An AI intelligent form automatic generation system based on a large model

Abstract

The application discloses an AI intelligent form automatic generation system based on a large model and relates to the technical field of natural language large models.The application maps form demand description text into a semantic feature vector, establishes multiple business fields according to a preset business type, inputs the semantic feature vector into the multiple business fields to obtain candidate fields, acquires a semantic concentration value of each candidate field, establishes an intent field tree based on the semantic concentration value, matches field nodes that exist in the intent field tree, outputs a demand feature correlation vector according to a matching result, acquires an included angle value between the demand feature correlation vector and a preset standard rule vector, intercepts a field constraint network from the intent field tree according to the size of the included angle value, sets field controls according to the types of field nodes in the field constraint network, and maps each field control in a form planning canvas based on the priority and the correlation relationship between the field nodes.

Description

Technical Field

[0001] This invention relates to the field of natural language processing large model technology, specifically to an AI-powered intelligent form automatic generation system based on a large model. Background Technology

[0002] In current enterprise digital business processes, forms serve as the core carrier for information collection and business flow, and are widely used in various scenarios such as administrative approval, business application, and data registration. Traditional form generation methods often rely on manual configuration of each field, manual sorting of validation rules, and interface layout. This not only results in long setup cycles and high labor costs, but also heavily depends on the business experience of the configuration personnel, making it prone to problems such as missing fields, conflicting rule logic, and insufficient layout rationality, and making it difficult to quickly respond to complex and ever-changing business needs.

[0003] Existing automatic form generation solutions largely rely on template matching and keyword extraction technologies, lacking sufficient semantic understanding depth. They fail to accurately interpret the multi-layered business intent within natural language requirements, easily overlooking implicit field requirements and logical connections. Furthermore, constraint rules often depend on pre-defined rule bases for enumeration and matching, limiting flexibility and scalability. When facing emerging business scenarios, rule gaps or logical deviations are prone to occur, resulting in generated forms that generally suffer from chaotic field hierarchies and ambiguous grouping logic, leading to poor information collection efficiency and a subpar user experience. In addition, most solutions lack closed-loop optimization mechanisms, failing to continuously iterate and generate results based on actual usage data, making it difficult to adapt to the dynamic evolution of business needs. Therefore, this paper proposes an AI-powered intelligent automatic form generation system based on a large-scale model. Summary of the Invention

[0004] The purpose of this invention is to provide an AI-powered intelligent form automatic generation system based on a large model to address the shortcomings of the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: An AI-powered intelligent form generation system based on a large model includes a multi-intent analysis module, an intent field planning module, a constraint rule construction module, and an automatic form generation module. The multi-intent analysis module is used to obtain form requirement description text, map the form requirement description text into semantic feature vector, establish multi-business fields according to preset business types, and input the semantic feature vector into the multi-business fields to obtain candidate fields. The intent field planning module is used to obtain the semantic concentration value of each candidate field, select core fields and auxiliary fields based on the semantic concentration value, and then build an intent field tree; The constraint rule construction module is used to match the related field nodes in the intent field tree, output the requirement feature association vector according to the matching result, obtain the angle value between the requirement feature association vector and the preset standard rule vector, and extract the field constraint network from the intent field tree according to the size of the angle value. The automatic form generation module is used to set up the form planning canvas, set up field controls according to the types of field nodes in the field constraint network, and map each field control to the form planning canvas based on the priority and association between field nodes.

[0006] Furthermore, the process of mapping the form requirement description text into semantic feature vectors and establishing multiple business domains based on preset business types includes: The form requirement description text is segmented into Chinese and named entity recognition is performed to remove meaningless stop words and modal particles. Business entities, attribute descriptions and logical relationship expressions in the text are identified to obtain a structured semantic unit sequence. Then, the semantic unit sequence is mapped to generate a semantic feature vector with dimension . Based on the preset business type system, retrieve the target feature vectors of each business type from the corresponding domain knowledge base, with each business type corresponding to one target feature vector; Based on the target feature vectors of each business type, business fields are constructed respectively. Each business field is centered on the corresponding target feature vector, and a standardized business field radius is set. The generated semantic feature vectors are sequentially input into all multi-business fields. The cosine similarity between the semantic feature vector and each target feature vector is calculated. The business field with the highest cosine similarity is selected as the main processing field, and the other business fields are used as auxiliary reference fields.

