A patent matching method based on large model and knowledge graph fusion

Abstract

The application discloses a patent matching method based on large model and knowledge graph fusion, comprising the following steps: obtaining a semantic vector by a pre-trained large model according to collected patent text data, obtaining a graph embedding vector by a pre-trained knowledge graph model, and fusing the semantic vector and the graph embedding vector into a joint representation vector; further comprising: obtaining a semantic vector by a pre-trained large model according to collected user demand data, obtaining a graph embedding vector by a pre-trained knowledge graph model, and obtaining a joint representation vector; performing patent preliminary screening according to the vector similarity between the joint representation vector of the user demand data and the patent text data, obtaining a correlation score by comparing the preliminary screened patent text data and the user demand data, fusing the vector similarity through the correlation score, and obtaining a patent recommendation sequence. The core problem of poor matching precision caused by shallow semantic understanding and insufficient knowledge utilization is solved.

Description

Technical Field

[0001] This invention belongs to the field of patent matching technology, specifically relating to a patent matching method based on the fusion of large models and knowledge graphs. Background Technology

[0002] Against the backdrop of global economic integration and the rapid development of the knowledge economy, intellectual property has become a strategic resource for national development and a core element of international competitiveness. With the completion of the inventory of existing patents by universities and research institutions in various provinces and cities, and the year-on-year increase in patent applications, enterprises are increasingly able to quickly and accurately find patents that meet their own technological development needs from a vast amount of patent information, which has become a key link in promoting technological innovation and industrial upgrading.

[0003] To address this challenge, patent matching technology has evolved from early simple keyword-based searches to iterations utilizing traditional natural language processing and machine learning methods. For example, Chinese patent CN118861437A discloses an intelligent matching system for enterprise technology needs and intellectual property information. This system uses natural language processing techniques such as word vector generation and named entity recognition to analyze text, and further utilizes machine learning algorithms such as K-means clustering and SVM classification to classify and match technology needs and patents. Such systems have improved the automation level of the matching process to a certain extent.

[0004] However, these existing technologies still have significant limitations, specifically: First, traditional word vector models have limited semantic representation capabilities, making it difficult to handle complex contextual relationships, polysemy of technical terms, and long-distance dependencies in patent texts, resulting in insufficient depth and accuracy of semantic understanding; Second, although they can identify entities in the text, they lack a structured understanding and application of the deep technical logic and relationships between entities, causing matching to remain at the level of surface text similarity and failing to reach the core technical connections; Finally, the aforementioned feature extraction and matching models are often isolated, failing to achieve effective synergy and integration of deep semantic information and structured knowledge information, which makes it difficult for the accuracy and reliability of the final matching results to meet the high standards of enterprises, resulting in a high rate of invalid pushes.

[0005] Therefore, how to achieve deep, cross-modal semantic understanding and information fusion from massive and complex patent texts and enterprise requirement descriptions, so as to complete high-precision and highly relevant intelligent matching, has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] This invention provides a patent matching method based on the fusion of large models and knowledge graphs. By using deep semantic understanding driven by large models and structured knowledge from knowledge graphs for collaborative fusion matching, it solves the problems of low accuracy and reliability in patent matching caused by shallow semantic understanding, weak knowledge association, and insufficient information fusion.

[0007] The technical solution adopted in this invention is as follows: A patent matching method based on the fusion of large models and knowledge graphs includes: Based on the collected patent text data, semantic vectors are obtained through a pre-trained large model, and graph embedding vectors are obtained through a pre-trained knowledge graph model. The semantic vectors and graph embedding vectors are then fused into a joint representation vector. The method further includes: Based on the collected user demand data, semantic vectors are obtained through a pre-trained large model, and graph embedding vectors are obtained through a pre-trained knowledge graph model, in order to obtain joint representation vectors. Patents are initially screened based on the joint representation vector of the user demand data and the vector similarity of the patent text data. A relevance score is obtained by comparing the initially screened patent text data with the user demand data. The vector similarity is then fused using the relevance score to obtain a patent recommendation sequence.

[0008] The method disclosed in this invention also has the following additional technical features: After collecting the patent text data, the following is also included: Based on the collected patent text data, preprocessing and cleaning are performed, and the classification number is used as the fragmentation key to fragment and store the cleaned patent text data.

[0009] The pre-training of large models specifically involves: A general large language model is used as the base model. Based on the fragmented patent text data, a masked language model is trained to identify named entities, extract task relationships, and the fragmented patent text data is fragmented, adjusted, and labeled. Based on the annotated patent text data, a large model is trained to obtain semantic vectors.

[0010] The pre-training of knowledge graph models involves the following steps: Based on the annotated patent text data, triples containing head entities, relations, and tail entities are generated to construct a knowledge graph. The knowledge graph is embedded using a graph neural network to generate a graph embedding vector for each entity node in the graph, thus obtaining a knowledge graph model.

[0011] The semantic vector and graph embedding vector are fused into a joint representation vector, specifically: Based on semantic vectors and graph embedding vectors, the input is a gated fusion network. The information richness of the semantic vectors and graph embedding vectors is compared through network layers, and the fusion weights are obtained through an activation function. When the information richness of the semantic vector is greater than that of the graph embedding vector, the fusion weights are positively adjusted. When the information richness of the semantic vector is less than that of the graph embedding vector, the fusion weight is negatively adjusted; Based on the fusion weights, the semantic vectors are weighted and fused with the graph embedding vectors to obtain a joint representation vector.

