Bidirectional granularity alignment method for university achievements and enterprise demands based on large model

Abstract

The invention discloses a college achievement and enterprise demand bidirectional granularity alignment method based on a large model. The method comprises the steps that an original data set of college scientific research achievements is collected and processed; performing fine-grained technical point disassembly on the abstract in the processed data; technical points required by enterprises are disassembled; matching and aligning technical points based on double-path representation fusion; and carrying out matching reordering and recommendation output. According to the method, a unified five-level technology level framework is established, granularity alignment of university achievements and enterprise requirements under the same technical standard is realized, and the matching accuracy is greatly improved. By means of a large model and cue word engineering, achievement technical feature aggregation and demand disassembly are completed, a traditional manual processing mode is replaced, and the docking efficiency is remarkably improved. Through the steps of author clustering, multi-dimensional feature weighted sorting and the like, it is ensured that the recommendation result has technical relevance and also conforms to the practical application value, and an efficient solution is provided for industry-university-research collaborative innovation.

Description

Technical Field

[0001] This invention relates to the field of information retrieval, specifically to a two-way granular alignment method for university research results and enterprise needs based on a large model, which is suitable for scenarios involving precise matching and efficient connection between university research results and enterprise technology needs. Background Technology

[0002] In the process of industry-academia-research collaborative innovation, the effective connection between university research results and enterprise technology needs is a core link. Currently, university research results are characterized by fragmented data formats and unclear technical features, while enterprise needs are vaguely expressed and lack structured breakdown, resulting in information asymmetry between the two.

[0003] Traditional matching methods rely heavily on keyword retrieval and lack a unified technical level standard, making it impossible to achieve precise alignment of technical points. This results in low matching accuracy and poor connection efficiency, making it difficult to transform and implement a large number of valuable scientific research results, and failing to meet the technical needs of enterprises in a timely manner, which seriously restricts the promotion of industry-university-research collaborative innovation.

[0004] Based on this, the present invention provides a bidirectional granular alignment method based on a large model, which achieves accurate matching and efficient connection between the two, and helps industry-academia-research collaborative innovation. Summary of the Invention

[0005] This invention aims to address the problems of low matching accuracy and poor matching efficiency in the existing connection between university research results and enterprise needs. It provides a two-way granular alignment method based on a large model to achieve accurate matching and efficient connection between the two, thereby facilitating collaborative innovation between industry, academia and research.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A two-way granular alignment method based on large-scale models between university research findings and enterprise needs includes: Perform the following steps using a computer device. Step 1: Collect and process the raw datasets of university research findings; Step 2: Decompose the abstract of the processed data into fine-grained technical points; Step 3: Break down the technical requirements of the enterprise; Step 4: Matching and aligning technical points based on dual-path representation fusion; Step 5: Matching, reordering, and recommending output.

[0007] Furthermore, the method for processing the original dataset of university research results in step 1 is as follows: Based on regularization rules, the target document data is structured and parsed to accurately extract the entity feature information associated with the document, including key attribute fields such as the pinyin identifier corresponding to the author's name and the English name of the college to which the author belongs, forming a feature set of the entity to be matched. Collect and integrate the name information of teachers from various colleges of the target universities, establish a standardized benchmark entity database as a reference benchmark for entity alignment, and ensure the integrity and authority of the benchmark data; Based on the pinyin mapping rules, the author pinyin features extracted from the documents are accurately matched with the standard pinyin of the teacher names in the benchmark entity database; combined with the English-Chinese correspondence rules of college names, the consistency of the English college names in the documents with the college affiliation information in the benchmark entity database is verified. By combining the above dual matching results, the document entities are accurately aligned with the target entities in the benchmark entity database, achieving entity normalization across data sources.

