Expert recommendation system and method, computer equipment and storage medium

Through an expert recommendation system based on BERT and LDA models, combined with deep learning and indicator recommendation modules, the problems of low efficiency and low accuracy of expert recommendation in school-enterprise cooperation are solved, and rapid and precise matching of experts is achieved, reducing resource consumption and manpower Cost has improved the R&D efficiency of school-enterprise cooperation.

CN118981530AInactive Publication Date: 2024-11-19HEFEI UNIV OF TECH

Patent Information

Application Number
CN202410853708.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-06-28
Publication Date
2024-11-19
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The existing expert recommendation methods are inefficient, the matching accuracy is not high, and it is impossible to quickly and accurately match experts who can solve corporate technical problems, especially in the school-enterprise cooperation scenario, and ignore the experts' research topic changes and scientific research capabilities. Traditional neural networks Training resources are expensive and the results are not ideal.

Method used

Using the enterprise demand classifier and the expert modeler based on the BERT model, combining the deep learning recommendation module and the indicator recommendation module, calculate the matching score through the vector relationship between the enterprise demand characteristics and expert characteristics, and quantify other expert indicators to calculate the recommendation. Priority, quickly lock in related research directions and latest research topics, evaluate scientific research capabilities, and optimize recommended results.

Benefits of technology

It achieves rapid and accurate matching of experts in the school-enterprise cooperation scenario, reduces search time and computing resource consumption, improves the accuracy and comprehensiveness of recommended results, saves labor costs, and improves the R&D efficiency of school-enterprise cooperation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118981530A_ABST
    Figure CN118981530A_ABST
Patent Text Reader

Abstract

The invention discloses an expert recommendation system and method, computer equipment and a storage medium, and belongs to the technical field of computer information processing. Aiming at the problems of low efficiency and low matching accuracy of an expert recommendation method in the prior art, enterprise demand data is input into an enterprise demand classifier based on a BERT model to generate enterprise demand features, and an expert modeler based on LDA outputs expert features; a deep learning recommendation module is constructed to calculate a vector relationship between enterprise demand features and expert features and a hiding relationship between enterprise demand hiding features and expert hiding features to obtain a matching score, and an index recommendation module is constructed to perform quantitative calculation on other indexes of experts to obtain a matching score. And the recommendation priority is calculated according to the matching score output by the deep learning recommendation module and the matching score output by the index recommendation module, so that experts capable of solving specific technical problems can be quickly and accurately matched, and the method is particularly suitable for expert recommendation in a school-enterprise cooperation scene.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of computer information processing, and more specifically, to an expert recommendation system, method, computer equipment and storage medium. Background Art

[0002] Industry-university-research cooperation refers to a model of close cooperation between industry, higher education institutions and scientific research institutions. Higher education institutions have rich knowledge resources, talent resources and R&D capabilities, while enterprises have economic foundations and market demand. Through the industry-university-research cooperation model of school-enterprise cooperation, both parties can share resources, complement each other's strengths, improve R&D efficiency and innovation capabilities, and promote a virtuous cycle of scientific and technological innovation and economic growth. In recent years, the central government of my country, including many provincial governments, has issued a number of relevant documents on promoting school-enterprise cooperation and promoting the development of industry-university-research cooperation, which also indirectly verifies the importance of school-enterprise cooperation.

[0003] When enterprises encounter technical problems in actual production, they usually seek cooperation from universities by publishing enterprise needs due to the lack of professional knowledge and relevant talents. However, existing expert recommendations are mostly focused on academic paper review and expert cooperation in similar research directions, and are not suitable for expert recommendations in school-enterprise cooperation scenarios. Manual screening of relevant experts by universities has the disadvantages of insufficient professional knowledge and ability, a lot of time and labor costs, and low accuracy.

[0004] In recent years, with the development of natural language processing, especially the emergence of pre-trained language models, new directions have been opened up for text-based recommendation algorithms. Pre-trained language models use large-scale unlabeled corpus data for pre-training to learn and generate the intrinsic representation of text, thereby providing an excellent feature extractor. Compared with traditional word vector-based models, it has stronger context understanding capabilities and better transfer effects, providing a new way for expert recommendations for enterprise needs with less labor consumption and high accuracy.

[0005] Enterprises usually face specific and urgent technical problems that require quick and accurate solutions from suitable experts. Traditional expert recommendation algorithms, such as those based on academic paper review and similar research directions, cannot fully meet this demand because they focus on matching academic achievements and research interests rather than solving technical problems in practical applications. In addition, manual screening of relevant experts by universities and scientific research institutions has the problem of insufficient professional knowledge and ability, and consumes a lot of time and manpower costs. Therefore, studying an expert recommendation algorithm for enterprise needs can help enterprises quickly locate experts who can solve specific technical problems, which is of great significance for accelerating problem solving in school-enterprise cooperation, shortening the R&D cycle, and reducing R&D costs.