[0012] Patent screening is performed based on the vector similarity between the joint representation vector of the user demand data and the vector similarity of the patent text data, specifically as follows: Based on the joint representation vector of the user demand data, the inner product similarity and Euclidean distance of the joint representation vector of the patent text data are obtained. The vector similarity is obtained by combining the inner product similarity and the Euclidean distance. Based on the vector similarity, the patent text data are sorted from largest to smallest to obtain a preliminary list containing multiple patent text data.

[0013] The preliminary screening list is as follows: The number of patent text data in the initial screening list is determined based on the information richness of the joint representation vector of the user demand data, and is negatively correlated with the information richness of the joint representation vector of the user demand data.

[0014] A relevance score is obtained by comparing the initially screened patent text data with the user demand data. This relevance score is then used to fuse vector similarity data. Specifically: Based on the user's request text, it is paired with each patent text data in the initial screening list to form multiple text pairs; Based on the multiple texts, a relevance score is obtained for each text pair by using a pre-trained large model operating in cross-encoder mode. The relevance score is weighted and fused with the corresponding vector similarity score to obtain a comprehensive score. The weight of the vector similarity score is determined based on the information richness of the joint representation vector of the user demand data and is positively correlated with the information richness of the joint representation vector of the user demand data.

[0015] The method also includes: Regularly collect newly generated patent text data and user feedback data on recommendation results; The knowledge graph model is incrementally updated using the newly generated patent text data. The patent semantic understanding model and the knowledge graph model are adjusted using the feedback behavior data.

[0016] The present invention also provides a computer-readable storage medium having a computer program stored thereon. The program implements the method when executed by the processor.

[0017] Due to the adoption of the above technical solution, the beneficial effects achieved by this invention are as follows: 1. In this invention, by using a pre-trained large-scale model and a knowledge graph model in parallel, two types of key information from patent text and enterprise requirements can be captured simultaneously: deep contextual semantic information provided by the large-scale model (solving problems of polysemy and complex sentence structure comprehension) and structured knowledge association information provided by the knowledge graph (revealing the logical relationships between technical terms and entities). Merging these two into a joint representation vector ensures that the final text representation not only contains rich semantics but also embeds domain knowledge, thereby fundamentally improving the system's depth and breadth of understanding of patents and technical requirements.

[0018] Furthermore, a two-stage matching strategy was employed. In the first stage, the joint representation vector was used for initial screening based on vector similarity, ensuring the feasibility of rapid retrieval within a large-scale patent database. In the second stage, the initial screening results were compared again with the user's demand text using a large model to obtain a relevance score, which was then used to fuse and correct the vector similarity results from the initial screening. This coarse-screening + fine-ranking matching mechanism effectively overcomes the potential biases of single-vector matching, comprehensively utilizing the efficiency advantages of semantic space and the accuracy advantages of deep semantic interaction, thereby significantly improving the overall accuracy and reliability of the final patent recommendation sequence.

[0019] Furthermore, the use of pre-trained models reduces the reliance on large-scale labeled data for specific tasks and enhances the model's generalization ability. Simultaneously, the unified representation and matching of patents and user needs enables the system to not only process structured patent data but also effectively understand unstructured, natural language-described enterprise technical needs, broadening the system's application scenarios and enhancing its practicality.

[0020] In summary, by using a dual-engine approach of large model and knowledge graph, and a two-stage matching process of initial vector screening and fine-tuning, the core problems of poor matching accuracy caused by shallow semantic understanding and insufficient knowledge utilization in existing technologies are addressed in a coordinated manner. Ultimately, this approach improves the depth of semantic understanding and enhances the accuracy and reliability of matching results. Attached Figure Description

[0021] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart of a patent matching method based on the fusion of large models and knowledge graphs according to one embodiment of the present invention. Detailed Implementation

[0022] To more clearly illustrate the overall concept of the present invention, a detailed description will be provided below with reference to the accompanying drawings and examples.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0024] like Figure 1 As shown, a patent matching method based on the fusion of large models and knowledge graphs includes: S000: Based on the collected patent text data, semantic vectors are obtained through a pre-trained large model, graph embedding vectors are obtained through a pre-trained knowledge graph model, and semantic vectors and graph embedding vectors are fused into a joint representation vector.

[0025] This step aims to address two core shortcomings of traditional patent matching systems: superficial semantic understanding and a lack of knowledge association. By constructing a novel patent text representation method, a solid foundation is laid for subsequent high-precision matching. Its core objective is to overcome the limitations of traditional text vector representation methods and create a joint representation vector that simultaneously contains deep semantic information and structured knowledge information, thereby achieving a deep understanding and comprehensive representation of the patent's technical content.

[0026] Deep semantic vector extraction employs a large-scale language model pre-trained with domain-adaptive technology as its semantic understanding engine. Building upon its general language understanding capabilities, the model undergoes further pre-training using massive amounts of patent texts and fine-tuning through multiple tasks such as patent classification and entity recognition, enabling it to deeply master the specialized terminology, legal context, and technical expression logic within the patent field.

[0027] The preprocessed patent text data is input into this model, and its deep Transformer encoder generates a high-dimensional, dense, context-aware semantic vector. This vector can effectively represent the complex technical logic, innovative details, and long-distance semantic dependencies in the patent text.

[0028] Structured knowledge embedding and extraction are used to construct a knowledge graph covering the patent technology field, and graph neural networks are used to learn its structural information. The knowledge graph stores technical entities (such as materials, methods, and devices) and their relationships (such as improvement, application, and inclusion) in the form of triples, forming a structured knowledge network.

[0029] For the same patent text, the technical entities involved are identified, and the graph embedding vectors corresponding to these entities are obtained from a graph neural network model. By performing pooling operations on these entity vectors, a comprehensive graph embedding vector representing the overall technical knowledge background of the patent is generated. This vector captures the position and association of the patent technology within the domain knowledge network.

