Intelligent matching method and system for intelligent supply chain platform

Abstract

The application discloses an intelligent matching method and system for an intelligent supply chain platform, and the method comprises the following steps: receiving user demand data and performing multi-modal semantic analysis to generate a semantic network comprising a technical field label, a technical problem vector and a technical means correlation graph; generating a first matching list and a second matching list through collaborative operation of a knowledge graph matching model and a natural language processing model; performing adaptive cross-validation and generating a candidate set; uniformly sorting the candidate set according to a sorting rule and outputting a matching result. Through the semantic network of the technical field label, the problem vector and the technical correlation graph, the limitation of traditional keyword matching is broken through, and multi-dimensional accurate mapping of demand and supply is realized; the double-model cooperation of the knowledge graph model and the natural language processing model solves the problem of semantic deviation of a traditional single model; and the adaptive cross-validation mechanism dynamically switches a fusion strategy through a deviation rate, thereby effectively improving the robustness in a complex scene.

Description

Technical Field

[0001] This invention belongs to the field of supply chain supply and demand matching technology, and in particular relates to an intelligent matching method and system for an intelligent supply chain platform. Background Technology

[0002] In today's digital age, the accurate matching of technological resources within supply chain platforms is crucial for enterprise innovation and development. However, existing technology matching methods often have shortcomings. Traditional methods primarily rely on single text matching techniques, such as keyword retrieval. When dealing with complex technological needs, these methods struggle to accurately understand the deep semantics and multi-dimensional characteristics of user requirements. For example, when users raise complex needs involving multiple technological intersections, simple text matching can hardly fully capture and analyze the key information, significantly reducing the accuracy and relevance of the matching results.

[0003] Furthermore, with the rapid development and upgrading of technology, how to incorporate new technological resources into the matching system in a timely manner is also an urgent problem to be solved. The existing mechanisms for building and updating technology knowledge graphs are relatively lagging behind and cannot reflect the latest trends in technological evolution in real time, thus making it impossible for the matching results to provide users with the most cutting-edge and practical technology resource recommendations.

[0004] In summary, the existing technology matching methods in supply chain platforms have significant shortcomings in terms of accuracy and timeliness, making it difficult to meet the diverse needs of users in complex and ever-changing technological scenarios. Summary of the Invention

[0005] The purpose of this invention is to address the problems existing in the prior art by providing an intelligent matching method and system for an intelligent supply chain platform. Through the synergy of knowledge graph and natural language processing dual models, incremental graph updates, and adaptive cross-validation mechanisms, it achieves accurate matching and personalized recommendations in complex technical scenarios, thereby solving the core problems of insufficient understanding of multimodal needs, low efficiency of cross-modal data fusion, and lagging adaptation to dynamic technology evolution in supply chain technical resource matching.

[0006] Based on this invention specification, one aspect of the present invention provides an intelligent matching method for a supply chain platform, comprising:

[0007] It receives user demand data and performs multimodal semantic parsing to generate a semantic network containing technical field labels, technical problem vectors, and technical means association graphs.

[0008] The first and second matching lists are generated through the collaborative operation of a knowledge graph matching model and a natural language processing model. The knowledge graph matching model adopts dynamic entity linking technology and calculates semantic similarity in real time based on the real-time updated knowledge graph. The natural language processing model integrates a transfer learning mechanism to map the patent text vector space to the knowledge graph embedding space.

[0009] Perform adaptive cross-validation and generate a candidate set:

[0010] The candidate set is uniformly sorted according to the sorting rules, and the matching results are output.

[0011] In the above technical solution, the semantic network of technical field labels, question vectors and technical association graphs breaks through the limitations of traditional keyword matching and realizes multi-dimensional accurate mapping of demand and supply; the dual-model collaboration of knowledge graph model and natural language processing model solves the semantic bias problem of traditional single model; at the same time, the adaptive cross-validation mechanism effectively improves the robustness in complex scenarios by dynamically switching the fusion strategy through the deviation rate.

[0012] As a further technical solution, the collaborative operation between the knowledge graph matching model and the natural language processing model includes:

[0013] By constructing a knowledge graph, entity recognition and relationship extraction are performed on patent data to generate a first matching list based on the knowledge graph.

[0014] Text vectors are generated and semantic similarity is calculated using a pre-trained BERT model based on the Transformer architecture, generating a second matching list based on natural language processing;

[0015] Dual-model collaboration is achieved through cross-modal feature interaction, specifically including:

[0016] Graph attention networks are used to embed entity relationships from knowledge graphs into the encoding layer of a natural language processing model.

[0017] The KL divergence minimization algorithm is used to achieve bidirectional alignment between knowledge graph features and text features.

[0018] In the above technical solutions, the graph attention network injects the topological relationships of the knowledge graph into the encoding layer, enhancing the ability to capture implicit technical associations; KL divergence minimization forces the knowledge graph to align with the text feature space, which can reduce conflicts between models.

[0019] As a further technical solution, the construction of the knowledge graph includes:

[0020] Receive imported patent data and literature data, and perform data cleaning and deduplication.

[0021] Extract entities from the cleaned data, including technical terms, inventors, and patent numbers;

[0022] Determine the technical reference relationships and technical association relationships between entities;

[0023] Constructing knowledge graphs using graph database technology;

[0024] When new patent data is detected, perform an incremental update:

[0025] Identify breakpoints in technological evolution, whereby breakpoints include nodes where the connectivity between new technology features and existing knowledge graphs is below a first threshold.

[0026] Initiate local subgraph reconstruction, preserving existing entity relationships with a correlation degree ≥ the second threshold of new technology features;

[0027] Synchronously update the domain adaptation layer parameters of the natural language processing model;

[0028] The implicit technical relationships parsed by the natural language processing model are injected back into the knowledge graph.

