Intelligent writing application method and device based on large model
The large-scale intelligent writing method solves the problem of time-consuming and labor-intensive traditional writing, realizes the efficient generation of standardized and uniform manuscripts, improves writing efficiency and professionalism, and forms a reusable knowledge base.
Patent Information
- Application Number
- CN202511174047.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-21
- Publication Date
- 2026-01-30
AI Technical Summary
Traditional writing methods are time-consuming and energy-intensive, rely on individual skill levels, and struggle to produce standardized and uniform texts that meet requirements. Existing tools also lack the ability to deeply understand the text.
The intelligent writing method based on a large model includes modules for input analysis and intent understanding, knowledge retrieval and material extraction, style and standard calibration, first draft output, continuation and expansion, and rewriting. It generates standard-compliant text through multi-round dialogue interaction and multi-task fine-tuning.
Improve manuscript writing efficiency, ensure writing standards and consistency, enhance professionalism, achieve an organic combination of intelligent assistance and human creation, and form a reusable knowledge base.
Smart Images
Figure CN121435922A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application relates to the technical field of artificial intelligence, and particularly provides a large model-based intelligent writing application method and device. BACKGROUND
[0002] Various official documents, meeting minutes, work reports, speech scripts and the like are important carriers for daily operation, transmission and decision implementation. The traditional writing mode mainly relies on manual writing, which not only consumes a large amount of time and energy, but also is easily affected by the personal level and experience of the writer, resulting in uneven writing quality and sometimes non-uniform and non-standard format and style. General writing auxiliary tools, such as simple template library and word processing software, can provide format reference or basic editing function to a certain extent, but they lack deep understanding ability and are difficult to generate text content meeting the requirements according to specific background, specific writing specifications and user intention.
[0003] With the rapid development of artificial intelligence technology, especially the progress made by large models in understanding and generating human language, people are thinking about how to apply this cutting-edge technology to the government writing scenario to realize the automation and intelligentization of the writing process, thereby reducing the workload of staff. SUMMARY
[0004] The application is aimed at the deficiencies of the prior art, and provides a large model-based intelligent writing application method with high practicability.
[0005] The further technical task of the application is to provide a large model-based intelligent writing application device with reasonable design and safety.
[0006] The technical scheme adopted by the application to solve the technical problem is:
[0007] The large model-based intelligent writing application method first performs input analysis and intention understanding, knowledge retrieval and material extraction by the writing module, calibrates the style and specifications after the production of the first draft, and then outputs the first draft;
[0008] The continuation module generates new content that is semantically coherent, stylistically consistent and meets the specifications on the basis of the writing module, and then enters the expansion module to perform in-depth and detailed elaboration and expansion on the text paragraph or sentence specified by the user;
[0009] Finally, the rewriting module adjusts the existing text content, changes the style, structure or expression method.
[0010] Further, in input analysis and intent understanding, first receive the subject keywords, writing framework requirements and additional instructions of user input, through the semantic understanding ability of the large model, multi-dimensional analysis is carried out on the input, the Embedding model of the large model is used to map the theme to the semantic space of the government affairs field,
[0011] The framework provided by the user is converted into an operable template, and the Markdown mark corresponding to the format of the manuscript is generated through prompt word engineering, and the fuzzy requirements are clarified through multi-round dialogue interaction.
[0012] Further, in knowledge retrieval and material extraction, the large model retrieves and extracts information from internal and external databases;
[0013] The internal database interfaces the historical file library in the system, uses dense retrieval to match similar documents with theme vectors;
[0014] The external database extracts relevant information segments from data sources, and the information segments are integrated to form the knowledge base required for generating the initial draft.
[0015] Further, in the initial draft generation, based on the parsed user intent and integrated knowledge information, the writing module drives the large model to start generating text, and for the content under each subheading, knowledge graph-based generation is adopted to automatically convert statistical data into standard expressions, and table contents are output in HTML format.
