Method and device for job-post matching retrieval by fusing user behavior and text semantics

Abstract

The application discloses a position and resume matching recall method and device fusing user behavior and text semantics. The method comprises the following steps: obtaining text corpus of a target object, the target object comprising a position and a resume; inputting the text corpus into a fine-tuned large language semantic vector model to generate a basic semantic vector of the target object; obtaining text corpus of a historical behavior object corresponding to the target object, the historical behavior object being an object having a historical interaction behavior with the target object and the interaction behavior satisfying a preset condition; inputting text corpus of a preset number of historical behavior objects into the large language semantic vector model to generate behavior semantic vectors of the historical behavior objects; and weighting and fusing the basic semantic vector and the behavior semantic vector of the target object according to a preset fusion weight to generate a fusion semantic vector of the target object. The embodiment of the application can improve the recall precision of the position or the resume.

Description

Technical Field

[0001] This application relates to the field of computer technology, as well as artificial intelligence and natural language processing technology, and in particular to a method, apparatus, device, program product and storage medium for matching people and jobs by integrating user behavior and text semantics. Background Technology

[0002] With the rapid development of online recruitment platforms, the number of job postings and resumes accumulated on these platforms has reached tens of millions. How to accurately match suitable individuals or positions for users (including recruiters and job seekers) from this massive amount of data has become an important research direction in the field of recommendation systems. Traditional recruitment recommendation methods mainly rely on calculating the similarity between job descriptions and resume text. In recent years, semantic vector technology based on large language models has demonstrated excellent performance in text matching tasks. By transforming high-dimensional text into low-dimensional dense vectors and then using vector retrieval technology to quickly retrieve similar items from massive amounts of data, it has been widely applied in recommendation systems, advertising systems, and other fields.

[0003] However, in practical applications, the aforementioned text matching method based on semantic vectors of large language models often suffers from low accuracy in job descriptions and resumes. This is because job descriptions contain information dimensions that the other does not have. For example, the benefits and corporate culture mentioned in job descriptions are almost never found in resumes, but this type of information has a significant impact on whether job seekers submit their resumes. Conversely, personal project experience and technical details included in resumes may not be fully mentioned in job descriptions. As a result, the accuracy of the retrieved job or resume matching is low and cannot meet user needs. Summary of the Invention

[0004] In view of this, embodiments of this application provide a method, apparatus, electronic device, computer-readable storage medium, and computer program product for human-job matching and recall that integrates user behavior and text semantics, thereby solving at least one technical problem.

[0005] This application provides a method for job matching and recall that integrates user behavior and text semantics, comprising: acquiring text corpus of a target object, the target object including job title and resume; inputting the text corpus into a fine-tuned large language semantic vector model to generate a basic semantic vector of the target object; acquiring text corpus of historical behavior objects corresponding to the target object, the historical behavior objects being objects that have historical interaction behavior with the target object and whose interaction behavior meets preset conditions; inputting a preset number of text corpus of the historical behavior objects into the large language semantic vector model to generate a behavior semantic vector of the historical behavior objects; weighting and fusing the basic semantic vector of the target object with the behavior semantic vector according to a preset fusion weight to generate a fused semantic vector of the target object; and performing vector retrieval based on the fused semantic vector to recall a preset number of candidate objects matching the target object from a candidate object library.

[0006] Optionally, according to the method of this application embodiment, obtaining the text corpus of the target object and obtaining the text corpus of the historical behavior object respectively include: obtaining the original description text of the corresponding object, wherein the original description text includes the entire content of the job description text or the work experience description text and basic information text in the resume; extracting core attribute words from the original description text through a large language model and based on preset prompt words, wherein the core attribute words include at least one of function, industry, sub-field, skill point, professional or language certificate; organizing the extracted core attribute words through natural language and concatenating them with the corresponding original description text to generate the text corpus.

[0007] Optionally, according to the method of this application embodiment, the fine-tuning process of the large language semantic vector model includes: constructing a training sample set, the training sample set including multiple triples, each triple including query text, positive samples and negative samples; wherein, the query text is job description text, the positive samples are resume texts that have been submitted by job seekers and received positive feedback from recruiters, and the negative samples include random resume texts that are not related to the job description text, and resume texts that have been submitted by job seekers but received negative feedback from recruiters; and using the training sample set to fine-tune the basic large language semantic vector model.

[0008] Optionally, according to the method of this application embodiment, the loss function used in the fine-tuning training is the self-knowledge distillation loss function; the training rounds of the fine-tuning training are determined based on the recall and precision metrics of the large language semantic vector model on the test set.

[0009] Optionally, according to the method of this application embodiment, the historical interaction behavior includes the recruiter marking the resume as appropriate, downloading, saving, forwarding, communicating by phone or sending an interview invitation, and the job seeker submitting or viewing job applications; the preset condition is that the interaction behavior belongs to a preset behavior set that the user is interested in; the preset number is K, where K ranges from 5 to 15.

[0010] Optionally, according to the method of this application embodiment, the preset fusion weight includes semantic weight and behavioral weight, wherein the semantic weight and the behavioral weight are both numbers between 0 and 1 and their sum is 1; the step of weighted fusion of the basic semantic vector of the target object and the behavioral semantic vector according to the preset fusion weight includes: configuring different behavioral weights for the corresponding behavioral semantic vectors according to different historical interaction behavior types; and adding the product of the semantic weight and the basic semantic vector to the product of the behavioral weight and the weighted sum of each behavioral semantic vector.

[0011] Optionally, according to the method of this application embodiment, when the historical behavior data of the target object meets the preset sparsity condition, the method further includes: calculating the semantic similarity between the basic semantic vector of the target object and the basic semantic vector of each object in the behavior rich object pool, wherein the behavior rich object pool is composed of behavior rich objects with a historical behavior data volume greater than a preset threshold; selecting a preset number of objects with the highest semantic similarity from the behavior rich object pool as similar objects of the target object; obtaining the behavior semantic vector of the historical behavior object corresponding to the similar object as the borrowed behavior semantic vector of the target object; and weightedly fusing the basic semantic vector of the target object and the borrowed behavior semantic vector according to the preset fusion weight to generate the fused semantic vector of the target object.

[0012] Optionally, according to the method of this application embodiment, calculating the semantic similarity between the basic semantic vector of the target object and the basic semantic vector of each object in the behavior rich object pool includes: when the target object is a job position, calculating the semantic similarity between the job position and other job positions in the behavior rich object pool of the job position dimension; when the target object is a resume, calculating the semantic similarity between the resume and other resumes in the behavior rich object pool of the resume dimension.

[0013] Optionally, according to the method of this application embodiment, the vector retrieval based on the fused semantic vector includes: generating fused semantic vectors for each candidate object and constructing a vector index; during online recall, retrieving a preset number of candidate objects with the closest cosine distance from the vector index based on the identifier of the target object, wherein the preset number of recalls is 200-300.

[0014] Optionally, according to the method of the embodiments of this application, the large language semantic vector model is the BGE model, or any one of the GTE-Qwen2 series model, multilingual-e5-base model, or m3e-base model.

