Business scheme generation method based on knowledge graph
By using a knowledge graph-based approach, standardized Word templates are automatically generated and populated, solving the problems of low efficiency, unstable quality, and cumbersome template filling in business solution writing, and achieving efficient and professional fault analysis report generation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ARMY ENG UNIV OF PLA
- Filing Date
- 2025-12-29
- Publication Date
- 2026-04-10
AI Technical Summary
Existing technologies suffer from problems such as low efficiency in writing business solutions, inconsistent quality, difficulty in knowledge reuse, and cumbersome template filling. This is especially true in industries such as power, telecommunications, manufacturing, and IT operations and maintenance, making it difficult to efficiently generate high-quality fault analysis reports.
By receiving standardized Word templates uploaded by users, the system obtains text describing the fault phenomenon, uses knowledge graphs to retrieve relevant nodes and large language models to generate structured content, and performs semantic matching to automatically fill the template while maintaining format integrity.
It has achieved highly efficient and automated business solution generation, ensuring professional content and consistent formatting, reducing human error, and improving writing efficiency and solution quality.
Smart Images
Figure CN121835642A_ABST
Abstract
Description
Technical Field
[0001] This invention mainly relates to the fields of computer data processing and artificial intelligence technology, specifically to natural language processing, knowledge graph applications, large language model generation, and document automation processing technology, and in particular to a business solution generation method based on knowledge graphs. Background Technology
[0002] In many industries, such as power, telecommunications, manufacturing, and IT operations, when equipment or systems malfunction, experienced engineers are required to prepare detailed fault analysis reports or business processing solutions. These solutions typically need to adhere to strict formatting standards and reference a wealth of professional knowledge, historical case studies, and solutions.
[0003] Currently, business plan development mainly relies on manual work, which presents the following major problems:
[0004] Inefficient: Writers need to spend a lot of time finding information, recalling historical cases, and doing tedious editing and typesetting work in standard templates.
[0005] Inconsistent quality: The quality of a solution is highly dependent on the personal experience and knowledge of the writer, resulting in significant differences in depth, accuracy, and standardization among solutions written by different individuals.
[0006] Difficulty in knowledge reuse: The vast amount of troubleshooting experience and professional knowledge accumulated by enterprises is often scattered in various documents or the minds of senior employees, making it difficult to form a structured knowledge system for effective reuse.
[0007] Template filling is cumbersome: Existing document generation tools can often only perform simple keyword replacements, making it difficult to accurately fill complex paragraph content into specific chapters of Word templates based on contextual semantics, and easily destroying the original format.
[0008] With the development of knowledge graph technology and Large Language Models (LLM), it has become possible to use structured knowledge to guide the generation of professional content. Therefore, there is an urgent need for a method that can combine an enterprise's private knowledge graph, use a large model to automatically generate high-quality business solution content, and intelligently populate it into a standardized Word template. Summary of the Invention
[0009] The purpose of this invention is to provide a business solution generation method based on knowledge graphs, so as to solve the problems of low solution writing efficiency, difficulty in knowledge reuse and rigid template filling in the above-mentioned background technology.
[0010] To achieve the above objectives, the present invention provides the following technical solution: a business solution generation method based on knowledge graphs, comprising the following steps:
[0011] S1. Receive a standardized report template file uploaded by the user, wherein the template file is in Word format;
[0012] S2. Obtain the user-input text describing the fault phenomenon;
[0013] S3. Based on the fault phenomenon input by the user, retrieve the most similar fault phenomenon node in the local knowledge base, and obtain its first-order and second-order neighbor nodes and their relationships in the knowledge graph to form a knowledge block;
[0014] S4. Take the fault phenomenon and its neighboring node knowledge blocks as input, and generate structured business solution content through the large model;
[0015] S5. Identify headings at all levels from the uploaded Word template, and semantically match the generated structured business solution content with the template headings to establish a mapping relationship;
[0016] S6. Insert the structured business solution content into the template after the matched title paragraph, while keeping the original paragraph structure, title level and layout of the template unchanged, so as to achieve automatic content filling;
[0017] S7. Save the autofilled Word document and provide a download function for the generated document.
