Multi-source expert network construction and industrial science and technology talent recommendation method based on large model

Abstract

The present invention relates to the field of artificial intelligence and big data retrieval and classification technology, in particular to a method for constructing a multi-source expert network and recommending industrial science and technology talents based on a large model. The present invention comprises: obtaining expert data from different sources, constructing a corresponding initial expert network for each data source; constructing a Transformer neural network tower corresponding to each initial expert network, assigning the characteristic parameters of the initial expert network to the corresponding Transformer neural network tower, and having the Transformer neural network tower encode each node of the corresponding initial expert network as an expert vector; performing dual-tower fusion on all Transformer neural network towers in sequence, obtaining a shared space of all Transformer neural networks, embedding the expert vectors of all Transformer neural networks into the shared space, and obtaining a fused expert representation; and constructing an expert cooperation network based on the fused expert representation. The present invention constructs a distributed data acquisition big data platform to obtain raw data, and adopts a dual-tower fusion method to successively fuse expert information from different data sources, thereby achieving high-precision homonym disambiguation and expert information fusion, and ensuring that the ultimately obtained expert representation is accurate and complete.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and big data retrieval and classification technology, and in particular to a method for constructing a multi-source expert network and recommending industrial science and technology talents based on a large model. Background Art

[0002] In today's era of rapid technological advancement and ongoing industrial restructuring, society's demand for talent is undergoing structural changes, making it increasingly difficult to find individuals with specific skills and knowledge in specific industries and technological fields. Due to the high costs of traditional talent recruitment methods such as job fairs and headhunting, the use of artificial intelligence and big data analytics to recommend industry and technological talent is more cost-effective and effective. In particular, the recent rise of pre-trained large-scale model technology makes data-based talent recommendations more intelligent and precise.

[0003] However, existing talent recommendation technologies have the following problems:

[0004] 1. During data collection, there was no emphasis on how to solve problems such as the continuous growth of open source data, inconsistent raw data formats, and uneven data quality, which affected subsequent data fusion and processing.

[0005] 2. The accuracy of disambiguating expert information by using a probabilistic graphical model is not high.

[0006] 3. The construction of the scientific and technological knowledge graph is based solely on keywords, similarities, and existing semantic relationships, which lacks scientificity and usability. Moreover, only constructing the scientific and technological knowledge graph may result in some talents engaged in industrial applications being unable to be classified, resulting in some talents being unable to be recommended.

[0007] 4. The talent classification module uses manual labeling and convolutional neural network training to classify talents. This method requires a large amount of manual labeling, which is very costly. It also converts a large amount of expert information into a low-dimensional vector, losing a large amount of high-dimensional information, resulting in low classification accuracy. In addition, the sample training method cannot learn newly added category labels.

[0008] 5. In the talent intelligent recommendation module, the use of knowledge graph retrieval requires users to have a certain technical foundation, which increases the usage threshold and is not very intelligent; semantic-based intelligent recommendation uses word segmentation and similarity calculation methods, which cannot understand the deep semantics of user needs, resulting in low recommendation results and matching degree. Summary of the Invention

[0009] In order to overcome the defect of low talent recommendation accuracy in the above-mentioned existing technology, the present invention proposes a method for constructing a multi-source expert network, which can integrate expert information from different sources to build an accurate and complete expert cooperation network that can display expert information from multiple angles.

[0010] The present invention proposes a method for constructing a multi-source expert network, comprising:

[0011] SA1. Obtain expert data from different sources and build corresponding initial expert networks for each data source;

[0012] SA2. Construct a Transformer neural network tower corresponding to each initial expert network, assign the characteristic parameters of the initial expert network to the corresponding Transformer neural network tower, and let the Transformer neural network tower encode each node of the corresponding initial expert network as an expert vector;

[0013] SA3: Perform dual-tower fusion on all Transformer neural network towers in sequence to obtain the shared space of all Transformer neural networks, embed the expert vectors of all Transformer neural networks into the shared space, and obtain the fused expert representation;

[0014] Expert representation includes: expert’s professional field, affiliation, research results and expert collaboration;

[0015] SA4. Build an expert cooperation network based on the integrated expert representations.

[0016] Preferably, the method for fusing all Transformer neural network towers to obtain expert representation includes: constructing a tower set to count all Transformer neural network towers; merging any two Transformer neural network towers in the tower set, replacing the fused towers in the tower set with the merged tower, traversing the tower set until only one Transformer neural network tower remains in the tower set, then obtaining the expert information space of the only Transformer neural network tower in the tower set as the shared space, and obtaining the fused expert representation.