[0029] In the above technical solutions, the incremental update mechanism enables the knowledge graph to dynamically respond to new technologies, avoiding the lag problem of traditional static knowledge bases. Local subgraph reconstruction (retaining only entities with a correlation degree ≥ a threshold) effectively reduces the computational cost of graph updates. Implicit relationships parsed by NLP are injected back into the knowledge graph, enhancing its predictability.

[0030] As a further technical solution, the adaptive cross-validation includes:

[0031] Construct a dual-channel verification matrix, where the main channel calculates the cosine similarity and the auxiliary channel calculates the technical feature coverage η = the number of intersections between the matching results and the user's required technical features / the total number of user's required technical features;

[0032] Configure a dynamic weight adjuster to adjust the weight in real time based on click feedback data from the user's historical operations.

[0033] In the above technical solution, by constructing a dual-channel verification matrix, the reliability of the matching results can be verified from multiple dimensions. Secondly, by dynamically adjusting the weights based on the click feedback data in the user's historical operations, the sorting rules of the matching results can be optimized in real time, improving the user experience. At the same time, when the deviation rate of the matching results exceeds the threshold, the feature fusion algorithm based on the attention mechanism is launched, which can dynamically adjust the feature weights and generate an optimized candidate set, further improving the accuracy of matching.

[0034] As a further technical solution, the attention-based feature fusion algorithm generates an optimized candidate set including:

[0035] Extract the knowledge graph feature vector V_KG from the first few patents in the first matching list;

[0036] Extract the NLP feature vector V_NLP from the first few patents in the second matching list;

[0037] The attention weight α for each patent is calculated using the following formula:

[0038] α = σ(W·[Tage,Ccite] + b);

[0039] σ is the Sigmoid function, W and b are training parameters, Tage is the patent age, Ccite is the number of citations, W and b are obtained through training with historical patent datasets, and the training objective is to minimize δ.

[0040] Among them, α is dynamically adjusted according to the technical field label of the patent;

[0041] Generate the fused feature vector: V_fused = α·V_KG + (1-α)·V_NLP;

[0042] Sort by ||V_fused||² norm, and select the top few items to form the optimized candidate set.

[0043] In the above technical solution, by calculating the attention weight of each patent and dynamically adjusting the weight according to the technical field label of the patent, the high relevance features can be highlighted and the accuracy of the matching results can be improved. By fusing knowledge graph feature vectors and NLP feature vectors to generate fused feature vectors and sorting them by norm, multiple features can be comprehensively considered to ensure the high quality of the optimized candidate set.

[0044] As a further technical solution, the sorting rules include:

[0045] First, sort the matching results in descending order based on their relevance;

[0046] When the relevance is the same, sort the patents in descending order by the number of patent citations.

[0047] When relevance and citation count are the same, the patents are sorted by application date, with priority given to patents with earlier application dates.

[0048] When the user is a research and development-intensive enterprise, the weight is adjusted as follows: Score = 0.5·SimKG + 0.2·log(SimNLP) + 0.3·(Ccite / Tage), where Score is the comprehensive score, SimKG is the knowledge graph similarity, and SimNLP is the natural language processing similarity.

[0049] The criteria for identifying R&D-intensive enterprises are: the number of patent searches in user historical data is greater than the fifth threshold times / year, and the frequency of technical feature updates is greater than the sixth threshold times / month.

[0050] In the above technical solutions, sorting by relevance, number of patent citations, and application time ensures the rationality and priority of the matching results. At the same time, adjusting the sorting weights for R&D-intensive enterprises can better meet their specific needs and improve user experience.

[0051] Based on the specification, this invention provides an intelligent matching system for a supply chain platform, comprising:

[0052] The data receiving and preprocessing module is used to receive user demand data and perform multimodal semantic parsing to generate a semantic network containing technical field labels, technical problem vectors, and technical means association graphs.

[0053] The dual-model matching module is used to generate a first matching list and a second matching list through the collaborative operation of a knowledge graph matching model and a natural language processing model.

[0054] The validation module performs adaptive cross-validation and generates a candidate set.

[0055] The output display module is used to uniformly sort the candidate set according to the sorting rules and output the matching results;

[0056] The data receiving and preprocessing module receives user requirement data through an API interface and connects to the dual-model matching module via a JSON-LD format data bus.

[0057] As a further technical solution, a heterogeneous model coprocessor is also included, wherein the heterogeneous model coprocessor integrates:

[0058] The knowledge graph dynamic embedding module uses graph neural network technology to generate entity vectors in real time.

[0059] The cross-modal alignment module achieves spatial mapping between text vectors and graph vectors through an adversarial generative network.

[0060] In the above technical solutions, the real-time generation of entity vectors using graph neural networks enables dynamic updates to the embedding representation of the knowledge graph, improving the real-time performance and accuracy of matching. Furthermore, the spatial mapping between text vectors and graph vectors via generative adversarial networks further enhances the effectiveness of cross-modal feature interaction and strengthens the overall matching performance.

[0061] As a further technical solution, a self-learning optimization engine is also included, wherein the self-learning optimization engine comprises:

[0062] A model bias monitor is used to track the deviation rate of the matching results between two models;

[0063] The parameter compensation matrix is ​​used to store the historical optimal weight combination;

[0064] Incremental training trigger: Retraining is initiated when the bias rate exceeds the fourth threshold three times consecutively.

[0065] Historical weight database stores the mapping relationship between user actions and optimal weight parameters;

[0066] The real-time feedback parser dynamically updates the weighting rules based on user click-through rates on matching results, filtering operation time, and final adoption results.

