A person-job matching method based on LDA model, dependency syntax and deep learning

Abstract

The present invention relates to the technical field of matching methods, and discloses a person-job matching method based on LDA model, dependency syntax and deep learning, including the following steps: S1, classifying the professional names into the professional secondary classification; S2, unifying the school names into the correct expression form; S3, normalizing the job titles by editing distance and repeated fields: formulating a standard job title vocabulary by dividing the job functions of the recruitment website, and calculating the editing distance and repeated fields between the job titles in the text information and the preset standard job titles. The present invention proposes a person-job matching method based on LDA model, dependency syntax and deep learning. The implicit Dirichlet distribution model, dependency syntax and deep learning model method of the present invention and this patent solve the above-mentioned defects of wasting manpower and time from the perspective of practicality, and realize the effective extraction of resume information.

Description

Technical Field

[0001] The present invention relates to the field of matching methods, and in particular to a person-job matching method based on LDA model, dependency syntax and deep learning. Background Art

[0002] At present, we have mastered the following types of resume parsing technologies:

[0003] 1. Text normalization processing. There are often multiple forms of expression in the description of job titles and school names in resumes, but they all refer to the same type of job or the same school in essence. From the historical recruitment data set, the job titles or school names with a number of occurrences greater than a preset threshold are sorted from large to small, and the job titles or school names are determined to be correct; the resume text data is cleaned by regular expressions, and matched by edit distance in the mapping vocabulary, or a standardized vocabulary is constructed and matched directly to obtain the normalized results of the job titles or school names.

[0004] 2. The keyword determination module first performs word segmentation on the resume information according to the preset resume information training library to obtain a resume word set; searches for the weights and relevances corresponding to the words in the resume word set in the preset resume information training library; generates a comprehensive result of the resume based on the corresponding weights and relevances in the found resume word set, and sorts the words in the resume from high to low according to the comprehensive result, and outputs the first word in the sorting as the keyword of the resume.

[0005] 3. Person-job matching based on label information, which includes one or more of job function classification, high-frequency keywords, skill keywords, industry segments, salary forecasts, required years of work experience, and required education level; matching calculations are performed based on the job labels and basic information in the job text information and resume text information; the final job and resume matching score is calculated by combining the keyword score and other information weighted scores.

[0006] The above technologies all have defects. There are insurmountable obstacles to achieving the goal of matching people with jobs at the practical level, mainly:

[0007] 1. Method 1 This normalization method is relatively vague in setting the preset threshold, requires a lot of preliminary text cleaning work, easily filters out uncommon jobs, and requires a lot of job title information for screening.

[0008] 2. Method 2 requires the preparation of a preset resume information training library, which requires a lot of time and manpower to summarize the keyword weights and relevance of different positions. In addition, this method is limited to the weights and relevance of words in the preset resume library, and cannot effectively extract keywords outside the preset resume library.

[0009] 3. Method 3 is mainly based on word matching in the text, ignoring the semantic matching in the text; for job descriptions and resume-related work experience, similar skills and experience are expressed in different ways, and keywords alone cannot match the job description and resume text well.

[0010] To solve the above problems, this application proposes a person-job matching method based on LDA model, dependency syntax and deep learning. Summary of the invention

[0011] (I) Purpose of the invention

[0012] In order to solve the technical problems existing in the background technology, the present invention proposes a person-job matching method based on LDA model, dependency syntax and deep learning. The implicit Dirichlet distribution model, dependency syntax and deep learning model method of the present invention and this patent are based on practicality to solve the above-mentioned defects of consuming manpower and time, and realize the effective extraction of resume information.

[0013] (II) Technical solution

[0014] To solve the above problems, the present invention provides a person-job matching method based on LDA model, dependency syntax and deep learning, comprising the following steps:

[0015] S1. Classify the professional names into the secondary professional classification;

[0016] S2. Unify the school names into the correct form;

[0017] S3, edit distance and repeated fields are used to normalize job titles:

[0018] A standard job title vocabulary is developed based on the division of job functions on the recruitment website, and the edit distance and repeated fields between the job titles in the text information and the preset standard job titles are calculated. The job titles are normalized using the above two methods: edit distance and repeated fields.

