Position information auditing method and device based on multi-modal information

Abstract

The invention discloses a position information auditing method and device based on multi-modal information. Comprising the steps of converting a position description text and enterprise entity information into a predefined structured data format, and generating a system cue word matched with the position description text to be audited based on structured enterprise information and structured position information, the system cue word comprises an auditing rule used for performing association analysis between the structured position information and the structured enterprise information on N target position auditing dimensions; the N target position auditing dimensions at least comprise consistency verification between structured position information and enterprise qualification text information, enterprise credibility evaluation based on enterprise historical release position records, and compliance identification of position description texts; and inputting the system cue word into a pre-trained large language model to obtain an audit result for the position description text to be audited. According to the invention, the identification accuracy of false recruitment positions can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of Internet, in particular to the technical field of Internet information service, and particularly relates to a position information auditing method, device, system, medium and product based on multi-modal information. BACKGROUND

[0002] With the popularization of Internet recruitment platforms, position information auditing has become a key link to protect the health of platform ecology and the rights and interests of job seekers. At present, general content auditing technologies based on artificial intelligence, such as keyword filtering and machine learning classification models, have been widely used in user-generated content (UGC) auditing in social, e-commerce and other fields.

[0003] However, when applied to this professional field of recruitment, the analysis dimension of such a solution is relatively single, and it is mainly limited to the surface features of the position description text itself, and it lacks deep insight into the information authenticity, rationality and background relevance, which leads to the inability to effectively identify strong hidden false recruitment positions. SUMMARY

[0004] Therefore, the embodiments of the present application provide a position information auditing method, device, system, medium and product based on multi-modal information, which can improve the identification accuracy of false recruitment positions.

[0005] In a first aspect, the embodiments of the present application provide a position information auditing method based on multi-modal information, which comprises: in the case of obtaining a position description text to be audited and its associated enterprise entity information, converting the position description text and the enterprise entity information into a predefined structured data format to obtain structured position information and structured enterprise information, wherein the enterprise entity information at least includes enterprise qualification text information and enterprise historical published position records; based on the structured enterprise information and the structured position information, generating a system prompt word matched with the position description text to be audited, wherein the system prompt word contains auditing rules for associating and analyzing the structured position information and the structured enterprise information in N target position auditing dimensions, and the N target position auditing dimensions at least include consistency verification between the structured position information and the enterprise qualification text information, enterprise credibility evaluation based on the enterprise historical published position records, and compliance identification of the position description text itself; inputting the system prompt word into a pre-trained large language model to enable the large language model to analyze in the N target position auditing dimensions based on the system prompt word, and obtain an auditing result for the position description text to be audited, wherein the auditing result at least contains a risk level.

[0006] Secondly, this application provides a job information review system based on multimodal information. The system includes: a conversion module, used to convert the job description text and associated enterprise entity information into a predefined structured data format to obtain structured job information and structured enterprise information, where the enterprise entity information includes at least enterprise qualification text information and historical job posting records; a generation module, used to generate system prompts matching the job description text to be reviewed based on the structured enterprise information and structured job information, where the system prompts include review rules for performing correlation analysis between the structured job information and structured enterprise information across N target job review dimensions, where the N target job review dimensions include at least consistency verification between the structured job information and enterprise qualification text information, enterprise credibility assessment based on historical job posting records, and compliance identification of the job description text itself; and a review module, used to input the system prompts into a pre-trained large language model, so that the large language model analyzes the N target job review dimensions based on the system prompts to obtain a review result for the job description text to be reviewed, where the review result includes at least a risk level.

[0007] Thirdly, embodiments of this application provide an electronic device, including: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, it implements the steps of the job information verification method based on multimodal information as described in the first aspect.

[0008] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer program instructions, which, when executed by a processor, implement the steps of the job information verification method based on multimodal information as described in the first aspect.

[0009] Fifthly, embodiments of this application provide a computer program product, which is stored in a non-volatile storage medium, and when executed by a processor, implements the steps of the job information verification method based on multimodal information as described in the first aspect.

[0010] In a sixth aspect, embodiments of this application provide a chip including a processor and a communication interface, the communication interface being coupled to the processor, the processor being used to run programs or instructions to implement the steps of the job information review method based on multimodal information as described in the first aspect.

[0011] This application provides a method, device, system, medium, and product for job information review based on multimodal information. It converts the job description text to be reviewed and its associated enterprise entity information into predefined structured data and generates system prompts containing specific review rules to guide a pre-trained large language model for analysis. This introduces multi-dimensional information and performs structured processing, providing a reliable data foundation for in-depth analysis. Based on this, the association analysis rules set in the system prompts effectively guide the large language model to perform cross-validation and association reasoning across multiple dimensions, including consistency verification, credibility assessment, and compliance identification. This structured analysis path effectively simulates the deep logical judgment process in manual review, possessing deep reasoning capabilities and achieving standardized and intelligent job information review. Furthermore, the large language model can effectively integrate multi-source information for multi-dimensional rule analysis, uncovering the implicit risks of authenticity, rationality, and background relevance in job information. It effectively identifies complex and hidden job information risks, improving the accuracy of identifying concealed fraudulent job postings, and ultimately outputs review results including risk levels. Attached Figure Description

[0012] 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.

[0013] Figure 1 This is a flowchart illustrating a job information verification method based on multimodal information provided in an embodiment of this application; Figure 2 This is an exemplary schematic diagram of structured job information provided in an embodiment of this application; Figure 3 This is an exemplary schematic diagram of structured enterprise information provided in an embodiment of this application; Figure 4 This is an exemplary schematic diagram of a first preset mapping table provided in an embodiment of this application; Figure 5 This is an exemplary schematic diagram of a preset job level weight table provided in an embodiment of this application; Figure 6 This is an exemplary schematic diagram of a job information review process based on multimodal information provided in an embodiment of this application; Figure 7 This is a schematic diagram of the structure of a job information verification system based on multimodal information provided in an embodiment of this application; Figure 8 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0014] 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, enabling those skilled in the art to better understand and implement the principles and spirit of this application. 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 based on the embodiments described herein without inventive effort are within the scope of protection of this application.