[0030] To achieve adaptive fusion of multi-source information, a gated fusion network is designed to dynamically integrate the two aforementioned vectors. The deep semantic vector and the integrated graph embedding vector are jointly input into the gated fusion network. This network analyzes the information content and complementarity of the two vectors through a small neural network, automatically calculating and outputting a dynamic fusion weight between 0 and 1. Finally, a joint representation vector is generated through a weighted summation. This mechanism can adaptively determine whether semantic information or knowledge graph information should be relied upon more in a specific patent text, thereby achieving optimal fusion.

[0031] In the highly specialized scenario of patent matching, relying solely on textual semantics can easily overlook the inherent connections between technical terms, while knowledge graphs alone cannot comprehend the complex logic of technical descriptions. Using the two in isolation cannot achieve optimal matching results. Therefore, deep fusion representation at the very beginning of the process is a prerequisite for overcoming subsequent matching accuracy bottlenecks.

[0032] The joint representation vector simultaneously encodes the contextual semantics and domain knowledge structure of the text, resulting in a more accurate understanding of technical terms. This effectively overcomes the ambiguity and polysemy inherent in traditional word vector models, enhancing the depth and robustness of semantic understanding. A more comprehensive patent feature representation is constructed, providing a richer and more comprehensive comparative foundation for subsequent similarity calculations. This fundamentally improves the quality of the raw data used for retrieval and ranking, serving as the cornerstone for the system to achieve high-precision matching.

[0033] The method further includes: S100: Based on the collected user demand data, semantic vectors are obtained through a pre-trained large model, and graph embedding vectors are obtained through a pre-trained knowledge graph model, in order to obtain a joint representation vector.

[0034] The core objective of this step is to construct a standardized vector representation of user technical requirements, ensuring that it resides in the same semantic space as the patent text representation, thus establishing a basis for equivalent comparison in subsequent accurate matching. By employing a feature engineering process entirely consistent with patent text processing, unstructured and potentially ambiguous user requirements are transformed into standardized vectors rich in semantic information and knowledge structure, thereby eliminating the representational discrepancies between user requirements and patent documents.

[0035] The deep semantic parsing of the requirement text utilizes the same domain-adapted large model used for processing patent texts, ensuring consistency in semantic understanding standards. The user-submitted natural language requirement description (e.g., "We need a high-efficiency, low-power battery management chip") is input into the large model. Leveraging its expertise learned in the patent domain, the model deeply understands the technical context, performance metrics (high efficiency, low power consumption), and application scenario (battery management chip) within the requirement, and outputs a deep semantic vector with the same dimension and distribution as the patent's semantic vector.

[0036] The technical entity association mining of the demand content utilizes the same knowledge graph model used for constructing the patent graph to identify and associate technical entities within the demand. It automatically identifies key entities mentioned in the user demand text, such as technical components (e.g., chips), technical methods (e.g., management), and performance parameters (e.g., low power consumption). Subsequently, it queries the pre-trained graph embedding library for the corresponding graph embedding vectors of these entities and generates a demand graph embedding vector that comprehensively reflects the technical background of the demand through pooling operations.

[0037] The collaborative fusion of demand representations employs the same gated fusion network structure and parameters as the patent-side network for vector fusion. The resulting deep semantic vector of the demand and the demand graph embedding vector are input into the aforementioned fixed gated fusion network. Based on its learned fusion rules of patent domain knowledge, this network dynamically calculates the fusion weights and ultimately outputs a joint representation vector of the user demand. This process ensures that even if the user demand description is brief or colloquial, its representation maintains spatial consistency with the detailed patent text representation at the algorithmic level.

[0038] If user requirements and patent data use different representation systems, invalid comparisons will occur. This step, through a symmetry processing flow, forces user requirements to be mapped into a precise vector space constructed from patent data, which is an absolute prerequisite for achieving any meaningful automated matching calculation.

[0039] Through a fully consistent processing pipeline, user requirements and a large number of patents are projected into the same joint vector space, making similarity comparison based on vector operations scientific and effective, ensuring the comparability of matched parties. Even if the user requirement description is incomplete or not professional enough, the injection of knowledge graphs can infer its potential technical orientation with the help of domain knowledge, improving the ability to interpret ambiguous requirements and enhancing the robustness and practicality of the system.

[0040] S200: Perform initial patent screening based on the joint representation vector of the user demand data and the vector similarity of the patent text data. Obtain a relevance score by comparing the initially screened patent text data with the user demand data. Then, fuse the vector similarity using the relevance score to obtain a patent recommendation sequence.

[0041] The core objective of this step is to significantly improve the final accuracy and reliability of patent recommendations through a two-stage hybrid matching mechanism, while ensuring system response efficiency. Its design aims to overcome the limitations of a single matching strategy: avoiding semantic biases that may arise from relying solely on vector similarity, and circumventing the computational impracticality of directly using deep learning models for full-database matching. Through synergistic optimization of efficiency and accuracy, it achieves the precise selection of patent sequences that best meet user needs from a massive pool of patents.

[0042] This vector-based, efficient initial screening utilizes approximate nearest neighbor search technology to perform rapid retrieval within a pre-built patent vector index. The joint representation vector of the user's needs is used as the query vector, and large-scale vector similarity calculations (using inner product or cosine similarity) are performed between this vector and the joint representation vectors of all patents in the patent database. This process is achieved using a high-efficiency vector index library (such as FAISS), enabling the retrieval of the Top-K most similar patents from millions of patents within milliseconds, forming an initial screening list. The core objective of this stage is to ensure retrieval efficiency and response speed.

