Expert intelligent informatization extraction method fusing correlation

Abstract

The invention relates to the field of artificial intelligence informatization management, in particular to an expert intelligent informatization extraction method fusing correlation. According to the method, expert information, expert relation information, project information and project relation information are utilized to construct word embedding representation fusing expert-project correlation, and then a convolutional neural network and a bidirectional long-short-term memory network are utilized to carry out richer embedding representation on input fusing expert-project correlation. And finally, scoring the experts through an attention mechanism. According to the method, not only are professional fields and comprehensive quality of experts in the expert extraction process considered, but also relation factors such as organization affiliation, a cooperation network, a social network and professional field overlap ratio among the experts are considered, and characteristics of the experts are extracted from multiple dimensions, so that more accurate expert characterization is obtained.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence information management, and specifically to an expert intelligent information extraction method integrating correlation. Background Art

[0002] With the increasing demand for digital governance, the rapid development of information technology and the increase in investment in information technology projects, the application and establishment of information technology projects have become increasingly frequent. Information technology management departments at all levels have gradually achieved comprehensive coverage of information technology systems to improve the management efficiency of information technology projects. As a key link to ensure the quality, fairness and innovation of information technology projects, the role of expert review has become increasingly prominent. Expert review is not only a comprehensive and professional evaluation process of the technical route, innovation, feasibility and expected benefits of information technology projects, but also an important mechanism to guide the rational allocation of resources and promote the development of information technology. However, to give full play to the value of expert review, the first task is to ensure that the most suitable expert team for the review needs of a specific project can be selected accurately, efficiently and fairly from a large number of experts in the field of information technology. Expert selection is not only related to the objectivity, fairness and impartiality of the review work, but also directly affects the quality and efficiency of project review. Therefore, establishing a scientific, rigorous, efficient and fair expert selection mechanism is the basis and prerequisite for ensuring that the role of expert review in information technology project management is effectively played.

[0003] At present, the traditional method of expert selection is manual selection or random selection. Among them, manual selection has problems such as excessive subjectivity, low efficiency, limited information processing ability, and lack of transparency; the random selection method uses a random function to randomly select experts within a period of time. Although it ensures the objectivity of expert selection, the number of times each expert is selected is not balanced in the short term, and the fairness of the experts cannot be guaranteed. Moreover, random selection does not take into account the relationship between experts, which affects the diversity of experts in the expert team, thereby affecting the fairness and impartiality of expert evaluation. Summary of the invention

[0004] In view of the above problems, the present invention proposes an information-based intelligent expert extraction method integrating correlation, which aims to construct a correlation matrix between experts and projects by mining the correlation between experts and projects, and build an extraction model integrating expert-project correlation, and finally obtain the expert's score for the project to be reviewed, and extract experts suitable for the project to be reviewed by the score size. The extraction method integrating expert-project correlation breaks through the limitations of traditional manual or single algorithm extraction methods in extraction accuracy, resource utilization, fairness and efficiency, and realizes the in-depth mining and precise matching of the complex relationship between the multi-dimensional information such as expert business field, professional field, years of work, research results, project review history and the needs of the project to be reviewed. This method can quickly screen out a combination of experts that highly meet the project needs, effectively avoid waste of resources, eliminate subjective bias and interpersonal relationship influence in manual selection, ensure the transparency of the process, traceability of results, and enhance the public's trust in the fairness of the review.

[0005] S1. Establishing association relationships The construction of association relationships mainly includes the association relationships between experts and between projects. The construction of expert association relationships mainly calculates the correlation between experts through expert association relationship features, thereby constructing an expert association relationship network and obtaining an expert relationship matrix for use in downstream tasks of the recommendation algorithm; Expert data comes from the expert database of the information project management platform. Part of the expert information in the expert database is obtained from the data of the talent expert platform, and part is entered through local expert applications. In order to integrate the relationship between experts, it is necessary to calculate the expert correlation through the expert correlation characteristic factor; Establishing the Expert-Project Fusion Correlation Matrix The expert-project fusion correlation matrix is ​​mainly composed of the expert relationship matrix, the project relationship matrix and the expert-project correlation. The expert correlation determined by the above steps obtains the expert relationship matrix, and each expert is vectorized through the BERT encoder. For the number of review experts required to be recommended for a single project, the vectorized representation of the required experts is spliced ​​into an expert feature matrix, and the obtained expert feature matrix is ​​then spliced ​​with the expert relationship matrix to form an expert matrix of fusion correlation; Since each expert reviews multiple historical projects within a year, in order to ensure the consistency of the dimensions of the project matrix and the expert matrix, the number of projects needs to be equal to the number of experts. Therefore, in the fusion expert-project correlation map, according to the value of the relationship between experts and projects, the project with the largest amount of funds is selected from the historical project set reviewed by each expert for vectorization to obtain the project feature matrix. As above, the project feature matrix and the project relationship matrix are spliced ​​row by row to form a project matrix of fusion correlation. Finally, the expert matrix of fusion expert correlation and the project matrix of fusion project correlation are added to obtain the expert-project correlation matrix; S2. Expert Extraction Model The expert extraction model is mainly composed of convolutional neural network (CNN), bidirectional long short-term memory (BiLSTM) network and self-attention mechanism. The expert-project correlation matrix and the feature matrix of the project to be reviewed are used as the input of the convolutional neural network. The expert features integrating the expert-project correlation are extracted through the convolutional neural network. The time series prediction network Bi-LSTM is introduced. Two unidirectional LSTMs are used to extract global features of the sequence in the forward and backward directions to make up for the shortcoming that the convolutional neural network can only extract local features. Then the attention mechanism is introduced to weight the data features. Finally, the probability of the expert for the project to be reviewed is obtained through the softmax layer as the expert score. The review experts for the corresponding project are extracted according to the size of the expert score and the number of experts required. The expert extraction list is obtained according to the expert scores for the project to be reviewed. According to the number of extracted people, the corresponding experts with higher scores are selected as the intelligent extraction results.