[0015] This application also proposes a person-job matching and recall device that integrates user behavior and text semantics, comprising: a first acquisition module for acquiring text corpus of a target object, the target object including job title and resume; a basic semantic vector generation module for inputting the text corpus into a fine-tuned large language semantic vector model to generate a basic semantic vector of the target object; a second acquisition module for acquiring text corpus of a preset number of historical behavior objects corresponding to the target object, the historical behavior objects being objects that have historical interaction behavior with the target object and whose interaction behavior meets preset conditions; a behavior semantic vector generation module for inputting the text corpus of the preset number of historical behavior objects into the large language semantic vector model to generate a behavior semantic vector of the historical behavior objects; a fusion semantic vector generation module for weighted fusion of the basic semantic vector of the target object and the behavior semantic vector according to a preset fusion weight to generate a fusion semantic vector of the target object; and a recall module for performing vector retrieval based on the fusion semantic vector to recall a preset number of candidate objects matching the target object from a candidate object library.

[0016] This application provides an electronic device, which includes a processor and a memory storing computer program instructions; the electronic device executes the computer program instructions to implement the method described above.

[0017] This application provides a computer program product, which includes computer program instructions that, when executed, implement the method described above.

[0018] This application provides a computer-readable storage medium storing computer program instructions, which, when executed by a processor, implement the method described above.

[0019] By weighted and fused the basic semantic vector of the target object with the behavioral semantic vector of its historical behavioral objects, a fused semantic vector that simultaneously represents textual semantic information and user behavioral preference information is generated. This effectively overcomes the problem of insufficient matching accuracy caused by the difference in dimensionality between job description and resume text information in simple text matching, and improves the accuracy of recall results. At the same time, vector retrieval enables efficient recall under massive data. Attached Figure Description

[0020] 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.

[0021] Figure 1 This is a schematic diagram of the system architecture of an embodiment of this application.

[0022] Figure 2 This is a flowchart of the person-job matching and recall method that integrates user behavior and text semantics according to an embodiment of this application.

[0023] Figure 3 This is a schematic diagram illustrating the processing of a human-job matching and recall method that integrates user behavior and text semantics, according to an embodiment of this application.

[0024] Figure 4 This is a schematic diagram of a vector fusion process based on job title, according to an embodiment of this application.

[0025] Figure 5 This is a schematic diagram of a vector fusion process based on a resume, according to an embodiment of this application.

[0026] Figure 6 This is a schematic structural block diagram of the person-job matching and recall method apparatus that integrates user behavior and text semantics according to an embodiment of this application.

[0027] Figure 7 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0028] The principles and spirit of this application will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided to make the principles and spirit of this application clearer and more thorough. The exemplary embodiments provided herein are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this application.

[0029] Embodiments of this application relate to terminal devices and / or servers. Implementations of this application can be a system, terminal, device, method, computer-readable storage medium, or computer program product, and can be specifically implemented as entirely hardware, entirely software, or a combination of hardware and software. Figure 1This diagram illustrates a system architecture according to an embodiment of the present application, including a terminal device 102 and a server 104. The terminal device 102 may include at least one of the following: a smartphone, tablet computer, laptop computer, desktop computer, smart TV, various wearable devices, augmented reality (AR) devices, virtual reality (VR) devices, etc. A client application, such as an app, mini-program, or browser-based client, can be installed on the terminal device 102. Users can input commands through the client, and the terminal device 102 can send request information containing the commands to the server 104. Upon receiving the request information, the server 104 performs corresponding processing and returns the processing result information to the terminal device 102. The server 104 may be a local server or a cloud server, and may be a single server or a server cluster, etc.

[0030] In this document, the terms "first," "second," "third," etc., are used only to distinguish one entity (or operation) from another in textual description, and do not require or imply any sequential order between these entities (or operations).

[0031] Figure 2 This is a flowchart of the person-job matching and recall method that integrates user behavior and text semantics according to an embodiment of this application. The method includes the following steps: S101: Obtain the text corpus of the target objects, including job titles and resumes; S102: Input the text corpus into the finely tuned large language semantic vector model to generate the basic semantic vector of the target object; S103: Obtain the text corpus of the historical behavior objects corresponding to the target object. The historical behavior objects are objects that have historical interaction behaviors with the target object and whose interaction behaviors meet preset conditions. S104: Input the text corpus of a preset number of historical behavior objects into the large language semantic vector model to generate the behavior semantic vectors of the historical behavior objects; S105: Based on the preset fusion weights, the basic semantic vector and the behavioral semantic vector of the target object are weighted and fused to generate the fused semantic vector of the target object; S106: Perform vector retrieval based on fused semantic vectors to recall a preset number of candidate objects that match the target object from the candidate object library.

[0032] This application provides a method for matching and recalling people based on job requirements by integrating user behavior and text semantics. The processing procedure of this technical solution is described in detail below.

[0033] First, in step S101, the text corpus of the target objects is obtained. The target objects include two types: job postings and resumes, corresponding to job seekers and employers in a recruitment scenario, respectively. It should be noted that the text corpus is not simply the original text content. In some embodiments of this application, the text corpus may optionally be core attribute words extracted from the original descriptive text, such as job functions, industries, sub-fields, skill points, professional or language certificates, etc. These core attribute words are then organized using natural language processing and concatenated with the original descriptive text to obtain a clearer structure and more prominent focus, thereby improving the accuracy of subsequent semantic vector generation. In step S102, the above text corpus is input into a fine-tuned large-scale semantic vector model to generate the basic semantic vector of the target objects. The large-scale semantic vector model referred to here is a natural language processing model capable of encoding input text into a low-dimensional dense vector representation. The basic semantic vector is used to represent the textual semantic information of the target objects themselves. Taking job titles as an example, when the target object is the job description of a Java development engineer, the finely tuned semantic vector model maps the job description text into a fixed-dimensional numerical vector, which represents the content meaning of the job in the semantic space; similarly, when the target object is a job seeker's resume, the resume text is mapped into the corresponding semantic vector.

[0034] In step S103, the text corpus of historical behavior objects corresponding to the target object is obtained. Here, historical behavior objects refer to objects that have historical interaction behaviors with the target object and whose interaction behaviors meet preset conditions. For example, when the target object is a job posting, the historical behavior objects could be the resumes of job seekers who submitted their resumes to that job and received positive feedback from the recruiter (such as being marked as suitable, having their resume downloaded, or receiving an interview invitation). These behaviors reflect the recruiter's approval of the resume, and therefore the text information carried by the resume can reflect characteristics that match the target job. Similarly, when the target object is a resume, the historical behavior objects could be the job postings that the job seeker actively applied for; the application behavior itself indicates the job seeker's interest and preference for that job. In step S104, the text corpus of a preset number of historical behavior objects is input into the same large language semantic vector model mentioned above to generate behavioral semantic vectors for the historical behavior objects. It is worth emphasizing that the behavioral semantic vectors here can be generated by the same model as the basic semantic vectors in step S102, ensuring that they are in the same vector space, laying the foundation for subsequent vector fusion operations. The preset number is set to control the amount of behavioral information involved in the fusion, and to avoid introducing noise or increasing the computational burden. It can usually be set according to the actual application scenario, such as taking the text corpus of historical behavioral objects corresponding to the most recent historical behaviors.

