A user portrait construction method combined with knowledge graph

Abstract

The present invention discloses a method for constructing user portraits in combination with a knowledge graph, specifically relating to the field of big data, including a construction architecture consisting of a knowledge graph construction module, a label extraction module, a talent portrait module, and a feedback module: wherein the knowledge graph construction module includes: named entity recognition and relationship extraction; the label extraction module includes label selection, data cleaning, and label storage; and a user feedback module: users will provide feedback on the labels of user portraits. The present invention uses the knowledge graph as a supplement when constructing a labeling system; uses an end-to-end method to construct the knowledge graph, thereby improving the efficiency of knowledge graph construction and reducing complexity; the independence of the sample library and the database can improve the credibility of the labeling system; the constructed labeling system adds the supplement and verification of external information, which is more accurate and reliable; the end-to-end knowledge graph construction method greatly improves the efficiency of knowledge graph construction and reduces complexity.

Description

Technical Field

[0001] The present invention relates to the field of big data technology, and more specifically, to a method for constructing a user portrait in combination with a knowledge graph. Background Art

[0002] With the popularization of artificial intelligence, computing power has been greatly improved, and more and more algorithms and application scenarios have been discovered, which has also spawned many related downstream applications. User portraits are a very typical one of them.

[0003] Due to their high complexity and high dependence on scenarios, current methods for constructing talent profiles are mostly still based on statistics, or simple linear and nonlinear models, and have not yet formed a complete and mature system.

[0004] The common construction methods currently have the following disadvantages:

[0005] 1. The construction method is single, mainly manual operation and simple statistics.

[0006] 2. The data source acquisition method is single, making it difficult to accurately and comprehensively construct labels. Summary of the Invention

[0007] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a method for constructing a user portrait in combination with a knowledge graph. The technical problem to be solved by the present invention is: how to make the construction of user portraits flexibly used.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a method for constructing a user profile in combination with a knowledge graph, comprising a construction architecture consisting of a knowledge graph construction module, a label extraction module, a talent profile module, and a feedback module:

[0009] Among them, the knowledge graph construction module includes: named entity recognition and relationship extraction;

[0010] Import a pre-trained end-to-end neural network model from the model library, extract relevant text from the sample library through the sample interface, input the model, and output triples of all entities and related relationships in the text;

[0011] The data obtained by deep learning through the feedback module is added to the original database for joint training to further improve the accuracy of the model. The pre-trained end-to-end neural network is imported from the model library to directly predict entity relationships. The external knowledge of entity relationships obtained in this way is supplemented in the training of word vectors.

[0012] The label extraction module includes label selection, data cleaning and label storage;

[0013] Tag selection: For users in different scenarios, different strategies for tag selection should be implemented based on user needs and the application direction of user profiles.

[0014] Data cleaning: The target label may directly correspond to the content of a field in the database, or it may be content within the content. In this case, the content corresponding to the field needs to be cleaned, and if necessary, an algorithm needs to be selected to operate on the relevant content. For example, if the user selects the label "Sentiment Analysis" for "Work Report", the "Work Report" field needs to be extracted from the relevant applications in the database, and then the corresponding content needs to be judged by the sentiment analysis algorithm to determine its emotional tendency and finally obtain the label.

[0015] Tag storage: The obtained tags are stored in the HBase database according to the user ID. Each user ID will correspond to multiple tags. The row key is the ID and the corresponding content is a dict with the key being the tag name and the value being the tag value.

[0016] The talent portrait module is used to organize the specific storage fields of user portraits;

[0017] User feedback module: Users will provide feedback on user profile labels. If the labels are incorrect, they will be fed back to the sample library or database according to the direction of label generation as a correction to the data and algorithm.

[0018] The specific construction steps between each module are as follows:

[0019] S1: Retrieve the talent resume database and clean and merge the data;

[0020] S2: Extract talent labels from structured fields, such as industry, field, professional skills, work experience, and educational experience, as labels for the talent;

[0021] S3: Build an end-to-end industry knowledge graph, starting with professional skills knowledge in various fields, and construct a domain-industry-skill graph network to serve as content supplement when building the label system;

[0022] S4: Retrieve: Talent resume database. For unstructured fields, we use NLP algorithms and knowledge graph relationship extraction algorithms to extract labels of skills, industry experience, and years of experience.

