Intelligent atlas-based multivariate heterogeneous data talent evaluation system and method thereof

Abstract

The invention discloses a multivariate heterogeneous data talent evaluation system and method based on an intelligent atlas, and the system comprises a multi-source heterogeneous data collection module, a heterogeneous knowledge atlas construction module and a personalized recommendation module, and forms a heterogeneous atlas neural network through obtaining and determining a heterogeneous network data source. Based on the knowledge distillation technology, complex social network entity links and figure portraits of figure relations are established, talent data after knowledge distillation are generated, and in combination with skill talent evaluation and intelligent recommendation of the knowledge maps, knowledge complementation among multiple heterogeneous knowledge maps is achieved; according to the method, the integrity of knowledge is effectively guaranteed by utilizing a distributed data storage mechanism, the figure portrait is good in accuracy and high in credibility, meanwhile, accurate semantic understanding is achieved, the talent evaluation accuracy is high, and intelligent recommendation for different individuals is achieved.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a multi-heterogeneous data talent evaluation system based on intelligent graphs and a method thereof. Background Art

[0002] Currently, knowledge graphs and their research and development in key technologies for talent evaluation are showing a positive development trend. Knowledge graphs are regarded as key technologies for the sustainable development of the digital economy, and they show great potential in promoting the integration of knowledge graphs and traditional industries.

[0003] In the field of skilled talent evaluation, the current stage is still at the traditional single data collection and flat evaluation level. There are still deficiencies in the construction of knowledge graphs of complex social network entity links and character portraits of character relationships. Cross-domain entity and relationship references are prone to inconsistency, resulting in system confusion. The accuracy and recall rate of entity links are low, and the completeness and accuracy of talent evaluation knowledge are average. At the same time, there is a lack of systematic and holistic talent evaluation and intelligent recommendation functions. For this reason, we propose a multi-heterogeneous data talent evaluation system and method based on intelligent graphs. Summary of the invention

[0004] In view of the above-mentioned existing problems in talent evaluation in knowledge graphs, the present invention is proposed.

[0005] Therefore, one of the purposes of the present invention is to provide a multi-heterogeneous data talent evaluation system based on intelligent graph, which builds complex social network entity links and character portraits of character relationships based on knowledge distillation technology, and combines the skill talent evaluation and intelligent recommendation of knowledge graph. The demand intention tree is constructed through an end-to-end deep memory network, and demand recommendations are automatically generated according to the demand preferences in the user's knowledge graph, which has the advantages of high accuracy and good effect of talent evaluation.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] On the one hand, the present invention provides a multi-heterogeneous data talent evaluation system based on intelligent graph, comprising:

[0008] The multi-source heterogeneous data acquisition module forms a heterogeneous graph neural network by acquiring and determining the source of heterogeneous network data based on data interfaces, data sharing, web crawlers and public resources, extracts document information based on the heterogeneous graph neural network to generate structured data, and specifically constructs a heterogeneous graph consisting of sentence nodes, word nodes and skill evaluation element nodes by integrating evaluation individuals, skill types, and reward and punishment behavior elements based on a graph attention mechanism to capture the correlation between sentences, and classifies sentences to generate structured data;

[0009] A heterogeneous knowledge graph construction module, which establishes complex social network entity links and character portraits of character relationships based on knowledge distillation technology to generate talent data after knowledge distillation, and constructs corresponding heterogeneous knowledge graphs of skilled talents based on the talent data, and generates evaluation results after performing talent evaluation based on the heterogeneous knowledge graphs of skilled talents; the heterogeneous knowledge graph construction module includes a knowledge distillation unit for structured data disambiguation feature, an analysis unit for talent evaluation analysis, and a generation unit for generating evaluation results;

[0010] The personalized recommendation module combines the skill talent evaluation and intelligent recommendation of the knowledge graph, builds a demand intention tree through an end-to-end deep memory network, and automatically generates demand information for personalized recommendations based on the demand preferences in the user's knowledge graph.

[0011] As a preferred solution of the present invention, the heterogeneous data is divided into text and structured data, and into news, paper, monograph and patent data, and the corresponding mapping mechanism is matched to different types of data, and then the mapped data is integrated into a heterogeneous knowledge graph of skilled talents, and the heterogeneous data uses the conversion tool provided by Jena;

[0012] According to the different heterogeneous knowledge graphs of skilled talents formed after mapping, different types of knowledge graphs are normalized using the DBpedia program and the Sedris program to obtain multi-type heterogeneous knowledge graphs.

