A big data-based intelligent matching method for scientific and technological achievements

Abstract

This invention discloses an intelligent matching method for scientific and technological achievements based on big data. Relating to the fields of big data and artificial intelligence, this method addresses the problems of existing matching methods, such as the difficulty in bridging semantic gaps, lack of assessment of engineering feasibility, and uninterpretable matching results. It achieves dynamic semantic alignment by constructing academic language subgraphs and industry language subgraphs and performing cross-modal comparative learning. Engineering features are extracted to infer technology maturity and manufacturing maturity, which are then introduced as penalty terms into the graph neural network matching calculation. Furthermore, a causal reasoning graph network is used to perform interventional causal strength estimation and pruning on candidate matching pairs, generating a visualized decision reasoning subgraph and outputting a matching evaluation report. This invention can improve the semantic accuracy, industrial feasibility, and interpretability of matching scientific and technological achievements with technological needs, and is suitable for intelligent matching scenarios in technology transfer platforms.

Description

Technical Field

[0001] This invention relates to the fields of big data and artificial intelligence technology, specifically to a method for intelligent matching of scientific and technological achievements based on big data. Background Technology

[0002] Technology transfer refers to the process by which scientific and technological achievements, such as patents and papers, generated by universities and research institutes are transferred to enterprises through licensing, transfer, or investment to achieve application. This process usually requires semantic matching between the technical description of the scientific and technological achievements and the technical needs of enterprises to discover supply and demand matching opportunities. This process relies heavily on intelligent matching technology based on big data. The well-known technologies involved in this matching process mainly include named entity recognition and relation extraction technology, knowledge graph construction and embedding technology, text vectorization technology, and graph neural network technology in natural language processing, which are used to extract entity relations from unstructured text and perform low-dimensional vector representation.

[0003] The closest existing technology to this invention is a technology achievement matching system based on deep semantic vector similarity. Its implementation involves: collecting texts of technology achievements and enterprise technology requirements, inputting them into a pre-trained language model to obtain semantic vectors, calculating the cosine similarity between the vectors as a matching score, sorting candidate technology achievements in descending order according to the score, and outputting a matching recommendation list. Its system structure consists of a text acquisition module, a semantic encoding module, a similarity calculation module, and a sorting output module connected sequentially, belonging to a single-path matching architecture.

[0004] The aforementioned existing technologies have the following shortcomings:

[0005] First, the use of a general pre-trained language model to uniformly encode academic and industrial texts fails to consider the structural differences in their terminology systems and expression paradigms, leading to mismatches and missed matches in the matching results.

[0006] Secondly, similarity calculation is based only on the surface semantic features of the text and does not assess the engineering status of the scientific and technological achievements, resulting in invalid matches with high superficial scores but difficult to implement in practice.

[0007] Third, it only outputs indivisible matching scores without revealing the reasoning behind the scores, making it a black box model that is difficult to gain the trust of technical managers and corporate decision-makers. Summary of the Invention

[0008] The technical problem this invention aims to solve is that existing big data-based methods for matching scientific and technological achievements with technological needs generally suffer from three major defects: difficulty in bridging the semantic gap, lack of assessment of engineering feasibility, and uninterpretable matching results. These defects lead to low accuracy in matching scientific and technological achievements with industry needs, poor feasibility, and low trust between supply and demand sides, thus hindering the efficiency of scientific and technological achievement transformation. This invention aims to provide an intelligent method for matching scientific and technological achievements that can dynamically align the semantics of academic and industrial language, integrate technological maturity and manufacturing maturity for engineering feasibility assessment, and output interpretable matching paths.

[0009] To address the aforementioned technical problems, this invention provides a method for intelligent matching of scientific and technological achievements based on big data, comprising: acquiring text of scientific and technological achievements and text of technical requirements;

[0010] An academic language subgraph is constructed based on the text of the scientific and technological achievements, and an industrial language subgraph is constructed based on the text of the technical requirements. Cross-modal comparative learning training is performed on the node embeddings in the academic language subgraph and the industrial language subgraph to obtain a unified semantic embedding space, thereby achieving dynamic alignment of academic language and industrial language at the semantic level.

[0011] Based on the text of the scientific and technological achievements, obtain the technology maturity level and manufacturing maturity level corresponding to the nodes of the scientific and technological achievements and form a maturity vector, and obtain the technology maturity level threshold and manufacturing maturity level threshold corresponding to the technology demand nodes.