[0007] Furthermore, the process of inputting semantic feature vectors into a multi-service domain to obtain candidate fields includes: For semantic feature vectors in the multi-business domain, denoted as semantic particles, in each iteration, the direction cosine similarity between the particle and the target feature vector is calculated, and it is also subject to the bias correction effect pointing towards the target feature vector. The iteration round is set to 10 rounds. After each iteration, the spatial coordinates and semantic intensity of all semantic particles are updated. After multiple iterations and convergence, the spatial distribution of candidate field seeds converges towards the global target. Redundant semantic particles that deviate from the core business gradually diffuse, decay and are eliminated, finally obtaining all effective candidate fields, as well as the relative coordinates of each candidate field in the multi-business domain.

[0008] Furthermore, the process of obtaining the semantic concentration value of the candidate field includes: The semantic concentration value is used to characterize the degree of cohesion and business relevance of candidate fields in the business field. It is obtained by taking the semantic intensity of the semantic particle corresponding to the candidate field and adding the cosine similarity between the target feature vector and the semantic feature vector corresponding to the candidate field.

[0009] Furthermore, the process of selecting core and auxiliary fields based on semantic concentration values, and then constructing the intent field tree, includes: Two levels of concentration thresholds are set, namely the first concentration threshold and the second concentration threshold, and the thresholds can be dynamically adjusted according to the business type; Candidate fields with semantic concentration values ​​higher than the first concentration threshold are classified as core fields, candidate fields with semantic concentration values ​​between the first and second concentration thresholds are classified as auxiliary fields, and candidate fields with semantic concentration values ​​lower than the second concentration threshold are considered redundant fields and are directly eliminated. Set information value indicators, use the core fields obtained from the screening as first-level nodes, and cluster them according to business type to form a first-level grouping structure; auxiliary fields are used as second-level nodes, and are attached to the first-level nodes of the corresponding business type according to their semantic relationship. After completing the adjustment of node hierarchy and grouping, the semantically duplicated field nodes are merged, and then the intent field tree is output.

[0010] Furthermore, the process of matching related field nodes in the intent field tree and outputting a requirement feature association vector based on the matching results includes: Iterate through all field nodes in the intent field tree, randomly combine and match any two field nodes, and try to generate existing validation expressions, dependency conditions or cascading filtering logic one by one. Based on the matching results of all field pairs, all candidate association constraint rules are output and imported into the compliance rule library. The compliance rule library stores standard validation rules, data security specifications and industry compliance requirements for forms in various business domains. Each standard rule is pre-coded into a standardized standard rule vector, and the same initial confidence score is set for each candidate association constraint rule.

[0011] Furthermore, the process of obtaining the angle between the demand feature association vector and the preset standard rule vector, and then extracting the field constraint network from the intent field tree based on the angle value, includes: Each candidate association constraint rule is converted into a requirement feature association vector with consistent dimensions through semantic encoding. The average direction angle between the corresponding requirement feature association vector and all standard rule vectors of the same type in the compliance rule base is calculated. An angle threshold is preset. Based on the relationship between the average direction angle and the angle threshold, the initial confidence score of each candidate association constraint rule is increased or decreased. Candidate association constraint rules with confidence scores lower than 0.2 * initial confidence score are eliminated. After filtering all candidate association constraint rules, the effective constraint relationships between the field nodes corresponding to the remaining candidate association constraint rules are extracted from the intent field tree to form a field constraint network.

[0012] Furthermore, the process of setting field controls according to the types of field nodes in the field constraint network, and mapping each field control to the form planning canvas based on the priority and association between field nodes, includes: Initialize the form planning canvas, match the form interaction controls according to the field type and data format of each field node in the field constraint network, and map and arrange each field control in the form planning canvas according to the priority level and association relationship of the field nodes and layout rules. The layout follows the principles of grouping and centralization, priority sorting, association and proximity, and grid adaptation. After the field controls are arranged, operation buttons are generated at the bottom of the form planning canvas, and the corresponding field control fill format prompt text is automatically labeled according to the association constraint rules in the field constraint network. The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention employs a multi-intent analysis mechanism based on multi-business field radiation and gravity correction. By guiding semantic particles to converge toward the core business through target feature vectors, it can mine explicit and implicit field requirements in natural language needs, effectively reducing redundant semantic interference. By combining semantic concentration grading and information value indicators to construct an intent field tree, it can realize the hierarchical sorting and business grouping of fields, effectively improving the completeness and business adaptability of form fields.

