Bidirectional matching method and system for scientific research talents and R&D tasks based on graph representation

Abstract

The present invention relates to the field of graph representation data analysis technology, and in particular to a method and system for bidirectional matching of scientific research talents and R&D tasks based on graph representation. A knowledge graph is constructed, and the knowledge graph is pre-trained based on the weight attributes of scientific research talents and the weight attributes of R&D tasks, and the node relationships hidden in the knowledge graph are captured to obtain a representation method of the knowledge graph; the characteristic data of scientific researchers are integrated and the relationship is mined according to the representation method of the knowledge graph to obtain a fusion representation of the professional and technical capabilities of scientific researchers; a bidirectional LSTM encoder is used to independently encode meta-paths from the perspective of scientific research talents and the perspective of R&D tasks, and knowledge graph information is modeled from the perspective of scientific research talents and the perspective of R&D tasks; a bidirectional recommendation training is performed through a bilateral quadruple loss function to obtain a bidirectional matching model, and this method can improve the recommendation matching accuracy of scientific research talents and R&D tasks.

Description

Technical Field

[0001] The present invention relates to the field of graph representation data analysis technology, and in particular to a method and system for bidirectional matching of scientific research talents and R&D tasks based on graph representation. Background Art

[0002] Scientific research is a scientific research model that can cross the boundaries of university disciplines and efficiently integrate internal advantages of multidisciplinary resources to carry out task-oriented research. Through the transformation and innovation of university scientific research paradigms and organizational models, and by integrating resources inside and outside the school, scientific research tasks with clear goals can be carried out in a more flexible organizational form. However, there are large differences between different disciplines and industries, and the knowledge data is huge and complex. The efficiency of scientific research organization is low by manual means. Therefore, it is necessary to use big data mining and other emerging artificial intelligence technologies for two-way matching to improve the matching accuracy. Taking the organized scientific research service platform of Sun Yat-sen University as an example, by automatically extracting and mining the features of the horizontal projects, vertical projects, scientific research talents and their academic achievements accumulated over many years, the scientific research image of the school's talents can be better grasped, so that relevant enterprises can query the relevant achievements of interest. However, due to the huge amount of data in technical texts, it is urgent to improve the matching accuracy through technical means such as data mining. In addition, the two-way recommendation system is different from the traditional one-way recommendation system. They are designed to meet the needs of both parties involved in the recommendation process at the same time. In the massive data, the data of industry-university-research cooperation has a very prominent problem of sparsity of interactive data. In view of the limitations of existing methods, we proposed a bidirectional method and system for scientific research talents and R&D tasks based on graph representation. Summary of the invention

[0003] The present invention overcomes the sparsity problem of interactive data in existing two-way recommendation systems, and provides a two-way matching method and system for scientific research talents and R&D tasks based on graph representation.

[0004] To achieve the above object, the present invention discloses a bidirectional matching method between scientific research talents and R&D tasks based on graph representation, comprising the following steps:

[0005] Acquire characteristic data of scientific research talents in a big data network, perform keyword extraction processing on the characteristic data to obtain a number of characteristic keywords, and determine weight attributes of scientific research talents according to the characteristic keywords;

[0006] Acquire the R&D task, use the prompt template to obtain the key technical words of the R&D task in the large language model as attributes and return them, and determine the weight attribute of the R&D task according to the returned order;

[0007] Construct a knowledge graph, and pre-train the knowledge graph based on the weight attributes of scientific research talents and the weight attributes of R&D tasks, capture the hidden node relationships of the knowledge graph, and obtain a representation method of the knowledge graph;

[0008] According to the representation method of knowledge graph, the characteristic data of scientific researchers are integrated and the relationship is mined to obtain the integrated representation of the professional and technical capabilities of scientific researchers;

[0009] Use a bidirectional LSTM encoder to independently encode meta-paths from the perspective of scientific research talents and the perspective of R&D tasks, so as to model knowledge graph information from the perspective of scientific research talents and the perspective of R&D tasks, and obtain the representation of scientific research talents' capabilities and the text representation of R&D tasks;

[0010] Performing two-way recommendation training through a bilateral quadruple loss function to obtain a two-way matching model, and performing two-way matching between scientific research talents and R&D tasks according to the two-way matching model;

[0011] Among them, the characteristic data include horizontal projects undertaken and participated by scientific research talents and the patents and paper results obtained.