[0012] Based on the node embeddings output from the unified semantic embedding space and the maturity vector, a heterogeneous graph containing technology achievement nodes and technology demand nodes is constructed. The heterogeneous graph is propagated through a graph neural network to obtain the hidden layer representation of the nodes. The similarity between technology achievement nodes and technology demand nodes is calculated based on the hidden layer representation of the nodes. An engineering feasibility gap is constructed based on the difference between the maturity vector and the technology maturity level threshold and the manufacturing maturity level threshold. The similarity and the engineering feasibility gap are weighted and combined to obtain a matching score.

[0013] Candidate matching pairs are determined based on the matching score. A local causal graph including technical principle nodes, application scenario nodes, and engineering gap nodes is constructed for the candidate matching pairs. Interventional causal strength estimation is performed on the candidate causal edges in the local causal graph. Candidate causal edges with causal strength exceeding a preset threshold are retained to obtain a decision reasoning subgraph.

[0014] A matching evaluation report is generated and output based on the decision reasoning subgraph.

[0015] Among them, the Technology Readiness Level is used to quantitatively characterize the development stage of scientific and technological achievements from basic research to practical application;

[0016] Manufacturing Readiness Level (MRL) is used to characterize the quantitative level of engineering manufacturing conditions that enable scientific and technological achievements to move from laboratory processes to large-scale, stable production.

[0017] Compared with the prior art, the beneficial effects of the present invention by adopting the above technical solution are as follows:

[0018] 1. This invention constructs academic language subgraphs and industry language subgraphs respectively, and introduces a cross-modal contrastive learning model to dynamically align the node embeddings of the two. This reduces the distance between semantically related cross-domain entities in a unified semantic embedding space and increases the distance between semantically unrelated entities. It fundamentally bridges the semantic gap between the academic language system and the industry language system, significantly improves the semantic accuracy of the matching results, and reduces the occurrence of mismatches and missed matches.

[0019] 2. This invention extracts engineering feature fields from the patent text, paper text, and third-party testing report text corresponding to the scientific and technological achievements, and deduces the technology maturity level and manufacturing maturity level. Then, it constructs an engineering feasibility gap and introduces it as a penalty item into the matching score calculation process of the graph neural network. This makes the matching result not only reflect the semantic relevance, but also the engineering feasibility, thereby improving the actual success rate of the transformation of scientific and technological achievements.

[0020] 3. This invention introduces a causal reasoning graph network mechanism to construct a local causal graph for candidate matching pairs and perform intervention-based causal strength estimation and pruning. This generates a visualized decision reasoning subgraph containing nodes of technical principles, application scenarios, and engineering gaps, revealing the traceable logical chain behind the matching score. This solves the black box problem of deep learning matching algorithms, improves the interpretability of matching results, reduces the trust barrier between supply and demand sides, and directly empowers technical due diligence work. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the overall system architecture in an embodiment of the present invention.

[0022] Figure 2 This is a flowchart of bilingual domain graph construction and dynamic semantic alignment in an embodiment of the present invention;

[0023] Figure 3 This is a flowchart illustrating the generation of interpretable matching paths based on causal reasoning in an embodiment of the present invention. Detailed Implementation

[0024] The following is in conjunction with the appendix Figure 1-3The specific embodiments of the present invention will be further described below. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0025] The technical solution of the present invention will be further described in detail below with reference to specific embodiments. The embodiments are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.

[0026] System overall architecture implementation method

[0027] In this embodiment, the intelligent matching method for scientific and technological achievements provided by the present invention is implemented collaboratively by a multi-source data acquisition module, a bilingual domain heterogeneous knowledge graph construction module, a dynamic semantic alignment module, an engineering feature extraction module, a graph neural network matching module that integrates technology maturity and manufacturing maturity, a causal reasoning graph network interpretable path generation module, and a matching evaluation report output module. The multi-source data acquisition module is connected to university patent databases, technology transfer platforms, academic literature databases, and enterprise bidding and technology bounty platforms, respectively, to collect the text of scientific and technological achievements and the text of technical requirements and output them to the next stage.

[0028] The bilingual domain heterogeneous knowledge graph construction module constructs academic language subgraphs and industry language subgraphs in parallel from the original text;

[0029] The dynamic semantic alignment module performs cross-modal comparative learning training on the node embeddings of the two subgraph branches, and the two are jointly optimized in an iterative update manner;

[0030] The engineering feature extraction module and the dynamic semantic alignment module output in parallel, and the results of both are input into the graph neural network matching module.

