Patent document adaptive processing method and device, and computer device

Abstract

This application discloses an adaptive processing method, apparatus, and computer device for patent documents. Based on the structured data of patent documents, it identifies the semantic boundaries and technical paragraphs of the patent documents, dividing them into several text blocks. It dynamically selects and invokes a text embedding model to generate semantic vectors for the text blocks based on the first features of the patent documents and text blocks. It extracts multi-dimensional features from the patent documents and encodes them into multi-dimensional feature vectors. The patent metadata, semantic vectors, multi-dimensional feature vectors, and patent documents are stored in a relational database, a vector database, and an object storage, respectively, establishing index connections between these storage systems. This application introduces a large language model for structure recognition and feature extraction, adaptive segmentation, and dynamic model selection, transforming messy patent documents into high-quality data assets with clear structure, rich semantics, and distinctive features. This significantly improves the quality and usability of patent data, providing rich feature support for the evaluation of high-value patents.

Description

Technical Field

[0001] This application relates to the fields of computer data processing and artificial intelligence technology, and in particular to a method, apparatus and computer equipment for adaptive processing of patent documents. Background Technology

[0002] With the acceleration of global technological innovation, the number of patent documents is growing exponentially. As a special type of legal and technical text, patent documents are characterized by complex structure, high specialization, and length. When building a high-value patent analysis system, high-quality preprocessing, feature extraction, and vectorization of massive amounts of patent documents are fundamental for subsequent analyses (such as value assessment, similarity retrieval, and trend prediction). However, existing patent data processing technologies have many limitations. Document structure recognition is rigid, and traditional processing methods typically rely on fixed regular expressions or position-based rules to extract key information such as titles, abstracts, and claims. However, patent document formats vary greatly across different countries and periods, making fixed rules difficult to adapt, leading to failures or errors in key information extraction. This severely affects the accuracy of subsequent analyses and cannot support in-depth mining and evaluation of high-value patents.

[0003] Therefore, there is an urgent need for an intelligent method that can deeply understand the semantics of patent documents and dynamically adjust the processing strategy according to the inherent characteristics of the documents in order to solve the above problems. Summary of the Invention

[0004] The main purpose of this application is to solve the technical problem that existing technologies cannot dynamically and intelligently process patent documents.

[0005] The first aspect of this application provides a method for adaptive processing of patent documents, including: Utilize large language models to parse current patent documents and obtain structured data; Based on the structured data, the semantic boundaries and technical paragraphs of the current patent document are identified, and the corresponding segmentation strategy is executed according to the preset decision logic to divide the current patent document into one or more text blocks; Extract the first feature of the current patent document and text block, where the first feature includes language type, technical field, technical complexity of text block and special elements, and dynamically call the corresponding text embedding model based on the first feature to generate semantic vector of text block; Natural language processing techniques are used to extract multidimensional features from current patent documents and encode them into multidimensional feature vectors. The patent metadata in the structured data is stored in a relational database, the semantic vectors and multidimensional feature vectors are stored in a vector database, the current patent document is stored in an object storage, and an index association is established between the storage.

[0006] A second aspect of this application provides a patent document adaptive processing apparatus, which specifically includes: The parsing module is used to parse the current patent document using a large language model to obtain structured data; The segmentation module is used to identify the semantic boundaries and technical paragraphs of the current patent document based on structured data, and execute the corresponding segmentation strategy according to the preset decision logic to divide the current patent document into one or more text blocks. The semantic vector generation module is used to extract the first feature of the current patent document and text block. The first feature includes the language type, technical field, technical complexity of the text block and special elements of the patent. Based on the first feature, the corresponding text embedding model is dynamically called to generate the semantic vector of the text block. The multidimensional feature extraction module is used to extract multidimensional features of the current patent document using natural language processing technology and encode them into multidimensional feature vectors. The storage module is used to store patent metadata from structured data into a relational database, semantic vectors and multidimensional feature vectors into a vector database, store the current patent document into an object storage, and establish index relationships between the various storages.

[0007] A third aspect of this application provides a computer device, comprising: a memory and at least one processor, wherein the memory stores instructions; and the at least one processor invokes the instructions in the memory to cause the computer device to execute the aforementioned patent document adaptive processing method.

[0008] The fourth aspect of this application provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the aforementioned patent document adaptive processing method.

[0009] This application provides a method for adaptive processing and multimodal vectorization of patent documents based on deep semantic understanding. It addresses the deep semantic understanding, adaptive segmentation, multimodal vectorization, and storage of unstructured patent documents, standardizing the storage of multi-source data and supporting incremental updates and quality control. It supports second-level retrieval and analysis of massive amounts of multimodal patent data. By introducing a Large Language Model (LLM) for structure recognition and feature extraction, combined with adaptive segmentation and a dynamic model selection mechanism, the quality and usability of patent data are significantly improved. Leveraging the powerful semantic understanding capabilities of the LLM, it can adapt to various patent document formats, accurately extracting structural information and greatly improving the accuracy and coverage of extraction. The adaptive segmentation strategy ensures the relative semantic integrity of each text block, avoiding the semantic fragmentation problem caused by traditional fixed-size segmentation. The dynamic model selection mechanism matches appropriate models to patents with different features, enabling the generated vectors to more accurately reflect their semantic information. Intelligent feature extraction generates multi-dimensional feature vectors containing innovation points, citation networks, etc., providing rich feature support for high-value patent evaluation.

[0010] This application can transform messy original patent documents into high-quality data assets with clear structure, rich semantics, and distinctive features, providing a solid foundation for subsequent intelligent evaluation. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of one embodiment of the patent document adaptive processing method in this application. Figure 2 This is a schematic diagram of one embodiment of the patent document adaptive processing device in this application. Figure 3 This is a schematic diagram of one embodiment of the computer device described in this application. Detailed Implementation

[0012] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms “comprising” or “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0013] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 The patent document adaptive processing method in this application embodiment includes: S100: Use a large language model to parse the current patent document and obtain structured data.

[0014] Specifically, this step is used for document acquisition and structure recognition: The current patent document is received, and structural analysis prompts are constructed. These prompts contain instructions, including layout analysis instructions, to instruct the Large Language Model (LLM) or multimodal large language model to intelligently identify the patent type, document structure (i.e., document structure type), key field locations, special elements, etc., of the current patent document, and are returned in JSON format. The output of the large language model is parsed into structured data, which guides subsequent segmentation and extraction operations. The structured data includes patent metadata, patent type, document structure format, and key field locations. Patent metadata includes the patent unique identifier, applicant information, and classification number.

[0015] The patent types include invention patents, utility model patents, and design patents.

[0016] Document structure formats include standard formats, PDF scans, structured PDFs, XML plain text, Word text, and other formats.

[0017] The unique identifier for a patent can be the application number or the publication number.

[0018] The classification number can be an IPC classification number or a main classification number, etc.

[0019] For current patent documents that are not in plain text format, the system first calls a high-precision OCR (Optical Character Recognition) engine or text extraction tool to recognize the text, while preserving layout information such as the coordinates of text blocks, font size, and table borders. This process not only converts images or PDFs into computer-readable text streams, but more importantly, it preserves the document's physical structure, which is crucial for distinguishing titles, body text, headers, and footers.

[0020] Key fields include the patent title, abstract, claims, description (technical field, background art, invention content, description of drawings, detailed embodiments), and drawings.

[0021] Structured data can also include the metadata skeleton of the current patent document and the coordinates or index information of each key area, which can be used to guide subsequent accurate extraction.

[0022] The following are examples: Example 1: Extract a pre-defined length of text from the beginning of the patent document, for example, extract the first 500 or 1000 characters of the current patent document as input to the large language model.

[0023] Construct a prompt, for example: "Analyze the structure and key parts of the following patent document: the first 500 characters of the document content: {content}. Please identify: 1. Patent type (invention, utility model, design); 2. Document structure (standard, special, hybrid); 3. Key field locations; 4. Special elements. Return results in JSON format."

[0024] Call the LLM interface to obtain structured data.

[0025] Parse the returned JSON to obtain data such as patent_type, structure_format, and key_fields.

[0026] Example 2: Example Prompt: "You are a senior patent document analysis expert. Please analyze the following patent document fragment and its corresponding layout information. Your task is to identify the document's internal structure. Please pay attention to the following points: 1. Determine the patent type (invention, utility model, design); 2. Identify the document's format (e.g., whether there are explicit [Abstract], [Claims], etc., or a mixed flow layout); 3. Locate the starting position or coordinate range of key fields (title, abstract, claims); 4. Identify special elements, such as complex nested tables, blocks of mathematical formulas, chemical structure diagrams, and hierarchical relationships of multi-level headings. Please output the results in JSON format, ensuring that it includes field names and their corresponding text content or page number coordinates."