[0019] S4, extracting text keywords: regularized extraction, LDA model, dependency syntax model and Roberta model. In the subsequent steps, the keywords extracted by various methods are comprehensively considered, sorted from large to small according to the weight of the words, and the top 20 keywords are output;

[0020] The dependency syntax model uses the pyltp library to perform word segmentation, part-of-speech tagging, and dependency syntax analysis on the text:

[0021] Locate keywords based on the core word list, which includes: responsible for participating verbs, record the positions of these core words and words in parallel relationship (COO) with these core words in the text, and then find words that have attributive relationship (ATT), adverbial relationship (ADV), verb-object relationship (VOB), fronted object (FOB) and subject-predicate relationship (SBV) with the core words;

[0022] Excluding the core words and stop words mentioned above, the words with the highest frequency of occurrence are counted, and the keywords are located according to the high frequency, and then the words that have ATT or VOB dependency relationships with the high frequency words are searched;

[0023] The core words and high-frequency words mentioned above are limited to verbs and nouns, and words of other parts of speech are not considered;

[0024] LDA model, call the LDA model of the gensim library to perform unsupervised classification of resumes in the resume database:

[0025] First, we clean the resumes in the resume database, remove prepositions and conjunctions without obvious meanings, remove high-frequency words with a frequency of more than 50% of the text, and remove low-frequency words with a frequency of less than 20 texts. We use the remaining words to build a dictionary and corpus. Then, we use the LDA model to divide the text into 25 topics, sort them from high to low according to the frequency of the words in different topics, and output the top 200 words with the highest probability.

[0026] S5. Calculation of person-job matching:

[0027] The calculation of the school dimension in the overall fit;

[0028] Calculation of overall matching degree of professional dimension;

[0029] Repeat the field dimension in the calculation of overall matching degree;

[0030] The length of relevant work experience is used in the calculation of overall matching degree;

[0031] Calculation of overall matching degree of relevant skill dimensions;

[0032] In the calculation of the overall matching degree of the text semantic similarity dimension, the cosine value of the text word is calculated after vectorization as the measure of text semantic similarity.

[0033] S6. Word embedding deep learning pre-training model based on PyTorch deep learning.

[0034] We use PyTorch to build a machine learning framework for resume information parsing. In order to accelerate the efficiency of machine learning, we use GPU as the hardware device of the accelerator, and therefore choose the PyTorch version (torch-gpu) that includes GPU acceleration in the Python language.

[0035] Preferably, in S1, according to the professional classification table of the undergraduate major catalog of ordinary colleges and universities, there are 93 secondary disciplines, and the 740 professional names in the professional catalog published by the Ministry of Education are word-vectorized through the Chinese roberta wwm ext pre-training library, and then the 740 sub-majors are divided into 93 major categories through the nearest neighbor algorithm (KNN).

[0036] Preferably, in S2, based on the public list of national higher education institutions obtained from the website of the Ministry of Education and the correct school names in QS500, some simple text cleaning is performed on the resumes uploaded by the applicants, including: removing the suffixes after "university" and "college" and other annotations in brackets; mapping the denoised school names with the correct school name vocabulary.

[0037] Preferably, in S3, if the minimum edit distance and the maximum repetition field point to the same standard job title, the correct job title is output; if the standard job titles mapped by the above two methods are not unified, the corresponding standard job titles are output according to different conditions: if some character strings of the resume job title completely overlap with the standard job title mapped by the maximum number of repeated characters, the job title mapped by the maximum number of repeated characters is output; if some character strings of the resume job title completely overlap with the standard job title mapped by the minimum edit distance, the job title mapped by the minimum edit distance is output; if the resume job title and the standard job title mapped by the two methods have no completely repeated character strings, the original job title in the resume is output.

