Word vector text similarity analysis method and system in recruitment field

Abstract

The embodiment of the invention provides a word vector text similarity analysis method and system in a recruitment field, and can solve the problems that in a core link of resume and post matching, a model has significant deviation in understanding of skills and post requirements of job seekers, so that the recommendation precision is reduced, and the misjudgment rate is increased. The method comprises the steps of obtaining a technical word set from a recruitment field corpus based on a pre-trained named entity recognition model; inputting the high-frequency technical keyword set into the large language model, performing category judgment on each technical keyword under the constraint of a technical keyword classification framework used for describing technical knowledge in the recruitment field, selecting a corresponding multi-round question prompt word template based on the category to which each technical keyword belongs, and performing multi-round multi-angle question on the large language model, and training a word meaning matching model on the basis of the generated refined training corpus used for representing the technical semantic relationship in the recruitment field, so as to obtain a plurality of question and answer fragments and context samples generated around each technical keyword.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method and system for word vector text similarity analysis in the recruitment field. Background Technology

[0002] In intelligent recruitment scenarios driven by natural language processing (NLP) technology, the application of general-purpose word vector models faces challenges of semantic drift and anisotropy. Semantic drift refers to the fact that, because general-purpose models are trained on broad corpora such as Wikipedia and news articles, they fail to fully learn the subtle semantic differences within vertical domains, leading to the calculation of excessively high semantic similarity for dissimilar technical terms or job functions. This bias is further compounded by the anisotropy problem inherent in word vector spaces, causing cosine similarity-based matching mechanisms to become even more distorted. The direct consequence is that, in the core stage of resume and job matching, the model's understanding of job seekers' skills and job requirements becomes significantly biased, mistakenly classifying mismatched resumes as matches or omitting qualified resumes, ultimately leading to decreased recommendation accuracy and increased misjudgment rate. Summary of the Invention

[0003] This application provides a word vector text similarity analysis method and system in the recruitment field, which can solve the problem that in the core link of resume and job matching, the model has a significant deviation in understanding the job seeker's skills and job requirements, mistakenly judging mismatched resumes as matches, or missing qualified resumes, which ultimately leads to a decrease in recommendation accuracy and an increase in the misjudgment rate.

[0004] The first aspect of this application provides a method for word vector text similarity analysis in the recruitment field, including: Based on the pre-trained named entity recognition model, target words related to technology are extracted from multiple job description texts and multiple resume texts in the recruitment domain corpus, resulting in a set of technical words containing multiple technical keywords. A set of high-frequency technical keywords is input into a large language model. Under the constraints of a technical keyword classification framework used to describe technical knowledge in the recruitment field, each technical keyword is classified. Based on its category, a corresponding multi-round question prompt template is selected, and multiple rounds of multi-angle questions are performed on the large language model to obtain multiple question-and-answer fragments and context samples generated around each technical keyword. The technical keyword classification framework includes at least one or more categories such as programming language, operating system, and tool name. A word sense matching model is trained on a refined training corpus generated based on the multiple question-and-answer fragments and context samples to represent the semantic relationships of technologies in the recruitment field. This enables the word sense matching model to output word vector representations of the technical keywords and to calculate the semantic similarity between different texts based on the word vector representations.

[0005] Optionally, a co-occurrence graph is constructed based on a corpus of recruitment data. The nodes of the co-occurrence graph are technical keywords, and the edge weights are used to characterize the co-occurrence strength of the technical keywords in the job description text and resume text. Combining the co-occurrence graph and user interaction information in the recruitment field, multiple groups of technical keywords with semantically similar relationships are identified. Technical keywords belonging to the same technical keyword group are determined as a set of skill synonyms. In the refined training corpus, synonym replacement is performed on the technical keywords to generate an extended training corpus, so as to obtain an enhanced training corpus set. The word sense matching model is trained or retrained based on the enhanced training corpus to obtain a target word sense matching model for recruitment scenarios. When calculating the text similarity between the job description text to be matched and the resume text to be matched, the target semantic matching model is invoked to obtain the word vector representations of the technical keywords in the job description text to be matched and the resume text to be matched. Based on the word vector representations, the technical semantic similarity between the job description text and the resume text is calculated, so that the technical semantic similarity is used as a component of the job and resume matching score or matching score.

[0006] Optional, also includes: Construct an annotated training set containing text from the recruitment field, and annotate multiple job description texts and multiple resume texts with technical terms related to the business domain; The sequence labeling model is trained based on the labeled training set so that it can identify technology-related target words in unlabeled text and output them as technology keywords, thereby obtaining the named entity recognition model.

[0007] Optionally, the step of inputting the set of high-frequency technical keywords into a large language model, and determining the category of each technical keyword under the constraints of a technical keyword classification framework used to describe technical knowledge in the recruitment field, includes: High-frequency technical keywords and their typical contextual fragments in the original corpus are input into the large language model. The large language model then outputs the target category label and confidence score of each technical keyword from the category set defined by the technical keyword classification framework. When the confidence level is lower than a preset threshold, a manual review or secondary questioning process is triggered to correct the target category label of the technical keyword. When the technical keyword is a programming language category, a multi-round question sequence is generated to guide the large language model to output the applicable scenarios, common supporting frameworks, commonly used libraries, typical job responsibilities, and competency level classifications of the programming language; and / or, When the technical keyword is a tool name category, generate a multi-round question sequence to guide the large language model to output the tool's functional positioning, common use cases, applicable job positions, and differences from other tools; and / or, When the technical keywords are in the test-related category, a multi-round question sequence is generated to guide the large language model to output test types, typical test steps, coverage objects, and collaboration methods with development positions. The training samples in the refined training corpus consist of answers generated from each multi-round question sequence, original technical keywords, and contextual fragments.

[0008] Optionally, the co-occurrence graph constructed based on the recruitment domain corpus, wherein the nodes of the co-occurrence graph are technical keywords, and the edge weights are used to characterize the co-occurrence strength of technical keywords in job description text and resume text, including: In job description text and resume text, multiple technical keywords that are in the same sentence, the same paragraph or a preset window range are regarded as a co-occurrence relationship, and the initial edge weight of the keyword pair is calculated based on the number of co-occurrences or statistical relevance. The process involves combining the co-occurrence graph and user interaction information in the recruitment field to identify multiple groups of technical keywords with semantically similar relationships. Technical keywords belonging to the same group are defined as a set of skill synonyms. Furthermore, synonym replacement is performed on these technical keywords in the refined training corpus to generate an expanded training corpus, resulting in an enhanced training corpus set, including: Based on the co-occurrence graph and combined with statistical information from the interaction logs of the recruitment system, the edge weights of technical keyword pairs that frequently appear simultaneously in user filtering, collection, application, and interview behaviors are increased, while the edge weights of technical keyword pairs that appear in negative feedback samples are decreased, in order to obtain the adjusted co-occurrence graph. In the adjusted co-occurrence graph, multiple technical keyword clusters are obtained based on edge weight threshold, community partitioning algorithm or clustering algorithm, and multiple technical keywords in the same technical keyword cluster are identified as a set of skill synonyms; When generating extended training corpora, for target technical keywords in refined training corpora, multiple sample variants are generated by replacing them with other technical keywords in the corresponding skill synonym set, while keeping the original semantic labels unchanged, so as to form training samples that are semantically equivalent but have different surface word forms.