[0023] S5: Integrate structured and unstructured tags, perform semantic alignment for secondary verification, and use them as talent labels;

[0024] S6: Extract keywords from new project cases and recommend talents based on talent tags.

[0025] In a preferred embodiment, the knowledge graph construction module includes an end-to-end model that uses remote supervision and pre-training models to directly obtain entities and relationships from input text to construct a knowledge graph; the text here includes company-related text information, numerical information, announcements, organizational structure and news data; but does not include standard data in the company database.

[0026] In a preferred embodiment, a trained model is imported into the deep learning prediction part. When text is input, it can directly enter the module to output the relationship between entities. The output result needs to pass through LSTM decoding and Softmax (normalized exponential function) layer to obtain the final entity relationship.

[0027] The pseudo code of the pre-trained end-to-end model is as follows:

[0028] gettexts\\Get company-related articles in the sample library

[0029] wordembedding\\The input text is passed through a shared word embedding (word embedding vector)

[0030] encodinginputbyBiLSTM\\Connect the bidirectional LSTM layer to encode the input

[0031] NERbyLSTM\\Named Entity Recognition with LSTM

[0032] relationsclassifierbyCNN\\Use CNN (convolutional neural network) for relation classification.

[0033] In a preferred embodiment, the knowledge graph construction module is achieved by sharing the underlying model parameters. During training, both tasks will update the shared parameters through the backpropagation algorithm to realize the dependency between the two subtasks.

[0034] In a preferred embodiment, when selecting the tag, after selecting the tag, it is necessary to determine the field to be selected in the database and extract the required tag from the field.

[0035] In a preferred embodiment, when the data is cleaned, an algorithm is selected to operate on the relevant content; if the user selects the label "Sentiment Analysis" of "Work Report", it is first necessary to extract the "Work Report" field from the relevant applications in the database, and then perform the sentiment analysis algorithm on the corresponding content to confirm its emotional tendency, and finally obtain the label.

[0036] In a preferred embodiment, the talent portrait module receives two parts of information;

[0037] ① Basic tag fields and tag values ​​obtained based on the database;

[0038] ② Various attributes expanded based on the knowledge graph and user entities;

[0039] Merge the labels obtained in ① and ② and store them together in the HBase database; the labels obtained from the knowledge graph need to be proofread, supplemented and modified for the labels stored in the database.

[0040] Technical effects and advantages of the present invention:

[0041] The present invention uses the knowledge graph as a supplement when constructing the labeling system; uses an end-to-end method to construct the knowledge graph, which improves the construction efficiency of the knowledge graph and reduces the complexity; the independence of the sample library and the database can improve the credibility of the labeling system; the constructed labeling system adds the supplement and verification of external information, which is more accurate and reliable; the end-to-end knowledge graph construction method greatly improves the construction efficiency of the knowledge graph and reduces the complexity. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is the overall architecture diagram of the present invention. DETAILED DESCRIPTION

[0043] 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.

[0044] The present invention provides a method for constructing a user profile in combination with a knowledge graph, including a construction architecture consisting of a knowledge graph construction module, a label extraction module, a talent profile module, and a feedback module:

[0045] Among them, the knowledge graph construction module includes: named entity recognition and relationship extraction;

[0046] Import a pre-trained end-to-end neural network model from the model library, extract relevant text from the sample library through the sample interface, input the model, and output triples of all entities and related relationships in the text;

[0047] The knowledge graph construction module includes an end-to-end model that uses distant supervision and pre-trained models to directly extract entities and relationships from input text to build a knowledge graph. This text includes company-related text information, numerical information, announcements, organizational structure, and news data, but does not include standard data in the company database.

[0048] Import the trained model into the deep learning prediction part. When text is input, it can directly enter this module to output the relationship between entities. The output results need to pass through LSTM decoding and Softmax (normalized exponential function) layer to obtain the final entity relationship.

