Training method of semantic analysis model for recalled positions and resumes

Abstract

The invention discloses a training method of a semantic analysis model for recalled positions and resumes. The training method of the semantic analysis model used for recalled positions and resumes comprises the steps that delivery behavior data, user resume data and position description data of a user are obtained, and the delivery behavior data comprise positions where the user delivers the resumes; based on the positions of the resumes delivered by the user, constructing matching samples of the positions of the resumes; based on a pre-configured language analysis model, identifying the delivery behavior data, the user resume data and the position description data corresponding to the matching samples, and screening out the matching samples corresponding to the user resume data which does not meet the position constraint requirements to obtain positive matching samples; in the samples which do not meet the position constraint requirements, negative matching samples are extracted; and performing fine tuning training on a pre-constructed semantic analysis model based on the positive matching sample and the negative matching sample to generate a semantic analysis model of the position and the resume.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method for training a semantic analysis model for recalling job postings and resumes. Background Technology

[0002] In today's recruitment industry, job recommendation systems have become crucial tools for job seekers to find suitable positions and for companies to quickly acquire talent. Recommendation systems typically operate through two phases: recall and ranking. The recall phase involves filtering a batch of candidate positions relevant to the user from a massive amount of job information. For example, user-job interaction data, such as browsing, clicking, saving, and submitting applications, is used to build user and job profiles. Recall and ranking mechanisms then recommend positions that the user might be interested in. However, in practical applications, newly posted jobs or resumes often lack sufficient historical interaction data, preventing the system from effectively using collaborative filtering or behavioral modeling methods to model them. In other words, existing recall methods still suffer from the cold start problem.

[0003] Due to a lack of historical interaction data, traditional recommendation systems struggle to identify suitable candidate groups. Consequently, during the recall phase, they fail to accurately push new job postings to potentially suitable users or new resumes to suitable positions. This delays the recruitment process for employers and slows down the job search for job seekers. Therefore, improving recall quality is a pressing issue that needs to be addressed in the absence of interaction data. Summary of the Invention

[0004] In view of this, embodiments of this application provide a method for training a semantic analysis model for recalling job titles and resumes, in order to solve at least one technical problem.

[0005] This application provides a training method for a semantic analysis model of job postings and resumes for recall, comprising: acquiring user application behavior data, user resume data, and job description data, wherein the application behavior data includes the job postings submitted by the user; constructing matching samples of resumes and job postings based on the job postings submitted by the user; identifying the application behavior data, user resume data, and job description data corresponding to the matching samples based on a pre-configured language analysis model, filtering out matching samples corresponding to user resume data that do not meet the job constraints, and obtaining positive matching samples; and extracting negative matching samples from the samples that do not meet the job constraints; and fine-tuning the pre-constructed semantic analysis model based on the positive and negative matching samples to generate a semantic analysis model of job postings and resumes.

[0006] This application provides a recall method based on a semantic analysis model of job titles and resumes, comprising: obtaining job information corresponding to newly published job titles; extracting skill requirements, industry experience requirements, and job responsibility requirements from the job information based on the semantic analysis model of job titles and resumes, and converting them into job semantic vectors, wherein the semantic analysis model of job titles and resumes is generated based on the training method of the semantic analysis model of job titles and resumes for recall provided in this application; based on the job semantic vectors, obtaining a first preset number of similar jobs to newly published job titles from a pre-built historical job database, and obtaining the received historical resumes of similar jobs to obtain a historical resume set; obtaining user resume data for each historical resume in the historical resume set; extracting project experience, work experience, and user skills from the user resume data of each historical resume based on the semantic analysis model of job titles and resumes, and converting them into semantic vectors of historical resumes; calculating the similarity between the job semantic vector and the semantic vector of each historical resume to obtain a matching score between each historical resume and the newly published job title; and based on the matching score between each historical resume and the newly published job title, taking the users corresponding to a second preset number of historical resumes with the highest scores as the user recall results.

[0007] This application provides an electronic device, which includes a processor and a memory storing computer program instructions; the processor executes the computer program instructions to implement the steps of the method described above.

[0008] This application provides a computer-readable storage medium storing computer program instructions, which, when executed by a processor, implement the steps of the method described above.

[0009] This application provides a computer program product, which includes computer program instructions. When the computer program instructions are executed by a processor, they implement the steps of the method described above.

[0010] The embodiments of this application construct positive and negative matching samples that better suit recruitment scenarios by acquiring user application behavior data, resume data, and job description data. On the one hand, by performing language analysis and job constraint screening on the application behavior samples, low-quality data such as random and malicious applications can be effectively eliminated, thereby ensuring the accuracy of positive matching samples. On the other hand, by introducing samples that do not meet the job constraint requirements as negative matching samples into the training, the model can better distinguish between matching and non-matching situations, enhancing the model's discriminative ability. Finally, the semantic analysis model of jobs and resumes, fine-tuned by the positive and negative matching samples, can not only more accurately understand the deep semantic relationship between user resumes and job descriptions, but also convert newly published jobs or resumes into accurate semantic vectors, thereby improving the quality of recall results in the recall stage. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings of the embodiments of this application will be briefly described below.

[0012] Figure 1 This is a schematic diagram of the system architecture of an embodiment of this application.

[0013] Figure 2 This is a flowchart illustrating a method for training a semantic analysis model for recalling job titles and resumes, as described in an embodiment of this application.

