A human resources resume screening method based on word classification

Abstract

A method for screening human resource resumes based on word classification, characterized in that: the specific steps of the method are: S101, collecting and saving the demand content of the employing department; S102, extracting demand keywords, extracting keywords from the collected demand content, and using the TF-IDF algorithm to extract keywords; S103, searching in the resume database, using the keyword field in the resume database as the entry, searching through the demand keywords, and finding the resumes with the keywords; S104, screening the search results, saving the results of the search in step S103, and sorting them according to the number of keyword matches. Its beneficial effects are: resume structured data can realize the convenience of data storage, making data storage more orderly, and facilitating database management and information retrieval.

Description

Technical Field

[0001] The invention relates to natural language processing, and in particular to a human resources resume screening method based on word classification. Background Art

[0002] In human resource management, resume screening is a very important link. With the development of Internet technology, the number of job seekers submitting resumes through online platforms has increased dramatically, which has put the human resource department under pressure to screen a large number of resumes. In order to improve efficiency, everyone is studying how to screen resumes based on word classification and automatically screen candidates by finding keywords in them using natural language processing (NLP) technology.

[0003] Currently, resume screening technology based on word classification mainly relies on NLP and machine learning algorithms. These technologies can identify key information in resumes, such as education background, work experience and skills, and match them with job requirements.

[0004] However, there are still many difficulties in the existing technology in accurately finding key information, efficiently performing structured operations on resumes, and realizing effective data integration. On the other hand, due to the large amount of resume data and the current shortage of graphics cards and computing resources, if a large model is used for processing, although better results may be achieved, the efficiency is low and cannot adapt to large-scale resume processing and screening. Summary of the invention

[0005] The present invention uses a structured resume and an efficient and accurate mapping model to achieve precise matching of resumes and employment needs. It also achieves code reusability and easy maintenance through flexible calling and modification of regular expressions, thereby providing a human resources resume screening method based on word classification.

[0006] A method for screening human resources resumes based on word classification, characterized in that the specific steps of the method are:

[0007] S101 The needs of the employing department, collect and save the needs;

[0008] S102 extracts demand keywords, extracts keywords from the collected demand content, and uses a TF-IDF algorithm to extract keywords;

[0009] S103 Resume database search, using the keyword field in the resume database as an entry, searching by the required keyword, and finding the resume where the keyword is located;

[0010] S104 filters the search results, saves the search results of step S103, and sorts them according to the number of keyword matches;

[0011] In step S103, the resume database contains four fields: resume name, resume content, keywords, and resume structured data. The generation process of the resume structured data and keywords is as follows:

[0012] S201 obtains resume data and collects resumes from the Internet or applicants;

[0013] S202 Structured resume: divide the resume into blocks according to the content, and the fields divided into blocks include basic information, professional skills, work experience, training experience and other content. The resume content after the blocks is saved in JSON format as structured resume data;

[0014] S203 extracts keywords from the mapping model, inputs the structured data from step S202 into the mapping model, and obtains keywords for the resume;

[0015] S204 saves the resume database, and saves the data generated in steps S202 and S203 in the database;

[0016] The method for dividing the resume into blocks according to the content is to use regular expressions to extract corresponding fields and corresponding corpora, and the regular expressions are used as variables of the function to divide the resume into blocks according to the content and classified and stored in the data table.

[0017] Furthermore, the structured resume data generation process is to call the regular expression in the data table in sequence through a function to obtain the key-value pairs of each block classification, where the key name is the block name, and the value is the corresponding corpus. Finally, all the key-value pairs in the resume are assembled and saved as structured resume data in json format.

[0018] Furthermore, the keyword field in step S103 is generated by a mapping model, and the mapping model is used to classify the collected corpus into corresponding keywords, and the generation process is as follows:

[0019] S301 collects resume corpus and marks corresponding keywords: collects corpus in resumes to form a corpus, and marks each corpus in the corpus with corresponding keywords; the keyword categories include job keywords, professional keywords and skill keywords, and the corpus comes from the resume database in step S204;

[0020] S302 Data preprocessing: Filter the punctuation marks in each sentence in the corpus and remove spaces;