[0012] More specifically, the knowledge graph is pre-trained based on the weight attributes of scientific research talents and the weight attributes of R&D tasks, the node relationships hidden in the knowledge graph are captured, and the representation method of the knowledge graph is obtained, which also includes:

[0013] For the i-th node entity e in the knowledge graph i , through multi-layer graph convolutional network encoding, the transmission of adjacency information between entities is realized:

[0014]

[0015] In the formula, is the node entity e at the l+1th layer i express; It is a relationship ij The weight coefficient of g are the parameters learned by the network; σ(·) is an activation function; is the node entity e at layer l i Indicates; n is the node entity e i The total number of .

[0016] Also includes:

[0017] Introducing convolutional TransE decoding into the knowledge graph building module;

[0018] For convolutional TransE decoding, the kernel mapping representation m is obtained based on C kernels of different sizes. c , and each kernel map represents m cThe concatenation constructs a comprehensive representation M of multi-scale features C ;

[0019] The scoring function of the convolution TransE method after nonlinear convolution is defined as follows:

[0020] θ(e s , e o )=f(vec(M(e s , e r ))W c ) o

[0021] In the formula, θ(e s , e o ) represents a scoring function used to evaluate e s , e o The rationality of the relationship between s , e o are the head node and tail node embeddings, e r is the corresponding relation embedding; W c is the weight coefficient of the convolution TransE decoding; M(e s , e r ) represents the s , e r Perform concatenation operations to output a matrix or high-dimensional tensor; vec(·) represents a vectorization operation to flatten the input matrix or high-dimensional tensor into a vector; f(·) is a nonlinear activation function.

[0022] Also includes:

[0023] During the training of the convolutional TransE model, the logistic sigmoid function is applied to the score:

[0024] p(e s , e r , e o )=σ(θ(e s , e o ))

[0025] In the formula, p(e s , e r , e o ) means e s , e r , e o The probability of the triple being true, the output value range is between 0 and 1, indicating the possibility of the triple existing in the knowledge graph; s , e o are the head node and tail node embeddings, e r is the corresponding relation embedding; θ(es , e o ) represents a scoring function used to evaluate e s , e o The rationality of the relationship between them; σ(·) is an activation function.

[0026] More specifically, the use of a bidirectional LSTM encoder to independently encode meta-paths from the perspective of scientific research talent and the perspective of R&D tasks also includes:

[0027] For the i-th node entity e on the path i , which is fed into a bidirectional LSTM encoder to learn the contextual representation:

[0028]

[0029] In the formula, and are the model encoding forward and backward hidden states of the i-th step; e i Represents the i-th node entity; Indicates that the model encodes the i-1th step forward hidden state; represents the model encoding the i+1th backward hidden state; f(·) is the function of the LSTM unit;

[0030] Finally, the aggregated representation of the two views is calculated by averaging the hidden states of the bidirectional LSTM encoder from both directions:

[0031]

[0032] In the formula, p c and p t are the meta-path fusion representations of talents and scientific research tasks respectively; m is the actual length of the meta-path, that is, the number of node entities contained; and are the model encoding forward and backward hidden states of the i-th step respectively.

[0033] More specifically, the scientific research talent ability representation and the R&D task text representation are input into the training model for training, and the training and validation data sets are split in the ratio of 8:2.