[0006] The current expert recommendation algorithm has three main shortcomings: (1) The utilization rate of expert personal information is low. The existing expert recommendation method mainly makes recommendations based on information about the expert's research field, while other information about the expert's achievements, scientific research projects, and scientific research capabilities are often not involved in the recommendation process. (2) Traditional expert recommendation algorithms ignore the changes in expert research topics. For example, an expert who previously studied inorganic materials may have turned to studying biomaterials. (3) Using traditional neural networks to extract features from corporate texts and experts and relying solely on similarity for recommendation ignores the rich semantic information between experts, and training is more resource-intensive, so the effect is not ideal. Summary of the invention

[0007] 1. Technical issues to be solved

[0008] In view of the problems of low efficiency and low matching accuracy of expert recommendation methods in the prior art, the present invention provides an expert recommendation system, method, computer device and storage medium, which can quickly and accurately match experts who can solve specific technical problems, and is particularly suitable for expert recommendation in school-enterprise cooperation scenarios.

[0009] 2. Technical solution

[0010] The purpose of the present invention is achieved through the following technical solutions.

[0011] An expert recommendation system, comprising: a data collection module, a classification module, a storage module, a modeling module, and a recommendation module;

[0012] The data collection module collects enterprise demand data and pre-processes the enterprise demand data;

[0013] The classification module constructs an enterprise demand classifier based on the BERT model, inputs enterprise demand data into the enterprise demand classifier based on the BERT model to generate enterprise demand features, and determines the demand field to which the enterprise demand belongs through the enterprise demand features;

[0014] The storage module stores expert information data and classifies the expert information data to obtain expert feature fields;

[0015] The modeling module constructs an expert modeler based on the LDA model through the expert's feature fields, and the LDA-based expert modeler outputs expert features;

[0016] The recommendation module includes a deep learning recommendation module and an indicator recommendation module; the deep learning recommendation module calculates the vector relationship between the enterprise demand characteristics and the expert characteristics and the hidden relationship between the enterprise demand hidden characteristics and the expert hidden characteristics to obtain a matching score, and the indicator recommendation module quantifies other expert indicators to obtain a matching score, and calculates the recommendation priority based on the matching score output by the deep learning recommendation module and the matching score output by the indicator recommendation module.

[0017] Furthermore, the step of inputting the enterprise demand data into the enterprise demand classifier based on the BERT model to generate enterprise demand features, and judging the demand field to which the enterprise demand belongs by the enterprise demand features comprises:

[0018] Input enterprise demand into the enterprise demand classifier based on the BERT model to generate enterprise demand features;

[0019] The vector dimension of enterprise demand features is converted into the number of categories of the demand field classification through the fully connected layer, and the possible vector set of the demand field to which the enterprise demand belongs is obtained;

[0020] The probability vector set is passed through the Softmax layer to obtain the probability vector of the demand field to which the enterprise demand belongs. The calculation formula of the probability vector is:

[0021] P = Softmax(FC(O))

[0022] Among them, P represents the probability vector, Softmax represents the activation function, FC represents the fully connected layer whose activation function is the ReLu activation function, and O represents the enterprise demand characteristics;

[0023] Determine the demand area to which the enterprise demand belongs through the probability vector.

[0024] Furthermore, the deep learning recommendation module includes a demand-expert encoder and a prediction module; the demand-expert encoder maps the enterprise demand characteristics and expert characteristics to the same dimensional space, performs feature extraction and learning, and obtains the vector relationship between the enterprise demand characteristics and the expert characteristics. The prediction module integrates the vector relationship between the enterprise demand characteristics and the expert characteristics and then reduces the dimension to output the matching score.

[0025] Furthermore, the demand-expert encoder converts the enterprise demand feature embedding into the word vector of the enterprise demand feature, and encodes the word vector of the enterprise demand feature through the encoder to obtain the enterprise demand hidden feature;

[0026] Requirements: The expert encoder converts the expert feature embedding into the word vector of the expert feature, and encodes the word vector of the expert feature through the encoder to obtain the expert hidden feature;

[0027] Calculate the hidden relationship between the hidden features of enterprise needs and the hidden features of experts to obtain the matching probability score. The calculation formula is:

[0028] Score(V e ,V s )=Sigmoid(FC(V S T V e ))

[0029] Among them, V e represents the hidden characteristics of enterprise demand, V s Represents expert hidden features, Score(Ve,Vs) represents the matching probability score, Sigmoid represents the Sigmoid activation function, T represents the transposition operation, and FC represents the fully connected layer function.