[0006] The key points and beneficial effects of this technical solution are: The present invention mainly solves the problems of excessive subjectivity, low efficiency, limited information processing capacity, insufficient transparency and unbalanced expert extraction in the process of manual extraction and random extraction. The word embedding representation integrating expert-project correlation is constructed by using the expert's own information, expert relationship information, project's own information and project relationship information. Then, the convolutional neural network and bidirectional long short-term memory network are used to make a richer embedding representation of the input integrating expert-project correlation. Finally, the experts are scored through the attention mechanism.

[0007] The present invention not only takes into account the professional field and comprehensive quality of experts in the expert extraction process, but also takes into account the organizational affiliation, cooperation network, social network and professional field overlap between experts and other relationship factors, extracts the characteristics of experts from multiple dimensions, and thus obtains more accurate expert representation. In addition, the importance of the expert's historical review projects for expert extraction is also taken into account. Therefore, the most representative review projects of experts in the past year are selected, and the collected structured information is used to determine whether there is a correlation between multiple projects, including industry fields, public support resources, project construction teams, and project funding sources. In the process of expert word embedding, the expert association relationship and the corresponding project association relationship are integrated to further strengthen the historical project representation characteristics of the expert. At the same time, the expert's professional research field and the reviewed projects have time attributes, so the present invention introduces a bidirectional long short-term memory network to extract the global features of the expert sequence and the project sequence, improving the problem that static word embedding does not consider the global attributes of experts and projects.

[0008] In this technical solution, by setting the extraction expert type, number of experts and extraction rules, combined with the collected structured expert information, the input data set of the model can be dynamically adjusted, so as to more accurately extract experts who meet the extraction conditions, thereby improving the efficiency of expert extraction. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 Schematic diagram of the correlation map integrating expert-item correlation; Figure 2 Schematic diagram of expert matrix integrating correlation; Figure 3 Schematic diagram of the expert extraction model integrating correlation. DETAILED DESCRIPTION

[0010] The following is a further explanation based on the case: 1. Establishing association relationships The construction of association relationships mainly includes the association relationships between experts and between projects. The construction of expert association relationships mainly calculates the correlation between experts through the expert association relationship characteristics, thereby constructing an expert association network and obtaining an expert relationship matrix for use in downstream tasks of the recommendation algorithm. The association relationships between experts are mainly reflected in organizational affiliation, cooperation history, social network, and overlap in professional fields.

[0011] Table 1 Expert correlation characteristics and values Expert i-th correlation feature factor f i (e m ,e n ) is expressed as follows: Among them, e m Indicates experts m, e n represents expert n. If the mth expert e m With the nth expert e n If there is an i-th correlation feature between the experts, the i-th correlation feature factor is 1. On the contrary, if there is no i-th correlation feature between the experts, the correlation feature factor value is 0.

[0012] Expert data comes from the expert database of the information project management platform. Part of the expert information in the expert database comes from the data obtained from the talent expert platform, and part of it is entered through local expert applications. The entered information can only be entered into the database after strict review by the competent department to ensure that the expert information in the expert database is true and reliable. In order to integrate the relationship between experts, it is necessary to calculate the expert correlation through the expert correlation characteristic factor. The calculation method is as follows: Among them, rel(e m ,e n ) represents the expert e m and experts n The correlation between i (e m ,e n ) represents expert e m and experts n The i-th characteristic factor of i is the weight of the i-th feature factor among experts, indicating the importance of the feature factor. i The value range of is [0,1] and the sum of the weights of all characteristic factors is 1. The larger the value, the more important the characteristic factor. Specifically, there are 4 expert correlation features in this paper, namely, the same work unit, a history of cooperation, joint participation in project review meetings, and similar professional research fields. Among them, similar professional research fields have the greatest impact on expert correlation, followed by a history of cooperation and joint participation in project reviews. The same work unit has the least impact on expert correlation, so the corresponding characteristic factor weights are 0.15, 0.3, 0.2, and 0.35, respectively. L is the number of correlation features, E is the number of correlation features, and L is the number of correlation features. m For experts m The set of all experts with whom you have a relationship.