[0043] Based on deep semantic interaction-based fine-tuning, patents in the initial screening list are re-ranked using a deep semantic matching model that is more computationally expensive but also more accurate. The user's original demand text is paired with the original text (such as abstracts and claims) of each patent in the initial screening list, forming multiple (demand, patent) text pairs. These text pairs are then fed into a larger model, ultimately outputting a precise relevance score. This score reflects the deep technical connection between the two texts more effectively than vector similarity.

[0044] The vector similarity score from the first stage is combined with the deep relevance score from the second stage to generate a comprehensive score for each patent. Finally, the initial screening list is reordered based on the comprehensive score to generate a high-quality patent recommendation sequence.

[0045] In real-world patent recommendation scenarios, simply pursuing efficiency or accuracy is not feasible. While vector retrieval is fast, it may miss key patents or introduce ranking biases due to the inherent limitations of the vector space. Conversely, deep semantic matching of the entire patent database, while accurate, incurs enormous computational costs, failing to meet real-time response requirements. Therefore, adopting a hybrid strategy is the inevitable technical choice for achieving a high-precision patent recommendation system suitable for industrial applications.

[0046] This design cleverly allocates computing resources, efficiently processing most irrelevant data while focusing valuable deep learning resources on the most promising candidate set, achieving a balance between throughput and recommendation quality. The final recommendation sequence is generated based on dual validation, and its reliability and technical relevance are far superior to those of single methods, better meeting users' actual decision-making needs and enhancing the reliability and persuasiveness of the recommendation results.

[0047] In a preferred embodiment of the present invention, after acquiring the patent text data, the method further includes: Based on the collected patent text data, preprocessing and cleaning are performed, and the classification number is used as the fragmentation key to fragment and store the cleaned patent text data.

[0048] The core objective of this step is to build a high-quality, highly available data infrastructure for subsequent large-scale data processing and intelligent analysis. Through systematic data cleaning and a technology-based distributed storage strategy, it addresses two fundamental challenges in processing massive amounts of patent data: inconsistent data quality and system scalability bottlenecks. This ensures that complex AI models at higher levels receive a stable and reliable data supply.

[0049] Multi-dimensional data cleaning and standardization establishes an automated cleaning rule base tailored to the characteristics of patent data. Structured deconstruction automatically identifies and separates different components of patent documents, including the title, abstract, claims, description, figures, and legal status metadata, laying the foundation for subsequent differentiated processing.

[0050] Text cleansing employs a combination of rule-based and machine learning methods to remove noisy data such as HTML / XML tags, special characters, and garbled text. Terminology standardization involves building a thesaurus and terminology database for the patent field to standardize technical terms that have the same meaning but differ in expression (e.g., standardizing GPU, graphics processor, and graphics processing unit as standard terms). Quality verification sets integrity verification rules to automatically mark and remove invalid data records with missing key fields (such as abstracts and claims) or severely formatted errors.

[0051] This intelligent sharded storage system, based on the technology field, employs distributed database technology and innovatively uses the patent classification system as the core basis for data distribution. The IPC International Classification number or CPC Joint Classification number of the patent is selected as the sharding key. These classification numbers naturally reflect the technical field to which the patent belongs (e.g., H01L Semiconductor Devices, G06F Electrical Digital Data Processing).

[0052] A range-based sharding algorithm is employed to distribute consecutive classification code ranges across different physical server nodes. For example, sharding boundaries R1 to R_M are defined such that patents with classification codes in the range [R1, R2) are stored in shard 1, those in the range [R2, R3) are stored in shard 2, and so on. Based on the primary classification code of each patent, the sharding mapping function `shard_id=f(classification_code)` automatically routes and stores it on the corresponding shard node.

[0053] Patent data is characterized by its massive volume, rapid growth, and clear technological divisions. Using a single database or randomly distributed storage method would lead to chaotic data management, poor query performance, and make subsequent batch processing for specific technological fields extremely difficult. Therefore, meticulous data governance and technology-semantic-based storage planning at the front end of the process are essential architectural requirements for supporting the efficient and stable operation of the entire system.

[0054] When the system needs to perform analysis or matching in a specific technical field, requests can be directly routed to a few shards storing relevant patents for processing, avoiding a full database scan, significantly reducing I / O overhead and computational latency, and greatly improving the efficiency of subsequent processing. Patents stored in the same shard naturally belong to the same or similar technical fields, which greatly facilitates the acquisition of data for directly reading specific domain data from distributed storage and performing domain-adaptive pre-training or fine-tuning of large models.

[0055] As one embodiment of this implementation method, the pre-training of the large model specifically involves: A general large language model is used as the base model. Based on the fragmented patent text data, a masked language model is trained to identify named entities, extract task relationships, and the fragmented patent text data is fragmented, adjusted, and labeled. Based on the annotated patent text data, a large model is trained to obtain semantic vectors.

[0056] This embodiment aims to construct a specialized semantic understanding model that is deeply adapted to the characteristics of the patent field. Its core objective is to transform a general-purpose language model into a domain expert model capable of accurately understanding patent terminology, technical logic, and legal context, laying the foundation for generating high-quality patent semantic vectors. This specialized training process addresses the knowledge gaps and semantic biases inherent in general-purpose models within the highly specialized field of patents.

[0057] Domain knowledge injection and language pattern learning are used to perform domain-adaptive pre-training on the base model using fragmented patent data.

[0058] Data sharding and scheduling: Patent text data is sharded according to technical fields (shards) to ensure the comprehensiveness and representativeness of the training data, based on the needs of the training task. Masked language model training: A masked language model task is used, randomly masking key content such as technical terms and features in the patent text, allowing the model to learn to predict the masked content based on context. A patent text-sensitive masking strategy is employed, assigning higher masking probabilities to key information such as technical terms and numerical parameters. The masking objective function is... , Among them, Represents the set of mask locations. For unmasked context, This is a patent text dataset.