[0030] Furthermore, the matching score of the deep learning recommendation module is obtained by combining the matching score of the vector relationship output of the enterprise demand features and the expert features and the matching probability score of the hidden relationship output of the hidden features of the enterprise demand and the expert hidden features.

[0031] Furthermore, other expert indicators are quantitatively calculated, including expert topics, academic achievements and related backgrounds. The calculation formula is:

[0032] S=(P+Q+G) / 3

[0033] Among them, S represents the matching score output by the indicator recommendation module, P represents the weight of the expert topic part, Q represents the weight of the academic achievement part, and G represents the weight of the relevant background part.

[0034] Furthermore, the recommendation priority is calculated by the matching score output by the deep learning recommendation module and the matching score output by the indicator recommendation module. The calculation formula is:

[0035] M=α×Score i +β×S i

[0036] Among them, M represents the recommendation priority, i represents the expert, and Score i represents the matching score output by the deep learning recommendation module of expert i, α represents the weight of the matching score output by the deep learning recommendation module, S i represents the matching score output by the indicator recommendation module of expert i, and β represents the weight of the matching score output by the indicator recommendation module.

[0037] An expert recommendation method comprises the following steps:

[0038] Constructing an expert recommendation system as described above;

[0039] Input enterprise demand data into the enterprise demand classifier based on the BERT model to generate enterprise demand features, determine the demand field to which the enterprise demand belongs through the enterprise demand features, and determine the relevant expert set based on the demand field to which the enterprise demand belongs;

[0040] The expert features and enterprise demand features output by the LDA-based expert modeler are input into the recommendation module to calculate the recommendation priority M of all experts in the relevant expert set;

[0041] Arrange in descending order according to the recommendation priority M. After the arrangement is completed, take the first K experts to obtain the K most suitable experts recommended for the enterprise needs.

[0042] A computer device comprises a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and the processor implements the above-mentioned method when executing the computer program.

[0043] A computer-readable storage medium stores a computer program, and the computer program executes the above method when executed by a processor.

[0044] 3. Beneficial effects

[0045] Compared with the prior art, the advantages of the present invention are:

[0046] (1) The expert recommendation system, method, computer device and storage medium of the present invention aim to solve the problem that the research directions of university experts are diverse and large in number, and the global search and matching method consumes a lot of resources and time. The enterprise demand classifier based on the BERT model is constructed. By preliminarily classifying enterprise needs, the expert group with relevant research directions can be quickly identified, thereby reducing search time and saving computing resources.

[0047] (2) The expert recommendation system, method, computer device and storage medium of the present invention address the problem that existing relevant expert recommendation algorithms ignore the dynamic changes of expert research topics, and construct an LDA-based expert modeler to extract the latest research topics of experts, thereby avoiding the negative impact of changes in expert research topics on feature coding.

[0048] (3) An expert recommendation system, method, computer device and storage medium of the present invention constructs a deep learning recommendation module to address the problem that traditional neural networks ignore the rich semantic information between features when extracting demand text and expert features, and training related models consumes a lot of computing resources. Compared with the deep learning recommendation module constructed by conventional neural networks, the deep learning recommendation module constructed based on BERT can not only deeply encode related features but also save certain training resources.

[0049] (4) The expert recommendation system, method, computer device and storage medium of the present invention are designed to address the problem that the current expert recommendation algorithm for school-enterprise cooperation is limited to simple text conversion and ignores a large amount of valuable information by calculating the matching degree through cosine similarity, resulting in unsatisfactory recommendation result indicators. A deep learning recommendation module based on a deep learning model is constructed, and the stronger feature extraction, fusion and dimensionality reduction capabilities of deep learning are used to obtain better recommendation effects.

[0050] (5) The expert recommendation system, method, computer device and storage medium of the present invention address the problem that existing expert recommendation algorithms ignore the scientific research capabilities of experts themselves, and propose an expert scientific research information indicator evaluation system, thereby constructing an indicator recommendation module to achieve the evaluation and quantification of the expert's scientific research capabilities, and use it as one of the recommendation indicators in the recommendation process, so that the final recommendation result is more comprehensive and the recommendation effect is effectively improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 This is a structural diagram of the expert recommendation system according to an embodiment of the present invention;

[0052] Figure 2 This is a flow chart of data training processing according to an embodiment of the present invention;

[0053] Figure 3 This is a structural diagram of a classification module according to an embodiment of the present invention;