[0013] The relationship between projects mainly covers multiple dimensions such as project field matching, shared infrastructure and resources, teams and participants, and common funding sources. For example, whether the projects focus on the same industry field, whether the projects use the same public support resources, whether the project construction team has the same personnel, and whether the project funding sources are the same. The relationship between the above projects is used as the feature of project correlation calculation, and the project relationship network is constructed through the reasonable combination of project correlation features. The project correlation features are shown in Table 2 below.

[0014] Table 2 Project correlation characteristics and values Similarly, the correlation between items is calculated based on the expert correlation characteristic factor calculation formula (1) and the expert correlation calculation formula (2).

[0015] Through the correlation between experts and projects, we can draw a correlation graph. For example, experts are taken as the main nodes, and the correlation between experts is taken as the edge, marked as Erel, and the value is the correlation between experts. Projects are taken as secondary nodes, and the correlation between projects is taken as the edge, marked as Prel, and the value is the correlation between projects. Based on the historical data of expert review projects, the relationship between experts and projects is established. If the expert has participated in the review of the project in the past year, there is a relationship between the expert and the project, marked as the edge E-Prel, and the value is the amount of funds involved in the project. The correlation graph integrating expert-project correlation is shown in the figure below: Figure 1 shown.

[0016] 2. Establish an expert-project integration correlation matrix The expert-project fusion correlation matrix is ​​mainly composed of the expert relationship matrix, the project relationship matrix and the expert-project association relationship. The expert association relationship determined through the above steps can obtain the expert relationship matrix. Each expert is vectorized through the BERT encoder. For the number of review experts required to be recommended for a single project, the vectorized representation of the required experts is spliced ​​into an expert feature matrix. The obtained expert feature matrix is ​​then spliced ​​with the expert relationship matrix to form an expert matrix of fusion correlation. For example, a certain project requires 6 review experts. The vector of each expert encoded by the BERT encoder is a 1×128-dimensional vector. The 6 experts can be spliced ​​to obtain a 6×128 expert feature matrix. The 6×128 expert matrix and the 6×6 expert relationship matrix are spliced ​​into a 6×134 fusion correlation matrix. The expert matrix of fusion correlation is as follows: Figure 2 shown.

[0017] Since each expert reviews multiple historical projects within a year, in order to ensure the consistency of the dimensions of the project matrix and the expert matrix, the number of projects needs to be equal to the number of experts. Therefore, in the fusion expert-project correlation map, according to the value of the relationship between experts and projects, E-Prel, the project with the largest amount of funds is selected from the historical project set reviewed by each expert for vectorization to obtain the feature matrix of the project. Similarly, the project feature matrix and the project relationship matrix are concatenated row by row to form a project matrix of fusion correlation. Finally, the expert matrix of fusion expert correlation and the project matrix of fusion project correlation are added to obtain the expert-project correlation matrix.

[0018] 3. Expert Extraction Model The expert extraction model is mainly composed of a convolutional neural network (CNN), a bidirectional long short-term memory network (BiLSTM), and a self-attention mechanism. Among them, the expert-project correlation matrix and the feature matrix of the project to be reviewed are used as the input of the convolutional neural network, and the expert features that integrate the expert-project correlation are extracted through the convolutional neural network. Since the expert review projects are time-sensitive, the time series prediction network Bi-LSTM is introduced. By using two unidirectional LSTMs to extract global features of the sequence in the forward and backward directions, the disadvantage that the convolutional neural network can only extract local features is compensated. Then the attention mechanism is introduced to assign weights to the data features, and finally the probability of the expert's evaluation of the project to be reviewed is obtained through the softmax layer as the expert's score. The review experts for the corresponding project are extracted according to the size of the expert score and the required number of experts. The expert extraction model is as follows: Figure 3 shown.

[0019] 4. Convolutional Neural Network Layer The convolutional neural network layer is mainly composed of an embedding layer, a convolution layer, a pooling layer, a fully connected layer, etc. The embedding layer takes the m×n fusion correlation matrix concatenated with the feature matrix and the relationship matrix as input. In order to retain as much feature information as possible, zero padding is used to keep the feature size consistent. The expert data to be processed is defined as E={e 1 ,e 2 ,...,e n}, the embedded data of this data after the embedding layer is H E ={h e1 ,h e2 ,...,h en}, where h en For experts n The embedded representation of .

[0020] The embedded data is convolved through the convolution layer, and the features after convolution are c i t The calculation formula is as follows: Where W t and b t are embedded data H i The main purpose of convolution is to extract features from input data and assign weights to features. The essence of convolution operation is matrix operation. The convolution result is obtained by convolving different input data matrices with convolution kernels in sequence. The convolution result is processed by activation function and then pooled. The activation function is Sigmoid function, which is expressed as follows: The pooling layer performs dimensionality reduction on the result after the activation function processing and compresses the subsequent input. The pooling layer mainly uses the maximum pooling method to retain the most significant features of experts and projects. By selecting the kernel size and step size, the fixed-size pooling result is calculated.