[0038] Preferably, in S4, the LDA model of the gensim library is called to perform unsupervised classification of resumes in the resume database:

[0039] First, we clean the resumes in the resume database, remove prepositions and conjunctions without obvious meanings, remove high-frequency words with a frequency of more than 50% of the text, and remove low-frequency words with a frequency of less than 20 texts. We use the remaining words to build a dictionary and corpus. Then, we use the LDA model to divide the text into 25 topics, sort them from high to low according to the frequency of the words in different topics, and output the top 200 words with the highest probability.

[0040] The LDA model generates topic classification as follows:

[0041] Document 1: word 11, word 12, ..., word 1n

[0042] Document 2: word 21, word 22, ..., word 2n

[0043] …

[0044] Document m: word m1, word m2, ..., word mn

[0045] See attached Figure 2

[0046] Where α is a k-dimensional vector, αk=1…K represents the prior probability of the topic for any document d: αk;

[0047] θFor any document d, its topic distribution is θd, that is, the probability that document d belongs to topic k is θ d ;

[0048] θ d ~Dirichlet k(α), θd follows the Dirichlet distribution;

[0049] Z is an N-dimensional vector from 1 to k, that is, for the nth word in document d, the probability of belonging to topic k is: Z dn ;

[0050] Z d＝1…M,n＝1…N ~Multi k(θ d ), Z d＝1…M,n＝1…N Subject to polynomial division;

[0051] W is an N-dimensional vector from 1 to V, that is, when document d belongs to topic k, the nth word generated is: W dn ;

[0052] W d＝1…W,n＝1…N It follows a multinomial distribution;

[0053] is a V-dimensional probability vector, indicating that for topic k, the word distribution is:

[0054] It follows the Dirichlet distribution;

[0055] β is a V-dimensional vector, β w＝1…V Represents the prior probability of a word for any topic k: β w ;

[0056] K: the number of topics;

[0057] V: the number of all words in the vocabulary;

[0058] M: the number of documents;

[0059] N: the number of words in all documents;

[0060] α->θ->Z: Sample from Dirichlet distribution α to generate topic distribution θ of document d d , from the multinomial distribution of topics θ dThe topic Z of the nth word in the sampled generated document d dn , i.e., the document-topic parameter matrix;

[0061] The word distribution of the topic is sampled from the Dirichlet distribution β From the multinomial distribution of words Generate the nth word W by sampling dn , i.e. the topic-word matrix;

[0062] The LDA model can choose Gibbs sampling or variational inference EM sampling to update the iterative document-topic parameter matrix Z dn and the topic-word matrix W dn ; Gibbs sampling uses the Mallet library in Java, which takes a long time to calculate but has high model accuracy; variational inference EM sampling is the default sampling method for the LDA model in gensim;

[0063] Roberta model, the word vector of this patent calls the Chinese Roberta wwmext pre-trained model under the transformers library to calculate the word vector to calculate the relationship between words and documents:

[0064] chinese Roberta wwm ext is a word vector for Chinese generated based on the Roberta model. After cleaning and segmenting the text data, the word frequency is counted, the text is characterized by word frequency (CountVectorizer), and candidate text keywords are generated; the text and candidate keywords are word-embedded (word embedding) through the chinese Roberta wwm ext pre-trained model. In order to select the keywords with the closest meaning to the text, this method will select the words with the highest similarity to the text from the candidate keywords, and select the words with the smallest similarity between keywords from the candidate keywords, where the similarity will be calculated by the cosine value.

[0065] Preferably, in S5, the school dimension is used in the calculation of the overall matching degree:

[0066] Set corresponding school weights according to the rankings of different schools.