[0034] To achieve the above-mentioned purpose, the present invention further discloses a two-way matching system of scientific research talents and R&D tasks based on graph representation, which is applied to any one of the two-way matching methods of scientific research talents and R&D tasks, and comprises:

[0035] Scientific research talent information database: saves the horizontal and vertical project data of scientific research talents, as well as the data of scientific research talents and their academic achievements; stores them according to different data types for specific data preprocessing modules to call and query;

[0036] Horizontal project preprocessing module: responsible for extracting keywords from the horizontal projects of researchers, including the project's brief introduction, name, and core achievement name;

[0037] Longitudinal project preprocessing module: responsible for extracting keywords from the longitudinal projects of researchers, including the project's achievement profile, name, and core achievement name;

[0038] Paper results preprocessing module: responsible for extracting keyword information from the names, keywords and abstracts of researchers' papers;

[0039] Patent results preprocessing module: responsible for extracting keyword information from the patent names, keywords and abstracts of researchers’ patents;

[0040] Scientific research task input module: The input module of the system is used by users to input the scientific research task requirements that need to be matched. The system matches researchers according to the requirements and returns the matching results in order of matching degree from high to low.

[0041] Also includes:

[0042] Prompt information generation module: responsible for managing the information templates that call the large language model interface, generating query instructions for R&D tasks according to the format requirements of the prompt templates, and returning the query instructions to the scientific research task information preprocessing module;

[0043] Scientific research task information preprocessing module: obtains key technical words of R&D tasks as attributes in the large language model;

[0044] Knowledge graph construction module: responsible for preprocessing the horizontal project data, vertical project data, academic achievement data and scientific research task information of scientific research talents into triple format, constructing knowledge graph for training, and completing and mining the potential relationship between entities;

[0045] LSTM-based bidirectional encoder module: responsible for using the bidirectional LSTM encoder to independently encode meta-paths from the perspective of talent and the perspective of scientific research tasks, modeling knowledge graph information from both perspectives, and generating representation information of scientific research talent vision and R&D task vision, and returning the representation information;

[0046] Matching prediction module: Calculate the similarity between the representation information of scientific research personnel vision and R&D task vision, and return them from high to low in terms of similarity to complete a recommendation;

[0047] Prediction training module: responsible for returning the recommendation results of scientific research tasks and the adoption results of users to the model, pre-training the model, and regularly retraining new data to optimize the prediction model;

[0048] Recommendation result output module: responsible for returning the matching results according to scientific research talent information and matching values ​​and displaying them to users.

[0049] The present invention solves the technical defects existing in the background technology, and has the following beneficial effects:

[0050] The characteristic data of scientific research talents are obtained in the big data network, and the characteristic data are processed by keyword extraction to obtain several characteristic keywords, and the weight attributes of scientific research talents are determined according to the characteristic keywords; the research and development tasks are obtained, and the key technical words of the research and development tasks are obtained as attributes in the large language model through the prompt template, and the weight attributes of the research and development tasks are determined according to the order of return; the knowledge graph is constructed, and the knowledge graph is pre-trained based on the weight attributes of scientific research talents and the weight attributes of research and development tasks, and the node relationships hidden in the knowledge graph are captured to obtain the representation of the knowledge graph; the characteristic data of scientific researchers are fused and the relationship is mined according to the representation of the knowledge graph to obtain the fusion representation of the professional and technical capabilities of scientific researchers; the bidirectional LSTM encoder is used to independently encode the meta-path from the perspective of scientific research talents and the perspective of research and development tasks, so as to model the knowledge graph information from the perspective of scientific research talents and the perspective of research and development tasks, and obtain the representation of scientific research talents and the text representation of research and development tasks; the bilateral recommendation training is carried out through the bilateral quadruple loss function to obtain the bidirectional matching model, and the scientific research talents and research and development tasks are bidirectionally matched according to the bidirectional matching model. The recommendation matching accuracy of scientific research talents and research and development tasks can be improved by this method. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, drawings of other embodiments can be obtained based on these drawings without paying creative work.