[0016] Further, in style and specification calibration, through multi-task fine-tuning of the large model to ensure that the text meets the specifications, the generation process is constrained by using a term library, and the model is forced to preferentially select standard expressions through prefix optimization;
[0017] Based on regular expression and large model joint judgment, load sensitive word detection module to replace expressions that may cause ambiguity;
[0018] Use the built-in constraints and post-processing capabilities of the large model to calibrate the style of the generated content.
[0019] Further, in the initial draft output, the generated initial draft is automatically typeset, the font, font size and alignment are adjusted, when multiple users edit at the same time, the modification suggestions are automatically merged and the controversial paragraphs are marked, and the history version and modification record are automatically saved, after the typesetting and editing of the initial draft are completed, the user downloads the initial draft to the local save.
[0020] Further, in the continuation module, it includes:
[0021] (1) Context understanding;
[0022] Once the module is started, the large model will first conduct an in-depth analysis of the input text, capturing the text's deeper meaning, core arguments, current emotional tone, narrative perspective, established writing style, and potential logical direction.
[0023] (2) Dynamic knowledge retrieval;
[0024] The process involves retrieving historical documents related to the topic of the text to be continued, calculating semantic similarity, extracting relevant paragraphs for reference, obtaining updated content from government portals via API, and leveraging the key information extraction capabilities of large models to extract relevant key information and embedding it into the text to be continued.
[0025] (3) Continuation strategy;
[0026] For argumentative texts, the large model generates subsequent content based on causal reasoning chains. If the original text is a list or outline, the module automatically identifies incomplete items and completes the content.
[0027] When the original text contains statistical indicators, the model uses numerical reasoning capabilities to generate analytical conclusions.
[0028] (4) Compliance verification;
[0029] During the continuation process, the large model simultaneously performs compliance checks.
[0030] Furthermore, the extension module includes:
[0031] (1) Input parsing;
[0032] Users specify the range of text to be expanded through the interactive interface. The large model first performs semantic segmentation on the input text and analyzes the logical relationship between the expanded part and the whole text.
[0033] (2) Generation of expansion strategies;
[0034] Based on a deep understanding of the target paragraph and the user's expansion instructions, the big model will dynamically generate specific expansion strategies and actively retrieve knowledge fragments in its training data that are highly related to the topic of the target paragraph.
[0035] (3) Multi-dimensional and multi-level content generation;
[0036] After clarifying the expansion strategy and preparing relevant knowledge materials, the large model begins to generate content from the following dimensions and levels.
[0037] Furthermore, the rewrite module includes:
[0038] (1) Instruction parsing;
[0039] First, the user inputs the text to be rewritten and the explicit rewriting instruction, and the large model identifies the specific aspect that the user wants to change: style, structure, or expression;
[0040] (2) Rewriting strategy execution;
[0041] If the goal is to change the style, the model will call its style transfer capability;
[0042] If the goal is to adjust the structure, the model will use its Seq2Seq generation capability to rearrange and combine sentences and paragraphs while keeping the content unchanged, and construct a new text that meets the new structure requirements;
[0043] If the goal is to optimize the expression, the model will handle it specifically; for simplification, the model will identify and delete redundant information, repetitive expressions, and minor details, and use more concise words and sentence patterns; for changing sentence patterns, the model will try to split long sentences into short sentences or combine multiple short sentences into complex sentences;
[0044] (3) Context fusion and style unification;
[0045] The rewritten text needs to be naturally integrated into the context of the original text, and the large model will continuously refer to the information before and after the original text when generating rewritten content.
[0046] The large model-based intelligent writing application device comprises at least one memory and at least one processor;
[0047] The at least one memory is used to store machine-readable programs;
[0048] The at least one processor is used to call the machine-readable programs and execute the large model-based intelligent writing application method.