[0027] The system inputs this prompt along with the document fragment into a pre-trained multimodal large language model. This model possesses visual understanding capabilities, enabling it to recognize document layout and make comprehensive judgments based on textual semantics. For example, when encountering an old, scanned patent document where an OCR error causes the "claims" section to be identified as a "claims request," traditional regular expression matching will fail. However, the large language model can infer that this is the "claims" section based on its position on the page (usually after the specification) and the legal stylistic features of the context, and correctly mark it.

[0028] The model outputs a JSON result containing high-precision structured information. The system parses this JSON to obtain the document's metadata skeleton.

[0029] For design patents, after identifying the patent type, the system automatically calls specialized visual branches such as image extraction and description extraction logic, uses object detection models (such as YOLO or Mask R-CNN) to locate the product design view in the attached drawings, and extracts a brief description of the view.

[0030] For invention patents containing mathematical formulas, the system uses a specialized formula recognition engine (such as LaTeX-OCR) or mathematical formula parser to convert the formula image into editable and calculable LaTeX code based on the coordinates of the marked location.

[0031] Furthermore, this embodiment may include a special error correction mechanism. The system compares the identified structured data with standard templates in the metadata database. If an anomaly is detected (e.g., the claims of an invention patent appear before the abstract), the system triggers a secondary confirmation process. By expanding the input text range or adjusting the Prompt parameter, the system requests the large language model to re-analyze, thereby ensuring extremely high accuracy in structure recognition. This parsing method based on deep learning and layout awareness completely solves the adaptability problem of traditional rule-based methods when dealing with documents with complex formats, laying a solid foundation for subsequent refined processing.

[0032] Traditional extraction methods typically use fixed rules, which are prone to failure when dealing with documents of different formats. This embodiment uses a large language model for document content analysis, which can not only extract text but also perceive document layout, thus overcoming the limitations of fixed rules.

[0033] S200: Based on structured data, identify the semantic boundaries and technical paragraphs of the current patent document, and execute the corresponding segmentation strategy according to the preset decision logic to divide the current patent document into one or more text blocks.

[0034] Specifically, this step is used for intelligent block segmentation: determine whether the patent type is an invention patent; if so, identify the technical blocks in the current patent document and segment the current patent document according to the technical blocks; if it is not an invention patent, determine whether the claim section contains a large number of claims or has a complex hierarchical structure; if so, treat each claim as an independent text block; for other identified content, use an adaptive semantic block segmentation method to identify semantic turning points in the text using semantic similarity and segment at the semantic turning points.

[0035] This step, through adaptive chunking decisions, ensures the semantic integrity and technical independence of each text block.

[0036] S300: Extract the first feature of the current patent document and text block, wherein the first feature includes the language type, technical field, technical complexity and whether it contains special elements of the patent, dynamically select and call the corresponding text embedding model or multimodal embedding model according to the first feature, and generate the semantic vector of the text block.

[0037] Specifically, after the text blocks are segmented, the system enters the dynamic model selection stage. In order to make accurate decisions, the system extracts the deep features of each text block.

[0038] This step is used for dynamic model selection and vectorization: extract features of each text block or the entire document, including special elements such as language type, technical field, technical complexity, and whether it contains mathematical formulas, dynamically select the corresponding text embedding model based on the first feature, and generate semantic vectors.

[0039] Among them, the language type of the text block or the current patent document can be determined by a language detection library (such as langdetect), such as Chinese, English or other languages.

[0040] By using keyword matching or lightweight classifiers, it can be determined that the current patent document belongs to a certain field such as AI, computer, biomedicine, or mechanical structure.

[0041] Technical complexity can be determined by evaluating metrics such as average sentence length, density of technical terms or term frequency (TF-IDF), and density of logical connectors, generating a technical complexity score between 0 and 1. The higher the score, the more complex the text.

[0042] Regular expressions can be used to scan text blocks for one or more special elements, such as mathematical formulas (e.g., LaTeX format mathematical formulas), chemical formulas, code snippets, tables, or multi-level nested lists.

[0043] You can also use a LaTeX syntax detector to determine whether a text block contains mathematical formulas.

[0044] Image detection algorithms can be used to determine whether a key attached image exists in a text block.

[0045] The system constructs a decision tree or rule engine, which can select the optimal model for vectorization from multiple candidate models (such as bge-m3 and nomic-embed-text) based on the primary features of the text block, such as language, domain, and complexity.

[0046] Depending on the selected model, text blocks can be converted into fixed-dimensional vectors (e.g., 1024-dimensional floating-point arrays). The generated semantic vectors capture not only the literal meaning of the text but also the underlying technical logic.

[0047] This step uses a dynamic model selection mechanism to select the optimal model for vectorization, which can achieve the optimal vectorized representation of patent documents.

[0048] S400: Utilizes natural language processing technology to extract multidimensional features from the current patent document and encodes them into multidimensional feature vectors.

[0049] Specifically, this step is used for intelligent feature extraction: extracting at least one of the following multidimensional information: technical keywords, innovation points, citation features, legal features, commercial value features, and multimodal features, and constructing a multidimensional feature vector.

[0050] S500: Stores patent metadata from structured data in a relational database, stores semantic vectors and feature vectors in a vector database, stores the current patent document in an object storage, and establishes index associations between the storage systems.

[0051] Specifically, this step involves storage and indexing: a multi-level storage architecture is built to support efficient multi-dimensional retrieval. Multi-source data (metadata, vectors, and raw text) is stored in a standardized manner, supporting incremental updates and quality control.

[0052] The unique nature of patent data lies in the fact that it requires precise structured queries (such as "finding invention patents filed by Huawei after 2020"), fuzzy semantic searches (such as "finding patents related to autonomous driving LiDAR image stabilization technology"), and fast reading of large files (such as viewing the original PDF). A single database cannot simultaneously meet these needs. Therefore, this embodiment designs a hybrid architecture that combines relational databases, vector databases, and object storage.

[0053] Storing metadata (such as application number, applicant information, IPC classification number) from structured data into a relational database (such as MySQL, PostgreSQL) supports structured queries.

[0054] Relational Database Management System (RDBMS): MySQL or PostgreSQL can be used as the metadata center. It stores data with a well-defined structure and standardized fields. The table structure design includes: a basic information table (patent ID, application number, publication number, title, applicant, inventor, IPC classification number, application date, legal status), an index table (such as an inverted index created by the applicant), and a processing status table (recording the document's processing stage, number of chunks, and model version used). Relational databases handle precise filtering and selection operations, responding to structured queries in milliseconds.

[0055] Semantic vectors and multidimensional feature vectors can be stored in vector databases (such as Milvus, Pinecone, Weaviate, or Faiss), supporting approximate nearest neighbor search for high-dimensional vectors.

[0056] Vector Database: Milvus or Weaviate can be used as the core search engine. Two main types of vectors are stored: semantic vectors of text blocks and multidimensional feature vectors of patents. The vector database internally employs the HNSW (Hierarchical Navigable Small World) indexing algorithm. HNSW is a graph-based indexing structure capable of rapidly performing near nearest neighbor (ANN) searches in high-dimensional space. The system establishes an independent Collection for each field (such as title vector, abstract vector, claim vector, and specification vector). During storage, the vector is stored along with its corresponding patent ID. During retrieval, the user-input query text is vectorized in real time, and the system quickly locates the K nearest nodes to the query vector in the HNSW index, returning a list of the corresponding patent IDs. To improve retrieval accuracy, the system also supports hybrid queries, i.e., while performing vector retrieval, filtering by IPC classification number or applicant using RDBMS, and taking the intersection of the two results, thus achieving a highly precise search through "semantic + filtering".

[0057] Store the current patent document, extracted figures, and parsed formula files to object storage (such as MinIO, S3, etc.).

[0058] Object storage: MinIO or AWS S3 is used to store raw files and multimodal objects. Original PDF files, high-resolution images extracted from PDFs (PNG / JPG format), parsed LaTeX formula files (.tex), and even intermediate logs from the processing are all stored in object storage as file objects. Object storage offers high scalability and high throughput, making it suitable for storing massive amounts of unstructured large files.

