Adaptive document generation method and system based on knowledge graph and large model

Abstract

The application relates to the field of artificial intelligence and natural language processing technology, and proposes a self-adaptive document generation method and system based on a knowledge graph and a large model.The application is based on a document generation request input by a user;extracts entities from multiple source data and constructs and updates a queryable knowledge graph;combines a user request, a current generation context and feedback of the knowledge graph to dynamically plan the structure, detail level and expression style of a document;based on a planning result, searches for related knowledge segments from the knowledge graph as conditions to drive a pre-trained large language model to generate coherent text content;structurally organizes the generated text content according to semantic roles, and renders the text content into a readable document format based on a dynamic template to output.The application realizes the organic combination of knowledge driving and language generation, can flexibly cope with document generation tasks in different scenarios through an adaptive mechanism, and improves the efficiency and effect of information transmission.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and natural language processing technology, and more specifically, to an adaptive document generation method and system based on knowledge graphs and large models. Background Technology

[0002] With the development of information technology, the demand for automated document generation technology is increasing in scenarios such as office automation, content creation, report writing, and knowledge management. Traditional document generation methods mainly rely on two technical paths: one is a system based on fixed templates and rules, which generates documents by filling in predefined structures and fields. While this method can ensure standardized formatting, the content is rigid and difficult to adapt to flexible and changing generation needs. It also heavily relies on manual template design and knowledge entry, resulting in insufficient scalability and adaptability. The other approach is the generation method based on large-scale pre-trained language models, which has emerged in recent years. This type of method can generate fluent and coherent natural language text, greatly improving the flexibility and diversity of content. However, the content generated by large models suffers from the illusion problem, which may produce factual errors or statements that are inconsistent with specific domain knowledge. Furthermore, the generation process lacks controllable structure, making it difficult to guarantee the logical hierarchy, information integrity, and stylistic consistency of the generated documents.

[0003] To overcome these problems, existing technologies attempt to combine external knowledge bases with large models. For example, retrieval-enhanced generation techniques incorporate relevant text fragments as references during generation. However, these methods typically rely on unstructured document retrieval, and the introduced knowledge lacks precise semantic associations and structured representations, making it difficult to support deep knowledge reasoning and consistency verification. Knowledge graphs, as a technology for efficiently organizing and managing structured knowledge, can clearly express entities, attributes, and the relationships between them, providing a precise and reasonable knowledge foundation for content generation. However, current solutions for deeply integrating knowledge graphs into the large model generation process still face challenges: the construction of knowledge graphs usually relies on complex pipeline models, requiring large amounts of labeled data and incurring high domain migration costs; knowledge injection methods are often rigid, making it difficult to achieve end-to-end adaptive control of the generated content in terms of macro-structure, level of detail, and language style; and the final document formatting is often separated from the content generation process, failing to achieve intelligent typesetting based on content semantics.

[0004] Therefore, existing technologies lack a document generation solution that can deeply integrate structured knowledge graphs with powerful generation capabilities, and achieve comprehensive self-adaptation across multiple dimensions such as factual accuracy, logical structure, style adaptability, and presentation professionalism based on user intent and context. Summary of the Invention

[0005] In view of this, the present invention proposes an adaptive document generation method and system based on knowledge graphs and large models to solve the problems existing in the prior art.

[0006] To achieve the above objectives, this invention proposes an adaptive document generation method based on knowledge graphs and large models, including: Receive document generation requests from the user; Based on the document generation request, entities, relationships, and attributes are extracted from multi-source data to construct and update a queryable knowledge graph; Based on the document generation request, the current generation context, and the feedback from the knowledge graph, the document's structure, level of detail, and expression style are dynamically planned. Based on the planning results, relevant knowledge fragments are retrieved from the knowledge graph as conditions to drive the pre-trained large language model to generate coherent text content. The generated text content is structured according to semantic roles and rendered into a readable document format based on dynamic templates for output.