[0021] S303: Establish a word frequency dictionary set: Use Jieba word segmentation to segment each sentence in the corpus, convert the words into the form of "word: word frequency", take the words as subscripts, and save them in the form of "word: [sequence number, word frequency]", so as to form a word frequency dictionary set;

[0022] S304 sentence vectorization: after step S303, all sentences in the corpus are vectorized. For each sentence, a vectorized zero matrix with the same dimension as the word frequency dictionary set is generated. Then, the words in the sentence after word segmentation by Jieba are mapped to the zero matrix. If a word appears once in the sentence, the value at the corresponding position in the zero matrix is ​​added by 1, thereby converting the zero matrix into a word frequency matrix. Finally, each value in the word frequency matrix is ​​divided by the quantity value of the sentence word quantity.

[0023] S305 constructs a neural network model: the structure of the neural network is five layers, the number of neurons in the input layer is the same as the number of elements in the word frequency dictionary set, the number of neurons in the second, third and fourth layers are 128, 128 and 64 respectively, and the number of neurons in the output layer is the same as the number of set keywords, that is, one neuron corresponds to one keyword;

[0024] S306: training the model by inputting the vectorized results of the sentences and the corresponding keywords into the neural network model in step S305 for iterative training 10,000 times or until the error value converges or stabilizes;

[0025] S307 saves the mapping model, saves the model trained and generated in step S306, and uses the mapping model for prediction later. The model will output the probability of each keyword, and then convert the probability into a binary label according to the threshold, and the label is the keyword.

[0026] Furthermore, in step S305, the neural network model uses Relu as an activation function, the loss function is binary cross entropy, and the optimizer uses stochastic gradient descent.

[0027] The beneficial effects of the present invention are:

[0028] 1. Resume structured data can facilitate data storage, making data storage more orderly and easy for database management and information retrieval; using block processing to divide the resume into different parts can more effectively perform specialized processing and analysis on each part. On the other hand, it also provides a basis for corpus induction. Structured processing provides a good data foundation for keyword extraction and corpus construction, thus laying the foundation for subsequent data analysis and model training.

[0029] 2. Flexible calling and modification of regular expressions can achieve code reusability. By cyclically calling regular expressions in the data table, they can be reused in different application scenarios, improving the reusability of the code. Flexible calling and modification of regular expressions makes the system easy to maintain and upgrade, providing a guarantee for the long-term development of the system.

[0030] 3. Through the application of neural network models, the efficiency of sentence processing is improved, efficient and fast sentence processing is achieved, and the speed and accuracy of keyword extraction are improved; at the same time, the neural network model can accurately extract keywords, providing a good foundation for accurately matching resumes and employment needs.

[0031] 4. The mapping model can process the corpus in structured resumes quickly and accurately, improving the efficiency of classification and keyword acquisition; the corpus classification achieved through the mapping model can obtain more accurate keywords, thereby improving the quality and efficiency of resume screening. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a flowchart of the HR resume screening method;

[0033] Figure 2 It is a structured flow chart of resume data;

[0034] Figure 3 It is a mapping model to generate a flow chart. DETAILED DESCRIPTION

[0035] In order to make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the drawings in the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments.

[0036] A human resources resume screening method based on word classification, the specific steps are:

[0037] S101 The needs of the employing department, collect and save the needs;

[0038] S102 extracts demand keywords, extracts keywords from the collected demand content, and uses a TF-IDF algorithm to extract keywords;

[0039] S103 Resume database search, using the keyword field in the resume database as an entry, searching by the required keyword, and finding the resume where the keyword is located;

[0040] S104 filters the search results, saves the search results of step S103, and sorts them according to the number of keyword matches.

[0041] The resume database in step S103 includes four fields: resume name, resume content, keywords, and resume structured data. The generation process of resume structured data and keywords is as follows:

[0042] S201 obtains resume data and collects resumes from the Internet or applicants;

[0043] S202 Structured resume: divide the resume into basic information, professional skills, work experience, training experience and other contents, and save it in json format, for example: {basic information: {name: Zhang San, gender: male...}, professional skills: proficient in office applications, work experience:...}, and save the resume content in the above format for subsequent use;

[0044] The above resume block method uses regular expressions to extract corresponding fields. For example, the regular expression is "^(basic information|personal information)" to extract the content in "basic information", and the regular expression "name|first name" is used to extract the "name" therein.