[0017] Preferably, the two Transformer neural network towers are merged as follows: traverse each expert vector in tower A and search for a paired expert vector of the expert vector of tower A in tower B; when the similarity between an expert vector B1 in tower B and a specified expert vector A1 of tower A is the largest and greater than a set similarity threshold, the expert vector B1 is paired with the expert vector A1; based on the pairing relationship, the expert information corresponding to the expert vectors in towers A and B is fused, and the fused Transformer neural network tower is constructed based on the fused expert representation.

[0018] The present invention proposes a method for recommending industrial and scientific talents based on a large model, which is characterized by comprising the following steps:

[0019] S1. Build a big data platform to collect, store, integrate, and eliminate homonyms in distributed talent data, forming a big data warehouse for industry and technology talent;

[0020] S2. Combining expert knowledge and large models to assist in building segmented industry knowledge graphs (IK) and technology knowledge graphs (SK), and automatically classifying talents into industry and technology fields;

[0021] S3. Create a scoring profile for each talent, marking the talent's technical ability score (TS) and background security score (BS);

[0022] S4. Combine the industry knowledge graph IK, technology knowledge graph SK and scoring portraits to develop and recommend talents.

[0023] Preferably, in S1, a distributed parallel crawling framework of scrapy combined with redis is used to obtain relevant information of scientific and technological experts and industry experts from various open source platforms, and a multi-source basic database of talents is constructed in hive. Then, the method for constructing a multi-source expert network is used to obtain expert representations after the expert data from different sources in the multi-source basic database of talents are integrated to form a big data warehouse of industry and scientific and technological talents.

[0024] Preferably, the construction of the subdivided industry knowledge graph IK and the science and technology knowledge graph SK in S2 includes the following steps:

[0025] S21. First, select a field and use expert knowledge in the selected field to build a seed knowledge graph;

[0026] S22. Use the pre-trained domain classification model to extract corresponding entities and upstream and downstream relationships from domain public information, and add the entities to the seed knowledge graph based on the upstream and downstream relationships; domain public information includes: patent texts, papers, monographs, reports, achievements, and several items of domain standards;

[0027] S23. Combine expert knowledge to manually proofread and revise the expanded knowledge graph to obtain the domain knowledge graph.

[0028] Preferably, the method for classifying talents in a selected field is: obtaining the domain knowledge graph corresponding to the selected field, the domain knowledge graph including the subdivided industry knowledge graph IK and the science and technology knowledge graph SK; constructing a classification task prompt for the talent information text based on the entities and relationships in the domain knowledge graph, the entities are classification labels, and the relationships represent the hierarchy of each label; using a pre-trained classification model combined with the domain knowledge graph to classify the talent information text, and obtain the industry category label and science and technology category label of the talent.

[0029] Preferably, in S4, the field to which the target talent belongs is first determined in the target knowledge graph, and the search scope on the target knowledge graph is obtained; the talent search request is input into the pre-trained search model, which understands the talent search request and generates a corresponding search strategy; the search strategy is then executed in the target knowledge graph to retrieve candidate talents; the pre-trained talent recommendation model ranks the candidate talents, selects the candidate talents as target talents according to the ranking order from high to low, and recommends them to the user; and the talent recommendation model is reinforced and trained based on customer feedback;

[0030] The similarity between candidate talents and talent search requests is calculated as follows: vectorizing the candidate talent profile and the talent search request, and calculating the similarity between the two vectors as the similarity PS between the candidate talents and the talent search request; the talent search request includes the search scope on the target knowledge graph;

[0031] The input of the talent recommendation model includes: annotated candidate talents and task requirement text. The annotations of candidate talents are: technical ability score TS, background security score BS and similarity PS.

[0032] Preferably, the input X1 of the retrieval model includes: talent demand text, knowledge graph scope and set conversion task demand text; the retrieval strategy output by the retrieval model is the knowledge graph retrieval language Y * The scope of the knowledge graph is the search scope on the target knowledge graph determined based on the field to which the target talent belongs;

[0033] The retrieval model consists of an input layer, N intermediate layers, and an output layer. Each intermediate layer contains an attention layer and a feedforward neural network layer. Several intermediate layers are connected sequentially, with the input of the first intermediate layer being the output of the input layer, and the output of the last intermediate layer being the input of the output layer. Between two adjacent intermediate layers, the input of the latter intermediate layer is the output of the previous intermediate layer.