[0049] Compared with the prior art, the large model-based intelligent writing application method and device of the present application has the following outstanding advantages:
[0050] (1) Improve the efficiency of document writing. Through the intelligent assistance of the large model, the time for traditional manual writing of documents is shortened, the workload of starting from scratch or repeatedly modifying is reduced, and the office efficiency is improved.
[0051] (2) Effectively ensure the writing specification and uniformity. The large model can generate documents that meet the standards, have uniform formats, and have standardized language by learning the background and writing specifications, which helps to avoid style differences or format errors caused by individual writing habits.
[0052] (3) Improve the professional level of documents. Through the learning ability of the large model on a large amount of high-quality documents, the large model can automatically generate professional documents with rigorous structure, accurate expression, and detailed data, and narrow the horizontal gap between regions.
[0053] (4) The system achieves an organic combination of intelligent assistance and human creation. It retains the function of human editing and forms a new writing mode of "AI first draft + human refinement", which takes into account both efficiency and quality.
[0054] (5) Repeated use of knowledge. During operation, the system continuously accumulates high-quality manuscript examples and reference data to form a reusable knowledge base, providing intelligent support for subsequent writing. Attached Figure Description
[0055] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0056] Figure 1 This is a flowchart illustrating the application method of intelligent writing based on large models. Detailed Implementation
[0057] To enable those skilled in the art to better understand the present invention, the present invention will be further described in detail below with reference to specific embodiments. Obviously, the described embodiments are merely 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] The following is a preferred embodiment:
[0059] The intelligent writing application method based on large models, firstly, the writing module:
[0060] (1) Input analysis and intent understanding;
[0061] First, the system receives user input including keywords, writing framework requirements, and additional instructions. Then, leveraging the semantic understanding capabilities of a large-scale model, it performs multi-dimensional analysis of the input.
[0062] Topic-level analysis: Using the embedding model of a large model, topics are mapped to the semantic space of the government affairs domain.
[0063] Framework structuring: Converts user-provided frameworks into operable templates. Generates Markdown tags that conform to the document's format through prompt engineering, including content formatting, writing focus of each paragraph, and key point hierarchy.
[0064] Refine instructions: Clarify vague requirements through multi-turn conversational interaction.
[0065] In this step, the large model analyzes the core writing theme and the themes and requirements of each part of the writing framework, accurately grasps the user's intent, and lays the foundation for subsequent text generation.
[0066] (2) Knowledge retrieval and material extraction;
[0067] The large model dynamically acquires information from the following data sources through knowledge retrieval and information extraction capabilities:
[0068] Internal database: Connects to the system's historical file library, using dense retrieval to match similar documents based on topic vectors.
[0069] External authoritative sources: Large models leverage their information retrieval capabilities to quickly and accurately extract relevant information fragments from massive data sources such as databases, document repositories, public reports, and news information. These information fragments are integrated to form the knowledge base needed to generate the initial draft.
[0070] User-owned private data: For user-uploaded local files, the model uses long text chunking to extract key points and avoids loss of context. For official document writing, users can upload similar document examples; for meeting minutes writing, users can upload meeting agendas, reports, speeches, etc.; for speech writing, users can upload documents for review, reference materials, etc.
[0071] The large model is not only responsible for retrieval in this process, but also for the preliminary extraction and integration of the retrieved information to ensure the effectiveness and relevance of the information.
[0072] (3) First draft generated;
[0073] Based on the parsed user intent and integrated knowledge information, the writing module drives the large model to begin generating text. The generation process follows the framework structure provided by the user.
[0074] Main content expansion: For the content under each subheading, a knowledge graph-based generation method is used.
[0075] Data embedding and formatting: Automatically converts statistical figures into standard representations, and outputs tabular content in HTML format for direct system rendering.
[0076] Throughout the generation process, the large model utilizes its information extraction capabilities to select the most relevant data, facts, and opinions from the integrated knowledge base. It then leverages its text generation capabilities to organize this information into fluent, coherent paragraphs that conform to linguistic conventions. The model automatically considers the logical connections between paragraphs, using appropriate transition words and phrases. Simultaneously, it ensures that the content remains focused on the theme and stays within the established framework.