[0067] In the above technical solution, the deviation rate of the matching results can be monitored in real time through the model deviation monitor and incremental training trigger, and retraining can be initiated when necessary to ensure continuous optimization of the model. By storing the mapping relationship between user operation behavior and optimal weight parameters, and dynamically updating the weight rules according to user feedback, the system can achieve self-learning and self-optimization, thereby improving the system's intelligence level.

[0068] Compared with existing technologies, the beneficial effects of this invention are: 1. By generating a semantic network through multimodal semantic parsing, user needs can be comprehensively and accurately understood. The collaborative operation of the knowledge graph matching model and the natural language processing model, combined with dynamic entity linking technology and transfer learning mechanism, can improve the accuracy and relevance of the matching results. The adaptive cross-validation and attention-based feature fusion algorithm can be optimized when the matching results of the two models have large deviations, further improving the matching quality. The nonlinear ranking function comprehensively considers multiple factors, making the ranking results more scientific and reasonable. The ranking rules can output the matching results that best meet the user's needs according to different situations, effectively meeting the user's patent and other technology-related needs on the supply chain platform, improving matching efficiency and accuracy, and helping users quickly find suitable technical resources; 2. By constructing a cross-modal interaction network between the technology knowledge graph and the BERT model, the deep integration of structured knowledge and text semantics is achieved. The graph attention mechanism embeds entity relationships from the knowledge graph into BERT's Transformer layer, strengthening the contextual association of technical terms. The KL divergence algorithm bidirectionally distills structured and textual features, reducing cross-modal information loss and improving the consistency and robustness of matching results, making it particularly suitable for accurate matching in complex technical scenarios. The incremental knowledge graph update engine combines data cleaning and entity relationship extraction to ensure the graph dynamically reflects technological evolution trends. Local subgraph reconstruction and domain adaptation layer parameters are updated synchronously to avoid knowledge gaps caused by newly added patent data. The implicit technical relationship back-injection mechanism captures non-explicit technical associations (such as cross-domain technical fusion), enhancing the graph's completeness and forward-looking nature, and providing more comprehensive knowledge support for the matching model. Attached Figure Description

[0069] Figure 1 A flowchart of an intelligent matching method for an intelligent supply chain platform provided in an embodiment of the present invention;

[0070] Figure 2 This is a structural diagram of an intelligent matching system for an intelligent supply chain platform provided in an embodiment of the present invention. Detailed Implementation

[0071] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0072] like Figure 1 As shown, the specific solution of the embodiment is as follows: Based on the specification of this invention, one aspect of this invention provides an intelligent matching method for a supply chain platform, comprising:

[0073] It receives user demand data and performs multimodal semantic parsing to generate a semantic network containing technical field labels, technical problem vectors, and technical means association graphs.

[0074] The first and second matching lists are generated through the collaborative operation of a knowledge graph matching model and a natural language processing model. The knowledge graph matching model adopts dynamic entity linking technology and calculates semantic similarity in real time based on the real-time updated knowledge graph. The natural language processing model integrates a transfer learning mechanism to map the patent text vector space to the knowledge graph embedding space.

[0075] Perform adaptive cross-validation and generate a candidate set:

[0076] The candidate set is uniformly sorted according to the sorting rules, and the matching results are output.

[0077] In the above technical solution, the semantic network of technical field labels, question vectors and technical association graphs breaks through the limitations of traditional keyword matching and realizes multi-dimensional accurate mapping of demand and supply; the dual-model collaboration of knowledge graph model and natural language processing model solves the semantic bias problem of traditional single model; at the same time, the adaptive cross-validation mechanism effectively improves the robustness in complex scenarios by dynamically switching the fusion strategy through the deviation rate.

[0078] In this embodiment, the natural language processing model that integrates the transfer learning mechanism adopts any one of the commercially available mature BERT, GPT, and XLNet.

[0079] In this embodiment, the collaborative operation between the knowledge graph matching model and the natural language processing model includes:

[0080] By constructing a knowledge graph, entity recognition and relationship extraction are performed on patent data to generate a first matching list based on the knowledge graph.

[0081] Text vectors are generated and semantic similarity is calculated using a pre-trained BERT model based on the Transformer architecture, generating a second matching list based on natural language processing;

[0082] Dual-model collaboration is achieved through cross-modal feature interaction, specifically including:

[0083] Graph attention networks are used to embed entity relationships from knowledge graphs into the encoding layer of a natural language processing model.

[0084] The KL divergence minimization algorithm is used to achieve bidirectional alignment between knowledge graph features and text features.

[0085] In the above technical solutions, the graph attention network injects the topological relationships of the knowledge graph into the encoding layer, enhancing the ability to capture implicit technical associations; KL divergence minimization forces the knowledge graph to align with the text feature space, which can reduce conflicts between models.

[0086] In this embodiment, the construction of the knowledge graph includes:

[0087] Receive imported patent data and literature data, and perform data cleaning and deduplication.

[0088] Extract entities from the cleaned data, including technical terms, inventors, and patent numbers;

[0089] Determine the technical reference relationships and technical association relationships between entities;

[0090] Constructing knowledge graphs using graph database technology;

[0091] When new patent data is detected, perform an incremental update:

[0092] Identify breakpoints in technological evolution, whereby breakpoints include nodes where the connectivity between new technology features and existing knowledge graphs is below a first threshold.

[0093] Initiate local subgraph reconstruction, preserving existing entity relationships with a correlation degree ≥ the second threshold of new technology features;

[0094] Synchronously update the domain adaptation layer parameters of the natural language processing model;

[0095] The implicit technical relationships parsed by the natural language processing model are injected back into the knowledge graph.

[0096] In the above technical solutions, the incremental update mechanism enables the knowledge graph to dynamically respond to new technologies, avoiding the lag problem of traditional static knowledge bases. Local subgraph reconstruction (retaining only entities with a correlation degree ≥ a threshold) effectively reduces the computational cost of graph updates. Implicit relationships parsed by NLP are injected back into the knowledge graph, enhancing its predictability.

