An intelligent document generation method, system, device and medium based on multi-model fusion

Abstract

The application discloses an intelligent document generation method and system based on multi-model fusion, equipment and medium, mainly related to the technical field of document generation, to solve the problem that the existing scheme has a single material dependent on the system, and the specification of the generated content is lacking. Including: obtaining a generation mode; when the generation mode is step-by-step generation, from a plurality of AI models configured by a multi-model scheduling engine, an AI model corresponding to the writing task requirement is obtained; the writing task requirement, writing style, writing logic and writing key terms are filled into a preset prompt template to obtain a prompt; the prompt is input into the AI model to obtain an initial intelligent document; when the generation mode is quick generation, a corresponding government document generation model is determined according to the writing style; the writing task requirement, writing logic and writing key terms are filled into the government document generation model to output the initial intelligent document.

Description

Technical Field

[0001] This application relates to the field of intelligent document generation technology, and in particular to an intelligent document generation method, system, device and medium based on multi-model fusion. Background Technology

[0002] In the field of official document writing, there have long been many pressing problems that need to be addressed. From the perspective of traditional writing methods, while template-based writing offers certain conveniences, its limitations are extremely apparent when faced with complex and ever-changing business scenarios, making it difficult to accurately adapt to various specific needs. Moreover, manual drafting is inefficient, and differences in style among different writers make it difficult to ensure consistency in document style. This not only hinders the standardization of government work but also reduces work efficiency to some extent.

[0003] While existing intelligent writing systems have provided new avenues for government document writing to some extent with technological advancements, they still have many shortcomings. On the one hand, the materials relied upon by these systems are relatively limited, failing to fully encompass the rich and diverse information in government work, resulting in a significant reduction in the richness and professionalism of the generated content. On the other hand, the generated content lacks standardization, failing to meet the rigorous and accurate requirements of government documents. It cannot conduct in-depth analysis and precise processing of materials from various formats and sources, leading to a significant deviation between the generated documents and actual business needs, making them difficult to directly apply to practical government work. Summary of the Invention

[0004] This application provides a method, system, device, and medium for generating intelligent official documents based on multi-model fusion, in order to solve the problems of existing solutions having relatively limited material support and a lack of standardization in the generated content.

[0005] Firstly, this application provides an intelligent document generation method based on multi-model fusion, the method comprising:

[0006] Obtain a heterogeneous collection of government documents in different styles, and use the heterogeneous collection of government documents to train a government document generation model corresponding to each style;

[0007] The multi-model scheduling engine is used to obtain writing task requirements and writing styles; keywords in the writing task requirements are extracted, and the writing logic and key writing terms corresponding to the keywords are obtained from the preset knowledge graph corresponding to the writing style.

[0008] Obtain the generation method; when the generation method is step-by-step generation, obtain the AI ​​model corresponding to the writing task requirements from several AI models configured by the multi-model scheduling engine; fill the writing task requirements, writing style, writing logic and key writing terms into the preset prompt template to obtain the prompt; input the prompt into the AI ​​model to obtain the initial intelligent document;

[0009] When the generation method is rapid generation, the corresponding government document generation model is determined according to the writing style; the writing task requirements, writing logic and key writing terms are filled into the government document generation model, and the initial intelligent document is output.

[0010] Obtain the multi-dimensional verification rules for official documents. When the initial intelligent document meets the multi-dimensional verification rules, determine the initial intelligent document as the final intelligent document.

[0011] When the multidimensional verification rules are not met, the unmet multidimensional verification rules are filled into the preset correction prompt word template to obtain the correction prompt word; from several AI models configured by the multi-model scheduling engine, the correction AI model is selected; the correction prompt word and the initial intelligent document are input into the correction AI model to obtain the final intelligent document.

[0012] In one implementation of this application, a heterogeneous set of government documents in different styles is obtained, and a government document generation model corresponding to each style is trained using the heterogeneous set of government documents, specifically including:

[0013] Build a multi-source data acquisition system and obtain source materials by connecting to the material through API interfaces;

[0014] Obtain structured and unstructured heterogeneous materials corresponding to each style;

[0015] Preset core metadata is extracted from the structured and unstructured heterogeneous materials corresponding to the current style using regular expressions;

[0016] Obtain the sample parsing text corresponding to the preset core metadata;

[0017] By utilizing pre-defined core metadata and sample parsing text, a trained semantic parsing model can be obtained.

[0018] The general government language is pre-trained based on the RoBERTa-wwm model. The input layer receives sample parsed text to obtain the trained RoBERTa-wwm model.