[0021] The purpose of the fully connected layer is to process the pooling result into the input of the bidirectional long short-term memory layer. The fully connected layer can integrate the local information of the data in the pooling layer. The output u of the nth neuron in the kth layer in the fully connected m k as follows: In the formula, x m k-1 It is represented as the eigenvalue of the mth eigenvector in the k-1th layer, w mn k is the weight corresponding to the eigenvalue of the mth eigenvector in the kth layer, b m k is the corresponding offset value.

[0022] 5. Bidirectional long short-term memory layer LSTM is a special recurrent neural network that can effectively solve the problem of long-term dependency. It controls the forgetting, updating and output of information through a gating mechanism, and is suitable for processing text sequence or time series data. BiLSTM combines forward and reverse LSTM, and can simultaneously consider the past and future contextual information of sequence data, which is crucial for understanding the temporal order relationship in the expert context, such as the expert's historical cooperation record, the projects involved in the review over a period of time, etc.

[0023] BiLSTM is composed of forward and reverse LSTM networks. The output state h at time t in the forward LSTM network is t f With the input at time t and the hidden state h at time t-1 t-1 f Related, expressed as follows: Similarly, the output state h at time t in the reverse LSTM network t b With the input at time t and the hidden state h at time t-1 t-1 b The calculation method is as follows: The output states of the forward and reverse LSTM networks together form the overall output state h of the entire BiLSTM t , expressed as follows: 6. Attention Mechanism The self-attention mechanism allows the model to dynamically assign different weights based on the importance of different parts of the input, thereby globally considering the information in the input sequence. In the expert extraction scenario, this means that the model can independently judge the importance of various attributes of the expert (such as work unit, cooperation history, and professional field) in determining relevance. By learning the relationship between elements in the input sequence, the self-attention mechanism can more finely capture which features are most critical to the expert's relevance judgment. For example, if two experts have a frequent cooperation history, then this feature may be given a higher weight. The self-attention mechanism enhances the model's ability to understand long sequence data, allowing the model to make more accurate decisions when faced with complex and nonlinear relationships, thereby improving the accuracy and efficiency of extracting experts.

[0024] The output of the feature encoding with time series characteristics of the BiLSTM network is weighted according to different feature encodings. The attention mechanism consists of input, attention score weight, and attention function. The output sequence of the BiLSTM network is used as the input of the attention mechanism, denoted as X={x 1 ,x 2 ,......,x n}. The attention score function is calculated as follows: Where q is the query and d is the length of the input sequence. The attention weight corresponding to each value is determined according to the scoring function, and the attention weight is calculated as follows: The expert score is calculated based on the input sequence and attention weights, and is calculated as follows: The expert extraction list is obtained by sorting the experts' scores on the items to be reviewed. The corresponding experts with higher scores are selected as the intelligent extraction results according to the number of people extracted. Specific embodiment: The specific implementation method mainly includes the following steps: 1. Data preprocessing Expert information and project information are mainly collected and managed by the provincial government information project management system. The collected expert information includes work units, professional research fields, review records, educational background, research results and other information, covering about 10,000 information experts across the province. Project information includes project fields, budget funds, funding sources, public support, construction teams, planning cycles and other information. A total of 4,926 project cases are entered as project data sources. Before vectorizing structured text, it is necessary to preprocess the above expert information and project information. By removing duplicate values, processing null values, removing special characters and other operations, 9,455 expert information and 4,926 project data are retained.

[0026] 2. Feature extraction Expert information is extracted by inputting the expert structured attribute data preprocessed in the above steps. For example, the attribute work unit of expert 1 is {China Electronics Wanwei Information Technology Co., Ltd.}, and the result after word segmentation is {"China Electronics", "Wanwei", "Information", "Technology", "Limited", "Liability", "Company"}. The data after word segmentation is encoded into a vector representation of the corresponding attribute by the BERT encoder. Multiple attributes of expert 1 are encoded by the BERT encoder in turn, and the vectors generated by the encoding of multiple attributes are spliced ​​into a vector to represent the characteristics of expert 1. For expert attributes, there are 4 categories in this embodiment, mainly covering basic information of experts, professional fields, social networks, and achievements and honors. There are 6 attributes such as age, gender, educational background, work experience, position and title, and expert type in the basic information. The professional field includes 3 attributes of research direction, published papers, and project experience. The social network mainly includes 2 attributes such as work unit and cooperation network. There are 2 attributes of award records and scientific research results in the achievements and honors. The above 13 expert attributes are encoded and spliced ​​in turn to obtain a 1×100-dimensional feature vector of expert 1. If 6 experts are needed to review the project, the size of the expert feature vector obtained by BERT encoding and concatenation is 6×100.