[0009] Optional, also includes: Technical keywords and their context fragments from refined and expanded training corpora are constructed as positive and negative sample pairs. The parameters of the word sense matching model are constrained by a contrastive learning loss function or a ranking-based loss function, so as to reduce the distance between word vector representations of technical keywords that are semantically similar but have different word forms, and increase the distance between word vector representations of technical keywords that are semantically different but have similar word forms.

[0010] A second aspect of this application provides a word vector text similarity analysis device for the recruitment field, including: The recognition unit is used to extract technology-related target words from multiple job description texts and multiple resume texts in the recruitment domain corpus based on a pre-trained named entity recognition model, and obtain a set of technical words containing multiple technical keywords. The analysis unit is used to input a set of high-frequency technical keywords into a large language model. Under the constraints of a technical keyword classification framework used to describe technical knowledge in the recruitment field, it determines the category of each technical keyword and selects the corresponding multi-round question prompt word template based on its category. It then performs multi-round, multi-angle questioning on the large language model to obtain multiple question-and-answer fragments and context samples generated around each technical keyword. The technical keyword classification framework includes at least one or more categories such as programming language, operating system, and tool name. The training unit is used to train a word sense matching model on a refined training corpus generated based on the multiple question-and-answer fragments and context samples to represent the semantic relationships of technologies in the recruitment field, so that the word sense matching model can output the word vector representation of the technology keywords and calculate the semantic similarity between different texts based on the word vector representation.

[0011] A third aspect of this application provides an electronic system including a memory and a processor, wherein the processor is used to execute a computer program stored in the memory to implement the steps of the above-described word vector text similarity analysis method in the recruitment field.

[0012] A fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method for analyzing word vector text similarity in the recruitment field.

[0013] In summary, the word vector text similarity analysis method for the recruitment field provided in this application extracts technology-related target words from multiple job description texts and resume texts in a recruitment field corpus based on a pre-trained named entity recognition model, obtaining a set of technical words containing multiple technical keywords. The set of high-frequency technical keywords is input into a large language model. Under the constraints of a technical keyword classification framework used to describe technical knowledge in the recruitment field, each technical keyword is categorized, and corresponding multi-round question prompt word templates are selected based on their respective categories. Multiple rounds of multi-angle questioning are performed on the large language model to obtain multiple question-and-answer fragments and context samples generated around each technical keyword. The technical keyword classification framework includes at least one or more categories such as programming languages, operating systems, and tool names. A word sense matching model is trained on a refined training corpus generated based on the multiple question-and-answer fragments and context samples to characterize the semantic relationships of technical aspects in the recruitment field. This enables the word sense matching model to output word vector representations of the technical keywords and to calculate the semantic similarity between different texts based on these word vector representations. Thus, the problem of semantic drift and anisotropy-induced semantic comprehension bias caused by general word vectors in recruitment scenarios is solved. First, the named entity recognition model ensures that the training and matching process focuses on truly distinguishable technical keywords, avoiding the dominance of general expressions such as teamwork and communication skills in the vector space, thus reducing the interference of general expressions on technical semantics. Second, the technical keyword classification framework and multi-turn question answering based on a large language model ensure that the training corpus contains a wealth of explicit knowledge about skill categories, application scenarios, differences in job responsibilities, and boundaries with similar skills, providing clear supervision signals for subsequent model learning. Finally, the word-meaning matching model trained on refined corpora makes the word vector space more consistent with the intuitive semantic distance relationships in the recruitment field: truly similar technical experiences cluster in the space, while skills with similar names but completely different responsibilities are separated, making the similarity calculation between job descriptions and resumes more consistent with the judgment of human recruitment experts. Overall, this method can significantly improve the accuracy of job-resume matching, reduce instances such as incorrectly recommending Java test engineers to Java development engineer positions, and improve the screening efficiency and credibility of intelligent recruitment systems with massive combinations of candidates and positions.

[0014] Accordingly, the word vector text similarity analysis device, electronic system, and computer-readable storage medium for the recruitment field provided in this embodiment of the invention also have the above-mentioned technical effects. Attached Figure Description

[0015] Figure 1 A flowchart illustrating a possible word vector text similarity analysis method for the recruitment field provided in this application embodiment; Figure 2A schematic structural block diagram of a possible word vector text similarity analysis device in the recruitment field provided in the embodiments of this application; Figure 3 A schematic diagram of the hardware structure of a possible word vector text similarity analysis device in the recruitment field provided in this application embodiment; Figure 4 A schematic structural block diagram of a possible electronic system provided for embodiments of this application; Figure 5 This is a schematic structural block diagram of a possible computer-readable storage medium provided for embodiments of this application. Detailed Implementation

[0016] This application provides a word vector text similarity analysis method and system in the recruitment field, which can solve the problem that in the core link of resume and job matching, the model has a significant deviation in understanding the job seeker's skills and job requirements, mistakenly judging mismatched resumes as matches, or missing qualified resumes, which ultimately leads to a decrease in recommendation accuracy and an increase in the misjudgment rate.

[0017] To resolve the above issues, please refer to [link / reference]. Figure 1 The flowchart of a word vector text similarity analysis method in the recruitment field provided in this application embodiment can specifically include: S110-S130.

[0018] S110, Based on the pre-trained named entity recognition model, extract technology-related target words from multiple job description texts and multiple resume texts in the recruitment domain corpus to obtain a set of technical words containing multiple technical keywords.

[0019] S120, input the set of high-frequency technical keywords into the large language model, under the constraint of the technical keyword classification framework used to describe technical knowledge in the recruitment field, classify each technical keyword, and select the corresponding multi-round question prompt word template based on its respective category, and perform multi-round multi-angle questioning on the large language model to obtain multiple question and answer fragments and context samples generated around each technical keyword. The technical keyword classification framework includes at least one or more categories such as programming language, operating system, and tool name.

[0020] S130, a word sense matching model is trained on the basis of a refined training corpus generated from the multiple question-and-answer fragments and context samples to represent the semantic relationships of technologies in the recruitment field, so that the word sense matching model can output the word vector representation of the technology keywords and calculate the semantic similarity between different texts based on the word vector representation.

[0021] Understandably, in recruitment scenarios, traditional general-purpose word vector models are usually trained on open-domain corpora. These models primarily rely on word co-occurrence patterns to learn semantic relationships, without deliberately distinguishing specific semantic roles such as job titles, skill sets, and tool names in the recruitment field. This easily leads to semantic drift and anisotropy problems: on the one hand, because Java and JavaScript often appear in similar technology stack descriptions, they are mistakenly grouped very close in the word vector space, and the model will consider them almost equivalent; on the other hand, Java development engineers and Java testing engineers share a large amount of vocabulary, and general-purpose models find it difficult to recognize the essential differences between development and testing in terms of job responsibilities, resulting in a high similarity between the two, thus causing misjudgments in job recommendations and resume screening. To address the aforementioned issues, we first utilize named entity recognition technology to extract key technical terms that truly embody technical semantics from job descriptions and resumes, filtering out a large amount of noise text irrelevant to skills. Then, we employ a large language model to perform refined classification and multi-round question-and-answer generation on these technical keywords, constructing a refined training corpus that explicitly expresses skill meanings, applicable scenarios, and similarity / conflict relationships. Based on this corpus, we train a specialized word-meaning matching model, enabling the model to actively narrow down truly semantically similar technical terms in the word vector space while separating easily confused but semantically different technical terms, fundamentally improving the accuracy of text similarity analysis in the recruitment field.

