An intelligent post matching method based on feature discrete coefficient and attention mechanism

Abstract

The application discloses an intelligent human-post matching method based on feature discrete coefficients and an attention mechanism, and comprises the following steps: obtaining a resume / post text corpus, pre-training a semantic extraction model, and performing model fine-tuning; dividing the resume / post text corpus into text features, numerical features and hierarchical features according to the description types of various fields; analyzing the resume / post text corpus according to different feature types to extract features and storing the features in a vector database; taking the latest N browsing records of a job seeker on a client, searching for corresponding feature values of the browsing records in the vector database, and calculating the interest portrait of the job seeker and the interest degrees of the features; for each candidate post recommendation item, obtaining a matching degree according to the similarity between the feature values of the post and the interest portrait of the corresponding feature and the interest degree of the feature; and sorting the matching degrees of the candidate items to obtain a recommendation result.

Description

Technical Field

[0001] This invention belongs to the field of job matching and recommendation, and in particular relates to an intelligent job matching method based on feature discrete coefficients and attention mechanisms. Background Technology

[0002] In recent years, with the development of artificial intelligence technology, the phenomenon that previously required professional recruiters and job seekers to manually select target candidates from a massive pool of candidates no longer exists. Instead, intelligent matching and recommendation between supply and demand parties is achieved by utilizing text descriptions or browsing history provided by both parties and deep learning methods such as text vector matching or sequence prediction.

[0003] The aforementioned text vector matching method primarily uses techniques like word2vec or TF-IDF to simply convert the entire resume text or specific fields into text vectors, and then uses vector similarity techniques for job matching. This method has several problems and drawbacks. First, the actual needs of both employers and job seekers often possess a degree of ambiguity and flexibility, which is difficult to accurately express in words. Furthermore, if the writer of the text has limited knowledge and writing skills, the meaning conveyed in the text is more likely to deviate from the author's true intention. Second, text vector generation methods like word2vec or TF-IDF only use word frequency statistics or co-occurrence information to generate word vectors, and then use these word vectors to calculate the text vector using averaging or simple weighting. The text vectors calculated in this way only contain shallow semantic information and cannot uncover the deep semantic information contained in the text. Therefore, when performing text semantic matching, the matching of similar texts may be inaccurate. Third, this method of indiscriminately converting different fields of the resume into vectors and then calculating similarity ignores the differences in the importance of different fields to the job matching model, which clearly contradicts real-world recruitment scenarios. Finally, this method overlooks a crucial fact: the interests of both employers and job seekers are not static. For example, if a candidate excels in a particular area, the hiring manager might lower or abandon requirements for other aspects. Alternatively, during peak or off-peak hiring seasons, both employers and job seekers may unintentionally raise or lower their overall requirements based on feedback from their applications. Because these changes are usually short-term or constantly evolving, neither party is inclined to revise their application materials. Therefore, methods relying solely on text vector matching for job matching are ill-suited to scenarios involving shifting interests. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an intelligent human-job matching method based on feature discrete coefficients and an attention mechanism.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: A first aspect of the present invention provides an intelligent human-job matching method based on feature discrete coefficients and an attention mechanism, the method comprising the following sub-steps:

[0006] S1. Obtain resume / job posting text corpus, generate similar sentence pairs through manual annotation, mark them as positive examples, then shuffle the similar sentence pairs as negative examples, and then input them as positive and negative examples into the pre-trained semantic extraction model for model fine-tuning;

[0007] S2, based on the description type of each field, divide the resume / job text corpus into text features, numerical features, and hierarchical features; convert the text features into semantic vectors, convert the numerical features and hierarchical features into numerical scalars, and store the semantic vectors and numerical scalars as feature values ​​in the vector database;

[0008] Specifically, based on the finely tuned pre-trained semantic extraction model, text features are transformed into semantic vectors. Based on the statistical analysis of resume / job posting texts, numerical and graded features are normalized into numerical scalars between 0 and 1. The semantic vectors and numerical scalars are stored as feature values ​​in the vector database.

[0009] S3, customize the browsing history hyperparameter N, take the applicant's most recent N browsing records on the client, search for the feature values ​​corresponding to the browsing records in the vector database, and obtain the applicant's interest profile and the interest degree corresponding to each feature;

[0010] S4. For each candidate job recommendation, the matching degree is obtained based on the similarity between each feature value of the job and the interest profile of the corresponding feature, and the interest degree of the feature.

[0011] S5. Sort the matching degree of each candidate option obtained in step S4 to obtain the recommendation result.