[0007] Furthermore, the construction and updating of the queryable knowledge graph specifically includes: An end-to-end joint extraction framework based on a large language model is adopted. By assembling structured prompts containing instructions, examples and text to be processed, the large language model can directly output formatted knowledge triples. The output triples are linked and normalized. By calculating the similarity of the embedding vectors and combining the context, the new entities are matched and fused with the existing entities in the knowledge base. The processed triples are imported into a graph database to build a knowledge network and a query interface that supports natural language question retrieval.

[0008] Furthermore, the process of driving the pre-trained large language model to generate coherent text content specifically includes: A knowledge retrieval-based conditional generation framework is adopted to convert the structured knowledge fragments retrieved from the knowledge graph into a prompt context containing semantic tags and generation constraints; The contextual prompts are input into the large language model, and an attention-guided mechanism is used during the generation process to ensure the consistency between the text description and the knowledge fragments. After paragraph generation, a consistency check is performed. If a deviation from the facts is found, a prompt for revision and regeneration is triggered.

[0009] Furthermore, the structure, level of detail, and style of expression of the dynamic programming document specifically include: A hierarchical reinforcement learning framework based on goal-driven and multi-granular planners is adopted to model document generation as a hierarchical decision-making process; Using a hierarchical strategy network, dynamic document outlines and control instructions are output in a rolling manner based on the encoded global, local and micro states; The policy network's decision-making is optimized online using reinforcement learning based on a multi-objective reward function that considers accuracy, coherence, adaptability, and efficiency.

[0010] Furthermore, rendering the generated text content into a readable document format specifically includes: A semantically aware structured template engine is used to automatically label text units with semantic roles; Based on the semantic roles and user style preferences, style mapping rules are dynamically selected and instantiated from the template library to generate exclusive temporary templates; Based on the temporary template, text, charts, and special content are integrated and rendered, and exported into various common document formats.

[0011] On the other hand, to achieve the above objectives, this invention proposes an adaptive document generation system based on knowledge graphs and large models, including a knowledge graph construction module, a large model integration module, an adaptive control module, and a document generation and rendering module. The knowledge graph construction module is used to extract entities, relationships and attributes from multi-source data and construct a queryable knowledge network. The large model integration module is used to access a pre-trained large language model to understand user intent and generate coherent text content. The adaptive control module is used to dynamically adjust the document structure, level of detail, and expression style based on user input, context, and knowledge graph feedback. The document generation and rendering module is used to organize the generated content according to a predetermined format and output it as a readable document.

[0012] Furthermore, the knowledge graph construction module adopts an end-to-end joint extraction framework based on a large language model, which transforms the multi-step extraction task into a structured prompting engineering problem; it converts the raw data into a plain text sequence and assembles structured prompts containing instructions, examples and the current text. Input the text sequence along with prompts into a large language model, enabling the model to simultaneously identify entity pairs and their relationships, and output descriptive features as attributes; constrain the model output to a triple format.

[0013] Furthermore, the knowledge graph construction module calculates the similarity between the embedding vectors of new entities and existing entities and performs entity linking and normalization by combining contextual features; the fused triples are imported into the graph database, and a unified query service interface supporting natural language question retrieval is encapsulated.

[0014] Furthermore, the large model integration module receives a document generation request containing metadata about the topic, audience, and style, and deconstructs the document generation request into an executable query intent; Automatically generate query statements for knowledge graphs and extract relevant entities, attributes, and relationship paths; Convert search results into a contextual prompt that blends knowledge and natural language, and set the generation target, knowledge point list, and expressions to be avoided; An attention-guided mechanism ensures that the generated content is consistent with the attribute values ​​of the knowledge graph, and a consistency check loop is used to detect fact deviations and regenerate the content.