[0008] Furthermore, in step 2, the summary of the processed data is broken down into fine-grained technical points: A five-level technical hierarchy framework is defined, including Level 1: Macro-level technical domain; Level 2: Core domain direction; Level 3: Specific task of direction; Level 4: Task implementation method; Level 5: Method optimization scheme, and the semantic boundaries and classification rules of each level are defined, focusing on the decomposition target with Level 5 as the endpoint; By performing deep semantic analysis on the abstract text using a large model, key features such as technical terms, method descriptions, and functional characteristics contained in the text are mined and extracted to form a structured set of technical features, providing data support for hierarchical mapping. Leveraging the cross-domain knowledge graph and semantic matching capabilities of the large model, the extracted technical features are mapped step by step to the fifth-level technical points; based on the fine-grained technical point decomposition results, a user, technical point, and frequency triplet dataset is constructed. A dual-path graph neural network representation learning method is adopted. The first path learns the structural representation of users and technical points based on the user-technical point bipartite graph, and the second path learns the semantic representation of technical points based on the technical point similarity graph. Finally, the comprehensive representation of technical points is obtained by fusion through neural networks.

[0009] Furthermore, the technical breakdown of the enterprise's requirements in step 3 is as follows: Based on the five levels defined above, a prompt word project was designed, and the output of the large model was strictly standardized to the fifth level among the five levels, thereby realizing fine-grained technical point analysis of enterprise needs.

[0010] Further, step 4 includes: After completing the five-level technical point breakdown for both enterprises and universities, the fifth-level technical point is taken as the core. Based on the learned user representation and comprehensive technical point representation, the number of user interactions on each technical point is predicted as the professional proficiency score. By calculating the representation similarity between the technical points required by enterprises and the technical points of universities, and combining it with the user's professional proficiency score, a precise two-way matching at the granular level is achieved.

[0011] Furthermore, the matching reordering and recommendation output in step 5 includes: First, the matching results are clustered and integrated according to the author dimension to form a matching set with the author as the unit. Then, this set is input into the re-ranking model, and a multi-dimensional feature weighting strategy is adopted to integrate core indicators such as semantic similarity of technical points, user professional proficiency score, the scale of the number of technical points associated with the author, and the quantitative score of the achievement maturity of the university to which the author belongs, to complete the optimization and ranking of the matching results. Finally, the ranking results are input into the large language model to generate recommendation interpretations with technical relevance and application value.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. A unified five-level technical hierarchy framework was established, which enabled the granular alignment of university research results and enterprise needs under the same technical standards, greatly improving the accuracy of matching and avoiding the ambiguity problem of traditional keyword retrieval.

[0013] 2. By leveraging large-scale models and prompt word engineering, we can aggregate the technical features of deliverables and decompose requirements, replacing traditional manual processing methods and significantly improving the efficiency of the process.

[0014] 3. Through steps such as author clustering and multi-dimensional feature weighted ranking, the recommendation results are ensured to be both technically relevant and of practical application value, providing an efficient solution for collaborative innovation between industry, academia and research. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the overall process of the method of the present invention. Detailed Implementation

[0016] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only 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.

[0017] The present invention discloses a method for aligning university research findings with enterprise needs at a bidirectional granularity based on a large model. This method includes the following steps: Step 1: Collect and process the raw datasets of scientific research results from universities to form a structured set of entity features, providing a data foundation for subsequent technical point decomposition; Step 2: Based on the processing results of Step 1, the summary in the data is decomposed into fine-grained technical points to generate a triplet dataset of user-technical point-frequency, and the technical point representation is learned. Step 3: Parallel to Step 2, break down the enterprise's requirements into technical points and output the fifth level of technical points to ensure that they are consistent with the enterprise's standards. Step 4: Based on the outputs of Step 2 and Step 3, calculate the similarity of technical points through dual-path representation fusion, and combine it with the user's professional proficiency score to achieve granular bidirectional matching and alignment. Step 5: Reorder and recommend the matching results from Step 4, cluster by author, fuse multidimensional features to optimize the sorting, and generate interpretable recommendation explanations.

[0018] Specifically, the method for processing the original dataset of university research results in step 1 is as follows: Based on regularization rules, the target document data is structured and parsed to accurately extract the entity feature information associated with the document, including key attribute fields such as the pinyin identifier corresponding to the author's name and the English name of the college to which the author belongs, forming a feature set of the entity to be matched.