[0052] Figure 1 This is a flowchart of the method for bidirectional matching of scientific research talents and R&D tasks based on graph representation;

[0053] Figure 2 This is a system block diagram of a two-way matching system for scientific research talents and R&D tasks based on graph representation. DETAILED DESCRIPTION

[0054] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0055] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0056] like Figure 1 As shown, the present invention discloses a two-way matching method between scientific research talents and R&D tasks based on graph representation, comprising the following steps:

[0057] S102: Acquire characteristic data of scientific research talents in a big data network, perform keyword extraction processing on the characteristic data to obtain a number of characteristic keywords, and determine weight attributes of scientific research talents according to the characteristic keywords;

[0058] It should be noted that appropriate big data networks should be selected, such as CNKI database, scientific research social platform, professional forum, etc. Use web crawlers, API interfaces or other data collection tools to collect characteristic data related to scientific research talents, such as personal information, research fields, published papers, project experience, etc. Natural language processing method models such as bag-of-words model, TF-IDF algorithm, topic model, etc. can be used to extract characteristic keywords, calculate the frequency of each keyword in the data set, and screen out important characteristic keywords based on word frequency or other related indicators. You can also use word vector model or semantic network to expand keywords to obtain richer semantic information; determine the weight according to the frequency of occurrence of keywords, and high-frequency keywords have higher weights; you can also use TF-IDF algorithm to calculate the weight of keywords, considering the frequency of keywords in the document and the distribution in the entire data set. The weights of each keyword are comprehensively calculated to obtain the final weight attribute of scientific research talents, and the weights are normalized so that they are within a certain range for easy comparison and analysis.

[0059] S104: Acquire the R&D task, obtain the key technical words of the R&D task in the large language model as attributes through the prompt template, and determine the weight attribute of the R&D task according to the returned order;

[0060] It should be noted that the R&D tasks are obtained as attributes of the key technical words of the R&D tasks in the pre-trained large language model through the prompt template, and the weights are set according to the order returned. For example, if you ask the large model, "There is a technical demand for "demand content", please answer the main technical keywords involved, and the system will automatically sort and return them according to the importance of the keywords from high to low." After the large model returns the keywords of the demand "K1, K2, K3, K4, K4", if there are more than the threshold N, the first N can be selected, and the system will automatically set the weights in order to the first level (such as 1.0), the second level (such as 0.9), the third level (such as 0.8), the fourth level (such as 0.7), the fifth level (such as 0.6), etc.

[0061] S106: construct a knowledge graph, and pre-train the knowledge graph based on the weight attributes of scientific research talents and the weight attributes of R&D tasks, capture the node relationships hidden in the knowledge graph, and obtain a representation method of the knowledge graph;

[0062] S108: According to the representation method of the knowledge graph, the characteristic data of the researchers are integrated and the relationship is mined to obtain the integrated representation of the professional and technical capabilities of the researchers;

[0063] It should be noted that after pre-training on the knowledge graph, a sub-graph fusion information representation based on a single-node path can be obtained. With the help of the knowledge graph representation method, the characteristic data of scientific researchers are effectively fused and relationship mined, and a fused representation of the professional and technical capabilities of scientific researchers is obtained.

[0064] S110: Use a bidirectional LSTM encoder to independently encode meta-paths from the perspective of scientific research talents and the perspective of R&D tasks, so as to model knowledge graph information from the perspective of scientific research talents and the perspective of R&D tasks, and obtain the representation of scientific research talents' capabilities and the text representation of R&D tasks;

[0065] It should be noted that a bidirectional LSTM encoder is used to process meta-paths because it can process sequence data and capture dependencies in sequences. By considering researchers and R&D tasks as the starting points of the sequence, the bidirectional LSTM is able to learn a dual-view representation of each meta-path.