[0015] Furthermore, the adaptive control module adopts a hierarchical reinforcement learning framework based on goal-driven and multi-granular planners, treating document generation as a hierarchical decision-making process; The system encodes state representations at three levels: global state, local state, and micro state, and utilizes graph neural networks to fuse knowledge graph structure information. The top-level planner outputs a dynamic document outline sequence, and the bottom-level executor generates specific control instructions. The evaluation feedback-driven hierarchical policy network is continuously fine-tuned online using a proximal policy optimization algorithm to achieve dynamic alignment between content and knowledge networks.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention achieves joint extraction and dynamic mapping of unstructured text to knowledge triples from a large language model through end-to-end prompts. Combined with a knowledge retrieval-enhanced generation framework, it uses graph-structured knowledge as explicit constraints and introduces a consistency check loop, significantly improving the factual accuracy and logical consistency of the content. Simultaneously, it employs a hierarchical reinforcement learning architecture to drive a target-aware dynamic planner, enabling multi-granular adaptive control of document structure, detail, and style, achieving online optimization of the generation strategy and knowledge alignment. Finally, it uses a semantically aware dynamic template engine to intelligently bind content and format, supporting lossless output in multiple formats. This systematically solves the problems of fragmented knowledge extraction, weak content controllability, and rigid formatting in traditional document generation, achieving an organic integration of knowledge-driven and language generation. Attached Figure Description

[0017] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. In the drawings: Figure 1 This is a schematic diagram of the adaptive document generation system based on knowledge graphs and large models of the present invention. Figure 2This is a schematic diagram of the adaptive document generation method based on knowledge graphs and large models proposed in this invention. Detailed Implementation

[0018] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] This embodiment proposes an adaptive document generation system and method based on knowledge graphs and large models, such as... Figure 1 As shown, the system includes a knowledge graph construction module, a large model integration module, an adaptive control module, and a document generation and rendering module.

[0020] Among them, the knowledge graph construction module is responsible for extracting entities, relationships and attributes from multi-source data and constructing them into a queryable knowledge network; The large model integration module connects to a pre-trained large language model to understand user intent and generate coherent text content. The adaptive control module adjusts the document's structure, level of detail, and expression style based on user input, contextual environment, and feedback from the knowledge graph. The document generation and rendering module organizes the generated content according to a predetermined format and outputs it as a readable document.

[0021] In a preferred embodiment, the specific steps for entity extraction and knowledge network construction in the knowledge graph construction module are as follows: We adopt an end-to-end joint extraction framework based on a large language model to transform the traditional pipeline-style multi-step extraction task into a structured generation prompting engineering problem, realizing a direct and dynamic mapping from unstructured text to structured knowledge triples.

[0022] The system converts raw data from documents, databases, and API interfaces into plain text sequences. For each text segment, the system provides a structured hint containing instructions, examples, and the current text. This hint explicitly instructs the model to output the entities, relations, and attributes appearing in the text in a predefined triple format.

[0023] The preprocessed text sequence, along with prompts, is input into a large language model. Based on its understanding of the prompts and semantic analysis of the text content, the model simultaneously identifies entity pairs and their relationships during the generation process, and outputs the descriptive features of the entities as attributes. The model's output is strictly constrained to the formats "(entity1, relation, entity2)" and "(entity, attribute name, attribute value)".

[0024] The newly extracted triples first enter the entity linking stage. This stage calculates the similarity between the embedding vectors of the new entity and existing entities in the knowledge base, and matches them with contextual features to determine whether the new entity is an alias or reference of an existing entity. If the match is successful, entity merging and information supplementation are performed; if it is a new entity, a new node is created. The same normalization process is applied to relations and attributes.

[0025] The merged triples are imported into a graph database, where entities are nodes, relations are edges, and attributes are auxiliary fields of the nodes. The system encapsulates a unified query service interface on top of the graph database, supporting retrieval based on natural language questions. These natural language questions are transformed into a graph query language using a dedicated parsing model, enabling multi-hop queries and inference within the underlying knowledge network, ultimately returning structured query results.