[0019] The trained sample parsing text, combined with the trained RoBERTa-wwm model, is used as the government document generation model.

[0020] In one implementation of this application, before obtaining the AI ​​model corresponding to the writing task requirement from several AI models configured by the multi-model scheduling engine when the generation method is step-by-step generation, the method further includes:

[0021] Keyword extraction is performed to meet the writing task requirements, and a preset sensitive word detection database is invoked.

[0022] Compare the data with the keywords and the preset sensitive word detection database;

[0023] When the keyword exists in the preset sensitive word detection database, a pop-up input box will appear, prompting you to retrieve the writing task requirements again.

[0024] In one implementation of this application, the AI ​​model corresponding to the writing task requirement is obtained from several AI models configured by the multi-model scheduling engine, specifically including:

[0025] Obtain the requirement type corresponding to the writing task requirements;

[0026] Based on the pre-defined demand type and the degree of adaptation between the AI ​​model, the most suitable AI model is determined to be the AI ​​model corresponding to the writing task requirements.

[0027] Alternatively, an AI model corresponding to the writing task requirement can be obtained from several AI models through external triggering.

[0028] In one implementation of this application, before obtaining the AI ​​model corresponding to the writing task requirement from several AI models configured by the multi-model scheduling engine, the method further includes:

[0029] It can adapt to and connect to multiple AI models, build a unified model interface layer, and configure one-click switching of AI models.

[0030] In one implementation of this application, a prompt is input into an AI model to obtain an initial intelligent document, specifically including:

[0031] Input the prompts into the AI ​​model to obtain the initial document framework;

[0032] Input the preset expanded prompts and initial document framework into the AI ​​model to create the initial intelligent document.

[0033] Secondly, this application provides an intelligent document generation system based on multi-model fusion, the system comprising:

[0034] The structured material processing module is used to acquire a set of heterogeneous government affairs materials of different styles, and to train a government affairs document generation model corresponding to each style using the set of heterogeneous government affairs materials.

[0035] The multi-model scheduling module is used to obtain writing task requirements and writing styles through the multi-model scheduling engine; extract keywords from the writing task requirements; and obtain the writing logic and key writing terms corresponding to the keywords from the preset knowledge graph corresponding to the writing style.

[0036] The dual-path generation architecture module is used to obtain the generation method. When the generation method is step-by-step generation, it obtains the AI ​​model corresponding to the writing task requirements from several AI models configured by the multi-model scheduling engine; it fills the writing task requirements, writing style, writing logic, and key writing terms into a preset prompt template to obtain prompts; it inputs the prompts into the AI ​​model to obtain the initial intelligent document; when the generation method is rapid generation, it determines the corresponding government document generation model based on the writing style; it fills the writing task requirements, writing logic, and key writing terms into the government document generation model to output the initial intelligent document.

[0037] The multi-dimensional writing control module is used to obtain the multi-dimensional verification rules of official documents. When the initial intelligent document meets the multi-dimensional verification rules, it is determined as the final intelligent document. When it does not meet the multi-dimensional verification rules, the unmet multi-dimensional verification rules are filled into the preset correction prompt word template to obtain the correction prompt words. The correction AI model is selected from several AI models configured by the multi-model scheduling engine. The correction prompt words and the initial intelligent document are input into the correction AI model to obtain the final intelligent document.

[0038] In one implementation of this application, the structured material processing module includes a government document generation model generation unit.

[0039] Used to build a multi-source data acquisition system, which obtains source materials by connecting to the API interface;

[0040] Obtain structured and unstructured heterogeneous materials corresponding to each style;

[0041] Preset core metadata is extracted from the structured and unstructured heterogeneous materials corresponding to the current style using regular expressions;

[0042] Obtain the sample parsing text corresponding to the preset core metadata;

[0043] By utilizing pre-defined core metadata and sample parsing text, a trained semantic parsing model can be obtained.

[0044] The general government language is pre-trained based on the RoBERTa-wwm model. The input layer receives sample parsed text to obtain the trained RoBERTa-wwm model.

[0045] The trained sample parsing text, combined with the trained RoBERTa-wwm model, is used as the government document generation model.

[0046] Thirdly, this application provides an intelligent document generation device based on multi-model fusion, the device comprising:

[0047] processor;

[0048] And a memory storing executable code, which, when executed, causes the processor to execute an intelligent document generation method based on multi-model fusion as described in any of the preceding claims.