[0097] In this embodiment, the adaptive cross-validation includes:

[0098] Construct a dual-channel verification matrix, where the main channel calculates the cosine similarity and the auxiliary channel calculates the technical feature coverage η = the number of intersections between the matching results and the user's required technical features / the total number of user's required technical features;

[0099] Configure a dynamic weight adjuster to adjust the weight in real time based on click feedback data from the user's historical operations.

[0100] In this embodiment, the weights are adjusted in real time according to the following rules:

[0101] When the click-through rate does not exceed the third threshold, α and β each account for half of the weight;

[0102] When the click-through rate exceeds the third threshold and the filtering time is longer than the preset time, α+0.1, β-0.1, where α is the weight of the knowledge graph model and β is the weight of the natural language processing model.

[0103] Calculate the deviation rate δ = |SimKG - SimNLP| between the two models.

[0104] When δ does not exceed the fourth threshold, the data from the first matching list and the second matching list are merged and a candidate set is formed after deduplication.

[0105] When δ exceeds the fourth threshold, an attention-based feature fusion algorithm is activated to generate an optimized candidate set.

[0106] The candidate set is ranked using a non-linear sorting function, which satisfies the following:

[0107] Score=0.4·SimKG+0.3·log(SimNLP)+0.3·(Ccite / Tage);

[0108] Score is the overall score, SimKG is the knowledge graph similarity, SimNLP is the natural language processing similarity, Tage = current year - patent application year + 1, and Ccite is the number of patent citations.

[0109] In the above technical solution, by constructing a dual-channel verification matrix, the reliability of the matching results can be verified from multiple dimensions. Secondly, by dynamically adjusting the weights based on the click feedback data in the user's historical operations, the ranking rules of the matching results can be optimized in real time, improving the user experience. At the same time, when the deviation rate of the matching results exceeds the threshold, the feature fusion algorithm based on the attention mechanism is launched, which can dynamically adjust the feature weights and generate an optimized candidate set, further improving the accuracy of matching. In addition, by classifying the candidate set through a non-linear ranking function, factors such as knowledge graph similarity, natural language processing similarity, patent age, and citation count can be comprehensively considered to ensure the rationality and relevance of the ranking results.

[0110] In this embodiment, the feature fusion algorithm based on the attention mechanism generates the optimized candidate set as follows:

[0111] Extract the knowledge graph feature vector V_KG from the first few patents in the first matching list;

[0112] Extract the NLP feature vector V_NLP from the first few patents in the second matching list;

[0113] The attention weight α for each patent is calculated using the following formula:

[0114] α = σ(W·[Tage,Ccite] + b);

[0115] σ is the Sigmoid function, W and b are training parameters, Tage is the patent age, Ccite is the number of citations, W and b are obtained through training with historical patent datasets, and the training objective is to minimize δ.

[0116] Among them, α is dynamically adjusted according to the technical field label of the patent;

[0117] Generate the fused feature vector: V_fused = α·V_KG + (1-α)·V_NLP;

[0118] Sort by ||V_fused||² norm, and select the top few items to form the optimized candidate set.

[0119] In the above technical solution, by calculating the attention weight of each patent and dynamically adjusting the weight according to the technical field label of the patent, the high relevance features can be highlighted and the accuracy of the matching results can be improved. By fusing knowledge graph feature vectors and NLP feature vectors to generate fused feature vectors and sorting them by norm, multiple features can be comprehensively considered to ensure the high quality of the optimized candidate set.

[0120] In this embodiment, the sorting rules include:

[0121] First, sort the matching results in descending order based on their relevance;

[0122] When the relevance is the same, sort the patents in descending order by the number of patent citations.

[0123] When relevance and citation count are the same, the patents are sorted by application date, with priority given to patents with earlier application dates.

[0124] When the user is a research and development-intensive enterprise, the weight is adjusted as follows: Score = 0.5·SimKG + 0.2·log(SimNLP) + 0.3·(Ccite / Tage);

[0125] The criteria for identifying R&D-intensive enterprises are: the number of patent searches in user historical data is greater than the fifth threshold times / year, and the frequency of technical feature updates is greater than the sixth threshold times / month.

[0126] In the above technical solutions, sorting by relevance, number of patent citations, and application time ensures the rationality and priority of the matching results. At the same time, adjusting the sorting weights for R&D-intensive enterprises can better meet their specific needs and improve user experience.

[0127] Example 1:

[0128] Based on the above-mentioned invention, this embodiment provides a technical solution for a specific and operable intelligent matching method for supply chain platforms, aiming to provide an efficient, intelligent and scalable solution for the accurate matching of technical resources in supply chain platforms, and to meet the diverse needs of users in complex technical scenarios.

[0129] During the user requirement reception phase, the system comprehensively analyzes the user-input requirement data using multimodal semantic parsing technology. Specifically, it utilizes lexical analysis, syntactic analysis, and semantic understanding algorithms from natural language processing, combined with multimodal data processing methods such as image recognition technology (if the user input includes images), to extract key information from the user requirements. This includes technology field tags (such as artificial intelligence, biomedicine, etc.), technical problem vectors (multidimensional feature representations describing technical challenges), and technical means correlation graphs (showing the relationships between different technical methods).

[0130] For example, when a user inputs a text requesting intelligent logistics route optimization, the system can identify that it belongs to the "logistics technology" field, extract technical problem vectors such as "shortening delivery time" and "vehicle load limit", and construct a correlation graph of related technical means such as "ant colony algorithm" and "genetic algorithm", thereby generating a complete semantic network and laying the foundation for subsequent accurate matching.