[0022] For example, thousands of representative job descriptions and resume samples can be collected first. Domain annotators can then manually annotate the names of programming languages ​​(e.g., Java, C, Python), operating systems (e.g., Linux, Windows Server), databases (e.g., MySQL, Oracle), front-end frameworks (e.g., Vue, React), testing tools (e.g., JUnit, Selenium), operations and monitoring tools (e.g., Prometheus), and cloud platform products (e.g., Alibaba Cloud). Common nouns and adjectives unrelated to technology can be labeled as non-technical entities. These annotated samples can then be used to train a sequence labeling model, for example, using a pre-trained language model-based annotation structure, enabling the model to automatically learn language patterns such as version numbers, abbreviations, English characters, and contextual verbs associated with technical terms. After training, the named entity recognition model was deployed on a large-scale recruitment corpus, batch-processing millions of job descriptions and resumes. All technology-related target words were extracted from each text, and their forms were standardized. For example, "Java" in "Java development" and "Java programmer" were grouped into the same technical keyword, and "Vue" in "Vue front-end framework" was identified as a framework-related skill. This method significantly reduces the vocabulary space that needs to be considered during training and matching, preventing the model from being interfered with by noise in a large amount of redundant natural language descriptions. It also ensures that all subsequent similarity learning and knowledge distillation revolve around the skill words that truly affect job matching, making the final word vectors more closely aligned with the technical semantics relevant to the recruitment field.

[0023] For example, firstly, a classification framework for describing technical knowledge in the recruitment field is constructed. This framework can at least include categories such as programming languages, operating systems, and tool names, and can be expanded to more granular categories such as databases, front-end frameworks, back-end frameworks, testing tools, and cloud service products as needed. Then, the resulting set of high-frequency technical keywords is input into a large language model. Under the constraints of the classification framework, each technical keyword is categorized. For example, the large language model can determine that Java belongs to programming languages, Linux belongs to operating systems, Jenkins belongs to continuous integration tools, and Vue belongs to front-end frameworks, and output the category label and its confidence score. For keywords with insufficient confidence, manual review or further questions can be triggered to ensure the reliability of the classification results. After obtaining the categories, design corresponding multi-round question prompt templates for each category of technical keywords. For example, for the programming language category, you can first ask about the typical application scenarios of the language, such as Java being mostly used for backend development or enterprise-level applications. Then, ask about common supporting frameworks, differences from similar languages, such as the differences between Java and JavaScript in runtime environment, type system, and typical job positions, as well as typical requirements for candidates' years of experience and project experience in recruitment scenarios. For the tool name category, you can conduct multi-round questions and answers around the tool's functional positioning, applicable scenarios, differences from similar tools, and what job roles it usually corresponds to when it appears in a resume. The answers to the multi-round questions are saved together with the original technical keywords and their context fragments to form question-and-answer fragments and context samples. These samples often explicitly contain statements such as Java is more inclined towards backend development, while JavaScript is usually used for browser-side scripts, Java development engineers are responsible for implementing business functions, while Java test engineers mainly design test cases and automation scripts, thus clearly marking the similarities and differences between easily confused skills at the corpus level. This process distills the expertise implicit in the large language model into the structured training corpus, which not only enhances the domain relevance of the corpus but also explicitly encodes the semantic boundaries between skills.

[0024] For example, a pre-trained language model can be used as the encoder to learn subtle semantic distinctions by constructing various types of training samples. For instance, in a refined corpus, different expressions describing the same skill can be considered semantically similar samples. For example, sentences like "Java backend development," "Java Web development engineer," and "proficient in using Spring and MyBatis frameworks for enterprise application development" can be used as similar samples, allowing the model to learn that these expressions all point to a relatively consistent skill direction. Simultaneously, easily confused but semantically different skill pairs can be constructed as semantically dissimilar samples, such as "Java programming development" and "Java automated testing," "Java" and "JavaScript," and "frontend framework Vue" and "backend framework Spring." During training, the loss function is designed so that after encoding these sentences or words, the distance between the former is compressed, while the distance between the latter is increased, thus forming a more reasonable geometric structure in the vector space. After training, the semantic matching model can be applied to actual job descriptions and resumes. First, a named entity recognition model is used to extract technical keywords. Then, the trained model encodes each technical keyword and its local context, resulting in a set of technical vector representations. Next, a matching strategy is used to compare the skill vectors on the job description and resume sides. For example, the similarity of corresponding skills is calculated and weighted according to skill category, frequency of occurrence, and job importance, thus obtaining a technical similarity score between the job description and resume. In this way, in actual matching scenarios, when a job requires a Java developer, the model significantly increases the similarity with resumes of Java backend developers familiar with Spring and Spring Boot, while decreasing the similarity with resumes focused on Java automation testing or functional testing. For frontend engineer (Vue) positions, the model prioritizes matching candidates with experience in frontend frameworks such as Vue and React, avoiding being misled by resumes with only Java backend development experience, thereby significantly reducing the false positive rate.

[0025] In one embodiment, it further includes: A co-occurrence graph is constructed based on a corpus of recruitment-related language. The nodes of the co-occurrence graph are technical keywords, and the edge weights are used to characterize the co-occurrence strength of technical keywords in job description texts and resume texts. Combining the co-occurrence graph and user interaction information in the recruitment field, multiple groups of technical keywords with semantically similar relationships are identified. Technical keywords belonging to the same technical keyword group are determined as a set of skill synonyms. In the refined training corpus, synonym replacement is performed on the technical keywords to generate an extended training corpus, so as to obtain an enhanced training corpus set. The word sense matching model is trained or retrained based on the enhanced training corpus to obtain a target word sense matching model for recruitment scenarios. When calculating the text similarity between the job description text to be matched and the resume text to be matched, the target semantic matching model is invoked to obtain the word vector representations of the technical keywords in the job description text to be matched and the resume text to be matched. Based on the word vector representations, the technical semantic similarity between the job description text and the resume text is calculated, so that the technical semantic similarity is used as a component of the job and resume matching score or matching score.

[0026] Understandably, by leveraging the co-occurrence relationships of skill terms in a large number of job descriptions and resumes, combined with real user interactions within the recruitment system, a set of synonyms for skills that are semantically equivalent or highly similar in recruitment contexts can be automatically mined. Then, by augmenting the training corpus through synonym replacement, the word-meaning matching model learns to group these synonyms together in the vector space. This addresses two typical problems: first, different companies may use different names for the same skill, such as front-end framework, Vue, React, or SPA framework, which may be scattered in a general model; second, there are implicit mapping relationships between certain skills, such as from Vue to front-end framework to React. Without explicit enhancement, the model often can only infer ambiguous relationships through limited co-occurrence, making it susceptible to noise interference. By constructing a co-occurrence graph, overlaying user interaction weights, forming a set of skill synonyms, amplifying the training data with synonyms, and finally training a word meaning matching model, the model not only utilizes statistical co-occurrence information but also introduces a higher-level signal from hiring decisions. This allows the model to learn which skills can be considered similar or substitutable from a recruitment perspective, thus making the calculation of job and resume similarity more consistent with the intuitive judgment of human resource recruitment.