[0077] (4) Style and standard calibration;
[0078] Multi-task fine-tuning of a large model ensures that the text conforms to government regulations.
[0079] Terminology accuracy: The generation process is constrained by a terminology library, and prefix tuning forces the model to prioritize the use of standardized expressions.
[0080] Sensitivity filtering: Based on the joint judgment of regular expressions and large models, a sensitive word detection module is loaded to replace expressions that may cause ambiguity.
[0081] Style consistency: Utilizing the built-in constraints and post-processing capabilities of the large model, style calibration is performed on the generated content. In particular, the writing style of the speech draft should match the speaker's style.
[0082] (5) Style and standard calibration;
[0083] After the initial draft is generated, the module supports the following collaborative operations:
[0084] One-click layout: The system can automatically layout the generated draft, adjusting the font, font size, alignment and other formats. Users can also edit the layout themselves.
[0085] Multi-user collaboration: When multiple users are editing simultaneously, the system automatically merges modification suggestions and marks disputed paragraphs, while automatically saving historical versions and modification records.
[0086] Local download: After completing the layout and editing of the first draft, users can download the draft to their local computer for saving.
[0087] The writing module is designed to connect with subsequent modules. If users are not satisfied with the first draft, they can use commands to guide the large model to regenerate or adjust specific parts, achieving initial iterative optimization.
[0088] Continue writing module:
[0089] The core task of the continuation writing module is to generate new content that is semantically coherent, stylistically consistent, and conforms to standards based on existing text. The specific steps are as follows:
[0090] (1) Contextual understanding;
[0091] Once the module starts, the large model will first conduct an in-depth analysis of the input text, capturing its deeper meaning, core arguments, current emotional tone, narrative perspective, established writing style, and potential logical direction. This results in a comprehensive understanding of the current state of the text.
[0092] (2) Dynamic knowledge retrieval;
[0093] The system retrieves historical documents related to the topic of the text to be continued, calculates semantic similarity, and extracts relevant paragraphs. Updated content from the portal website is obtained via API, and the key information extraction capabilities of the large model are used to extract relevant key information and embed it into the continued text.
[0094] (3) Continuation strategy;
[0095] For argumentative texts, the large model generates subsequent content based on causal reasoning chains. If the original text is a list or outline, the module automatically identifies incomplete items and completes the content. When the original text contains statistical indicators, the model utilizes numerical reasoning capabilities to generate analytical conclusions.
[0096] (4) Compliance verification;
[0097] During the continuation process, the large model simultaneously runs compliance checks to ensure that the statements conform to the established standards. Internal verification is also performed to avoid obvious logical contradictions or common-sense errors.
[0098] Extended module:
[0099] The expansion module provides in-depth and detailed elaboration and expansion of user-specified text paragraphs or sentences, making their content richer, their arguments more thorough, and their expression clearer. The specific steps are as follows:
[0100] (1) Input parsing;
[0101] Users specify the range of text to be expanded through an interactive interface. The large model first performs semantic segmentation on the input text, analyzes the logical relationship between the expanded part and the whole text, and adds content such as causal arguments and actionable steps according to the different logical relationships.
[0102] (2) Generation of expansion strategies;
[0103] Based on a deep understanding of the target paragraph and the user's expansion instructions, the large model dynamically generates specific expansion strategies. These strategies may include: supplementing background information, explaining key terms, listing specific examples, analyzing causes or effects, providing solutions, and elaborating from different perspectives. While generating strategies, the large model actively retrieves knowledge fragments from its training data that are highly relevant to the topic of the target paragraph.