[0013] As a preferred solution of the present invention, wherein: the construction of a heterogeneous graph consisting of sentence nodes, word nodes and skill evaluation element nodes is to construct a skilled talent character node database, specifically based on the character relationship matching algorithm in the character graph character relationship matching model of machine learning technology to extract skilled talent character information, character relationship information and character technical information in public social media and competition media, and then use the learning training and verification process of machine learning technology to train and verify the constructed skilled talent character relationship matching model; the skilled talent character portrait database contains character skill information, skilled talent competition information and skilled talent achievements, and at the same time, the above-extracted character technical information, skilled talent competition information and skilled talent achievement information are classified and added to construct a character relationship knowledge base.

[0014] As a preferred solution of the present invention, the steps of constructing the character relationship matching model are as follows: the training process and the testing process are the offline model training and verification process; data reading and processing; extracting character information; storing the training set in the graph character relationship matching model, inputting the training set, completing the model training through the training process, and obtaining the character matching result.

[0015] As a preferred solution of the present invention, the knowledge distillation unit performs word segmentation and named entity recognition on the text by calling the LTP system in the training of the graph character relationship matching model. Specifically, a word segmentation user dictionary and a part-of-speech tagging user dictionary are added to the LTP system. The method is divided into two stages. In the first stage, the manually annotated training corpus is sorted and learned by using four types of disambiguation features to obtain a sorting model for predicting the target entity of the referential item in the corpus. After the prediction is completed, the second stage is entered. The second stage uses the loss function optimization model of knowledge distillation to optimize the accuracy of the prediction result.

[0016] The loss function optimization model of the knowledge distillation is as follows:

[0017] L=α*L soft +β*L hard

[0018]

[0019] Among them, L is the cross entropy loss function, L soft is the teacher model sofemax soft label, L hard is the hard label of the student model studentloss, α and β are the corresponding coefficients respectively;

[0020] H stands for L soft The divergence of the soft label of the teacher model sofemax, the probability distribution of the output after high temperature, K represents L hard The cross entropy of the student model studentloss hard label, G j is the loss function model parameter of knowledge distillation, a is a hyperparameter, and is the ratio of the output distribution divergence of the control true label cross entropy and the loss function model distribution of knowledge distillation;

[0021] n is the loss function model parameter of the jth knowledge distillation, is the smoothed value of the teacher model sofemax soft label, is the probability of the teacher model sofemax outputting a soft label at the temperature T stage, is the smoothed value of the hard label of the student loss of the student model, The probability of hard label output for the student model student loss.

[0022] As a preferred solution of the present invention, the four types of disambiguation features are four hypothetical relationships, specifically, the text of the entry matching any entity reference has semantic relevance with the text to be disambiguated, the entity reference and the matching entity are of the same type, there is a correlation between the entities in the text to be disambiguated, and the entities in the knowledge base are incomplete;

[0023] By using the corresponding hypothetical relationships among the four types of disambiguation features, a series of candidate entities, i.e., entries, for each reference in the disambiguation text are sorted to obtain matching entries, thereby completing entity recognition after optimizing the disambiguation features of the structured data.

[0024] As a preferred solution of the present invention, the multi-source heterogeneous data acquisition module further includes a heterogeneous data comparison unit and a data processing unit. The heterogeneous data comparison unit is used to compare the social relationship map between skilled talents and match it with the talent database; if it already exists, it is completed to form a skilled talent metadata ontology data set; if it does not exist, it is recorded in the skilled talent metadata ontology set, and finally manually sorted and analyzed;

[0025] The data processing unit uses the metadata database to fuse the metadata knowledge after information extraction according to the characteristic values ​​in the metadata database. The skill data fusion method includes: reorganizing the non-standard data from different sources through the relationship mapping method to construct a centralized and unified data set, and normalizing the centralized data set, including content structuring, normalizing data elements, cleaning, filtering to remove invalid information and data duplication, and uniformly mapping to the basic entity set.

[0026] As a preferred solution of the present invention, wherein: the personality recommendation module includes a human-computer interaction model building unit, a human-computer interaction analysis unit, and a recommendation unit;

[0027] The human-computer interaction model building unit is used to build a multi-round human-computer interaction model to support demand acquisition based on multi-round interactions;

[0028] The multi-round human-computer interaction model analysis unit is used to obtain the user's real-time needs in a multi-round interaction manner according to the needs and preference knowledge in the user knowledge graph during the process of obtaining real-time interaction needs, and then generate the demand intention tree based on the intention tree demand model after analyzing;

[0029] The recommendation unit is used to obtain and automatically generate user demand information based on the demand preference information in the user knowledge graph, and then push the user demand information to the corresponding user terminal for the user to make personalized selections.