[0031] The graph neural network matching module internally sets up an engineering feasibility gap penalty node, which participates in the matching score calculation as a negative weight;

[0032] The candidate matching pairs output by this module are further input into the causal reasoning graph network interpretable path generation module;

[0033] The final results from each stage are aggregated into the matching evaluation report output module and presented in report form to technology managers, corporate decision-makers, and technology management departments.

[0034] Implementation methods for constructing bilingual domain heterogeneous knowledge graphs and dynamic semantic alignment

[0035] First, we collected texts of scientific and technological achievements and technical requirements from university patent databases, academic literature databases, and enterprise bidding platforms. Then, we performed word segmentation, noise reduction, and format standardization on the collected texts.

[0036] Secondly, a pre-trained large language model is used to perform named entity recognition and relation extraction on the pre-processed scientific and technological achievement text to obtain theoretical entities, mechanism entities and experimental parameter entities. Based on the extracted entities and their relations, an academic language subgraph is constructed, which is denoted as the academic subgraph.

[0037] Meanwhile, a pre-trained large language model is used to perform named entity recognition and relation extraction on the pre-processed technical requirement text to obtain entities of process pain points, mass production indicators, and cost constraints. Based on the extracted entities and their relations, an industry language subgraph is constructed, which is denoted as the industry subgraph.

[0038] Subsequently, knowledge graph embedding encoding is performed on the nodes in the academic and industry subgraphs respectively to obtain the initial vector representation of each node. Anchor point sample sets are constructed based on a small number of existing labeled and aligned term pairs. For each academic entity anchor point, its corresponding positive sample industry entity and several negative sample industry entities are sampled. The contrastive loss function is calculated and the embedding parameters are updated through backpropagation, so that the distance between positive sample pairs in the embedding space is reduced and the distance between negative sample pairs is increased. The contrastive loss function used is as follows:

[0039] ;

[0040] in, An embedding vector representing an academic entity;

[0041] This represents the embedding vector of the positive sample industry entity that is semantically aligned with the academic entity.

[0042] Indicates the first in the candidate industry entity set Embedding vectors of each entity (including positive and negative samples);

[0043] The cosine similarity function represents the similarity between two vectors;

[0044] This represents the temperature coefficient, used to control the smoothness of the probability distribution, and has a preset constant value greater than zero.

[0045] This represents the total number of candidate industry entities participating in the comparison calculation.

[0046] Repeat the above contrastive learning training process until the loss function converges to obtain a unified semantic embedding space after dynamic alignment. The system continuously receives new terminology data and uses an incremental learning mechanism to periodically fine-tune the trained alignment model to adapt to the dynamic evolution of terminology in the scientific and technological field and avoid the decay of semantic alignment effect due to changes in the terminology system.

[0047] In this embodiment, the anchor sample set for contrastive learning can be obtained by domain experts through manual annotation of the correspondence between a small number of academic terms and industry terms. For example, the surface energy regulation mechanism of materials in academic texts can be annotated to be semantically aligned with the insufficient coating adhesion in industry texts. In this way, an initial anchor sample set can be constructed to start the contrastive learning training process.

[0048] Implementation of graph neural network matching based on the maturity of engineering feature extraction and fusion technology and manufacturing maturity.

[0049] In this embodiment, feature fields related to the engineering status are extracted from the patent text, paper text and third-party test report text corresponding to the scientific and technological achievement text. The feature fields include the number of pilot-scale verifications, experimental environment level, mass production equipment dependence and yield data. The extracted engineering feature fields are input into a pre-trained few-shot learning classification model to deduce the technology maturity level and manufacturing maturity level corresponding to each scientific and technological achievement node, thus forming a maturity vector.

[0050] At the same time, based on the technical indicators recorded in the technical requirements document, the technology maturity level threshold and manufacturing maturity level threshold corresponding to the technical requirement node are determined.

[0051] Furthermore, a heterogeneous graph containing result nodes and demand nodes is constructed. The node features are formed by concatenating the node embeddings output from the aforementioned unified semantic embedding space with the maturity vector. The candidate matching relationships between nodes are represented as edges. Multi-layer neighbor aggregation propagation is performed on the node features in this heterogeneous graph to obtain the hidden layer representation of nodes that integrates semantic information and engineering information. The aggregation propagation formula is as follows:

[0052] ;

[0053] in, Indicates the first Layer nodes The hidden layer representation vector;

[0054] Indicates the first Layer trainable weight matrix;

[0055] The neighbor aggregation function can be implemented using either mean aggregation or attention-weighted aggregation.