[0067] Preferably, in S5, the professional dimension is used in the calculation of the overall matching degree:

[0068] The professional requirements in the job description and the professional background of the applicant in the resume are extracted. If the professional requirements in the job and the professional name in the resume belong to the same category (professional secondary classification), the applicant's professional background is considered to be consistent with the requirements in the job description.

[0069] Preferably, in S5, the calculation of the overall matching degree of the repeated field dimension is:

[0070] After word segmentation of the job description and resume, the proportion of repeated words in the total vocabulary is counted. The total vocabulary is the number of all words that appear in the job description.

[0071] Preferably, in S5, the overall matching degree of the working experience duration is calculated as follows: the LDA model is used to determine whether the job description and the working experience description in the resume are the same topic, the working experience durations of all the same topics are accumulated to obtain the relevant working years, the requirements for the relevant working experience years in the job description are extracted through regularization, the working years requirements in the job description and the working years in the resume are compared, if the working years in the resume are greater than the requirements in the job description, it is considered that the working years in the resume meet the job requirements.

[0072] Preferably, in S5, the calculation of the overall matching degree of the relevant skill dimension is as follows: skill keywords in the job description and resume text are extracted according to the LDA model, dependency syntax and deep model, and the overlapping skills between the two are compared. For skills that are required to be "proficient" in the job, the weight is 1; for skills that are required to be "prioritized", the weight is 0.9. For skills that do not appear in the job requirements but appear in the resume, the skills in the job-skill keyword vocabulary are matched, and the weight is 1. The job-skill keyword vocabulary is a database sorted by job classification based on historical job description data. Finally, all skills are weighted averaged to obtain the final skill matching score.

[0073] The above technical solution of the present invention has the following beneficial technical effects:

[0074] By combining the advantages of LDA model, dependency syntax and deep learning algorithm in a pioneering way, a variety of model methods are introduced. On the one hand, regular expression templates are used to extract common specific text expressions, and on the other hand, keywords are extracted for general text content; text keywords are extracted through a variety of methods to meet the diversity and professionalism of keywords, effectively avoiding the construction of a corpus that requires a lot of manual processing, and combining the good adaptability of machine models and the accuracy of information extraction of deep learning models. It can effectively and accurately extract keywords from resume information and job information. It has good practicality; through this method, accurate matching of resume information and job descriptions can be achieved based on limited job description samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 This is a flow chart of a person-job matching method based on LDA model, dependency syntax and deep learning proposed by the present invention.

[0076] Figure 2 This is a diagram of how the LDA model generates topic classification in a person-job matching method based on the LDA model, dependency syntax and deep learning proposed in the present invention.

[0077] Figure 3 This is a basic flow chart of deep learning for extracting keywords from the entire resume information and job description in a person-job matching method based on the LDA model, dependency syntax and deep learning proposed by the present invention. DETAILED DESCRIPTION

[0078] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings. It should be understood that these descriptions are only exemplary and are not intended to limit the scope of the present invention. In addition, in the following description, the description of well-known structures and technologies is omitted to avoid unnecessary confusion of the concept of the present invention.

[0079] like Figure 1-3 As shown, the present invention proposes a person-job matching method based on LDA model, dependency syntax and deep learning, which is characterized by comprising the following steps:

[0080] S1. Classify the professional names into the secondary professional classification;

[0081] S2. Unify the school names into the correct form;

[0082] S3, edit distance and repeated fields are used to normalize job titles:

[0083] A standard job title vocabulary is developed based on the division of job functions on the recruitment website, and the edit distance and repeated fields between the job titles in the text information and the preset standard job titles are calculated. The job titles are normalized using the above two methods: edit distance and repeated fields.