[0035] In step S105, the basic semantic vector and behavioral semantic vector of the target object are weighted and fused according to preset fusion weights to generate a fused semantic vector of the target object. Specifically, this fusion operation is not a simple vector concatenation or averaging, but rather a weighted summation by configuring weight parameters for each type of vector. The semantic weights corresponding to the basic semantic vector and the behavioral weights corresponding to the behavioral semantic vector are both between 0 and 1, and their sum is 1, to balance the contribution of the semantics of the text itself and user behavior preferences to the final fused semantic vector. The fused semantic vector generated by this step represents both the textual semantic information and user behavior preference information of the target object. Finally, in step S106, vector retrieval is performed based on the fused semantic vector to recall a preset number of candidate objects that match the target object from the candidate object library. This step can be achieved by calculating the similarity (e.g., cosine similarity) between the fused semantic vector of the target object and the pre-calculated candidate object vectors in the candidate object library, sorting them from high to low similarity, and selecting a preset number of candidate objects as the recall result, which is then returned to the subsequent ranking stage of the recommendation system for further processing.

[0036] This application achieves weighted fusion of the target object's basic semantic vector and the behavioral semantic vector of historical behavior objects. The fused vector includes both the target object's own textual semantic information and the user's preference information reflected through historical interaction behavior. Since historical behavior objects come from the other party that actually interacted with the target object (job corresponding to resume, resume corresponding to job), their textual information includes content dimensions not present in the target object's original text, such as the job seeker's personal project experience not found in job text, and job benefits not found in resume text. Therefore, this application compensates for the lack of information dimensions in single text semantic matching by using a cross-text type substitution fusion method, significantly improving the accuracy and precision of the job matching recall results. Furthermore, by completing the recall through vector retrieval based on the fused semantic vector, it fully utilizes the millisecond-level response advantage of vector retrieval technology under massive data, making this application highly engineering-ready in actual online recommendation scenarios.

[0037] In some embodiments of this application, optionally, obtaining the text corpus of the target object and obtaining the text corpus of historical behavior objects respectively include: Retrieve the original description text of the corresponding object. The original description text includes the full content of the job description text or the work experience description text and basic information text in the resume. Using a large language model and based on preset prompts, core attribute words are extracted from the original description text. The core attribute words include at least one of the following: function, industry, sub-field, skill point, professional or language certificate. The extracted core attribute words are organized using natural language and then concatenated with the corresponding original descriptive text to generate a text corpus.

[0038] In this embodiment, when obtaining the text corpus of the corresponding object, the original text content is not used directly. Instead, the original descriptive text of the object is obtained first. It should be noted that the content range of the original descriptive text differs depending on the type of the target object and the historical behavior object: when the corresponding object is a job posting, the original descriptive text includes the entire content of the job description text; when the corresponding object is a resume, the original descriptive text includes the work experience description text and basic information text (such as education, major, etc.) in the resume. Among these, the work experience description text is the closest to the job description text in terms of language style and expression, and it is the information that recruiters are most concerned about. Therefore, the description texts of the two most recent work experiences can be used, which can cover the key information related to job matching and avoid the interference of irrelevant information in the full resume text on the semantic vector.

[0039] Furthermore, core attribute words are extracted from the original description text using a large language model and based on preset prompts. Specifically, the preset prompts are instruction texts pre-constructed based on business experience in the recruitment industry, used to guide the large language model to focus on the most critical attribute dimensions in the job matching process. These core attribute words include at least one of the following: job function, industry sub-field, skill points, and professional or language certifications. For example, for a job description text for a Java development engineer, inputting it into the large language model will guide it to extract "Job function: Java development engineer; Industry sub-field: Internet / Computer software / Backend development / Microservices; Skill points: Spring Boot, Redis, MySQL; Major: Computer Science and Technology or related majors". Similarly, for a job seeker's resume text, inputting it into the large language model will guide it to extract "Job function: Java development engineer; Industry sub-field: Internet / Backend development; Skill points: Spring Boot, Docker; Major: Computer Science and Technology". Subsequently, the extracted core attribute words are organized using natural language, for example, into the format of "The core requirements of this position include: function as a Java development engineer, industry as an internet company, sub-specialty as backend development, and core skills including Spring Boot and Redis." The organized core attribute word text is then concatenated with the corresponding original description text to generate the final text corpus used as input to the large language semantic vector model. Through this process, the core attribute words are explicitly incorporated into the semantic vector generation process, thus effectively guiding the encoding process of the vector model.

[0040] In addition, it should be emphasized that the large language model in this application can be an existing known model, such as open source models like ChatGLM, Qwen, and Baichuan, or a closed source model that is called via API.

[0041] This application utilizes a large language model to extract core attribute words during the text corpus construction stage, explicitly organizes them in a natural language manner, and then concatenates them into the original text. This allows core attribute words to directly participate in the semantic vector encoding process in text form, thereby guiding the model to focus on key information dimensions in person-job matching. This avoids the model's attention being diverted to redundant text unrelated to matching due to a lack of recruitment industry background, thus improving the accuracy of basic semantic vector and behavioral semantic vector generation. At the same time, since the organization and extraction process of core attribute words is decoupled from the model training process, when business requirements change and the focus dimension needs to be adjusted, only the prompt words need to be modified to complete the debugging, without retraining the model. This significantly reduces the maintenance and optimization costs of the solution in engineering applications.

[0042] In some embodiments of this application, optionally, the fine-tuning process of the large language semantic vector model includes: Construct a training sample set, which includes multiple triples, each triple consisting of query text, positive sample, and negative sample. The query text is the job description text. Positive samples are resumes that have been submitted by job seekers and received positive feedback from recruiters. Negative samples include random resumes that are not related to the job description text, as well as resumes that have been submitted by job seekers but received negative feedback from recruiters. The basic large language semantic vector model is fine-tuned using the training sample set.

[0043] In this embodiment, to fine-tune the large language semantic vector model, a training sample set for fine-tuning training needs to be constructed first. This training sample set includes multiple training samples in the form of triples. Each triple specifically includes three components: query text, positive samples, and negative samples. The query text is the job description text, i.e., the description content corresponding to the job used as a query condition in a recruitment scenario. Positive samples are resumes submitted by job seekers and received positive feedback from recruiters. Positive feedback here refers to actions taken by recruiters that indicate their satisfaction or interest in the resume, such as marking the resume as suitable, downloading the resume attachment, saving the resume, forwarding the resume, or sending an interview invitation. Such actions indicate that the resume and the corresponding job have a high degree of matching in the recruiter's actual judgment, thus constituting a positive sample. Negative samples are further divided into two parts: The first part consists of random resume texts that are not related to the job description, i.e., resumes randomly selected from the resume database. These negative samples are easily distinguishable from the query text due to their significant semantic differences and are called normal negative samples. The second part consists of resume texts from job seekers who submitted their resumes but received negative feedback from recruiters. These negative samples are characterized by the fact that job seekers believed they were qualified for the position and therefore submitted their resumes, making them semantically relevant to the job. However, recruiters judged them as mismatched based on higher screening criteria or certain detailed dimensions (such as specific areas of expertise or skills), resulting in negative feedback. These samples are difficult to distinguish from positive samples because they appear similar but are not actually a match and are called hard negative samples. Subsequently, the basic large language semantic vector model is fine-tuned using the constructed training sample set. Specifically, the triplet loss function guides the model to narrow the distance between the query text and positive samples in the vector space while simultaneously widening the distance between the query text and negative samples, thereby enhancing the model's semantic discrimination ability in job-person matching scenarios within the recruitment industry.