[0034] The calculation process formula of the input layer is expressed as:

[0035] PE(pos,d)=sin(d×(pos-2 (-key) ))

[0036] X2=padding(PE(pos,d),X1)

[0037] X0=L in (X2)

[0038] Among them, PE represents position encoding, pos represents the position information of the word in X1, d represents the set hyperparameter, key represents the dimension of position encoding; sin represents the sine function; X2 is the transition value, L inrepresents the linear transformation adopted by the input layer, and X0 represents the output of the input layer;

[0039] Let the input of the middle layer be recorded as X, and the attention layer in the middle layer calculates the multi-head self-attention weight matrix A of X: The calculation process formula of the attention layer in the middle layer is expressed as:

[0040]

[0041] Q=W Q X

[0042] K=W K X

[0043] V=W V X

[0044] Q, K and V represent query vector, key vector and value vector respectively. T represents transpose. K represents the dimension of K;

[0045] The calculation process formula of the feedforward neural network layer in the middle layer is expressed as follows:

[0046] F=W2×ReLU(W1×X+b1)+b2 (6)

[0047] F is the feedforward neural network matrix output by the feedforward neural network layer, W1 and W2 represent the two weight matrices of the feedforward neural network, b1 and b2 represent the two bias terms of the feedforward neural network, and ReLU represents the ReLU activation function;

[0048] A is converted into the weight matrix A after linear transformation, residual connection and layer normalization 1 ; F will be converted into matrix F after residual connection and layer normalization 1 ; A 1 and F 1 Combined into the output sequence X of the intermediate layer through residual connection 1 ; Let the output sequence X of the last intermediate layer be 1 Denoted as X * , which serves as the input sequence of the output layer;

[0049] The calculation process formula of the output layer is as follows:

[0050] Y * =σ(W O ·X * +b o )

[0051] Among them, σ is the activation function, W O is the weight matrix of the output layer, b o is the bias vector of the output layer.

[0052] The present invention proposes an industry and scientific and technological talent recommendation system based on a large model, including a memory and a processor, wherein a computer program is stored in the memory, and the processor is connected to the memory, and the processor is used to execute the computer program to implement the industry and scientific and technological talent recommendation method based on the large model.

[0053] The advantages of the present invention are:

[0054] (1) The method for constructing a multi-source expert network proposed in this invention achieves efficient and accurate expert data integration and homonym disambiguation. This invention uses a distributed data acquisition big data platform and dual-tower fusion technology to sequentially fuse expert data from different sources. This method ensures high-precision integration of expert information and effective disambiguation of homonymous experts, thereby generating accurate and complete expert representations, reducing information redundancy and confusion, and improving the quality and credibility of expert network construction.

[0055] (2) The present invention proposes a method for recommending industry and scientific and technological talents based on a large model, which realizes the construction of knowledge graphs in subdivided fields and the precise positioning of talents. The present invention innovatively constructs both scientific and technological knowledge graphs and subdivided industry knowledge graphs at the same time, providing detailed domain knowledge support for talent classification. In the present invention, a complete and highly targeted knowledge system is formed by presetting a seed knowledge graph, combining the extraction of public information in the field with manual proofreading; the precise graph construction method enables talents to be finely classified in the dimensions of science and technology and industry, greatly improving the applicability of the recommendation system and the matching degree of user needs.

[0056] (3) The present invention utilizes a large pre-trained model to achieve efficient and accurate talent classification. This invention abandons the traditional talent classification method that relies on manual labeling and instead uses a pre-trained classification model for automated classification. This not only captures newly added category labels in real time but also significantly improves classification accuracy, demonstrating strong model adaptability and self-learning capabilities, effectively responding to the rapidly changing dynamics of the industry and technology talent market.

[0057] (4) The present invention combines large models with knowledge graph technology to realize an intelligent and explainable recommendation system. The present invention cleverly combines the high-level semantic understanding of large models with the accuracy of knowledge graphs to design a system based on retrieval models and talent recommendation models. This not only improves the explainability of recommendation results and facilitates users to understand the recommendation logic, but also introduces a user feedback mechanism to continuously optimize model performance through reinforcement learning. This design effectively solves the potential bias problem of the recommendation system, improves the recommendation accuracy, and enables the model to continuously learn and upgrade in actual applications, enhancing the practicality of the system and user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1This is a flow chart of a method for constructing a multi-source expert network;

[0059] Figure 2 This is the expert data fusion flow chart;

[0060] Figure 3 A flowchart of a method for recommending industrial and scientific talents based on a large model;

[0061] Figure 4 To seed the knowledge graph for the industry in the field of artificial intelligence;

[0062] Figure 5 Build a knowledge graph for the science and technology seeds in the field of artificial intelligence;

[0063] Figure 6 This is a module diagram of an industrial technology talent recommendation system based on a large model;

[0064] Figure 7 This is the architecture diagram of the talent information collection and fusion module. DETAILED DESCRIPTION

[0065] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0066] Reference Figure 1 , a method for constructing a multi-source expert network proposed in this embodiment includes:

[0067] SA1. Obtain expert data from different sources and build corresponding initial expert networks for each data source;

[0068] The data for the initial expert network all comes from a single data source. Nodes in the initial expert network correspond one-to-one with experts in the data source. Node attributes include metadata such as the expert's field of expertise, affiliation, and research achievements. Connections between nodes are established based on collaborative behaviors, forming directed or undirected edges. Edge weights reflect the frequency, depth, or influence of collaboration. The Louvain community detection algorithm is used to identify expert groups within the network and calculate the size of each group. Degree centrality, closeness centrality, and betweenness centrality are calculated for each node.