[0012] A second aspect of the present invention provides an intelligent human-job matching device based on feature discrete coefficients and an attention mechanism, comprising one or more processors for the aforementioned intelligent human-job matching method based on feature discrete coefficients and an attention mechanism.

[0013] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon that, when executed by a processor, is used to implement the above-described intelligent human-job matching method based on feature discrete coefficients and attention mechanisms.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0015] (1) This invention uses a pre-trained and fine-tuned sentence_transformer to perform vector transformation on text features, which can greatly increase the accuracy of text feature vector representation. For numerical features and graded features, the value range is analyzed and reasonably transformed, so that each feature can be accurately and reasonably represented without training with samples.

[0016] (2) When calculating the feature update vector, a multi-dimensional attention mechanism including interest, duration and similarity is introduced. The user's interest attention score for a record is extracted from the duration of the user's browsing of a record, and used as the interest attention score for all features under that record. The memory attention score for a record is extracted from the time since the user finished browsing a record, and used as the memory attention score for all features under that record. The feature similarity attention score is extracted from the similarity between the user's browsing record features and the feature profile.

[0017] (3) The comprehensive attention score is calculated by the root mean square of the attention scores of different subclasses, so that more attention is paid to feature records with very high individual attention scores when generating new portraits.

[0018] (4) When calculating the matching degree between candidate resumes / jobs and target job / resumes, use the discrete coefficient to generate feature matching weights (i.e. feature interest) to ensure that more weight is given to features that users are interested in (showing relatively stable similarity).

[0019] (5) This invention can ensure high recommendation accuracy without collecting users’ original data for training, thus ensuring users’ data security and privacy. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of the method of the present invention.

[0022] Figure 2 This is a schematic diagram of the device of the present invention. Detailed Implementation

[0023] The present invention will now be described in detail with reference to the accompanying drawings. Unless otherwise specified, the features of the following embodiments and implementations can be combined with each other.

[0024] like Figure 1As shown, this invention proposes an intelligent person-job matching method based on feature discrete coefficients and an attention mechanism. The method specifically includes the following sub-steps:

[0025] S1. Obtain recruitment documents (i.e., resume / job posting text corpus), generate similar sentence pairs through manual annotation, mark them as positive examples, shuffle the similar sentence pairs as negative examples, and then input them as positive and negative examples into the pre-trained semantic extraction model for model fine-tuning.

[0026] In step S1 above, recruitment-related documents can be obtained through online collection or self-generation. Collection of recruitment-related documents is preferred; if recruitment-related documents are insufficient, they can be supplemented by collection from various professional fields. After completing the document collection, the most important step is to annotate the recruitment-related documents with sentence pairs. This annotation method involves selecting pairs of semantically similar sentences / paragraphs from the corpus and placing them in the same list as a training corpus. To reduce the workload of manual selection, the corpus can be first divided into sentences or paragraphs, then converted into sentence vectors. The most semantically similar sentences are then found from the corpus by calculating vector similarity. Finally, the corpus is manually reviewed to obtain the resume / job posting text corpus. In this example, the pre-trained semantic extraction model uses the sentence-transformer model.

[0027] S2, based on the description type of each field, divide the resume / job posting text corpus into text features, numerical features, and hierarchical features; and analyze the resume / job posting text corpus according to different feature types.

[0028] Specifically, the text features are converted into sentence vectors using the pre-trained semantic extraction model fine-tuned in step S1.

[0029] For numerical features, the data range of the numerical features is first inferred, and then the maximum and minimum values ​​are normalized to convert them into numerical scalars between 0 and 1.

[0030] For example, numerical features are easy to identify. If their value range is clear, they can be directly normalized using the minimum-maximum normalization method. If the value range is unclear, the range needs to be determined by expert experience before minimum-maximum normalization can be performed.

[0031] For hierarchical features, based on the number of levels, they are normalized to their maximum and minimum values ​​in the form of an arithmetic sequence, and then converted into numerical scalars between 0 and 1.

[0032] For example, a graded feature refers to a feature whose attributes are limited, easily exhaustive, and have varying degrees of importance, such as education level, professional title level, salary level, etc. For graded features, they can be simply labeled from 1 to N (N is the number of attributes) in ascending order, and then the labeling results are normalized by dividing by N.

[0033] For each resume / job posting, the different feature transformation results are stored in the vector database ES.

[0034] This invention employs a pre-trained and fine-tuned sentence_transformer for text feature vector transformation, which can significantly increase the accuracy of text feature vector representation. For numerical and rank features, analysis and reasonable transformation are performed based on different value ranges and the number of ranks, enabling each feature to obtain an accurate and reasonable representation without requiring sample training.