[0054] Figure 4 This is a structural diagram of an expert modeler based on LDA according to an embodiment of the present invention;

[0055] Figure 5 This is a structural diagram of the LDA model according to an embodiment of the present invention;

[0056] Figure 6 This is a structural diagram of a deep learning recommendation module according to an embodiment of the present invention;

[0057] Figure 7 It is a graph of other indicators composed by experts in the embodiment of the present invention;

[0058] Figure 8 The figure is a flow chart of the expert recommendation method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0059] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0060] Example

[0061] In the prior art, when enterprises encounter technical problems in actual production, they usually seek cooperation from universities by publishing enterprise needs due to the lack of professional knowledge and relevant talents. However, most of the existing expert recommendations focus on academic paper review and expert cooperation in similar research directions, and are not suitable for expert recommendations in school-enterprise cooperation scenarios. Therefore, this embodiment is described by an expert recommendation system applied to school-enterprise cooperation scenarios.

[0062] like Figure 1 As shown in FIG. 1 , an expert recommendation system provided by this embodiment includes a data collection module, a classification module, a storage module, a modeling module, and a recommendation module.

[0063] The data collection module is used to collect enterprise demand data and pre-process the enterprise demand data. The classification module is used to build an enterprise demand classifier based on the BERT model, input enterprise demand data into the enterprise demand classifier based on the BERT model to generate enterprise demand features, and determine the demand field to which the enterprise demand belongs through the enterprise demand features. The storage module is used to store expert information data and classify the expert information data to obtain the expert's feature field. The modeling module builds an expert modeler based on the LDA model through the expert's feature field, and the expert modeler based on LDA outputs expert features. The recommendation module includes a deep learning recommendation module and an indicator recommendation module; the deep learning recommendation module calculates the vector relationship between enterprise demand features and expert features and the hidden relationship between enterprise demand hidden features and expert hidden features to obtain a matching score, and the indicator recommendation module quantifies other expert indicators to obtain a matching score, and calculates the recommendation priority based on the matching score output by the deep learning recommendation module and the matching score output by the indicator recommendation module.

[0064] Specifically in this embodiment, in the data collection module, the enterprise demand data is first collected and pre-processed. Figure 2As shown, in this embodiment, enterprise demand data is collected by crawling. Specifically, several seed pages with high relevance to the topic are set as initial data, and the crawler is directed to crawl various overview pages containing enterprise demand information to obtain corresponding text data. It should be noted that in this embodiment, collecting and obtaining enterprise demand data by crawling is a prior art. Then, the crawled text data is parsed to obtain the original keyword information, and the keyword information is stored in a text format according to classification as training material for training and verifying the enterprise demand classifier. In this embodiment, in order to ensure the accuracy and fluency of the training material, the training material needs to be preprocessed to obtain the enterprise demand training set. In this embodiment, the preprocessing of the training material text includes: (1) construction of a stop word library. The general stop word library draws on the existing stop word list, while in this embodiment, the domain stop word library is constructed by continuous updating and improvement. The idea is to add words greater than a certain threshold α to the domain stop word library. (2) Remove special characters and punctuation marks, remove special characters, punctuation marks and symbols in the text, and unify the spaces and line breaks in the text for subsequent processing. (3) Spelling correction, correct spelling errors in the text as needed. Thus, after preprocessing the training material text, the enterprise demand training set is obtained.

[0065] like Figure 3 As shown, in the classification module, an enterprise demand classifier based on the BERT model is constructed, and the enterprise demand classifier of the BERT model is trained through the enterprise demand training set to obtain the trained enterprise demand classifier of the BERT model. Furthermore, the enterprise demand is input into the preprocessor, and after the preprocessor preprocesses the enterprise demand, the enterprise demand is input into the enterprise demand classifier based on the BERT model, and the enterprise demand feature O is generated by the enterprise demand classifier of the BERT model. The vector dimension of the enterprise demand feature O is converted into the number of categories of the enterprise demand field classification through the fully connected layer, and a possibility vector y representing the demand field to which the enterprise demand belongs is obtained. The possibility vectors y of multiple different categories representing the demand field to which the enterprise demand belongs form a possibility vector set Y, Y=(y 1 ,y 2 ,y 3 ,…,ynclas s ), where nclass is the number of categories. Then, the probability vector set Y can be used to obtain the probability vector P of the demand field to which the enterprise demand belongs through the Softmax layer. In this embodiment, the calculation formula of the probability vector P is:

[0066] P = Softmax(FC(O))

[0067] Wherein, P represents a probability vector, Softmax represents an activation function, FC represents a fully connected layer whose activation function is a ReLu activation function, and O represents the enterprise demand feature. Therefore, in this embodiment, after obtaining the probability vector P, taking the category where argmax(P) is located can obtain the most likely demand field of the enterprise demand.