[0131] Upon entering the matching phase, the system initiates collaborative computation between the knowledge graph matching model and the natural language processing model. The knowledge graph matching model is based on a pre-constructed technical knowledge graph, which integrates massive amounts of patent data, literature data, and other technical resources, and uses dynamic entity linking technology to connect the latest technological evolution information in real time.

[0132] It is important to note that the data source for the pre-built technical knowledge graph here comes from the client's own company. The patent data collected by the client's own company is more relevant to their own field compared to public resources.

[0133] For example, in the field of new energy vehicle battery technology, the knowledge graph contains entities such as battery materials, energy density, and charging time, as well as the relationships between them. The model generates a first matching list by calculating the semantic similarity between the user's demand semantic network and the entities and relationships in the knowledge graph. Simultaneously, a natural language processing model uses a pre-trained BERT model to encode the patent text, generating text vectors, and then maps these vectors to the knowledge graph embedding space through transfer learning, thus generating a second matching list. Taking a user demand related to "high-efficiency battery management system" as an example, the BERT model can understand its semantic relevance to "battery monitoring system" and "battery balancing system," and find the corresponding positions in the embedding space with the knowledge graph entities, thereby generating a matching list. This dual-model collaborative approach fully leverages the structured knowledge advantages of the knowledge graph and the text understanding capabilities of the natural language processing model, improving the comprehensiveness and accuracy of the matching results.

[0134] To ensure the reliability of the matching results, the system performs adaptive cross-validation. A dual-channel validation matrix is ​​constructed. The main channel uses cosine similarity to compare the matching results of the two models, while the auxiliary channel analyzes the coverage of technical elements to check whether the matching results encompass the key technical elements of the user's needs. For example, for a requirement involving intelligent warehouse robot navigation technology, the system verifies whether the matching results include key technical elements such as robot localization, path planning, and obstacle avoidance. When the deviation rate δ of the matching results between the two models is ≥ 0.25, it indicates a significant difference, and the system initiates a feature fusion algorithm based on an attention mechanism. This algorithm learns the importance weights of different features (such as entity relationship features in knowledge graphs and text semantic features in natural language processing models) to the matching results, generating an optimized candidate set. For example, in some emerging technology fields, knowledge graphs may not be updated in a timely manner, while natural language processing models can better capture novel information in text. In this case, the attention mechanism will assign higher weights to natural language processing features, thereby optimizing the candidate set and improving the matching quality.

[0135] The system applies a non-linear sorting function to classify the optimized candidate set.

[0136] Function: Score = 0.4·SimKG + 0.3·log(SimNLP) + 0.3·(Ccite / Tage)

[0137] This formula comprehensively considers SimKG (Knowledge Graph Similarity), SimNLP (Natural Language Processing Similarity), Ccite (Patent Citation Count), and Tage (Patent Age Factor). SimKG reflects the relevance of the matching result to the user's needs within the knowledge graph structure; SimNLP measures the semantic similarity of the text; Ccite reflects the patent's influence and value; and Tage considers the patent's timeliness. For example, a patent that is highly relevant to the user's needs in the knowledge graph (SimKG=0.8), has high semantic similarity (SimNLP=0.7), is frequently cited (Ccite=100), and is relatively new (Tage=2) would have the following comprehensive score:

[0138] The score = 0.4 × 0.8 + 0.3 × log(0.7) + 0.3 × (100 / 2) = 0.32 + 0.3 × (-0.1549) + 0.3 × 50 ≈ 0.32 - 0.0465 + 15 = 15.2735, which will be ranked higher. According to the sorting rules, the results are first sorted in descending order of relevance; when relevance is the same, they are sorted in descending order of patent citation count; and finally, they are sorted by the chronological order of patent application.

[0139] In addition, for R&D-intensive enterprises, the weighting formula has been adjusted as follows:

[0140] Score=0.5·SimKG+0.2·log(SimNLP)+0.3·(Ccite / Tage),

[0141] This formula emphasizes the weight of knowledge graph similarity, which better aligns with the focus of these companies on core technology paths, ensuring that the final matching results best meet user needs and helping users quickly locate high-quality technology resources.

[0142] like Figure 2 As shown, based on the same technical concept as the above embodiments, the present invention provides an intelligent matching system for a supply chain platform, comprising:

[0143] The data receiving and preprocessing module is used to receive user demand data and perform multimodal semantic parsing to generate a semantic network containing technical field labels, technical problem vectors, and technical means association graphs.

[0144] The dual-model matching module is used to generate a first matching list and a second matching list through the collaborative operation of a knowledge graph matching model and a natural language processing model.

[0145] The validation module performs adaptive cross-validation and generates a candidate set.

[0146] The output display module is used to uniformly sort the candidate set according to the sorting rules and output the matching results;

[0147] The data receiving and preprocessing module receives user requirement data through an API interface and connects to the dual-model matching module via a JSON-LD format data bus.

[0148] This embodiment also includes a heterogeneous model coprocessor, which integrates:

[0149] The knowledge graph dynamic embedding module uses graph neural network technology to generate entity vectors in real time.

[0150] The cross-modal alignment module achieves spatial mapping between text vectors and graph vectors through an adversarial generative network.

[0151] In the above technical solutions, the real-time generation of entity vectors using graph neural networks enables dynamic updates to the embedding representation of the knowledge graph, improving the real-time performance and accuracy of matching. Furthermore, the spatial mapping between text vectors and graph vectors via generative adversarial networks further enhances the effectiveness of cross-modal feature interaction and strengthens the overall matching performance.

[0152] This embodiment also includes a self-learning optimization engine, which comprises:

[0153] A model bias monitor is used to track the deviation rate of the matching results between two models;

[0154] The parameter compensation matrix is ​​used to store the historical optimal weight combination;

[0155] Incremental training trigger: Retraining is initiated when the bias rate exceeds the fourth threshold three times consecutively.