[0035] S3, customize the browsing history hyperparameter N, take the applicant's most recent N browsing records on the client, find the feature value Vrf corresponding to each feature in the vector database ES, and obtain the applicant's interest profile Ff under each feature based on the attention mechanism.

[0036] This invention introduces a multi-dimensional attention mechanism, including interest, duration, and similarity, when calculating the feature update vector. It extracts the user's interest attention score for a record from the duration of the user's browsing of that record, and uses it as the interest attention score for all features under that record. It also extracts the memory attention score for a record from the time since the user finished browsing it, and uses it as the memory attention score for all features under that record. Finally, it extracts the feature similarity attention score from the similarity between each feature of the user's browsing record and the feature profile.

[0037] The specific calculation process for the interest profile Ff is as follows:

[0038] S31. For each browsing record, the difference between its browsing start time and browsing end time is normalized using softmax, and then used as the interest and attention score Wr_i for each browsing record;

[0039] S32, for each browsing record, the reciprocal of the difference between its browsing end time and the current time is normalized by softmax, and then used as the memory attention score Wr_m for each browsing record;

[0040] S33, calculate the average value of the same features in different browsing records. For each browsing record, calculate the similarity Srf between each feature and the corresponding average value of the features.

[0041] S34, For each feature, calculate the coefficient of variation Df of its similarity Srf under different browsing records;

[0042] S35, For each feature, use its discrete coefficient Df to calculate its updated weight Uf using the negative natural exponent;

[0043] S36. Softmax normalization is performed on the update weights Uf of different features to obtain the interest level If of the feature. The feature interest level is generated using discrete coefficients to ensure that more weight is given to features that users are interested in (showing relatively stable similarity).

[0044] S37. For each feature, the similarity Srf calculated in step S33 under different records is calculated using softmax normalization, and the result is used as the similarity attention score Wrf_s for each feature under different records.

[0045] S38. For each feature in each record, calculate the root mean square average of the similarity attention score Wrf_s, the memory attention score Wr_m, and the interest attention score Wr_i as the comprehensive attention score Wrf.

[0046] S39, for each feature, calculate and update the interest profile Ff' of that feature based on its feature value Vrf corresponding to each feature in different browsing records, its comprehensive attention score Wrf in that record, and its update weight Uf. The method of this invention calculates the comprehensive attention score by using the root mean square of the attention scores of different subclasses, thus giving more attention to feature records with very high individual attention scores when generating a new profile.

[0047] The formula for calculating the updated interest profile Ff' is as follows:

[0048] Ff'=Ff×(1-Uf)+Vrf×Wrf×Uf.

[0049] S4. For each job candidate recommendation, calculate the matching degree M between the candidate recommendation and the comprehensive profile based on the similarity Sf between each feature value Vif of the job and the interest profile Ff of the corresponding feature, and then calculate the matching degree M between the candidate recommendation and the comprehensive profile based on the interest degree If of the feature.

[0050] In this example, cosine similarity is used to calculate the similarity Sf between each feature value Vif of the job and the interest profile Ff of the corresponding feature.

[0051] The matching degree M is calculated by weighted summation of feature similarity Sf and feature interest If.

[0052] S5. Sort the matching degree M values ​​of each candidate option obtained in step S4 to obtain the recommendation results.

[0053] In particular, the method of the present invention can ensure high recommendation accuracy without collecting a large amount of users' original resume data for deep learning training, thus ensuring users' data security and privacy.

[0054] Example 1

[0055] Retrieve resumes to be matched and process the retrieved resumes.

[0056] A personal resume includes the following fields: date of birth, city, gender, honors and awards (Note a), skills and languages ​​(Note a), highest degree, school name (Note a), major name (Note a), major course name (Note a), company name (Note a), company industry (Note a), job title (Note a), job description (Note a), project name (Note a), position held (Note a), project description (Note a).

[0057] The job description includes: job type, job title, industry sector, salary range, professional background, work area, and job description.

[0058] The birth date in the resume can be converted into age and then the data range of numerical features can be inferred. In this example, the data range of age is inferred to be 16-70, and then minimum-maximum normalization is performed.

[0059] The highest degree in the resume can be mapped to 1, 2, 3, 4, 5, and 6 respectively for below associate degree, associate degree, bachelor's degree, master's degree, doctorate, and postdoctoral degree, and then normalized to 0, 0.2, 0.4, 0.6, 0.8, and 1.0.

[0060] The salary ranges for the positions are “0-2k”, “2-5k”, “5-10k”, “10-20k”, “20-50k”, and “above 50k”, which can be mapped to 1, 2, 3, 4, 5, and 6 respectively, and then normalized to 0, 0.2, 0.4, 0.6, 0.8, and 1.0.