[0068] Therefore, in this embodiment, in order to address the problem that the research directions of university experts are diverse and large in number, and the global search and matching method consumes a lot of resources and time, a business demand classifier based on the BERT model is constructed. By preliminarily classifying business needs, the expert group in the relevant research direction can be quickly locked in, thereby reducing search time and saving computing resources.

[0069] like Figure 4 As shown, in the modeling module, an expert modeler based on the LDA model is constructed. Specifically, in this embodiment, all papers published by experts in the university expert database in the past five years are collected, and the titles and abstracts of all papers are extracted as the corpus for training the expert modeler based on the LDA model. For example, for the set of papers published by expert A in the past five years, P = {p 1 , p 2 , p 3 , ..., p n}, concatenate the title and abstract of the nth paper to form the training corpus t x , forming the final corpus T of expert A = {t 1 , t 2 ,t 3 ,...,t n}.

[0070] It should be noted that, in this embodiment, the LDA (Linear Discriminant Analysis, latent Dirichlet distribution) model is a topic generation model based on three-layer Bayesian probability, which extracts the topic information of the document through the probability distribution between document-topic-word. For the generation of each document, the topic of the document is first selected with a certain probability, and then the words corresponding to each topic are selected with a certain probability. The whole document is generated by continuously iterating the above process, and both the document to topic and the topic to word obey the multinomial distribution.

[0071] like Figure 5 As shown in the figure, it is the structure diagram of the LDA model. Among them, α represents the prior parameter of the Dirichlet distribution of the topic under each document, β represents the prior parameter of the Dirichlet distribution of each topic word, D represents the total number of documents, θ d Represents the training corpus t d The topic probability distribution, z i Represents the training corpus t dThe topic corresponding to the i-th word in , K represents the number of topics, represents the probability distribution of words corresponding to topic K, ω i represents the i-th word in document d, and N represents the number of words. It should be noted that in this embodiment, the hidden variables θ, The calculation is performed by the variational method.

[0072] In this embodiment, the expert modeler based on the LDA model is used to perform text clustering on the corpus T of expert A to obtain the topic distribution of the expert's research content. Among them, each topic is composed of the distribution of words, which reflects the importance of different topics in the expert's research content. The expert's research topic is obtained by arranging it in the order of probability obtained by the expert modeler based on the LDA model as the input of the demand-expert encoder.

[0073] In this embodiment, the research topics of experts are extracted through the LDA model, and then the research topics are input into the BERT model to generate expert features, and the expert modeler based on LDA outputs the expert features. Therefore, this embodiment can extract the latest research topics of experts by constructing an expert modeler based on LDA, thereby avoiding the negative impact of changes in expert research topics on feature encoding.

[0074] In the recommendation module, the recommendation module is constructed by combining the BERT-based deep learning recommendation module with the indicator recommendation module based on the expert scientific research information indicator evaluation system. According to the input enterprise needs, the K experts who best meet the enterprise needs are recommended by combining the output indicators of the deep learning recommendation module and the indicator recommendation module.

[0075] Specifically, Figure 6 As shown, it is a structural diagram of the deep learning recommendation module. In this embodiment, the deep learning recommendation module includes a demand-expert encoder and a prediction module. The demand-expert encoder maps the enterprise demand features and the expert features to the same dimensional space, extracts and learns the features, and obtains the vector relationship between the enterprise demand features and the expert features. The prediction module outputs the matching score through feature fusion and dimensionality reduction based on the feature vector relationship between the enterprise demand features and the expert features, which is used as part of the recommendation composition.

[0076] It is worth noting that the demand-expert encoder uses the embedding layer of the BERT model to perform word segmentation and word embedding representation on the input enterprise demand features and expert features, and converts them into word vector representations in the latent semantic space. Specifically, for the input enterprise demand features, the demand-expert encoder converts the enterprise demand feature embedding into the word vector of the enterprise demand features, and encodes the word vector of the enterprise demand features through the Transformer encoder of the BERT model to obtain the enterprise demand hidden features containing hidden semantics; for the input expert features, the demand-expert encoder converts the expert feature embedding into the word vector of the expert features, and encodes the word vector of the expert features through the Transformer encoder of the BERT model to obtain the expert hidden features.