[0156] Historical weight database stores the mapping relationship between user actions and optimal weight parameters;

[0157] The real-time feedback parser dynamically updates the weighting rules based on user click-through rates on matching results, filtering operation time, and final adoption results.

[0158] In the above technical solution, the deviation rate of the matching results can be monitored in real time through the model deviation monitor and incremental training trigger, and retraining can be initiated when necessary to ensure continuous optimization of the model. By storing the mapping relationship between user operation behavior and optimal weight parameters, and dynamically updating the weight rules according to user feedback, the system can achieve self-learning and self-optimization, thereby improving the system's intelligence level.

[0159] In this embodiment, the workflow of the incremental training trigger includes:

[0160] When new patent data is added to the database, the technical feature words in its claims are extracted and matched with the core entities of the knowledge graph for similarity matching;

[0161] If the matching degree is lower than the preset threshold, the local subgraph reconstruction algorithm is started to retain the entity relationship chains with an association degree ≥ 0.6 in the original graph;

[0162] At the same time, the adaptation layer parameters of the natural language processing model are updated to ensure that the semantic similarity fluctuation rate of the outputs of the new and old models is ≤5%.

[0163] In the above technical solution, the incremental training trigger intelligently decides the scope of local graph reconstruction based on the technical feature matching degree of the new patent data (retaining entity chains with a correlation degree ≥ 0.6), avoiding computational redundancy caused by full updates. The synchronous update of the adaptation layer parameters controls the semantic similarity fluctuation rate (≤ 5%), ensuring a smooth transition of the system to the new data environment, balancing efficiency and stability.

[0164] In this embodiment, the parameter compensation matrix generates the optimal weight combination in the following manner:

[0165] Record the model weight parameters corresponding to the user's final adoption result in historical matching tasks;

[0166] The random forest algorithm is used to analyze the association rules between weight parameters and user enterprise type, technology field and demand complexity; a multi-dimensional weight mapping table is generated, and when a new task is triggered, the nearest weight combination is automatically matched according to the real-time input features.

[0167] In the above technical solution, the parameter compensation matrix analyzes the association rules between weights and user features in historical tasks using the random forest algorithm to generate a multi-dimensional weight mapping table. When a new task is triggered, the optimal weight combination is automatically matched based on real-time input features, realizing intelligent configuration for personalized recommendations. This significantly shortens the model tuning cycle, improves the service response speed for large-scale enterprise users, enhances the personalization and adaptability of matching results, further improves the user experience and the system's intelligence level, and enables the system to better meet the complex needs of diverse user groups in different scenarios, thereby enhancing the system's competitiveness and practicality in the market.

[0168] Example 2:

[0169] The intelligent matching system in this embodiment adopts a modular design, including a data receiving and preprocessing module, a dual-model matching module, a verification module, a candidate grading module, and an output display module. The data receiving and preprocessing module receives user request data and uses multimodal semantic parsing technology to generate a semantic network. For example, it parses user-uploaded request information containing multiple modalities such as text and images, extracts key information, and constructs a semantic network. The dual-model matching module includes a knowledge graph matching model and a natural language processing model. The former uses dynamic entity linking technology and a real-time updated technology evolution graph to calculate semantic similarity, while the latter integrates a transfer learning mechanism to map the patent text vector space to the knowledge graph embedding space, collaboratively generating a matching list. The verification module performs adaptive cross-validation, constructs a dual-channel verification matrix, and sets a dynamic weight adjuster. It adjusts the model weight parameters in real time based on the user's historical operation data to optimize the matching results. The candidate grading module applies a non-linear ranking function to grade the candidate set, comprehensively considering multiple factors to score each candidate result. The output display module uniformly sorts the results after cross-validation according to the ranking rules, presenting the final matching results to the user in an intuitive and clear manner, such as displaying the matched patent information and relevance scores in a list format.

[0170] The system also includes a heterogeneous model coprocessor, integrating a knowledge graph dynamic embedding module and a cross-modal alignment module. The knowledge graph dynamic embedding module uses graph neural networks to generate entity vectors in real time. For example, when processing new technical knowledge graph data, the graph neural network can dynamically generate low-dimensional vector representations for each technical entity based on the graph's topology and entity attributes, preserving the topological features of its structured relationships, enabling the knowledge graph to adapt to constantly changing technological environments. The cross-modal alignment module uses a generative adversarial network to achieve spatial mapping between text vectors and graph vectors, addressing the semantic gap between different modalities. For example, when fusing text vectors generated by a natural language processing model with knowledge graph vectors, the generative adversarial network forces consistency in cross-modal spatial mapping through adversarial training of the generator and discriminator, ensuring that text vectors are accurately mapped to the knowledge graph vector space, guaranteeing consistency and coherence in semantic representation, and improving the system's ability to process heterogeneous data and its matching effect.

[0171] The self-learning optimization engine is a crucial component of the system, comprising a model bias monitor, a parameter compensation matrix, an incremental training trigger, a historical weight database, and a real-time feedback parser. The model bias monitor continuously tracks the differences between the two model outputs, recording biases during the matching process. The parameter compensation matrix stores historically optimal weight combinations, providing a reference for model parameter adjustments. The incremental training trigger initiates retraining when the bias rate exceeds 0.25 three consecutive times, ensuring the model can adapt to changes in data and requirements promptly. The historical weight database stores the mapping relationship between user actions and optimal weight parameters, accumulating user interaction data to support personalized services. The real-time feedback parser dynamically updates weight rules based on user click-through rates on matching results, filtering operation duration, and final adoption results. For example, if a user frequently clicks on a certain type of matching result while ignoring others, the system analyzes their behavioral patterns and adjusts the model weights to make subsequent matching results better match user preferences and needs, achieving system self-optimization and continuous improvement, and enhancing user experience and matching accuracy.