[0030] As a preferred solution of the present invention, the human-computer interaction model construction unit uses a deep memory network to construct a language understanding layer, wherein the memory of the memory network is used to encode the text sequence, and the memory network pointer can identify the user's needs through the attention mechanism;

[0031] The multi-round human-computer interaction model analysis unit is used to perform vector initialization encoding on the demand preferences in the user knowledge graph to generate a preference vector, and uses the GRU model to generate corresponding interaction strategy information for the preference vector and the interaction state. The recommendation unit generates corresponding push information content based on the interaction strategy information.

[0032] On the one hand, the present invention provides a method for a multi-heterogeneous data talent evaluation system based on intelligent graph, comprising:

[0033] Step 1: Based on data interfaces, data sharing, web crawlers and public resources, a heterogeneous graph neural network is formed by acquiring and determining heterogeneous network data sources, and document information is extracted according to the heterogeneous graph neural network to generate structured data;

[0034] Step 2: Based on the knowledge distillation technology, complex social network entity links and character portraits of character relationships are established to generate talent data after knowledge distillation, and corresponding heterogeneous knowledge graphs of skilled talents are constructed based on the talent data. Talent evaluation is performed based on the heterogeneous knowledge graphs of skilled talents to generate evaluation results;

[0035] Step 3: Combine the skill talent evaluation and intelligent recommendation of the knowledge graph, build a demand intention tree through an end-to-end deep memory network, and automatically generate demand information for personalized recommendations based on the demand preferences in the user's knowledge graph.

[0036] The beneficial effects of the present invention are as follows: the present invention has the functions of complex social network entity linking and character portrait of character relationships; based on the knowledge distillation technology, a cross-domain knowledge fusion network is performed to realize the entity linking of the complex network, to prevent the system confusion caused by inconsistent cross-domain entity and relationship references, and to realize the complex social network entity linking based on character relationships; through the association representation model of preference targets and texts, external knowledge is introduced to approximate the distribution between domain knowledge, to realize the knowledge complementarity between multiple heterogeneous knowledge graphs, and to use the distributed data storage mechanism to effectively ensure the integrity of knowledge, so that the character portrait has good accuracy and high trustworthiness.

[0037] It also has the functions of skill talent evaluation and intelligent recommendation combined with knowledge graph; it builds a demand intention tree through an end-to-end deep memory network, automatically generates demand recommendations based on demand preferences in the user's knowledge graph, and prevents noise caused by unactivated information through the gating structure of the dynamic attention mechanism, thereby achieving precise semantic understanding, high accuracy in talent evaluation, and realizing intelligent recommendations for different individuals. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative labor. Among them:

[0039] Figure 1 It is a schematic diagram of the modular structure of the system of the present invention;

[0040] Figure 2 A schematic diagram of the process of performing knowledge distillation in the system of the present invention;

[0041] Figure 3 A schematic diagram of a multi-round human-computer interaction processing scenario of a personalized recommendation module in the system of the present invention;

[0042] Figure 4 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solution and advantages of the embodiment of the present invention clearer, the technical solution of the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings of the embodiment of the present invention. Obviously, the described embodiment is a part of the embodiment of the present invention, not all of the embodiments. Based on the described embodiment of the present invention, all other embodiments obtained by ordinary technicians in this field belong to the scope of protection of the present invention.

[0044] In order to fully and comprehensively evaluate the true level of skilled talents and improve the accuracy and efficiency of talent screening, this embodiment, based on the talent intelligent evaluation system, studies the technologies of intelligent collection of multi-source heterogeneous data, complex social network entity links and character portraits based on character relationships, and skilled talent evaluation and intelligent recommendation combined with knowledge graphs, to solve technical problems such as data collection, building character portraits, and artificial intelligence. It not only ensures the fairness of the skilled talent evaluation process and the authenticity of the results, but also conducts reliable analysis of subsequent talent predictions and career planning, thereby improving the efficiency and quality of the entire skilled talent evaluation process.