[0069] SA2. Construct a Transformer neural network tower corresponding to each initial expert network, assign the characteristic parameters of the initial expert network to the corresponding Transformer neural network tower, and let the Transformer neural network tower encode each node of the corresponding initial expert network as an expert vector;

[0070] SA3: Perform dual-tower fusion on all Transformer neural network towers in sequence to obtain the shared space of all Transformer neural networks, embed the expert vectors of all Transformer neural networks into the shared space, and obtain the fused expert representation;

[0071] SA4. Build an expert cooperation network based on the integrated expert representations.

[0072] Specifically, expert representation includes: the expert's professional field, affiliated institution, research results, expert cooperation relationship, etc.

[0073] Reference Figure 2 , the above SA3 specifically includes the following sub-steps:

[0074] SA31. Build a tower set to count all Transformer neural network towers.

[0075] SA32. Merge any two Transformer neural network towers in the tower set, store the merged tower in the tower set, and delete the merged tower from the tower set;

[0076] The way to merge two Transformer neural network towers includes the following steps:

[0077] Traverse each expert vector in Tower A and search for a paired expert vector in Tower B. When the similarity between an expert vector B1 in Tower B and a specified expert vector A1 in Tower A is the largest and greater than the set similarity threshold, the expert vector B1 is paired with the expert vector A1.

[0078] Combined with the pairing relationship, the expert information corresponding to the expert vectors in Tower A and Tower B are fused, and the fused Transformer neural network tower is constructed based on the fused expert representation.

[0079] SA33. Determine whether there is only one Transformer neural network tower in the tower set; if not, return to step SA32; if yes, obtain the expert information space of the only Transformer neural network tower in the tower set as the shared space, and obtain the fused expert representation.

[0080] Reference Figure 3, this embodiment proposes a method for recommending industrial and scientific talents based on a large model, comprising the following steps:

[0081] S1. Build a big data platform to collect, store, integrate, and eliminate homonyms in distributed talent data, forming a big data warehouse for industry and technology talent;

[0082] In this step, the distributed parallel crawling framework of Scrapy combined with Redis is used to obtain relevant information of scientific and industrial experts from various open source platforms. Kafka is combined with HDFS and Hive to perform parallel real-time storage and offline warehousing of data, which improves the efficiency of data collection. Relevant information includes: basic information, project results, award titles, scientific and technological news, social comments and background information, etc. Basic information includes identity, age, etc. Background information includes resume, religious beliefs, etc.

[0083] Specifically, in this step, first, the data is captured and directly stored in Kafka, structured data is extracted from the captured data and stored in Hive, unstructured data is extracted from the captured data and stored in HDFS, and then the structured extraction model is used to extract the unstructured data from HDFS into Hive; thereby, a multi-source basic database of talents is constructed in Hive, and then the above-mentioned multi-source expert network construction method is used to obtain expert representations after the expert data from different sources in the multi-source basic database of talents are integrated, thereby forming a big data warehouse of industrial and scientific talents.

[0084] In this embodiment, the structured extraction model uses a BERT-like model, specifically Roberta. The structured extraction model is trained by first constructing a structured dataset using a pretrained Qwen-32B-Chat model, which includes unstructured data and corresponding structured data. The structured extraction model is then trained on the structured dataset. The structured dataset is constructed by obtaining unstructured data and inputting it into the Qwen-32B-Chat model, which then outputs the structured data corresponding to the unstructured data.

[0085] S2. Combining expert knowledge and large models to assist in building segmented industry knowledge graphs IK and technology knowledge graphs SK, and automatically classifying talents into industry and technology fields. That is, combining segmented industry knowledge graphs IK to classify talents into industry fields, and combining technology knowledge graphs SK to classify talents into technology fields.

[0086] In this step, the construction of the subdivided industry knowledge graph IK and the science and technology knowledge graph SK can adopt the domain knowledge graph construction method provided by the present invention, which includes the following steps S21-S23.

[0087] S21. First, select a field and use expert knowledge in the selected field to build a seed knowledge graph; the industry and technology seed knowledge graph in the field of artificial intelligence, the corresponding seed graph example is as follows Figure 4 , Figure 5 shown.