[0059] Intelligent Index Association: The key to the system lies in how to connect three heterogeneous storage systems. The core link is the globally unique patent ID (usually a standardized application number). When storing vectors in the vector database, the patent ID and text block ID are attached. In the RDBMS, the patent ID is used to associate with the file path in the object storage.

[0060] Establish index relationships between different storage systems and tightly bind relational data, vector data, and file object data using globally unique identifiers (such as patent IDs) to facilitate the management of heterogeneous databases.

[0061] When a user initiates a search request: The system first analyzes the query intent to determine whether it is a structured query or a semantic query.

[0062] If it is a semantic query, the system performs an ANN search in the vector database to obtain a list of patent IDs and similarity scores.

[0063] The system retrieves the corresponding metadata (title, applicant, abstract, etc.) from the RDBMS in batches based on the list of patent IDs.

[0064] If a user needs to view the original text or attached images, the system streams the file from object storage to the front end in real time using the file path pointer in the metadata.

[0065] To accelerate retrieval, the system also incorporates a multi-level caching mechanism. For frequently accessed queries (such as high-frequency search terms), the system caches their vectors and results in Redis, directly hitting the cache without accessing the vector database. For frequent access to metadata, a caching layer is also deployed on top of the RDBMS. This carefully designed hybrid storage and intelligent indexing scheme ensures that the system maintains sub-second response times even when dealing with hundreds of millions of patent data points, providing users with a smooth interactive experience.

[0066] This application implements a patent document adaptive processing and multimodal vectorization method based on deep semantic understanding. It receives the original patent document, performs deep analysis using a Large Language Model (LLM), and intelligently identifies the document structure type, key field positions, and special elements. Based on the identified structural information and semantic boundaries, it executes an intelligent block segmentation strategy, including segmentation by technical blocks, by claim blocks, or adaptive semantic segmentation. According to the patent's language, technical field, complexity, and whether it contains mathematical formulas, it dynamically selects the optimal text embedding model for vectorization. It intelligently extracts technical keywords, innovation points, citation features, and legal features from the patent content to construct multidimensional feature vectors. Finally, it stores metadata, text block vectors, feature vectors, and the original file in a relational database, a vector database, and an object storage, respectively, and updates the data through an incremental processing mechanism. By endowing the data processing flow with deep understanding and adaptive decision-making capabilities, this application solves the problems of rigid structure recognition, fragmented semantic segmentation, and inappropriate model matching in traditional patent processing, providing a high-quality data foundation for the mining of high-value patents.

[0067] This application implements a method for deep semantic understanding, adaptive segmentation, multimodal vectorization, and storage of unstructured patent documents. It standardizes the storage of multi-source data (metadata, vectors, and original text) and supports incremental updates and quality control. It enables second-level retrieval and analysis of massive amounts of multimodal patent data. By introducing a Large Language Model (LLM) for structure recognition and feature extraction, combined with adaptive segmentation and dynamic model selection mechanisms, it significantly improves the quality and usability of patent data.

[0068] Deep understanding and high adaptability: Utilizing the powerful semantic understanding capabilities of large language models, it can adapt to patent documents of various formats, accurately extract structural information, and greatly improve the accuracy and coverage of extraction.

[0069] Semantic integrity guarantee: The adaptive chunking strategy ensures the relative semantic integrity of each text block, avoiding the semantic fragmentation problem caused by traditional fixed-size chunking.

[0070] Vector representation optimization: The dynamic model selection mechanism matches suitable models to patents with different features, so that the generated vectors can more accurately reflect their semantic information.

[0071] Deep Feature Mining: Intelligent feature extraction generates multi-dimensional feature vectors containing innovation points, citation networks, and other features, providing rich feature support for the evaluation of high-value patents.

[0072] This application can transform messy original patent documents into high-quality data assets with clear structure, rich semantics, and distinctive features, providing a solid foundation for subsequent intelligent evaluation.

[0073] In one embodiment, step S200 involves identifying the semantic boundaries and technical paragraphs of the current patent document based on structured data, and executing a corresponding segmentation strategy according to a preset decision logic to divide the current patent document into one or more text blocks, including: If the patent type is an invention patent, then identify the technical blocks and divide paragraphs in the current patent document that belong to the same technical feature into a text block; If the claims section is identified as containing several claims or having a hierarchical structure, then the text content corresponding to each claim or each level will be treated as an independent text block. For other identified text content, semantic similarity is used to identify semantic turning points in the text, and the identified other text content is segmented at the semantic turning points.

[0074] Specifically, in semantic retrieval systems based on vector databases, chunking is crucial. If a text chunk is too large, containing multiple unrelated technical points, the semantic vector during retrieval will be interfered with by noise, leading to a decrease in matching accuracy. If a text chunk is too small, it lacks sufficient contextual information, resulting in incomplete semantic expression and an inability to accurately answer the query. More importantly, traditional fixed-size (e.g., 512 token) sliding window segmentation methods completely ignore the semantic boundaries of the content. For example, traditional segmentation methods may cut off a detailed description of a technical solution or forcibly splice background technology with specific implementation methods. In short, existing solutions are highly prone to fragmenting a complete technical solution or disrupting the citation relationships of claims.

[0075] To address the aforementioned issues, this embodiment abandons the traditional fixed-size segmentation method and proposes an adaptive segmentation strategy based on content and semantics. The system intelligently selects and switches between three strategies according to the patent type, structure, and semantic flow within the text.

[0076] To determine if a patent is an invention patent, the first step is to identify "technical blocks." A technical block typically refers to a collection of paragraphs or texts that solve a specific sub-problem or achieve a specific function. By analyzing paragraph headings or using clustering algorithms, paragraphs that span the claims and specification and belong to the same technical feature are grouped into a single text block.

[0077] The specific implementation includes: Title hierarchy analysis: The system scans the various levels of headings in the document (such as "1. A...", "2. According to claim 1...", or "First embodiment", "Second embodiment"). These headings typically naturally delineate the boundaries of the technology.

[0078] Clustering algorithm assistance: For paragraphs without explicit titles, the system segments the text into fine-grained segments based on natural paragraphs. Then, it uses a sentence embedding model (such as SBERT) to calculate the vector representation of each segment, followed by clustering algorithms (such as K-Means or DBSCAN) to cluster these segments. Paragraphs belonging to the same category are considered to belong to the same technical block, and the system groups them into the same text block.

[0079] This block-based approach ensures that each text block fully describes an independent technical concept, making it ideal for comparative analysis of technical solutions.

[0080] When the system detects that the claim portion is particularly long or complex, it will trigger special claim processing logic to divide each independent claim or dependent claim into a separate text block.

[0081] In addition, the reference relationships between claims can be preserved as metadata.

[0082] For other cases (e.g., detailed descriptions in the instruction manual), the system can employ intelligent segmentation based on semantic boundaries or adaptive semantic segmentation. It uses models such as BERT to calculate the semantic similarity between adjacent sentences or paragraphs, identifies semantic inflection points (i.e., semantic boundaries) where similarity drops sharply, and segments the identified text content at the semantic boundaries corresponding to these inflection points.

[0083] Its core idea is that the optimal block boundary is where the semantic content undergoes a sudden change.

[0084] Semantic similarity calculation: The system first segments the document into a sequence of sentences or short paragraphs. Then, it uses a pre-trained semantic model (such as BERT or RoBERTa) to calculate the semantic similarity between two adjacent sentences / paragraphs (usually using cosine similarity).

[0085] Boundary recognition: The system plots a similarity curve. In coherent text, adjacent sentences are usually highly similar; however, when the topic changes (e.g., from "background technology" to "invention content"), the similarity drops sharply (i.e., "semantic inflection point").

[0086] Segmentation execution: The system sets a similarity threshold or finds a local minimum point. When the similarity is below the threshold, the system determines that this is a semantic boundary and performs segmentation at this point.

[0087] In addition, the system will dynamically adjust the size of the text blocks to avoid the excessively long context window affecting the model's attention mechanism.

[0088] For example, dynamic resizing: To prevent certain text blocks from becoming too long and affecting the attention mechanism of the retrieval model, the system also sets a maximum length limit (e.g., max_tokens=1000). If no obvious semantic turning point is detected when the maximum length is reached, the system will force a split, but will prioritize breaking at punctuation marks to maintain fluency.