[0061] Except for the date of birth and highest degree in the resume, the other fields can be transformed into semantic vectors using a finely tuned sentence-transformer model.

[0062] For the field labeled 'a', each field can store multiple values. The final vector result of this field is to use the sentence-transformer model to calculate the semantic vector of each value, and then average the semantic vectors of different values ​​to obtain the transformation result.

[0063] In this implementation example, the processing results for each resume will be stored in the vector database ES.

[0064] In this example, a resume / job browsing history of a recruiter / job seeker is R, where Ri is the i-th browsing record; a complete set of resume / job features is F, where Fj is the j-th feature, and the feature value of the j-th feature Fj in the i-th browsing record Ri is Vij; a complete talent / job interest profile is P, and the feature value of the corresponding feature Fj in the interest profile P is Vj.

[0065] A complete browsing history record Ri contains the following information: start browsing time Ti_s, end browsing time Ti_e, and feature value Vij corresponding to feature Fj.

[0066] In this implementation example, the initial value of feature Vj comes from the conversion result of the field description corresponding to the job posting of the recruiter / job seeker.

[0067] Calculate the difference Ti_se between the start browsing time Ti_s and the end browsing time Ti_e; perform softmax normalization on Ti_se to obtain the browsing record interest and attention score Wi_i (Note: In this example, the first i in Wi_i is short for interest, representing the degree of interest).

[0068] Calculate the difference Ti_ce between the current time (image update time) Tc and the end browsing time Ti_e; calculate the browsing history memory attention score Wm_i by performing softmax normalization on the reciprocal of Ti_ce.

[0069] For the same features in different browsing records, calculate their average value. For each browsing record, calculate the similarity Sij between each feature and the corresponding average value.

[0070] The similarity Sij under browsing records Ri is normalized by softmax to obtain the similarity attention score Ws_ij for each feature in the browsing records.

[0071] The root mean square average of the similarity attention score Ws_ij, the memory attention score Wm_i, and the interest attention score Wi_i is calculated as the comprehensive attention score Wrf.

[0072] For feature Fj, calculate the weighted sum of the feature value Vij with W_ij to obtain the component Vj_u to be updated for feature Fj. Calculate the discrete coefficients of the similarity attention scores Ws_ij in different Ri for feature Fj, denoted as Dj. Calculate the update weights Uj for each feature by applying the negative natural exponent to the discrete coefficients Dj. For feature Fj, calculate the updated feature value Vj according to the formula Vj=Vj*(1-Uj)+Vj_u*Uj. Apply softmax normalization to the update weights Uj to obtain the interest level of each feature as feature Ij.

[0073] The set of candidate resumes / jobs is S, and Sk is the Kth candidate resume / job; a complete set of resume / job features is F, and Fj is its jth feature, with the feature value of feature Fj in Sk being Vkj; the profile of the recruiter / job seeker's talent / job requirements is P, with the feature value of feature Fj in profile P being Vj.

[0074] For the Kth candidate resume / job post Sk, calculate the similarity Skj between Vkj and Vj; calculate the weighted sum of Skj and Ij for the Kth candidate resume / job post Sk to obtain the matching degree Mk between the Kth candidate resume / job post Sk and job post P. Sort Mk in descending order as the recommendation result for candidate resumes / job posts S.

[0075] Corresponding to the aforementioned embodiments of the intelligent human-job matching method based on feature discrete coefficients and attention mechanisms, the present invention also provides embodiments of an intelligent human-job matching device based on feature discrete coefficients and attention mechanisms.

[0076] See Figure 2 The present invention provides an intelligent human-job matching device based on feature discrete coefficients and attention mechanism, comprising one or more processors for implementing the intelligent human-job matching method based on feature discrete coefficients and attention mechanism in the above embodiments.

[0077] The embodiment of the intelligent human-job matching device based on feature discrete coefficients and attention mechanism of the present invention can be applied to any device with data processing capabilities, such as a computer. The device embodiment can be implemented in software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, such as... Figure 2 The diagram shown is a hardware structure diagram of any data processing-capable device in which the intelligent human-job matching device based on feature discrete coefficients and attention mechanism of the present invention is located. (Except for...) Figure 2 In addition to the processor, memory, network interface, and non-volatile memory shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0078] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0079] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0080] This invention also provides a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements the intelligent human-job matching method based on feature discrete coefficients and attention mechanism described in the above embodiments.

[0081] The computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0082] The above embodiments are only used to illustrate the design concept and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, all equivalent changes or modifications made based on the principles and design ideas disclosed in the present invention are within the protection scope of the present invention.