[0027] Project feature extraction mainly extracts the features of the projects with the most project funds that have been reviewed by experts in the past year. For example, expert 1 reviewed 16 projects in the past year, and selected the project with the largest project funds. The project name is {"Government Service Integration Platform Construction Project"}, and the result after word segmentation is {"Government Affairs", "Service", "Integration", "Platform", "Construction", "Project"}. The data after word segmentation is encoded into a project vector representation of the corresponding attributes through the BERT encoder. The attributes of the project mainly include 7 attributes: project name, project industry, project construction content, project budget, project funding source, project team staffing, and project public support resources. Similarly, the remaining 5 experts participated in the review of the largest project funds in the past year and obtained a project feature vector of size 6×100 after encoding through BERT.

[0028] 3. Correlation calculation In this embodiment, the expert correlation is first calculated, and the correlation factor of the above-mentioned experts who have been subjected to feature extraction is determined based on the expert correlation feature. For example, the above-mentioned 6 experts who have undergone feature extraction are recorded as E={e 1 ,e 2 ,e 3 ,e 4 ,e 5 ,e 6}, Expert 1 and e 2 There are characteristics of working in the same unit and participating in project review together, experts e 1 and e 5 Have a history of cooperation, experts 2 and e 3 There is a history of cooperation between them and their professional research fields are similar. 3 and e 4 They have a history of cooperation and have participated in reviews together. Experts e3 and e5 have similar professional research fields and have participated in project reviews together. Experts e3 and e6 have similar professional fields, have participated in review meetings together, and have a history of cooperation. The expert correlation characteristic factors determined by formula (1) are shown in Table 3.

[0029] Table 3 Expert correlation characteristic factors Then, the correlation between experts was calculated according to formula (3), and the weights of the correlation features of the same work unit, cooperation history, joint review, and similar professional fields were set to 0.15, 0.3, 0.2, and 0.35 respectively. The expert e 1 -e 6 The correlation results are shown in Table 4 below, thus obtaining an expert relationship matrix of size 6×6. 1There are 10 experts with related relationships. According to formula (2), we can calculate expert e 1 and experts 2 The correlation index q(e 1 ,e 2 ) is about 1.4, Expert 1 and e 5 The correlation index q(e 1 ,e 5 ) is about 1.3, and then the expert e is calculated according to formula (3) 1 Respectively with e 3 and e 5 The correlations are 0.22 and 0.21. 1 With e 3 、e 5 Although the correlation characteristics between them are different, the correlation between them is basically equal through calculation, which shows that the expert e 1 With e 3 、e 5 There is a similar correlation between them.

[0030] Table 4 Expert correlation Similarly, the item correlations are shown in Table 5, and a 6×6 item relationship matrix is ​​calculated.

[0031] Table 5 Project correlation 4. Fusion Expert-Project Relevance In this embodiment, the fusion of expert-project correlation is to fuse the feature matrix and relationship matrix of the expert-project obtained in the above steps 2 and 3. Specifically, the expert feature matrix obtained by the 6 experts is a feature representation of size 6×100, and the size of the relationship matrix between the experts calculated is 6×6. The number of row vectors of the two matrices is consistent. Through the row splicing method, the matrix representing the expert features and the expert relationship is spliced ​​into a 6×106-dimensional fusion expert correlation matrix. Similarly, the project feature matrix and the project relationship matrix are spliced ​​into a 6×106-dimensional fusion project correlation matrix. Finally, the corresponding elements of the expert matrix and the project matrix with the same two dimensions are added to obtain an expert-project matrix that integrates the correlation between experts and projects. In the process of integrating correlation, not only the characteristics and relationships of experts are considered, but also the characteristics and relationships of historical projects reviewed by experts are integrated. The obtained expert-project matrix can more comprehensively reflect the information of experts and provide a basis for the subsequent matching process.

[0032] 5. Model parameters The expert-project correlation matrix and the project to be reviewed are used as the input of the convolutional neural network. The convolutional neural network is used to extract expert features with project attributes. A bidirectional long short-term memory network is introduced to extract global features of experts to ensure that the expert features have temporal sequence. Specifically, the convolutional neural network includes an input layer, a convolution layer, a pooling layer, and a fully connected layer. There are n experts in the input layer, and the fusion correlation sequence of each expert is represented by s n =[r 1 ,r 2 ,…,r m ], then the fusion correlation sequence of n experts is expressed as V=[s 1 ,s 2 ,…,s n ] T , which is the input of the convolutional neural network. The above sequence is convolved through the convolution layer to obtain the data sequence H e =[h 1 ,h 2 ,…,h s ], after the convolution layer, in order to reduce the dimension of the sequence, it is necessary to perform a pooling operation on the convolutional sequence. This paper adopts the maximum pooling method to pool h s The small blocks obtained by pooling are spliced ​​and the maximum value is taken. The maximum value after splicing constitutes a vector C=[c 1 ,c 2 ,…,c i ]. The data after the pooling layer is integrated through the fully connected layer, and the output of the jth neuron in the lth layer of the fully connected layer is represented as y j l The data output by the fully connected layer is input into the bidirectional long short-term memory network for forward and reverse feature extraction to extract the temporal features. The output of the feature encoding with temporal characteristics of the BiLSTM network is weighted according to the attention mechanism and different feature encodings. The weighted sequence is calculated according to formula (11) to calculate the input expert score.