[0013] 2. This invention constructs a field constraint network through vector angle matching and phototropic decay mechanism, which can automatically generate validation, cascading and dependency rules between fields. Combined with a compliance rule library, it ensures the rationality and compliance of rule logic, effectively avoiding omissions and logical conflicts in manually configured rules. It is equipped with an automated grid layout generation capability, which can quickly output clear, user-friendly and implementable forms. At the same time, it has the ability to self-iterate with data feedback, and can continuously optimize the generation quality to adapt to dynamic changes in business. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0015] Figure 1 This is a system block diagram of the present invention.

[0016] Figure 2 This is a flowchart of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 and Figure 2 As shown, an AI-powered intelligent form automatic generation system based on a large model includes a multi-intent analysis module, an intent field planning module, a constraint rule construction module, and a form automatic generation module. The multi-intent analysis module is used to obtain form requirement description text, map the form requirement description text into semantic feature vector, establish multi-business fields according to preset business types, and input the semantic feature vector into the multi-business fields to obtain candidate fields; The intent field planning module is used to obtain the semantic concentration value of each candidate field, select core fields and auxiliary fields based on the semantic concentration value, and then build an intent field tree; The constraint rule construction module is used to match the related field nodes in the intent field tree, output the requirement feature association vector according to the matching result, obtain the angle value between the requirement feature association vector and the preset standard rule vector, and extract the field constraint network from the intent field tree according to the size of the angle value. The automatic form generation module is used to set up the form planning canvas, set up field controls according to the types of field nodes in the field constraint network, and map each field control to the form planning canvas based on the priority and association between field nodes.

[0019] The working principle of the present invention is illustrated below through examples: Users can upload a text description of their form requirements in natural language to the Multi-Intent Analysis module. For example, "I need to build a summer vacation departure registration form for students to submit their summer vacation departure applications. It needs to include the student's name, student ID, college name, departure date, counselor's name, return date, and reason for departure." Users can also choose to indicate the business type of the form. If the user does not indicate it, the Multi-Intent Analysis module will automatically identify it. It should be noted that the form requirement description text supports multiple input formats, including plain text description, speech-to-text text, and text recognized from screenshots of historical forms. All of these are uniformly converted into a standardized text format before further processing to ensure compatibility with different input channels.

[0020] Furthermore, the multi-intent analysis module receives the form's requirement description text and performs candidate field extraction and semantic localization processing: The form requirement description text is segmented into Chinese and named entity recognition is performed to remove meaningless stop words and modal particles. Business entities, attribute descriptions and logical relationship expressions in the text are identified to obtain a structured semantic unit sequence. Then, the semantic unit sequence is mapped to generate a semantic feature vector with dimension . Based on the preset business type system, the target feature vectors of each business type in the corresponding domain knowledge base are retrieved. The preset business types cover mainstream application scenarios such as grade summary, activity / exam registration, and work report. Each business type corresponds to a unique target feature vector. The target feature vector is generated by large-scale model aggregation and encoding of the core collection target, set of required fields, and business value orientation of the form in the domain knowledge base, representing the core task direction of the form. Based on the target feature vectors of each business type, corresponding business fields are constructed. Each business field is centered on the corresponding target feature vector, and a standardized business field radius is set. The generated semantic feature vectors are sequentially input into all multi-business fields. The cosine similarity between the semantic feature vector and each target feature vector is calculated. The business field with the highest cosine similarity is selected as the main processing field, and the other business fields are used as auxiliary reference fields, which are only used to supplement edge semantic information. Finally, the extraction and coordinate positioning of candidate fields are completed in the main processing field. For semantic feature vectors in a multi-business domain, denoted as semantic particles, in each iteration, the direction cosine similarity between the semantic particle and the target feature vector is calculated. At the same time, it is subject to a bias correction effect pointing towards the target feature vector. The magnitude of the bias correction is negatively correlated with the current direction cosine similarity of the semantic particle: the lower the direction cosine similarity, the further the semantic particle deviates from the core task of the form, and the larger the corresponding bias correction magnitude, which pushes the semantic particle to shift rapidly towards the target feature vector; the higher the direction cosine similarity, the higher the semantic particle's fit with the core business objective, and the smaller the corresponding bias correction magnitude, which keeps the particle's current semantic position stable. The iteration rounds are set to 10 rounds. After each iteration, the spatial coordinates and semantic strength of all semantic particles are updated. After multiple iterations and convergence, the spatial distribution of candidate field seeds always converges toward the global goal of "completing the core task of the form". Redundant semantic particles that deviate from the core of the business gradually diffuse, decay and are eliminated, and finally all effective candidate fields are obtained, as well as the relative coordinates, semantic strength and other attribute parameters of each candidate field in multiple business fields. It should be noted that if there are candidate field seeds with high semantic strength in the auxiliary reference field that are not covered by the main processing field, they will be included as supplementary candidate fields in subsequent processing to avoid field omissions due to business type identification bias and improve the comprehensiveness of candidate fields.