[0026] In the large model integration module, a conditional generation framework based on knowledge retrieval is adopted. By retrieving precise structured knowledge fragments from the knowledge graph and using them as explicit conditions and constraints in the generation process, the large language model is guided to narrate within the boundaries of factual correctness and logical consistency.

[0027] Specifically, this module receives a document generation request from the adaptive control module. This request contains metadata such as topic, audience, and style. Employing a lightweight instruction parsing model, it first deconstructs the natural language request into an executable query intent. The system then automatically generates a precise knowledge graph-oriented query statement based on the intent to extract the entities, attributes, and relationship paths most relevant to the currently generated paragraph.

[0028] The structured results retrieved from the knowledge graph are converted into a prompting context in a hybrid format that incorporates knowledge and natural language. This format embeds key entities and relationships within a guided narrative framework using semantic tags and highlighted information. Simultaneously, the prompting context explicitly defines the generation goal of the text segment, the list of knowledge points to be covered, and the expressions to be avoided, thus establishing clear task boundaries and content baselines for subsequent generation.

[0029] The constructed dynamic contextual prompts are fed into a pre-trained large language model, enabling the model to generate text under strong constraints. During generation, an attention-guided mechanism ensures that the model's descriptions of specific entities or relationships maintain a high degree of consistency with the attribute values ​​provided in the knowledge graph. After a single paragraph is generated, a consistency check loop is used to extract key information from the generated text and compare it with the original retrieved knowledge. If a discrepancy is found, prompt revision and paragraph regeneration based on the differences are triggered.

[0030] After completing the current paragraph, the system uses the generated content as historical context and predicts the knowledge focus required for the next paragraph, starting a new cycle. Style parameters provided by the adaptive control module are converted into actionable style tokens and sentence templates, which are then integrated into the prompting construction stage to influence the model's output tendencies in vocabulary selection, sentence structure, and rhetorical devices.

[0031] In a preferred embodiment, the adaptive control module employs a hierarchical reinforcement learning framework based on goal-driven and multi-granular planners. It treats document generation as a hierarchical decision-making process, constructing an intelligent control core with long-term planning capabilities and short-term execution feedback to dynamically regulate the document's macro-structure, meso-level paragraph organization, and micro-level language style. The feedback from the knowledge graph is considered a crucial component of the environmental state, transforming user needs and context into a multi-objective reward function, enabling the system to autonomously learn optimal content organization and expression strategies in different situations.

[0032] Specifically, this module receives the initial request from the user, the complete context of the currently generated document, and the knowledge coverage and entity relationship network returned by the knowledge graph. It encodes these into three levels of state representation: a global state captures the overall document progress, topic consistency, and style fit; a local state describes the logical development, information density, and knowledge association strength of the current chapter or paragraph; and a micro-state tracks sentence-level rhetorical features, terminology accuracy, and readability metrics. By leveraging a graph neural network to fuse the structural information of the knowledge graph, the text context is transformed into a vector sequence through a semantic encoder.

[0033] The hierarchical strategy network comprises a top-level planner and a bottom-level executor. The top-level planner, based on the global state, outputs a dynamic sequence of document outlines, including the target knowledge coverage points, expected level of detail, and style tags for each section. The bottom-level executor, based on the current local state and the current paragraph target issued by the top-level planner, generates specific control instructions in real time, including the query focus for knowledge retrieval, the intensity parameters for text expansion or compression, and adjustment values ​​for sentence complexity and formality. Simultaneously, this planning process is continuously updated; after each chapter or paragraph is completed, the planner re-evaluates subsequent planning based on the latest state.

[0034] After each text segment is generated, the module uses a knowledge graph to verify factual accuracy, calculates text coherence scores using a pre-trained evaluator, compares the user-defined style with the style vector of the generated text to calculate the fit, and evaluates efficiency by combining generation length and information entropy. The hierarchical policy network is continuously fine-tuned online using a proximal policy optimization algorithm. Feedback from the knowledge graph directly participates in the calculation of accuracy and completeness rewards, ensuring dynamic alignment between content and the knowledge network.