[0049] Fourthly, this application provides a non-volatile computer storage medium storing computer instructions thereon, which, when executed, implement an intelligent document generation method based on multi-model fusion as described in any of the preceding claims.

[0050] As can be seen from the above technical solutions, this application has the following advantages:

[0051] By acquiring heterogeneous sets of government affairs materials in different styles and using them to train corresponding government document generation models, the system can fully cover the rich and diverse information in government work. In practical applications, the system can accurately call upon the appropriate materials and models according to different writing task requirements and writing styles, significantly improving the richness and professionalism of the generated government documents. This better meets the actual needs of government work, provides government personnel with more comprehensive and accurate information references, and improves the efficiency and quality of government work.

[0052] After acquiring writing task requirements and writing style through a multi-model scheduling engine, keywords are extracted and writing logic and key terms are obtained from a pre-set knowledge graph. Then, combined with different generation methods, corresponding models are used to generate initial intelligent documents. Furthermore, the system sets up multi-dimensional verification rules for documents to rigorously verify the initial intelligent documents. When the rules are not met, corrections are made using pre-set correction prompt templates and a correction AI model to ensure that the final output intelligent document meets the rigorous and accurate requirements of government documents. This standardized processing flow allows the generated documents to more accurately meet actual business needs, effectively reducing deviations from actual work, and can be directly applied to actual government work, improving the practicality and reliability of government document generation. Attached Figure Description

[0053] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0054] Figure 1 This is a flowchart of an intelligent document generation method based on multi-model fusion provided in an embodiment of this application.

[0055] Figure 2 This is a schematic diagram of the internal structure of an intelligent document generation system based on multi-model fusion provided in an embodiment of this application.

[0056] Figure 3 This is a schematic diagram of the internal structure of an intelligent document generation device based on multi-model fusion provided in an embodiment of this application. Detailed Implementation

[0057] 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, and 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.

[0058] Those skilled in the art should understand that the embodiments described below are merely preferred embodiments of this disclosure and do not imply that this disclosure can only be implemented through these preferred embodiments. These preferred embodiments are merely used to explain the technical principles of this disclosure and are not intended to limit the scope of protection of this disclosure. Based on the preferred embodiments provided by this disclosure, all other embodiments obtained by those skilled in the art without creative effort should still fall within the scope of protection of this disclosure.

[0059] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0060] The technical solutions proposed in the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0061] The embodiment provides an intelligent document generation method based on multi-model fusion, such as Figure 1 As shown in the embodiments of this application, the method mainly includes the following steps:

[0062] Step 110: Obtain a set of heterogeneous government affairs materials of different styles, and use the set of heterogeneous government affairs materials to train and obtain a government affairs document generation model corresponding to each style.

[0063] In some embodiments, a heterogeneous set of government documents in different styles is obtained, and a government document generation model corresponding to each style is trained using the heterogeneous set of government documents, specifically including:

[0064] Build a multi-source data acquisition system and obtain source materials by connecting to the material through API interfaces;

[0065] Obtain structured and unstructured heterogeneous materials corresponding to each style;

[0066] Preset core metadata is extracted from the structured and unstructured heterogeneous materials corresponding to the current style using regular expressions;

[0067] Obtain the sample parsing text corresponding to the preset core metadata;

[0068] By utilizing pre-defined core metadata and sample parsing text, a trained semantic parsing model can be obtained.

[0069] The general government language is pre-trained based on the RoBERTa-wwm model. The input layer receives sample parsed text to obtain the trained RoBERTa-wwm model.

[0070] The trained sample parsing text, combined with the trained RoBERTa-wwm model, is used as the government document generation model.

[0071] Based on the above description, this step involves constructing a multi-source data acquisition system to accurately acquire heterogeneous government affairs materials of different styles. Regular expressions are then used to effectively extract core metadata, and a semantic parsing model is trained by combining sample parsing text. Simultaneously, a general government affairs language pre-training is performed based on the RoBERTa-wwm model, ultimately constructing a government document generation model corresponding to each style. This process ensures the comprehensiveness and accuracy of the model training data, enabling the generated government documents to better meet actual work needs. This directly improves the relevance and practicality of government document generation, helping staff quickly obtain document content that meets specific style requirements and improving government processing efficiency.

[0072] Step 120: Obtain writing task requirements and writing style through a multi-model scheduling engine; extract keywords from the writing task requirements, and obtain the writing logic and key writing terms corresponding to the keywords from the preset knowledge graph corresponding to the writing style.