[0021] Furthermore, the multi-intent analysis module transmits all candidate fields and their attribute parameters to the intent field planning module for field hierarchy sorting and intent field tree construction; Calculate the semantic concentration value corresponding to each candidate field. The semantic concentration value is used to characterize the degree of cohesion and business relevance of the candidate field in the business field. The calculation method is as follows: take the semantic intensity of the semantic particle corresponding to the candidate field and add it together with the cosine similarity between the target feature vector and the semantic feature vector corresponding to the candidate field. Two concentration thresholds are set: a first concentration threshold and a second concentration threshold. The thresholds can be dynamically adjusted according to the business type. For example, for financial and approval forms with high rigor requirements, the first concentration threshold is set to 0.75 and the second concentration threshold is set to 0.45; for general and registration forms, the first concentration threshold is set to 0.7 and the second concentration threshold is set to 0.4. Candidate fields with semantic concentration values ​​higher than the first concentration threshold are classified as core fields. Core fields are key information items that must be included to achieve the form's business objectives and carry the core data collection function of the form. Candidate fields with semantic concentration values ​​between the first and second concentration thresholds are classified as auxiliary fields. Auxiliary fields are used to supplement the contextual information of core fields and improve the completeness of form information and the filling experience. Candidate fields with semantic concentration values ​​lower than the second concentration threshold are considered redundant fields and are directly eliminated.

[0022] Furthermore, an intent field tree is constructed. First, an information value index is set to quantify the practical value of the field in actual business operations. The formula for calculating the information value index is as follows: ; in is the information value indicator for the field, with a value range of 0-1; F is the average fill frequency of the corresponding field in similar historical forms. C represents the highest field completion frequency among all fields in similar forms; C represents the contribution of the corresponding field to data integrity, that is, the percentage increase in the overall information completeness of the form after the field is completed. The field with the highest contribution among similar forms; w1 and w2 are the weight coefficients corresponding to the filling frequency, completeness contribution, and business importance, respectively, and satisfy w1+w2+w3=1, with default values ​​of 0.3, 0.3, and 0.4 respectively; The core fields obtained from the screening are used as first-level nodes, and clustered and grouped according to business type to form first-level grouping structures such as "basic information group", "business information group", "financial information group" and "supplementary explanation group"; auxiliary fields are used as second-level nodes, and are attached to the first-level nodes of the corresponding business type according to their semantic relationship. During the node mounting process, a region association detection is performed on each auxiliary field: a detection region with a radius of r is set with the semantic coordinates of the auxiliary field as the center, and the number of core fields in the region is counted. If the number is greater than or equal to 3, the auxiliary field is determined to be located in a strongly associated region, indicating that the corresponding field is closely associated with multiple first-level nodes. For auxiliary fields located in strongly related areas, move them to the edge of the first-level group structure and display them as key fields under the corresponding first-level group. At the same time, increase the visual weight of their interactive controls, such as increasing the width of the input box by 20%, adding light background highlighting prompts, and setting them as required fields by default, to guide users to fill them in first. Otherwise, determine that the corresponding auxiliary field is not located in a strongly related area and perform regional association detection on the next auxiliary field. After adjusting the node hierarchy and grouping, the intent field tree is structurally optimized by merging semantically redundant field nodes, adjusting the rationality of the grouping logic, and supplementing missing general basic fields. Finally, an intent field tree with a clear priority hierarchy and a clear grouping structure is output. Each field node in the intent field tree carries parameters such as semantic concentration value, information value index, group to which it belongs, and required attributes, providing a basis for subsequent constraint rule construction and form layout.