[0104] (3) Multi-dimensional and multi-level content generation;
[0105] After clarifying the expansion strategy and preparing relevant knowledge materials, the large model begins to generate content from the following dimensions and levels: concept explanation and deepening, background information supplementation, case and data support, logical chain construction, multi-angle analysis, detail supplementation, and adding appropriate transition words or connecting sentences.
[0106] Rewrite module:
[0107] This involves adjusting existing text content, changing its style, structure, or expression to better suit a specific occasion, audience, or communication purpose. The specific steps are as follows:
[0108] (1) Instruction parsing;
[0109] First, the system receives the text to be rewritten from the user, along with explicit rewriting instructions. The large model then identifies the specific aspects the user wishes to change: writing style, structure, or expression.
[0110] (2) Rewrite the strategy execution;
[0111] The model contains pre-trained style vectors, enabling it to precisely control the stylistic features of the output text. If the goal is to change the writing style, the model will utilize its style transfer capabilities.
[0112] If the goal is to adjust the structure, the model will utilize its Seq2Seq generation capabilities to rearrange and combine sentences and paragraphs while keeping the content unchanged, constructing new text that meets the requirements of the new structure. This may involve splitting, reorganizing, and reconnecting the original text.
[0113] If the goal is to optimize expression, the model will take targeted actions. For simplification, the model will identify and remove redundant information, repetitive expressions, and minor details, using more concise words and sentence structures. For sentence structure modification, the model will attempt to break long sentences into shorter ones, or combine multiple short sentences into compound sentences, and adjust active or passive voice, etc. These operations are all based on a deep understanding of the original text's semantics, ensuring that the optimized expression is clearer, more accurate, or better meets specific requirements.
[0114] (3) Contextual integration and stylistic consistency;
[0115] The rewritten text needs to blend naturally into the context of the original text to avoid abrupt changes in style or logical breaks. When generating rewritten content, the large model continuously refers to the context of the original text to ensure that the tone, word choice, and referential relationships of the rewritten part are consistent with the context.
[0116] Based on the above method, the large-model-based intelligent writing application device in this embodiment includes: at least one memory and at least one processor;
[0117] The at least one memory is used to store a machine-readable program;
[0118] The at least one processor is used to call the machine-readable program to execute the intelligent writing application method based on a large model.
[0119] The above-described specific embodiments are merely specific examples of the present invention. The patent protection scope of the present invention includes, but is not limited to, the above-described specific embodiments. Any technical solution that conforms to the above-described specific embodiments of the present invention and any appropriate changes or substitutions made by those skilled in the art should fall within the patent protection scope of the present invention.
[0120] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A large model-based intelligent writing application method, characterized in that, First, the writing module performs input analysis and intent understanding, knowledge retrieval and material extraction. After the production of the first draft, style and specification calibration is performed, followed by first draft output; The continuation module generates new content that is semantically coherent, stylistically consistent, and compliant with specifications based on the writing module. Then it enters the expansion module, which elaborates and expands on the user-specified text passages or sentences in depth and detail. Finally, the rewriting module adjusts the existing text content, changing the style, structure, or expression method.
2. The large model-based intelligent writing application method according to claim 1, characterized in that, During input analysis and intent understanding, the system first receives the user's input topic keywords, writing framework requirements, and additional instructions. Through the semantic understanding capabilities of the large model, the input is analyzed in multiple dimensions. The Embedding model of the large model maps the topic to the semantic space of the government affairs field, The user-provided framework is converted into an operable template, and the Markdown mark corresponding to the format of the manuscript is generated through prompt word engineering. Through multiple rounds of dialogue-style interaction, ambiguous requirements are clarified.
3. The large model-based intelligent writing application method according to claim 2, characterized in that, In knowledge retrieval and material extraction, the large model uses its knowledge retrieval and information extraction capabilities to obtain information from internal and external databases. The internal database interfaces with the system's historical file library, using dense retrieval to match similar documents based on topic vectors. The external database extracts relevant information segments from data sources. These information segments are integrated to form the knowledge base required for generating the first draft.