[0045] Reference Figure 1-Figure 4 , is an embodiment of the present invention, which provides an aspect that the present invention provides a multi-heterogeneous data talent evaluation system based on intelligent graph, including:

[0046] The multi-source heterogeneous data acquisition module forms a heterogeneous graph neural network by acquiring and determining the source of heterogeneous network data based on data interfaces, data sharing, web crawlers and public resources, extracts document information based on the heterogeneous graph neural network to generate structured data, and specifically constructs a heterogeneous graph consisting of sentence nodes, word nodes and skill evaluation element nodes by integrating evaluation individuals, skill types, and reward and punishment behavior elements based on a graph attention mechanism to capture the correlation between sentences, and classifies sentences to generate structured data;

[0047] The heterogeneous knowledge graph construction module generates talent data after knowledge distillation by establishing complex social network entity links and character portraits of character relationships based on knowledge distillation technology, and constructs corresponding heterogeneous knowledge graphs of skilled talents based on talent data, and generates evaluation results after talent evaluation based on the heterogeneous knowledge graphs of skilled talents; the heterogeneous knowledge graph construction module includes a knowledge distillation unit for structured data disambiguation feature, an analysis unit for talent evaluation analysis, and a generation unit for generating evaluation results;

[0048] The personalized recommendation module combines the skill talent evaluation and intelligent recommendation of the knowledge graph, builds a demand intention tree through an end-to-end deep memory network, and automatically generates demand information for personalized recommendations based on the demand preferences in the user knowledge graph. After extracting and annotating the corresponding dialogue samples and user knowledge graph, it is necessary to build a multi-round human-computer interaction model to support demand acquisition based on multi-round interactions. The multi-round human-computer interaction model needs to obtain the user's real-time needs in a multi-round interactive manner according to the demand and preference knowledge in the user knowledge graph during the process of real-time interactive demand acquisition. Therefore, building a multi-round human-computer interaction model based on the user knowledge graph is the core issue of demand intention tree construction. After completing the model training, the model can interact and identify needs in real time, and build a demand intention tree based on the intention tree demand model.

[0049] Specifically, in this embodiment, heterogeneous data is divided into text and structured data, and into news, paper, monograph and patent data. Different types of data are matched with corresponding mapping mechanisms, and then the mapped data is integrated into a heterogeneous knowledge graph of skilled talents. The heterogeneous data uses Jena's built-in conversion tool;

[0050] According to the different heterogeneous knowledge graphs of skilled talents formed after mapping, different types of knowledge graphs are normalized using the DBpedia program and the Sedris program to obtain multi-type heterogeneous knowledge graphs.

[0051] Specifically, this embodiment constructs a heterogeneous graph consisting of sentence nodes, word nodes and skill evaluation element nodes. In order to build a skilled talent character node database, the character relationship matching algorithm in the character graph character relationship matching model based on machine learning technology is used to extract skilled talent character information, character relationship information and character technical information in public social media and competition media, and then the constructed skilled talent character relationship matching model is trained and verified by the learning, training and verification process of machine learning technology; the skilled talent character portrait database contains character skill information, skilled talent competition information and skilled talent achievements, and at the same time, the above-extracted character technical information, skilled talent competition information and skilled talent achievement information are classified and added to build a character relationship knowledge base.

[0052] Specifically, in this embodiment, the steps of constructing the character relationship matching model are as follows: training process and testing process and offline model training and verification process; data reading and processing; extracting character information; storing the training set in the graph character relationship matching model, inputting the training set, completing the model training through the training process, and obtaining the character matching result.

[0053] It should be emphasized in this embodiment that the knowledge distillation unit performs word segmentation and named entity recognition on the text by calling the LTP (Language Technology Platform) system in the training of the graph character relationship matching model. Specifically, a word segmentation user dictionary and a part-of-speech tagging user dictionary are added to the LTP system. The method is divided into two stages. In the first stage, the manually annotated training corpus is sorted and learned by using four types of disambiguation features to obtain a sorting model for predicting the target entity of the referential item in the corpus. After the prediction is completed, the second stage is entered. In the second stage, the loss function optimization model of knowledge distillation is used to optimize the accuracy of the prediction result.