[0014] Figure 3 This is a flowchart of the recall method based on a semantic analysis model of job title and resume according to an embodiment of this application.

[0015] Figure 4 Another flowchart of the recall method based on the semantic analysis model of job title and resume in this application embodiment.

[0016] Figure 5 This is a schematic diagram of an electronic device used to implement the embodiments of this application. Detailed Implementation

[0017] The features and exemplary embodiments of various aspects of this application will be described in detail below. To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only intended to explain this application and not to limit it. For those skilled in the art, this application can be implemented without some of these specific details. The following description of the embodiments is merely to provide a better understanding of this application by illustrating examples.

[0018] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0019] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0020] Various modifications and variations can be made to this application without departing from its spirit or scope, which will be apparent to those skilled in the art. Therefore, this application is intended to cover modifications and variations falling within the scope of the corresponding claims (the claimed technical solutions) and their equivalents. It should be noted that the implementation methods provided in the embodiments of this application can be combined with each other without contradiction.

[0021] Embodiments of this application relate to terminal devices and / or servers. Those skilled in the art will recognize that embodiments of this application can be implemented as a system, apparatus, device, method, computer-readable storage medium, or computer program product. Therefore, this disclosure can be specifically implemented in at least one of the following forms: entirely hardware, entirely software, or a combination of hardware and software.

[0022] It should be noted that the acquisition, storage, use, and processing of data in this application embodiment all comply with the relevant provisions of laws and regulations.

[0023] Figure 1 A schematic diagram of a system architecture according to an embodiment of this application is shown. For example... Figure 1 As shown, the system includes terminal devices 102a and 102b and a server 104. The terminal devices can include at least one of the following: smartphones, tablets, laptops, desktop computers, smart TVs, various wearable devices, augmented reality (AR) devices, virtual reality (VR) devices, etc. Clients can be installed on the terminal devices. For example, the client can be a dedicated client for performing specific functions (such as an application app), or a client with embedded multiple application applets (with different functions), or a client that logs in via a browser. Users can perform operations on the terminal devices. For example, a user can open the client installed on the terminal device to access the recruitment platform, or a user can open the browser installed on the terminal device to access the recruitment platform. The recruitment platform can be a job search website accessed by job seekers or a recruitment website accessed by recruiters.

[0024] For example, when a job seeker accesses the job search website homepage via terminal device 102a, the recruitment platform can retrieve several job postings to display on the homepage. When a recruiter accesses the recruitment website homepage via terminal device 102b, the recruitment platform can retrieve several resumes to display on the homepage.

[0025] Figure 2 The flowchart illustrates a method for training a semantic analysis model for recalling job positions and resumes according to an embodiment of this application. The method includes steps 201 to 204.

[0026] Step 201: Obtain user application behavior data, user resume data, and job description data. The application behavior data includes the job positions that the user applied for.

[0027] Step 202: Based on the job postings submitted by users, construct a matching sample of resumes and job postings.

[0028] Step 203: Based on the pre-configured language analysis model, identify the application behavior data, user resume data and job description data corresponding to the matching samples respectively, filter out the matching samples corresponding to user resume data that do not meet the job constraints, and obtain positive matching samples; and extract negative matching samples from the samples that do not meet the job constraints.

[0029] Step 204: Fine-tune the pre-built semantic analysis model based on positive and negative matching samples to generate semantic analysis models for job postings and resumes.

[0030] The above steps will be described in detail below with reference to specific embodiments, as shown in the following figures.

[0031] First, step 201 is involved, where application behavior data can directly reflect a user's job-seeking intentions.

[0032] User resume data can include hard skills such as programming languages ​​and database skills, soft skills such as communication skills and teamwork skills, educational background, project experience, and professional certifications.

[0033] Job description data can include elements such as job responsibilities, required skills, preferred qualifications, educational and work experience requirements, and geographical restrictions. During the data acquisition phase, resumes and job texts can be structured, for example, by splitting resumes into skill fields, project fields, and education fields, and job descriptions into skill requirements, job responsibilities, and educational requirements fields, to facilitate subsequent semantic model recognition.

[0034] In step 202, each application can be considered a candidate "resume-job" matching pair. However, users may engage in mass applications, random applications, or malicious applications, resulting in some application samples not accurately reflecting the user's matching intent. Therefore, it is necessary to further utilize a pre-configured language analysis model to identify the data corresponding to the matching samples. The language analysis model can be a pre-trained large language model or an industry-specific semantic model, which can be used to determine the degree of matching between the resume and the job description in terms of constraints such as skills, experience, and education. Through model identification, samples that do not meet the job constraints can be filtered out. For example, the job requires "proficiency in Python," but the user's resume does not reflect this skill; the job requires "Master's degree or above," but the user's resume only shows an associate's degree. The filtered samples will no longer be used as positive matching samples in training. After the filtering process, the resulting matching samples are the positive matching samples.

[0035] Furthermore, during the sample construction process, it is also necessary to extract negative matching samples. Negative matching samples can be unreasonable submission samples under pre-defined user behavior characteristics. For example, a user submitting resumes to multiple different positions in a very short period of time can be regarded as "mass application," and the corresponding matching samples are used as negative matching samples. Another type is samples that do not meet the job constraints, namely, "resume-job" combinations that the language analysis model determines are not a match. These negative matching samples can help the model learn "which situations should not be considered a match" during training, thereby improving the model's discrimination ability.