[0044] By fine-tuning the training samples using triples that include positive samples, normal negative samples, and hard negative samples, the large language semantic vector model can fully learn the real matching logic between job postings and resumes in the recruitment industry. In particular, by introducing hard negative samples that have been submitted but received negative feedback, the model's ability to distinguish easily confused samples that are semantically similar overall but do not match in detail is effectively enhanced, avoiding the problem of insufficient model discrimination ability caused by relying solely on random negative samples. At the same time, the positive samples come from real submissions and positive feedback behaviors, so that the matching relationships learned by the model are consistent with user preferences and screening criteria in actual business scenarios, rather than just shallow semantic similarity at the text level, thereby improving the contribution of semantic vectors to matching accuracy in the recall stage.

[0045] In some embodiments of this application, optionally, the loss function used for fine-tuning training is the self-knowledge distillation loss function; the number of training rounds for fine-tuning training is determined based on the recall and precision metrics of the large language semantic vector model on the test set.

[0046] During fine-tuning training, this application employs a self-knowledge distillation loss function. It should be noted that the self-knowledge distillation loss function is a training strategy that does not require the introduction of an external large teacher model. Its core idea lies in improving the model's learning performance through multi-task collaborative optimization or self-distillation mechanisms between multi-layer representations within the model. Compared to traditional triplet loss functions (such as triplet loss) or contrastive learning loss functions (such as InfoNCE), the self-knowledge distillation loss function can fully utilize the model's own intermediate layer representations to guide the training process in the absence of external supervision signals, thereby achieving superior semantic representation capabilities with limited samples.

[0047] Furthermore, the specific approach to determining the termination criteria for fine-tuning training was as follows: The initial training epochs were set to three. After each complete epoch, the trained model was validated and evaluated on a test set. Specifically, job semantic vectors were used to recall resumes from the test set. The model's performance was comprehensively evaluated based on both precision and recall. If the precision and recall metrics did not reach the expected level, the training epochs were increased until the metrics no longer significantly improved or reached a business-acceptable standard. Based on the parameter configuration during the actual fine-tuning process, the final number of training epochs was five. In addition, during training, auxiliary parameters such as the training group size, maximum query length, and learning rate were dynamically adjusted based on the actual machine hardware resources and training sample size to ensure the stability and convergence efficiency of the training process.

[0048] By employing a self-knowledge distillation loss function, the fine-tuning process can still achieve good vector representation capabilities while maintaining low training costs. By determining the fine-tuning termination condition based on the model's precision and recall metrics on the test set rather than a fixed number of training epochs, the training process becomes more data-driven, effectively preventing underfitting due to too few epochs or overfitting due to too many epochs. This improves the generalization performance and recall of the large language semantic vector model in real-world application scenarios.

[0049] In some embodiments of this application, the large language semantic vector model may optionally be the BGE model, or any one of the GTE-Qwen2 series models, the multilingual-e5-base model, or the m3e-base model.

[0050] In the implementation of this method, the large language semantic vector model, as the core component for encoding text corpora into dense vectors, is not limited to a single fixed model but can be selected from a variety of open-source models with semantic vector encoding capabilities. Specifically, the model can be the BGE model, which is one of the more commonly used basic models in the field of semantic vectors, possessing good Chinese and multilingual processing capabilities and performing excellently in multiple semantic matching benchmark tests. In addition, the GTE-Qwen2 series models can also be used; the multilingual-e5-base model can also be used, which supports the generation of multilingual semantic vectors; and the m3e-base model can also be used. All of the above models have the ability to process Chinese text and some English text, and are all built on the Transformer architecture. Their core function is to map the input text into a fixed-dimensional dense vector representation, and the output results can be directly used for downstream vector retrieval and similarity calculation tasks. In the implementation of this solution, any model can be flexibly selected from the above models as the implementation basis of the large language semantic vector model, based on the deployment environment, hardware resource conditions, inference performance requirements, and the specific performance of each model on the recruitment industry corpus, according to the actual business scenario. Then, in accordance with the fine-tuning process of this application, the model can be fine-tuned and trained using training samples from the recruitment industry to further adapt it to the characteristics of the corpus of job descriptions and resume texts in the recruitment field.

[0051] In some embodiments of this application, optionally, historical interaction behaviors include recruiters marking resumes as appropriate, downloading, saving, forwarding, communicating by phone or sending interview invitations, and job seekers submitting or viewing job applications. The preset condition is that the interactive behavior belongs to a set of preset behaviors that the user is interested in; The preset quantity is K, where K ranges from 5 to 15.

[0052] In the actual operation of recruitment platforms, target objects (job positions or resumes) will engage in various types of interactions with other parties. For example, when the target object is a job position, the recruiter, as the manager of that position, will perform a series of actions on resumes submitted by the recruiter or job seeker. These actions include, but are not limited to, marking the resume as suitable, downloading resume attachments, saving the resume, forwarding the resume to other colleagues, communicating with the job seeker by phone, or sending interview invitations. These actions reflect, to varying degrees, the recruiter's recognition and interest in the job seeker corresponding to the resume. When the target object is a resume, the job seeker, as the owner of that resume, will perform a series of actions on job positions on the platform, including actively submitting their resume to a specific job or viewing the detailed information of a specific job. These actions reflect the job seeker's interest in the job and their job application intention.

[0053] The aforementioned interactive behaviors constitute the specific content of historical interactive behaviors. Based on this, the preset condition is limited to interactive behaviors belonging to a preset set of behaviors that the user is interested in. It should be noted that different types of interactive behaviors reflect different levels of user interest. For example, a job seeker's viewing of a position only indicates initial interest, while submitting an application indicates that they believe they are qualified for the position after careful consideration; the latter shows a significantly higher level of confidence in their interest than the former. Similarly, a recruiter's marking a resume as suitable and sending an interview invitation demonstrates a significantly higher level of approval than downloading or saving it. Therefore, the preset set of behaviors can be set according to actual business needs, including behaviors such as submitting applications, sending interview invitations, and marking resumes as suitable, while excluding behaviors such as exposure and simple browsing. This ensures that the historical behavior objects participating in subsequent integration have sufficient representativeness of user preferences. After determining the historical behavior objects that meet the preset condition, a preset number, i.e., the top K objects, needs to be selected from them according to their degree of interest, where K ranges from 5 to 15. The ranking criteria can be based on the timeliness of the behavior (e.g., more recent behaviors have higher weight) or the intensity of the behavior itself (e.g., an interview invitation is better than a download). By taking the top K behaviors, we can focus on the few key behaviors that best represent the user's current preferences, avoid introducing too much noise, and effectively control computational costs.