[0077] Therefore, in the indicator recommendation module based on the expert scientific research information indicator evaluation system, the input is the hidden features of enterprise demand and the hidden features of experts output by the BERT encoder, and the matching degree between the enterprise demand hidden features and the expert hidden features is measured by calculating the hidden relationship between the vectors of the enterprise demand hidden features and the expert hidden features, that is, the matching probability score. In this embodiment, the matching probability score is represented by the vector dot product, and the calculation formula is:

[0078] Score(V e ,V s )=Sigmoid(FC(V S T V e ))

[0079] Among them, V e represents the hidden characteristics of enterprise demand, V s Represents the expert hidden feature, Score(Ve,Vs) represents the matching probability score, Sigmoid represents the Sigmoid activation function, T represents the transposition operation, FC represents the fully connected layer function, and its neuron activation function is ReLU. By setting up multiple fully connected layer superposition structures, the high-dimensional sparse features are compressed and mapped to low-dimensional dense features, and finally the features are converted into a matching score Score with a value between [0,1] through the sigmoid layer, that is, the degree of matching between the enterprise demand and the expert. In this embodiment, the closer the Score is to 1, the higher the matching degree is, and the closer it is to 0, the lower the matching degree is. In this embodiment, Score is taken as the matching score output by the deep learning recommendation module.

[0080] Thus, the matching score of the deep learning recommendation module is obtained by combining the matching score of the vector relationship output of the enterprise demand features and the expert features and the matching probability score of the hidden relationship output of the enterprise demand hidden features and the expert hidden features. Therefore, in view of the problem that the traditional neural network ignores the rich information at the semantic level between the features when extracting the demand text and the expert features, and that training the relevant models will consume a lot of computing resources, in this embodiment, a deep learning recommendation module is specially constructed. Compared with the deep learning recommendation module constructed by the conventional neural network, the deep learning recommendation module constructed based on BERT can not only deeply encode the relevant features, but also save certain training resources. In addition, by constructing a deep learning recommendation module based on a deep learning model, the stronger feature extraction, fusion and dimensionality reduction capabilities of deep learning can be used to obtain better recommendation effects.

[0081] It is worth noting that in this embodiment, other expert indicators are also comprehensively considered and involved in the recommendation process, which can alleviate the unexplainability of deep learning recommendations to a certain extent. Figure 7 As shown in Figure 2, other expert indicators include: academic achievements, projects participated in, professional title, etc. Thus, an output indicator is obtained through quantification, which is combined with the output indicator of the deep learning recommendation module as the final recommendation result.

[0082] In this embodiment, in order to facilitate normalization, three weights x, y, and z are first given. Therefore, in terms of related topics, the number and level of topics that experts have participated in are mainly considered, and evaluation is performed through indicator quantification. Three weights are given according to the overall distribution of the number of topics completed by experts in the database, which are: 0-3, 4-6, and more than 7, and the corresponding weights are P1=(x, y, z); for the quantification of the level of the topic, the level of the topic is divided into national major projects, provincial and ministerial projects, and local projects. Considering that this part is difficult to quantify, but those who can undertake national major projects and provincial and ministerial projects are usually experts with strong comprehensive abilities, so the weight of this part of the experts is divided into three grades according to whether they have undertaken national major projects, whether they have undertaken provincial and ministerial projects, and whether they have not undertaken any projects. The corresponding weights are P2=(x, y, z), and finally the weight P of the expert topic part is obtained by P=P1+P2.

[0083] In terms of academic achievements, the main indicators are quantified based on the number of papers and patents published by experts. Among them, the papers are weighted according to the average number of papers published by experts in the database, and the weights Q1 = (x, y, z) are given according to the number of papers published by experts 0-10, 10-20, and more than 20; the number of patents is also given three weights Q2 = (x, y, z) according to different levels 0-3, 4-6, and more than 7. Finally, the weight Q of the academic achievement part is obtained through Q = Q1 + Q2.

[0084] In terms of relevant background, the main indicators are quantified based on the titles and length of time the experts have been employed. The weights are divided into three levels according to the titles of the experts: lecturer, associate professor, and professor, corresponding to G1 = (x, y, z) respectively; the weights are divided into three weights G2 = (x, y, z) according to the length of time the experts have been employed for 0-4 years, 4-9 years, and 10 years and above, and finally the weight G of the relevant background is obtained through G = G1 + G2.

[0085] The indicator quantification calculation is performed for the above three weights, and the matching score output by the indicator recommendation module is set to S. For the output matching score S, we have:

[0086] S=(P+Q+G) / 3

[0087] In this embodiment, the closer S is to 1, the stronger the scientific research ability of the expert is, and the more worthy of recommendation is when the similarity with the enterprise's needs is close.