[0019] The system collects and integrates the name information of teachers from various colleges of the target universities, establishes a standardized benchmark entity database as a reference benchmark for entity alignment, and ensures the integrity and authority of the benchmark data.

[0020] Based on the pinyin mapping rules, the author pinyin features extracted from the documents are accurately matched with the standard pinyin of the teacher names in the benchmark entity database; combined with the English-Chinese correspondence rules of college names, the consistency of the English college names in the documents with the college affiliation information in the benchmark entity database is verified. By combining the above dual matching results, the document entities are accurately aligned with the target entities in the benchmark entity database, achieving entity normalization across data sources.

[0021] In step 2, the abstract of the processed data is broken down into fine-grained technical points: A clear five-level technology hierarchy framework is established (Level 1: Macro-level technology domain; Level 2: Core domain direction; Level 3: Specific task of direction; Level 4: Task implementation method; Level 5: Method optimization scheme), and the semantic boundaries and classification rules of each level are defined, focusing on the decomposition target with Level 5 as the endpoint.

[0022] First, this method processes two types of input information in parallel. As shown in the flowchart, on the one hand, a large model is used to perform deep semantic parsing on the abstracts of university research results, and based on a pre-defined five-level technical hierarchy framework, the technical content described in the text is decomposed into a set of fine-grained technical points. On the other hand, the same technical standard framework is used to parse enterprise requirement documents, also outputting standardized fifth-level technical points. Based on this, the system traverses all documents of each author, counts the frequency of each technical point, and generates a (user, technical point, frequency) triplet dataset. The output of this step (triplet data) provides a direct data foundation for subsequent association mining. Specifically, the technical point decomposition formula is: (1) Breakdown of technical points:

[0023] TechPoints(d) is the output of the model, representing the set of technical points extracted from the abstract of document d. LLM decompose The large language model represents the task of decomposing technical points. Abstract(d) represents the abstract text corresponding to document d, which is the input to the technical point decomposition model. d represents a document. fn-gand This represents a five-level granularity decomposition prompt.

[0024] (2) Triple generation:

[0025] in D represents the number of times author a appears at technical point t. a This represents the set of all publications by author a. I() is an indicator function, which has a value of 1 when the condition in parentheses is true, and 0 otherwise.

[0026] After obtaining the triplet data, the system enters the representation learning phase. This phase consists of two parallel paths, aiming to learn vector representations of technologies and users: The first part is structural representation learning, which constructs a user-technology point bipartite graph based on the triplet data generated in the first step. Through the message passing mechanism of a graph neural network (GNN), neighbor information is aggregated to learn user representations and technology point structural representations that reflect the cooperation network and interest preferences. The second part is semantic representation learning, which uses a large model to calculate the semantic similarity between pairs of technology points, constructing a semantic similarity graph of technology points. Similarly, a graph neural network is used to learn on this graph to obtain the semantic representations of the technology points.

[0027] The formula for learning the user-technology point bipartite graph is:

[0028]

[0029] in , Let N(t) and N(a) represent the node representation vectors of technique point t and author a in the l-th layer of the graph neural network, respectively. N(t) and N(a) represent the sets of all author neighbor nodes connected to technique point t and all technique point neighbor nodes connected to author a in the bipartite graph, respectively. t D a Let t and a represent the node degrees (i.e. the number of connecting edges) of the technical point t and the author a, respectively. Let σ be the trainable weight matrix of the l-th layer of the graph neural network, and σ be the non-linear activation function.

[0030] The formula for learning the similarity graph of technical points is as follows:

[0031]

[0032] in Let be the attention coefficient between technical point t and its neighbor t', and a be a trainable attention vector used to calculate the correlation between the representations of the two nodes.