[0027] For example, the technical keywords extracted in the previous steps are first used as nodes in the graph structure, with each node corresponding to a skill term appearing in job descriptions or resumes. Then, the co-occurrence relationships between these technical keywords are statistically analyzed in a massive corpus. For instance, if "Java" and "Spring" frequently co-occur in the same section of a job description, or "Vue" and "front-end engineer" repeatedly co-occur in the project experience section of numerous resumes, an edge can be created between these two nodes in the co-occurrence graph, and the edge weight can be set to the corresponding co-occurrence strength. In practice, co-occurrence windows can be defined at the sentence, paragraph, or sliding window levels. For example, multiple sets of technical keywords appearing in the same paragraph or a fixed-length text window are considered as one co-occurrence. Then, the edge weights are normalized according to the number of co-occurrences, node mutual information, or other statistical correlation indicators. By constructing a co-occurrence graph on the entire recruitment corpus, clear skill clusters can be automatically presented in the graph structure. For example, Java, Spring, MyBatis, and backend development constitute one cluster, while Vue, React, front-end frameworks, and single-page applications constitute another.

[0028] For example, various interaction data can be extracted from the recruitment system. For instance, when screening resumes, companies often place Vue front-end engineers and those familiar with React in the same candidate pool; candidates frequently apply for both Vue and React front-end development positions; or in system recommendations, if a job requires familiarity with mainstream front-end frameworks, users' click rates for resumes with Vue and React experience are very similar. These behaviors implicitly suggest that Vue and React have a strong substitutive relationship for many positions in recruitment scenarios. This embodiment can convert these interaction records into additional weights, increasing the weights of relevant edges in the co-occurrence graph. For example, if a skill pair frequently appears in resumes matching high-conversion-rate positions, the weight of that edge is increased; conversely, for skill pairs that frequently appear in negative feedback scenarios, such as a job requiring Java back-end development but the system recommending junior PHP developers, followed by frequent rejections by the company, the edge weights of these skill pairs in the graph can be decreased. Subsequently, community partitioning or clustering algorithms can be used to identify multiple closely related groups of technical keywords on the adjusted co-occurrence graph. For example, one group might contain JS, JavaScript, and front-end scripting languages, while another group might contain Linux system operation and maintenance, Linux administration, and Linux platform maintenance engineers. For these groups, the skill words within each group are considered as a set of synonyms for skills that are highly similar or substitutable in recruitment semantics. Compared to relying solely on plain text co-occurrence, this set of synonyms more closely reflects the real-world usage scenarios of recruitment decisions and can significantly reduce the misclassification of skills that only occasionally co-occur in certain documents but are not substitutable in recruitment as synonyms.

[0029] For example, the obtained skill synonym set can be used to automatically expand the refined training corpus already used to describe the semantic relationships of technologies. The target technology keywords in each training sample can be scanned to see if they belong to a certain skill synonym set. If they do, when generating new training samples, the technology keyword is replaced with other skill words from the same set. For instance, in a sample describing a candidate as familiar with Vue front-end development and able to independently build a single-page application, if it is identified that Vue and React are strong substitutes for each other in a certain type of role, a new training sample can be constructed: a candidate is familiar with React front-end development and can independently build a single-page application, while keeping the original label unchanged. Similarly, for synonym sets such as Linux system administration, Linux operations engineer, and Linux platform maintenance, multiple variations can be generated through replacement. In this way, a single original sample is automatically expanded into multiple samples with equivalent semantics but different superficial word forms, allowing the word meaning matching model to repeatedly see skill combinations with different names and similar responsibilities during the training phase. It significantly alleviates the problem of certain expressions appearing too infrequently in the original corpus, allowing the model to fully perceive that "the same job requirement can be expressed in different natural languages," thereby actively narrowing the vector representations of these expressions in the word vector space without changing the original supervision labels.

[0030] For example, the aforementioned expanded training corpus can be used to perform initial training or incremental retraining of the word-semantic matching model, allowing the model to absorb the knowledge contained in the skill synonym set during parameter updates. Paired samples generated by synonym replacement can be considered semantically equivalent positive examples, while skill combinations with significantly different responsibilities or opposite job directions can be considered negative examples. For instance, Vue front-end development experience and React front-end development experience are equivalent positive examples, while Vue front-end development experience and Java back-end development experience are considered highly discriminative samples for certain specific job roles. During training, by designing an appropriate loss function, the model can compress the word vector representations of these equivalent samples more closely, and widen the vector distance between samples with vastly different responsibilities. Since the training corpus contains a large number of variants obtained through synonym replacement, the model will gradually learn that even if a resume uses React while the job description mentions Vue, if the surrounding context describes similar responsibilities—for example, both emphasizing single-page applications, component-based development, and proficiency in modern front-end engineering tools—then from a vector perspective, these two skill sets should be similar. This training process implements a mechanism that inversely constrains the vector space structure from the set of synonyms, enabling the target word sense matching model to more accurately express the substitutability relationships between skills in the recruitment field, rather than being influenced solely by pure statistical co-occurrence.

[0031] For example, the aforementioned trained target semantic matching model can be applied to an intelligent recruitment system. Whenever it's necessary to recommend candidates for a specific position or evaluate the match between a resume and the job, the named entity recognition model described earlier is first used to extract technical keywords from the job description, such as Java backend, Spring Boot, microservice architecture, and distributed transactions. Simultaneously, candidate technical terms are extracted from the resume, such as Java microservices, Spring Cloud, and distributed system development experience. Then, these technical keywords and their local context are input into the target semantic matching model, which outputs their representations in vector space. The skill vectors from the job description and resume are then compared pairwise to obtain the technical semantic similarity between each pair of skills. Next, based on the importance of the skill in the job description, its position (e.g., core skills and bonus points required for the job), and the depth of the skill demonstrated by the candidate in their resume (e.g., project experience, years of experience), these similarities are weighted and fused to obtain the overall technical semantic similarity score between the job and the resume. This score can be used directly as the job matching score or combined with scores from other dimensions as a component of the comprehensive matching score. The model trained using the aforementioned synonym expansion and graph structure constraints demonstrates a significant advantage at this stage. For example, for front-end engineer positions requiring familiarity with mainstream front-end frameworks, if the resume states that the candidate primarily uses React and has also used Vue in projects, the model will give a high technical semantic similarity score. However, for resumes that only state familiarity with Photoshop and UI design tools, the technical similarity score will be significantly lower. This effectively reduces missed matches even when the tools used are different but the essential skills are the same, and also avoids false matches where the co-occurrence frequency is high but the responsibilities differ greatly.

[0032] In one embodiment, it further includes: Construct an annotated training set containing text from the recruitment field, and annotate multiple job description texts and multiple resume texts with technical terms related to the business domain; The sequence labeling model is trained based on the labeled training set so that it can identify technology-related target words in unlabeled text and output them as technology keywords, thereby obtaining the named entity recognition model.

[0033] In one embodiment, the step of inputting a set of high-frequency technical keywords into a large language model, and determining the category of each technical keyword under the constraints of a technical keyword classification framework used to describe technical knowledge in the recruitment field, includes: High-frequency technical keywords and their typical contextual fragments in the original corpus are input into the large language model. The large language model then outputs the target category label and confidence score of each technical keyword from the category set defined by the technical keyword classification framework. When the confidence level is lower than a preset threshold, a manual review or secondary questioning process is triggered to correct the target category label of the technical keyword. When the technical keyword is a programming language category, a multi-round question sequence is generated to guide the large language model to output the applicable scenarios, common supporting frameworks, commonly used libraries, typical job responsibilities, and competency level classifications of the programming language; and / or, When the technical keyword is a tool name category, generate a multi-round question sequence to guide the large language model to output the tool's functional positioning, common use cases, applicable job positions, and differences from other tools; and / or, When the technical keywords are in the test-related category, a multi-round question sequence is generated to guide the large language model to output test types, typical test steps, coverage objects, and collaboration methods with development positions. The training samples in the refined training corpus consist of answers generated from each multi-round question sequence, original technical keywords, and contextual fragments.