[0049] The pseudo code of the pre-trained end-to-end model is as follows:

[0050] gettexts\\Get company-related articles in the sample library

[0051] wordembedding\\The input text is passed through a shared word embedding (word embedding vector)

[0052] encodinginputbyBiLSTM\\Connect the bidirectional LSTM layer to encode the input

[0053] NERbyLSTM\\Named Entity Recognition with LSTM

[0054] relationsclassifierbyCNN\\Use CNN (convolutional neural network) for relation classification;

[0055] The data obtained by deep learning through the feedback module is added to the original database for joint training to further improve the accuracy of the model. The pre-trained end-to-end neural network is imported from the model library to directly predict entity relationships. The external knowledge of entity relationships obtained in this way is supplemented in the training of word vectors.

[0056] The knowledge graph construction module is built by sharing the underlying model parameters. During training, both tasks will use the backpropagation algorithm to update the shared parameters to achieve the dependency between the two subtasks.

[0057] The label extraction module includes label selection, data cleaning and label storage;

[0058] Tag selection: For users in different scenarios, different strategies for tag selection are required based on user needs and the application direction of user profiles. After selecting tags, it is necessary to determine the fields to be selected in the database and extract the required tags from the fields.

[0059] Data cleaning: The target label may directly correspond to the content of a field in the database, or it may be content within the content. In this case, the content corresponding to the field needs to be cleaned, and if necessary, an algorithm needs to be selected to operate on the relevant content. For example, if the user selects the label "Sentiment Analysis" for "Work Report", the "Work Report" field needs to be extracted from the relevant applications in the database, and then the corresponding content needs to be judged by the sentiment analysis algorithm to determine its emotional tendency and finally obtain the label.

[0060] Tag storage: The obtained tags are stored in the HBase database according to the user ID. Each user ID will correspond to multiple tags. The row key is the ID and the corresponding content is a dict with the key being the tag name and the value being the tag value.

[0061] The talent portrait module is used to organize the specific storage fields of user portraits;

[0062] The talent profile module receives two parts of information;

[0063] ① Basic tag fields and tag values ​​obtained based on the database;

[0064] ② Various attributes expanded based on the knowledge graph and user entities;

[0065] Merge the labels obtained in ① and ② and store them together in the HBase database; the labels obtained from the knowledge graph need to be proofread, supplemented, and modified with the labels stored in the database;

[0066] User feedback module: Users will provide feedback on user profile labels. If the labels are incorrect, they will be fed back to the sample library or database according to the direction of label generation as a correction to the data and algorithm.

[0067] The specific construction steps between each module are as follows:

[0068] S1: Retrieve the talent resume database and clean and merge the data;

[0069] S2: Extract talent labels from structured fields, such as industry, field, professional skills, work experience, and educational experience, as labels for the talent;

[0070] S3: Build an end-to-end industry knowledge graph, starting with professional skills knowledge in various fields, and construct a domain-industry-skill graph network to serve as content supplement when building the label system;

[0071] S4: Retrieve: For unstructured fields such as work experience and project experience, we use NLP algorithms such as NER, topic models, and keyword extraction, as well as relationship extraction algorithms from knowledge graphs to extract labels for skills, industry experience, and years of experience.

[0072] S5: Integrate structured and unstructured tags, perform semantic alignment for secondary verification, and use them as talent labels;

[0073] S6: Extract keywords from new project cases and recommend talents based on talent tags.

[0074] The labeling system constructed by the present invention adds supplementation and verification of external information, which is more accurate and reliable; the end-to-end knowledge graph construction method greatly improves the construction efficiency of the knowledge graph and reduces the complexity.

[0075] Finally, a few points should be explained: First, in the description of this application, it should be noted that, unless otherwise specified or limited, the terms "mounted," "connected," and "connected" should be understood in a broad sense, and may refer to mechanical or electrical connections, internal communication between two components, or direct connection. "Up," "down," "left," and "right" are only used to indicate relative positional relationships. When the absolute positions of the objects being described change, the relative positional relationships may also change.

[0076] Secondly: The drawings of the embodiments disclosed in the present invention only involve structures related to the embodiments disclosed in the present invention. Other structures may refer to conventional designs. The same embodiment and different embodiments of the present invention may be combined with each other without conflict.