[0073] It's important to note that this step, through a multi-model scheduling engine, accurately captures the writing task requirements and style, achieving initial positioning for personalized writing. Extracting keywords from the writing task requirements and retrieving writing logic and key terms from the corresponding pre-defined knowledge graph directly improves the accuracy and professionalism of the writing content. On one hand, the writing logic selected based on clear requirements and style ensures a reasonable and clear article structure, conforming to the expression norms of specific scenarios. On the other hand, the accurate extraction and application of key terms enhances the professional depth of the content, avoiding misunderstandings caused by misuse or omission of terminology. Overall, this solution effectively reduces blind exploration and repeated revisions during the writing process, making content production more efficient and better suited to the needs, laying a solid foundation for subsequent optimization and improvement.

[0074] Step 130: Obtain the generation method; When the generation method is step-by-step generation, obtain the AI ​​model corresponding to the writing task requirements from several AI models configured by the multi-model scheduling engine; fill the writing task requirements, writing style, writing logic, and key writing terms into the preset prompt template to obtain prompts; input the prompts into the AI ​​model to obtain the initial intelligent document; When the generation method is rapid generation, determine the corresponding government document generation model according to the writing style; fill the writing task requirements, writing logic, and key writing terms into the government document generation model to output the initial intelligent document.

[0075] When the generation method is step-by-step generation, before obtaining the AI ​​model corresponding to the writing task requirements from several AI models configured by the multi-model scheduling engine, the method further includes:

[0076] Keyword extraction is performed to meet the writing task requirements, and a preset sensitive word detection database is invoked.

[0077] Compare the data with the keywords and the preset sensitive word detection database;

[0078] When the keyword exists in the preset sensitive word detection database, a pop-up input box will appear, prompting you to retrieve the writing task requirements again.

[0079] Specifically, the AI ​​model corresponding to the writing task requirements is obtained from several AI models configured by the multi-model scheduling engine, including:

[0080] Obtain the requirement type corresponding to the writing task requirement; determine the most suitable AI model as the AI ​​model corresponding to the writing task requirement based on the degree of adaptation between the preset requirement type and the AI ​​model; or, obtain the AI ​​model corresponding to the writing task requirement from several AI models through external triggering.

[0081] Before obtaining the AI ​​model corresponding to the writing task requirements from several AI models configured by the multi-model scheduling engine, the method further includes:

[0082] It can adapt to and connect to multiple AI models, build a unified model interface layer, and configure one-click switching of AI models.

[0083] The process involves inputting prompts into the AI ​​model to obtain initial intelligent documents, specifically including:

[0084] Input the prompts into the AI ​​model to obtain the initial document framework; input the preset expanded prompts and the initial document framework into the AI ​​model to obtain the initial intelligent document.

[0085] Based on the above description, a dual-mode design of step-by-step generation and rapid generation, combined with sensitive word detection and dynamic model adaptation mechanisms, achieves standardization and controllability in the intelligent document generation process. In the step-by-step generation mode, the system first pre-filters the required keywords through a sensitive word detection database. When a sensitive word is detected, a pop-up window confirms the requirements, effectively mitigating content compliance risks. Subsequently, the system selects a model based on the suitability of the requirement type and the AI ​​model, or external trigger commands. Combined with the one-click switching function of the unified model interface layer, this ensures the flexibility and accuracy of model calls. During the model input stage, the system embeds requirements, style, logic, and terminology into preset prompt templates to generate structured instructions. Through a secondary input mechanism of initial framework and expanded prompts, the generated initial intelligent document conforms to the format specifications of government documents and contains complete content elements. The rapid generation mode directly matches the writing style to a dedicated government model, simplifying the input process while ensuring the complete transmission of key information (writing logic and terminology). Both modes rely on a unified interface layer to manage multiple model resources, avoiding incompatibility issues with input parameters caused by model switching. This ensures that the generated results maintain professionalism while reducing the need for manual intervention, thus improving the output efficiency and quality stability of the initial draft of official documents.

[0086] Step 140: Obtain the multi-dimensional verification rules of the document. When the initial intelligent document meets the multi-dimensional verification rules, determine the initial intelligent document as the final intelligent document. When the multi-dimensional verification rules are not met, fill the unmet multi-dimensional verification rules into the preset correction prompt word template to obtain the correction prompt word. Select the correction AI model from several AI models configured by the multi-model scheduling engine. Input the correction prompt word and the initial intelligent document into the correction AI model to obtain the final intelligent document.