[0088] S22. Use the pre-trained domain classification model to extract corresponding entities and upstream and downstream associations from the public information in the domain, and add the entities to the seed knowledge graph based on the upstream and downstream associations to obtain the expanded knowledge graph; specifically, in this step, the automatic expansion and update of the seed knowledge graph is continuously realized to enrich the content of the knowledge graph.

[0089] Public information in the field includes: patent texts, papers, monographs, reports, achievements, field standards, etc.

[0090] S23. Manually proofread and revise the expanded knowledge graph using expert knowledge to obtain a domain knowledge graph. In this step, the revised content of the expanded knowledge graph can also be fed back to the domain classification model. The domain classification model is updated in real time based on the revised content, thereby improving the classification accuracy of the domain classification model. The revised content specifically includes: incorrect updated content that was deleted during manual proofreading, etc.

[0091] In specific implementation, when using the domain knowledge graph construction method to construct the subdivided industry knowledge graph IK, the selected field is the subdivided industry field, and the field classification model adopts the industry classification model; when using the domain knowledge graph construction method to construct the science and technology knowledge graph SK, the selected field is the science and technology field, and the field classification model adopts the science and technology classification model.

[0092] In this embodiment, the method for classifying talents in a selected field includes the following steps S24-S25, and the selected field can be set to an industrial field and a scientific and technological field.

[0093] S24. Obtain the domain knowledge graph corresponding to the selected domain, and construct a classification task prompt for the talent information text based on the entities and relationships in the domain knowledge graph. The entities are classification labels, and the relationships represent the hierarchy of each label. The classification task adopts the knowledge graph link classification strategy, with each link as a classification task. Each talent information text can be divided into multiple knowledge graph entity nodes, thereby improving classification accuracy and reducing classification difficulty.

[0094] S25. Use the pre-trained classification model combined with the domain knowledge graph to classify the talent information text and obtain the talent's industry category label and technology category label.

[0095] During specific implementation, a parallel classification mechanism can be used so that the classification model can combine multiple different field knowledge graphs to realize the automatic classification of talent information text in different fields. For example, talent information text, segmented industry knowledge graph IK and science and technology knowledge graph SK can be input into the classification model to realize the automatic classification of talent information text in segmented industry fields and science and technology fields.

[0096] The classification model provided in this embodiment has 14 billion parameters, including an input layer, a large number of intermediate layers and an output layer; the input layer of the classification model concatenates the input text and text position encoding and performs padding to obtain an input sequence; each intermediate layer is based on the transformer architecture, including a parallel attention layer and a feedforward neural network layer, and merges the two outputs through a layer normalization layer and a residual layer as the input of the next intermediate layer, and continuously accumulates the intermediate layers until the parameters reach the 14 billion level; the output layer uses an activation function to predict the classification label of the text.

[0097] S3. Create a scoring profile for each talent, marking the talent's technical ability score (TS) and background security score (BS);

[0098] Specifically, this implementation utilizes a pre-built technical capability evaluation model to score talent's technical capabilities. Specifically, the input indicators of the technical capability evaluation model can include: the talent's research field and interests, award list, achievement information, academic evaluation, relationship network, industry category tags, and technology category tags. The output of the technical capability evaluation model is a technical capability score (TS). In practice, the technical capability score (TS) can be set to a single score or a combination of scores for each field corresponding to the industry category tags and technology category tags.

[0099] In this implementation, a pre-built background security assessment model is used to score talent's background security. Specifically, the model's input indicators include the talent's basic information, educational background, criminal record, inappropriate social media posts, cultural background information, and real-time news. The output of the background security assessment is a background security score (BS). This background security score can provide early warning of political, religious, legal, and business risks.

[0100] In this embodiment, both the technical capability evaluation model and the background security evaluation model adopt a multi-indicator hierarchical evaluation model, thereby realizing the technical capability and background security of computing talents using open source big data.

[0101] S4. Combine the industry knowledge graph IK, technology knowledge graph SK and scoring portraits to develop and recommend talents.

[0102] In this step, the target talent's domain is first identified within the target knowledge graph, obtaining the search scope within the target knowledge graph. The talent search request is then fed into the search model, which interprets the request and generates a corresponding search strategy. This search strategy is then executed within the target knowledge graph to retrieve candidate talents. The talent recommendation model then ranks the candidate talents based on their technical capability score (TS), background security score (BS), and similarity to the talent search request. The candidate is then selected as the target talent and recommended to the user based on the ranking order. Both the search model and the talent recommendation model are pre-trained.

[0103] The talent search request includes: the search scope on the target knowledge graph; the target knowledge graph can be the segmented industry knowledge graph IK or the science and technology knowledge graph SK, or it can be the segmented industry knowledge graph IK and the science and technology knowledge graph SK.