[0054] In some embodiments of this application, these K behavior objects can also be arranged in chronological order to obtain the K historical interaction objects that are closest to the current time and meet preset conditions.

[0055] By specifying historical interaction behavior types as various positive actions taken by recruiters on resumes and job seekers submitting or viewing job applications, the source and scope of behavioral data are clarified. By setting the preset conditions that the interaction behavior belongs to a preset set of behaviors that users are interested in, low-confidence behavioral noise is effectively eliminated, ensuring that the historical behavior objects participating in vector fusion can truly and reliably reflect user preferences. By limiting the preset number to the top K ranked by the degree of interest, with K ranging from 5 to 15, sufficient behavioral information is ensured to support the effectiveness of preference representation, while avoiding the introduction of redundant information or reduction of real-time computing efficiency by too many historical behaviors, achieving a reasonable balance between recall accuracy and computational performance.

[0056] In some embodiments of this application, optionally, the preset fusion weights include semantic weights and behavioral weights, both of which are numbers between 0 and 1, and their sum is 1; according to the preset fusion weights, the basic semantic vector of the target object and the behavioral semantic vector are weighted and fused, including: Different behavior weights are configured for the corresponding behavior semantic vectors based on different types of historical interaction behaviors; The product of semantic weight and basic semantic vector is added to the product of behavior weight and weighted sum of each behavior semantic vector.

[0057] When performing weighted fusion based on preset fusion weights, the preset fusion weights specifically include two types of parameters: semantic weights and behavioral weights. Semantic weights correspond to the contribution of the target object's basic semantic vector to the fusion result, while behavioral weights correspond to the contribution of the historical behavior object's behavioral semantic vector to the fusion result. Both are values ​​between 0 and 1, and their sum equals 1. For example, if the semantic weight α is set to 0.6, the behavioral weight β is correspondingly 0.4, indicating that in the final generated fused semantic vector, the target object's own textual semantic information accounts for 60%, while the user's preference information reflected through historical interaction behavior accounts for 40%. By adjusting the relative values ​​of these two, the tendency of the recall strategy between faithfulness to textual content and catering to user preferences can be flexibly controlled. In the specific calculation of weighted fusion, different behavioral weights need to be configured for the corresponding behavioral semantic vectors according to different types of historical interaction behavior. It should be noted that this behavioral weight is different from the aforementioned behavioral weight parameter β, but refers to the individual weight used when performing weighted summation within the behavioral semantic vector. For example, when the target object is a job posting, it may correspond to multiple historical behavior objects (i.e., multiple resumes from which the recruiter gave positive feedback). If the recruiter sends an interview invitation to one of the resumes, the behavioral semantic vector of that resume is assigned a higher individual weight (e.g., 1.0). If only another resume is downloaded without further contact, a lower individual weight is assigned (e.g., 0.5). The value of this individual weight can be preset according to the degree of recognition reflected by the behavior type. After configuring the individual weights of each behavioral semantic vector, the specific calculation formula for weighted fusion is as follows: when the target object is a job posting, the fused job vector = α1 × job basic semantic vector + β1 × Σ(wi × behavioral semantic vector of the i-th positive feedback resume), where α1 is the weight of the job basic semantic vector, β1 is the resume behavior weight, and wi is the individual weight corresponding to the behavior of the i-th positive feedback resume. When the target object is a resume, the fused resume vector = α2 × resume basic semantic vector + β2 × Σ(ni × the behavioral semantic vector of the i-th positive feedback position), where α2 is the resume basic semantic weight, β2 is the position behavioral weight, and ni is the individual weight corresponding to the behavior of the i-th applied position.

[0058] By configuring different individual weights for the corresponding behavioral semantic vectors based on different types of historical interaction behaviors, the fusion process can finely distinguish the differences in preference intensity implied by different behaviors. For example, an interview invitation has a higher degree of certainty in preference expression compared to downloading. This differentiated weight configuration makes the fused semantic vectors more accurately reflect the user's true preference level, thereby improving the quality of recall ranking. At the same time, by adjusting the relative values ​​of semantic weights and behavioral weights, the strategy can be flexibly adjusted between text matching and preference matching based on the results of online experiments, enhancing the adaptability of the solution in different business scenarios.

[0059] Optionally, in some embodiments of this application, when the historical behavior data of the target object meets a preset sparsity condition, the method further includes: Calculate the semantic similarity between the base semantic vector of the target object and the base semantic vector of each object in the behavior rich object pool. The behavior rich object pool consists of behavior rich objects whose historical behavior data volume is greater than a preset threshold. Select a predetermined number of objects with the highest semantic similarity from the pool of objects with rich behavior as similar objects to the target object; Obtain the behavioral semantic vector of the historical behavioral objects corresponding to similar objects, and use it as the borrowed behavioral semantic vector of the target object; Based on the preset fusion weights, the basic semantic vector of the target object and the semantic vector of the borrowing behavior are weighted and fused to generate the fused semantic vector of the target object.

[0060] In real-world recruitment scenarios, not all target objects possess abundant historical behavioral data. For example, newly posted positions or resumes of newly registered users often have zero or extremely limited historical behavioral data due to insufficient exposure and interaction records on the platform. This situation constitutes a pre-defined sparsity condition. When the historical behavioral data of a target object meets the pre-defined sparsity condition—that is, when the number of historical behavioral objects is insufficient to support weighted fusion based on real behavior—the following borrowing process is initiated. First, the semantic similarity between the target object's basic semantic vector and the basic semantic vectors of each object in the behavior-rich object pool is calculated. The behavior-rich object pool is a pre-built database composed of behavior-rich objects with historical behavioral data exceeding a pre-defined threshold. For example, positions and resumes with more than 10 historical positive feedback behaviors are selected to construct a job behavior-rich pool and a resume behavior-rich pool, respectively. Semantic similarity can be calculated using methods such as cosine distance between vectors; a higher similarity value indicates that the two objects are closer at the textual semantic level. Second, a pre-defined number of objects with the highest semantic similarity are selected from the behavior-rich object pool as similar objects to the target object. It is important to emphasize that this matching process follows the principle of matching similar categories. That is, when the target object is a job title, only job titles with similar semantics are searched in the job title behavior rich pool; when the target object is a resume, only resumes with similar semantics are searched in the resume behavior rich pool.

[0061] Subsequently, the behavioral semantic vectors of the historical behavior objects corresponding to these similar objects are obtained, and these borrowed behavioral semantic vectors are used as the borrowed behavioral semantic vectors of the target object. Taking a newly posted Java development engineer job as an example, this job has no historical behavioral data yet. Through semantic similarity, several Java development engineer jobs with the closest text descriptions are found in a pool of jobs with rich behavioral data. These similar jobs all have rich records of recruiter behavior. The behavioral semantic vectors of the positive feedback resumes corresponding to these similar jobs are used as the borrowed behavioral semantic vectors of the new job. Finally, according to the preset fusion weights, the borrowed behavioral semantic vectors are weighted and fused with the target object's own basic semantic vectors to generate the fused semantic vector of the target object. Through the above processing, even objects with sparse or no behavior can obtain fused vectors that reflect user preference information, instead of simply degenerating into pure text semantic vectors.