[0087] It should be noted that this step achieves controllable quality in intelligent document generation through multi-dimensional verification rules and dynamic correction mechanisms. The system first automatically verifies the initial document based on preset rules (such as format specifications, terminology accuracy, and logical consistency). If deviations are found, it automatically embeds specific problem items into correction prompt templates to generate structured instructions and calls a dedicated correction AI model for targeted correction. This closed-loop processing mechanism ensures that the final document meets standards in terms of standardization, professionalism, and completeness, avoiding redundant manual modifications. By separating the initial generation and correction optimization stages, the system retains the efficiency advantages of the first round of generation while overcoming the limitations of a single model output through rule-driven iterative optimization. The selection logic of the correction model and the interface compatibility of the multi-model scheduling engine allow for the adaptation of dedicated models to verification needs in different scenarios (such as legal provision verification and data consistency verification), improving the applicability to complex government scenarios. Overall, this design combines pre-defined rules with dynamic correction, enabling intelligent documents to complete quality self-checks and corrections before output, reducing subsequent manual review costs, while ensuring that the output meets the standardization requirements of government documents.

[0088] As described above, this embodiment acquires a collection of heterogeneous government affairs materials of different styles and uses them to train corresponding government document generation models, which can fully cover the rich and diverse information in government work. In practical applications, the system can accurately call upon the corresponding materials and models according to different writing task requirements and writing styles, significantly improving the richness and professionalism of the generated government documents, better meeting the actual needs of government work, providing government personnel with more comprehensive and accurate information references, and improving the efficiency and quality of government work. After acquiring the writing task requirements and writing styles through a multi-model scheduling engine, keywords are extracted and writing logic and key terms are obtained from a preset knowledge graph. Then, combined with different generation methods, the corresponding models are used to generate initial intelligent documents. Furthermore, the system sets multi-dimensional verification rules for documents to strictly verify the initial intelligent documents. When the rules are not met, corrections are made through preset correction prompt word templates and correction AI models to ensure that the final output intelligent documents meet the requirements of rigorous and accurate government documents. This standardized processing procedure enables the generated documents to more accurately meet actual business needs, effectively reducing deviations from actual work, and can be directly applied to actual government affairs, thus improving the practicality and reliability of government document generation.

[0089] In addition, this application Figure 2 This application provides an intelligent document generation system based on multi-model fusion. For example... Figure 2 As shown in the embodiments of this application, the system mainly includes:

[0090] The structured material processing module 210 is used to obtain a set of heterogeneous government affairs materials of different styles, and to train a government affairs document generation model corresponding to each style using the set of heterogeneous government affairs materials.

[0091] The structured material processing module 210 includes a government document generation model generation unit.

[0092] Used to build a multi-source data acquisition system, which obtains source materials by connecting to the API interface;

[0093] Obtain structured and unstructured heterogeneous materials corresponding to each style;

[0094] Preset core metadata is extracted from the structured and unstructured heterogeneous materials corresponding to the current style using regular expressions;

[0095] Obtain the sample parsing text corresponding to the preset core metadata;

[0096] By utilizing pre-defined core metadata and sample parsing text, a trained semantic parsing model can be obtained.

[0097] The general government language is pre-trained based on the RoBERTa-wwm model. The input layer receives sample parsed text to obtain the trained RoBERTa-wwm model.

[0098] The trained sample parsing text, combined with the trained RoBERTa-wwm model, is used as the government document generation model.

[0099] It should be noted that an automated training system for generating multi-style government documents was constructed through the structured material processing module 210, achieving efficient utilization of heterogeneous government materials and improving model adaptability. The module first connects to the API interface through a multi-source data acquisition system, integrating government materials from different sources and styles (such as regulatory documents, meeting minutes, notices, and announcements), solving the problem of insufficient coverage from a single channel. In the material processing stage, regular expressions are used to extract core metadata (such as document titles, document numbers, body paragraph structure, policy keywords, etc.) from structured and unstructured materials, ensuring standardized extraction of key information and avoiding subjective bias from manual annotation. In the sample parsing text generation stage, structured training samples are constructed through the association mapping between metadata and the original text, providing the model with clearer semantic feature input. The training of the semantic parsing model further enhances the recognition ability of text elements, while the pre-training of general government language based on the RoBERTa-wwm model improves the model's adaptability to government language scenarios through semantic understanding optimization of large-scale government texts. The final government document generation model, which combines sample parsed text with a pre-trained model, retains the textual features of specific styles (such as the rigor of legal documents and the conciseness of notices) while ensuring generalization across styles through the fundamental capabilities of a general language model. This design eliminates the need for extensive manual annotation during model training and reduces the interference of heterogeneous materials on model performance through style classification. The final output model can be directly used for the automatic generation of government documents of the corresponding style, improving the model's practicality and accuracy in government scenarios.