[0018] Preferably, the specific method for obtaining the fault description text input by the user is as follows:
[0019] Perform a preprocessing operation on the input fault description text to remove leading and trailing spaces, and then perform URL-safe encoding on the preprocessed text;
[0020] The encoded query parameters are sent to the backend fault matching interface via an HTTP request;
[0021] The backend fault matching interface splits the query string into keywords by delimiters, then iterates through all fault phenomenon entities in the database, filters out fault phenomena that contain the input keywords, and returns a matching list.
[0022] Preferably, the specific method for retrieving the most similar fault phenomenon node from the local knowledge base based on the fault phenomenon input by the user, and obtaining its first-order and second-order neighbor nodes and their associations in the knowledge graph to form a knowledge block is as follows:
[0023] The Sentence Transformer model is used to convert user description text into high-dimensional semantic vectors.
[0024] Calculate the cosine similarity between the high-dimensional semantic vector and the vectors of all fault phenomenon entity nodes in the knowledge graph, and select the node with the highest similarity as the anchor node in the graph.
[0025] Centered on the anchor node, perform a multi-hop query in the graph database to obtain all first-order neighbor nodes directly connected to the anchor node; and obtain the second-order neighbor nodes associated with the first-order neighbor nodes.
[0026] Anchor points, first-order neighbors, second-order neighbors, and their mutual edge attributes are concatenated according to the entity-relation-entity triple format or the natural language narrative format to form a specially structured text paragraph, namely a knowledge block.
[0027] Preferably, the Sentence Transformer model generates sentence embedding vectors through a pre-trained Transformer encoder and pooling layers.
[0028] Preferably, the structured business solution content is generated by taking the fault phenomenon and its neighboring node knowledge blocks as input and using the large model: a prompt word containing instructions, context and output requirements is constructed and input into the large language model, and a business processing solution is generated based on the fault phenomenon and its neighboring node knowledge blocks.
[0029] Preferably, the generated results are represented in key-value pairs, where each key corresponds to an existing node type in the knowledge base, and each value corresponds to the content of the corresponding node in the knowledge graph.
[0030] Preferably, the specific method for identifying headings at all levels from the uploaded Word template and semantically matching the generated structured business solution content with the template headings to establish a mapping relationship is as follows:
[0031] The uploaded Word document is parsed, paragraph text is read, and headings from level one to level four are identified and formed into a heading set according to regular expression matching rules. Then, the heading set and the structured content fields generated by the large model are semantically encoded, and the cosine similarity between the heading vector and the field vector is calculated. The heading-field mapping relationship is established only when the similarity exceeds a preset threshold.
[0032] Compared with the prior art, the significant advantages of this invention are:
[0033] High efficiency and automation: The process of knowledge retrieval, content generation and document editing is automated, which greatly shortens the time for writing business solutions.
[0034] Content professionalism and standards: The content is generated based on the enterprise's private knowledge graph, which ensures that the knowledge referenced in the solution is accurate and professional, and reduces human error and bias.
[0035] Intelligent semantic fill: Unlike simple keyword replacement, this invention accurately inserts the generated content into the corresponding chapter of the template through semantic matching, and perfectly preserves the complex layout format of the Word template through technical means. The generated document can be used without secondary processing.
[0036] User-friendly interface: Features such as template management and error input autocomplete are provided, which lowers the barrier to entry for using the system. Attached Figure Description
[0037] Figure 1 This is a flowchart of a knowledge graph-based business solution generation method provided in an embodiment of the present invention. Detailed Implementation
[0038] The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings in the embodiments of the present invention.
[0039] like Figure 1 As shown, this invention provides a business solution generation method based on a knowledge graph. This method relies on a pre-built enterprise private domain knowledge graph, which stores entities such as equipment information, fault phenomena, fault causes, solutions, related cases, and involved tools, as well as their interrelationships. The specific implementation steps are as follows:
[0040] Step S1: Receive Standardized Report Template File: Users upload a blank or semi-blank standardized Word report template (.docx format) through the front-end system interface. The system supports drag-and-drop upload and file selection upload. After successful upload, the file is temporarily stored in a specified location on the server, the interface displays the file name, and provides options to download and review or delete the template. The template has pre-set standard chapter titles, such as "I. Description of Fault Phenomenon", "II. Analysis of Fault Causes", "III. Suggested Solutions", etc., and the corresponding fonts, font sizes, and paragraph styles are set.