[0035] The planner sends control commands output in real time to the knowledge graph construction module and the large model integration module. When the planner determines that the information density of the current paragraph is insufficient, it requests the knowledge graph module to provide more granular entity attributes and relationship paths, and instructs the large model to add explanatory content. After generation, the new content, along with evaluation feedback, is sent back to the adaptive control module to update the state representation and trigger the next round of decision-making. Simultaneously, the system supports real-time user interaction; for example, if a user highlights a section requesting expansion, this interaction will be directly injected as a high-priority reward signal, prompting the policy network to quickly adjust subsequent planning.

[0036] As a preferred embodiment, the document generation and rendering module adopts a semantically aware structured template engine and a dynamic template mechanism that can understand the semantic roles of content, intelligently binding abstract format requirements with specific text content to achieve precise adaptation between content and presentation.

[0037] Specifically, this module receives a continuous text stream, along with document structure instructions from the adaptive control module. It then segments the text stream logically according to chapters, paragraphs, lists, and chart descriptions. A lightweight semantic analyzer is subsequently used to label each unit with its semantic role. A structured template library is built, where page layout, font, font size, and color are defined for each template. Style mapping rules are also defined to specify the specific formatting to be used for content with different semantic roles. Based on the document type and the user's initial style preferences, the engine selects a base template and dynamically instantiates and combines the required style rule fragments according to the semantic tag set of the current content, generating a unique temporary template with multiple formatting rules for the current document.

[0038] The segmented and semantically labeled text content is sequentially input into the corresponding areas defined in the temporary template. For key entities or data points in the text, if the knowledge graph provides corresponding visualization suggestions, the rendering engine calls the built-in chart generation component to convert the relevant data into a chart of the specified type and automatically inserts it into the corresponding positions of the data description labels in the text. At the same time, it generates and inserts the chart title and explanatory text. Special content such as formulas and code is processed using a dedicated typesetting mode.

[0039] After rendering is complete, the document forms a complete Document Object Model (DOM) in memory, containing all styles and embedded objects. This model can be exported losslessly to various common formats, such as editable document formats, fixed-layout formats, and fluid layout formats suitable for web pages. Before final output, the system provides a visual preview and automatically detects and corrects formatting errors, ensuring that the output document has a professional look that is ready to read or deliver.

[0040] The method steps proposed in this embodiment are the same as the implementation principle of the above system, specifically as follows: Figure 2 As shown, this embodiment organically combines knowledge-driven learning with language generation, ensuring both the accuracy and structure of the content while maintaining the naturalness and fluency of the text. Through an adaptive mechanism, it can flexibly handle document generation tasks in different scenarios, improving the efficiency and effectiveness of information delivery.

[0041] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. An adaptive document generation method based on a knowledge graph and a large model, characterized in that, Includes the following steps: Receive document generation requests from the user; Based on the document generation request, entities, relationships, and attributes are extracted from multi-source data to construct and update a queryable knowledge graph; Based on the document generation request, the current generation context, and the feedback from the knowledge graph, the document's structure, level of detail, and expression style are dynamically planned. Based on the planning results, relevant knowledge fragments are retrieved from the knowledge graph as conditions to drive the pre-trained large language model to generate coherent text content. The generated text content is structured according to semantic roles and rendered into a readable document format based on dynamic templates for output.

2. The method according to claim 1, characterized in that, The construction and updating of the queryable knowledge graph specifically includes: An end-to-end joint extraction framework based on a large language model is adopted. By assembling structured prompts containing instructions, examples and text to be processed, the large language model can directly output formatted knowledge triples. The output triples are linked and normalized. By calculating the similarity of the embedding vectors and combining the context, the new entities are matched and fused with the existing entities in the knowledge base. The processed triples are imported into a graph database to build a knowledge network and a query interface that supports natural language question retrieval.