[0104] During specific implementation, we can further obtain user feedback on talent recommendations, and further conduct reinforcement learning on the talent recommendation model based on the feedback information to improve the accuracy of the final talent recommendation.

[0105] Specifically, the similarity between the candidate talent and the talent search request in S4 is calculated as follows: vectorizing the candidate talent profile and the talent search request, and calculating the similarity between the two vectors as the similarity PS between the candidate talent and the talent search request.

[0106] The input to the talent recommendation model includes annotated candidate candidates and task requirements. Candidates are annotated with a technical capability score (TS), a background security score (BS), and a similarity score (PS). The task requirements are the ranking tasks, such as "perform a weighted ranking of these scored candidate candidates."

[0107] The input X1 of the retrieval model includes: talent demand text, knowledge graph scope and set transformation task demand text; the retrieval strategy output by the retrieval model is the knowledge graph retrieval language Y * The scope of the knowledge graph is the search scope on the target knowledge graph determined based on the field to which the target talent belongs.

[0108] The talent demand text is the user's talent demand query, such as "Help me find some industry talent for intelligent voice." The conversion task demand text reminds the search model of the conversion task it needs to complete, such as "Convert the above talent demand text into the search language within the scope of the following knowledge graph."

[0109] The retrieval model specifically consists of an input layer, N intermediate layers, and an output layer. Each intermediate layer contains an attention layer and a feedforward neural network layer. The intermediate layers are connected sequentially, with the input of the first intermediate layer being the output of the input layer, and the output of the last intermediate layer being the input of the output layer. Between two adjacent intermediate layers, the input of the latter intermediate layer is the output of the previous one.

[0110] The input layer uses formula (1) to positionally encode the words in X1:

[0111] PE(pos,d)=sin(d×(pos-2 (-key) )) (1)

[0112] In formula (1), PE represents position encoding, pos represents the position information of the word in X1, d represents a hyperparameter, key represents the dimension of position encoding; sin represents the sine function;

[0113] The input layer merges PE into X1 and obtains a sequence of uniform size X2 after padding operation. Then, a linear transformation layer is used to convert X2 into sequence X0.

[0114] Let the input of the middle layer be recorded as X, and the attention layer in the middle layer uses equations (2) to (5) to calculate the multi-head self-attention weight matrix A of X:

[0115]

[0116] Q=W Q X (3)

[0117] K=W K X (4)

[0118] V=W V X (5)

[0119] In formulas (2) to (5), Q, K, and V represent the query vector, key vector, and value vector, respectively. Q 、W K and W V Represents the weight matrices of query, key and value respectively; T represents transpose, d K represents the dimension of K;

[0120] The attention layer in the middle layer uses a linear change layer, a residual connection and a layer normalization layer to convert A into a weight matrix A 1 ;

[0121] The feedforward neural network layer in the middle layer uses formula (6) to perform nonlinear transformation on X to obtain the feedforward neural network matrix F:

[0122] F=W2×ReLU(W1×X+b1)+b2 (6)

[0123] In formula (6), W1 and W2 represent the two weight matrices of the feedforward neural network, b1 and b2 represent the two bias terms of the feedforward neural network, and ReLU represents the ReLU activation function;

[0124] The feedforward neural network layer in the middle layer uses a residual connection and a layer normalization layer to transform F into a matrix F 1 ;

[0125] A 1 and F 1 The output sequence X of the intermediate layer is combined through a residual connection 1 , X 1 As the input sequence of the next intermediate layer;

[0126] Let the output sequence X of the last intermediate layer be 1 Denoted as X * , which serves as the input sequence of the output layer;

[0127] The output layer uses formula (7) to convert X * Mapped to the output space, we get the word vector Y * And as the generated knowledge graph retrieval language, that is, the retrieval strategy:

[0128] Y * =σ(W O ·X * +b o ) (7)

[0129] In formula (7), σ is the activation function, W O is the weight matrix of the output layer, b o is the bias vector of the output layer.

[0130] Reference Figure 6 ,This implementation also proposes an industrial and technological talent recommendation system based on a ,big model, including: a talent information collection and fusion module, an industry and ,technological field classification module, a talent portrait module, and a talent retrieval and recommendation module.

[0131] Talent information collection and fusion module: Use distributed data capture technology to aggregate multi-source heterogeneous data of talents into the big data platform for storage, fusion and name disambiguation. Specifically, the talent information collection and fusion module is used to implement the above step S1. Its structure is as follows: Figure 7 shown.