[0088] Therefore, by constructing an indicator recommendation module, it is possible to evaluate and quantify the scientific research capabilities of experts, and use it as one of the recommendation indicators in the recommendation process, making the final recommendation results more comprehensive and effectively improving the recommendation effect.

[0089] Finally, for a specific enterprise demand, after obtaining the output indicators of the expert i deep learning recommendation module and the indicator recommendation module, a linear weighted method is used to obtain the recommendation priority M. Since both output indicators play a positive role in the final effect of expert recommendation, in this embodiment, a positive weighted method is used to calculate the recommendation priority M, and the calculation formula is:

[0090] M=α×Score i +β×S i

[0091] Among them, M represents the recommendation priority. The closer M is to 1, the higher the recommendation. i represents the expert. Score i represents the matching score output by the deep learning recommendation module of expert i, α represents the weight of the matching score output by the deep learning recommendation module, S i represents the matching score output by the indicator recommendation module of expert i, and β represents the weight of the matching score output by the indicator recommendation module. In this embodiment, α+β=1 is set, and the ideal values ​​of α and β are finally determined to be 0.8 and 0.2 through multiple experiments.

[0092] like Figure 8 As shown, this embodiment also provides an expert recommendation method, which is to recommend K most suitable experts for a specific enterprise demand.

[0093] Constructing an expert recommendation system as described above;

[0094] Input enterprise demand data into the enterprise demand classifier based on the BERT model to generate enterprise demand features, determine the demand field to which the enterprise demand belongs through the enterprise demand features, and determine the relevant expert set based on the demand field to which the enterprise demand belongs;

[0095] The expert features and enterprise demand features output by the LDA-based expert modeler are input into the recommendation module to calculate the recommendation priority M of all experts in the relevant expert set;

[0096] Arrange in descending order according to the recommendation priority M. After the arrangement is completed, take the first K experts to obtain the K most suitable experts recommended for the enterprise needs.

[0097] Therefore, the expert recommendation method provided in this embodiment can quickly and accurately match an enterprise with experts who can solve specific technical problems, and is particularly suitable for expert recommendation in school-enterprise cooperation scenarios.

[0098] This embodiment also provides a computer device. A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the expert recommendation method when executing the computer program.

[0099] This embodiment also provides a computer-readable storage medium. A computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, an expert recommendation method described in this embodiment is executed. The computer-readable storage medium can be any tangible medium containing or storing a program, which can be used by or in combination with an instruction execution system, device or device; the program code contained in the computer-readable medium can be transmitted using any appropriate medium, including but not limited to wireless, wire, optical cable, RF, etc., or any suitable combination of the above.

[0100] The above schematically describes the invention and its implementation methods, which is not restrictive. Without departing from the spirit or basic features of the invention, the invention can be implemented in other specific forms. What is shown in the accompanying drawings is only one of the implementation methods of the invention. The actual structure is not limited thereto, and any figure mark in the claims should not limit the claims involved. Therefore, if a person of ordinary skill in the art is inspired by it, without departing from the purpose of the invention, a structural method and an embodiment similar to the technical solution are designed without creativity, which should all belong to the protection scope of the present invention. In addition, the word "including" does not exclude other elements or steps, and the word "one" before the element does not exclude the inclusion of "multiple" elements. The multiple elements stated in the product claim can also be implemented by one element through software or hardware. The words first, second, etc. are used to indicate the name, and do not indicate any specific order.

Claims

1. An expert recommendation system, It is characterized in that include: Data collection module, classification module, storage module, modeling module, recommendation module; The data collection module collects enterprise demand data and pre-processes the enterprise demand data; The classification module constructs an enterprise demand classifier based on the BERT model, inputs enterprise demand data into the enterprise demand classifier based on the BERT model to generate enterprise demand features, and determines the demand field to which the enterprise demand belongs through the enterprise demand features; The storage module stores expert information data and classifies the expert information data to obtain expert feature fields; The modeling module constructs an expert modeler based on the LDA model through the expert's feature fields, and the LDA-based expert modeler outputs expert features; The recommendation module includes a deep learning recommendation module and an indicator recommendation module; the deep learning recommendation module calculates the vector relationship between the enterprise demand characteristics and the expert characteristics and the hidden relationship between the enterprise demand hidden characteristics and the expert hidden characteristics to obtain a matching score, and the indicator recommendation module quantifies other expert indicators to obtain a matching score, and calculates the recommendation priority based on the matching score output by the deep learning recommendation module and the matching score output by the indicator recommendation module.