[0084] S4, extracting text keywords: regularized extraction, LDA model, dependency syntax model and Roberta model. In the subsequent steps, the keywords extracted by various methods are comprehensively considered, sorted from large to small according to the weight of the words, and the top 20 keywords are output;

[0085] The dependency syntax model uses the pyltp library to perform word segmentation, part-of-speech tagging, and dependency syntax analysis on the text:

[0086] Locate keywords based on the core word list, which includes: responsible for participating verbs, record the positions of these core words and words in parallel relationship (COO) with these core words in the text, and then find words that have attributive relationship (ATT), adverbial relationship (ADV), verb-object relationship (VOB), fronted object (FOB) and subject-predicate relationship (SBV) with the core words;

[0087] Excluding the core words and stop words mentioned above, the words with the highest frequency of occurrence are counted, and the keywords are located according to the high frequency, and then the words that have ATT or VOB dependency relationships with the high frequency words are searched;

[0088] The core words and high-frequency words mentioned above are limited to verbs and nouns, and words of other parts of speech are not considered;

[0089] LDA model, call the LDA model of the gensim library to perform unsupervised classification of resumes in the resume database:

[0090] First, we clean the resumes in the resume database, remove prepositions and conjunctions without obvious meanings, remove high-frequency words with a frequency of more than 50% of the text, and remove low-frequency words with a frequency of less than 20 texts. We use the remaining words to build a dictionary and corpus. Then, we use the LDA model to divide the text into 25 topics, sort them from high to low according to the frequency of the words in different topics, and output the top 200 words with the highest probability.

[0091] S5. Calculation of person-job matching:

[0092] The calculation of the school dimension in the overall fit;

[0093] Calculation of overall matching degree of professional dimension;

[0094] Repeat the field dimension in the calculation of the overall match;

[0095] The length of relevant work experience is used in the calculation of overall matching degree;

[0096] Calculation of overall matching degree of relevant skill dimensions;

[0097] In the calculation of the overall matching degree of the text semantic similarity dimension, the cosine value of the text word is calculated after vectorization as the measure of text semantic similarity.

[0098] S6. Word embedding deep learning pre-training model based on PyTorch deep learning.

[0099] We used PyTorch to build a machine learning framework for resume information parsing. To speed up the efficiency of machine learning, we used GPU as the hardware device for the accelerator, and thus chose the PyTorch version (torch-gpu) with GPU acceleration in Python.

[0100] The basic process of deep learning for extracting keywords from the entire resume information and job description is as follows: Figure 1 :

[0101] In an optional embodiment, in S1, according to the professional classification table of the undergraduate major catalog of ordinary colleges and universities, there are 93 secondary disciplines. The 740 professional names in the professional catalog published by the Ministry of Education are word-vectorized through the Chinese Erotica wwm ext pre-training library, and then the 740 sub-majors are divided into 93 major categories through the nearest neighbor algorithm (KNN).

[0102] In an optional embodiment, in S2, based on the public list of national higher education institutions obtained from the website of the Ministry of Education and the correct school names in QS500, some simple text cleaning is performed on the resume uploaded by the applicant, including: removing the suffixes after "university" and "college" and other annotations in brackets; mapping the denoised school names with the correct school name vocabulary.

[0103] In an optional embodiment, in S3, if the minimum edit distance and the maximum repetition field point to the same standard job title, the correct job title is output; if the standard job titles mapped by the above two methods are not unified, the corresponding standard job titles are output according to different conditions: if some character strings of the resume job title completely overlap with the standard job title mapped by the maximum number of repeated characters, the job title mapped by the maximum number of repeated characters is output; if some character strings of the resume job title completely overlap with the standard job title mapped by the minimum edit distance, the job title mapped by the minimum edit distance is output; if the resume job title and the standard job title mapped by the two methods have no completely repeated character strings, the original job title in the resume is output.

[0104] In an optional embodiment, in S4, the LDA model of the gensim library is called to perform unsupervised classification of resumes in the resume database:

[0105] First, we clean the resumes in the resume database, remove prepositions and conjunctions without obvious meanings, remove high-frequency words with a frequency of more than 50% of the text, and remove low-frequency words with a frequency of less than 20 texts. We use the remaining words to build a dictionary and corpus. Then, we use the LDA model to divide the text into 25 topics, sort them from high to low according to the frequency of the words in different topics, and output the top 200 words with the highest probability.