[0066] S112: Performing two-way recommendation training through a bilateral quadruple loss function to obtain a two-way matching model, and performing two-way matching between scientific research talents and R&D tasks according to the two-way matching model;

[0067] Among them, the characteristic data include horizontal projects undertaken and participated by scientific research talents and the patents and paper results obtained.

[0068] More specifically, the knowledge graph is pre-trained based on the weight attributes of scientific research talents and the weight attributes of R&D tasks, the node relationships hidden in the knowledge graph are captured, and the representation method of the knowledge graph is obtained, which also includes:

[0069] For the i-th node entity e in the knowledge graph i , through multi-layer graph convolutional network encoding, the transmission of adjacency information between entities is realized:

[0070]

[0071] In the formula, is the node entity e at the l+1th layer i express; It is a relationship ij The weight coefficient of g is a parameter learned by the network; σ(·) is an activation function, such as ReLU; is the node entity e at layer l i Indicates; n is the node entity e i The total number of .

[0072] It should be noted that in the knowledge graph, entities and relations can be represented by vectors, and the weighted knowledge graph representation method can better capture the complex relationships between entities by assigning weights to relations. Based on the graph convolutional network encoding method, weights can be used to represent the importance of different relations. Then, for the node entity e i (For example, the i-th scientific research talent), through multi-layer graph convolutional network coding, the transmission of adjacency information between entities is realized.

[0073] In order to make the knowledge capture of the knowledge graph more complete, convolutional TransE decoding is introduced into the knowledge graph construction module;

[0074] For convolutional TransE decoding, the kernel mapping representation m is obtained based on C kernels of different sizes. c , and each kernel map represents m c The concatenation constructs a comprehensive representation M of multi-scale features C ;

[0075] The scoring function of the convolution TransE method after nonlinear convolution is defined as follows:

[0076] θ(e s ,e o )=f(vec(M(e s ,e r ))W c ) o

[0077] In the formula, θ(es ,e o ) represents a scoring function used to evaluate e s ,e o The rationality of the relationship between s ,e o are the head node and tail node embeddings, e r is the corresponding relation embedding; W c is the weight coefficient of the convolution TransE decoding; M(e s ,e r ) represents the s ,e r Perform concatenation operations to output a matrix or high-dimensional tensor; vec(·) represents a vectorization operation to flatten the input matrix or high-dimensional tensor into a vector; f(·) is a nonlinear activation function, such as ReLU, tanh, etc.

[0078] During the training of the convolutional TransE model, the logistic sigmoid function is applied to the score:

[0079] p(e s ,e r ,e o )=σ(θ(e s ,e o ))

[0080] In the formula, p(e s ,e r ,e o ) means e s ,e r ,e o The probability of the triple being true, the output value range is between 0 and 1, indicating the possibility of the triple existing in the knowledge graph; s ,e o are the head node and tail node embeddings, e r is the corresponding relation embedding; θ(e s ,e o ) represents a scoring function used to evaluate e s ,e o The rationality of the relationship between them; σ(·) is an activation function.

[0081] More specifically, the use of a bidirectional LSTM encoder to independently encode meta-paths from the perspective of scientific research talent and the perspective of R&D tasks also includes:

[0082] For the i-th node entity e on the path i , which is fed into a bidirectional LSTM encoder to learn the contextual representation:

[0083]

[0084] In the formula, and are the model encoding forward and backward hidden states of the i-th step; e i Represents the i-th node entity; Indicates that the model encodes the i-1th step forward hidden state; represents the model encoding the i+1th backward hidden state; f(·) is the function of the LSTM unit;

[0085] Finally, the aggregated representation of the two views is calculated by averaging the hidden states of the bidirectional LSTM encoder from both directions:

[0086]

[0087] In the formula, p c and p t are the meta-path fusion representations of talents and scientific research tasks respectively; m is the actual length of the meta-path, that is, the number of node entities contained; and are the model encoding forward and backward hidden states of the i-th step respectively.