[0172] When performing alignment operations, the cross-modal alignment module first constructs semantic anchors in the knowledge graph vector space. These semantic anchors are representative locations determined based on key entities and technical concepts in the knowledge graph. Then, the text vectors output by the BERT model are projected onto the neighborhood of these anchors through a non-linear transformation, allowing the text vectors to find appropriate positions in the knowledge graph vector space. Next, the KL divergence algorithm is used to minimize the distribution difference between the graph vectors and the text vectors, continuously adjusting the transformation parameters until the cosine similarity between the two in the latent space is ≥0.85. For example, when processing a batch of new technical text data, this method ensures that the text vectors and knowledge graph vectors are highly semantically consistent, solving the problem of heterogeneous data fusion. This ensures that the matching model can stably output high-quality matching results when handling complex technical scenarios, such as in cross-domain technology fusion scenarios, accurately matching the technical needs described in the text with relevant technical entities in the knowledge graph, providing users with comprehensive and accurate technical resource recommendations.

[0173] When new patent data is added to the database, the incremental training trigger begins. First, technical feature words from the new patent claims, such as "high-efficiency solar cell" and "conversion efficiency improvement," are extracted and matched against the core entities of the knowledge graph. If the matching degree is lower than a preset threshold, it indicates that the new patent data brings new technical features or concepts, at which point a local subgraph reconstruction algorithm is activated. This algorithm only reconstructs local subgraphs related to the new patent data, preserving entity relationship chains with a correlation degree ≥ 0.6 in the original graph, avoiding large-scale updates to the entire knowledge graph and saving computational resources. Simultaneously, the adaptation layer parameters of the natural language processing model are updated to ensure that the semantic similarity fluctuation rate between the old and new model outputs is ≤ 5%, maintaining system performance stability. For example, if a new battery technology patent is added to the database in the new energy field, the trigger will decide whether to initiate local subgraph reconstruction and model parameter updates based on the matching of its technical feature words with existing battery technology entities in the knowledge graph. This ensures that the system can promptly incorporate new technology information, providing users with the latest technology matching services without causing significant fluctuations in system performance or affecting user experience due to updates.

[0174] The parameter compensation matrix accumulates a large number of effective configurations from practical applications by recording the model weight parameters corresponding to the final adoption results of users in historical matching tasks. Then, a random forest algorithm is used to analyze the association rules between the weight parameters and the user's enterprise type, technology field, and demand complexity. For example, it was found that for R&D-intensive enterprises in the electronics and information field, higher knowledge graph weight parameters better meet their needs when dealing with complex technological innovation requirements. Based on these analysis results, a multi-dimensional weight mapping table is generated. When a new task is triggered, the system automatically matches the nearest weight combination based on real-time input features (such as the user's industry and demand type), achieving intelligent configuration for personalized recommendations. This approach not only significantly shortens the model tuning cycle but also quickly responds to the diverse service needs of large-scale enterprise users, improves the personalization and adaptability of matching results, further enhances the user experience and the system's intelligence level, and strengthens the system's practicality and competitiveness in the market.

[0175] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent matching method for a supply chain platform, characterized in that, include: It receives user demand data and performs multimodal semantic parsing to generate a semantic network containing technical field labels, technical problem vectors, and technical means association graphs. The first and second matching lists are generated through the collaborative operation of a knowledge graph matching model and a natural language processing model. The knowledge graph matching model adopts dynamic entity linking technology and calculates semantic similarity in real time based on the real-time updated knowledge graph. The natural language processing model integrates a transfer learning mechanism to map the patent text vector space to the knowledge graph embedding space. Perform adaptive cross-validation on the first and second matching lists and generate a candidate set: the adaptive cross-validation includes: Construct a dual-channel verification matrix, where the main channel calculates the cosine similarity and the auxiliary channel calculates the technical feature coverage η = the number of intersections between the matching results and the user's required technical features / the total number of user's required technical features; Calculate the deviation rate δ = |SimKG - SimNLP| for the matching results of the two models, where SimKG is the knowledge graph similarity and SimNLP is the natural language processing similarity. When δ does not exceed the fourth threshold, the data from the first matching list and the second matching list are merged and a candidate set is formed after deduplication. When δ exceeds the fourth threshold, an attention-based feature fusion algorithm is activated to generate an optimized candidate set. The attention-based feature fusion algorithm generates an optimized candidate set including: Extract the knowledge graph feature vector V_KG from the first few patents in the first matching list; Extract the NLP feature vector V_NLP from the first few patents in the second matching list; The attention weight α for each patent is calculated using the following formula: α = σ(W·[Tage,Ccite] + b); σ is the Sigmoid function, W and b are training parameters, Tage is the patent age, Ccite is the number of citations, W and b are obtained through training with historical patent datasets, and the training objective is to minimize δ. Among them, α is dynamically adjusted according to the technical field label of the patent; Generate the fused feature vector: V_fused = α·V_KG + (1-α)·V_NLP; Sort by ||V_fused||² norm, and select the top few items to form the optimized candidate set; The candidate set is uniformly sorted according to the sorting rules, and the matching results are output.

2. The intelligent matching method for a supply chain platform according to claim 1, characterized in that, The collaborative operation between the knowledge graph matching model and the natural language processing model includes: By constructing a knowledge graph, entity recognition and relationship extraction are performed on patent data to generate a first matching list based on the knowledge graph. Text vectors are generated and semantic similarity is calculated using a pre-trained BERT model based on the Transformer architecture, generating a second matching list based on natural language processing; Dual-model collaboration is achieved through cross-modal feature interaction, specifically including: Graph attention networks are used to embed entity relationships from knowledge graphs into the encoding layer of a natural language processing model. The KL divergence minimization algorithm is used to achieve bidirectional alignment between knowledge graph features and text features.