[0036] Finally, the pre-built semantic analysis model is fine-tuned based on positive and negative matching samples. The semantic analysis model can be a large-scale pre-trained semantic matching model, such as BGE-M3 or RoBERTa. Through fine-tuning, the model can learn a more accurate semantic mapping relationship between job descriptions and user resumes, enabling it to quickly select a batch of candidate jobs or resumes most relevant to the user during the recall phase, serving as input for the subsequent ranking phase.

[0037] In another alternative implementation, the sample screening and model fine-tuning stages can be adjusted according to different business needs. For example, when screening matching samples, a multi-level filtering strategy can be adopted: the first layer uses rules to filter obviously unreasonable samples, such as those with completely mismatched education, years of experience, or location; the second layer uses a large language model to perform semantic-level deep judgment, thereby ensuring and improving the quality of positive matching samples. Regarding the construction of negative matching samples, in addition to mass-applied samples and samples that do not meet constraints, "difficult-to-match samples" can be introduced, such as samples that are highly similar in skills but differ in key responsibilities or soft skill requirements. These difficult-to-match samples can improve the model's ability to distinguish between similar job positions.

[0038] In addition, during model fine-tuning, a multi-task learning approach can be adopted. For example, in addition to the similarity discrimination training task, there are also job category prediction or skill set prediction training tasks, so that the model can learn semantic matching while also mastering the ability to distinguish job categories and skill distributions.

[0039] Optionally, different training methods can be selected for datasets of different sizes: when the amount of data is large, a contrastive learning method can be used; when the amount of data is small, a distillation strategy can be adopted to transfer the discrimination results of the large model to a lightweight model so that it can run with low latency in actual recommendation.

[0040] The above implementation methods not only filter out low-quality data such as mass applications and random submissions, improving the accuracy of training samples, but also enhance the model's ability to recognize complex job constraints and semantic differences, thereby improving the model's discriminative ability in recruitment scenarios. The final semantic analysis model for job postings and resumes can more accurately understand the deep semantic relationship between user resumes and job descriptions, solving the problem of poor recall performance in cold-start scenarios, thus improving the relevance of recommended content and user experience.

[0041] In some embodiments of this application, the application behavior data includes embedded data within a first associated time window corresponding to the resume application time. The first associated time window includes a first preset time period before the resume application time and a second preset time period after the resume application time. Based on the job position submitted by the user, a matching sample of resume and job position is constructed, including: extracting the length of time the user browses the job position and the number of application behaviors within the predetermined time window from the embedded data within the associated time period of each resume application time; filtering out application behavior data when the length of time is less than a preset time threshold and the number of application behaviors within the first associated time window is greater than a first preset number; and constructing a matching sample of resume and job position based on the resumes included in the filtered application behavior data.

[0042] Specifically, delivery behavior data includes not only the user's delivery action itself, but also the tracking data related to that delivery action.

[0043] For example, when a user submits a resume to a job on a recruitment platform, the time of submission and the data points within the first time window associated with that time are recorded. Based on the data points, the user's interactive behavior before and after submission, such as browsing, dwelling, and clicking, can be reflected, thus providing more accurate user behavior characteristics for sample construction.

[0044] Furthermore, in terms of acquiring event tracking data, it's possible to record the depth of user dwell time on job detail pages, the number of times they click on different fields in the job description, and even whether they viewed information such as company profile and benefits. This more detailed event tracking data can further enrich user behavior characteristics, making the filtering results more accurate.

[0045] In the sample construction phase, based on the job postings submitted by users, for each resume submission time, the length of time a user spends browsing the job postings can be extracted, as well as the number of times a user submits a resume within that time window.

[0046] If the browsing time is less than a preset time threshold, and the number of submissions within that time window exceeds a first preset number, then the submission behavior is considered to be low-quality behavior, such as random or mass submissions. Therefore, such submission behavior data will be filtered out and will not be used as valid training samples for subsequent modeling.

[0047] After the above filtering process, matching samples of resumes and job descriptions are constructed based on the remaining user resume data and job description data. Compared with the unfiltered data, this yields positive matching samples that more accurately reflect the matching relationship between users and jobs, providing a reliable data foundation for training the semantic model.

[0048] In another alternative implementation, the time threshold and the number of times threshold can be flexibly configured according to the actual business scenario.

[0049] For example, the browsing time threshold can be set to 10 seconds, 20 seconds, or 30 seconds. The number of times threshold can be adjusted according to the activity level of the user group and the traffic characteristics of the recruitment platform, for example, set to 3 times, 5 times, or more, in order to more accurately distinguish between normal application behavior and mass application behavior.

[0050] The above implementation methods can eliminate low-quality submissions at the submission behavior data level, thereby improving the accuracy and reliability of training samples. The resume-job matching samples obtained through screening can more realistically reflect the user's actual job-seeking intentions, ensuring that the semantic model learns higher-quality matching relationships during training. Ultimately, the semantic analysis model of jobs and resumes trained can effectively improve the accuracy of job recommendations during the recall stage, avoiding interference caused by low-quality samples.

[0051] In some embodiments of this application, constructing a matching sample of resumes and positions based on the positions submitted by users further includes: for each position for which users submit resumes, obtaining the tracking data within a second associated time window corresponding to the time when the resume is received, the second associated time window including a first preset time period before the time when the resume is received and a second preset time period after the time when the resume is received; determining the number of resumes received based on the tracking data within the second associated time window; and filtering out the submission behavior data corresponding to the position if the number of resumes received within the second associated time window is greater than a second preset number.