[0059] This process forces the model to learn the specialized vocabulary, technical expressions, and grammatical structures of patent texts. Through training with a large amount of patent corpus, the model masters the unique language patterns of patent documents.

[0060] Multi-task collaborative deep semantic understanding training, based on domain pre-training, further enhances the model's deep understanding capabilities through a multi-task learning framework.

[0061] The named entity recognition task trains the model to identify and label technical entities in patent texts, such as technical terms, technical problems, technical effects, inventors, etc., enhancing the model's sensitivity to technical elements. The relation extraction task trains the model to identify relationships between technical entities, such as solutions, implementations, improvements, etc., enhancing the model's ability to understand the internal logic of technical solutions.

[0062] Data segmentation and annotation: Based on the model's performance in specific technical fields (segments), the distribution of training data is dynamically adjusted, and difficult-to-understand patent texts are highlighted and reinforced during training. Domain language pattern mastery: Through training with a large amount of patent corpus, the model masters the language patterns unique to patent documents. Joint training and optimization: The above multiple tasks share the model's underlying parameters and are trained and optimized synchronously. The multi-task loss function is a weighted summation. , Among them, each This is the task weight, which is dynamically adjusted based on the importance of the task. These are the parameters for patent classification, named entity recognition, and relation extraction tasks, respectively. This ensures that the semantic representation obtained by the model can simultaneously meet multiple patent understanding needs, resulting in a more general and robust patent semantic understanding capability.

[0063] Patent texts are characterized by high specialization, legal rigor, and technical complexity. While general-purpose language models perform well in general domains, they often fail to accurately understand the precise meanings of technical terms and grasp the legal scope of protection for technical solutions when processing patent texts. Therefore, it is essential to transform general-purpose models into patent-domain expert models through specialized domain-adaptive training; otherwise, subsequent semantic vector generation and patent matching will be based on unreliable semantic understanding.

[0064] The specially trained model can accurately understand the technical terms and legal expressions unique to the patent field, and the generated semantic vectors can truly reflect the technical content of the patent, providing a reliable guarantee for accurate matching and significantly improving the accuracy of semantic understanding. Through segmentation adjustment and data annotation mechanisms, the training process is ensured to be targeted and effective, guaranteeing the feasibility of implementing the technical solution.

[0065] As another embodiment of this implementation, the pre-training of the knowledge graph model specifically includes: Based on the annotated patent text data, triples containing head entities, relations, and tail entities are generated to construct a knowledge graph. The knowledge graph is embedded using a graph neural network to generate a graph embedding vector for each entity node in the graph, thus obtaining a knowledge graph model.

[0066] The core objective of this embodiment is to construct a structured knowledge system capable of deeply representing the associations between patent technologies. By transforming unstructured patent text into a machine-understandable and reasonable knowledge network, and learning low-dimensional distributed representations for the technical entities within it, rich structured knowledge features are provided for subsequent semantic matching, compensating for the shortcomings of pure text semantic understanding.

[0067] The automated construction of knowledge triples is based on a large model that has been fine-tuned through multiple tasks, which automatically extracts structured knowledge from patent texts.

[0068] Entity recognition and linking leverages the named entity recognition capabilities of large models to identify entities such as technical components (e.g., transistors), technical methods (e.g., chemical vapor deposition), materials (e.g., graphene), and performance indicators (e.g., conversion efficiency) in patent texts, and links them to existing entity nodes in the knowledge graph or creates new nodes.

[0069] Relation extraction and verification leverage the relation extraction capabilities of a large model to determine the relationship type between identified entity pairs, forming candidate triples. A rule-based and statistical filtering mechanism ensures the reliability of the relationships. The main relation types include: technology hierarchy relationships, technology function relationships, and technology dependency relationships.

[0070] Knowledge fusion and denoising are performed by fusing identical entities from different patents to eliminate representational differences; low-frequency noise relationships are removed through statistical significance testing to ensure the quality of the knowledge graph.

[0071] Graph neural network embedding learning employs a message-passing-based graph neural network to learn the distributed representation of nodes in a knowledge graph.

[0072] Graph structure modeling: modeling knowledge graphs as heterogeneous graphs. ,in For a set of entity nodes, Let be the set of edges. It is a set of relation types.

[0073] Neighborhood information aggregation, for each entity node It aggregates its multi-hop neighborhood information through a multi-layer graph convolutional network. In the first... The update formula for the layer is: , in Represents a node In the Layer representation, Indicates the relationship With nodes Connected neighbor set, The normalization constant is A weight matrix specific to the relationship.

[0074] A multi-relationship graph attention mechanism is introduced, which dynamically learns the importance of different neighbor nodes: , in For relationship Specific attention vectors, This indicates vector concatenation.

[0075] The target design is optimized by adopting a translation model-based approach, and the loss function is designed to encourage positive sample triples to score higher than negative samples. , in This is the triplet scoring function. For the positive sample set, This is a set of negative samples constructed by replacing the head and tail entities. This is the interval hyperparameter.

[0076] The quality of knowledge graph construction directly depends on the entity recognition and relation extraction capabilities acquired by the large model during the multi-task fine-tuning stage. Graph embedding vectors focus on the structured relationships between technical entities, complementing the semantic vectors generated by the large model to jointly constitute a joint representation.

[0077] Based on the fragmentation of patent technology fields, a distributed graph partitioning method is used to partition a large-scale knowledge graph, enabling parallel graph neural network training. An incremental graph learning strategy is adopted, so when new patent data is added, only the affected knowledge subgraphs need to be locally retrained, greatly improving model update efficiency.