[0023] Furthermore, the constraint rule construction module receives the intent field tree, performs field association matching and constraint rule filtering, and constructs a field constraint network: Iterate through all field nodes in the intent field tree, randomly combine and match any two field nodes, and attempt to generate existing validation expressions, dependency conditions, or cascading filtering logic one by one. For example, between the "Country" field and the "City" field, generate a cascading dropdown relationship, that is, when the corresponding country is selected, the city option will only display the list of cities within that country; between "Start Date" and "End Date", generate a sequence validation rule, that is, the value of the end date must be later than the start date; between "Actual Order Payment Amount" and "Requested Refund Amount", generate a numerical upper limit validation rule, that is, the value of the requested refund amount cannot be greater than the actual order payment amount; between the "Do You Need an Invoice?" and "Invoice Title" fields, generate explicit and implicit dependency rules, that is, the "Invoice Title" field will only be displayed and set as a required field when "Do You Need an Invoice?" is selected as "Yes". Based on the matching results of all field pairs, all candidate association constraint rules are output. Each candidate association constraint rule contains information such as rule type, associated fields, and preliminary logical expression. Then, it is introduced into the compliance rule library, which stores standard validation rules, data security specifications and industry compliance requirements for forms in various business domains. Each standard rule is pre-coded into a standardized standard rule vector, and the same initial confidence score is set for each candidate association constraint rule. Each candidate association constraint rule is converted into a dimension-consistent requirement feature association vector through semantic encoding. The average directional angle between the corresponding requirement feature association vector and all standard rule vectors of the same type in the compliance rule library is calculated. The preset angle threshold is 20°, which can be adjusted within the range of 15°-25° according to the strictness of business compliance. For candidate association constraint rules whose average directional angle is less than or equal to the angle threshold, it indicates that they meet the business logic and compliance requirements. Positive gain processing is applied to them: the confidence score of the candidate association constraint rule is increased, and the gain is equal to the ratio between the angle threshold and the angle size. The smaller the angle, the greater the gain. For candidate association constraint rules whose average directional angle is greater than the angle threshold, it indicates that they deviate significantly from common sense and compliance requirements, and have logical inconsistencies or compliance risks. Therefore, they are subject to attenuation processing: during the iteration process, the confidence score of the candidate association constraint rule is gradually reduced, and the attenuation magnitude is equal to the ratio between the angle threshold and the angle size. A total of 3 attenuation iterations are set. If the confidence score is lower than 0.2 * the initial confidence score after 3 rounds of attenuation, the candidate association constraint rule is directly removed; otherwise, it is retained. After filtering all candidate association constraint rules, the effective constraint relationships between the field nodes corresponding to the remaining candidate association constraint rules are extracted from the intent field tree to form a field constraint network. Each constraint edge in the field constraint network is accompanied by attributes such as confidence score, constraint type, and rule parameters. The higher the confidence score, the stronger the rationality and compliance of the corresponding association constraint rule, and the higher the execution priority during the form execution process.