Claims

1. An intelligent person-job matching method based on feature discrete coefficients and attention mechanism, characterized in that, The method includes the following sub-steps: S1. Obtain resume / job posting text corpus, generate similar sentence pairs through manual annotation, mark them as positive examples, then shuffle the similar sentence pairs as negative examples, and then input them as positive and negative examples into the pre-trained semantic extraction model for model fine-tuning; S2, based on the description type of each field, divide the resume / job text corpus into text features, numerical features, and hierarchical features; convert the text features into semantic vectors, convert the numerical features and hierarchical features into numerical scalars, and store the semantic vectors and numerical scalars as feature values ​​in the vector database; S3, customize the browsing history hyperparameter N, take the applicant's most recent N browsing records on the client, search for the feature values ​​corresponding to the browsing records in the vector database, and obtain the applicant's interest profile and the interest degree corresponding to each feature; S4. For each candidate job recommendation, the matching degree is obtained based on the similarity between each feature value of the candidate job and the interest profile of the corresponding feature, and the interest degree of the feature. S5. Sort the matching degree of each candidate obtained in step S4 to obtain the recommendation result; Specifically, step S3 includes the following sub-steps: S31. For each browsing record, the difference between its browsing start time and browsing end time is normalized using softmax, and then used as the interest and attention score Wr_i for each browsing record; S32, for each browsing record, the reciprocal of the difference between its browsing end time and the current time is normalized by softmax, and then used as the memory attention score Wr_m for each browsing record; S33, calculate the average value of the same features in different browsing records. For each browsing record, calculate the similarity Srf between each feature and the corresponding average value of the features. S34, For each feature, calculate the coefficient of variation Df of its similarity Srf under different browsing records; S35, For each feature, calculate its updated weight Uf using its discrete coefficients Df; S36, perform softmax normalization calculation on the update weights Uf of different features to obtain the interest level If of the feature; S37. For each feature, the similarity Srf calculated in step S33 under different browsing records is calculated using softmax normalization, and the result is used as the similarity attention score Wrf_s for each feature under different browsing records. S38, For each feature in each browsing record, obtain the comprehensive attention score Wrf based on the similarity attention score Wrf_s, the memory attention score Wr_m and the interest attention score Wr_i of the browsing record; S39, for each feature, calculate and update the interest profile Ff of that feature based on its feature value Vrf corresponding to each feature in different browsing records, its comprehensive attention score Wrf in that record, and its update weight Uf. The updated interest profile is denoted as Ff'. The calculation formula for the updated interest profile Ff' is as follows: Ff'= Ff × (1-Uf) + Vrf × Wrf × Uf.

2. The intelligent person-job matching method based on feature discrete coefficients and attention mechanism according to claim 1, characterized in that, The process of extracting features from resume / job posting text corpora based on fine-tuned pre-trained semantic extraction models and feature type analysis includes: For text features, the pre-trained semantic extraction model finely tuned in step S1 is used to convert the text features into semantic vectors. For numerical features, first infer the data range of the numerical features, and then perform maximum and minimum value normalization processing; For hierarchical features, based on the different numbers of levels, they are normalized by maxima and minima in the form of an arithmetic sequence. The extracted features are stored in the vector database ES.

3. The intelligent person-job matching method based on feature discrete coefficients and attention mechanism according to claim 1, characterized in that, The method for calculating the updated weight Uf using its coefficient of variation Df is as follows: calculate the updated weight Uf using the negative natural exponent with its coefficient of variation Df.

4. The intelligent person-job matching method based on feature discrete coefficients and attention mechanism according to claim 1, characterized in that, The process of obtaining the comprehensive attention score Wrf based on the similar attention score Wrf_s, the memory attention score Wr_m, and the interest attention score Wr_i of the record includes: calculating the root mean square average of the similar attention score Wrf_s, the memory attention score Wr_m, and the interest attention score Wr_i of the record as the comprehensive attention score Wrf.

5. The intelligent person-job matching method based on feature discrete coefficients and attention mechanism according to claim 1, characterized in that, The matching degree is calculated by weighting and summing the similarity between each feature value of the job and the interest profile of the corresponding feature, and the interest degree of the feature.

6. An intelligent human-job matching device based on feature discrete coefficients and an attention mechanism, characterized in that, It includes one or more processors for implementing the intelligent human-job matching method based on feature discrete coefficients and attention mechanism as described in any one of claims 1-5.

7. A computer-readable storage medium having a program stored thereon, characterized in that, When executed by the processor, the program is used to implement the intelligent human-job matching method based on feature discrete coefficients and attention mechanism as described in any one of claims 1-5.

Images