[0054] The loss function optimization model for knowledge distillation is as follows:

[0055] L=α*L soft +β*L hard

[0056]

[0057] Among them, L is the cross entropy loss function, L soft is the teacher model sofemax soft label, L hard is the hard label of the student model studentloss, α and β are the corresponding coefficients respectively;

[0058] H stands for L soft The divergence of the soft label of the teacher model sofemax, the probability distribution of the output after high temperature, K represents L hardThe cross entropy of the student loss hard label of the student model, G j is the loss function model parameter of knowledge distillation, a is a hyperparameter, and is the ratio of the output distribution divergence of the control true label cross entropy and the loss function model distribution of knowledge distillation;

[0059] n is the loss function model parameter of the jth knowledge distillation, is the smoothed value of the teacher model sofemax soft label, is the probability of the teacher model sofemax outputting a soft label at the temperature T stage, is the smoothed value of the hard label of the student loss of the student model, The probability of outputting hard labels for the student loss of the student model. From the perspective of the relevance of entity extraction and entity linking tasks, the knowledge distillation framework is used to effectively improve the accuracy of character portraits in the field of vocational skill talent evaluation.

[0060] Specifically, in this embodiment, the four types of disambiguation features are four hypothetical relationships, specifically, the text of the entry matching any entity reference has semantic relevance with the text to be disambiguated, the entity reference and the matching entity are of the same type, there is a correlation between the entities in the text to be disambiguated, and the entities in the knowledge base are incomplete;

[0061] By using the corresponding hypothetical relationships among the four types of disambiguation features, a series of candidate entities, i.e., entries, for each reference in the disambiguation text are sorted to obtain matching entries, thus completing entity recognition after optimizing the disambiguation features of structured data.

[0062] Based on the above, it can be seen that this embodiment has the functions of complex social network entity linking and character portrait of character relationships; based on the knowledge distillation technology, cross-domain knowledge fusion network is carried out to realize the entity linking of complex networks, to prevent the system confusion caused by inconsistent cross-domain entity and relationship references, and to realize the complex social network entity linking based on character relationships; through the association representation model of preference targets and texts, external knowledge is introduced to approximate the distribution between domain knowledge, to realize knowledge complementarity between multiple heterogeneous knowledge graphs, and to use the distributed data storage mechanism to effectively ensure the integrity of knowledge, and the character portraits are accurate and trustworthy.

[0063] In order to solve the problems in the process of collecting multi-source heterogeneous data, the collection sources are diverse, such as using web crawlers to collect network data, using APIs to obtain third-party data, using sensors and Internet of Things devices to collect real-time data, etc.; the data forms are diverse, such as structured data, semi-structured data, unstructured data, etc., to solve the current common data collection technology problems such as low data collection efficiency, high error rate, difficulty in processing big data, difficulty in collecting complex data, and poor scalability; specifically, in this embodiment, the multi-source heterogeneous data collection module also includes a heterogeneous data comparison unit and a data processing unit. The heterogeneous data comparison unit is used to compare the social relationship map between skilled talents and match it with the talent database; if it already exists, it is completed to form a skilled talent metadata ontology data set. If it does not exist, it is recorded in the skilled talent metadata ontology set, and finally manually sorted and analyzed;

[0064] The data processing unit uses the metadata database to fuse the metadata knowledge after information extraction according to the characteristic values ​​in the metadata database. The skill data fusion method includes: reorganizing the non-standard data from different sources by the relational mapping method to build a centralized and unified data set, and normalizing the centralized data set, including content structuring, normalizing data elements, cleaning, filtering to remove invalid information and data duplication, and uniformly mapping to the basic entity set.

[0065] Specifically, in this embodiment, the personalized recommendation module includes a human-computer interaction model construction unit, a human-computer interaction analysis unit, and a recommendation unit; the human-computer interaction model construction unit is used to construct a multi-round human-computer interaction model to support demand acquisition based on multi-round interactions; the multi-round human-computer interaction model analysis unit is used to obtain the user's real-time needs in a multi-round interaction manner according to the needs and preference knowledge in the user knowledge graph during the process of real-time interaction demand acquisition, and then generate a demand intention tree based on the intention tree demand model; the recommendation unit is used to obtain and automatically generate user demand information based on the demand preference information in the user knowledge graph, and then push the user demand information to the corresponding user terminal for the user to make personalized selections.

[0066] In addition, the human-computer interaction model construction unit uses a deep memory network to construct a language understanding layer, in which the memory network memory is used to encode text sequences, and the memory network pointer can identify user needs through the attention mechanism; the multi-round human-computer interaction model analysis unit is used to vector initialize the encoding of the demand preferences in the user knowledge graph to generate a preference vector, and uses the GRU model to generate corresponding interaction strategy information for the preference vector and the interaction status. The recommendation unit generates corresponding push information content based on the interaction strategy information.