[0034] In one embodiment, it further includes: In job description text and resume text, multiple technical keywords that are in the same sentence, the same paragraph or a preset window range are regarded as a co-occurrence relationship, and the initial edge weight of the keyword pair is calculated based on the number of co-occurrences or statistical relevance. Based on the co-occurrence graph and combined with statistical information from the interaction logs of the recruitment system, the edge weights of technical keyword pairs that frequently appear simultaneously in user filtering, collection, application, and interview behaviors are increased, while the edge weights of technical keyword pairs that appear in negative feedback samples are decreased, in order to obtain the adjusted co-occurrence graph. In the adjusted co-occurrence graph, multiple technical keyword clusters are obtained based on edge weight threshold, community partitioning algorithm or clustering algorithm, and multiple technical keywords in the same technical keyword cluster are identified as a set of skill synonyms; When generating extended training corpora, for target technical keywords in refined training corpora, multiple sample variants are generated by replacing them with other technical keywords in the corresponding skill synonym set, while keeping the original semantic labels unchanged, so as to form training samples that are semantically equivalent but have different surface word forms.

[0035] Understandably, one approach is to first use statistical co-occurrence relationships from job descriptions and resumes to connect frequently appearing skill words into a graph. Then, this graph is corrected using real recruitment interactions. Finally, clustering is performed on the graph to obtain a set of highly relevant or even interchangeable skill synonyms from a recruitment perspective. These synonyms are then systematically injected into the training corpus. By replacing synonyms, a large number of semantically equivalent sample variants are generated, allowing the word-meaning matching model to naturally classify these skills as belonging to the same category during the training phase. The goal is to enable the model to provide more reasonable similarity judgments for skill groups such as Vue and React, Linux system administration and Linux backend development and Java microservice development when matching actual job positions, while avoiding mistaking skill pairs that only appear occasionally in the text as synonyms.

[0036] For example, firstly, the list of technical keywords in each text has been obtained through named entity recognition in the previous steps. These technical keywords are then grouped according to a certain co-occurrence window. For instance, all technical words within a natural sentence can be considered as a co-occurrence relationship, or the text can be scanned by natural paragraphs or by a fixed-length sliding window. Within each window, all distinct technical keywords are recorded as co-occurring once between each pair. Then, the entire recruitment corpus is traversed, accumulating the co-occurrence frequency of each pair of technical keywords. Appropriate statistical relevance indicators are selected based on actual needs. For example, co-occurrence frequency can be used directly, or indicators that better highlight the correlation beyond random collisions, such as point mutual information, can be used. These statistical values ​​are then normalized and used as the initial weights of the edges in the co-occurrence graph. For instance, if "Java" and "SpringBoot" appear frequently together in a large number of backend job descriptions and resumes, the edge weight between them will be high; while "Java" and "Photoshop" may occasionally co-occur in some full-stack or cross-functional jobs, but with a lower frequency, the initial weight of this edge will be significantly lower than the former.

[0037] For example, various interactive behaviors can be collected from the recruitment system, such as companies selecting certain resumes to enter the candidate pool and focus on when screening candidates, candidates clicking to save or apply after browsing the job postings, employers giving statuses such as scheduling interviews, passing, and rejection during the interview stage, as well as records of positive and negative feedback from companies regarding the recommendation results. For skill pairs that frequently appear together in user screening, collection, application, and successful interview samples—for example, if a job requires familiarity with mainstream front-end frameworks like Vue or React, and the resumes of hired or final-interview candidates often include "proficient in React, with experience in multiple single-page application projects"—then the edge weights between Vue and React can be increased in the co-occurrence graph. This indicates that these two skills are often considered close or substitutable abilities in recruitment decisions. Conversely, for skill pairs appearing in negative feedback samples—for example, if a job explicitly requires Java backend development experience, but the system has recommended numerous resumes with only basic PHP development experience, and these resumes have been repeatedly marked as completely mismatched or automatically eliminated by the company—then the edge weights between Java and PHP can be decreased in the co-occurrence graph. In some scenarios, these edges can even be directly suppressed to prevent the model from mistakenly treating them as close skills. Through this adjustment of the co-occurrence graph, For example, a threshold for edge weights can be set first, retaining only edges with weights greater than this threshold to filter out skill relationships that co-occur only in a very small number of instances or only appear in noisy text, reducing interference during subsequent clustering. Then, on this sparse but more reliable graph, community partitioning algorithms, graph clustering algorithms, or other aggregation methods suitable for large-scale graph data are used to identify several subsets of nodes with tight internal connections and relatively weak external connections. Each subset is a cluster of technical keywords; technical keywords within the same cluster are highly correlated in text co-occurrence and have been repeatedly verified in recruitment interactions to be easily accepted or considered together.

[0038] For example, for each sample in the refined training corpus, the system can first identify the target technical keywords that appear within them, determine which skill synonym set each keyword belongs to, and then replace each skill keyword with other members from the same set to generate a new sample. For instance, the original sample might be a job requirement of being familiar with the Vue front-end framework and able to independently develop single-page applications, with the corresponding label highly matching the front-end development engineer position. When it's identified that Vue and React belong to the same skill synonym set, a new sample can be generated requiring familiarity with the React front-end framework and the ability to independently develop single-page applications, while retaining the original label. Similarly, on the resume side, "responsible for Linux system maintenance" can be replaced with "responsible for Linux operation and maintenance management," and "Java microservice project development" can be replaced with "Java back-end service development," etc. When generating these sample variants, the original semantic labels are not changed; that is, the skill name is changed, but the overall ability requirements and job role remain the same. This continuously sends a signal to the model during the training phase that even if the specific wording differs, these samples should be considered as belonging to the same category in terms of recruitment semantics. In this way, the semantic matching model continuously encounters batches of samples with different superficial word forms but consistent labels during training, spontaneously narrowing their representations in the vector space. Ultimately, it treats different expressions in the same skill synonym set as semantically equivalent or highly similar vectors. Therefore, by utilizing co-occurrence graphs and interaction logs, it can finely distinguish between skill pairs that only occasionally appear together and those genuinely considered similar in recruitment decisions, resulting in a more reliable set of skill synonyms. Furthermore, by performing synonym replacement amplification on refined training corpora, the coverage of training data with different skill naming methods, spelling habits, and expression styles is significantly improved, allowing the model to learn to recognize scenarios with different names but equivalent abilities during training. Additionally, in subsequent job and resume similarity calculations, because the model has fully corrected the vector space structure on the augmented corpus, it will naturally give higher similarity scores for skill pairs like Vue and React, while lowering the similarity scores for word pairs with different responsibilities or semantics, such as Java backend development and Java testing, thus directly improving the accuracy and reliability of recruitment matching results.