[0089] By organically combining the three strategies described above, this embodiment ensures that each generated text block is semantically cohesive. For example, in a patent concerning "deep learning image segmentation," the system can accurately separate the "convolutional layer structure" from the "loss function definition," or store "claim 1" and "claim 2" independently. This lays a solid foundation for the quality of subsequent vectorization generation and solves the semantic fragmentation problem caused by traditional fixed-window segmentation.

[0090] This embodiment is based on structured data and dynamically selects to divide the data into blocks according to technical blocks, claims, or semantic boundaries, which can ensure semantic integrity.

[0091] In one embodiment, the identification technology block divides paragraphs belonging to the same technical feature in the current patent document into a text block, including: A large language model is used to perform full-text semantic analysis on the current patent document to construct a triplet knowledge graph containing technical issues, technical means, and technical effects. Based on the knowledge graph, the technical logic flow in the current patent document is identified, and the dependency relationship between core technical nodes and supporting descriptive content is determined. Based on the aforementioned dependencies, a semantic aggregation and segmentation strategy is executed to aggregate content scattered across different chapters but pointing to the same technical feature into one or more logically independent text blocks.

[0092] The step of identifying the technical logic flow in the current patent document based on the knowledge graph and determining the dependency relationship between core technical nodes and supporting descriptive content includes: Identify the claim tree structure in the current patent document and map each independent claim and its dependent claims to a core skeleton node in the knowledge graph; For the instruction manual section, a large language model is used to extract the entities and functions corresponding to each paragraph and map them as leaf nodes. Calculate the semantic relevance and text distance between leaf nodes and core skeleton nodes, and aggregate the leaf nodes that reach the preset relevance threshold and have the closest text distance to the corresponding core skeleton node.

[0093] Specifically, a large language model is used to perform a full semantic scan of the current patent document to construct a knowledge graph based on "problem-means-effect".

[0094] Technical issues: Identify the pain points or defects mentioned in the background technology (e.g., "the heat dissipation efficiency of the prior art is low").

[0095] Technical means: Identify the specific structures, algorithms or chemical components described in the specific implementation (e.g., "setting up liquid cooling pipes").

[0096] Technical effect: Identify beneficial results directly related to the means (e.g., "increase heat dissipation speed by 30%).

[0097] The system extracts the above elements into triples <node A, relation, node B>, such as <low heat dissipation efficiency, solution, install liquid cooling pipes>, <install liquid cooling pipes, produce effect, improve heat dissipation speed>. This series of triples constitutes the underlying semantic skeleton of the document.

[0098] After constructing the graph, the system identifies the technical logic flow of the document. This is not merely the order of the text, but the framework of the technical concept.

[0099] Identification of core technical nodes: The system typically identifies the technical features in the independent claims as "core skeleton nodes." These nodes constitute the core of the patent's protection scope.

[0100] Supporting content identification: Detailed descriptions in the specific implementation methods of the instruction manual (such as method steps, circuit diagrams, and chemical reaction steps) are identified as "leaf nodes" or "supporting nodes".

[0101] Calculating semantic relevance and dependency aggregation is key to achieving "cross-chapter aggregation". This embodiment achieves logical aggregation through the following calculations: Semantic relevance calculation: Calculate the semantic similarity between each leaf node (paragraph) and the core skeleton node (claims) in the specification. For example, this can be achieved by comparing keyword overlap, entity consistency, or utilizing pre-trained sentence vector similarity.

[0102] Text distance calculation: The system calculates the physical distance of leaf nodes in the document, which is usually used to assist in judgment because similar paragraphs often discuss the same topic.

[0103] Dependency determination: If the semantic relevance between a specification paragraph (leaf node) and a claim (core skeleton node) exceeds a second preset threshold (e.g., 0.85), and the paragraph mainly explains the features in the claim, then the paragraph is determined to depend on the claim.

[0104] Based on the above dependencies, the system performs an aggregation operation: Instead of strictly adhering to physical paragraph boundaries, multiple specification paragraphs determined to depend on claim 1 are logically "stitched" together with the text content of claim 1.

[0105] This stitching creates a larger text block. This text block contains: legal definitions (claims) + technical explanations (details of the specification). This greatly improves the accuracy of subsequent searches. When a user searches for "how feature A works," the search results directly hit this aggregated block, allowing the user to see both the legal boundaries and the explanation of the principles.

[0106] In one embodiment, step S200 specifically includes: Based on structured data, the text units in the current patent document are constructed into a semantic graph, where the text unit is a sentence or paragraph, the nodes in the semantic graph represent text units, and the weight of the edge represents the semantic similarity or reference relationship between text units. By using graph neural networks to traverse semantic graphs and learn features, the local and global semantic coherence of text units are calculated and evaluated, and semantic inflection points are determined. Based on the output of the graph neural network, a segmentation strategy is selected to divide the current patent document into one or more semantically complete text blocks.

[0107] Graph neural networks, for example, are built based on the GNN model.

[0108] More specifically, if the current patent document is an invention patent and the graph neural network identifies a cluster center for the technology blocks or a cluster structure exists, then the current patent document is divided into blocks according to the technology blocks to ensure the integrity of each technology block. Each technology block corresponds to a technical solution, including a description of the technical problem, the technical solution, and the technical effect, which is used to solve a technical problem and achieve a certain technical effect by using a certain technical means.

[0109] If the claims section is hierarchically complex, and the graph neural network detects strong dependencies but semantic independence between the claims, then each claim is treated as an independent text block. The graph of reference relationships between claims can also be preserved as metadata.

[0110] If the text content does not fall under the above categories, such as ordinary instruction manual paragraphs, an adaptive segmentation method based on semantic boundaries is adopted. This method uses the changes in node similarity output by the graph neural network to find semantic turning points or semantic turning point probability distributions, and segments other text content at the trough of the sharp drop in semantic similarity.

[0111] In addition, the size of the text block can be dynamically adjusted to fit the context window constraints of the subsequent vectorization model.

[0112] This application introduces a Large Language Model (LLM) for structure recognition and feature extraction, combines a Graph Neural Network (GNN) for complexity assessment, utilizes an adaptive segmentation strategy and a dynamic model selection mechanism, and integrates multimodal information from images and formulas to achieve adaptive semantic segmentation. This ensures the semantic integrity and technical independence of each text block, significantly improving the quality of patent data, vectorization representation capabilities, and the intelligence level of the retrieval system.

[0113] In one embodiment, technical complexity is obtained through the following steps: Using the node features generated by the graph neural network, the graph structure complexity index of the subgraph corresponding to the text block is calculated. The graph structure complexity index includes at least one of edge density, node degree distribution entropy, and clustering coefficient. The graph structure complexity index is combined into a technical complexity score with a value between 0 and 1.

[0114] Specifically, the graph structure complexity index mathematically quantifies the degree of technical entanglement in text. For example, a description of a complex mechanical structure containing numerous "wherein...connected to..." will inevitably have a high graph structure complexity; while a simple background introduction will have a lower complexity. The system combines these indices into a technical complexity value between 0 and 1.

[0115] In one embodiment, dynamically invoking the corresponding text embedding model based on the first feature specifically includes: Multiple alternative text embedding models are preset. The alternative text embedding models include at least a first model, a second model, and a third model. The first model is suitable for processing complex technical documents and documents containing special elements. The second model is suitable for general scenarios and has a faster processing speed. The third model is suitable for multimodal data processing. If the language of the current patent document is detected to be Chinese and the technical field belongs to a cutting-edge interdisciplinary field, or if the text block contains special elements, or if the technical complexity is higher than the preset threshold, then the first model is selected. If a text block is detected to contain an attached diagram or structural diagram, and / or if cross-modal detection is required or if the user's retrieval needs include text search based on images, then the third model is selected. If the text block is another regular technical document, then select the second model.

[0116] Specifically, this embodiment details how to dynamically select the optimal text embedding model from multiple candidate models based on the features of the text block, transforming the text block into a high-dimensional semantic vector. This is a key step in solving the problem that "a single model cannot adapt to all scenarios".

[0117] Existing patent vectorization systems often suffer from a "one-size-fits-all" problem, using the same model to process all documents. However, the patent world is extremely diverse: Language differences: Chinese, English, Japanese, German, etc., have huge differences in semantic space.

[0118] Domain differences: Patents in the biomedical field are full of gene sequences and chemical formulas; patents in the computer field are full of algorithmic logic; and patents in the mechanical field involve structural parameters. It is difficult for a general model to be proficient in all fields at the same time.

[0119] Differences in complexity: Some patents are all easy-to-understand descriptions, while others contain highly complex mathematical derivations and logical proofs.