[0088] More specifically, the scientific research talent ability representation and the R&D task text representation are input into the training model for training, and the training and validation data sets are split in the ratio of 8:2.

[0089] In addition, the characteristic data of scientific researchers are integrated and relationship mined according to the representation method of knowledge graph to obtain the integrated representation of the professional and technical capabilities of scientific researchers, specifically:

[0090] Use deep learning models to generate embedding vectors for entities and relations in the knowledge graph. The embedding vectors are used to capture the semantic information of entities and relations. The deep learning models include TransE and RotatE.

[0091] Generate embedding vectors based on entities and relationships in the knowledge graph. Use the LSTM model to encode the sequence of scientific researchers' project experience and achievements to generate text vectors of scientific researchers' project experience and achievements.

[0092] Perform weighted averaging of the vectors of the text vectors based on the attention mechanism to obtain the achievements of researchers in various projects;

[0093] Add an attention mechanism to the output layer of the LSTM model so that the LSTM model pays different attention to different projects when generating the final project experience representation, and outputs projects with a higher attention level than the preset attention level.

[0094] Use graph neural networks to mine relationship paths in knowledge graphs. Relationship paths reflect the relationship between researchers and projects and achievements.

[0095] The mined relationship path information, the achievements of researchers in various projects, and the projects with a higher degree of attention than the preset degree of attention are integrated, and a fusion representation of the professional and technical capabilities of researchers is generated through weighted summation or product operations;

[0096] Among them, the fusion representation is a vector containing multiple dimensions, each dimension representing the ability strength of scientific researchers in different aspects.

[0097] It should be noted that by utilizing the attention mechanism based on the representation method of the knowledge graph, the project experience and achievement information of scientific researchers can be effectively integrated, the relationship between their professional capabilities can be explored, and a fusion representation of professional and technical capabilities with deep understanding can be generated.

[0098] like Figure 2 As shown, the present invention also discloses a two-way matching system of scientific research talents and R&D tasks based on graph representation, which is applied to any one of the two-way matching methods of scientific research talents and R&D tasks, including:

[0099] Scientific research talent information database: saves the horizontal and vertical project data of scientific research talents, as well as the data of scientific research talents and their academic achievements; stores them according to different data types for specific data preprocessing modules to call and query;

[0100] Horizontal project preprocessing module: responsible for extracting keywords from the horizontal projects of researchers, including the project's brief introduction, name, and core achievement name;

[0101] Longitudinal project preprocessing module: responsible for extracting keywords from the longitudinal projects of researchers, including the project's achievement profile, name, and core achievement name;

[0102] Paper results preprocessing module: responsible for extracting keyword information from the names, keywords and abstracts of researchers' papers;

[0103] Patent results preprocessing module: responsible for extracting keyword information from the patent names, keywords and abstracts of researchers’ patents;

[0104] Scientific research task input module: The input module of the system is used by users to input the scientific research task requirements that need to be matched. The system matches researchers according to the requirements and returns the matching results in order of matching degree from high to low.

[0105] Also includes:

[0106] Prompt information generation module: responsible for managing the information templates that call the large language model interface, generating query instructions for R&D tasks according to the format requirements of the prompt templates, and returning the query instructions to the scientific research task information preprocessing module;

[0107] Scientific research task information preprocessing module: obtains key technical words of R&D tasks as attributes in the large language model;

[0108] Knowledge graph construction module: responsible for preprocessing the horizontal project data, vertical project data, academic achievement data and scientific research task information of scientific research talents into triple format, constructing knowledge graph for training, and completing and mining the potential relationship between entities;

[0109] LSTM-based bidirectional encoder module: responsible for using the bidirectional LSTM encoder to independently encode meta-paths from the perspective of talent and the perspective of scientific research tasks, modeling knowledge graph information from both perspectives, and generating representation information of scientific research talent vision and R&D task vision, and returning the representation information;

[0110] Matching prediction module: Calculate the similarity between the representation information of scientific research personnel vision and R&D task vision, and return them from high to low in terms of similarity to complete a recommendation;

[0111] Prediction training module: responsible for returning the recommendation results of scientific research tasks and the adoption results of users to the model, pre-training the model, and regularly retraining new data to optimize the prediction model;

[0112] Recommendation result output module: responsible for returning the matching results according to scientific research talent information and matching values ​​and displaying them to users.