[0056] Represents a node The set of neighboring nodes in a heterogeneous graph;

[0057] Indicates the first Layer bias terms;

[0058] This represents a non-linear activation function.

[0059] For each candidate matching pair, calculate the maturity difference between the supply-side outcome node and the demand-side node to construct the engineering feasibility gap. The calculation formula is as follows:

[0060] ;

[0061] in, Indicates the supply-side achievement milestones. Indicates the demand-side node;

[0062] This indicates the level of technological maturity of the supplier's deliverables;

[0063] This indicates the threshold for the technology maturity level required by the demander.

[0064] Indicates the manufacturing maturity level of the supplier's output;

[0065] This indicates the threshold for the manufacturing maturity level required by the demander.

[0066] , These are the pre-set weighting coefficients for the difference in technological maturity and the difference in manufacturing maturity.

[0067] The matching score is obtained by weighting the semantic similarity of the node's hidden layer representation with the engineering feasibility gap. The calculation formula is as follows:

[0068] ;

[0069] in, Indicates supply node With demand nodes Match score between them;

[0070] This represents the similarity between the hidden layer representations of two nodes.

[0071] This represents the aforementioned engineering feasibility gap value;

[0072] , The system uses pre-defined semantic similarity trade-off coefficients and engineering gap penalty trade-off coefficients to sort candidate matching pairs in descending order based on matching scores and outputs a candidate matching list.

[0073] Implementation of Explainable Matching Path Generation Based on Causal Inference Graph Networks

[0074] For the candidate matching pairs ranked first in the aforementioned candidate matching list, a local causal graph is constructed. The node types include technical principle nodes, application scenario nodes, engineering gap nodes, and evidence nodes associated with the above three types of nodes. Candidate causal edges are established between nodes according to preset domain prior knowledge.

[0075] For each candidate causal edge in the local causal graph, an intervention operation is performed using a structural causal model. The difference between the expected matching score and the expected score of the factor node under the two intervention states of presence and absence is calculated as the causal strength of the causal edge. The calculation formula is as follows:

[0076] ;

[0077] in, Indicates causal edge The causal strength value;

[0078] The aforementioned matching score is used as the outcome variable in the causal model.

[0079] Indicates the factor node Apply intervention to make it take the value of an existing state;

[0080] Indicates the factor node Apply intervention to make it take the value of non-existent;

[0081] This represents the expected value operator.

[0082] By setting a causal strength threshold, only candidate causal edges with causal strength exceeding the threshold in the local causal graph are retained, resulting in a simplified decision reasoning subgraph. Each path in the decision reasoning subgraph is sorted according to its causal strength, generating a traceable logical chain from the technical principle node through the application scenario node to the engineering gap node. Finally, the decision reasoning subgraph, the strength values ​​of each causal edge, and the traceable logical chain are encapsulated into a matching evaluation report and presented in a visual form for technical managers and corporate decision-makers to review and approve.

[0083] Implementation examples of the overall method flow

[0084] To facilitate understanding, we will take a patent for a battery cathode material proposed by a university and a mass production technology requirement document for battery materials released by a new energy company as examples to illustrate the overall process of this invention. First, we will collect the patent document and the company's requirement document respectively.

[0085] Named entity recognition is performed on the patent text to obtain academic entities related to material mechanisms and experimental parameters, and an academic subgraph is constructed.

[0086] Named entity recognition is performed on enterprise demand text to obtain industry entities involving mass production indicators and cost constraints, and an industry subgraph is constructed.

[0087] Through cross-modal contrastive learning training, the mechanism entity in the patent and the industry pain point entity in the demand are mapped to similar positions in a unified semantic embedding space. Furthermore, feature fields such as the number of pilot verifications and yield rate data are extracted from the experimental data of the patent and third-party testing reports to deduce the technology maturity level and manufacturing maturity level corresponding to the scientific and technological achievement.