[0120] To achieve the best balance between accuracy and speed, this embodiment proposes a dynamic model selection mechanism.

[0121] Vectorization serves as a bridge connecting unstructured text and computer algorithms. This embodiment constructs a model pool containing various heterogeneous models, pre-configuring multiple high-performance embedding models. The first model is a deep semantic model or a high-precision model, used to process documents with high technical complexity. It is a multilingual, multi-functional, high-performance model that supports dense, sparse, and multi-vector retrieval. The first model is particularly suitable for processing Chinese technical documents, complex logical reasoning, and scenarios containing mathematical formulas. For example, large-scale parameter models based on Transformer or BERT architectures, or bge-m3 (BAAI General Embedding) and gte-large models optimized for Chinese technical patents, are suitable for processing documents with high technical complexity, containing long and difficult sentences, and abstract concepts. The first model typically has a huge number of parameters (e.g., 7B or more) and a long context window, enabling it to capture complex logical relationships and long-distance dependencies. Although the first model has a slow inference speed and high computational cost, it has strong representational capabilities and can ensure a deep understanding of highly complex content.

[0122] The second model is a general-purpose, efficient, or high-speed model suitable for general scenarios. It is a lightweight model with fewer parameters and extremely fast inference speed, suitable for rapid vectorization in general scenarios, but its ability to understand extremely complex technical semantics is slightly inferior to the first model. For example, a lightweight dual-encoder model is suitable for general scenarios and has a fast processing speed; another example is the fast and versatile nomic-embedded-text model.

[0123] The third model is a multimodal alignment model, such as a CLIP-based architecture, a graph-text dual-tower model, or a graph-text multimodal encoder (such as Chinese-CLIP, OpenCLIP), specifically designed to generate vectors containing image and text alignment features.

[0124] In addition, the system can be extended to access specialized models such as ProtBERT (for processing protein sequences) for the biological field or ChemBERT for the chemical field.

[0125] Cutting-edge interdisciplinary fields include artificial intelligence, computer science, quantum computing, and bioinformatics.

[0126] If the technical complexity score of the current patent document is found to be higher than the preset threshold H1, or contains complex mathematical formulas, or the technical field belongs to a cutting-edge interdisciplinary field (such as quantum computing or bioinformatics), then the first model is selected to ensure the accuracy of deep semantic understanding. If a document is detected to contain important figures (key figures) or structural diagrams, and cross-modal detection is required (e.g., user search requirements include "searching for text by image"), then the third model is called to perform joint encoding of images and text to generate a multimodal vector that integrates image and text features. For other standard technical documents or scenarios that do not fall under either of the above two conditions, choose the second model in exchange for faster processing speeds and lower storage costs.

[0127] The preset threshold H1 can be any one of the following values: 0.8, 0.85, 0.9, etc.

[0128] For example, if the language is Chinese and the domain belongs to a highly abstract field such as "quantum computing" or "GAN network", or if mathematical formulas or other special elements are detected, or if the technical complexity score is >0.8 or 0.75, then the bge-m3 model is selected to ensure deep semantic understanding.

[0129] Otherwise, choose the nomic-embed-text model for faster processing speed. For example, if the text block is a simple English background technical introduction or a general mechanical structure description, then directly choose the nomic-embed-text model to maximize processing throughput.

[0130] This embodiment uses a fast model for simple content and a precise model for complex content, thereby maximizing the system's processing efficiency and economy while ensuring the accuracy of patent retrieval and analysis.

[0131] In one embodiment, in a batch processing scenario, text blocks are grouped according to the model selected for each text block, and text blocks using the same model are grouped into a text block set and sent to the corresponding model for vectorization processing.

[0132] Specifically, in batch processing mode, to maximize GPU utilization, the system adopts a "bucket aggregation" strategy. The system first scans all text blocks to be processed, statistically analyzes the model distribution, groups text blocks that need to use the same type together, and sends them in batches to the corresponding model service, thereby improving GPU utilization and throughput.

[0133] More specifically, text blocks are assigned to different queues based on the selected model. For example, all complex text blocks requiring the first model are placed in the "depth queue," while ordinary text blocks are placed in the "high-speed queue." The system loads the corresponding model weights for each queue and performs batch inference on the entire queue at once. This strategy avoids the overhead of frequent model switching and significantly improves throughput.

[0134] For example, all text blocks assigned to the "first model" are placed in queue A, and text blocks assigned to the "second model" are placed in queue B. The system then initiates batch processing tasks for the first and second models respectively. For instance, the 64 text blocks in queue A are packaged into a large batch and sent to the GPU for inference all at once. This processing method greatly improves GPU utilization and significantly reduces the overall processing time.

[0135] This embodiment, through dynamic selection and multimodal fusion, means that the system no longer mechanically generates vectors, but rather "teaches according to aptitude," matching appropriate computing resources to each patent and incorporating the cognitive habits of human experts when reading patents through "text and image comparison," thereby generating vectors with extremely high semantic fidelity and retrieval usability.

[0136] In one embodiment, multidimensional features are extracted from the current patent document using natural language processing techniques, and these multidimensional features are encoded into multidimensional feature vectors, including: The core technical terms are extracted from the current patent documents using a large language model. Based on the distribution of the core technical terms in the current patent documents, the weight of each core technical term is calculated. The core technical terms and weights are vectorized to generate feature vectors and weight vectors of the core technical terms. The distribution includes the frequency and position of occurrence. Based on a pre-built innovation vocabulary, the innovative descriptive fragments in the current patent document are identified through semantic analysis, and the innovative descriptive fragments are vectorized to generate innovation point vectors. Based on the citation data of the current patent documents, citation indicators are calculated, and citation indicators are vectorized to generate citation feature vectors. The citation indicators include at least one of citation strength, the number of times cited by the core patent, and the number of patent families. Extract legal and commercial indicators from the current patent document, and vectorize the legal and commercial indicators to generate legal feature vectors. The legal and commercial features include at least one of the following: IPC classification number, number of claims, hierarchical depth of dependent claims, legal status, and remaining patent life. The feature vectors of core technical terms, weight vectors, innovation point vectors, citation feature vectors, and legal feature vectors are combined to generate multidimensional feature vectors.

[0137] Specifically, traditional patent features are often limited to metadata (such as applicant, citation count) or simple textual statistics (such as TF-IDF). These features are superficial and cannot reflect the "quality" or "innovation level" of a patent.

[0138] This embodiment proposes to extract deep technical connotations from patent documents, extract multi-dimensional features, and construct a "multi-dimensional feature vector." This multi-dimensional feature vector is a composite vector with multiple dimensions, integrating information from semantics, innovation, law, and commerce, and can uniquely and comprehensively represent the core features of a patent. It aims to transcend traditional text vectors and provide unique data assets for the evaluation of high-value patents.

[0139] Intelligent identification of core technical terms (i.e., technical keywords): Use LLM to extract core technical terms (such as "deep learning", "image segmentation", "convolutional neural network", "carbon nanotube", "negative refractive index", "convolutional kernel", "LiDAR", "point cloud segmentation", "Kalman filter", etc.) and calculate the weight of each core technical term between 0 and 1.

[0140] In one specific embodiment, the weights can be calculated based on at least one of the following indicators: whether the core technical terms appear in the claims (especially in the independent claims where they have the highest weight), the frequency of their appearance in the specification, the distribution entropy of the core technical terms in the entire patent text (the more concentrated the distribution, the more important they may be), the scarcity (IDF) of the core technical terms in the current technical field, and the specificity of the core technical terms (general terms have lower weights, specialized terms have higher weights). This ultimately generates a weighted term vector. A higher weight indicates a more core term.

[0141] In another specific embodiment, the frequency and location of each core technical term in the claims and specification are statistically analyzed (e.g., whether it appears in the main claim, dependent claims, specification, or one or more other locations), and a weight is calculated based on the frequency and location of occurrence.

[0142] More specifically, for any core technical term, its frequency weight is the ratio of its occurrence frequency to the total frequency, where the total frequency is the sum of the occurrence frequencies of all core technical terms. The occurrence frequency includes the frequency of occurrence in core positions and the frequency of occurrence in non-core positions.

[0143] For any given core technical term, its position weight is the ratio of the number of times it appears in a core position to the total number of times it appears in a core position. The total number of times it appears in a core position is the sum of the number of times all core technical terms appear in core positions.

[0144] For any core technical term, its weight is the average of its frequency weight and its position weight.