[0062] By constructing a pool of objects with rich behaviors and performing semantic similarity matching between objects of the same type based on basic semantic vectors, job postings or resumes with sparse behaviors can borrow behavioral information from similar objects of the same type that are semantically similar in text and rich in behaviors. This effectively alleviates the problem of missing behavioral data caused by newly posted job postings or newly registered users in recruitment scenarios and prevents the recall effect from significantly decreasing due to cold start.

[0063] In some embodiments of this application, optionally, calculating the semantic similarity between the base semantic vector of the target object and the base semantic vectors of each object in the behavior-rich object pool includes: When the target object is a job title, the semantic similarity between the job title and other job titles is calculated in the behavior rich object pool of the job title dimension. When the target object is a resume, the semantic similarity between the resume and other resumes is calculated in the behavior rich object pool of the resume dimension.

[0064] In this embodiment, when the historical behavior data of a target object meets a preset sparsity condition, it is necessary to calculate the semantic similarity between the target object's basic semantic vector and the basic semantic vectors of each object in the behavior-rich object pool to find the behavior-rich object whose textual semantics are closest to the target object. This calculation process does not involve cross-type mixed matching of job titles and resumes; instead, it is performed in the corresponding pool of objects of the same type according to the specific type of the target object. Specifically, when the target object is a job title, in the pre-built behavior-rich object pool at the job dimension, the basic semantic vector of the job title is compared with the basic semantic vector of each job title in the pool to calculate the semantic similarity score between the job title and each behavior-rich job title. This behavior-rich object pool at the job dimension only includes behavior-rich jobs with a historical behavior data volume greater than a preset threshold and does not include any resume objects. Correspondingly, when the target object is a resume, in the pre-built behavior-rich object pool at the resume dimension, the basic semantic vector of the resume is compared with the basic semantic vector of each resume in the pool to calculate the semantic similarity score between the resume and each behavior-rich resume. The behavioral rich object pool for this resume dimension only includes behavioral rich resumes with a historical behavioral data volume greater than a preset threshold, and does not include any job objects.

[0065] By employing the aforementioned matching strategy, we ensure that semantic similarity calculations are performed between objects with consistent representations. This type of semantic similarity calculation between similar objects is more accurate and reliable than cross-type job and resume similarity calculations.

[0066] In some embodiments of this application, vector retrieval may optionally be performed based on fused semantic vectors, including: Generate fused semantic vectors for each candidate object and construct vector indexes; During online recall, the candidate objects with the closest cosine distance to the target object are retrieved from the vector index based on the target object's identifier. The preset recall number is 200-300.

[0067] After generating the fused semantic vectors of the target and candidate objects, efficient online retrieval needs to be achieved based on these vectors. First, the fused semantic vectors of each candidate object are generated offline or near-line. The candidate object, referred to here, is the set of objects that serve as the retrieval target during the retrieval phase, and its type is the opposite of the target object: when the target object is a job posting, the candidate objects are all resumes in the resume database; when the target object is a resume, the candidate objects are all job postings in the job database. A corresponding fused semantic vector is generated for each candidate object; this process can be completed in batches offline, without consuming online request processing time. Subsequently, a vector index is constructed using the generated fused semantic vectors of each candidate object. It should be noted that a vector index is an efficient data structure for high-dimensional vectors. It partitions and organizes the vector space through specific indexing algorithms (such as hierarchical navigable small-world graphs, productization, etc.), enabling nearest neighbor retrieval in sets of millions or even tens of millions of vectors to return search results within milliseconds, without traversing all vectors to calculate similarity one by one, thus significantly improving retrieval efficiency under massive datasets.

[0068] During the online recall phase, when a recall request including the target object identifier is received, the corresponding fused semantic vector is read from the in-memory database based on the identifier. This vector is then input into the vector index, and the similarity between this vector and the candidate object vectors in the index is calculated using cosine distance. The vectors are then sorted from closest to furthest (i.e., from highest to lowest similarity) according to their cosine distance, and the candidate objects closest to this index are retrieved from the index, representing a preset recall number. This preset recall number is set to 200 to 300. This range ensures that the recall phase can select a moderately sized candidate set from the candidate object library. This set includes enough potential matching objects to guarantee the recall rate, while avoiding excessive computational pressure on the subsequent sorting phase due to an excessively large recall number.

[0069] The implementation methods and advantages of the embodiments of this application have been described above through multiple examples. The specific processing procedures of the embodiments of this application are described in detail below with reference to specific examples.

[0070] Figure 3 This is a schematic diagram illustrating the processing procedure of a person-job matching and recall method that integrates user behavior and text semantics, according to an embodiment of this application. Combined with... Figure 3As shown, suppose an online recruitment platform M posts a Java development engineer job posting. The job description includes job responsibilities, required technology stack (e.g., Spring Boot, Redis, MySQL), qualifications (e.g., computer-related major, 3+ years of work experience), and benefits. Meanwhile, the platform's resume database stores millions of resumes. The platform aims to use the recall method described in this application to filter the most relevant resumes from this massive database for the subsequent ranking stage. The following describes the process step by step.

[0071] Step 1: Obtain the complete original description text of the Java Developer job title, which serves as the original description text for the target object. Next, using a large language model and based on preset prompts, extract core attribute words from the input text, including: job title, industry, sub-field, skill points, professional qualifications, and language certifications. This yields: Job title: Java Developer; Industry: Internet / Computer Software; Sub-field: Backend Development / Microservice Architecture; Skill points: Spring Boot, Redis, MySQL, Docker; Professional qualifications: Computer Science and Technology or related majors. Then, organize the extracted core attribute words using natural language to form supplementary text emphasizing the core attributes, resulting in: Job title: Java Developer; Industry: Internet; Sub-field: Backend Development; Core skills include Spring Boot, Redis, MySQL, and Docker. Finally, concatenate this organized core attribute word text with the original job description text to generate the job title's text corpus. Input this text corpus into a fine-tuned large language semantic vector model to generate the job title's basic semantic vector Vj. In the platform's offline processing flow, each resume in the resume database also generates its own basic semantic vector Vc in the same way as described above.

[0072] Step 2: Retrieve resume records from the database that have historical interactions with the job posting. These interactions include various positive feedback actions taken by the recruiter after the job seeker submitted their resume, such as marking the resume as suitable, downloading the resume attachment, saving the resume, and sending interview invitations. Assume that since the job posting, over 200 resumes have been received, and the recruiter has given positive feedback to several of them. The platform sorts these resumes that received positive feedback according to their level of interest (e.g., interview invitations have the highest weight, followed by downloads), and selects the top K resumes, where K is 10, representing the 10 resumes that best reflect the recruiter's preferences. For each of the 10 resumes, its original description text is obtained, including the work experience description text (prioritizing the descriptions of the two most recent work experiences) and basic information text (such as education, major, etc.). Core attribute words are extracted from these texts using the same large language model based on preset prompts (e.g., the extraction result is "Job Title: Java Development Engineer; Industry: Internet; Specialization: Backend Development; Skills: Spring Boot, Docker, MySQL; Major: Computer Science and Technology"). The extracted core attribute words are then organized in natural language and concatenated with the original description text to generate the text corpus for each historical behavior object. Subsequently, the text corpus of these 10 resumes is input into the same finely tuned large language semantic vector model as in step 1 to generate behavioral semantic vectors Vci for each historical behavior object, where i=1,2,...,10. These behavioral semantic vectors reside in the same vector space as the basic semantic vectors.