[0033] The model parameters are set as follows: In the convolutional neural network, there are two convolution layers. The first convolution kernel size is 2, the stride is 1, and the number of convolution kernels is 16; the second convolution kernel size is 3, the stride is set to 1, and the number of convolution kernels is 32. The corresponding pooling layer also requires two layers, the size of the pooling layer is 2, and the stride is 1.

[0034] In the bidirectional long short-term memory network, the number of hidden neurons is set to 30, the number of network layers is 1, the learning rate is 0.01, the training batch size is 256, and the number of iterations is 24.

[0035] 6. Expert rating list This embodiment collects 473 project data and 3922 expert data through the provincial government information project management system, divides the data into training set and test set according to the ratio of 7:3, tests the model through the above model and model parameters, obtains the expert evaluation score for the input project to be reviewed, sorts the expert evaluation list, and extracts the top n experts as the evaluation experts of the project to be reviewed. For example, the project to be reviewed needs expert evaluation of the project establishment, and the basic information of the project, the field to which it belongs, the construction content, the source of project funds, the project budget, the staffing of the project team, the public support resources of the project and other information are collected through the provincial government information project management system. According to the setting of the expert extraction type, the number of expert extractions and the expert extraction rules, the extraction setting requires 2 university experts, 3 head enterprise experts and the expert work units cannot be the same. Through the set expert type, the model can automatically select the corresponding expert data set. In this embodiment, the university expert data is 1600 and the head enterprise expert data is 524. Since the experts' work units cannot be the same before extraction, if experts from different work units are used as the input of the entire extraction model, the expert correlation matrix will be sparse, and there are too many random combination possibilities between different units to accurately determine the input of the model. Therefore, in this embodiment, experts from the same unit are selected as input, so that it can be ensured that there is at least one correlation relationship between the experts input each time, and the expert scores of the same unit are output through the model. After obtaining the expert score list of each unit, the expert scores are compared between different units according to the set number of experts extracted, and the experts from different units with the highest final scores are selected to participate in the project review.

[0036] 7. Experimental analysis and comparison In essence, what is finally obtained by expert extraction is the expert score ranking. The expert model is evaluated according to the universal indicator NDCG for measuring ranking quality. In order to fully illustrate the advanced nature of this case, the following two rounds of experiments are set up in this case.

[0037] Experiment 1: The technical solution of this case (denoted as P 1 ) and the technical solution without using the expert-item fusion correlation matrix (denoted as P 2 ) was used for experiments. The experimental data source was provided by the government information project management system. The experimental model parameters used the model parameters of step 5 in the embodiment. The number of experts extracted was set to 5 and 10 respectively. Two groups of experiments were conducted. The experimental results are shown in Table 6 below.

[0038] Table 6 P 1 With P 2 Average NDCG value Experiment 2: The technical solution of this case (denoted as P 1) and the technical solution without adding the bidirectional long short-term memory layer (denoted as P 3 ) was used for experiments. The experimental data source was provided by the government information project management system. The experimental model parameters used the model parameters of step 5 in the embodiment. The number of experts extracted was set to 5 and 10 respectively. Two groups of experiments were conducted. The experimental results are shown in Table 7 below.

[0039] Table 7 P 1 With P 3 Average NDCG value From the above experimental results, it can be seen that the technical solution using the expert-item correlation matrix has a significant improvement in the average NDCG index compared with the technical solution not using it, and it is also improved compared with the technical solution without adding the bidirectional long short-term memory layer. In addition, the performance of this technical solution is affected by the number of extracted experts.

[0040] To sum up, the advanced nature of this case is mainly reflected in the in-depth exploration of the relationship between experts and projects before expert extraction; secondly, a bidirectional long short-term memory layer is added on the basis of the convolutional neural network when extracting expert features, which makes up for the defect that the convolutional neural network only focuses on local features and improves the accuracy and fairness of expert extraction.

[0041] Compared with the existing methods, the innovations of the method described in this paper are mainly: 1. The patent with publication number CN118115032A discloses an expert extraction method based on intelligent talent evaluation. The method extracts experts through the steps of data collection, data preprocessing, feature engineering, model selection and training, model evaluation and optimization, expert extraction, result analysis and feedback. Compared with the above-mentioned patent, the distinguishing technical features of this application are: (1) In this case, we first establish an expert-project fusion correlation matrix, fuse the expert association relationship and the project association relationship to obtain the expert’s feature representation vector. In the extraction process, we take the expert relationship into consideration instead of treating the experts as independent entities, thus ensuring the relevance and richness of the expert’s features.