[0100] The multi-model scheduling module 220 is used to obtain writing task requirements and writing style through the multi-model scheduling engine; extract keywords from the writing task requirements; and obtain the writing logic and key writing terms corresponding to the keywords from the preset knowledge graph corresponding to the writing style.

[0101] It should be noted that the multi-model scheduling module 220 achieves precise matching and knowledge reuse between writing task requirements and writing styles, improving the efficiency of pre-processing for official document generation. The module first obtains the writing task requirements (such as document type, theme, and length) and writing style (such as rigorous style for regulations and concise style for notices) through the multi-model scheduling engine, clarifying the generation objective. Subsequently, the system extracts keywords from the task requirements (such as "budget approval" and "epidemic prevention and control") and searches based on the pre-defined knowledge graph corresponding to the writing style, quickly locating the writing logic associated with the keywords (such as the "basis-measures-implementation" structure of regulations) and key terms (such as proper nouns and standardized expressions in policy documents). This process avoids the inefficiency of manually sorting out the relationship between requirements and knowledge, and the structured storage of the knowledge graph ensures the consistent reuse of writing logic and terminology across different tasks. The final output logic and terminology can be directly used in subsequent model generation stages, reducing generation errors caused by misunderstandings of requirements or lack of knowledge, and improving the professionalism and compliance of the initial draft of the official document.

[0102] The dual-path generation architecture module 230 is used to obtain the generation method. When the generation method is step-by-step generation, it obtains the AI ​​model corresponding to the writing task requirements from several AI models configured by the multi-model scheduling engine; it fills the writing task requirements, writing style, writing logic, and key writing terms into a preset prompt template to obtain prompts; it inputs the prompts into the AI ​​model to obtain the initial intelligent document; when the generation method is rapid generation, it determines the corresponding government document generation model according to the writing style; it fills the writing task requirements, writing logic, and key writing terms into the government document generation model to output the initial intelligent document.

[0103] The multi-dimensional writing control module 240 is used to obtain the multi-dimensional verification rules of official documents. When the initial intelligent document meets the multi-dimensional verification rules, it is determined that the initial intelligent document is the final intelligent document. When it does not meet the multi-dimensional verification rules, the unmet multi-dimensional verification rules are filled into the preset correction prompt word template to obtain the correction prompt word. The correction AI model is selected from several AI models configured by the multi-model scheduling engine. The correction prompt word and the initial intelligent document are input into the correction AI model to obtain the final intelligent document.

[0104] As described above, this embodiment acquires a collection of heterogeneous government affairs materials of different styles and uses them to train corresponding government document generation models, which can fully cover the rich and diverse information in government work. In practical applications, the system can accurately call upon the corresponding materials and models according to different writing task requirements and writing styles, significantly improving the richness and professionalism of the generated government documents, better meeting the actual needs of government work, providing government personnel with more comprehensive and accurate information references, and improving the efficiency and quality of government work. After acquiring the writing task requirements and writing styles through a multi-model scheduling engine, keywords are extracted and writing logic and key terms are obtained from a preset knowledge graph. Then, combined with different generation methods, the corresponding models are used to generate initial intelligent documents. Furthermore, the system sets multi-dimensional verification rules for documents to strictly verify the initial intelligent documents. When the rules are not met, corrections are made through preset correction prompt word templates and correction AI models to ensure that the final output intelligent documents meet the requirements of rigorous and accurate government documents. This standardized processing procedure enables the generated documents to more accurately meet actual business needs, effectively reducing deviations from actual work, and can be directly applied to actual government affairs, thus improving the practicality and reliability of government document generation.

[0105] The above are method embodiments of this application. Based on the same inventive concept, this application also provides an intelligent document generation device based on multi-model fusion. Figure 3 As shown, the device includes: a processor; and a memory storing executable code thereon, which, when executed, causes the processor to perform an intelligent document generation method based on multi-model fusion as described in the above embodiments.