[0045] The regular expressions mentioned above are saved in the data table by block classification. This method can realize the use of regular expressions as variables, and can realize the cyclic call of different regular expressions in the data table, and realize the flexible call and modification of regular expressions. This method can achieve the following effects:

[0046] 1. Flexibility and maintainability: Storing regular expressions in a data table makes them configurable variables, which can be modified, added, or deleted at any time without modifying the code. This flexibility makes maintenance and updating easier;

[0047] 2. Reusability: By cyclically calling different regular expressions in the data table, the same pattern can be reused in different contexts, so there is no need to write a new regular expression for each specific matching requirement, thereby improving the reusability of the code;

[0048] 3. Performance optimization: Storing regular expressions in a data table avoids repeatedly compiling the same expression. Compiling regular expressions is a relatively time-consuming operation, so by reusing compiled expressions, the execution efficiency of the code can be improved.

[0049] 4. Easy to maintain and expand: If you need to add a new regular expression, you only need to add a row of records to the data table. This is much more convenient than adding a new regular expression directly in the code.

[0050] During use, the regular expressions in the data table are called in sequence through the function to obtain the key-value pairs of each block classification, where the key name is the block name and the value is the corresponding corpus. Finally, all the key-value pairs in the resume are assembled and saved as resume structured data in json format.

[0051] S203 extracts keywords from the mapping model, inputs the structured data from step S202 into the mapping model, and obtains keywords for the resume;

[0052] S204 saves the resume database, and saves the data generated in steps S202 and S203 in the database. The database used can be MySQL, sqlserver, oracle and other databases.

[0053] like Figure 3 As shown in the figure, the mapping model is used to classify the collected resume corpus into corresponding keywords. The generation process is as follows:

[0054] S301 collects resume corpus and marks corresponding keywords: collects corpus in resumes to form a corpus, and marks each corpus in the corpus with corresponding keywords;

[0055] The keyword categories mentioned above include job keywords, professional keywords and skill keywords. The corpus in the resume mentioned above comes from the resume database in step S204.

[0056] S302 Data preprocessing: Filter the punctuation marks in each sentence in the corpus and remove spaces;

[0057] S303: Establish a word frequency dictionary set: Use Jieba word segmentation to segment each sentence in the corpus, convert the words into the form of "word: word frequency", take the words as subscripts, and save them in the form of "word: [sequence number, word frequency]", so as to form a word frequency dictionary set;

[0058] S304 Sentence vectorization: After step S303, all sentences in the corpus are vectorized. For each sentence, a vectorized zero matrix with the same dimension as the word frequency dictionary set is generated, and then the words in the sentence after word segmentation are mapped to the zero matrix. If a word appears once in the sentence, the value at the corresponding position in the zero matrix is ​​increased by 1, thereby converting the zero matrix into a word frequency matrix. Finally, each value in the word frequency matrix is ​​divided by the quantity value of the sentence word quantity.

[0059] To explain the process of sentence vectorization, for example, the existing word frequency dictionary has 1,000 words. There is a corpus called "responsible for the sales of the company's online resources". After Jieba word segmentation, it will be divided into "responsible, company, online, resources, of, sales, work". In the word frequency dictionary: "responsible: [0:80]", "company: [3:21]", "online: [31:10]", "resources: [55:30]", "of: [101:8]", "sales: [208:6]", "work: [210:8]". Generate a 1*1000 zero matrix. The process of vectorizing the above sentence will first form a word frequency matrix of [1, 0, 0, 1, 0, 0, ... 0], and finally divide it by the number of words in the sentence 7 to get the sentence vectorization matrix [1 / 7, 0, 0, 1 / 7, 0, 0, ... 0].

[0060] Through the above sentence vectorization processing, the semantics in the sentence can be converted into specific numerical expressions.

[0061] S305 constructs a neural network model: the structure of the neural network is five layers. The number of neurons in the input layer is the same as the number of elements in the word frequency dictionary set. The number of neurons in the second, third and fourth layers are 128, 128 and 64 respectively. The number of neurons in the output layer is the same as the number of set keywords, that is, one neuron corresponds to one keyword. Relu is used as the activation function, the loss function is binary cross entropy, and the optimizer uses stochastic gradient descent (SGD).