3. The method according to claim 1, characterized in that, The process of driving the pre-trained large language model to generate coherent text content specifically includes: A knowledge retrieval-based conditional generation framework is adopted to convert the structured knowledge fragments retrieved from the knowledge graph into a prompt context containing semantic tags and generation constraints; The contextual prompts are input into the large language model, and an attention-guided mechanism is used during the generation process to ensure the consistency between the text description and the knowledge fragments. After paragraph generation, a consistency check is performed. If a deviation from the facts is found, a prompt for revision and regeneration is triggered.

4. The method according to claim 1, characterized in that, The structure, level of detail, and style of expression of the dynamic programming document specifically include: A hierarchical reinforcement learning framework based on goal-driven and multi-granular planners is adopted to model document generation as a hierarchical decision-making process; Using a hierarchical strategy network, dynamic document outlines and control instructions are output in a rolling manner based on the encoded global, local and micro states; The policy network's decision-making is optimized online using reinforcement learning based on a multi-objective reward function that considers accuracy, coherence, adaptability, and efficiency.

5. The method according to claim 1, characterized in that, The step of rendering the generated text content into a readable document format specifically includes: A semantically aware structured template engine is used to automatically label text units with semantic roles; Based on the semantic roles and user style preferences, style mapping rules are dynamically selected and instantiated from the template library to generate exclusive temporary templates; Based on the temporary template, text, charts, and special content are integrated and rendered, and exported into various common document formats.

6. An adaptive document generation system based on knowledge graphs and large models, characterized in that, It includes a knowledge graph construction module, a large model integration module, an adaptive control module, and a document generation and rendering module; The knowledge graph construction module is used to extract entities, relationships and attributes from multi-source data and construct a queryable knowledge network. The large model integration module is used to access a pre-trained large language model to understand user intent and generate coherent text content. The adaptive control module is used to dynamically adjust the document structure, level of detail, and expression style based on user input, context, and knowledge graph feedback. The document generation and rendering module is used to organize the generated content according to a predetermined format and output it as a readable document.

7. The system according to claim 6, characterized in that, The knowledge graph construction module adopts an end-to-end joint extraction framework based on a large language model, which transforms the multi-step extraction task into a structured prompting engineering problem; it converts the raw data into a plain text sequence and assembles a structured prompt containing instructions, examples and the current text. Input the text sequence along with prompts into a large language model, enabling the model to simultaneously identify entity pairs and their relationships, and output descriptive features as attributes; constrain the model output to a triple format.

8. The method according to claim 7, characterized in that, The knowledge graph construction module calculates the similarity between the embedding vectors of new entities and existing entities and performs entity linking and normalization by combining contextual features; it imports the fused triples into the graph database and encapsulates a unified query service interface that supports natural language question retrieval.

9. The method according to claim 6, characterized in that, The large model integration module receives a document generation request containing metadata about the topic, audience, and style, and deconstructs the document generation request into an executable query intent. Automatically generate query statements for knowledge graphs and extract relevant entities, attributes, and relationship paths; Convert search results into a contextual prompt that blends knowledge and natural language, and set the generation target, knowledge point list, and expressions to be avoided; An attention-guided mechanism ensures that the generated content is consistent with the attribute values ​​of the knowledge graph, and a consistency check loop is used to detect fact deviations and regenerate the content.

10. The method according to claim 6, characterized in that, The adaptive control module adopts a hierarchical reinforcement learning framework based on goal-driven and multi-granular planners, and regards document generation as a hierarchical decision-making process. The system encodes state representations at three levels: global state, local state, and micro state, and utilizes graph neural networks to fuse knowledge graph structure information. The top-level planner outputs a dynamic document outline sequence, and the bottom-level executor generates specific control instructions. The evaluation feedback-driven hierarchical policy network is continuously fine-tuned online using a proximal policy optimization algorithm to achieve dynamic alignment between content and knowledge networks.

Images