[0106] The LDA model generates topic classification as follows:

[0107] Document 1: word 11, word 12, ..., word 1n

[0108] Document 2: word 21, word 22, ..., word 2n

[0109] …

[0110] Document m: word m1, word m2, ..., word mn

[0111] See attached Figure 2

[0112] Where α is a k-dimensional vector, αk=1…K represents the prior probability of the topic for any document d: αk;

[0113] θFor any document d, its topic distribution is θd, that is, the probability that document d belongs to topic k is θ d ;

[0114] θ d ~Dirichlet k(α), θd follows the Dirichlet distribution;

[0115] Z is an N-dimensional vector from 1 to k, that is, for the nth word in document d, the probability of belonging to topic k is: Z dn ;

[0116] Z d＝1…M,n＝1…N ~Multi k(θ d ), Z d＝1…M,n＝1…N Subject to polynomial division;

[0117] W is an N-dimensional vector from 1 to V, that is, when document d belongs to topic k, the nth word generated is: W dn ;

[0118] W d＝1…W,n＝1…N It follows a multinomial distribution;

[0119] is a V-dimensional probability vector, indicating that for topic k, the word distribution is:

[0120] It follows the Dirichlet distribution;

[0121] β is a V-dimensional vector, β w＝1…V Represents the prior probability of a word for any topic k: β w ;

[0122] K: the number of topics;

[0123] V: the number of all words in the vocabulary;

[0124] M: the number of documents;

[0125] N: the number of words in all documents;

[0126] α->θ->Z: Sample from Dirichlet distribution α to generate topic distribution θ of document d d, from the multinomial distribution of topics θ d The topic Z of the nth word in the sampled generated document d dn , that is, the document-topic parameter matrix

[0127] The word distribution of the topic is sampled from the Dirichlet distribution β From the multinomial distribution of words Generate the nth word W by sampling dn , i.e. the topic-word matrix;

[0128] The LDA model can choose Gibbs sampling or variational inference EM sampling to update the iterative document-topic parameter matrix Z dn and the topic-word matrix W dn ; Gibbs sampling uses the Mallet library in Java, which takes a long time to calculate but has high model accuracy; variational inference EM sampling is the default sampling method for the LDA model in gensim;

[0129] Roberta model, the word vector of this patent calls the Chinese Roberta wwmext pre-trained model under the transformers library to calculate the word vector to calculate the relationship between words and documents:

[0130] chinese Roberta wwm ext is a word vector for Chinese generated based on the Roberta model. After cleaning and segmenting the text data, the word frequency is counted, the text is characterized by word frequency (CountVectorizer), and candidate text keywords are generated; the text and candidate keywords are word-embedded (word embedding) through the chinese Roberta wwm ext pre-trained model. In order to select the keywords with the closest meaning to the text, this method will select the words with the highest similarity to the text from the candidate keywords, and select the words with the smallest similarity between the keywords from the candidate keywords, so as to achieve high relevance and diversity of keywords, and the similarity will be calculated by the cosine value.

[0131] In an optional embodiment, in S5, the school dimension is used in the calculation of the overall matching degree:

[0132] Set corresponding school weights according to the rankings of different schools.

[0133] In an optional embodiment, in S5, the professional dimension is used in the calculation of the overall matching degree:

[0134] The professional requirements in the job description and the professional background of the applicant in the resume are extracted. If the professional requirements in the job and the professional name in the resume belong to the same category (professional secondary classification), the applicant's professional background is considered to be consistent with the requirements in the job description.

[0135] In an optional embodiment, in S5, the calculation of the overall matching degree of the repeated field dimension is:

[0136] After word segmentation of the job description and resume, the proportion of repeated words in the total vocabulary is counted. The total vocabulary is the number of all words that appear in the job description.