[0106] Specifically, the server acquires a collection of heterogeneous government documents in different styles, and uses this collection to train a document generation model corresponding to each style. Through a multi-model scheduling engine, it obtains the writing task requirements and writing style; extracts keywords from the writing task requirements, and retrieves the writing logic and key terminology corresponding to the keywords from a pre-defined knowledge graph corresponding to the writing style; it obtains the generation method; when the generation method is step-by-step generation, it selects the AI ​​model corresponding to the writing task requirements from several AI models configured in the multi-model scheduling engine; it fills the writing task requirements, writing style, writing logic, and key terminology into a pre-defined prompt template to obtain prompts; and it inputs the prompts into the AI... The system generates an initial intelligent document. When the generation method is rapid generation, the corresponding government document generation model is determined based on the writing style. The writing task requirements, writing logic, and key writing terms are filled into the government document generation model, and the initial intelligent document is output. The system obtains the multi-dimensional verification rules for the document. When the initial intelligent document meets the multi-dimensional verification rules, it is determined as the final intelligent document. When the multi-dimensional verification rules are not met, the unmet multi-dimensional verification rules are filled into the preset correction prompt word template to obtain correction prompt words. The system selects a correction AI model from several AI models configured by the multi-model scheduling engine. The correction prompt words and the initial intelligent document are input into the correction AI model to obtain the final intelligent document.

[0107] As described above, this embodiment acquires a collection of heterogeneous government affairs materials of different styles and uses them to train corresponding government document generation models, which can fully cover the rich and diverse information in government work. In practical applications, the system can accurately call upon the corresponding materials and models according to different writing task requirements and writing styles, significantly improving the richness and professionalism of the generated government documents, better meeting the actual needs of government work, providing government personnel with more comprehensive and accurate information references, and improving the efficiency and quality of government work. After acquiring the writing task requirements and writing styles through a multi-model scheduling engine, keywords are extracted and writing logic and key terms are obtained from a preset knowledge graph. Then, combined with different generation methods, the corresponding models are used to generate initial intelligent documents. Furthermore, the system sets multi-dimensional verification rules for documents to strictly verify the initial intelligent documents. When the rules are not met, corrections are made through preset correction prompt word templates and correction AI models to ensure that the final output intelligent documents meet the requirements of rigorous and accurate government documents. This standardized processing procedure enables the generated documents to more accurately meet actual business needs, effectively reducing deviations from actual work, and can be directly applied to actual government affairs, thus improving the practicality and reliability of government document generation.

[0108] In addition, this application embodiment also provides a non-volatile computer storage medium storing executable instructions, which, when executed, implement the intelligent document generation method based on multi-model fusion as described above.

[0109] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for generating intelligent official documents based on multi-model fusion, characterized in that, The method includes: Obtain a heterogeneous collection of government documents in different styles, and use the heterogeneous collection of government documents to train a government document generation model corresponding to each style; The multi-model scheduling engine is used to obtain writing task requirements and writing styles; keywords in the writing task requirements are extracted, and the writing logic and key writing terms corresponding to the keywords are obtained from the preset knowledge graph corresponding to the writing style. The generation method is determined as follows: When the generation method is step-by-step generation, the AI ​​model corresponding to the writing task requirements is obtained from several AI models configured by the multi-model scheduling engine; the writing task requirements, writing style, writing logic, and key writing terms are filled into the preset prompt template to obtain prompts; the prompts are input into the AI ​​model to obtain the initial intelligent document; when the generation method is rapid generation, the corresponding government document generation model is determined according to the writing style; the writing task requirements, writing logic, and key writing terms are filled into the government document generation model to output the initial intelligent document. The system retrieves the multi-dimensional verification rules for official documents. When the initial intelligent document meets the multi-dimensional verification rules, it is determined to be the final intelligent document. When the multi-dimensional verification rules are not met, the unmet multi-dimensional verification rules are filled into the preset correction prompt word template to obtain the correction prompt word. The system selects the correction AI model from several AI models configured by the multi-model scheduling engine. The system inputs the correction prompt word and the initial intelligent document into the correction AI model to obtain the final intelligent document.

2. The intelligent document generation method based on multi-model fusion according to claim 1, characterized in that, Obtain a heterogeneous collection of government documents in different styles, and use this collection to train a government document generation model corresponding to each style. Specifically, this includes: Build a multi-source data acquisition system and obtain source materials by connecting to the material through API interfaces; Obtain structured and unstructured heterogeneous materials corresponding to each style; Preset core metadata is extracted from the structured and unstructured heterogeneous materials corresponding to the current style using regular expressions; Obtain the sample parsing text corresponding to the preset core metadata; By utilizing pre-defined core metadata and sample parsing text, a trained semantic parsing model can be obtained. The general government language is pre-trained based on the RoBERTa-wwm model. The input layer receives sample parsed text to obtain the trained RoBERTa-wwm model. The trained sample parsing text, combined with the trained RoBERTa-wwm model, is used as the government document generation model.