[0077] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A user portrait construction method combined with knowledge graph, including a construction architecture consisting of a knowledge graph construction module, a label extraction module, a talent portrait module and a feedback module. Its characteristics are: Among them, the knowledge graph construction module includes: named entity recognition and relationship extraction; Import a pre-trained end-to-end neural network model from the model library, extract relevant text from the sample library through the sample interface, input the model, and output triples of all entities and related relations in the text; The data obtained by deep learning through the feedback module is added to the original database for joint training to further improve the accuracy of the model; the pre-trained end-to-end neural network is imported from the model library to directly predict entity relationships; the external knowledge of entity relationships obtained in this way is supplemented into the training of word vectors; The label extraction module includes label selection, data cleaning and label storage; Tag selection: For users in different scenarios, different strategies for tag selection should be implemented based on user needs and the application direction of user profiles. Data cleaning: The target label may be directly the content corresponding to a field in the database, or it may be the content in the content. In this case, the content corresponding to the field needs to be cleaned; Tag storage: The obtained tags are stored in the HBase database according to the user ID. Each user ID will correspond to multiple tags. The corresponding content under the rowkey ID is a key tag name, and the value is a dict of the tag value. The talent portrait module is used to organize the specific storage fields of user portraits; Feedback module: Users will provide feedback on user profile labels. If the labels are incorrect, they will be fed back to the sample library or database according to the direction of label generation as a correction to the data and algorithm. The specific construction steps between each module are as follows: S1: Retrieve the talent resume database and clean and merge the data; S2: Extract the talent's label, industry, field, professional skills, work experience, and educational experience from the structured field as the talent's label; S3: Build an end-to-end industry knowledge graph, starting with professional skills knowledge in various fields, and construct a domain-industry-skill graph network to serve as content supplement when building the label system; S4: Retrieve: Talent resume database. For unstructured fields, we use NLP algorithms and knowledge graph relationship extraction algorithms to extract labels of skills, industry experience, and years of experience. S5: Integrate structured and unstructured tags, perform semantic alignment for secondary verification, and use them as talent labels; S6: Extract keywords from new project cases and recommend talents based on talent tags.

2. The method for constructing a user profile in combination with a knowledge graph according to claim 1, characterized in that: The knowledge graph construction module includes an end-to-end model that uses remote supervision and pre-trained models to directly obtain entities and relationships from input text to construct a knowledge graph; the text here includes company-related text information, numerical information, announcements, organizational structure and news data; but does not include standard data in the company database.

3. The method for constructing a user profile in combination with a knowledge graph according to claim 2, characterized in that: Import the trained model into the deep learning prediction part. When text input comes in, it directly enters this module to output the relationship between entities. The output results need to pass through LSTM decoding and Softmax layer to obtain the final entity relationship. The pseudo code of the pre-trained end-to-end model is as follows: gettexts\\Get company-related articles in the sample library; wordembedding\\The input text is processed through a shared word embedding; encodinginputbyBiLSTM\\Connect a bidirectional LSTM layer to encode the input; NERbyLSTM\\Named entity recognition using LSTM; relationsclassifierbyCNN\\Use CNN for relation classification.

4. The method for constructing a user profile in combination with a knowledge graph according to claim 3, characterized in that: The knowledge graph construction module is achieved by sharing the underlying model parameters. During training, both tasks will update the shared parameters through the backpropagation algorithm to realize the dependency between the two subtasks.

5. The method for constructing a user profile in combination with a knowledge graph according to claim 1, characterized in that: When selecting the tag, after selecting the tag, it is necessary to determine the field to be selected in the database and extract the required tag from the field.

6. The method for constructing a user profile in combination with a knowledge graph according to claim 1, characterized in that: During the data cleaning, an algorithm is selected to operate on relevant content.

7. The method for constructing a user profile in combination with a knowledge graph according to claim 6, characterized in that: If the user selects the tag "Sentiment Analysis" for "Work Report," the first step is to extract the "Work Report" field from the relevant applications in the database, then run the sentiment analysis algorithm on the corresponding content to determine its sentiment tendency and finally obtain the tag.

8. The method for constructing a user profile in combination with a knowledge graph according to claim 1, characterized in that: The talent profile module receives two parts of information; ① Basic tag fields and tag values ​​obtained based on the database; ② Various attributes expanded based on the knowledge graph and user entities; Merge the labels obtained in ① and ② and store them in the HBase database. The labels obtained from the knowledge graph need to be proofread, supplemented and modified for the labels stored in the database.

Images