[0033] The formula for constructing the similarity matrix is:

[0034] in For semantic similarity between technical points, t i ,t j These represent the i-th and j-th technical points, respectively. LLM_similarity represents the semantic similarity calculation prompt word, and LLM_similarity represents the large language model performing the semantic similarity evaluation task. This represents the semantic similarity measurement prompt.

[0035] The technical breakdown of the enterprise's needs in step 3 is as follows: Based on the five levels defined above, a prompt word project was designed, and the output of the large model was strictly standardized to the fifth level among the five levels, thereby realizing fine-grained technical point analysis of enterprise needs.

[0036] Step 4 includes: After breaking down the technical points into five levels for both enterprises and universities, the fifth-level technical point is used as the core. Based on the learned user representation and comprehensive technical point representation, the number of user interactions on each technical point is predicted as a professional proficiency score. By calculating the representational similarity between the technical points required by enterprises and those of universities, the structural and semantic representations of the technical points are fused to obtain the final comprehensive representation of the technical points. Combined with the user's professional proficiency score, granular, two-way precise matching is achieved.

[0037] The formula for representation fusion and proficiency prediction is: (1) Representation fusion:

[0038] MLP fuse It is a multilayer perceptron. This refers to the technology point structure representation vector learned from the user-technology point bipartite graph. The semantic representation vector of the technology point learned from the semantic graph of the technology point.

[0039] (2) Proficiency prediction:

[0040] in The model's predicted output represents the predicted proficiency score of author a at technical point t. The author-user representation vector learned from the user-technology point bipartite graph. This is the final comprehensive representation vector of the technique point.

[0041] Step 5, the matching reordering and recommendation output, includes: First, the matching results are clustered and integrated according to the author dimension, forming a matching set with the author as the unit. Then, this set is input into the re-ranking model, employing a multi-dimensional feature weighting strategy to integrate core indicators such as semantic similarity of technical points, user professional proficiency score, the scale of the number of technical points associated with the author, and the quantitative score of the maturity of the achievements of the affiliated university, to optimize the ranking of the matching results. Finally, the ranked results are input into a large language model to generate recommendation explanations with technical relevance and application value. The re-ranking algorithm formula is:

[0042] The constraints are as follows: + + =1, with preferred weights of 0.5, 0.3, and 0.2, R final The final matching recommendation score, For semantic similarity of technical points, To quantify the maturity of the results, This represents the predicted professional proficiency score of author a at the corresponding technical point t.

[0043] The optimal parameter range for this implementation is shown in the table below:

[0044] In this implementation, the key hyperparameters involved in model training and their functions are as follows: Parameter d user (User representation dimension) and d tech The length of the node embedding vector in the above graph neural network is determined by the (technical point representation dimension), which directly affects the model's ability to characterize the complex features of users and technical points; the parameter L (number of graph neural network layers) controls the depth of model information aggregation. The more layers there are, the more extensive high-order neighborhood information the nodes can capture.

[0045] The following are examples: I. Data 1. Enterprise requirements input Enterprise core needs: Improve the accuracy of personalized recommendation systems, involving three key technical requirements: Technical Point 1: User Behavior Data Collection and Cleaning Technology Technical Point 2: User Feature Extraction and Profile Modeling Technology Technical Point 3: Recommendation Algorithm Optimization and A / B Testing Techniques 2. Input of data on university research achievements Target literature: "Genetic Algorithm for Bayesian Knowledge Tracing: A Practical Application" published by Professor Hu Xuegang (Author ID: 23090636200) Basic information about the publication: The institution is the School of Computer Science and Information; the research fields are artificial intelligence and machine learning; the core application scenario is personalized learning systems for online education; and the publication year is 2022 (indexed by Scopus). II. Operation Process Step 1: Data Processing of University Scientific Research Results Regularization rules based on the above method and verification with the benchmark entity library: Extracting document entity features: Author's pinyin identifier (Hu, XueGang), and the English name of the affiliated college (School of Computer and Information). Matching with the standardized benchmark entity library: The author was confirmed as Professor Hu Xuegang through the pinyin mapping rules, and the consistency check of the college affiliation information passed, thus completing the entity normalization.