[0145] Innovation point identification: Based on a pre-built "innovation trigger word dictionary (innovation word list)", which includes trigger words such as "first proposed", "originality", "breakthrough", "overcame the problem", "significantly improved", etc., combined with the semantic judgment of LLM, innovative descriptions are extracted from the text and vectors are generated.

[0146] In one specific embodiment, the system locates sentences in a document that contain descriptions of innovation. These sentences are then encoded using a specially trained "innovation semantic model." This model focuses not only on the sentences themselves but also on their context, extracting semantic triples of "technical problem-solution-technical effect." These triples are mapped to points in a vector space, forming innovation point features. This allows users to retrieve all patents that "solve a specific technical problem," rather than just those containing a particular keyword.

[0147] Citation Feature Analysis: This analyzes the citation and cited status of patents. Citation data includes external citation data and / or internal citation data within the current patent document. Based on the citation data, a citation network can be constructed. For example, if patent A cites patent B, the system will build a directed graph. Graph algorithms (such as PageRank or HITS) are used to calculate the graph features or node features of the patent in the citation network, thereby obtaining one or more citation indicators such as citation strength, the number of times it is cited by core patents, scientific relevance, and the number of patent families. If no citation data exists, prior art bottlenecks mentioned in the text are extracted based on the "background technology" section as potential citation relationship features.

[0148] Citation strength: The number of backreferences, representing the breadth of the technology.

[0149] Number of times cited by core patents: If a patent is cited by a large number of high-value patents, it indicates that it has strong fundamental value.

[0150] Scientific relevance: The number of non-patent documents (such as papers) cited represents the depth of its scientific foundation.

[0151] Patents are not isolated but exist within a knowledge citation network. The system combines external databases (such as citation databases) or internal citation data based on the "background technology" section of the current patent document to construct a citation relationship graph for the patent. Graph algorithms (such as PageRank) are used to calculate the patent's node centrality, i.e., its importance in the technology network. Simultaneously, the system analyzes its preceding patents (whether they cite foundational core patents) and following patents (whether they are cited by a large number of subsequent patents). These graph indicators are quantified into feature vectors. If a patent is cited by a large number of high-value patents, its citation feature vector will exhibit high potential energy.

[0152] Legal and commercial characteristics: This involves extracting one or more of the following: IPC classification number, number of claims (scope of protection; generally, more claims indicate a more complex scope of protection), legal status (valid, invalid, under substantive examination, rejected, withdrawn, etc.), remaining patent life, and the depth of dependent claims. These hard indicators are normalized and encoded into vectors.

[0153] Multimodal feature fusion: The feature vectors of the attached figures, the structure vectors of the formulas, and the feature vectors of the text are concatenated or weighted by an attention mechanism.

[0154] Multidimensional feature vector fusion: All the feature vectors extracted above are fused through weighted concatenation or an attention mechanism to form a fixed-dimensional high-dimensional vector, i.e., a multidimensional feature vector. A multidimensional feature vector can uniquely and comprehensively represent the core features of a patent.

[0155] In one specific embodiment, the system fuses all the aforementioned vectors through a multi-gating mechanism. Considering that the importance of different dimensions varies across different scenarios, the system employs an attention mechanism to dynamically allocate weights. For example, when evaluating "technological advancement," the weights of the innovation point vector and the citation network vector are increased; when evaluating "legal stability," the weights of legal features are increased. The fused high-dimensional vector (e.g., 2048 dimensions) is the multidimensional feature vector of this patent.

[0156] In another specific embodiment, the core technical terminology feature vector, weight vector, innovation point vector, citation feature vector, and legal feature vector are concatenated end to end. If the concatenated dimension is too high, PCA (principal component analysis) or a neural network can be used to compress it to a fixed dimension (e.g., 256 or 512 dimensions).

[0157] Traditional patent features are often limited to the title, abstract, and IPC classification number. This embodiment proposes a "multi-dimensional feature vector" designed to comprehensively capture the core technology, legal attributes, and market value of a patent. This is achieved by extracting deep, multi-dimensional features of the patent and fusing them into a unique and comprehensive multi-dimensional feature vector.

[0158] This multidimensional feature vector has powerful application value, containing deep semantics. Therefore, by calculating the cosine similarity of the multidimensional feature vector, it can find patents that are textually unrelated but technically extremely similar (cross-language and cross-expression similarity). Since the multidimensional feature vector includes legal and commercial characteristics, it can also be used to predict patent value. Through this multidimensional feature vector, the system transforms unstructured documents into calculable and measurable digital assets. Searches using this multidimensional feature vector can not only find similar texts but also high-quality patents with comparable technical strength and innovation levels. Cluster analysis using this multidimensional feature vector can intuitively identify core technological barriers and hot topics within an industry.

[0159] In one specific embodiment, the method of extracting multidimensional features from the current patent document using natural language processing technology and encoding the multidimensional features into a multidimensional feature vector further includes: The image encoder is used to extract the feature vector of the attached image, and the formula parser is used to extract the structural feature vector of the formula. The feature vectors in the attached figures, the structure vectors of the formulas, and the feature vectors in the text are concatenated or weighted and fused using an attention mechanism to obtain a multimodal fusion feature vector. The feature vectors of core technical terms, weight vectors, innovation point vectors, citation feature vectors, legal feature vectors, and multimodal fusion feature vectors are combined to generate multidimensional feature vectors for patents.

[0160] In one embodiment, the patent document adaptive processing method further includes an incremental processing step: Monitor the pending catalog. If a new patent document is detected, query the database to see if there are any identical patents with the same patent unique identifier or similar patents with similar content. If it does not exist, the newly added patent document will be used as the current patent document, and the complete processing flow including steps S100 to S500 will be executed. If they exist, compare the document content or hash values ​​of the old and new versions, use a large language model to analyze the document differences, and identify the type and extent of the changes. If the change is a critical change, the entire processing flow, including steps S100 to S500, is re-executed; if the change is a minor change, only the affected text blocks or features are updated; if the change is an irrelevant change, the processing is skipped.

[0161] Specifically, patent data is dynamic; legal status is updated, patent documents may be corrected (such as correcting erroneous bibliographic entries), and even the patent documents themselves may be modified during the granting process. Reprocessing the entire database every time data changes would be computationally costly. Therefore, this embodiment designs a sophisticated incremental update mechanism.

[0162] For incremental processing, the system detects new data through file monitoring, for example, by using a file monitoring service (such as Watchdog) to monitor the directory to be processed in real time. When a new file or a change in file hash value is detected, the incremental processing flow is triggered. The system queries the RDBMS based on the file name or the application number parsed internally. If it is a new patent, it directly enters the full processing flow (S100-S500).

[0163] If the update involves an existing patent (i.e., an update to old data), the system proceeds to the difference analysis stage. The old and new versions of the text are aligned, and an LLM comparison is used to determine whether it is a "critical change" (such as claims amendments, abstract changes, etc.) or a "minor change" (such as typos, formatting adjustments, or non-critical text corrections). Depending on the situation, a full reprocessing or incremental update is performed to conserve computational resources.

[0164] This is not just string-level diffing (like Linux's diff tool), but semantic-level diffing analysis using a large language model.

[0165] The Large Language Model (LLM) was fed into two versions of patent documents, old and new, and was asked to analyze them. Where is the change located? What is the nature of the change? (A. Modification of key technical terms; B. Addition or deletion of the scope of the claims; C. Correction of typos; D. Format adjustment)

[0166] Importance level of the change (high / medium / low).

[0167] Based on the LLM analysis results: If the change is critical (such as amendments to the claims or changes to the technical solution), the system determines that the core technology of the patent has changed and the original vector representation has become invalid. The system marks the patent as "to be recalculated," inserts it into a high-priority queue, re-executes the entire process from S100 to S500, and updates the version number in the RDBMS.

[0168] For minor changes (such as address changes or corrections of spelling errors by the applicant), these changes do not affect the technical semantics. The system only updates the corresponding metadata fields in the RDBMS without triggering the vectorization process, which greatly saves GPU resources.

[0169] If the change is irrelevant (such as adding or removing a page watermark), the system will skip the process.

[0170] The core solution of this embodiment is as follows: monitor the pending processing directory, and when a new or modified patent document is detected, use a hash algorithm to quickly determine whether it has been processed. If it is a new document, execute the entire S100-S500 process; if it is an update to an existing document, use LLM to perform difference analysis and identify the type of change (such as claim amendments, specification errata, etc.). If it is a change to a core field, trigger a full reprocessing; if it is a change to a non-core field, only update the affected vectors or features.