[0088] Based on the enterprise's requirement text, the required technology maturity level threshold and manufacturing maturity level threshold are determined. The node embeddings output from the unified semantic embedding space are concatenated with the maturity vector and then input into a graph neural network for propagation. The semantic similarity and engineering feasibility gap are calculated, and a weighted combination is used to obtain a matching score. Based on this, the patent and the requirement are identified as candidate matching pairs. For the candidate matching pairs, a local causal graph is constructed and interventional causal strength estimation is performed. After pruning, a decision reasoning subgraph is obtained, which shows the technical principle fit path between the material mechanism of the patent and the enterprise's mass production pain points, as well as the existing engineering gaps. Finally, a matching evaluation report is generated and output for technical managers to refer to.

[0089] Alternative implementation methods

[0090] In another embodiment of the present invention, in addition to adopting a cross-modal contrastive learning model, the dynamic semantic alignment step can also adopt a cross-graph entity alignment network based on graph attention mechanism. By constructing virtual alignment edges between academic language subgraphs and industry language subgraphs, and using graph attention mechanism to learn cross-graph attention weights between nodes, the semantic alignment effect can be achieved. This alternative method can also achieve semantic alignment between academic language and industry language.

[0091] In another embodiment of the present invention, in addition to using a few-shot learning classification model to deduce the technology maturity level and manufacturing maturity level, the engineering feasibility assessment stage can also adopt a hybrid assessment method based on a rule base and an expert knowledge base. That is, a rule base for judging engineering features labeled by domain experts is pre-built, the extracted engineering feature fields are scored by rule matching, and the scoring results are weighted and fused with the semantic matching score output by the graph neural network to obtain the matching score. This alternative method is suitable for small and medium-scale deployment scenarios.

[0092] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for intelligent matching of scientific and technological achievements based on big data, characterized in that, Includes the following steps: Obtain texts of scientific and technological achievements and technical requirements. An academic language subgraph is constructed based on the text of the scientific and technological achievements, and an industrial language subgraph is constructed based on the text of the technical requirements. Cross-modal comparative learning training is performed on the node embeddings in the academic language subgraph and the industrial language subgraph to obtain a unified semantic embedding space. Based on the text of the scientific and technological achievements, obtain the technology maturity level and manufacturing maturity level corresponding to the nodes of the scientific and technological achievements and form a maturity vector, and obtain the technology maturity level threshold and manufacturing maturity level threshold corresponding to the technology demand nodes. Based on the node embeddings output from the unified semantic embedding space and the maturity vector, a heterogeneous graph containing nodes of scientific and technological achievements and nodes of technological needs is constructed. The heterogeneous graph is then propagated through a graph neural network to obtain the hidden layer representation of the nodes. The similarity between the scientific and technological achievement nodes and the technology requirement nodes is calculated based on the hidden layer representation of the nodes. An engineering feasibility gap is constructed based on the difference between the maturity vector and the technology maturity level threshold and the manufacturing maturity level threshold. The similarity score and the engineering feasibility gap are weighted and combined to obtain a matching score. Candidate matching pairs are determined based on the matching scores, and a local causal graph including technical principle nodes, application scenario nodes, and engineering gap nodes is constructed for the candidate matching pairs. Interventional causal strength estimation is performed on candidate causal edges in the local causal graph; Candidate causal edges with causal strength exceeding a preset threshold are retained to obtain a decision reasoning subgraph; A matching evaluation report is generated and output based on the decision reasoning subgraph.

2. The method according to claim 1, characterized in that, The method for constructing an academic language subgraph based on the scientific and technological achievement text is as follows: Named entity recognition and relation extraction are performed on the scientific and technological achievement text to obtain theoretical entities, mechanism entities and experimental parameter entities, and the academic language subgraph is constructed based on the theoretical entities, the mechanism entities and the experimental parameter entities. The method for constructing an industry language subgraph based on the technical requirement text is as follows: Named entity recognition and relation extraction are performed on the technical requirement text to obtain process pain point entities, mass production indicator entities, and cost constraint entities. The industry language subgraph is then constructed based on the process pain point entities, the mass production indicator entities, and the cost constraint entities.

3. The method according to claim 1, characterized in that, The cross-modal contrastive learning training includes: for academic entity nodes in the academic language subgraph, sampling positive sample industry entity nodes and multiple negative sample industry entity nodes that are semantically aligned with the academic entity node; The contrast loss value is calculated based on the similarity between the embedding vector of the academic entity node and the embedding vector of the positive sample industry entity node, and the similarity between the embedding vector of the academic entity node and each embedding vector in the candidate industry entity node set containing the positive sample industry entity node and the negative sample industry entity node. The node embedding parameters are updated by backpropagation based on the contrastive loss value, which reduces the distance between positive sample pairs and increases the distance between negative sample pairs in the embedding space.