[0113] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0114] The units described above as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units; they may be located in one place or distributed on multiple network units; some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0115] In addition, all functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may be separately used as a unit, or two or more units may be integrated into one unit; the above-mentioned integrated units may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0116] Those skilled in the art can understand that: all or part of the steps of implementing the above method embodiments can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above method embodiments; and the aforementioned storage medium includes: mobile storage devices, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), disks or optical disks, and other media that can store program codes.

[0117] Alternatively, if the above-mentioned integrated unit of the present invention is implemented in the form of a software function module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present invention can be essentially or partly reflected in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the methods of each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as mobile storage devices, ROM, RAM, magnetic disks or optical disks.

[0118] The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A two-way matching method for scientific research talents and R&D tasks based on graph representation, characterized in that: The following steps are involved: Acquire characteristic data of scientific research talents in a big data network, perform keyword extraction processing on the characteristic data to obtain a number of characteristic keywords, and determine weight attributes of scientific research talents according to the characteristic keywords; Acquire the R&D task, use the prompt template to obtain the key technical words of the R&D task in the large language model as attributes and return them, and determine the weight attribute of the R&D task according to the returned order; Construct a knowledge graph, and pre-train the knowledge graph based on the weight attributes of scientific research talents and the weight attributes of R&D tasks, capture the hidden node relationships of the knowledge graph, and obtain a representation method of the knowledge graph; According to the representation method of knowledge graph, the characteristic data of scientific researchers are integrated and the relationship is mined to obtain the integrated representation of the professional and technical capabilities of scientific researchers; Use a bidirectional LSTM encoder to independently encode meta-paths from the perspective of scientific research talents and the perspective of R&D tasks, so as to model knowledge graph information from the perspective of scientific research talents and the perspective of R&D tasks, and obtain the representation of scientific research talents' capabilities and the text representation of R&D tasks; Performing two-way recommendation training through a bilateral quadruple loss function to obtain a two-way matching model, and performing two-way matching between scientific research talents and R&D tasks according to the two-way matching model; The knowledge graph is pre-trained based on the weight attributes of scientific research talents and the weight attributes of R&D tasks to capture the node relationships hidden in the knowledge graph and obtain a representation method of the knowledge graph, which also includes: For the first Node Entity , through multi-layer graph convolutional network encoding, the transmission of adjacency information between entities is realized: ， In the formula, It is Layer node entity express; It's a relationship The weight coefficient of are the parameters learned by the network; is an activation function; It is Layer node entity express; Is a node entity Total number of; Also includes: Introducing convolutional TransE decoding into the knowledge graph building module; For convolutional TransE decoding, the kernel mapping representation is obtained based on 𝐶 kernels of different sizes. , and characterize each kernel map Concatenate and construct a comprehensive representation of multi-scale features ; The scoring function of the convolution TransE method after nonlinear convolution is defined as follows: ， In the formula, Represents a scoring function used to evaluate the rationality of the relationship between them; are the head node and tail node embeddings, yes and The corresponding relation embedding; is the weight coefficient of the convolutional TransE decoding; Representatives Perform concatenation operations to output a matrix or high-dimensional tensor; Represents a vectorized operation, flattening the input matrix or high-dimensional tensor into a vector; is a non-linear activation function.