[0039] In one embodiment, it further includes: Technical keywords and their context fragments from refined and expanded training corpora are constructed as positive and negative sample pairs. The parameters of the word sense matching model are constrained by a contrastive learning loss function or a ranking-based loss function, so as to reduce the distance between word vector representations of technical keywords that are semantically similar but have different word forms, and increase the distance between word vector representations of technical keywords that are semantically different but have similar word forms.

[0040] In one embodiment, it further includes: New recruitment data is aggregated according to a preset time period. The new recruitment data includes newly collected job description texts and resume texts. Named entity recognition and high-frequency technical keyword statistics are performed on the newly added recruitment data to obtain a set of new technical keywords; For the newly added set of technical keywords, a refined training corpus is repeatedly generated, and a co-occurrence graph and a set of skill synonyms are constructed to obtain an incremental training corpus set. Under the constraint of keeping the main direction of the existing word vector space unchanged, the incremental training corpus is used to incrementally train the new parameters or some parameters of the word sense matching model. During the incremental training process, the new word vectors are embedded into a subspace compatible with the existing word vector space through orthogonal projection, so as to keep the geometric relationship between the original technical keywords basically unchanged while adding new word representations, without having to retrain all historical corpora.

[0041] In some cases, during actual recruitment, some candidates' resumes deliberately pile up skill tags, such as listing a large number of keywords in the skills section like Java, Python, C, Go, Rust, Vue, React, Kubernetes, Hadoop, and Spark. However, only one or two languages ​​or frameworks actually appear in the project descriptions, and some even lack corresponding projects altogether. Typically, word vector models only look at the number of words listed, interpreting these resumes as having a broad technical range and being highly similar to multiple job positions. This causes recommendation systems to heavily favor resumes with piled-up skills, while truly solid but simply articulated candidates are suppressed. To address this issue, one embodiment further includes: After extracting technical keywords from the resume text, the skill list area and project experience area are distinguished. In the project experience area, the frequency and position of each technical keyword in the relevant project descriptions, job descriptions, and technical detail descriptions are statistically analyzed, as well as the co-occurrence pattern with verb phrases such as "responsible for," "implemented," "built," or "optimized." Based on the above statistical results, a skill support score is calculated for each technical keyword. The support score can comprehensively consider factors such as the frequency of the skill in the project, whether it appears near the description of technical difficulties, and whether it co-occurs with descriptions of specific indicator improvement or troubleshooting. When constructing training samples and calculating the technical semantic similarity between the job and the resume, the skill support score is mapped to a skill credibility weight. The word vector contribution of skill keywords with credibility below a preset threshold is reduced in weight, so that only skills truly supported by project experience are given higher weight in the job requirements.

[0042] Understandably, the above embodiments, based on existing semantic matching models, consider skill credibility vectors and re-evaluate technical keywords based on the support level of skill words in the project context. Skills that only appear in the skill list but lack project support are assigned lower credibility, and the word vector weights of low-credibility skills are attenuated when constructing training samples and calculating similarity. This way, even if a resume is filled with a large number of skill words, these skills will not contribute excessively to the matching process. For example, a resume might list proficiency in Hadoop and Spark, but the project experience is entirely in web management backend development using Spring Boot and MySQL. After support modeling, Hadoop and Spark will hardly appear in the project area, significantly reducing credibility. When matching for a big data platform development engineer position, even if the skill descriptions appear similar, the similarity on the technical vector side will be suppressed due to low credibility, thus reducing the chances of skill-stuffed resumes being mistakenly included in the candidate list.

[0043] In some cases in the recruitment field, certain skills or roles superficially share a large number of words, such as product manager and project manager, R&D engineer and implementation engineer. At the corpus level, they share many common words like requirements, clients, systems, testing, and deployment. However, their actual job responsibilities have clear opposing or complementary relationships. If only general word vectors are used, these roles are often seen as highly similar, leading to a situation where positions requiring in-depth R&D coding are filled with resumes that are more implementation-oriented or business-oriented. To address this issue, one embodiment further includes: Based on job categories and hiring department feedback records in the recruitment system, a set of skill pairs or job tag pairs with obvious opposing or complementary responsibilities are identified. These skill pairs or job tag pairs include, but are not limited to, requirements analysis and deep coding positions, implementation and delivery positions and architecture design positions, and pre-sales consulting positions and back-end development positions. By statistically analyzing the interaction information such as mutually exclusive applications, reasons for screening and elimination, and failed cases of job category switching between these positions, the above skill pairs or job tag pairs are marked as responsibility exclusion samples. When training the semantic matching model based on refined training corpus, responsibility exclusion samples are constructed as training pairs with negative relational labels. During the training process, a specific loss function is used to impose exclusion constraints on these samples, so that the distance between the skills or job tags marked as responsibilities-opposing in the word vector space is greater than a preset lower limit. When calculating the similarity between a job and a resume, if there are multiple high-weight skills marked as responsibilities-opposing between the overall resume vector and the job tag vector, the overall matching score of the resume is further reduced.

[0044] Understandably, when constructing the training corpus, a specific label of opposing responsibilities can be added, marking certain roles or skill pairs as having an exclusionary relationship. During training, this not only brings semantically similar skill vectors closer together but also explicitly uses negative constraints to push opposing skill vectors further apart, forming a more structured space in terms of recruitment semantics: it has both similarity and exclusionary relationships, rather than just two levels of similarity and dissimilarity. For example, for a position like Senior Backend Development Engineer, after modeling negative relationships, the system will maintain a significant distance between the skill vectors corresponding to roles like Implementation Engineer and Pre-sales Consultant. Even if these resumes frequently contain words similar to the job description, such as requirements analysis, system launch, and customer communication, the overall vector similarity will ultimately be suppressed due to the exclusionary relationship of responsibilities, thus allowing candidates who are truly engaged in deep coding to rank higher.

[0045] In some cases, the same skill is expressed very differently in Chinese and English resumes for internet positions, multinational corporations, or foreign companies. For example, container orchestration versus Kubernetes, distributed message queues versus Kafka, automated build tools versus Maven, etc. Furthermore, many Chinese resumes contain mixed descriptions, such as "familiar with the Spring Boot framework" versus "understanding microservice architecture and DevOps processes." General-purpose word vectors are often trained either solely on Chinese or solely on English corpora, making it difficult to accurately represent the correspondence between Chinese descriptions and English nouns within the same vector space. This leads to severe distortion in the similarity calculation between cross-language candidates and job positions. However, because the model performs normally in monolingual tests, this problem is difficult to identify in conventional evaluations. To address the above issues, one embodiment further includes: This study identifies mixed Chinese and English texts in recruitment corpora, aligning sentences containing English skill names with several preceding and following Chinese descriptions to form aligned segments. Co-occurrence patterns of English skill names and Chinese technical phrases are statistically analyzed, such as experience in Kubernetes cluster operation and maintenance, familiarity with container orchestration tools like Kubernetes, and responsibility for maintaining the company's Kafka message queue. Based on these aligned segments, cross-lingual technical alignment samples are constructed, treating each pair of Chinese technical phrases and English skill names as semantically equivalent positive sample pairs, while other English skill names with significantly different responsibilities are treated as negative samples. When training the semantic matching model, a cross-lingual alignment loss term is introduced, enabling the model to map semantically equivalent Chinese and English technical expressions to similar vector regions when encoding these samples, while mapping skills with different responsibilities to more distant regions. When calculating the similarity between job descriptions and resumes, even when the job description is in Chinese but the skills in the resume are primarily expressed in English, the two can still be accurately aligned using this cross-lingual technical concept subspace, thus avoiding underestimation of the matching degree due to language differences.