[0067] After this embodiment is implemented through software, it can support a knowledge graph triple number of no less than 100,000; the number of AI users that can be connected at the same time is no less than 200; the accuracy and recall rate of entity links are greater than 85% and 90% respectively; and the overall response speed of the system is less than 500ms.

[0068] Based on the above, this embodiment has the function of skill talent evaluation and intelligent recommendation combined with knowledge graph; constructs a demand intention tree through an end-to-end deep memory network, automatically generates demand recommendations based on demand preferences in the user's knowledge graph, and prevents noise caused by unactivated information through the gated structure of the dynamic attention mechanism, thereby achieving precise semantic understanding, high accuracy in talent evaluation, and realizing intelligent recommendations for different individuals; and finally realizing interactive feedback with users to confirm demands, thereby realizing the evaluation and intelligent recommendation of knowledge graphs for professional skill talents.

[0069] Understandably, the personalized recommendation module uses a deep memory network to build a language understanding layer, in which the memory of the memory network is used to encode text sequences, and the memory network pointer can identify the user's needs through the attention mechanism; the demand preferences in the user's knowledge graph are vector-initialized and encoded to realize the function of the potential demand reasoning layer. The interaction strategy is controlled through the dialogue state management layer. The GRU (Gate Recurrent Unit) model is used to generate corresponding interaction strategies based on preference vectors and interaction states. Finally, the interaction generation layer generates corresponding reply content based on the interaction strategy. The multi-round dialogue and demand tree construction model is studied in an end-to-end form, and trained by back propagation learning with a supervised learning method. After training, the model is used for real-time online interaction to realize a multi-round interaction process based on the user's knowledge graph, thereby completing the construction of a demand intention tree based on multi-round interactions, and realizing the core technology of skill talent evaluation and intelligent recommendation.

[0070] In addition, this embodiment also provides a method for a multi-heterogeneous data talent evaluation system based on an intelligent graph, which is as follows:

[0071] Step 1: Based on data interfaces, data sharing, web crawlers and public resources, a heterogeneous graph neural network is formed by acquiring and determining the source of heterogeneous network data. Document information is extracted based on the heterogeneous graph neural network to generate structured data.

[0072] Step 2: Based on the knowledge distillation technology, complex social network entity links and character portraits of character relationships are established to generate talent data after knowledge distillation, and the corresponding heterogeneous knowledge graph of skilled talents is constructed based on the talent data. Talent evaluation is performed based on the heterogeneous knowledge graph of skilled talents to generate evaluation results;

[0073] Step 3: Combine the skill talent evaluation and intelligent recommendation of the knowledge graph, build a demand intention tree through an end-to-end deep memory network, and automatically generate demand information for personalized recommendations based on the demand preferences in the user's knowledge graph.

[0074] In summary, the present invention studies the technologies of intelligent collection of multi-source heterogeneous data, complex social network entity links and character portraits based on character relationships, skill talent evaluation and intelligent recommendation combined with knowledge graphs. The knowledge graph-based system can analyze users' potential needs and provide strong support for personalized recommendations, intelligent question and answer, and talent management.

[0075] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function according to the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium.

[0076] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine different embodiments or examples described in this specification and the features of different embodiments or examples, unless they are contradictory.

[0077] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include at least one of the features. In the description of this application, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0078] Any process or method description in the flow chart or otherwise described herein can be understood to represent a module, fragment or portion of a code including one or more executable instructions for implementing the steps of a specific logical function or process. And the scope of the preferred embodiment of the present application includes other implementations, in which the functions may not be performed in the order shown or discussed, including in a substantially simultaneous manner or in a reverse order according to the functions involved.

[0079] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, which can be embodied in any computer-readable medium for use by an instruction execution system, apparatus or device (such as a computer-based system, a system including a processor or other system that can fetch instructions from an instruction execution system, apparatus or device and execute instructions), or used in combination with these instruction execution systems, apparatuses or devices.

[0080] It should be understood that the various parts of the present application can be implemented with hardware, software, firmware or a combination thereof. In the above embodiments, multiple steps or methods can be implemented with software or firmware stored in a memory and executed by a suitable instruction execution system. All or part of the steps of the above embodiment method can be completed by instructing the relevant hardware through a program, which can be stored in a computer-readable storage medium, and when the program is executed, it includes one of the steps of the method embodiment or a combination thereof.