[0078] Understanding patent technologies requires not only in-depth textual semantic analysis but also a grasp of the complex networks of relationships between technical entities. Traditional methods, which rely solely on knowledge graphs constructed from textual co-occurrence or simple rules, struggle to accurately capture the unique technical logic and legal context of the patent field. Combining deep language understanding with large-scale models and structured representation learning from graph neural networks is an essential technical path for building a high-quality patent knowledge system.

[0079] By learning graph embedding vectors, the system can discover deep connections such as technological alternatives and technological evolution paths that are difficult to identify with traditional text matching, which is expected to improve the success rate of matching complex technological needs. Recommendation results based on knowledge graphs can trace the matched technological paths, enhancing the system's credibility and practicality.

[0080] This embodiment injects structured technical understanding capabilities into the patent matching system through systematic knowledge graph construction and embedding learning, forming a powerful technical synergy with the semantic understanding module, and jointly improving the overall performance and application value of the system.

[0081] In a preferred embodiment of the present invention, the semantic vector and the graph embedding vector are fused into a joint representation vector, specifically as follows: Based on semantic vectors and graph embedding vectors, the input is a gated fusion network. The information richness of the semantic vectors and graph embedding vectors is compared through network layers, and the fusion weights are obtained through an activation function. When the information richness of the semantic vector is greater than that of the graph embedding vector, the fusion weights are positively adjusted. When the information richness of the semantic vector is less than that of the graph embedding vector, the fusion weight is negatively adjusted; Based on the fusion weights, the semantic vectors are weighted and fused with the graph embedding vectors to obtain a joint representation vector.

[0082] The core objective of this embodiment is to construct an adaptive, dynamically balanced multi-source information fusion mechanism to address the problem of uneven distribution of semantic and knowledge information in different patent texts. By intelligently evaluating the relative information value of the two vectors, optimal feature fusion is achieved, thereby generating a joint representation vector that accurately reflects the essence of the patent technology.

[0083] Information richness assessment and weight generation: Design a dedicated gating fusion network to dynamically analyze the information content of two vectors.

[0084] Feature concatenation and transformation, transforming semantic vectors And graph embedding vector Input the network after splicing into the evaluation network: , in Represents element-wise product. The vector magnitude is represented by the differential feature, which helps the network better distinguish the information characteristics of the two vectors.

[0085] Learning the relative information value of two vectors using a multilayer perceptron: , , , in, , , These are the weight vectors, , , These are the bias vectors.

[0086] When semantic vector information is richer When the graph embedding vector information is richer .

[0087] Based on the learned fusion weights, accurate vector synthesis is performed. Weighted fusion calculation: , in This is the final generated joint representation vector.

[0088] It should be noted that the joint representation vector also needs to be normalized: , Ensure that the magnitude of the output vector is uniform to facilitate subsequent similarity calculation.

[0089] The technical characteristics of patent texts differ significantly: fundamental patents often contain complex theoretical descriptions and innovative technical logic (rich in semantic information), while improvement patents may involve the combined application of numerous existing technical entities (rich in knowledge information). Using fixed fusion weights cannot accommodate this diversity; only dynamic adaptive fusion can achieve optimal representation results across all types of patents.

[0090] By tailoring a fusion strategy for each patent, the joint representation vector can most accurately reflect its technical essence, which is expected to improve the accuracy of subsequent matching. This embodiment achieves an intelligent balance between semantic and knowledge information through an innovative gating fusion mechanism, providing a more accurate and robust feature representation foundation for the patent matching system.

[0091] In a preferred embodiment of the present invention, patent screening is performed based on the vector similarity between the joint representation vector of the user demand data and the vector similarity of the patent text data. Specifically: Based on the joint representation vector of the user demand data, the inner product similarity and Euclidean distance of the joint representation vector of the patent text data are obtained. The vector similarity is obtained by combining the inner product similarity and the Euclidean distance. Based on the vector similarity, the patent text data are sorted from largest to smallest to obtain a preliminary list containing multiple patent text data.

[0092] The core objective of this implementation method is to achieve rapid and comprehensive screening of potentially relevant patents from a massive patent database while ensuring retrieval efficiency. By designing a multi-faceted and complementary similarity measurement strategy, the potential biases of a single similarity indicator are overcome, ensuring that the initial screening results have both high recall and reasonable precision, thus providing a high-quality candidate set for the subsequent fine-tuning stage.

[0093] Multi-dimensional similarity calculation employs two similarity measurement methods with complementary characteristics to evaluate the correlation between vectors from different perspectives.

[0094] Inner product similarity calculation captures the correlation between vectors in direction and is sensitive to semantic similarity. , in This represents the query (user demand) vector. Represents the patent vector. For vector dimensions.

[0095] Euclidean distance calculation measures the absolute distance between vectors in space and is sensitive to differences in numerical distribution. .

[0096] Convert the Euclidean distance into a similarity score, making it comparable to the inner product similarity in numerical range. .

[0097] The two similarity scores are combined into a unified evaluation index through weighted fusion.

[0098] , in To balance the weights, the value range is [0,1], which is used to adjust the relative importance of the two similarities.

[0099] Automatically adjust based on the characteristics of the query vector value: , in For the sigmoid function, and These are learnable parameters. When the query vector has a large magnitude (rich in information), it tends to rely on inner product similarity; when the magnitude is small, it relies more on Euclidean distance.

[0100] This paper utilizes an approximate nearest neighbor search technique to achieve fast retrieval of a large-scale vector library. It employs fused similarity as a distance metric to perform an approximate nearest neighbor search within a pre-built vector index. Candidate patents are then sorted in descending order of fused similarity score.