[0024] Furthermore, the constraint rule building module sends the field constraint network to the form auto-generation module, which then performs automated layout generation of the form interface. Initialize the form planning canvas, which is a two-dimensional visual layout space. The default canvas width is adapted to the desktop display, and mobile adaptation rules are preset. It supports one-click switching of multi-terminal display effects after generation. Based on the field type and data format of each field node in the field constraint network, the system matches corresponding form interaction controls: text fields match single-line input boxes, long text description fields match multi-line text fields, numeric fields match numeric input boxes with configurable unit selectors, enumerated option fields match dropdown selection boxes or radio button groups, date and time fields match date pickers, Boolean judgment fields match checkboxes or toggle controls, and attachment fields match file upload controls, etc. Each field control is automatically associated with the basic validation rules of the corresponding field, such as mobile phone number format validation, ID card number format validation, and number range validation, etc. Based on the priority hierarchy and relationships of field nodes, the field controls are mapped and arranged on the form planning canvas according to layout rules. The layout follows the core principles below: First, the principle of grouping and centralization: field controls under the same group are arranged in the same area of ​​the canvas, the group name is displayed in bold title form at the top of the area, and different groups are separated by a 24px spacing to ensure a clear form structure and reduce the user's information recognition cost. Second, the priority sorting principle; within the same group, field controls are arranged in descending order of information value indicators, with core fields located at the beginning of the group and auxiliary fields arranged at the end; key fields that have been moved to the edge of the first level are placed in prominent positions within the group, such as the first row of the group or occupying an entire row, to highlight their business importance. Third, the principle of proximity; field controls with constrained relationships are arranged adjacently on the canvas. For example, “Start Date” and “End Date” are displayed side by side, “Province”, “City” and “District” are arranged horizontally in sequence, and “Event Name” and “Location” are arranged vertically adjacently. This reduces the visual jump cost for users when filling in information and facilitates real-time interactive feedback for cascading and validation logic. Fourth, the grid adaptation principle; allocate corresponding grid widths according to the importance of field information and the length of input content: core fields and long text fields occupy the full width of 12 columns by default; short text, numeric, date, and option fields can occupy 6 or 4 columns to achieve multi-field parallel arrangement, improving form space utilization while ensuring a good filling experience; After the field controls are arranged, operation buttons such as submit and reset are automatically generated at the bottom of the form planning canvas. Based on the association constraint rules in the field constraint network, the corresponding field control's fill format prompt text is automatically marked, the overall validation logic when the form is submitted is configured, and finally a complete smart form that can be directly previewed, exported, and integrated into the business system is generated. At the same time, users can make secondary fine-tuning on the generated results to meet personalized customization needs.

[0025] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An AI-powered intelligent form automatic generation system based on a large model, characterized in that, It includes a multi-intent analysis module, an intent field planning module, a constraint rule construction module, and an automatic form generation module; The multi-intent analysis module is used to obtain form requirement description text, map the form requirement description text into semantic feature vector, establish multi-business fields according to preset business types, and input the semantic feature vector into the multi-business fields to obtain candidate fields. The intent field planning module is used to obtain the semantic concentration value of each candidate field, select core fields and auxiliary fields based on the semantic concentration value, and then build an intent field tree; The constraint rule construction module is used to match the related field nodes in the intent field tree, output the requirement feature association vector according to the matching result, obtain the angle value between the requirement feature association vector and the preset standard rule vector, and extract the field constraint network from the intent field tree according to the size of the angle value. The automatic form generation module is used to set up the form planning canvas, set up field controls according to the types of field nodes in the field constraint network, and map each field control to the form planning canvas based on the priority and association between field nodes.

2. The AI-powered intelligent form automatic generation system based on a large model according to claim 1, characterized in that, The process of mapping form requirement description text into semantic feature vectors and establishing multiple business domains based on preset business types includes: The form requirement description text is segmented into Chinese and named entity recognition is performed to remove meaningless stop words and modal particles. Business entities, attribute descriptions and logical relationship expressions in the text are identified to obtain a structured semantic unit sequence. Then, the semantic unit sequence is mapped to generate a semantic feature vector with dimension. According to the preset business type system, the target feature vectors of each business type in the corresponding domain knowledge base are retrieved, and each business type corresponds to a target feature vector. Based on the target feature vectors of each business type, business fields are constructed respectively. Each business field is centered on the corresponding target feature vector, and a standardized business field radius is set. The generated semantic feature vectors are sequentially input into all multi-business fields. The cosine similarity between the semantic feature vector and each target feature vector is calculated. The business field with the highest cosine similarity is selected as the main processing field, and the other business fields are used as auxiliary reference fields.