[0052] Specifically, the moment a resume is received for a job corresponds to a second associated time window. By collecting the data points within the second time window, the application status of the job in a short period of time can be reflected.

[0053] Based on the tracking data within the second associated time window, the number of resumes received for that position during that time period can be counted. If the number of resumes received within this time window exceeds a preset threshold, it is determined that the application behavior for that position during that period is abnormal, such as a concentrated influx of resumes due to external promotion, fraudulent activity, or fake applications. In this case, the application behavior data corresponding to that position will be filtered out and no longer included in the construction of the matching sample. This avoids affecting the overall quality of the training sample due to an abnormally high number of resumes received by a single position in a short period of time.

[0054] In another alternative implementation, the specific threshold for the second preset number of times can be flexibly set according to the business scenario and historical statistical results.

[0055] Optionally, this threshold can be tiered for different job categories. For example, the threshold for popular positions or positions in large companies can be appropriately relaxed, while the threshold for niche positions or highly specialized positions can be set more strictly. Furthermore, in addition to counting the number of resumes submitted, a comprehensive judgment can be made by combining information such as resume source channels, user group characteristics, and submission time distribution. For example, if the number of resumes from the same channel increases abnormally within a very short period, it can be considered an abnormal submission and the corresponding data can be filtered out.

[0056] By implementing the above methods, low-quality applications resulting from a large influx of resumes for a job posting within a short period are identified and eliminated, thereby improving the authenticity and representativeness of the sample data. The filtered samples more accurately reflect users' true application intentions and the actual matching situation of the job postings, providing a high-quality data foundation for subsequent fine-tuning and training of the semantic model, thus improving the accuracy of recommendation recall.

[0057] In some embodiments of this application, based on a pre-configured language analysis model, the application behavior data, user resume data, and job description data corresponding to the matching samples are identified, and user resume data that does not meet the job constraint requirements are filtered out. This includes: identifying user resume data and job description data based on a pre-configured language analysis model; extracting project experience, work experience, and user skills from user resume data; and extracting industry experience requirements, responsibility requirements, and skill requirements from job description data; generating job constraint requirements for each job description data including industry experience requirements, responsibility requirements, and / or skill requirements; and comparing the resume data and job constraint requirements for each matching sample to determine user resume data that does not meet the job constraint requirements.

[0058] Specifically, by invoking a pre-configured language analysis model, semantic parsing can be performed on user resume data and job description data. From user resume data, the model can identify and extract information such as project experience, work experience, and user skills; from job description data, the model can identify and extract information such as industry experience requirements, job responsibilities, and skill requirements. Thus, core semantic elements for subsequent matching and determination can be obtained.

[0059] Furthermore, the industry experience requirements, job responsibilities, and / or skill requirements identified in each job description are transformed into structured job constraints. These constraints typically include a set of essential skills, minimum educational or work experience requirements, and mandatory industry experience, serving as a benchmark for determining whether a candidate resume meets the criteria. Subsequently, in each constructed resume-job matching sample, the resume data is compared with the job constraints. For example, if the job constraints include "more than three years of experience in the internet industry" and "proficiency in Python programming," and the candidate resume does not demonstrate relevant industry experience or skills, then the resume is determined not to meet the job constraints and is removed from the matching sample. This semantic understanding-based comparison method reduces misjudgments caused by relying solely on keyword matching.

[0060] In another alternative implementation, the language analysis model can employ a pre-trained large-scale language model and be fine-tuned by incorporating specialized corpora from the recruitment field to make it more accurate in identifying entities such as job responsibilities, industry experience, and skill names.

[0061] The methods for generating job constraints include, for example, extracting elements through a language analysis model and setting rules such as mandatory conditions and priority conditions to determine their attribute categories.

[0062] Through the above implementation methods, a language analysis model is used to perform structured recognition and semantic constraint comparison between resumes and job descriptions, accurately filtering out low-quality samples that do not meet the job constraint requirements. In this way, the retained matching samples can more realistically reflect the matching relationship between resumes and jobs, providing high-quality data support for subsequent fine-tuning of the semantic model, thereby effectively improving the accuracy of recommendation recall.

[0063] In some embodiments of this application, the method further includes a step of determining negative matching samples to ensure that semantic features of both matching and non-matching functions can be learned simultaneously during the training phase. The method further includes determining negative matching samples: extracting the user's expected job function from user resume data based on a pre-configured language analysis model, and extracting the actual job function from the job postings submitted by the user; performing a matching analysis on the user's expected job function and the actual job function based on the pre-configured language analysis model, and marking matching samples where the actual job function does not meet the user's expected job function as negative matching samples.

[0064] Specifically, a pre-configured language analysis model is used to parse user resume data and extract the desired job functions that users explicitly or implicitly express in their resumes. For example, if a user marks "Desired position: Front-end development engineer" in the "Job objective" section of their resume, or repeatedly emphasizes "algorithm research" when describing their project experience, the user's desired job function can be identified.

[0065] At the same time, a language analysis model was used to parse the job descriptions corresponding to the resumes submitted by users, and the actual functions of the job were extracted from the job text.

[0066] The actual job functions can be determined by the job category, job tags, or frequently appearing keywords in the job responsibilities in the recruitment information, such as "backend development", "marketing", "human resource management", etc. In this way, the set of functions expected by users and the actual set of functions of the applied positions can be obtained.