[0171] In one embodiment, the patent document adaptive processing method further includes multimodal content extraction: Use image recognition models to extract patent figures or design images from current patent documents and generate semantic description vectors for the patent figures or design images; The LaTeX mathematical formula converter is used to convert the mathematical formulas in the current patent document into standard mathematical markup language, or the mathematical formula parser is used to extract the symbol sequence of the mathematical formulas in the current patent document and the tree structure representation therein; the obtained standard mathematical markup language or symbol sequence is then vectorized. The extracted patent drawings or design images and the parsed formula files are stored as multimodal objects in the object storage, and an index association is established between them and other storage. The formula files include mathematical markup language or symbol sequences of mathematical formulas.

[0172] In one embodiment, the patent document adaptive processing method further includes an adaptive quality control step: After processing the current patent document, the results are checked for completeness, consistency, and technical rationality. If quality issues are found during the inspection, the corresponding reprocessing strategy is triggered according to the issue type, including using advanced extraction rules, performing cross-validation, or calling domain expert models for correction.

[0173] Specifically, to ensure data quality, the system introduces an adaptive quality control mechanism.

[0174] After processing, the system will perform integrity checks (such as whether the title and summary are empty), consistency checks (such as whether the metadata and text content conflict), and rationality checks (such as whether the technical description is logically consistent).

[0175] Integrity check: Automatically checks whether all required fields are empty. For example, it checks whether the patent title, abstract, IPC classification number, and generated vectors have all been successfully generated and are not empty. If any missing fields are found, the system determines it as "integrity failure".

[0176] Consistency checks: For example, check if the "patent type" extracted by the LLM conflicts with the terms "utility model," "invention," or "design" appearing in the document. Check if the extracted "applicant" metadata is consistent with that recorded in the document text. Check if the timestamp of the text block is later than the document's application date (logical reasonableness).

[0177] Reasonableness check: The system calls upon the LLM (Limited Language Management) to conduct a logical review of the extracted technical features and generated abstract. For example, it might ask the LLM: "Based on the extracted content, is the technical solution of this patent logically consistent? Are there any obvious contradictions?" If a problem is detected, the system will automatically trigger a reprocessing process. For example, if "missing key information" is detected, a more advanced extraction model (such as an LLM with a larger number of parameters) will be called to re-extract the information; if "inconsistent internal information" is detected, cross-validation will be performed.

[0178] Adaptive quality control: After processing, the system automatically performs integrity, consistency, and reasonableness checks. For example, it uses vector retrieval results for backtracking verification to check whether the generated vectors can find their own legal family patents in similarity searches. If quality anomalies are detected (such as abnormal vector normalization, or mismatch between classification numbers and text content), a reprocessing strategy is triggered: automatically adjusting the prompt, calling an expert model with a larger number of parameters, or indicating manual intervention.

[0179] After processing is complete, integrity and consistency checks can be performed on the results. If an anomaly is detected, a reprocessing strategy is triggered.

[0180] In addition to incremental updates, adaptive quality control is crucial for ensuring the data lifeline. After each processing step (whether full or incremental), the system initiates a QC check.

[0181] QC checks include: Integrity check: Check whether required fields (such as ID, title) are empty and whether the vector dimensions meet expectations.

[0182] Consistency checks: Logical rules are used to determine if data is contradictory. For example, the application date cannot be earlier than the priority date, and the IPC classification number should roughly match the technical field in the title (verified using a simple classifier). For example, if the IPC classification is "textiles," but the title and abstract are about "semiconductors," a warning will be issued.

[0183] Reasonableness check (based on vector backtracking): This is an advanced verification method. The system uses the multi-dimensional feature vectors of the generated patents to retrieve the top 5 most similar patents in the vector database. If the search results contain patents from the same family as the original patent, or highly related technology patents, it indicates that the vector quality is good. If the search results are completely irrelevant, or completely unrelated categories appear, it indicates that there may be a serious bias in the vectorization process (such as model collapse or prompt error).

[0184] If the QC check fails, the system will trigger an adaptive correction process. Different remedial measures will be automatically taken depending on the error type: If the error type is extraction failure (e.g., field is empty), the system will automatically try to replace the Prompt or call a more powerful LLM model (e.g., switch from GPT-3.5 to GPT-4) to retry.

[0185] If the error type is poor vector quality, the system checks whether the blocks are too large, tries to adjust the block parameters, and then re-blocks and vectorizes.

[0186] If the error type is a logical contradiction, the system marks the patent document as "manual review" and pushes it to the management backend for expert processing.

[0187] This closed-loop mechanism transforms the system from a one-way data pipeline into an intelligent system with perception, judgment, and self-healing capabilities. Over time, the system accumulates more and more error patterns and correction strategies, continuously improving its accuracy and robustness, truly achieving intelligent and automated data governance.

[0188] This embodiment enables self-evolution, incremental updates, and quality control, ensuring the freshness and high accuracy of data assets.

[0189] This application, based on deep semantic understanding and multimodal fusion, breaks through the bottleneck of information extraction. It abandons the rigid, rule-based extraction model of traditional methods and utilizes a multimodal large language model for deep analysis of patent documents. In particular, it introduces the ability to process attached figures and mathematical formulas, enabling the identification of structured information that cannot be covered by text. Through the fusion of multimodal features, the representation of patent data becomes more comprehensive, solving the information loss problem caused by traditional technologies that only process text, and laying the foundation for advanced applications such as "patent search by image" and "formula retrieval."

[0190] Adaptive segmentation based on graph neural networks ensures semantic integrity and coherence. Unlike traditional fixed-size segmentation, this application utilizes GNNs to model the semantic structure of the text, accurately identifying the boundaries of technical blocks and the independent semantic units of the claims. This content-aware adaptive segmentation strategy ensures the relative semantic independence and integrity of each text block, avoiding the mechanical truncation of technical concepts, greatly improving the purity and accuracy of vectorized representations, and thus significantly enhancing the precision of semantic retrieval.

[0191] The dynamic model selection mechanism achieves a balance between processing speed and quality. This application innovatively introduces the concept of "technical complexity scoring" and dynamically selects the optimal model from multiple candidate models based on the specific characteristics of the document (language, domain, formula, complexity). High-precision models are used for complex documents, while efficient models are used for routine documents. This ensures both the processing quality of key patents and efficiency in large-scale data processing, solving the problem that a single model cannot adapt to diverse scenarios.

[0192] By constructing multi-dimensional feature vectors, a quantitative representation of deep features is achieved. This application not only extracts semantic vectors from the text but also generates multi-dimensional feature vectors containing technical keywords, innovative points, citation networks, legal stability, and commercial value through intelligent feature extraction. These deep feature vectors enable computers to understand and evaluate patents from multiple dimensions, just like human experts, providing strong data support for the automatic discovery of high-value patents, infringement risk analysis, and technology route prediction.

[0193] A closed-loop quality control and incremental update mechanism enhances the system's robustness and maintainability. The system's built-in adaptive quality control and incremental processing mechanisms enable it to self-detect errors and adapt to dynamic data changes. Through discrepancy analysis, the system can intelligently handle changes in the legal status of patents or text corrections, avoiding redundant calculations and saving computational resources. Simultaneously, the quality feedback loop continuously optimizes extraction strategies, ensuring high-quality data asset transfer.

[0194] This application also provides a patent document adaptive processing apparatus, which includes: Parsing module 100 is used to parse the current patent document using a large language model to obtain structured data; The segmentation module 200 is used to identify the semantic boundaries and technical paragraphs of the current patent document based on the structured data, and execute the corresponding segmentation strategy according to the preset decision logic to divide the current patent document into one or more text blocks; The semantic vector generation module 300 is used to extract the first feature of the current patent document and text block, wherein the first feature includes the language type, technical field, technical complexity of the text block and special elements of the patent, and dynamically calls the corresponding text embedding model according to the first feature to generate the semantic vector of the text block. The multidimensional feature extraction module 400 is used to extract multidimensional features of the current patent document using natural language processing technology and encode them into a multidimensional feature vector. The storage module 500 is used to store the patent metadata in the structured data to a relational database, store the semantic vector and multidimensional feature vector to a vector database, store the current patent document to an object storage, and establish index associations between the storages.

[0195] Since the embodiments of the device part correspond to the embodiments of the above method, the description of the patent document adaptive processing device provided in this application is based on the above method embodiments. This application will not repeat the description here, as it has the same beneficial effects as the above patent document adaptive processing method.