[0042] (2) The above-mentioned patent disclosure does not mention a specific feature extraction model. This case clarifies that the technical solution of the expert extraction model consists of a convolutional neural network (CNN), a bidirectional long short-term memory network (BiLSTM), and a self-attention mechanism module, which ultimately obtains expert scores to select the most appropriate review experts.

[0043] 2. The patent with publication number CN109886770A discloses a regional and professionally scalable method for extracting bid evaluation experts. This method mainly extracts experts through a random extraction model. When the number of extractors is insufficient, the extraction range can be expanded to the superior region or related professions, and the extraction weight is mainly determined by the number of historical evaluations of the experts. Compared with the above-mentioned patents, the main innovation of this application is that the above-mentioned patent 2 extracts experts through a random function combined with the number of historical evaluations of experts. Although this method can ensure the objectivity of expert extraction, the number of times each expert is selected is not balanced in the short term, and it cannot be guaranteed that the extracted experts are as unrelated as possible, thus affecting the fairness of expert extraction. In the implementation process of this case, the correlation between experts, the correlation between projects, and the correlation between experts and projects are considered. By constructing an expert-project correlation map, an expert-project fusion correlation matrix is ​​obtained, and the correlation of experts is fully considered. The experts with the smallest correlation and the highest score are selected as the extraction results, ensuring that the correlation between the extracted experts is as small as possible.

[0044] 3. The patent with publication number CN114510918A discloses an expert matching method and device, which extracts the first keyword of the project text through the TFIDF algorithm, extracts the second keyword of the project text through the TextRank algorithm, and then determines the project label based on the extracted first keyword and second keyword, calculates the similarity between the project label and the expert label of the target expert collected in advance, and finally adjusts the weights of multiple similarities to determine the matching degree between the project and the expert. Compared with the above-mentioned patents, the main innovations of this application are: (1) Both the TFIDF algorithm and the TextRank algorithm described in the above patent 3 are heavily dependent on the word segmentation results of the text. The core is to use high-frequency words that appear in the text as keywords to label the project. This paper mainly uses convolutional neural networks to extract features from the review items. It does not rely on manually designed features and can extract higher-level semantic features.

[0045] (2) The matching degree between the project and the expert in the above-mentioned Patent 3 mainly depends on the expert labels of the target experts collected in advance. Improper setting of expert labels and inaccurate collection of expert labels will directly affect the results of expert extraction. This paper mainly constructs a feature matrix that can represent experts by integrating the correlation features of experts and projects, and then extracts the local features of the expert sequence through a convolutional neural network, introduces a bidirectional long short-term memory network to add expert temporal features, and assigns weights to data features through a self-attention mechanism, and finally obtains the expert's fitness score for the project to be reviewed.

[0046] In summary, this technical solution is innovative.

[0047] Glossary Expert extraction: Expert extraction refers to a method of selecting suitable experts or professionals from a larger expert pool or professional talent pool to participate in a work process in certain highly professional jobs or activities based on specific standards and needs.

[0048] Convolutional Neural Networks (CNN): A type of feedforward neural network with deep structure and convolutional computing. It has representation learning capabilities and can perform translation-invariant classification of input information according to its hierarchical structure.

[0049] Long Short-Term Memory (LSTM): a time-recurrent neural network designed to solve the long-term dependency problem of general recurrent neural networks.