[0081] In addition, each functional unit in each embodiment of the present application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into one module. The above-mentioned integrated module can be implemented in the form of hardware or in the form of a software functional module. If the above-mentioned integrated module is implemented in the form of a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. The storage medium can be a read-only memory, a disk or an optical disk, etc.

[0082] The above are only specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of various changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A multi-heterogeneous data talent evaluation system based on intelligent graph, characterized in that: include: The multi-source heterogeneous data acquisition module forms a heterogeneous graph neural network by acquiring and determining the source of heterogeneous network data based on data interfaces, data sharing, web crawlers and public resources, extracts document information based on the heterogeneous graph neural network to generate structured data, and specifically constructs a heterogeneous graph consisting of sentence nodes, word nodes and skill evaluation element nodes by integrating evaluation individuals, skill types, and reward and punishment behavior elements based on a graph attention mechanism to capture the correlation between sentences, and classifies sentences to generate structured data; The heterogeneous knowledge graph construction module builds complex social network entity links and character portraits of character relationships based on knowledge distillation technology to generate talent data after knowledge distillation, and builds corresponding heterogeneous knowledge graphs of skilled talents based on the talent data, and generates evaluation results after talent evaluation based on the heterogeneous knowledge graphs of skilled talents; The heterogeneous knowledge graph construction module includes a feature knowledge distillation unit for structured data disambiguation, an analysis unit for talent evaluation analysis, and a generation unit for generating evaluation results; The personalized recommendation module combines the skill talent evaluation and intelligent recommendation of the knowledge graph, builds a demand intention tree through an end-to-end deep memory network, and automatically generates demand information for personalized recommendations based on the demand preferences in the user's knowledge graph.

2. A multi-heterogeneous data talent evaluation system based on intelligent graph as claimed in claim 1, characterized in that: The heterogeneous data is divided into text and structured data, and into news, papers, monographs and patents. The corresponding mapping mechanisms are matched to different types of data, and then the mapped data is integrated into a heterogeneous knowledge graph of skilled talents. The heterogeneous data uses the conversion tool provided by Jena. According to the different heterogeneous knowledge graphs of skilled talents formed after mapping, different types of knowledge graphs are normalized using the DBpedia program and the Sedris program to obtain multi-type heterogeneous knowledge graphs.

3. The multi-heterogeneous data talent evaluation system based on intelligent graph according to claim 1, characterized in that: The method constructs a heterogeneous graph consisting of sentence nodes, word nodes and skill evaluation element nodes to construct a skilled talent character node database. Specifically, the character relationship matching algorithm in the character graph character relationship matching model based on machine learning technology extracts skilled talent character information, character relationship information and character technical information in public social media and competition media, and then uses the learning training and verification process of machine learning technology to train and verify the constructed skilled talent character relationship matching model; the skilled talent character portrait database contains character skill information, skilled talent competition information and skilled talent achievements. At the same time, the above-extracted character technical information, skilled talent competition information and skilled talent achievement information are classified and added to construct a character relationship knowledge base.

4. The multi-heterogeneous data talent evaluation system based on intelligent graph according to claim 1, characterized in that: The steps of constructing the character relationship matching model are as follows: the training process and the testing process are combined with the offline model training and verification process; data reading and processing; extracting character information; storing the training set in the graph character relationship matching model, inputting the training set, completing the model training through the training process, and obtaining the character matching result.

5. The multi-heterogeneous data talent evaluation system based on intelligent graph according to claim 4, characterized in that: The knowledge distillation unit performs word segmentation and named entity recognition on the text by calling the LTP system in the training of the graph character relationship matching model. Specifically, a word segmentation user dictionary and a part-of-speech tagging user dictionary are added to the LTP system. The method is divided into two stages. In the first stage, the manually annotated training corpus is sorted and learned by using four types of disambiguation features to obtain a sorting model for predicting the target entity of the referential item in the corpus. After the prediction is completed, the second stage is entered. The second stage uses the loss function optimization model of knowledge distillation to optimize the accuracy of the prediction result. The loss function optimization model of the knowledge distillation is as follows: L=α*L soft +β*L hard Among them, L is the cross entropy loss function, L soft is the teacher model sofemax soft label, L hard is the hard label of the student model studentloss, α and β are the corresponding coefficients respectively; H is for L soft The divergence of the soft label of the teacher model sofemax, the probability distribution of the output after high temperature, K represents L hard The cross entropy of the student loss hard label of the student model, G j is the loss function model parameter of knowledge distillation, a is a hyperparameter, and is the ratio of the output distribution divergence of the control true label cross entropy and the loss function model distribution of knowledge distillation; n is the loss function model parameter of the jth knowledge distillation, is the smoothed value of the teacher model sofemax soft label, is the probability of the teacher model sofemax outputting a soft label at the temperature T stage, is the smoothed value of the hard label of the student loss of the student model, The probability of hard label output for the student model student loss.