[0132] Industry and Technology Classification Module: This module uses a large model to assist in the construction of segmented industry knowledge graphs and technology knowledge graphs, and categorizes talent data by industry and technology. The module pre-stores expert knowledge, domain classification models corresponding to segmented industry domains, domain classification models corresponding to technology domains, and classification models, enabling the construction of the segmented industry knowledge graph IK and technology knowledge graph SK through steps S21-S25.

[0133] Talent Profiling: This module creates a scoring profile for each talent, leveraging data from patents and papers to assess their technical capabilities in industry and technology. It also leverages open-source social media big data to assess their background and security, including their culture, language, beliefs, and political leanings. Specifically, the Talent Profiling module pre-stores technical capability and background assessment models to calculate these scores.

[0134] Talent Search and Recommendation Module: This module provides talent search and recommendation services that integrate large models and knowledge graphs, and strengthens the talent recommendation model through human feedback. Specifically, the module pre-stores a search model and a talent recommendation model. The search model generates a search strategy based on the talent search request, and the target knowledge graph executes this search strategy to retrieve candidate talents. The talent recommendation model ranks candidate talents based on their technical capability score (TS), background security score (BS), and similarity to the talent search request. Candidates are selected as target talents based on the ranking order and recommended to users.

[0135] The model mentioned in the present invention can be directly implemented using the existing model architecture.

[0136] Of course, it will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, but also encompasses the same or similar structures that can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and it is intended that all variations that fall within the meaning and range of equivalents of the claims be encompassed within the present invention. Any reference signs in the claims should not be construed as limiting the claim to which they relate.

[0137] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

[0138] The technology, shape, and structure not described in detail in the present invention are all well-known technologies.

Claims

1. A method for constructing a multi-source expert network, characterized in that: include: SA1. Obtain expert data from different sources and build corresponding initial expert networks for each data source; SA2. Construct a Transformer neural network tower corresponding to each initial expert network, assign the characteristic parameters of the initial expert network to the corresponding Transformer neural network tower, and let the Transformer neural network tower encode each node of the corresponding initial expert network as an expert vector; SA3: Perform dual-tower fusion on all Transformer neural network towers in sequence to obtain the shared space of all Transformer neural networks, embed the expert vectors of all Transformer neural networks into the shared space, and obtain the fused expert representation; Expert representation includes: expert’s professional field, affiliation, research results and expert collaboration; SA4, build an expert cooperation network based on the integrated expert representations; A method for fusing all Transformer neural network towers to obtain expert representations includes: constructing a tower set to count all Transformer neural network towers; merging any two Transformer neural network towers in the tower set, replacing the fused towers in the tower set with the merged tower, traversing the tower set until only one Transformer neural network tower remains in the tower set, then obtaining the expert information space of the only Transformer neural network tower in the tower set as a shared space, and obtaining the fused expert representation.

2. The method for constructing a multi-source expert network according to claim 1, wherein: The two Transformer neural network towers are merged as follows: traverse each expert vector in Tower A and search for a paired expert vector of Tower A's expert vector in Tower B; when the similarity between an expert vector B1 in Tower B and a specified expert vector A1 in Tower A is the largest and greater than the set similarity threshold, the expert vector B1 is paired with the expert vector A1; based on the pairing relationship, the expert information corresponding to the expert vectors in Tower A and Tower B is fused, and the fused Transformer neural network tower is constructed based on the fused expert representation.

3. A method for recommending industrial and scientific talents based on a large model using the method for constructing a multi-source expert network as described in any one of claims 1-2, characterized in that: The following steps are involved: S1. Build a big data platform to collect, store, integrate, and eliminate homonyms in distributed talent data, forming a big data warehouse for industry and technology talent; S2. Combining expert knowledge and large models to assist in building segmented industry knowledge graphs (IK) and technology knowledge graphs (SK), and automatically classifying talents into industry and technology fields; S3. Create a scoring profile for each talent, marking the talent's technical ability score (TS) and background security score (BS); S4. Combine the industry knowledge graph IK, technology knowledge graph SK and scoring portraits to develop and recommend talents.

4. The method for recommending industrial and scientific talents based on a large model according to claim 3, characterized in that: In S1, a distributed parallel crawling framework combining scrapy and redis is used to obtain relevant information of scientific and technological experts and industry experts from various open source platforms, and a multi-source basic database of talents is constructed in hive. Then, the method for constructing a multi-source expert network as described in any one of claims 1-2 is used to obtain expert representations after the expert data from different sources in the multi-source basic database of talents are integrated to form a big data warehouse of industry and scientific and technological talents.