[0067] Next, a matching analysis is performed on the user's expected job functions and the actual job functions based on a language analysis model. When it is determined that the actual job function in a resume-job matching sample does not meet the user's expected function, the matching sample is marked as a negative match. For example, if the user's expected function is "data analyst," but the actual job function is "sales manager," this is a typical mismatch and should be treated as a negative match. Similarly, if the user's expected function is "front-end development," but the actual job function is "back-end development," although there is some overlap in skill requirements, the job function does not meet the user's expectations and should therefore also be considered a negative match.

[0068] In another optional implementation, the matching analysis process can also be standardized by combining a functional ontology library or a job tag system to address the differences in the descriptions of different job titles. For example, "full-stack development engineer" may include both front-end and back-end functions; during analysis, it can be broken down into standardized tags and then compared with the user's expected functions one by one. When the semantic similarity is lower than a preset threshold, or when there is no inclusion relationship between standardized tags, the submission is determined to be inconsistent with the user's expectations, and the corresponding sample is marked as a negative match sample.

[0069] By implementing the above methods, introducing negative matching samples where the expected function does not match the actual function can help improve the model's ability to make judgments in actual recall and increase the matching degree of the recommendation results with the user's true intent.

[0070] In some embodiments of this application, the method further includes expanding the positive matching sample: for each job posting submitted by the user, based on the industry of the job, obtaining a set of candidate jobs with similar job names; extracting industry experience requirements and / or skill requirements from the job description data of each candidate job; based on a pre-configured language analysis model, identifying and extracting candidate jobs that have the same industry experience requirements and / or skill requirements as the job and have different job names, thus obtaining expanded jobs; and constructing a positive matching sample based on the expanded jobs and the user's resume.

[0071] Specifically, for each job posting submitted by a user, a set of candidate jobs with similar names is retrieved based on the industry to which the job belongs. For example, when a user submits a job posting for "Front-end Development Engineer," candidate jobs such as "Web Development Engineer" and "Front-end Design Engineer" will be retrieved within the same industry.

[0072] Extract industry experience and / or skill requirements from the job description data for each candidate position, such as programming languages, database technologies, development frameworks, and industry background.

[0073] Based on a pre-configured language analysis model, semantic recognition and comparison are performed on the industry experience and / or skill requirements of candidate positions. Candidate positions with the same industry experience and / or skill requirements as the target position, but with different job titles, are selected as extended positions. For example, if the target position requires "familiarity with JavaScript and over three years of internet front-end development experience," the candidate position "Web Development Engineer" also includes these requirements in its description. Although the job titles are different, its industry experience and skill requirements are consistent with the target position, therefore it can be considered an extended position.

[0074] Based on the matching relationship between extended job titles and user resumes, new positive matching samples are constructed. Thus, even if users do not directly apply for these extended job titles, these candidate job titles can be included in the positive matching sample range, enriching the coverage of the training dataset and improving the model's generalization ability when faced with differences in job descriptions.

[0075] In another alternative implementation, the candidate job set can be obtained using various strategies. For example, the candidate set can be constructed based on keyword similarity of job names, industry tag mapping relationships, or job knowledge graphs. Furthermore, the screening of expanded jobs can introduce a semantic similarity threshold; for example, a candidate job is only included in the expanded job list if the skill set similarity between the candidate job and the target job exceeds a preset threshold. Further, cross-validation can be performed using job responsibility fields to ensure that expanded jobs are not only similar in skill requirements but also consistent with the target job in terms of responsibilities, thereby avoiding erroneous expansion due to vague descriptions.

[0076] Through the above implementation methods, more high-quality positive matching samples can be automatically expanded on the basis of the original delivery samples, which helps to alleviate the problem of sparse training data. This allows the semantic analysis model of job titles and resumes to learn richer job expression forms and skill combinations during the fine-tuning process, thereby improving the accuracy of the model in the recommendation recall stage.

[0077] Based on the embodiments provided in this application, this application also provides a recall method based on a semantic analysis model of job titles and resumes, combined with... Figure 3 As shown, the method includes steps 301 to 307.

[0078] Step 301: Obtain the job information corresponding to the newly published job.

[0079] Step 302: Based on the semantic analysis model of job postings and resumes, extract skill requirements, industry experience requirements, and job responsibility requirements from the job information and transform them into job semantic vectors.

[0080] The semantic analysis models for job titles and resumes are generated based on the training method for the semantic analysis models of job titles and resumes used for recall provided in the embodiments of this application.

[0081] Step 303: Based on the job semantic vector, obtain a first preset number of similar jobs for newly published jobs from the pre-built historical job database, and obtain the received historical resumes for similar jobs to obtain a historical resume set.

[0082] Step 304: Obtain user resume data for each historical resume in the historical resume set.

[0083] Step 305: Based on the semantic analysis model of job titles and resumes, extract project experience, work experience, and user skills from the user resume data of each historical resume and transform them into semantic vectors of historical resumes.

[0084] Step 306: Calculate the similarity between the semantic vector of the job posting and the semantic vector of each historical resume, and obtain the matching score between each historical resume and the newly posted job posting.

[0085] Step 307: Based on the matching score of each historical resume and the newly posted job, the users corresponding to the second preset number of historical resumes with the highest scores are used as the user recall results.