[0101] Inner product similarity is sensitive to vector orientation, making it suitable for capturing semantic relevance, but it may introduce bias when vector magnitudes differ significantly. Euclidean distance is sensitive to differences in absolute position and can identify vectors with similar numerical distributions, but it is not sensitive enough to changes in orientation. Combining the two achieves a balance between semantic relevance and numerical distribution similarity.

[0102] As a preferred embodiment of this implementation, the preliminary screening list is as follows: The number of patent text data in the initial screening list is determined based on the information richness of the joint representation vector of the user demand data, and is negatively correlated with the information richness of the joint representation vector of the user demand data.

[0103] The core objective of this embodiment is to establish an intelligent and adaptive retrieval scope control mechanism. By dynamically adjusting the size of the initial screening list, it achieves optimal allocation of computing resources while ensuring retrieval quality. This design aims to address the risks of over-retrieval or missed detection caused by a fixed number of searches, and can automatically adjust the processing strategy based on the information quality of the query request.

[0104] Information richness metric assessment can be achieved using any one of the following dimensions, or a combination thereof.

[0105] Vector magnitude analysis calculates the L2 norm of the query vector, reflecting the overall information strength; Information entropy calculation analyzes the uniformity of information distribution across all dimensions of a vector. Sparsity evaluation identifies the number of salient features in a vector.

[0106] The optimal number of retrievals is determined by nonlinear mapping based on information richness scores.

[0107] Perform the base mapping. , in, and These represent the minimum and maximum number of searches, respectively. The attenuation coefficient is... For information richness.

[0108] Adaptive adjustments are made, taking into account real-time system load and query history, to further optimize the retrieval scale. , in, and These represent the average response time and the current system load, respectively. This represents the success rate of historical queries.

[0109] It should be noted that the adaptively adjusted retrieval scale must satisfy boundary constraints and rounding. .

[0110] Traditional fixed-quantity retrieval strategies have significant drawbacks: for queries with explicit information, a fixed large quantity leads to a waste of computational resources; for queries with vague information, a fixed small quantity may result in the omission of important patents. This approach cannot adapt to the vast differences in query quality in real-world applications.

[0111] High information richness query vectors have clear representations and distinct semantic focus, requiring only a small search scope to hit the target patent. Low information richness query vectors have vague representations and scattered semantics, requiring a larger search scope to avoid missed detections. This achieves an intelligent balance between search accuracy and efficiency.

[0112] In a preferred embodiment of the present invention, a relevance score is obtained by comparing the initially screened patent text data with the user demand data, and the vector similarity is then fused using the relevance score. Specifically: Based on the user's request text, it is paired with each patent text data in the initial screening list to form multiple text pairs; Based on the multiple texts, a relevance score is obtained for each text pair by using a pre-trained large model operating in cross-encoder mode. The relevance score is weighted and fused with the corresponding vector similarity score to obtain a comprehensive score. The weight of the vector similarity score is determined based on the information richness of the joint representation vector of the user demand data and is positively correlated with the information richness of the joint representation vector of the user demand data.

[0113] The core objective of this implementation is to construct a multi-layered, adaptive patent ranking optimization mechanism. By combining the advantages of deep semantic understanding and vector space similarity, it significantly improves the accuracy and reliability of recommendation results while ensuring computational efficiency. This design aims to address the potential limitations of semantic understanding in a single similarity metric, thereby achieving intelligent matching.

[0114] A cross-encoder architecture is used to achieve deep semantic matching between queries and patents, and a relevance score is calculated. The weights of different similarity scores are dynamically adjusted based on query quality. , in Scaling factor This is the baseline threshold.

[0115] Receive a comprehensive score. , in The vector similarity scores from the initial screening. This represents the relevance score.

[0116] Traditional patent search systems often rely on a single similarity measurement method, which cannot meet the complex technical matching requirements. While vector similarity is computationally efficient, it may ignore deep semantic relationships, and cross-encoders, although highly accurate, have enormous computational costs. Therefore, establishing a hierarchical fusion ranking mechanism is an inevitable choice to achieve a balance between accuracy and efficiency.

[0117] High information richness queries, with their high vector similarity reliability, are assigned higher weights. Low information richness queries rely on the deep understanding of the cross-encoder. The comprehensive score integrates surface similarity and deep relevance, providing double verification for the recommendation results, enhancing their credibility, and significantly improving both the reliability of the results and user satisfaction.

[0118] As a preferred embodiment of the present invention, the patent matching method based on the fusion of large models and knowledge graphs further includes: Regularly collect newly generated patent text data and user feedback data on recommendation results; The knowledge graph model is incrementally updated using the newly generated patent text data. The patent semantic understanding model and the knowledge graph model are adjusted using the feedback behavior data.

[0119] The core objective of this implementation method is to build an intelligent system with continuous learning and self-optimization capabilities. By establishing a complete closed-loop mechanism for data collection, model updates, and feedback applications, the system can adapt to rapidly changing technological environments and user needs, maintaining long-term accuracy and practicality.

[0120] Multi-source intelligent data collection and processing: Newly published patent data is crawled regularly through the patent office API interface. An incremental identification mechanism based on publication time is adopted to ensure data integrity, and automatic deduplication and conflict detection are performed to avoid data redundancy.

[0121] User feedback behavior collection includes at least: Explicit feedback, including user satisfaction ratings and relevant / irrelevant markers; Implicit feedback includes click-through rate, dwell time, download behavior, and ignored behavior. Session data, query sequences, filter conditions, and sorting preferences.