[0137] Preferably, in S5, the overall matching degree of the working experience duration is calculated as follows: the LDA model is used to determine whether the job description and the working experience description in the resume are the same topic, the working experience durations of all the same topics are accumulated to obtain the relevant working years, the requirements for the relevant working experience years in the job description are extracted through regularization, the working years requirements in the job description and the working years in the resume are compared, if the working years in the resume are greater than the requirements in the job description, it is considered that the working years in the resume meet the job requirements.

[0138] In an optional embodiment, in S5, the calculation of the overall matching degree of the relevant skill dimension is as follows: skill keywords in the job description and resume text are extracted according to the LDA model, dependency syntax and deep model, and the overlapping skills between the two are compared. For skills that are required to be "proficient" in the job, the weight is 1; for skills that are required to be "prioritized", the weight is 0.9. For skills that do not appear in the job requirements but appear in the resume, the skills in the job-skill keyword vocabulary are matched, and the weight is 1. The job-skill keyword vocabulary is a database sorted by job classification based on historical job description data. Finally, all skills are weighted and averaged to obtain the final skill matching score.

[0139] In the present invention, this method extracts English skill nouns appearing in resumes through regular expressions; this method realizes unsupervised classification of resumes through the LDA model, calculates the weights of word sets in the resume database for resumes of different categories, and realizes the automation of keyword extraction; and extracts resume information in a targeted manner through dependency syntax and deep learning methods.

[0140] It should be understood that the above specific embodiments of the present invention are only used to illustrate or explain the principles of the present invention, and do not constitute a limitation of the present invention. Therefore, any modifications, equivalent substitutions, improvements, etc. made without departing from the spirit and scope of the present invention should be included in the protection scope of the present invention. In addition, the appended claims of the present invention are intended to cover all changes and modifications that fall within the scope and boundaries of the appended claims, or the equivalent forms of such scope and boundaries.

Claims

1. A person-job matching method based on LDA model, dependency syntax and deep learning, characterized in that: The following steps are involved: S1. Classify the professional names into the secondary professional classification; S2. Unify the school names into the correct form; S3, edit distance and repeated fields are used to normalize job titles: A standard job title vocabulary is developed based on the division of job functions on the recruitment website, and the edit distance and repeated fields between the job titles in the text information and the preset standard job titles are calculated. The job titles are normalized using the above two methods: edit distance and repeated fields. S4, extracting text keywords: regularized extraction, LDA model, dependency syntax model and Roberta model. In the subsequent steps, the keywords extracted by various methods are comprehensively considered, sorted from large to small according to the weight of the words, and the top 20 keywords are output; The dependency syntax model uses the pyltp library to perform word segmentation, part-of-speech tagging, and dependency syntax analysis on the text: Locate keywords based on the core word list, which includes: responsible for participating verbs, record the positions of these core words and words in parallel relationship with these core words in the text, and then find words that have attributive relationship, adverbial relationship, verb-object relationship, prepositioned object and subject-predicate relationship with the core words; Excluding the above core words and stop words, find the words with the highest frequency of occurrence, locate the keywords based on the high frequency, and then find the words that have a subject-predicate relationship or a verb-object relationship with the high frequency words; The core words and high-frequency words mentioned above are limited to verbs and nouns, and words of other parts of speech are not considered; LDA model, call the LDA model of the gensim library to perform unsupervised classification of resumes in the resume database: First, we clean the resumes in the resume database, remove prepositions and conjunctions without obvious meanings, remove high-frequency words with a frequency of more than 50% of the text, and remove low-frequency words with a frequency of less than 20 texts. We use the remaining words to build a dictionary and corpus. Then, we use the LDA model to divide the text into 25 topics, sort them from high to low according to the frequency of the words in different topics, and output the top 200 words with the highest probability. S5. Calculation of person-job matching: The calculation of the school dimension in the overall fit; Calculation of overall matching degree of professional dimension; Repeat the field dimension in the calculation of overall matching degree; The length of relevant work experience is used in the calculation of overall matching degree; Calculation of overall matching degree of relevant skill dimensions; In the calculation of the overall matching degree of the text semantic similarity dimension, the cosine value of the text word is calculated after vectorization as the measure of text semantic similarity; S6, word embedding deep learning pre-training model based on PyTorch deep learning; Use PyTorch to build a machine learning framework for resume information parsing.