3. The intelligent document generation method based on multi-model fusion according to claim 1, characterized in that, When the generation method is step-by-step generation, before obtaining the AI ​​model corresponding to the writing task requirements from several AI models configured by the multi-model scheduling engine, the method further includes: Keyword extraction is performed to meet the writing task requirements, and a preset sensitive word detection database is invoked. Compare the data with the keywords and the preset sensitive word detection database; When the keyword exists in the preset sensitive word detection database, a pop-up input box will appear, prompting you to retrieve the writing task requirements again.

4. The intelligent document generation method based on multi-model fusion according to claim 1, characterized in that, From several AI models configured by the multi-model scheduling engine, obtain the AI ​​model corresponding to the writing task requirements, specifically including: Obtain the requirement type corresponding to the writing task requirements; Based on the pre-defined demand type and the degree of adaptation between the AI ​​model, the most suitable AI model is determined to be the AI ​​model corresponding to the writing task requirements. Alternatively, an AI model corresponding to the writing task requirement can be obtained from several AI models through external triggering.

5. The intelligent document generation method based on multi-model fusion according to claim 1, characterized in that, Before obtaining the AI ​​model corresponding to the writing task requirements from several AI models configured by the multi-model scheduling engine, the method further includes: It can adapt to and connect to multiple AI models, build a unified model interface layer, and configure one-click switching of AI models.

6. The intelligent document generation method based on multi-model fusion according to claim 1, characterized in that, Input the prompt into the AI ​​model to obtain the initial intelligent document, which includes: Input the prompts into the AI ​​model to obtain the initial document framework; Input the preset expanded prompts and initial document framework into the AI ​​model to create the initial intelligent document.

7. An intelligent document generation system based on multi-model fusion, characterized in that, The system includes: The structured material processing module is used to acquire a set of heterogeneous government affairs materials of different styles, and to train a government affairs document generation model corresponding to each style using the set of heterogeneous government affairs materials. The multi-model scheduling module is used to obtain writing task requirements and writing styles through the multi-model scheduling engine; extract keywords from the writing task requirements; and obtain the writing logic and key writing terms corresponding to the keywords from the preset knowledge graph corresponding to the writing style. The dual-path generation architecture module is used to obtain the generation method. When the generation method is step-by-step generation, it obtains the AI ​​model corresponding to the writing task requirements from several AI models configured by the multi-model scheduling engine; it fills the writing task requirements, writing style, writing logic, and key writing terms into a preset prompt template to obtain prompts; it inputs the prompts into the AI ​​model to obtain the initial intelligent document; when the generation method is rapid generation, it determines the corresponding government document generation model based on the writing style; it fills the writing task requirements, writing logic, and key writing terms into the government document generation model to output the initial intelligent document. The multi-dimensional writing control module is used to obtain the multi-dimensional verification rules of official documents. When the initial intelligent document meets the multi-dimensional verification rules, it is determined as the final intelligent document. When it does not meet the multi-dimensional verification rules, the unmet multi-dimensional verification rules are filled into the preset correction prompt word template to obtain the correction prompt words. The correction AI model is selected from several AI models configured by the multi-model scheduling engine. The correction prompt words and the initial intelligent document are input into the correction AI model to obtain the final intelligent document.

8. The intelligent document generation system based on multi-model fusion according to claim 7, characterized in that, The structured material processing module includes a government document generation model generation unit. Used to build a multi-source data acquisition system, which obtains source materials by connecting to the API interface; Obtain structured and unstructured heterogeneous materials corresponding to each style; Preset core metadata is extracted from the structured and unstructured heterogeneous materials corresponding to the current style using regular expressions; Obtain the sample parsing text corresponding to the preset core metadata; By utilizing pre-defined core metadata and sample parsing text, a trained semantic parsing model can be obtained. The general government language is pre-trained based on the RoBERTa-wwm model. The input layer receives sample parsed text to obtain the trained RoBERTa-wwm model. The trained sample parsing text, combined with the trained RoBERTa-wwm model, is used as the government document generation model.

9. An intelligent document generation device based on multi-model fusion, characterized in that, The device includes: processor; And a memory having executable code stored thereon, which, when executed, causes the processor to perform an intelligent document generation method based on multi-model fusion as described in any one of claims 1-6.

10. A non-volatile computer storage medium, characterized in that, It stores computer instructions, which, when executed, implement the intelligent document generation method based on multi-model fusion as described in any one of claims 1-6.

Images