[0196] Figure 3This is a schematic diagram of the structure of a computer device 5000 provided in an embodiment of this application. The computer device 5000 can vary significantly due to different configurations or performance. It may include one or more central processing units (CPUs) 510 (e.g., one or more processors) and a memory 520, and one or more storage media 530 (e.g., one or more mass storage devices) for storing application programs 533 or data 532. The memory 520 and storage media 530 can be temporary or persistent storage. The program stored in the storage media 530 may include one or more modules (not shown in the diagram), each module may include a series of instruction operations on the computer device 5000. Furthermore, the processor 510 may be configured to communicate with the storage media 530 and execute the series of instruction operations in the storage media 530 on the computer device 5000.

[0197] Computer device 5000 may also include one or more power supplies 540, one or more wired or wireless network interfaces 550, one or more input / output interfaces 560, and / or one or more operating systems 531, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 3 The computer device structure shown does not constitute a limitation on the computer device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0198] This application also provides a computer device including a memory and a processor. The memory stores computer-readable instructions, which, when executed by the processor, cause the processor to perform the steps of the patent document adaptive processing method in the above embodiments. This application also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. The computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the steps of the patent document adaptive processing method.

[0199] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0200] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0201] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for adaptive processing of patent documents, characterized in that, The adaptive processing method for patent documents includes: Utilize large language models to parse current patent documents and obtain structured data; Based on the structured data, the semantic boundaries and technical paragraphs of the current patent document are identified, and the corresponding segmentation strategy is executed according to the preset decision logic to divide the current patent document into one or more text blocks; Extract the first feature of the current patent document and text block. The first feature includes language type, technical field, technical complexity of the text block and special elements. Dynamically call the corresponding text embedding model according to the first feature to generate the semantic vector of the text block. The multidimensional features of the current patent document are extracted using natural language processing techniques and encoded into a multidimensional feature vector. The patent metadata in the structured data is stored in a relational database, the semantic vector and multidimensional feature vector are stored in a vector database, and the current patent document is stored in an object storage, establishing index associations between the storages.

2. The patent document adaptive processing method according to claim 1, characterized in that, The process involves identifying the semantic boundaries and technical paragraphs of the current patent document based on the structured data, and executing a corresponding segmentation strategy according to a preset decision logic to divide the current patent document into one or more text blocks, including: Determine whether the patent type is an invention patent. If the patent type is an invention patent, identify the technical block and divide the paragraphs in the current patent document that belong to the same technical feature into a text block. If the claims section is identified as containing several claims or having a hierarchical structure, then the text content corresponding to each claim or each level will be treated as an independent text block. For other identified text content, semantic similarity is used to identify semantic turning points in the text, and the identified other text content is segmented at the semantic turning points.

3. The patent document adaptive processing method according to claim 2, characterized in that, The identification technology block divides paragraphs belonging to the same technical feature in the current patent document into a text block, including: A large language model is used to perform full-text semantic analysis on the current patent document to construct a triplet knowledge graph containing technical issues, technical means, and technical effects. Based on the triplet knowledge graph, the technical logic flow in the current patent document is identified, and the dependency relationship between core technical nodes and supporting descriptive content is determined. Based on the aforementioned dependencies, a semantic aggregation and segmentation strategy is executed to aggregate content scattered across different chapters but pointing to the same technical feature into one or more logically independent text blocks.

4. The patent document adaptive processing method according to claim 3, characterized in that, The step of identifying the technical logic flow in the current patent document based on the triplet knowledge graph, and determining the dependency relationship between core technical nodes and supporting descriptive content, includes: Identify the claim tree structure in the current patent document and map each independent claim and its dependent claims to a core skeleton node in the knowledge graph; For the instruction manual section, a large language model is used to extract the entities and functions corresponding to each paragraph and map them as leaf nodes. Calculate the semantic relevance and text distance between leaf nodes and core skeleton nodes, and aggregate the leaf nodes that reach the preset relevance threshold and have the closest text distance to the corresponding core skeleton node.

5. The patent document adaptive processing method according to claim 2, characterized in that, The identification technology block divides paragraphs belonging to the same technical feature in the current patent document into a text block, including: Based on structured data, text units in the current patent document are constructed into a semantic graph, where text units are sentences or paragraphs, nodes in the semantic graph represent text units, and the weights of the edges represent the semantic similarity or reference relationships between text units. By using graph neural networks to traverse semantic graphs and learn features, the local and global semantic coherence of text units are calculated and evaluated, and semantic inflection points are determined. Based on the output of the graph neural network, a segmentation strategy is selected to divide the current patent document into one or more semantically complete text blocks.

6. The patent document adaptive processing method according to any one of claims 1-5, characterized in that, The step of dynamically invoking the corresponding text embedding model based on the first feature specifically includes: If the language of the current patent document is detected to be Chinese and the technical field belongs to a cutting-edge interdisciplinary field, or if the text block contains special elements, or if the technical complexity is higher than the preset threshold, then the first model is selected. If a text block is detected to contain an attached diagram or structural diagram, and / or if cross-modal detection is required or if the user's retrieval needs include text search based on images, then the third model is selected. If the text block is another regular technical document, then select the second model; The first model is suitable for processing complex technical documents and documents containing special elements; the second model is suitable for general scenarios and has a faster processing speed; and the third model is suitable for multimodal data processing.

7. The patent document adaptive processing method according to any one of claims 1-5, characterized in that, The step of extracting multidimensional features from the current patent document using natural language processing technology and encoding them into a multidimensional feature vector includes: The core technical terms are extracted from the current patent document using a large language model. Based on the distribution of the core technical terms in the current patent document, the weight of each core technical term is calculated. The core technical terms and weights are vectorized to generate feature vectors and weight vectors of the core technical terms. The distribution includes the frequency and position of occurrence. Based on a pre-built innovative vocabulary, innovative descriptive fragments in the current patent document are identified through semantic analysis, and the innovative descriptive fragments are vectorized to generate innovative point vectors. Based on the citation data of the current patent document, a citation index is calculated, and the citation index is vectorized to generate a citation feature vector. The citation index includes at least one of citation strength, the number of times it is cited by the core patent, and the number of patent families. Extract the legal and commercial indicators of the current patent document, and vectorize the legal and commercial indicators to generate a legal feature vector. The legal and commercial features include at least one of the following: classification number, number of claims, hierarchical depth of dependent claims, legal status, and remaining patent life. The feature vectors of the core technical terms, weight vectors, innovation point vectors, citation feature vectors, and legal feature vectors are combined to generate multidimensional feature vectors.

8. The patent document adaptive processing method according to any one of claims 1-5, characterized in that, The adaptive patent document processing method further includes an incremental processing step: Monitor the pending catalog. If a new patent document is detected, query the database to see if there are any identical patents with the same patent unique identifier or similar patents with similar content. If it does not exist, the newly added patent document will be used as the current patent document, and the current patent document will be parsed using a large language model to obtain structured data and subsequent steps. If they exist, compare the document content or hash values ​​of the old and new versions, use a large language model to analyze the document differences, and identify the type and extent of the changes. If the change is critical, the process of parsing the current patent document using a large language model to obtain structured data and subsequent steps will be repeated; if the change is minor, only the affected text blocks or features will be updated. If the change is considered irrelevant, then skip the processing.

9. A patent document adaptive processing device, characterized in that, The patent document adaptive processing device includes: The parsing module is used to parse the current patent document using a large language model to obtain structured data; The segmentation module is used to identify the semantic boundaries and technical paragraphs of the current patent document based on the structured data, and execute the corresponding segmentation strategy according to the preset decision logic to divide the current patent document into one or more text blocks; The semantic vector generation module is used to extract the first feature of the current patent document and text block. The first feature includes language type, technical field, technical complexity of text block and special elements. Based on the first feature, the corresponding text embedding model is dynamically called to generate the semantic vector of the text block. The multidimensional feature extraction module is used to extract multidimensional features of the current patent document using natural language processing technology and encode them into multidimensional feature vectors; The storage module is used to store the patent metadata in the structured data to a relational database, store the semantic vector and multidimensional feature vector to a vector database, store the current patent document to an object storage, and establish index associations between the storages.

10. A computer device, characterized in that, The computer device includes: a memory and at least one processor, wherein the memory stores instructions; The at least one processor invokes the instructions in the memory to cause the computer device to perform the patent document adaptive processing method as described in any one of claims 1-8.

Images