[0046] Understandably, the above method doesn't simply perform general Chinese-English alignment. Instead, it constructs a cross-lingual technical concept alignment subspace based on technical keywords in the recruitment field. By mining the co-occurrence relationships between Chinese descriptions and English skill names in a large-scale recruitment corpus, and combining alignment samples of job requirements and project descriptions, a specific cross-lingual constraint is introduced into the semantic matching model. This ensures that Chinese technical phrases and English skill names are close together in the same job context and separated in different responsibility contexts. For example, if a job posting states "familiarity with distributed message queues and preference for Kafka experience," and a resume states "responsible for Kafka cluster maintenance, optimizing message backlog issues, and using the Confluent toolchain," the cross-lingual alignment subspace will aggregate distributed message queues and Kafka in the same area, resulting in a high similarity between the technical vectors of the job posting and the resume. Conversely, a resume that only mentions RabbitMQ will maintain a moderate similarity to Kafka in the same subspace but will not be completely equivalent, thus reflecting differentiation in the ranking.

[0047] In some cases, during recruitment, the skill names of Java junior engineers, Java senior engineers, and Java architects highly overlap, and even their project keywords are very similar. However, the depth of knowledge, decision-making ability, and architectural perspective required for these positions are completely different. General word vector models typically only learn that they are all related to Java and Spring, crowding the vectors of all three together. This can easily lead to a resume for a Java junior engineer being mistakenly perceived as a high match for a Java architect position in the vector space. To address this issue, one embodiment further includes: The system automatically identifies feature words and phrases related to job levels from job description corpora, such as being responsible for the development of a module, independently undertaking the design of a system, leading a team to complete a project, being responsible for the selection and architecture design of key technologies, and being responsible for the final quality of the project. Jobs associated with these level features are then labeled into different level categories. When constructing training samples, texts of different levels of jobs under the same technology stack are constructed into sample pairs of the same skill but different levels, such as Java junior development engineer and Java architect. When training the semantic matching model, level decomposition constraints are introduced for these sample pairs, so that when the model encodes relevant text, skill-related components are mapped to a shared skill subspace, and job level-related components are mapped to an independent level subspace. When calculating the similarity between job and resume, the consistency of the technology stack is first calculated in the skill subspace, and then the degree of level matching is calculated in the level subspace. A penalty factor is applied to resumes with very similar technology stacks but obviously mismatched levels, thereby avoiding the situation where junior candidates are mistakenly rated as highly matched for senior positions.

[0048] Understandably, by explicitly introducing a job level axis into the word vector space, the skill content dimension and the job level dimension are decomposed and modeled. By mining responsibility words, performance indicators, and management responsibilities in job descriptions, information such as whether the candidate leads a team, is responsible for architecture design, and is responsible for key technology selection is injected into the training as level labels. This allows the model to differentiate levels when encoding the same technology stack, provided that the basic skills are similar. For example, a Java architect position requires experience in designing large-scale distributed systems, being responsible for core architecture and technology selection, and leading a team to complete system refactoring. On the other hand, a junior Java developer's resume mainly describes completing module development and fixing defects under the guidance of a mentor. In the skill subspace, both are highly related to Java, Spring, or microservices. However, in the level subspace, the former has a lot of strong responsibility features related to being responsible for, designing, and leading teams, while the latter mainly has subordinate features related to being guided and assisting. Finally, the comprehensive similarity calculation will significantly reduce the weight of cases where the levels do not match, making the recommendation results more in line with the expectations of enterprises for different levels of positions.

[0049] Please see Figure 2 One embodiment of the word vector text similarity analysis device in the recruitment field in this application may include: The identification unit 201 is used to extract technology-related target words from multiple job description texts and multiple resume texts in the recruitment domain corpus based on a pre-trained named entity recognition model, and obtain a set of technical words containing multiple technical keywords. Analysis unit 202 is used to input a set of high-frequency technical keywords into a large language model. Under the constraints of a technical keyword classification framework used to describe technical knowledge in the recruitment field, it determines the category of each technical keyword and selects the corresponding multi-round question prompt word template based on its category. It then performs multi-round, multi-angle questioning on the large language model to obtain multiple question-and-answer fragments and context samples generated around each technical keyword. The technical keyword classification framework includes at least one or more categories such as programming language, operating system, and tool name. Training unit 203 is used to train a word sense matching model on a refined training corpus generated based on the multiple question-and-answer fragments and context samples to represent the semantic relationships of technologies in the recruitment field, so that the word sense matching model can output the word vector representation of the technology keywords and can calculate the semantic similarity between different texts based on the word vector representation.

[0050] above Figure 2 The word vector text similarity analysis device for the recruitment field in this application embodiment has been described from the perspective of modular functional entities. The following is a detailed description of the word vector text similarity analysis device for the recruitment field in this application embodiment from the perspective of hardware processing. Please refer to... Figure 3 One embodiment of the word vector text similarity analysis device 300 in the recruitment field of this application includes: The system includes an input device 301, an output device 302, a processor 303, and a memory 304, wherein the number of processors 303 can be one or more. Figure 3 Taking a processor 303 as an example. In some embodiments of this application, the input device 301, output device 302, processor 303, and memory 304 can be connected via a bus or other means, wherein... Figure 3 Taking the example of a connection between China and Israel via a bus.

[0051] Among them, by calling the operation instructions stored in memory 304, processor 303 is also used to execute Figure 1 Any of the methods in the corresponding embodiments.

[0052] Please see Figure 4 , Figure 4 A schematic diagram of an embodiment of the electronic system provided in this application.

[0053] like Figure 4 As shown, this application provides an electronic system including a memory 410, a processor 420, and a computer program 411 stored in the memory 410 and executable on the processor 420. When the processor 420 executes the computer program 411, it can achieve... Figure 1 Any of the corresponding implementation methods in the embodiments.

[0054] Since the electronic system described in this embodiment is the equipment used to implement the word vector text similarity analysis device in the recruitment field of this application embodiment, those skilled in the art can understand the specific implementation method and various variations of the electronic system of this embodiment based on the method described in this application embodiment. Therefore, how the electronic system implements the method in this application embodiment will not be described in detail here. As long as those skilled in the art implement the method in this application embodiment, the equipment used is within the scope of protection of this application.

[0055] Please see Figure 5 , Figure 5 This is a schematic diagram illustrating an embodiment of a computer-readable storage medium provided in this application.

[0056] like Figure 5 As shown, this embodiment provides a computer-readable storage medium 500 on which a computer program 511 is stored. When the computer program 511 is executed by a processor, it can perform... Figure 1 Any of the corresponding implementation methods in the embodiments.

[0057] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0058] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0059] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0060] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0061] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0062] This application also provides a computer program product, which includes computer software instructions that, when executed on a processing device, cause the processing device to perform actions such as... Figure 1The process of the word vector text similarity analysis method in the recruitment field in the corresponding embodiment.