Claims

1. An expert intelligent information extraction method integrating correlation, characterized in that These include: S1. Establishing association relationships The construction of association relationships mainly includes the association relationships between experts and between projects; The construction of expert association relationship is mainly to calculate the correlation between experts through the expert association relationship characteristics, so as to construct the expert association relationship network and obtain the expert relationship matrix for the downstream tasks of the recommendation algorithm. The expert i-th correlation feature factor f i (e m ,e n ) is expressed as follows: Among them, e m Indicates experts m, e n represents expert n; if the mth expert e m With the nth expert e n If there is an i-th correlation feature between experts, the i-th correlation feature factor is 1; otherwise, if there is no i-th correlation feature between experts, the correlation feature factor value is 0; Expert data comes from the expert database of the information project management platform. Part of the expert information in the expert database is obtained from the data of the talent expert platform, and part is entered through local expert application. In order to integrate the relationship between experts, it is necessary to calculate the expert correlation through the expert correlation characteristic factor. The calculation method is as follows: Among them, rel(e m ,e n ) represents the expert e m and experts n The correlation between i (e m ,e n ) represents expert e m and experts n The i-th characteristic factor of i is the weight of the i-th feature factor among experts, indicating the importance of the feature factor; α i The value range of is [0,1] and the sum of the weights of all characteristic factors is 1. The larger the value, the more important the characteristic factor is. Establishing the Expert-Project Fusion Correlation Matrix The expert-project fusion correlation matrix is ​​mainly composed of the expert relationship matrix, the project relationship matrix and the expert-project correlation relationship; The expert relationship matrix is ​​obtained by determining the expert association relationship through the above steps. Each expert is vectorized through the BERT encoder. According to the number of review experts required to be recommended for a single project, the vectorized representations of the required experts are spliced ​​into an expert feature matrix. The obtained expert feature matrix is ​​then spliced ​​with the expert relationship matrix to form an expert matrix with fusion correlation. Since each expert reviews multiple historical projects within a year, in order to ensure the consistency of the dimensions of the project matrix and the expert matrix, the number of projects needs to be equal to the number of experts. Therefore, in the fusion expert-project correlation map, according to the value of the relationship E-Prel between experts and projects, the project with the largest amount of funds is selected from the historical project set reviewed by each expert for vectorization to obtain the feature matrix of the project. As above, the project feature matrix and the project relationship matrix are spliced ​​row by row to form a fusion correlation project matrix; Finally, the expert matrix integrating expert correlation and the project matrix integrating project correlation are added to obtain the expert-project correlation matrix; S2. Expert Extraction Model The expert extraction model is mainly composed of convolutional neural network CNN, bidirectional long short-term memory network BiLSTM and self-attention mechanism; The expert-project correlation matrix and the feature matrix of the project to be reviewed are used as the input of the convolutional neural network, and the expert features integrating the expert-project correlation are extracted through the convolutional neural network. The time series prediction network Bi-LSTM is introduced to extract global features of the sequence in the forward and backward directions by using two unidirectional LSTMs to make up for the shortcoming that the convolutional neural network can only extract local features. Then the attention mechanism is introduced to assign weights to the data features, and finally the probability of the expert for the project to be reviewed is obtained through the softmax layer as the expert score, and the review experts of the corresponding project are extracted according to the size of the expert score and the required number of experts. (1) Convolutional Neural Network Layer The convolutional neural network layer is mainly composed of an embedding layer, a convolution layer, a pooling layer, a fully connected layer, etc. The embedding layer takes the m×n fusion correlation matrix concatenated with the feature matrix and the relationship matrix as input. In order to retain as much feature information as possible, zero padding is used to keep the feature size consistent. The expert data to be processed is defined as E={e1,e2,...,e n }, the embedded data of this data after the embedding layer is H E ={h e1 ,h e2 ,...,h en }, where h en For experts n Embedding representation of ; The embedded data is convolved through the convolution layer, and the features after convolution are c i t The calculation formula is as follows: Where W t and b t are embedded data H i The main purpose of convolution is to extract features from input data and assign weights to features. The essence of convolution operation is matrix operation. The convolution result is obtained by convolving different input data matrices with convolution kernels in turn. The convolution result is processed by activation function and then pooled. The activation function selects Sigmoid function, which is expressed as follows: The pooling layer performs dimensionality reduction on the result after the activation function processing and compresses the subsequent input. The pooling layer mainly adopts the maximum pooling method to retain the most significant features of experts and projects. By selecting the kernel size and step size, the fixed-size pooling result is calculated. The purpose of the fully connected layer is to process the pooling result into the input of the bidirectional long short-term memory layer. The fully connected layer can integrate the local information of the data in the pooling layer. The output u of the nth neuron in the kth layer in the fully connected m k as follows: In the formula, x m k-1 It is represented as the eigenvalue of the mth eigenvector in the k-1th layer, w mn k is the weight corresponding to the eigenvalue of the mth eigenvector in the kth layer, b m k is the corresponding bias value; (2) Bidirectional long short-term memory layer LSTM controls the forgetting, updating, and output of information through a gating mechanism, and is suitable for processing text sequence or time series data; BiLSTM combines forward and backward LSTM, and can simultaneously consider the past and future contextual information of sequence data, which is crucial for understanding the temporal order relationship in the expert context; BiLSTM is composed of forward and backward LSTM networks. The output state h at time t in the forward LSTM network t f With the input at time t and the hidden state h at time t-1 t-1 f Related, expressed as follows: Similarly, the output state h at time t in the reverse LSTM network t b With the input at time t and the hidden state h at time t-1 t-1 b The calculation method is as follows: The output states of the forward and reverse LSTM networks together form the overall output state h of the entire BiLSTM t , expressed as follows: (3) Attention Mechanism The output of the feature encoding with time series characteristics of the BiLSTM network is weighted according to different feature encodings. The attention mechanism consists of input, attention score weight and attention function. The output sequence of the BiLSTM network is used as the input of the attention mechanism, denoted as X={x1,x2,......,x n }; The attention score function is calculated as follows: Where q is the query and d is the length of the input sequence; The attention weight corresponding to each value is determined according to the scoring function. The attention weight is calculated as follows: The expert score is calculated based on the input sequence and attention weights, and is calculated as follows: Sort the items to be reviewed by the experts according to their scores to obtain an expert extraction list; According to the number of people drawn, the corresponding experts with higher scores are selected as the intelligent extraction results.

Images