5. The method for recommending industrial and scientific talents based on a large model according to claim 3, characterized in that: The construction of the subdivided industry knowledge graph IK and the technology knowledge graph SK in S2 includes the following steps: S21. First, select a field and use expert knowledge in the selected field to build a seed knowledge graph; S22. Use the pre-trained domain classification model to extract corresponding entities and upstream and downstream relationships from domain public information, and add the entities to the seed knowledge graph based on the upstream and downstream relationships; domain public information includes: patent texts, papers, monographs, reports, achievements, and several items of domain standards; S23. Combine expert knowledge to manually proofread and revise the expanded knowledge graph to obtain the domain knowledge graph.

6. The method for recommending industrial and scientific talents based on a large model according to claim 3, characterized in that: The method for classifying talents in a selected field is as follows: obtain the domain knowledge graph corresponding to the selected field, which includes the subdivided industry knowledge graph IK and the technology knowledge graph SK; construct a classification task prompt for the talent information text based on the entities and relationships in the domain knowledge graph, where the entities are classification labels and the relationships represent the hierarchy of each label; use a pre-trained classification model combined with the domain knowledge graph to classify the talent information text and obtain the industry category label and technology category label of the talent.

7. The method for recommending industrial and scientific talents based on a large model according to claim 3, characterized in that: In S4, the target talent's domain is first determined in the target knowledge graph, and the search scope on the target knowledge graph is obtained. The talent search request is input into the pre-trained search model, which understands the talent search request and generates a corresponding search strategy. The search strategy is then executed in the target knowledge graph to retrieve candidate talents. The pre-trained talent recommendation model ranks the candidate talents, selects the candidate as the target talent based on the ranking order from high to low, and recommends it to the user. The talent recommendation model is then trained based on customer feedback. The similarity between the candidate talent and the talent search request is calculated as follows: vectorizing the candidate talent profile and the talent search request, and calculating the similarity between the two vectors as the similarity PS between the candidate talent and the talent search request; The talent search request includes the search scope on the target knowledge graph; The input of the talent recommendation model includes: annotated candidate talents and task requirement text. The annotations of candidate talents are: technical ability score TS, background security score BS and similarity PS.

8. The method for recommending industrial and scientific talents based on a large model according to claim 7, characterized in that: The input X1 of the retrieval model includes: talent demand text, knowledge graph scope and set transformation task demand text; the retrieval strategy output by the retrieval model is the knowledge graph retrieval language Y * The scope of the knowledge graph is the search scope on the target knowledge graph determined based on the field to which the target talent belongs; The retrieval model consists of an input layer, N intermediate layers, and an output layer. Each intermediate layer contains an attention layer and a feedforward neural network layer. Several intermediate layers are connected sequentially, with the input of the first intermediate layer being the output of the input layer, and the output of the last intermediate layer being the input of the output layer. Between two adjacent intermediate layers, the input of the latter intermediate layer is the output of the previous intermediate layer. The calculation process formula of the input layer is expressed as: PE(pos,d)=sin(d×(pos-2) (-key) )) X2=padding(PE(pos,d),X1) X0=L in (X2) Among them, PE represents position encoding, pos represents the position information of the word in X1, d represents the set hyperparameter, key represents the dimension of position encoding; sin represents the sine function; X2 is the transition value, L in represents the linear transformation used in the input layer, and X0 represents the output of the input layer; Let the input of the middle layer be X, and the attention layer in the middle layer calculates the multi-head self-attention weight matrix A of X: The calculation process formula of the attention layer in the middle layer is expressed as: Q=W Q X K=W K X V=W V X Q, K and V represent query vector, key vector and value vector respectively. T represents transpose. K represents the dimension of K; The calculation process formula of the feedforward neural network layer in the middle layer is expressed as follows: F=W2×ReLU(W1×X+b1)+b2 (6) F is the feedforward neural network matrix output by the feedforward neural network layer, W1 and W2 represent the two weight matrices of the feedforward neural network, b1 and b2 represent the two bias terms of the feedforward neural network, and ReLU represents the ReLU activation function; A is converted into the weight matrix A after linear transformation, residual connection and layer normalization 1 ; F will be converted into matrix F after residual connection and layer normalization 1 ; A 1 and F 1 Combined into the output sequence X of the intermediate layer through residual connection 1 ; Let the output sequence X of the last intermediate layer be 1 Denoted as X * , which serves as the input sequence of the output layer; The calculation process formula of the output layer is as follows: Y * =σ(W O ·X * +b o ) Among them, σ is the activation function, W O is the weight matrix of the output layer, b o is the bias vector of the output layer.

9. An industry and technology talent recommendation system based on a large model, characterized by: It includes a memory and a processor, the memory stores a computer program, the processor is connected to the memory, and the processor is used to execute the computer program to implement the large model-based industrial and scientific and technological talent recommendation method as described in any one of claims 3-8.

Images