2. An expert recommendation system according to claim 1, It is characterized in that The step of inputting enterprise demand data into an enterprise demand classifier based on the BERT model to generate enterprise demand features, and determining the demand field to which the enterprise demand belongs through the enterprise demand features comprises: Input enterprise demand into the enterprise demand classifier based on the BERT model to generate enterprise demand features; The vector dimension of enterprise demand features is converted into the number of categories of the demand field classification through the fully connected layer, and the possible vector set of the demand field to which the enterprise demand belongs is obtained; The probability vector set is passed through the Softmax layer to obtain the probability vector of the demand field to which the enterprise demand belongs. The calculation formula of the probability vector is: P = Softmax(FC(O)) Among them, P represents the probability vector, Softmax represents the activation function, FC represents the fully connected layer whose activation function is the ReLu activation function, and O represents the enterprise demand characteristics; Determine the demand area to which the enterprise demand belongs through the probability vector.

3. An expert recommendation system according to claim 2, It is characterized in that The deep learning recommendation module includes a demand-expert encoder and a prediction module; the demand-expert encoder maps the enterprise demand characteristics and expert characteristics to the same dimensional space, performs feature extraction and learning, and obtains the vector relationship between the enterprise demand characteristics and the expert characteristics. The prediction module integrates the vector relationship between the enterprise demand characteristics and the expert characteristics and then reduces the dimension to output the matching score.

4. An expert recommendation system according to claim 3, It is characterized in that Demand--The expert encoder converts the enterprise demand feature embedding into the word vector of the enterprise demand feature, and encodes the word vector of the enterprise demand feature through the encoder to obtain the hidden features of the enterprise demand; Requirements: The expert encoder converts the expert feature embedding into the word vector of the expert feature, and encodes the word vector of the expert feature through the encoder to obtain the expert hidden feature; Calculate the hidden relationship between the hidden features of enterprise needs and the hidden features of experts to obtain the matching probability score. The calculation formula is: Score(V e ,V s )=Sigmoid(FC(V S T V e )) Among them, V e represents the hidden characteristics of enterprise demand, V s Represents expert hidden features, Score(Ve,Vs) represents the matching probability score, Sigmoid represents the Sigmoid activation function, T represents the transposition operation, and FC represents the fully connected layer function.

5. An expert recommendation system according to claim 4, It is characterized in that The matching score of the deep learning recommendation module is obtained by combining the matching score of the vector relationship output of the enterprise demand features and the expert features and the matching probability score of the hidden relationship output of the hidden features of the enterprise demand and the expert hidden features.

6. An expert recommendation system according to claim 5, It is characterized in that Quantitative calculation of other expert indicators, including expert topics, academic achievements and related backgrounds, is performed using the following calculation formula: S=(P+Q+G) / 3 Among them, S represents the matching score output by the indicator recommendation module, P represents the weight of the expert topic part, Q represents the weight of the academic achievement part, and G represents the weight of the relevant background part.

7. An expert recommendation system according to claim 6, It is characterized in that The recommendation priority is calculated by the matching score output by the deep learning recommendation module and the matching score output by the indicator recommendation module. The calculation formula is: M=α×Score i +β×S i Among them, M represents the recommendation priority, i represents the expert, and Score i represents the matching score output by the deep learning recommendation module of expert i, α represents the weight of the matching score output by the deep learning recommendation module, S i represents the matching score output by the indicator recommendation module of expert i, and β represents the weight of the matching score output by the indicator recommendation module.

8. An expert recommendation method, The following steps are involved: Constructing an expert recommendation system based on any one of claims 1-7; Input enterprise demand data into the enterprise demand classifier based on the BERT model to generate enterprise demand features, determine the demand field to which the enterprise demand belongs through the enterprise demand features, and determine the relevant expert set based on the demand field to which the enterprise demand belongs; The expert features and enterprise demand features output by the LDA-based expert modeler are input into the recommendation module to calculate the recommendation priority M of all experts in the relevant expert set; Arrange in descending order according to the recommendation priority M. After the arrangement is completed, take the first K experts to obtain the K most suitable experts recommended for the enterprise needs.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program executable on the processor. It is characterized in that When the processor executes the computer program, the method according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, It is characterized in that When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is executed.

Citation Information

Patent Citations

  • Enterprise demand-oriented expert recommendation method and apparatus, medium and device

    CN112487161A

  • Question and answer community expert recommendation method and system

    CN113254782A

  • Intelligent industry policy recommendation method and system, electronic equipment and medium

    CN114297472A

  • Expert recommendation method and device, computer equipment and storage medium

    CN115292598A

Cited By

  • Patent and expert joint matching method and system for enterprise technology attack demand

    CN119168335A

  • Scientific research project expert matching method based on dynamic academic cooperation network graph representation learning

    CN119226819A