's demand and display it to the user. Wherein, the specific processing flow of each module included in the above natural language-based zero-code worksheet dynamic generation system can refer to the relevant introduction in the previous natural language-based zero-code worksheet dynamic generation method part, which will not be repeated here.

[0167] The zero-code platform table dynamic generation device of the present application can be applied to a natural language-based zero-code worksheet dynamic generation equipment, such as a terminal: mobile phone, computer, etc. Optionally, Figure 3 The hardware structure block diagram of the natural language-based zero-code worksheet dynamic generation equipment is shown, which can be referred to in the previous Figure 3The hardware structure of the natural language based zero code worksheet dynamic generation device can include at least one processor 1, at least one communication interface 2, at least one memory 3 and at least one communication bus 4. In the present application, the number of the processor 1, the communication interface 2, the memory 3 and the communication bus 4 is at least one, and the processor 1, the communication interface 2 and the memory 3 complete the communication with each other through the communication bus 4. The processor 1 can be a central processing unit CPU, or a specific integrated circuit ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the present application, etc.; the memory 3 can include a high-speed RAM memory, and can also include a non-volatile memory, etc., for example, at least one disk memory; wherein the memory stores a program, and the processor can call the program stored in the memory, and the program is used to: realize each processing flow in the terminal based natural language zero code worksheet dynamic generation scheme described above. The present application also provides a readable storage medium, which can store a program suitable for the processor to execute, and the program is used to: realize each processing flow in the terminal based natural language zero code worksheet dynamic generation scheme described above.

[0168] Finally, it should also be noted that, in this document, the relationship terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or device. Without more limitations, the element defined by the statement "including a" does not exclude the presence of another identical element in the process, method, article or device including the element. The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments, and the same or similar parts between the embodiments can be referred to each other. The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to the embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. The various embodiments can be combined with each other. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A natural language-based zero-code worksheet dynamic generation method, characterized in that, The method comprises: determining attribute parameters of a target table to be processed, wherein the attribute parameters of the target table comprise a name, content setting information and image data of a target worksheet to be generated; performing multi-dimensional cross-modal semantic fusion analysis on the attribute parameters of the target table to determine target worksheet structured demand information corresponding to user demand; performing similarity matching analysis on the target worksheet structured demand information and a preset dynamic industry knowledge graph, and calculating matching degrees of the target worksheet structured demand information and each subgraph in the preset dynamic industry knowledge graph; determining at least one candidate industry subgraph corresponding to the target worksheet structured demand information according to the matching degrees of the target worksheet structured demand information and each subgraph in the preset dynamic industry knowledge graph; instantiating an initial worksheet structure template based on each of the determined candidate industry subgraphs to obtain a target worksheet structure template, wherein the target worksheet structure template comprises field definitions of the worksheet, connection logic of the worksheet and an external data source, and cross-table reference formulas based on preset dynamic industry knowledge graph relationships; presenting the target worksheet structure template to a user and capturing all operations of the user on the target worksheet structure template; dynamically analyzing all operations of the user on the target worksheet structure template, dynamically optimizing the target worksheet structure template, and generating a target worksheet structure finally determined by the user; after the user confirms the final target worksheet structure, presenting the final target worksheet structure to the user again, automatically configuring an external data source corresponding to the target worksheet structure, and generating a dynamic worksheet corresponding to the demand of the user and displaying the dynamic worksheet to the user.

2. The method of claim 1, wherein, The multi-dimensional cross-modal semantic fusion analysis on the attribute parameters of the target table to determine target worksheet structured demand information corresponding to user demand comprises: performing multi-dimensional cross-modal semantic fusion analysis on the attribute parameters of the target table to obtain semantic analysis results of the target table; based on the semantic analysis results of the target table, determining target worksheet structured demand information corresponding to user demand.

3. The method of claim 2, wherein, The multi-dimensional cross-modal semantic fusion analysis on the attribute parameters of the target table to obtain semantic analysis results of the target table comprises: performing deep semantic analysis on the name, content setting information of the worksheet to be generated included in the attribute parameters of the target table, and extracting business entities, intents and field logical constraint information of the worksheet to be generated as text semantic analysis results of the target table; The image data included in the attribute parameters of the target table is analyzed by using a preset worksheet visual analysis model, and the table header structure, data area and table style features included in the attribute parameters of the target table are extracted as the text visual analysis result of the target table, wherein the preset worksheet visual analysis model takes the image data included in the attribute parameters of a training target table as a training sample, takes the table header structure, data area and table style features contained in the image data included in the attribute parameters of the training target table as a sample label, and is trained to obtain; The text semantic analysis result of the target table is aligned and fused with the text visual analysis result of the target table to generate a semantic analysis result of the target table.

4. The method of claim 3, wherein, The process of analyzing the image data included in the attribute parameters of the target table by using the preset worksheet visual analysis includes: Performing target detection and OCR recognition processing on the image data included in the attribute parameters of the target table; Jointly encoding the recognized text information and the control position information of the image data included in the attribute parameters of the target table to determine the table physical structure and logical hierarchy included in the attribute parameters of the target table; Based on the table physical structure and logical hierarchy included in the attribute parameters of the target table, the table header structure, data area and table style features included in the attribute parameters of the target table are extracted.