[0086] Specifically, based on the semantic analysis model of job titles and resumes, skill requirements, industry experience requirements, and job responsibility requirements are extracted from job information, and the extracted results are transformed into job semantic vectors. The semantic analysis model of job titles and resumes is a fine-tuned semantic model that can extract and encode key information in job descriptions into a unified semantic vector space, so that jobs with different expressions can also be reasonably represented in the vector space.

[0087] Based on the job's semantic vector, a pre-built historical job database is used to retrieve a first predetermined number of historical jobs most similar to the newly posted job. Similar jobs can be identified by calculating the similarity between their semantic vectors. Simultaneously, resumes received for these jobs in the past are also retrieved, forming a historical resume set.

[0088] For example, user resume data includes fields such as the user's skills, work experience, project experience, education background, and certificates.

[0089] A semantic analysis model based on job titles and resumes extracts information such as project experience, work experience, and user skills from each user's historical resume data and transforms it into a semantic vector for the historical resume. Based on this, resume information from different users is uniformly mapped to the same semantic space as the job title vector, thus supporting cross-domain matching.

[0090] Calculate the similarity between the semantic vector of the newly posted job and the semantic vector of historical resumes to obtain a matching score between each historical resume and the new job. The similarity can be calculated using cosine similarity or other vector similarity algorithms.

[0091] Based on the matching score, historical resumes are sorted, and users corresponding to the second preset number of historical resumes with the highest scores are selected as the final user recall results.

[0092] By pushing these users as potential candidates to companies, the visibility and matching effectiveness of newly posted positions can be improved.

[0093] In some embodiments of this application, combined with Figure 4 As shown, the method may also include steps 401 to 407.

[0094] Step 401: Obtain the user resume data corresponding to the newly published resume.

[0095] Step 402: Based on the semantic analysis model of job titles and resumes, extract project experience, work experience, and user skills from the user resume data and transform them into user resume semantic vectors.

[0096] Step 403: Based on the job semantic vector, obtain a third preset number of similar resumes of newly published resumes from the pre-built historical resume database, and determine the historical positions submitted by similar resumes to obtain a historical job set.

[0097] Step 404: Obtain the job information for each historical job in the historical job set.

[0098] Step 405: Based on the semantic analysis model of job postings and resumes, extract skill requirements, industry experience requirements, and job responsibility requirements from the job information of historical job postings, and transform the semantic vector of historical job postings.

[0099] Step 406: Calculate the similarity between the semantic vector of the user's resume and the semantic vector of each historical job posting to obtain the matching score between the historical job posting and the newly posted resume.

[0100] Step 407: Based on the matching score of each historical job posting and the newly published resume, the fourth preset number of historical jobs with the highest scores are used as the job recall results.

[0101] Specifically, based on the semantic analysis model of job titles and resumes, user resume data is parsed to extract at least one of project experience, work experience, and user skills. The extracted semantic features are then transformed into semantic vectors of user resumes, thereby uniformly representing the user's professional abilities and experience and ensuring the comparability of subsequent calculations.

[0102] Based on the semantic vectors of user resumes, a third preset number of similar resumes are retrieved from a pre-built historical resume database. The determination of similar resumes is typically achieved through semantic similarity calculation, which compares the semantic vectors of newly posted resumes with the semantic vectors of all resumes in the historical resume database, selecting the top-scoring resumes as similar resumes. Simultaneously, the historical job postings of these similar resumes are identified, resulting in a historical job posting set.

[0103] Next, retrieve the job information for each historical job in the historical job set. This job information includes, for example, skill requirements, industry experience requirements, job responsibilities, educational and work experience requirements, and geographic requirements.

[0104] Based on this, a semantic analysis model based on job titles and resumes is used to parse the job information of historical job titles, extract skill requirements, industry experience requirements, and job responsibilities, and transform the extracted results into semantic vectors of historical job titles. Based on this, job information and resume information are unified into the same semantic space, which facilitates direct vector matching.

[0105] Next, the similarity between the semantic vector of the user's resume and the semantic vector of each historical job posting is calculated to obtain the matching score between the historical job posting and the newly posted resume. Similarity calculation can use cosine similarity, dot product, or other semantic matching metrics.

[0106] Finally, based on the matching score between each historical job posting and the newly posted resume, the resumes are sorted, and the fourth-preset number of historical jobs with the highest matching scores are used as the final job retrieval results and recommended to users. In this way, even in the cold start phase where there is insufficient interaction data, users whose resumes have been posted can quickly obtain job recommendations that highly match their experience and skills.

[0107] In another optional implementation, the third and fourth preset quantities can be flexibly set according to business needs. This ensures the coverage of similar resume searches and the accuracy of the final job recommendations.

[0108] By implementing the above methods, the association information between historical resumes and historical job postings can be fully utilized to map newly published resumes to the existing job semantic space. This solves the problem that cold-start users have difficulty obtaining effective recommendations when there is a lack of behavioral data. It not only improves the relevance and accuracy of job retrieval, but also shortens the time from registration to obtaining effective recommendations for new users, thus improving the overall user experience of the recruitment platform.

[0109] Corresponding to the method embodiments of this application, this application also provides a training method apparatus for a semantic analysis model of recalled job positions and resumes. The training method apparatus for a semantic analysis model of recalled job positions and resumes in the embodiments of this application can correspond to the execution subject of the training method for a semantic analysis model of recalled job positions and resumes provided in the embodiments of this application. For specific details of the operation and / or function of each module / unit of the training method apparatus for a semantic analysis model of recalled job positions and resumes, please refer to the description of the corresponding part in the method provided in the above embodiments of this application.