[0041] Step S2: Obtain the fault description text: The user enters the observed fault phenomenon in the text input box on the system interface. To improve efficiency, the system integrates an auto-completion function. When the user enters, for example, "main pump temperature," the system front-end first performs a preprocessing operation on the entered fault phenomenon text, removing leading and trailing spaces, and then performs URL-secure encoding on the preprocessed text. The encoded query parameters are sent to the back-end fault matching interface via an HTTP request. The back-end segments the query string into keywords using space, comma, semicolon, and other delimiters, then iterates through all fault phenomenon entities in the database, filters out fault phenomena containing the input keywords, and returns a matching list. For example, if the user enters "main pump temperature," the system recommends "main pump bearing temperature too high" and "main pump coolant temperature abnormal." The user clicks to select an item from the recommendation list, and the system determines the text finally selected by the user as the "fault phenomenon description text."
[0042] Step S3: Retrieve Knowledge Graph to Form Knowledge Blocks After receiving the user-defined description of the fault phenomenon, the system performs the following operations:
[0043] This architecture uses a Sentence Transformer model to convert user-described text into high-dimensional semantic vectors. It generates sentence embedding vectors through a pre-trained Transformer encoder and pooling layers. The core of the Transformer encoder is based on an HFL / Chinese-Macbert-Base pre-trained model, employing a Transformer architecture to process the input text. The maximum sequence length of the encoder is set to 128 tokens, effectively capturing long-distance dependencies in the Chinese context. After the Transformer encoder, the model uses mean pooling to aggregate the token vectors of each sentence into fixed-length sentence vectors. This maps each sentence into a dense semantic space, facilitating subsequent similarity calculations.
[0044] Calculate the cosine similarity between the high-dimensional semantic vector and the vectors of all "fault phenomena" entity nodes in the knowledge graph, and select the node with the highest similarity as the anchor node in the graph.
[0045] Using the anchor node as the center, perform a multi-hop query in the graph database. First, obtain all first-order neighbor nodes directly connected to the anchor node (such as fault causes and solutions); then obtain the associated nodes of the first-order neighbor nodes, i.e., second-order neighbor nodes (such as maintenance tools and production units).
[0046] The retrieved anchor points, first-order neighbors, second-order neighbors, and their mutual edge attributes are concatenated according to the "entity-relationship-entity" triple format or natural language narrative format to form a structured text paragraph specific to the fault, namely "knowledge block".
[0047] Step S4: Generate a structured business solution from the large model: Construct a prompt word input containing instructions, context, and output requirements into the large language model. Example prompt word: "You are a senior equipment maintenance expert. Based on the following user-input fault symptoms and related knowledge blocks, generate a detailed business processing solution. The output format should be JSON key-value pairs, where the key must be one of the following categories: ['Fault Overview', 'Possible Causes', 'Troubleshooting Steps', 'Solution', 'Required Resources'].
[0048] Step S5: Semantic matching of template title and generated content
[0049] The system uses a Python library to parse the Word template uploaded in step S1 and iterate through the document paragraphs.
[0050] Identify paragraph texts with heading-level features and form a heading set (e.g., [“I. Description of Fault Phenomenon”, “II. Analysis of Fault Causes”, “III. Suggested Solutions”...]).
[0051] The text in the title set and the key fields (such as "fault overview" and "possible causes") in the JSON data generated by the large model are vectorized and encoded, so that texts from different sources and with different expressions are mapped to a unified semantic space. The Sentence Transformer semantic encoding model is used for deep semantic representation of the text. This model is based on the Transformer encoding architecture, models the contextual relationships between words through a self-attention mechanism, and introduces a sentence-level semantic aggregation strategy on the basis of a pre-trained language model. This maps variable-length text into dense semantic vectors of fixed dimensions, thereby accurately characterizing the overall semantic features of the text.
[0052] Calculate the cosine similarity between each pair of title vectors and key vectors. Set a threshold of 0.85; if the similarity is higher than the threshold, establish a mapping. For example, if the similarity between "II. Fault Cause Analysis" and "Possible Causes" is 0.92, then a mapping relationship is established.
[0053] Step S6: Automatic Content Filling and Formatting. Based on the mapping relationship established in S5, manipulate the Word document object again. Locate the matching heading paragraph (e.g., locate the paragraph "II. Fault Cause Analysis"). Insert a new paragraph after this paragraph and write the mapped content value (i.e., the specific text corresponding to "Possible Causes") into the new paragraph. When writing, explicitly specify that the new paragraph adopts the predefined "Normal" style in the template, thereby ensuring that the inserted content is completely consistent with the original style of the template in terms of font, line spacing, indentation, etc., and will not affect the style of the heading itself.