[0062] S306 trains the model, iterating the neural network model in step S305 for 10,000 times or until the error value converges or stabilizes.

[0063] S307 saves the mapping model, saves the model trained and generated in step S306, and uses the mapping model for prediction later. The model will output the probability of each keyword, and then convert the probability into a binary label (ie, keyword) according to the threshold.

[0064] Through the above operations, you can achieve fast matching during the resume search process and filter out the required resumes.

[0065] The above shows and describes the basic principle and main features of the present invention and the advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments.

Claims

1. A human resources resume screening method based on word classification, characterized by: The specific steps of the method are: S101 The needs of the employing department, collect and save the needs; S102 extracts demand keywords, extracts keywords from the collected demand content, and uses a TF-IDF algorithm to extract keywords; S103 Resume database search, using the keyword field in the resume database as an entry, searching by the required keyword, and finding the resume containing the keyword; S104 filters the search results, saves the search results of step S103, and sorts them based on the number of keyword matches; In step S103, the resume database contains four fields: resume name, resume content, keywords, and resume structured data. The generation process of the resume structured data and keywords is as follows: S201 obtains resume data and collects resumes from the Internet or applicants; S202 Structured resume: divide the resume into blocks according to the content, and the fields divided into blocks include basic information, professional skills, work experience, training experience and other content. The resume content after the blocks is saved in JSON format as structured resume data; S203 extracts keywords from the mapping model, inputs the structured data from step S202 into the mapping model, and obtains keywords for the resume; S204 saves the resume database, and saves the data generated in steps S202 and S203 in the database; The method of dividing the resume into blocks according to the content is to use regular expressions to extract the corresponding fields and the corresponding corpus, and the regular expressions are used as variables of the function to divide the resume into blocks according to the content and classify them and save them in the data table; In step S103, the keyword field is generated by a mapping model. The mapping model is used to classify the collected corpus into corresponding keywords. The generation process is as follows: S301 collects resume corpus and marks corresponding keywords: collects corpus in resumes to form a corpus, and marks each corpus in the corpus with corresponding keywords; the keyword categories include job keywords, professional keywords and skill keywords, and the corpus comes from the resume database in step S204; S302 Data preprocessing: Filter the punctuation marks in each sentence in the corpus and remove spaces; S303: Establish a word frequency dictionary set: Use Jieba word segmentation to segment each sentence in the corpus, convert the words into the form of "word: word frequency", take the words as subscripts, and save them in the form of "word: [sequence number, word frequency]", so as to form a word frequency dictionary set; S304 sentence vectorization: after step S303, all sentences in the corpus are vectorized. For each sentence, a vectorized zero matrix with the same dimension as the word frequency dictionary set is generated. Then, the words in the sentence after word segmentation by Jieba are mapped to the zero matrix. If a word appears once in the sentence, the value at the corresponding position in the zero matrix is ​​added by 1, thereby converting the zero matrix into a word frequency matrix. Finally, each value in the word frequency matrix is ​​divided by the quantity value of the sentence word quantity. S305 constructs a neural network model: the structure of the neural network is five layers, the number of neurons in the input layer is the same as the number of elements in the word frequency dictionary set, the number of neurons in the second, third and fourth layers are 128, 128 and 64 respectively, and the number of neurons in the output layer is the same as the number of set keywords, that is, one neuron corresponds to one keyword; S306: training the model, iteratively training the neural network model in step S305 for 10,000 times or until the error value converges or stabilizes; S307 saves the mapping model, saves the model trained and generated in step S306, and uses the mapping model for prediction later. The model will output the probability of each keyword, and then convert the probability into a binary label according to the threshold, and the label is the keyword.

2. The method for screening human resource resumes based on word classification according to claim 1 is characterized in that: The structured resume data generation process is to call the regular expression in the data table in sequence through the function to obtain the key-value pairs of each block classification, where the key is the block name and the value is the corresponding corpus. Finally, all the key-value pairs in the resume are assembled and saved as structured resume data in json format.

3. The method for screening human resource resumes based on word classification according to claim 2 is characterized in that: In step S305, the neural network model uses Relu as the activation function, the loss function is binary cross entropy, and the optimizer uses stochastic gradient descent.

Images