4. The method according to claim 1, characterized in that, Also includes: The system continuously receives new terminology data and fine-tunes the alignment model trained by the cross-modal contrastive learning through incremental learning, thereby updating the unified semantic embedding space.

5. The method according to claim 1, characterized in that, The method for obtaining the technology maturity level and manufacturing maturity level corresponding to the scientific and technological achievement nodes is as follows: Engineering feature fields are extracted from the patent text, paper text and third-party test report text corresponding to the scientific and technological achievement text. The engineering feature fields include the number of pilot-scale verifications, experimental environment level, dependence on mass production equipment and yield rate data. The engineering feature fields are input into a few-shot learning classification model to obtain the technology maturity level and the manufacturing maturity level.

6. The method according to claim 1, characterized in that, The methods for constructing engineering feasibility gaps are as follows: Calculate the absolute value of the difference between the technology maturity level of the scientific and technological achievement node and the technology maturity level threshold, and multiply it by a preset first weighting coefficient; Calculate the absolute value of the difference between the manufacturing maturity level of the technological achievement node and the manufacturing maturity level threshold, and multiply it by a preset second weighting coefficient; add the two product results to obtain the engineering feasibility gap.

7. The method according to claim 1, characterized in that, The weighted combination of the similarity and the engineering feasibility gap is as follows: Multiply the similarity by a preset semantic trade-off coefficient to obtain the first product; The engineering feasibility gap is multiplied by a preset gap trade-off coefficient to obtain a second product; the matching score is obtained by subtracting the second product from the first product.

8. The method according to claim 1, characterized in that, The local causal graph also includes evidence nodes that are associated with the technical principle node, the application scenario node, and the engineering gap node, respectively, and the candidate causal edges are established between the nodes according to preset domain prior knowledge; The method for intervening in the causal strength estimation of candidate causal edges in the local causal graph is as follows: apply an intervention operation with a value of existence and an intervention operation with a value of non-existence to the factor node corresponding to the candidate causal edge, respectively, calculate the difference between the expected values ​​of the matching scores under the two intervention operations, and take the difference as the causal strength of the candidate causal edge.

9. A smart matching system for scientific and technological achievements based on big data, characterized in that, It includes a multi-source data acquisition module, a bilingual domain heterogeneous knowledge graph construction module, a dynamic semantic alignment module, an engineered feature extraction module, a graph neural network matching module, a causal reasoning graph network interpretable path generation module, and a matching evaluation report output module; The multi-source data acquisition module is used to acquire texts of scientific and technological achievements and texts of technical requirements. The bilingual domain heterogeneous knowledge graph construction module is connected to the multi-source data acquisition module and is used to construct an academic language subgraph based on the scientific and technological achievement text and an industrial language subgraph based on the technical demand text. The dynamic semantic alignment module is connected to the bilingual domain heterogeneous knowledge graph construction module and is used to perform cross-modal comparative learning training on the node embeddings in the academic language subgraph and the industrial language subgraph to obtain a unified semantic embedding space. The engineering feature extraction module is used to obtain the technology maturity level and manufacturing maturity level corresponding to the technology achievement node based on the technology achievement text and form a maturity vector, and to obtain the technology maturity level threshold and manufacturing maturity level threshold corresponding to the technology requirement node. The graph neural network matching module is connected to the dynamic semantic alignment module and the engineering feature extraction module, respectively. It is used to construct a heterogeneous graph containing nodes of scientific and technological achievements and nodes of technical needs based on the node embeddings output by the unified semantic embedding space and the maturity vector. The graph neural network propagation is performed on the heterogeneous graph to obtain the hidden layer representation of the nodes. The matching score is calculated based on the hidden layer representation of the nodes and the maturity vector, and candidate matching pairs are determined. The causal reasoning graph network interpretable path generation module is connected to the graph neural network matching module. It is used to construct a local causal graph including technical principle nodes, application scenario nodes and engineering gap nodes for the candidate matching pairs. It performs intervention-type causal strength estimation on the candidate causal edges in the local causal graph, retains the candidate causal edges whose causal strength exceeds a preset threshold, and obtains a decision reasoning subgraph. The matching evaluation report output module is connected to the causal reasoning graph network interpretable path generation module, and is used to generate and output a matching evaluation report based on the decision reasoning subgraph.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the intelligent matching method for scientific and technological achievements based on big data as described in any one of claims 1 to 8.

Images