3. The AI-powered intelligent form automatic generation system based on a large model according to claim 2, characterized in that, The process of inputting semantic feature vectors into a multi-service domain to obtain candidate fields includes: For semantic feature vectors in the multi-business domain, denoted as semantic particles, in each iteration, the direction cosine similarity between the particle and the target feature vector is calculated, and it is also subject to the bias correction effect pointing towards the target feature vector. The iteration round is set to 10 rounds. After each iteration, the spatial coordinates and semantic intensity of all semantic particles are updated. After multiple iterations and convergence, the spatial distribution of candidate field seeds converges towards the global target. Redundant semantic particles that deviate from the core business gradually diffuse, decay and are eliminated, finally obtaining all effective candidate fields, as well as the relative coordinates of each candidate field in the multi-business domain.

4. The AI-powered intelligent form automatic generation system based on a large model according to claim 3, characterized in that, The process of obtaining the semantic concentration value of the candidate field includes: The semantic concentration value is used to characterize the degree of cohesion and business relevance of candidate fields in the business field. It is obtained by taking the semantic intensity of the semantic particle corresponding to the candidate field and adding the cosine similarity between the target feature vector and the semantic feature vector corresponding to the candidate field.

5. The AI-powered intelligent form automatic generation system based on a large model according to claim 4, characterized in that, The process of selecting core and auxiliary fields based on semantic concentration values, and then building an intent field tree, includes: Two levels of concentration thresholds are set, namely the first concentration threshold and the second concentration threshold, and the thresholds can be dynamically adjusted according to the business type; Candidate fields with semantic concentration values ​​higher than the first concentration threshold are classified as core fields, candidate fields with semantic concentration values ​​between the first and second concentration thresholds are classified as auxiliary fields, and candidate fields with semantic concentration values ​​lower than the second concentration threshold are considered redundant fields and are directly eliminated. Set information value indicators, use the core fields obtained from the screening as first-level nodes, and cluster them according to business type to form a first-level grouping structure; auxiliary fields are used as second-level nodes, and are attached to the first-level nodes of the corresponding business type according to their semantic relationship. After completing the adjustment of node hierarchy and grouping, the semantically duplicated field nodes are merged, and then the intent field tree is output.

6. The AI-powered intelligent form automatic generation system based on a large model according to claim 5, characterized in that, The process of matching related field nodes in the intent field tree and outputting a demand feature association vector based on the matching results includes: Iterate through all field nodes in the intent field tree, randomly combine and match any two field nodes, and try to generate existing validation expressions, dependency conditions or cascading filtering logic one by one. Based on the matching results of all field pairs, all candidate association constraint rules are output and imported into the compliance rule library. The compliance rule library stores standard validation rules, data security specifications and industry compliance requirements for forms in various business domains. Each standard rule is pre-coded into a standardized standard rule vector, and the same initial confidence score is set for each candidate association constraint rule.

7. The AI-powered intelligent form automatic generation system based on a large model according to claim 6, characterized in that, The process of obtaining the angle between the demand feature association vector and the preset standard rule vector, and then extracting the field constraint network from the intent field tree based on the angle value, includes: Each candidate association constraint rule is converted into a requirement feature association vector with consistent dimensions through semantic encoding. The average direction angle between the corresponding requirement feature association vector and all standard rule vectors of the same type in the compliance rule base is calculated. An angle threshold is preset. Based on the relationship between the average direction angle and the angle threshold, the initial confidence score of each candidate association constraint rule is increased or decreased. Candidate association constraint rules with confidence scores lower than 0.2 * initial confidence score are eliminated. After filtering all candidate association constraint rules, the effective constraint relationships between the field nodes corresponding to the remaining candidate association constraint rules are extracted from the intent field tree to construct the field constraint network.

8. The AI-powered intelligent form automatic generation system based on a large model according to claim 7, characterized in that, The process of setting field controls based on the types of field nodes in the field constraint network, and mapping each field control to the form planning canvas based on the priority and relationship between field nodes, includes: Initialize the form planning canvas, match the form interaction controls according to the field type and data format of each field node in the field constraint network, and map and arrange each field control in the form planning canvas according to the priority level and association relationship of the field nodes and layout rules. The layout follows the principles of grouping and centralization, priority sorting, association and proximity, and grid adaptation. After the field controls are arranged, operation buttons are generated at the bottom of the form planning canvas, and the corresponding field control's fill format prompt text is automatically labeled according to the association constraint rules in the field constraint network.

Images