[0046] Step 2: Fine-grained breakdown of technical points in university research abstracts A five-level technical hierarchy framework is adopted, and the literature abstract is decomposed through a large model + prompt word engineering approach: First-level macro-level technology field: Artificial Intelligence Second-level core area: Machine Learning Specific task at the third level: Personalized recommendation system Level 4 task implementation method: Genetic algorithm Five-level optimization scheme: Multi-skill BKT based on genetic algorithm Generate triplet datasets: (Hu Xuegang, Genetic Algorithm, 1), (Hu Xuegang, Personalized Recommendation System, 1), (Hu Xuegang, User Feature Matching, 1) Step 3: Breaking down the technical requirements of the enterprise The prompts are designed according to a five-level technical hierarchy, forcibly outputting the fifth-level technical points: Enterprise requirements breakdown results: (User behavior data cleaning solution, 1), (User profile feature extraction algorithm, 1), (Recommendation algorithm optimization strategy, 1) Step 4: Dual-path representation fusion and matching Technical point representation learning: Through a dual-path graph neural network, we learn Professor Hu Xuegang's user structure representation (based on user-technical point bipartite graph) and the semantic representation of technical points such as genetic algorithms (based on technical point similarity graph), and then fuse them to obtain a comprehensive representation.

[0047] Similarity calculation and proficiency scoring: Semantic similarity: Matching degree between enterprise demand technology points and university technology points (User behavior data cleaning - genetic algorithm data preprocessing module: 0.85; User feature extraction - personalized feature matching: 0.83; Recommendation algorithm optimization - genetic algorithm optimization strategy: 0.80) Professional proficiency score: Based on Professor Hu Xuegang's triplet counts on relevant technical points, the predicted score is 0.82. Step 5: Matching Reordering and Recommendation Output A multi-dimensional feature weighting strategy is adopted (semantic similarity weight λ1=0.5, proficiency weight λ2=0.3, achievement maturity weight λ3=0.2): Author clustering and integration: forming a matching set with Professor Hu Xuegang as the core. Quantification of core indicators: Mean semantic similarity of technical points: / 3=0.827 User's professional proficiency score: 0.82 Results maturity metric score: 0.93 (based on transformation experience) Final matching score calculation: 0.827×0.5+0.82×0.3+0.93×0.2=0.824 (i.e. 82.4%) Recommendation Interpretation Generation: Professor Hu Xuegang has made significant contributions to user preference analysis and collaborative filtering recommendation. His patent, "A Collaborative Filtering Recommendation Method Based on User Preferences," effectively improves the accuracy of personalized recommendation systems through NHSM similarity calculation and the construction of a new similarity matrix. The research involves techniques such as genetic algorithm optimization and user feature matching, which are highly compatible with the user behavior data processing, feature extraction, and recommendation algorithm optimization needs of enterprises. He has extensive experience in technology transfer and can provide theoretical support and practical application solutions for technological requirements.

[0048] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.

[0049] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.

[0050] In another embodiment provided in this application, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute any of the mobile source emission prediction methods based on time-series feature migration described in the above embodiments.

[0051] It is understood that the systems, devices, and storage media provided in the embodiments of the present invention correspond to the methods provided in the embodiments of the present invention, and the explanations, examples, and beneficial effects of the relevant content can be referred to the corresponding parts of the above methods.

[0052] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0053] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0054] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0055] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for bidirectional granular alignment of university research findings and enterprise needs based on a large model, characterized in that, Includes the following steps, Step 1: Collect and process the raw datasets of scientific research results from universities to form a structured set of entity features, providing a data foundation for subsequent technical point decomposition; Step 2: Based on the processing results of Step 1, the summary in the data is decomposed into fine-grained technical points to generate a triplet dataset of user-technical point-frequency, and the technical point representation is learned. Step 3: Decompose the enterprise's needs into technical points and output the fifth level of technical points to ensure that they are consistent with the enterprise's standards. Step 4: Based on the outputs of Step 2 and Step 3, calculate the similarity of technical points through dual-path representation fusion, and combine it with the user's professional proficiency score to achieve granular bidirectional matching and alignment. Step 5: Reorder and recommend the matching results from Step 4, cluster by author, fuse multidimensional features to optimize the sorting, and generate interpretable recommendation explanations.