[0122] The incremental evolution of the knowledge graph automatically identifies technical terms and innovative concepts in new patents, performs entity disambiguation and merging based on semantic similarity, and establishes temporal relationships between new and old entities. Incremental relation extraction captures new technology associations, adjusts the weights of old relations based on time decay, and automatically constructs and maintains technology evolution paths. A streaming graph neural network algorithm is employed, with local subgraph re-embedding to avoid full graph retraining, ensuring smooth evolution of the embedding space and maintaining semantic continuity.

[0123] End-to-end model tuning based on feedback signals. User feedback is used as a weakly supervised signal for online learning of the semantic understanding model. User satisfaction is used as a reward signal for online parameter updates of the policy gradient, achieving reinforcement learning of the recommendation policy.

[0124] Knowledge graph updates trigger semantic model retraining, and user feedback drives multi-model joint optimization, transforming the patent matching system from a static tool into a dynamic evolution, ensuring continuous value output in a rapidly changing technological environment.

[0125] The present invention also provides a computer-readable storage medium having a computer program stored thereon. The program implements the method when executed by the processor.

[0126] Therefore, the patent matching method based on the fusion of large models and knowledge graphs can achieve any effect, which will not be elaborated here.

[0127] For any parts not mentioned in this invention, existing technologies can be used or referenced.

[0128] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0129] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A patent matching method based on the fusion of large models and knowledge graphs, characterized in that, include: Based on the collected patent text data, semantic vectors are obtained through a pre-trained large model, and graph embedding vectors are obtained through a pre-trained knowledge graph model. The semantic vectors and graph embedding vectors are then fused into a joint representation vector. The method further includes: Based on the collected user demand data, semantic vectors are obtained through a pre-trained large model, and graph embedding vectors are obtained through a pre-trained knowledge graph model, in order to obtain joint representation vectors. Patents are initially screened based on the joint representation vector of the user demand data and the vector similarity of the patent text data. A relevance score is obtained by comparing the initially screened patent text data with the user demand data. The vector similarity is then fused using the relevance score to obtain a patent recommendation sequence.

2. The patent matching method based on the fusion of large models and knowledge graphs according to claim 1, characterized in that, After collecting the patent text data, the following is also included: Based on the collected patent text data, preprocessing and cleaning are performed, and the classification number is used as the fragmentation key to fragment and store the cleaned patent text data.

3. The patent matching method based on the fusion of large models and knowledge graphs according to claim 2, characterized in that, The pre-training of large models specifically involves: A general large language model is used as the base model. Based on the fragmented patent text data, a masked language model is trained to identify named entities, extract task relationships, and the fragmented patent text data is fragmented, adjusted, and labeled. Based on the annotated patent text data, a large model is trained to obtain semantic vectors.

4. The patent matching method based on the fusion of large models and knowledge graphs according to claim 3, characterized in that, The pre-training of knowledge graph models involves the following steps: Based on the annotated patent text data, triples containing head entities, relations, and tail entities are generated to construct a knowledge graph. The knowledge graph is embedded using a graph neural network to generate a graph embedding vector for each entity node in the graph, thus obtaining a knowledge graph model.

5. The patent matching method based on the fusion of large models and knowledge graphs according to claim 1, characterized in that, The semantic vector and graph embedding vector are fused into a joint representation vector, specifically: Based on semantic vectors and graph embedding vectors, the input is a gated fusion network. The information richness of the semantic vectors and graph embedding vectors is compared through network layers, and the fusion weights are obtained through an activation function. When the information richness of the semantic vector is greater than that of the graph embedding vector, the fusion weights are positively adjusted. When the information richness of the semantic vector is less than that of the graph embedding vector, the fusion weight is negatively adjusted; Based on the fusion weights, the semantic vectors are weighted and fused with the graph embedding vectors to obtain a joint representation vector.

6. The patent matching method based on the fusion of large models and knowledge graphs according to claim 1, characterized in that, Patent screening is performed based on the vector similarity between the joint representation vector of the user demand data and the vector similarity of the patent text data, specifically as follows: Based on the joint representation vector of the user demand data, the inner product similarity and Euclidean distance of the joint representation vector of the patent text data are obtained. The vector similarity is obtained by combining the inner product similarity and the Euclidean distance. Based on the vector similarity, the patent text data are sorted from largest to smallest to obtain a preliminary list containing multiple patent text data.

7. The patent matching method based on the fusion of large models and knowledge graphs according to claim 6, characterized in that, The preliminary screening list is as follows: The number of patent text data in the initial screening list is determined based on the information richness of the joint representation vector of the user demand data, and is negatively correlated with the information richness of the joint representation vector of the user demand data.

8. The patent matching method based on the fusion of large models and knowledge graphs according to claim 1, characterized in that, A relevance score is obtained by comparing the initially screened patent text data with the user demand data. This relevance score is then used to fuse vector similarity data. Specifically: Based on the user's request text, it is paired with each patent text data in the initial screening list to form multiple text pairs; Based on the multiple texts, a relevance score is obtained for each text pair by using a pre-trained large model operating in cross-encoder mode. The relevance score is weighted and fused with the corresponding vector similarity score to obtain a comprehensive score. The weight of the vector similarity score is determined based on the information richness of the joint representation vector of the user demand data and is positively correlated with the information richness of the joint representation vector of the user demand data.

9. The patent matching method based on the fusion of large models and knowledge graphs according to claim 1, characterized in that, Also includes: Regularly collect newly generated patent text data and user feedback data on recommendation results; The knowledge graph model is incrementally updated using the newly generated patent text data. The patent semantic understanding model and the knowledge graph model are adjusted using the feedback behavior data.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 9.

Images