[0073] Step 3: According to the preset fusion weights, the basic semantic vector of the job generated in Step 1 is weighted and fused with the 10 behavioral semantic vectors generated in Step 2. In this embodiment, based on online experimental experience, the platform sets the semantic weight α to 0.6 and the behavioral weight β to 0.4. That is, in the fusion result, the textual semantic information of the job itself contributes 60%, and the recruiter's preference information reflected by historical behaviors contributes 40%. At the same time, according to the specific type of historical interaction behavior, different individual behavioral weights wi are configured for each behavioral semantic vector. For example, the weight corresponding to sending an interview invitation is 1.0, the weight corresponding to marking as suitable is 0.8, and the weight corresponding to downloading attachments is 0.6. The final fused semantic vector of the job is calculated using the formula VJ=α×Vj+β×∑(wi×Vci), where VJ is the final fused semantic vector of the job.

[0074] Step 4: In the offline or near-line processing stage, a corresponding fusion semantic vector has been generated for each resume in the resume database according to the methods described in Steps 1 to 3 above. A vector index is constructed based on these resume fusion semantic vectors. This vector index uses efficient nearest neighbor retrieval algorithms such as hierarchical navigable small world graphs, enabling millisecond-level response times even with tens of millions of data points. In the online recall stage, when the platform receives a request to recall resumes for the Java development engineer position, the platform reads the fusion semantic vector VJ generated in Step 3 from the database based on the position's identifier. This vector is then input into the vector index. By calculating the cosine distance between VJ and the fusion semantic vectors of each resume in the index, the resumes are sorted from closest to furthest based on their cosine distance. A preset number of resumes with the closest distance are selected as the recall result. In this embodiment, the preset number of resumes is set to 300, meaning the size of the recalled candidate resume set is controlled within 300 to facilitate fine-tuning in the subsequent sorting stage. At this point, the recall process for this position is complete. For other positions or resumes on the platform (as target objects), the processing procedure is completely consistent with the above steps.

[0075] Step 5: Figure 4 This is a schematic diagram illustrating a job-based vector fusion process according to an embodiment of this application. Combined with... Figure 4 As shown, for newly posted positions without significant historical behavior, or newly registered job seeker resumes, when the historical behavior data of the target object meets the preset sparsity condition, i.e., the number of historical behavior objects is insufficient (e.g., less than 5), the platform adaptively adjusts steps 2 and 3 above. Specifically, for positions meeting the sparsity condition, assuming the Java development engineer position is a newly posted position with no job seekers applying and no positive feedback from the recruiter, the behavioral semantic vector of the historical behavior object corresponding to this position cannot be obtained in step 2. At this point, the platform initiates a leveraged processing flow. The platform pre-builds a pool of rich behavioral objects, specifically including a pool of rich behavioral objects at the position dimension and a pool at the resume dimension. The pool at the position dimension consists of positions with historical behavior data exceeding a preset threshold (e.g., more than 10 positive feedback behaviors). The platform calculates the cosine similarity between the basic semantic vector V_j of this position and the basic semantic vectors of each position in the pool at the position dimension, selecting the top 3 positions with the highest semantic similarity as similar objects for this position.

[0076] Suppose that the three most semantically similar Java developer job titles are Job A, Job B, and Job C, each with rich historical behavioral data from recruiters. The platform obtains the behavioral semantic vectors of the historical behavioral objects (i.e., resumes that received positive feedback for each job) corresponding to these three jobs, and uses these borrowed behavioral semantic vectors as the borrowed behavioral semantic vector for this job. Subsequently, following the same fusion method as in step 3, the platform calculates V_J = α × V_j + β × ∑(wi × V_ci), where V_ci is the behavioral semantic vector borrowed from the historical behavioral objects of similar jobs, generating the fused semantic vector for this job. This ensures that the newly posted job title can also obtain a fused semantic vector that simultaneously represents textual semantics and user preferences, avoiding a decline in recall effectiveness due to the cold start problem.

[0077] Step 6: Processing when the resume is the target object. The previous steps were explained using a job posting as the target object. When the target object is a resume, the processing logic is completely symmetrical. The platform obtains the text corpus of a resume and generates its basic semantic vector V_c. It also obtains the historical behavior objects corresponding to the resume, i.e., the text corpus of jobs actively applied for by the job seeker, and generates a behavioral semantic vector V_ji. A fused semantic vector for the resume is generated according to V_C = α × V_c + β × ∑(wi × V_ji), where V_C is the final fused resume vector, and wi is the time weight (the more recent the application, the higher the weight). During online retrieval, the fused semantic vector of the resume is used to retrieve a preset number of the closest jobs from the job vector index, which are then used as candidate matching jobs for the resume. Figure 5 This is a schematic diagram illustrating a vector fusion process based on a resume, according to an embodiment of this application. Combined with... Figure 5 As shown, in a cold start scenario, a newly registered user's resume with no application activity can still generate a fused semantic vector by searching for resumes with similar textual semantics in the resume-dimensional behavior enrichment object pool and borrowing their behavior vectors. All the above vector fusion operations are performed in a unified vector space, ensuring the comparability of fused semantic vectors generated under different object types and processing paths.

[0078] Correspondingly, this application also provides a person-job matching and recall device that integrates user behavior and text semantics, for reference. Figure 6 ,include: The first acquisition module 110 is used to acquire the text corpus of the target object, which includes job titles and resumes; The basic semantic vector generation module 120 is used to input the text corpus into the fine-tuned large language semantic vector model to generate the basic semantic vector of the target object. The second acquisition module 130 is used to acquire a preset number of text corpora of historical behavior objects corresponding to the target object. The historical behavior objects are objects that have historical interaction behaviors with the target object and whose interaction behaviors meet preset conditions. The behavior semantic vector generation module 140 is used to input a preset number of text corpora of the historical behavior objects into the large language semantic vector model to generate the behavior semantic vectors of the historical behavior objects. The fusion semantic vector generation module 150 is used to perform weighted fusion of the basic semantic vector of the target object and the behavioral semantic vector according to the preset fusion weight, so as to generate the fusion semantic vector of the target object. The recall module 160 is used to perform vector retrieval based on the fused semantic vector and recall a preset number of candidate objects that match the target object from the candidate object library.

[0079] Based on at least one of the above embodiments, the electronic device in the embodiments of this application may be a user terminal device, a server, other computing devices, or a cloud server. Figure 7 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of this application. The electronic device may include a processor 601 and a memory 602 storing computer program instructions. The processor 601 reads and executes the computer program instructions stored in the memory 602 to implement the process or function of any of the methods in the above embodiments.