2. The method for bidirectional granular alignment of university research findings and enterprise needs based on a large model, as described in claim 1, is characterized in that: The specific method for processing the original dataset of university research results in step 1 is as follows: Based on regularization rules, the target document data is structured and parsed to accurately extract the entity feature information associated with the document, including key attribute fields such as the pinyin identifier corresponding to the author's name and the English name of the college to which the author belongs, forming a feature set of the entity to be matched. Collect and integrate the name information of teachers from various colleges of the target universities, establish a standardized benchmark entity database as a reference benchmark for entity alignment, and ensure the integrity and authority of the benchmark data; Based on the pinyin mapping rules, the author pinyin features extracted from the documents are accurately matched with the standard pinyin of the teacher names in the benchmark entity database; combined with the English-Chinese correspondence rules of college names, the consistency of the English college names in the documents with the college affiliation information in the benchmark entity database is verified. By combining the above dual matching results, the document entities are accurately aligned with the target entities in the benchmark entity database, achieving entity normalization across data sources.

3. The method for bidirectional granular alignment of university research findings and enterprise needs based on a large model, as described in claim 2, is characterized in that: In step 2, the abstract of the processed data is broken down into fine-grained technical points: A clear five-level technology hierarchy framework is established, including Level 1: Macro-level technology domain; Level 2: Core areas of expertise; Level 3: Specific tasks and directions; Level 4: Task Implementation Methods; Level 5: Method optimization scheme, and define the semantic boundaries and classification rules of each level, focusing on the decomposition target with the 5th level as the endpoint; By performing deep semantic analysis on the abstract text using a large model, key features such as technical terms, method descriptions, and functional characteristics contained in the text are mined and extracted to form a structured set of technical features, providing data support for hierarchical mapping. Based on the cross-domain knowledge graph and semantic matching capabilities of the large model, the extracted technical features are mapped to the fifth level technical points step by step; Based on the results of fine-grained technical point decomposition, a dataset of triplet groups of user, technical point, and frequency is constructed. A dual-path graph neural network representation learning method is adopted. The first path learns the structural representation of users and technical points based on the user-technical point bipartite graph, and the second path learns the semantic representation of technical points based on the technical point similarity graph. Finally, the comprehensive representation of technical points is obtained by fusion through neural networks.

4. The method for bidirectional granular alignment of university research findings and enterprise needs based on a large model, as described in claim 3, is characterized in that: Step 2 involves fine-grained breakdown of technical points to generate a user-technical-frequency triplet dataset. the following: (1) Breakdown of technical points: Where TechPoints(d) is the output of the model, representing the set of technical points extracted from the abstract of document d; LLM decompose The large language model represents the task of decomposing technical points. Abstract(d) represents the abstract text corresponding to document d, which is the input to the technical point decomposition model. d represents a document, P fn-gand The key words represent the breakdown of the five technical points; (2) Triple generation: in D represents the number of times author a appears at technical point t. a This represents the set of all publications by author a. I() is an indicator function, which has a value of 1 when the condition in parentheses is true, and 0 otherwise.