[0110] Corresponding to the method embodiments of this application, this application also provides a training method and apparatus for a semantic analysis model of job positions and resumes for recall. The job recall apparatus based on the semantic analysis model of job positions and resumes in the embodiments of this application can correspond to the execution subject of the job recall method based on the semantic analysis model of job positions and resumes provided in the embodiments of this application. For specific details of the operation and / or function of each module / unit of the job recall apparatus based on the semantic analysis model of job positions and resumes, please refer to the description of the corresponding part in the method provided in the above embodiments of this application.

[0111] The electronic device in this application embodiment may be a user terminal device, a server, other computing devices, or a cloud server. Figure 5 The diagram illustrates the hardware structure of an electronic device according to an embodiment of this application. The electronic device may include a processor 501 and a memory 502 storing computer program instructions. When the processor 501 executes the computer program instructions, it implements the process or function of any of the methods described above.

[0112] Specifically, processor 501 may include a central processing unit (CPU), or an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application. Memory 502 may include a mass storage device for data or instructions. For example, memory 502 may be at least one of the following: a hard disk drive (HDD), read-only memory (ROM), random access memory (RAM), floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, universal serial bus (USB) drive, or other physical / tangible memory storage device. Alternatively, memory 502 may include removable or non-removable (or fixed) media. Furthermore, memory 502 may be internal or external to the integrated gateway disaster recovery device. Memory 502 may be non-volatile solid-state memory. In other words, typically memory 502 includes a tangible (non-transitory) computer-readable storage medium (such as a memory device) encoded with computer-executable instructions, and when the software is executed (e.g., by one or more processors), it can perform the operations described in the methods of the embodiments of this application. The processor 501 implements the process or function of any of the methods described in the above embodiments by reading and executing computer program instructions stored in the memory 502.

[0113] In one example Figure 5The illustrated electronic device may also include a communication interface 503 and a bus 510. The processor 501, memory 502, and communication interface 503 are connected via bus 510 and communicate with each other. Communication interface 503 is primarily used to enable communication between modules, devices, units, and / or equipment in the embodiments of this application. Bus 510 may include hardware, software, or both, and can couple components of the online data traffic billing device together. For example, the bus may include at least one of the following: Accelerated Graphics Port (AGP) or other graphics bus, Enhanced Industry Standard Architecture (EISA) bus, Front Side Bus (FSB), HyperTransport (HT) Interconnect, Industry Standard Architecture (ISA) bus, Infinite Bandwidth Interconnect, Low Pin Count (LPC) bus, memory bus, Microchannel Architecture (MCA) bus, Peripheral Component Interconnect (PCI) bus, PCI-Express (PCI-X) bus, Serial Advanced Technology Attachment (SATA) bus, Video Electronics Standards Association Local (VLB) bus, or other suitable buses. Bus 510 may include one or more buses. Although specific buses are described or illustrated in the embodiments of this application, any suitable bus or interconnection method may be considered in the embodiments of this application.

[0114] In conjunction with the methods in the above embodiments, this application also provides a computer-readable storage medium storing computer program instructions, which, when executed by a processor, implement the process or function of any of the methods in the above embodiments.

[0115] In addition, this application also provides a computer program product that stores computer program instructions, which, when executed by a processor, implement the process or function of any of the methods described above.

[0116] The flowcharts and / or block diagrams of methods, apparatuses, systems, and computer program products according to embodiments of this application have been exemplarily described above, and related aspects have been described. It should be understood that each block or combination thereof in the flowcharts and / or block diagrams may be implemented by computer program instructions, by dedicated hardware performing a specified function or action, or by a combination of dedicated hardware and computer instructions. For example, these computer program instructions may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to form a machine that enables the implementation of the function / action specified in each block or combination thereof in the flowcharts and / or block diagrams, executable via such processor. Such a processor may be a general-purpose processor, a dedicated processor, a special-purpose application processor, or a field-programmable logic circuit.

[0117] The functional blocks shown in the structural block diagrams of this application can be implemented as hardware, software, firmware, or a combination thereof. When implemented in hardware, they can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc.; when implemented in software, they are programs or code segments used to perform the required tasks. Programs or code segments can be stored in memory or transmitted over a transmission medium or communication link via data signals carried on a carrier wave. Code segments can be downloaded via computer networks such as the Internet or intranets.

[0118] It should be noted that this application is not limited to the specific configurations and processes described above or shown in the figures. The above descriptions are merely specific embodiments of this application. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the described systems, devices, modules, or units can be referred to the corresponding processes in the method embodiments, and need not be repeated here. It should be understood that the scope of protection of this application is not limited thereto. Any person skilled in the art can conceive of various equivalent modifications or substitutions within the scope of the technology disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application.

Claims

1. A method for training a semantic analysis model for recalling job postings and resumes, characterized in that, include: The system acquires user application behavior data, user resume data, and job description data, wherein the application behavior data includes the job positions submitted by the user in their resume; Based on the job positions submitted by the user, construct a matching sample of resumes and job positions; Based on a pre-configured language analysis model, the application behavior data, user resume data, and job description data corresponding to the matching samples are identified respectively. Matching samples corresponding to user resume data that do not meet the job constraints are filtered out to obtain positive matching samples; and negative matching samples are extracted from the samples that do not meet the job constraints. The pre-built semantic analysis model is fine-tuned and trained based on the positive and negative matching samples to generate semantic analysis models for job postings and resumes.