Claims

1. A method for word vector text similarity analysis in the recruitment field, characterized in that, include: Based on the pre-trained named entity recognition model, target words related to technology are extracted from multiple job description texts and multiple resume texts in the recruitment domain corpus, resulting in a set of technical words containing multiple technical keywords. A set of high-frequency technical keywords is input into a large language model. Under the constraints of a technical keyword classification framework used to describe technical knowledge in the recruitment field, each technical keyword is classified. Based on its category, a corresponding multi-round question prompt template is selected, and multiple rounds of multi-angle questions are performed on the large language model to obtain multiple question-and-answer fragments and context samples generated around each technical keyword. The technical keyword classification framework includes at least one or more categories such as programming language, operating system, and tool name. A word sense matching model is trained on a refined training corpus generated based on the multiple question-and-answer fragments and context samples to represent the semantic relationships of technologies in the recruitment field. This enables the word sense matching model to output word vector representations of the technical keywords and to calculate the semantic similarity between different texts based on the word vector representations.

2. The method according to claim 1, characterized in that, Also includes: A co-occurrence graph is constructed based on a corpus of recruitment-related language. The nodes of the co-occurrence graph are technical keywords, and the edge weights are used to characterize the co-occurrence strength of technical keywords in job description texts and resume texts. Combining the co-occurrence graph and user interaction information in the recruitment field, multiple groups of technical keywords with semantically similar relationships are identified. Technical keywords belonging to the same technical keyword group are determined as a set of skill synonyms. In the refined training corpus, synonym replacement is performed on the technical keywords to generate an extended training corpus, so as to obtain an enhanced training corpus set. The word sense matching model is trained or retrained based on the enhanced training corpus to obtain a target word sense matching model for recruitment scenarios. When calculating the text similarity between the job description text to be matched and the resume text to be matched, the target semantic matching model is invoked to obtain the word vector representations of the technical keywords in the job description text to be matched and the resume text to be matched. Based on the word vector representations, the technical semantic similarity between the job description text and the resume text is calculated, so that the technical semantic similarity is used as a component of the job and resume matching score or matching score.

3. The method according to claim 1, characterized in that, Also includes: Construct an annotated training set containing text from the recruitment field, and annotate multiple job description texts and multiple resume texts with technical terms related to the business domain; The sequence labeling model is trained based on the labeled training set so that it can identify technology-related target words in unlabeled text and output them as technology keywords, thereby obtaining the named entity recognition model.

4. The method according to claim 1, characterized in that, The process involves inputting a set of high-frequency technical keywords into a large language model, and then, under the constraints of a technical keyword classification framework used to describe technical knowledge in the recruitment field, determining the category of each technical keyword, including: High-frequency technical keywords and their typical contextual fragments in the original corpus are input into the large language model. The large language model then outputs the target category label and confidence score of each technical keyword from the category set defined by the technical keyword classification framework. When the confidence level is lower than a preset threshold, a manual review or secondary questioning process is triggered to correct the target category label of the technical keyword. When the technical keyword is a programming language category, a multi-round question sequence is generated to guide the large language model to output the applicable scenarios, common supporting frameworks, commonly used libraries, typical job responsibilities, and competency level classifications of the programming language; and / or, When the technical keyword is a tool name category, generate a multi-round question sequence to guide the large language model to output the tool's functional positioning, common use cases, applicable job positions, and differences from other tools; and / or, When the technical keywords are in the test-related category, a multi-round question sequence is generated to guide the large language model to output test types, typical test steps, coverage objects, and collaboration methods with development positions. The training samples in the refined training corpus consist of answers generated from each multi-round question sequence, original technical keywords, and contextual fragments.

5. The method according to claim 2, characterized in that, The co-occurrence graph constructed based on the recruitment domain corpus is used, where the nodes are technical keywords and the edge weights represent the co-occurrence strength of the technical keywords in the job description and resume texts, including: In job description text and resume text, multiple technical keywords that are in the same sentence, the same paragraph or a preset window range are regarded as a co-occurrence relationship, and the initial edge weight of the keyword pair is calculated based on the number of co-occurrences or statistical relevance. The process involves combining the co-occurrence graph and user interaction information in the recruitment field to identify multiple groups of technical keywords with semantically similar relationships. Technical keywords belonging to the same group are defined as a set of skill synonyms. Furthermore, synonym replacement is performed on these technical keywords in the refined training corpus to generate an expanded training corpus, resulting in an enhanced training corpus set, including: Based on the co-occurrence graph and combined with statistical information from the interaction logs of the recruitment system, the edge weights of technical keyword pairs that frequently appear simultaneously in user filtering, collection, application, and interview behaviors are increased, while the edge weights of technical keyword pairs that appear in negative feedback samples are decreased, in order to obtain the adjusted co-occurrence graph. In the adjusted co-occurrence graph, multiple technical keyword clusters are obtained based on edge weight threshold, community partitioning algorithm or clustering algorithm, and multiple technical keywords in the same technical keyword cluster are identified as a set of skill synonyms; When generating extended training corpora, for target technical keywords in refined training corpora, multiple sample variants are generated by replacing them with other technical keywords in the corresponding skill synonym set, while keeping the original semantic labels unchanged, so as to form training samples that are semantically equivalent but have different surface word forms.

6. The method according to claim 2, characterized in that, Also includes: Technical keywords and their context fragments from refined and expanded training corpora are constructed as positive and negative sample pairs. The parameters of the word sense matching model are constrained by a contrastive learning loss function or a ranking-based loss function, so as to reduce the distance between word vector representations of technical keywords that are semantically similar but have different word forms, and increase the distance between word vector representations of technical keywords that are semantically different but have similar word forms.

7. The method according to any one of claims 1 to 6, characterized in that, Also includes: New recruitment data is aggregated according to a preset time period. The new recruitment data includes newly collected job description texts and resume texts. Named entity recognition and high-frequency technical keyword statistics are performed on the newly added recruitment data to obtain a set of new technical keywords; For the newly added set of technical keywords, a refined training corpus is repeatedly generated, and a co-occurrence graph and a set of skill synonyms are constructed to obtain an incremental training corpus set. Under the constraint of keeping the main direction of the existing word vector space unchanged, the incremental training corpus is used to incrementally train the new parameters or some parameters of the word sense matching model. During the incremental training process, the new word vectors are embedded into a subspace compatible with the existing word vector space through orthogonal projection, so as to keep the geometric relationship between the original technical keywords basically unchanged while adding new word representations, without having to retrain all historical corpora.

8. A word vector text similarity analysis device for the recruitment field, characterized in that, include: The recognition unit is used to extract technology-related target words from multiple job description texts and multiple resume texts in the recruitment domain corpus based on a pre-trained named entity recognition model, and obtain a set of technical words containing multiple technical keywords. The analysis unit is used to input a set of high-frequency technical keywords into a large language model. Under the constraints of a technical keyword classification framework used to describe technical knowledge in the recruitment field, it determines the category of each technical keyword and selects the corresponding multi-round question prompt word template based on its category. It then performs multi-round, multi-angle questioning on the large language model to obtain multiple question-and-answer fragments and context samples generated around each technical keyword. The technical keyword classification framework includes at least one or more categories such as programming language, operating system, and tool name. The training unit is used to train a word sense matching model on a refined training corpus generated based on the multiple question-and-answer fragments and context samples to represent the semantic relationships of technologies in the recruitment field, so that the word sense matching model can output the word vector representation of the technology keywords and calculate the semantic similarity between different texts based on the word vector representation.

9. An electronic system comprising a memory and a processor, characterized in that, When the processor executes the computer program stored in the memory, it implements the steps of the word vector text similarity analysis method in the recruitment field as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the word vector text similarity analysis method in the recruitment field as described in any one of claims 1 to 7.

Images