2. The person-job matching method based on LDA model, dependency syntax and deep learning according to claim 1 is characterized in that: In S1, according to the professional classification table of the undergraduate major catalog of ordinary colleges and universities, there are 93 secondary disciplines. The 740 professional names in the professional catalog are word-vectorized through the pre-training library, and then the 740 sub-majors are divided into 93 major categories through the nearest neighbor algorithm.

3. The person-job matching method based on LDA model, dependency syntax and deep learning according to claim 1 is characterized in that: In S2, some simple text cleaning is performed on the resumes uploaded by the applicants, including: removing the suffixes after "university" and "college" and other annotations in brackets; and mapping the denoised school names with the correct school name vocabulary.

4. The person-job matching method based on LDA model, dependency syntax and deep learning according to claim 1 is characterized in that: In S3, if the minimum edit distance and maximum repetition fields point to the same standard job name, the correct job name is output; if the standard job names mapped by the above two methods are not unified, the corresponding standard job names are output according to different conditions: if some character strings of the resume job name completely overlap with the standard job name mapped by the maximum number of repeated characters, the job name mapped by the maximum number of repeated characters is output; if some character strings of the resume job name completely overlap with the standard job name mapped by the minimum edit distance, the job name mapped by the minimum edit distance is output; if there are no completely repeated character strings between the resume job name and the standard job name mapped by the two methods, the original job name in the resume is output.

5. The person-job matching method based on LDA model, dependency syntax and deep learning according to claim 1 is characterized in that: In S5, the school dimension is used in the calculation of overall matching: Set corresponding school weights according to the rankings of different schools.

6. The person-job matching method based on LDA model, dependency syntax and deep learning according to claim 1 is characterized in that: In S5, the professional dimension is used in the calculation of overall matching degree: The professional requirements in the job description and the professional background of the applicant in the resume are extracted. If the professional requirements in the job and the professional name in the resume belong to the same category, the applicant's professional background is considered to be consistent with the requirements in the job description.

7. The person-job matching method based on LDA model, dependency syntax and deep learning according to claim 1 is characterized in that: In S5, the calculation of the overall match degree of the repeated field dimension is: After word segmentation of the job description and resume, the proportion of repeated words in the total vocabulary is counted. The total vocabulary is the number of all words that appear in the job description.

8. The person-job matching method based on LDA model, dependency syntax and deep learning according to claim 1, characterized in that: In S5, the overall matching degree of working experience is calculated as follows: the LDA model is used to determine whether the job description and the working experience description in the resume are the same topic, the working experience duration of all the same topics is accumulated to obtain the relevant working years, and the requirements for the relevant working experience years in the job description are extracted through regularization. The working years requirements in the job description and the working years in the resume are compared. If the working years in the resume are greater than the requirements in the job description, it is considered that the working years in the resume meet the job requirements.

9. The person-job matching method based on LDA model, dependency syntax and deep learning according to claim 1, characterized in that: In S5, the calculation of the overall matching degree of the relevant skill dimension is as follows: based on the LDA model, dependency syntax and deep model, skill keywords in the job description and resume text are extracted, and the overlapping skills between the two are compared. For skills that are required to be "proficient" in the job, the weight is 1; for skills that are required to be "prioritized", the weight is 0.

9. For skills that do not appear in the job requirements but appear in the resume, the skills in the job-skill keyword vocabulary are matched, and the weight is 1. The job-skill keyword vocabulary is a database organized by job classification based on historical job description data. Finally, all skills are weighted averaged to obtain the final skill matching score.

Images