2. The method according to claim 1, characterized in that, The submission behavior data includes embedded data within the first associated time window corresponding to the resume submission time. The first associated time window includes a first preset time period before the resume submission time and a second preset time period after the resume submission time. The process of constructing a matching sample of resumes and positions based on the job postings submitted by the user includes: Extract the duration of time a user spends browsing the job postings and the number of times they submit a resume within the specified time window from the event tracking data for each resume submission time. If the duration is less than a preset time threshold and the number of delivery actions within the first associated time window is greater than a first preset number, the delivery action data is filtered out. Based on the resumes included in the application behavior data after filtering, a matching sample of resumes and job titles is constructed.

3. The method according to claim 1, characterized in that, The process of constructing a matching sample of resumes and positions based on the positions submitted by the user also includes: For each job posting a user submits a resume, the tracking data for the job within the second associated time window corresponding to the time the resume is received is obtained. The second associated time window includes a first preset time period before the time the resume is received and a second preset time period after the time the resume is received. Based on the data points collected within the second associated time window, the number of resumes received is determined. If the number of resumes received within the second associated time window exceeds the second preset number of times, the application data corresponding to the position will be filtered out.

4. The method according to claim 1, characterized in that, Based on a pre-configured language analysis model, the application behavior data, user resume data, and job description data corresponding to the matching samples are identified, and user resume data that does not meet the job constraints are filtered out, including: Based on a pre-configured language analysis model, user resume data and job description data are identified. Project experience, work experience, and user skills are extracted from user resume data, while industry experience requirements, job responsibilities, and skill requirements are extracted from job description data. For each of the job description data, including industry experience requirements, job responsibilities, and / or skill requirements, job constraints are generated. For each matching sample, compare the resume data with the job constraints to identify user resume data that does not meet the job constraints.

5. The method according to claim 1, characterized in that, The method also includes determining negative matching samples: Based on a pre-configured language analysis model, the expected job functions of users are extracted from the user resume data, and the actual job functions of the positions submitted by users are extracted from the job postings. Based on a pre-configured language analysis model, the expected and actual functions of the user are matched and analyzed. Matching samples in which the actual functions do not meet the user's expected functions are marked as negative matching samples.

6. The method according to claim 1, characterized in that, The method also includes expanding positive matching samples: For each job posting submitted by a user, based on the industry of the job, obtain a set of candidate jobs with similar job names to the job title; Industry experience requirements and / or skill requirements extracted from the job description data for each candidate position; Based on a pre-configured language analysis model, and based on the industry experience and / or skill requirements corresponding to the candidate positions, candidate positions that have the same industry experience and / or skill requirements as the stated positions but have different job names are identified and extracted to obtain extended positions. Based on the expanded job titles and user resumes, positive matching samples are constructed.

7. A recall method based on a semantic analysis model of job titles and resumes, characterized in that, include: Get the job information corresponding to the newly published job; Based on the semantic analysis model of job postings and resumes, skill requirements, industry experience requirements, and job responsibility requirements are extracted from the job information and transformed into job semantic vectors. The semantic analysis model of job postings and resumes is generated based on the training method of the semantic analysis model of job postings and resumes for recall as described in any one of claims 1 to 6. Based on the job semantic vector, a first preset number of similar jobs of the newly published job are obtained from a pre-built historical job database, and the received historical resumes of the similar jobs are obtained to obtain a historical resume set. Obtain the user resume data for each historical resume in the historical resume set; Based on the semantic analysis model of job titles and resumes, project experience, work experience, and user skills are extracted from the user resume data of each historical resume and transformed into a semantic vector of the historical resume. Calculate the similarity between the semantic vector of the job posting and the semantic vector of each historical resume to obtain a matching score between each historical resume and the newly posted job posting; Based on the matching score between each historical resume and the newly posted job, the users corresponding to the second preset number of historical resumes with the highest scores are used as the user recall results.

8. The method according to claim 7, characterized in that, The method further includes: Retrieve user resume data corresponding to newly published resumes; Based on the semantic analysis model of job titles and resumes, project experience, work experience, and user skills are extracted from the user resume data and transformed into a semantic vector of the user resume. Based on the job semantic vector, a third preset number of similar resumes of the newly published resume are obtained from a pre-built historical resume database, and the historical positions submitted by the similar resumes are determined to obtain a historical job set. Obtain the job information for each historical job in the historical job set; Based on the semantic analysis model of job titles and resumes, skill requirements, industry experience requirements, and job responsibility requirements are extracted from the job information of the historical job titles, and the semantic vector of the historical job titles is transformed. Calculate the similarity between the semantic vector of the user's resume and the semantic vector of each historical job posting to obtain the matching score between the historical job posting and the newly posted resume; Based on the matching score between each historical job posting and the newly posted resume, the fourth preset number of historical jobs with the highest scores are used as the job recall results.

9. An electronic device, characterized in that, The electronic device includes a processor and a memory storing computer program instructions; when the electronic device executes the computer program instructions, it implements the method as described in any one of claims 1 to 8.

10. A computer program product, characterized in that, It includes computer program instructions that, when executed by a processor, implement the method as described in any one of claims 1 to 8.

Images