[0054] Step S7: Save and Download the Document. Save the completed Word document object as a new .docx file to the server output directory. The system generates a download URL for this file and returns it to the front end. Users can click the "Download Solution" button on the interface to obtain the final completed business solution document.
[0055] The above-described specific implementation is only a preferred implementation of the present invention. Of course, the present invention may have other various embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art can make various corresponding changes and modifications according to the present invention, but these corresponding changes and modifications should all fall within the protection scope of the claims of the present invention.
Claims
1. A business solution generation method based on knowledge graph, characterized in that, Includes the following steps: S1. Receive a standardized report template file uploaded by the user, wherein the template file is in Word format; S2. Obtain the user-input text describing the fault phenomenon; S3. Based on the fault phenomenon input by the user, retrieve the most similar fault phenomenon node in the local knowledge base, and obtain its first-order and second-order neighbor nodes and their relationships in the knowledge graph to form a knowledge block; S4. Take the fault phenomenon and its neighboring node knowledge blocks as input, and generate structured business solution content through the large model; S5. Identify headings at all levels from the uploaded Word template, and semantically match the generated structured business solution content with the template headings to establish a mapping relationship; S6. Insert the structured business solution content into the template after the matched title paragraph, while keeping the original paragraph structure, title level and layout of the template unchanged, so as to achieve automatic content filling; S7. Save the autofilled Word document and provide a download function for the generated document.
2. The knowledge graph-based business solution generation method as described in claim 1, characterized in that, The specific method for obtaining the user-input description of the fault phenomenon is as follows: Perform a preprocessing operation on the input fault description text to remove leading and trailing spaces, and then perform URL-safe encoding on the preprocessed text; The encoded query parameters are sent to the backend fault matching interface via an HTTP request; The backend fault matching interface splits the query string into keywords by delimiters, then iterates through all fault phenomenon entities in the database, filters out fault phenomena that contain the input keywords, and returns a matching list.
3. The knowledge graph-based business solution generation method as described in claim 1, characterized in that, The specific method for retrieving the most similar fault phenomenon node from the local knowledge base based on the fault phenomenon input by the user, and obtaining its first-order and second-order neighbor nodes and their relationships in the knowledge graph to form a knowledge block is as follows: The Sentence Transformer model is used to convert user description text into high-dimensional semantic vectors. Calculate the cosine similarity between the high-dimensional semantic vector and the vectors of all fault phenomenon entity nodes in the knowledge graph, and select the node with the highest similarity as the anchor node in the graph. Centered on the anchor node, perform a multi-hop query in the graph database to obtain all first-order neighbor nodes directly connected to the anchor node; and obtain the second-order neighbor nodes associated with the first-order neighbor nodes. Anchor points, first-order neighbors, second-order neighbors, and their mutual edge attributes are concatenated according to the entity-relation-entity triple format or the natural language narrative format to form a specially structured text paragraph, namely a knowledge block.
4. The knowledge graph-based business solution generation method as described in claim 3, characterized in that, The SentenceTransformer model generates sentence embedding vectors through a pre-trained Transformer encoder and pooling layers.
5. The knowledge graph-based business solution generation method as described in claim 1, characterized in that, Taking the fault phenomenon and its neighboring node knowledge blocks as input, the large model generates a structured business solution. Specifically, it constructs a prompt word containing instructions, context, and output requirements and inputs it into the large language model, generating a business processing solution based on the fault phenomenon and its neighboring node knowledge blocks.
6. The knowledge graph-based business solution generation method as described in claim 1, characterized in that, The generated results are represented in key-value pairs, where each key corresponds to an existing node type in the knowledge base, and each value corresponds to the content of the corresponding node in the knowledge graph.
7. The knowledge graph-based business solution generation method as described in claim 1, characterized in that, The specific method for identifying headings at all levels from the uploaded Word template and semantically matching the generated structured business solution content with the template headings to establish a mapping relationship is as follows: The uploaded Word document is parsed, paragraph text is read, and headings from level one to level four are identified and formed into a heading set according to regular expression matching rules. Then, the heading set and the structured content fields generated by the large model are semantically encoded, and the cosine similarity between the heading vector and the field vector is calculated. The heading-field mapping relationship is established only when the similarity exceeds a preset threshold.