3. The intelligent matching method for a supply chain platform according to claim 2, characterized in that, The construction of the knowledge graph includes: Receive imported patent data and literature data, and perform data cleaning and deduplication. Extract entities from the cleaned data, including technical terms, inventors, and patent numbers; Determine the technical reference relationships and technical association relationships between entities; Constructing knowledge graphs using graph database technology; When new patent data is detected, perform an incremental update: Identify breakpoints in technological evolution, whereby breakpoints include nodes where the connectivity between new technology features and existing knowledge graphs is below a first threshold. Initiate local subgraph reconstruction, preserving existing entity relationships with a correlation degree ≥ the second threshold of new technology features; Synchronously update the domain adaptation layer parameters of the natural language processing model; The implicit technical relationships parsed by the natural language processing model are injected back into the knowledge graph.

4. The intelligent matching method for a supply chain platform according to claim 1, characterized in that, The adaptive cross-validation includes: Construct a dual-channel verification matrix, where the main channel calculates the cosine similarity and the auxiliary channel calculates the technical feature coverage η = the number of intersections between the matching results and the user's required technical features / the total number of user's required technical features; Configure a dynamic weight adjuster to adjust the weight in real time based on click feedback data from the user's historical operations.

5. The intelligent matching method for a supply chain platform according to claim 1, characterized in that, The sorting rules include: First, sort the matching results in descending order based on their relevance; When the relevance is the same, sort the patents in descending order by the number of patent citations. When relevance and citation count are the same, the patents are sorted by application date, with priority given to patents with earlier application dates. When the user is a research and development-intensive enterprise, the weight is adjusted as follows: Score = 0.5·SimKG + 0.2·log(SimNLP) + 0.3·(Ccite / Tage), where Score is the comprehensive score, SimKG is the knowledge graph similarity, and SimNLP is the natural language processing similarity. The criteria for identifying R&D-intensive enterprises are: the number of patent searches in user historical data is greater than the fifth threshold times / year, and the frequency of technical feature updates is greater than the sixth threshold times / month.

6. An intelligent matching system for a supply chain platform, characterized in that, include: The data receiving and preprocessing module is used to receive user demand data and perform multimodal semantic parsing to generate a semantic network containing technical field labels, technical problem vectors, and technical means association graphs. The dual-model matching module is used to generate a first matching list and a second matching list through the collaborative operation of a knowledge graph matching model and a natural language processing model. The collaborative operation between the knowledge graph matching model and the natural language processing model includes: By constructing a knowledge graph, entity recognition and relationship extraction are performed on patent data to generate a first matching list based on the knowledge graph. The text vectors are generated and semantic similarity is calculated using a pre-trained BERT model based on the Transformer architecture. This generates a second matching list based on natural language processing. Specifically, the natural language processing model encodes the patent text using the pre-trained BERT model to generate text vectors, and maps them to the knowledge graph embedding space through a transfer learning mechanism to generate a second matching list. Dual-model collaboration is achieved through cross-modal feature interaction, specifically including: Graph attention networks are used to embed entity relationships from knowledge graphs into the encoding layer of a natural language processing model. Achieve bidirectional alignment between knowledge graph features and text features using the KL divergence minimization algorithm; The validation module performs adaptive cross-validation and generates a candidate set. The adaptive cross-validation includes: Construct a dual-channel verification matrix, where the main channel calculates the cosine similarity and the auxiliary channel calculates the technical feature coverage η = the number of intersections between the matching results and the user's required technical features / the total number of user's required technical features; Calculate the deviation rate δ = |SimKG - SimNLP| for the matching results of the two models, where SimKG is the knowledge graph similarity and SimNLP is the natural language processing similarity. When δ does not exceed the fourth threshold, the data from the first matching list and the second matching list are merged and a candidate set is formed after deduplication. When δ exceeds the fourth threshold, an attention-based feature fusion algorithm is activated to generate an optimized candidate set. The attention-based feature fusion algorithm generates an optimized candidate set including: Extract the knowledge graph feature vector V_KG from the first few patents in the first matching list; Extract the NLP feature vector V_NLP from the first few patents in the second matching list; The attention weight α for each patent is calculated using the following formula: α = σ(W·[Tage,Ccite] + b); σ is the Sigmoid function, W and b are training parameters, Tage is the patent age, Ccite is the number of citations, W and b are obtained through training with historical patent datasets, and the training objective is to minimize δ. Among them, α is dynamically adjusted according to the technical field label of the patent; Generate the fused feature vector: V_fused = α·V_KG + (1-α)·V_NLP; Sort by ||V_fused||² norm, and select the top few items to form the optimized candidate set; The output display module is used to uniformly sort the candidate set according to the sorting rules and output the matching results; The data receiving and preprocessing module receives user requirement data through an API interface and connects to the dual-model matching module via a JSON-LD format data bus.

7. The intelligent matching system for a supply chain platform according to claim 6, characterized in that, It also includes a heterogeneous model coprocessor, which integrates: The knowledge graph dynamic embedding module uses graph neural network technology to generate entity vectors in real time. The cross-modal alignment module achieves spatial mapping between text vectors and graph vectors through an adversarial generative network.

8. The intelligent matching system for a supply chain platform according to claim 6, characterized in that, It also includes a self-learning optimization engine, which comprises: A model bias monitor is used to track the deviation rate of the matching results between two models; The parameter compensation matrix is ​​used to store the historical optimal weight combination; Incremental training trigger: Retraining is initiated when the bias rate exceeds the fourth threshold three times consecutively. Historical weight database stores the mapping relationship between user actions and optimal weight parameters; The real-time feedback parser dynamically updates the weighting rules based on user click-through rates on matching results, filtering operation time, and final adoption results.

Images