5. The method for bidirectional granular alignment of university research findings and enterprise needs based on a large model, as described in claim 4, is characterized in that: In step 2, technical point representations are learned. Includes the following steps, After obtaining the triplet data, the system enters the representation learning phase; This phase is divided into two parallel paths, with the aim of learning vector representations of technology and users: the first part is structural representation learning, which constructs a bipartite graph of user-technology points based on the triplet data generated in the first step; The first part is semantic representation learning. By aggregating neighbor information through the message passing mechanism of graph neural networks, user representations and technical point structure representations that can reflect the cooperative network and interest preferences are learned. The second part is semantic representation learning. The semantic similarity between pairs of technical points is calculated using a large model to construct a semantic similarity graph of technical points. Similarly, graph neural networks are used to learn on this graph to obtain the semantic representation of technical points. The user-technology point bipartite graph learning formula is as follows: in , Let N(t) and N(a) represent the node representation vectors of technique point t and author a in the l-th layer of the graph neural network, respectively. N(t) and N(a) represent the sets of all author neighbor nodes connected to technique point t and all technique point neighbor nodes connected to author a in the bipartite graph, respectively. t D a Let represent the node degree (i.e., the number of connecting edges) of technical point t and author a, respectively. Let be the trainable weight matrix of the l-th layer graph neural network, and σ be the non-linear activation function; And it also includes the formula for learning the similarity graph of technical points: in Let be the attention coefficient between technical point t and its neighbor t', and let a be a trainable attention vector used to calculate the correlation between the representations of the two nodes. The formula for constructing the similarity matrix is: in For semantic similarity between technical points, t i ,t j These represent the i-th and j-th technical points, respectively. LLM_similarity represents the semantic similarity calculation prompt word, and LLM_similarity represents the large language model performing the semantic similarity evaluation task. These are the prompts for semantic similarity calculation.

6. The method for bidirectional granular alignment of university research findings and enterprise needs based on a large model, as described in claim 5, is characterized in that: The technical breakdown of the enterprise's needs in step 3 is as follows: Based on the five levels defined above, a prompt word project was designed, and the output of the large model was strictly standardized to the fifth level among the five levels, thereby realizing fine-grained technical point analysis of enterprise needs.

7. The method for bidirectional granular alignment of university research findings and enterprise needs based on a large model, as described in claim 6, is characterized in that: Step 4 includes: After completing the five-level technical point breakdown for both enterprises and universities, the fifth-level technical point is taken as the core. Based on the learned user representation and comprehensive technical point representation, the number of user interactions on each technical point is predicted as the professional proficiency score. By calculating the representation similarity between the technical points required by enterprises and the technical points of universities, and combining it with the user's professional proficiency score, a precise two-way matching at the granular level is achieved.

8. The method for bidirectional granular alignment of university research findings and enterprise needs based on a large model, as described in claim 7, is characterized in that: Step 4 involves calculating the similarity of technical points through dual-path representation fusion and combining it with the user's professional proficiency score to achieve granular bidirectional matching and alignment, including: The formula for representation fusion and proficiency prediction is: (1) Representation fusion: MLP fuse It is a multilayer perceptron. This refers to the technology point structure representation vector learned from the user-technology point bipartite graph. The semantic representation vector of technical points learned from the semantic graph of technical points; (2) Proficiency prediction: in The model's predicted output represents the predicted proficiency score of author a at technical point t. The author-user representation vector learned from the user-technology point bipartite graph. This is the final comprehensive representation vector of the technique point.

9. The method for bidirectional granular alignment of university research findings and enterprise needs based on a large model, as described in claim 8, is characterized in that: Step 5, the matching reordering and recommendation output, includes: First, the matching results are clustered and integrated according to the author dimension to form a matching set with the author as the unit. Then, this set is input into the re-ranking model, and a multi-dimensional feature weighting strategy is adopted to integrate core indicators such as semantic similarity of technical points, user professional proficiency score, the scale of the number of technical points associated with the author, and the quantitative score of the achievement maturity of the university to which the author belongs, to complete the optimization and ranking of the matching results. Finally, the ranking results are input into the large language model to generate recommendation interpretations with technical relevance and application value.

10. The method for bidirectional granular alignment of university research findings and enterprise needs based on a large model, as described in claim 9, is characterized in that: Step 5 also includes, The reordering algorithm formula is: The constraints are as follows: + + =1; R final The final matching recommendation score, For semantic similarity of technical points, To quantify the maturity of the results, This represents the predicted professional proficiency score of author a at the corresponding technical point t.

Images