5. The method of claim 1, wherein, The method further includes: Real-time crawling of authoritative information in each industry; Extracting industry entity relationship information from the crawled authoritative information in each industry by using a preset training language analysis model and constructing an industry entity triple; Integrating the industry entity triple into the original dynamic industry knowledge graph to update the preset dynamic industry knowledge graph in real time.

6. The method of claim 5, wherein, The integration of the industry entity triple into the original dynamic industry knowledge graph includes: Mapping the entities and relationships in the industry entity triple to the corresponding continuous vector space of the dynamic industry knowledge graph; Finding the nearest subgraph to the structured demand representation corresponding to the industry entity triple in the vector space and embedding the entities and relationships in the industry entity triple.

7. The method of claim 1, wherein, The method further includes: while generating a target worksheet structure finally determined by a user, generating a data blood relationship report corresponding to the target worksheet structure, wherein the data blood relationship report includes the source, calculation logic and basis node in the dynamic industry knowledge graph of each field of the target worksheet structure.

8. A natural language based zero code worksheet dynamic generation system, characterized in that, The system is applied to the natural language-based zero-code worksheet dynamic generation method of any one of claims 1-7 and includes a multi-modal input interface module, a cross-modal semantic fusion analysis module, a dynamic industry knowledge graph module, a graph neural network matching processing module, a dynamic template generation module, an interaction optimization module and a deployment module; wherein the cross-modal semantic fusion analysis module includes a text analysis submodule, a visual analysis submodule and a multi-modal fusion submodule. The multi-modal input interface module receives and processes input data of a user, determines attribute parameters of a target table to be processed, and transmits the attribute parameters to the cross-modal semantic fusion analysis module. The text analysis submodule adopts a pre-training language model based on a Transformer to perform deep semantic analysis on the name and content setting information of the worktable to be generated included in the attribute parameters of the target table, uses a multi-head attention mechanism to capture long-distance dependency relationships, extracts business entities, intents, and logical constraint information among fields of the worktable to be generated as a text semantic analysis result of the target table, and transmits the text semantic analysis result to the multi-modal fusion submodule. The visual analysis submodule uses a preset worktable visual analysis model to analyze image data included in the attribute parameters of the target table, extracts table header structures, data regions, and table style features included in the attribute parameters of the target table as a text visual analysis result of the target table, and transmits the text visual analysis result to the multi-modal fusion submodule. The multi-modal fusion submodule performs alignment and fusion processing on the text semantic analysis result of the target table and the text visual analysis result of the target table based on a multi-modal fusion mechanism of cross attention, dynamically allocates weights by using an algorithm, generates target worktable structured demand information corresponding to the user demand, and transmits the target worktable structured demand information to the graph neural network matching processing module. The graph neural network matching processing module is responsible for similarity matching analysis of the target worktable structured demand information and a preset dynamic industry knowledge graph, calculates matching degrees of the target worktable structured demand information and each subgraph in the preset dynamic industry knowledge graph, and transmits the matching degrees to the dynamic industry knowledge graph module. The dynamic industry knowledge graph module determines at least one candidate industry subgraph corresponding to the target worktable structured demand information according to the matching degrees of the target worktable structured demand information and each subgraph in the preset dynamic industry knowledge graph, and transmits the at least one candidate industry subgraph to the dynamic template generation module. The dynamic industry knowledge graph module further includes a knowledge acquisition submodule, a knowledge storage submodule, and a knowledge updating submodule. The knowledge acquisition submodule is responsible for establishing an automatic knowledge extraction pipeline to continuously extract knowledge from industry websites, standard document libraries, and API documents. A relation joint extraction model based on BERT is used to extract industry entity triples. The knowledge storage submodule is responsible for storing each industry knowledge using a graph database to support complex graph query and reasoning. The knowledge updating submodule is responsible for implementing a knowledge freshness evaluation mechanism based on time decay to automatically update outdated knowledge at regular intervals and ensure the timeliness of the dynamic industry knowledge graph. The dynamic template generation module is responsible for instantiating an initial worktable structure template based on each of the candidate industry subgraphs, obtaining a target worktable structure template, and transmitting the target worktable structure template to the interaction optimization module. The interaction optimization module presents the target worksheet structure template to the user and captures all operations of the user on the target worksheet structure template; dynamically analyzes all operations of the user on the target worksheet structure template, dynamically optimizes the target worksheet structure template, until a target worksheet structure finally determined by the user is generated and transmitted to the deployment module; The deployment module presents the final target worksheet structure to the user again after the user confirms the final target worksheet structure, automatically configures the external data source corresponding to the target worksheet structure, and generates and displays a dynamic worksheet corresponding to the demand of the user to the user.

9. A natural language based zero code worksheet dynamic generation device, characterized in that, Comprise: One or more processors, and a memory; The memory has computer readable instructions stored therein, and the computer readable instructions are executed by the one or more processors to implement the steps of the natural language based zero code worksheet dynamic generation method according to any one of claims 1 to 7.

10. A readable storage medium characterized by: The readable storage medium has computer readable instructions stored therein, and the computer readable instructions are executed by one or more processors to make one or more processors implement the steps of the natural language based zero code worksheet dynamic generation method according to any one of claims 1 to 7.

Images