[0080] Specifically, processor 601 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 602 may include a mass storage device for data or instructions. For example, memory 602 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 602 may include removable or non-removable (or fixed) media. Furthermore, memory 602 may be internal or external to the integrated gateway disaster recovery device. Memory 602 may be non-volatile solid-state memory. In other words, typically memory 602 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.

[0081] As an example, Figure 7 The illustrated electronic device may also include a communication interface 603 and a bus 610. The processor 601, memory 602, and communication interface 603 are connected via bus 610 and communicate with each other. Bus 610 may include hardware, software, or both, and may couple components of an online data traffic metering device together. 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 bus. Bus 610 may include one or more buses. Although specific buses are described or shown in the embodiments of this application, any suitable bus or interconnection method is contemplated in the embodiments of this application.

[0082] 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.

[0083] 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.

[0084] The flowcharts and / or block diagrams of methods, terminals, 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 such that these instructions, executed via such processor, enable the implementation of the function / action specified in each block or combination thereof in the flowcharts and / or block diagrams. Such a processor may be a general-purpose processor, a dedicated processor, a special-purpose application processor, or a field-programmable logic circuit.

[0085] 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.

[0086] 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, terminals, 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 equivalent modifications or substitutions that can be conceived by those skilled in the art within the scope of the technology disclosed in this application should be covered within the scope of protection of this patent application.

Claims

1. A method for person-job matching and recall that integrates user behavior and text semantics, characterized in that, include: Acquire text corpus of target objects, including job titles and resumes; The text corpus is input into a finely tuned large language semantic vector model to generate the basic semantic vector of the target object; Obtain the text corpus of historical behavior objects corresponding to the target object, wherein the historical behavior objects are objects that have historical interaction behaviors with the target object and whose interaction behaviors meet preset conditions; A preset number of text corpora of the historical behavior objects are input into the large language semantic vector model to generate behavioral semantic vectors of the historical behavior objects. According to the preset fusion weight, the basic semantic vector of the target object and the behavioral semantic vector are weighted and fused to generate the fused semantic vector of the target object; Vector retrieval is performed based on the fused semantic vector to recall a preset number of candidate objects that match the target object from the candidate object library.

2. The method according to claim 1, characterized in that, The acquisition of the text corpus of the target object and the acquisition of the text corpus of the historical behavior object respectively include: Obtain the original description text of the corresponding object, which includes the full content of the job description text or the work experience description text and basic information text in the resume; Using a large language model and based on preset prompt words, core attribute words are extracted from the original description text. The core attribute words include at least one of the following: function, industry, sub-field, skill point, professional or language certificate. The extracted core attribute words are organized using natural language and then concatenated with the corresponding original descriptive text to generate the text corpus.

3. The method according to claim 1, characterized in that, The fine-tuning process of the large language semantic vector model includes: Construct a training sample set, which includes multiple triples, each triple including query text, positive sample, and negative sample; The query text is a job description text, the positive sample is a resume text that has been submitted by job seekers and received positive feedback from recruiters, and the negative sample includes random resume texts that are not related to the job description text, as well as resume texts that have been submitted by job seekers but received negative feedback from recruiters. The training sample set is used to fine-tune the basic large language semantic vector model.

4. The method according to claim 3, characterized in that, The loss function used in the fine-tuning training is the self-knowledge distillation loss function; the number of training rounds in the fine-tuning training is determined based on the recall and precision metrics of the large language semantic vector model on the test set.

5. The method according to claim 1, characterized in that, The historical interaction behaviors include recruiters marking resumes as appropriate, downloading, saving, forwarding, communicating by phone or sending interview invitations, and job seekers submitting or viewing job applications. The preset condition is that the interactive behavior belongs to a preset set of behaviors that the user is interested in; The preset quantity is K, where K ranges from 5 to 15.

6. The method according to claim 1 or 5, characterized in that, The preset fusion weights include semantic weights and behavioral weights, where both semantic weights and behavioral weights are numbers between 0 and 1, and their sum is 1. The step of weightedly fusing the basic semantic vector of the target object with the behavioral semantic vector according to a preset fusion weight includes: Different behavior weights are configured for the corresponding behavior semantic vectors based on different types of historical interaction behaviors; The product of the semantic weight and the basic semantic vector is added to the product of the behavior weight and the weighted sum of each behavior semantic vector.

7. The method according to claim 1, characterized in that, When the historical behavior data of the target object meets a preset sparsity condition, the method further includes: Calculate the semantic similarity between the base semantic vector of the target object and the base semantic vector of each object in the behavior rich object pool, wherein the behavior rich object pool consists of behavior rich objects whose historical behavior data volume is greater than a preset threshold. Select a predetermined number of objects with the highest semantic similarity from the pool of behavior-rich objects as similar objects to the target object; Obtain the behavioral semantic vector of the historical behavioral object corresponding to the similar object, and use it as the borrowed behavioral semantic vector of the target object; According to the preset fusion weight, the basic semantic vector of the target object and the borrowing behavior semantic vector are weighted and fused to generate the fused semantic vector of the target object.

8. The method according to claim 7, characterized in that, The calculation of the semantic similarity between the base semantic vector of the target object and the base semantic vectors of each object in the behavior-rich object pool includes: When the target object is a job position, the semantic similarity between the job position and other job positions is calculated in the behavior rich object pool of the job position dimension. When the target object is a resume, the semantic similarity between the resume and other resumes is calculated in the behavior rich object pool of the resume dimension.

9. The method according to claim 1, characterized in that, The vector retrieval based on the fused semantic vector includes: Generate fused semantic vectors for each candidate object and construct vector indexes; During online recall, candidate objects with the closest cosine distance to the target object are retrieved from the vector index based on the identifier of the target object. The preset recall number is 200-300.

10. The method according to claim 1, characterized in that, The large language semantic vector model is the BGE model, or any one of the GTE-Qwen2 series models, the multilingual-e5-base model, or the m3e-base model.

11. A person-job matching and recall device that integrates user behavior and text semantics, characterized in that, include: The first acquisition module is used to acquire the text corpus of the target objects, including job titles and resumes; The basic semantic vector generation module is used to input the text corpus into the finely tuned large language semantic vector model to generate the basic semantic vector of the target object. The second acquisition module is used to acquire a preset number of text corpora of historical behavior objects corresponding to the target object. The historical behavior objects are objects that have historical interaction behaviors with the target object and whose interaction behaviors meet preset conditions. The behavior semantic vector generation module is used to input a preset number of text corpora of the historical behavior objects into the large language semantic vector model to generate behavior semantic vectors of the historical behavior objects. The semantic vector fusion generation module is used to perform weighted fusion of the basic semantic vector of the target object and the behavioral semantic vector according to a preset fusion weight, so as to generate the fusion semantic vector of the target object. The recall module is used to perform vector retrieval based on the fused semantic vector and recall a preset number of candidate objects that match the target object from the candidate object library.

12. An electronic device, characterized in that, The electronic device is a terminal device or a server. 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 1-10.

13. A computer program product, characterized in that, It includes computer program instructions that, when executed, implement the method as described in any one of claims 1-10.

14. A computer-readable storage medium, characterized in that, It stores computer program instructions that, when executed, implement the method as described in any one of claims 1-10.

Images