6. A multi-heterogeneous data talent evaluation system based on intelligent graph as claimed in claim 5, characterized in that: The four types of disambiguation features are four hypothetical relationships, specifically, the text of the entry matching any entity reference has semantic relevance with the text to be disambiguated, the entity reference and the matching entity are of the same type, there is a correlation between the entities in the text to be disambiguated, and the entities in the knowledge base are incomplete; By using the corresponding hypothetical relationships among the four types of disambiguation features, a series of candidate entities, i.e., entries, for each reference in the disambiguation text are sorted to obtain matching entries, thereby completing entity recognition after optimizing the disambiguation features of the structured data.

7. The multi-heterogeneous data talent evaluation system based on intelligent graph according to claim 1, characterized in that: The multi-source heterogeneous data acquisition module also includes a heterogeneous data comparison unit and a data processing unit. The heterogeneous data comparison unit is used to compare the social relationship map between skilled talents and match it with the talent database; if it already exists, it is completed to form a skilled talent metadata ontology data set; if it does not exist, it is recorded in the skilled talent metadata ontology set, and finally manually sorted and analyzed; The data processing unit uses the metadata database to fuse the metadata knowledge after information extraction according to the characteristic values ​​in the metadata database. The skill data fusion method includes: reorganizing the non-standard data from different sources through the relationship mapping method to construct a centralized and unified data set, and normalizing the centralized data set, including content structuring, normalizing data elements, cleaning, filtering to remove invalid information and data duplication, and uniformly mapping to the basic entity set.

8. The multi-heterogeneous data talent evaluation system based on intelligent graph according to claim 1, characterized in that: The personality recommendation module includes a human-computer interaction model building unit, a human-computer interaction analysis unit, and a recommendation unit; The human-computer interaction model building unit is used to build a multi-round human-computer interaction model to support demand acquisition based on multi-round interactions; The multi-round human-computer interaction model analysis unit is used to obtain the user's real-time needs in a multi-round interaction manner according to the needs and preference knowledge in the user knowledge graph during the process of obtaining real-time interaction needs, and then generate the demand intention tree based on the intention tree demand model after analyzing; The recommendation unit is used to obtain and automatically generate user demand information based on the demand preference information in the user knowledge graph, and then push the user demand information to the corresponding user terminal for the user to make personalized selections.

9. The multi-heterogeneous data talent evaluation system based on intelligent graph according to claim 8, characterized in that: The human-computer interaction model construction unit uses a deep memory network to construct a language understanding layer, wherein the memory of the memory network is used to encode text sequences, and the memory network pointer can identify user needs through an attention mechanism; The multi-round human-computer interaction model analysis unit is used to perform vector initialization encoding on the demand preferences in the user knowledge graph to generate a preference vector, and uses the GRU model to generate corresponding interaction strategy information for the preference vector and the interaction state. The recommendation unit generates corresponding push information content based on the interaction strategy information.

10. A method for a multi-dimensional heterogeneous data talent evaluation system based on intelligent graphs as claimed in claim 1, characterized in that: include: Step 1: Based on data interfaces, data sharing, web crawlers and public resources, a heterogeneous graph neural network is formed by acquiring and determining heterogeneous network data sources, and document information is extracted according to the heterogeneous graph neural network to generate structured data; Step 2: Based on the knowledge distillation technology, complex social network entity links and character portraits of character relationships are established to generate talent data after knowledge distillation, and corresponding heterogeneous knowledge graphs of skilled talents are constructed based on the talent data. Talent evaluation is performed based on the heterogeneous knowledge graphs of skilled talents to generate evaluation results; Step 3: Combine the skill talent evaluation and intelligent recommendation of the knowledge graph, build a demand intention tree through an end-to-end deep memory network, and automatically generate demand information for personalized recommendations based on the demand preferences in the user's knowledge graph.

Images