2. According to claim 1, a method for bidirectional matching of scientific research talents and R&D tasks based on graph representation is characterized in that: Also includes: During the training of the convolutional TransE model, the logistic sigmoid function is applied to the score: ， In the formula, express The probability of the triple being true, with an output value ranging from 0 to 1, indicating the possibility of the triple existing in the knowledge graph; are the head node and tail node embeddings, yes and The corresponding relation embedding; Represents a scoring function used to evaluate the rationality of the relationship between them; is an activation function.

3. According to claim 1, a method for bidirectional matching of scientific research talents and R&D tasks based on graph representation is characterized in that: The use of a bidirectional LSTM encoder to independently encode meta-paths from the perspective of scientific research talent and the perspective of R&D tasks also includes: For the first Node Entity , which is fed into a bidirectional LSTM encoder to learn the contextual representation: ， ， In the formula, and They are model encoding Step forward and backward hidden states; Indicates Node entities; Representation model encoding Step forward hidden state; Representation model encoding Step backward hidden state; is a function of the LSTM unit; Finally, the aggregated representation of the two views is calculated by averaging the hidden states of the bidirectional LSTM encoder from both directions: ， ， In the formula, and They are the meta-path fusion representations of talents and scientific research tasks respectively; is the actual length of the meta-path, that is, the number of node entities contained; and They are model encoding Step forward and backward hidden states.

4. According to claim 1, a method for bidirectional matching of scientific research talents and R&D tasks based on graph representation, characterized in that: The scientific research talent ability representation and the R&D task text representation are input into the training model for training, and the training and verification data sets are split in the ratio of 8:

2.

5. A two-way matching system for scientific research talents and R&D tasks based on graph representation, applied to the two-way matching method for scientific research talents and R&D tasks described in any one of claims 1 to 4, characterized in that: include: Scientific research talent information database: saves the horizontal and vertical project data of scientific research talents, as well as the data of scientific research talents and their academic achievements; stores them according to different data types for specific data preprocessing modules to call and query; Horizontal project preprocessing module: responsible for extracting keywords from the horizontal projects of researchers, including the project's brief introduction, name, and core achievement name; Longitudinal project preprocessing module: responsible for extracting keywords from the longitudinal projects of researchers, including the project's achievement profile, name, and core achievement name; Paper results preprocessing module: responsible for extracting keyword information from the names, keywords and abstracts of researchers' papers; Patent results preprocessing module: responsible for extracting keyword information from the patent names, keywords and abstracts of researchers’ patents; Scientific research task input module: The input module of the system is used by users to input the scientific research task requirements that need to be matched. The system matches researchers according to the requirements and returns the matching results in order of matching degree from high to low.

6. According to claim 5, a two-way matching system for scientific research talents and R&D tasks based on graph representation is characterized in that: Also includes: Prompt information generation module: responsible for managing the information templates that call the large language model interface, generating query instructions for R&D tasks according to the format requirements of the prompt templates, and returning the query instructions to the scientific research task information preprocessing module; Scientific research task information preprocessing module: obtains key technical words of R&D tasks as attributes in the large language model; Knowledge graph construction module: responsible for preprocessing the horizontal project data, vertical project data, academic achievement data and scientific research task information of scientific research talents into triple format, constructing knowledge graph for training, and completing and mining the potential relationship between entities; LSTM-based bidirectional encoder module: responsible for using the bidirectional LSTM encoder to independently encode meta-paths from the perspective of talent and the perspective of scientific research tasks, modeling knowledge graph information from both perspectives, and generating representation information of scientific research talent vision and R&D task vision, and returning the representation information; Matching prediction module: Calculate the similarity between the representation information of scientific research personnel vision and R&D task vision, and return them from high to low in terms of similarity to complete a recommendation; Prediction training module: responsible for returning the recommendation results of scientific research tasks and the adoption results of users to the model, pre-training the model, and regularly retraining new data to optimize the prediction model; Recommendation result output module: responsible for returning the matching results according to scientific research talent information and matching values ​​and displaying them to users.

Images