4. The large model-based intelligent writing application method according to claim 3, characterized in that, In the first draft generation, based on the user's intent and integrated knowledge information, the writing module drives the large model to start generating text. For each subheading, the content is generated based on the knowledge graph, automatically converting statistical data into standard expressions, and outputting table content in HTML format.
5. The large model-based intelligent writing application method according to claim 4, characterized in that, In the style and specification calibration, the large model is fine-tuned to ensure that the text meets the specifications. The term library is used to constrain the generation process, and the prefix is optimized to force the model to prefer standard expressions. Based on regular expressions and the joint judgment of the large model, the sensitive word detection module is loaded to replace expressions that may cause ambiguity. The built-in constraints and post-processing capabilities of the large model are used to calibrate the style of the generated content.
6. The large model-based intelligent writing application method according to claim 5, characterized in that, In the first draft output, the generated first draft is automatically typeset, with adjustments to font, font size, and alignment. When multiple users edit simultaneously, the system automatically merges modification suggestions and marks controversial passages, while automatically saving historical versions and modification records. After completing the typesetting and editing of the first draft, the user downloads the first draft to the local computer for saving.
7. The large model-based intelligent writing application method according to claim 6, characterized in that, In the continuation module, it includes: (1) Context understanding; After the module starts, the large model will first analyze the input text in depth, capturing the deep meaning, core arguments, current emotional tone, narrative perspective, established writing style, and potential logical direction of the text. (2) Dynamic knowledge retrieval; Retrieval of historical documents related to the topic of the text to be continued, calculation of semantic similarity, extraction of reference passages, API access to government portal website updates, and extraction of relevant key information using the large model's key information extraction capabilities. (3) Continuation strategy; For expository texts, the large model generates subsequent content based on causal reasoning chains. If the original text is a list or an outline, the module automatically identifies incomplete items and completes the content. When the original text contains statistical indicators, the model calls on numerical reasoning capabilities to generate analysis conclusions. (4) Compliance verification; During the continuation process, the large model runs compliance detection in parallel.
8. The large model-based intelligent writing application method according to claim 7, characterized in that, In the expansion module, it includes: (1) Input analysis; The user specifies the text range that needs to be expanded through the interactive interface. The large model first performs semantic segmentation on the input text and analyzes the logical relationship between the expanded part and the full text. (2) Expansion strategy generation; Based on the deep understanding of the target paragraph and the user's expansion instructions, the large model dynamically generates specific expansion strategies. The large model actively retrieves knowledge fragments highly related to the theme of the target paragraph from its training data. (3) Multi-dimensional and multi-level content generation; After the expansion strategy is clear and the relevant knowledge materials are prepared, the large model begins content generation from the following dimensions and levels.
9. The large model-based intelligent writing application method according to claim 8, characterized in that, In the rewriting module, it includes: (1) Instruction analysis; First, receive the user's input text and clear rewriting instructions. The large model identifies the specific aspects the user wants to change: style, structure, or expression. (2) Rewriting strategy execution; If the goal is to change the style, the model will use its style transfer capabilities. If the goal is to adjust the structure, the model will use its Seq2Seq generation capabilities to rearrange and combine sentences and paragraphs while keeping the content unchanged, and construct a new text that meets the new structure requirements. If the goal is to optimize the expression, the model will handle it specifically; for simplification, the model will identify and delete redundant information, repetitive expressions, and minor details, using more concise words and sentence patterns; for changing sentence patterns, the model will try to split long sentences into short sentences or combine multiple short sentences into complex sentences. (3) Context fusion and style unification; The rewritten text needs to naturally integrate into the context of the original text. When generating rewritten content, the large model will continuously refer to the information before and after the original text.
10. A large model-based intelligent writing application device, characterized by, It includes: At least one memory and at least one processor; The at least one memory is used to store machine-readable programs; The at least one processor is used to call the machine-readable programs to execute the method of any one of claims 1-9.