[0015] In this document, terms such as first, second, and third are used only to distinguish one entity (or operation) from another entity (or operation), and are not intended to require or imply any order or relationship between these entities (or operations).

[0016] Before describing the technical solutions provided in the embodiments of this application, in order to facilitate understanding of the embodiments of this application, this application first specifically explains the problems existing in the related technologies: In related technologies, general-purpose content moderation technologies based on artificial intelligence are commonly used in fields such as social media and e-commerce, employing keyword recognition and filtering or machine learning classification models trained on general corpora. However, recruitment, as a highly specialized vertical field, presents significant limitations with such general solutions. Specifically, general-purpose content moderation technologies typically only perform superficial compliance checks on job description text, such as filtering obviously prohibited words, but cannot deeply integrate and combine corporate entity information and external market information for multi-dimensional, in-depth correlation analysis and cross-validation. For example, it is difficult to identify the fraud risk implied in "a newly established company with low registered capital posting a remote job with a salary far exceeding the market average." Although general-purpose large language models possess powerful natural language understanding capabilities, without targeted domain knowledge injection and specific task guidance, directly applying them to professional moderation can easily lead to "illusion" problems, making it difficult to independently and reliably complete complex compliance reasoning.

[0017] Due to the limitations of the aforementioned technical solutions, the recruitment industry still heavily relies on manual review of job postings. Reviewers need to make subjective judgments based on a combination of job content, company information, market knowledge, and past experience. While this method can handle complex situations to some extent, its quality and efficiency are highly dependent on the reviewer's personal experience and expertise, making standardization and scalability difficult. Faced with a massive volume of job postings, manual review is costly, inefficient, and prone to inconsistent review standards due to subjective factors, becoming a bottleneck for platform operations.

[0018] In conclusion, existing technical solutions cannot simultaneously achieve high efficiency, low cost, and strong adaptability while ensuring the accuracy of the review process.

[0019] In view of the inventors’ above-mentioned research findings, the embodiments of this application provide a method, device, system, medium and product for job information review based on multimodal information, which aims to solve at least one of the above-mentioned technical problems.

[0020] The following description, in conjunction with the accompanying drawings, details the job information verification method based on multimodal information provided in this application through specific embodiments and application scenarios.

[0021] Figure 1 This is a flowchart illustrating a job information review method based on multimodal information provided in an embodiment of this application. The executing entity of this job information review method based on multimodal information can be a job information review system based on multimodal information (hereinafter referred to as "the system").

[0022] The following example, using a multimodal information-based job information verification system as the implementing entity, illustrates the multimodal information-based job information verification method of this application. It should be noted that the aforementioned implementing entity and application scenario do not constitute a limitation on this application.

[0023] like Figure 1 As shown, the job information review method based on multimodal information provided in this application embodiment may include steps 110-130.

[0024] Step 110: After obtaining the job description text to be reviewed and its associated enterprise entity information, convert the job description text and enterprise entity information into a predefined structured data format to obtain structured job information and structured enterprise information. The enterprise entity information includes at least enterprise qualification text information and enterprise historical job posting records. Step 120: Based on structured enterprise information and structured job information, generate system prompt words that match the job description text to be reviewed. The system prompt words contain review rules for performing correlation analysis between structured job information and structured enterprise information across N target job review dimensions. Step 130: Input the system prompt words into the pre-trained large language model so that the large language model can analyze the N target job review dimensions based on the system prompt words to obtain the review results for the job description text to be reviewed. The review results include at least the risk level.

[0025] The job information review method based on multimodal information provided in this application converts the job description text to be reviewed and its associated enterprise entity information into predefined structured data, and generates system prompts containing specific review rules to guide a pre-trained large language model for analysis. This introduces multi-dimensional information and performs structured processing, providing a reliable data foundation for in-depth analysis. Based on this, the association analysis rules set in the system prompts effectively guide the large language model to perform cross-validation and association reasoning in multiple dimensions such as consistency verification, credibility assessment, and compliance identification. This structured analysis path effectively simulates the deep logical judgment process in manual review, possessing deep reasoning capabilities and achieving standardized and intelligent job information review. Furthermore, the large language model can effectively integrate multi-source information for multi-dimensional rule analysis, uncovering the implicit risks of authenticity, rationality, and background relevance in job information. It effectively identifies complex and hidden job information risks, improves the accuracy of identifying hidden fraudulent job postings, and ultimately outputs review results including risk levels.

[0026] The specific implementation of the above steps will be described in detail below with reference to specific embodiments.

[0027] In step 110, after obtaining the job description text to be reviewed and its associated enterprise entity information, the job description text and enterprise entity information are converted into a predefined structured data format to obtain structured job information and structured enterprise information.

[0028] In step 110, the job description text can be job information filled in by HR, containing multiple job attribute information, such as job name, job description, job content, function type, salary range, work address, responsibilities, qualifications, team introduction, company benefits, and whether it is a remote position. The company qualification text information can be key text information extracted after recognizing the business license image, which may include, but is not limited to: company name, unified social credit code, registered address, company size, business scope, and establishment date. Historical job posting records can be obtained from the company database, which records the specific review status of job postings. The predefined structured data format can be standardized using JSON Schema to ensure the uniformity and parsability of data fields. Thus, through parallel extraction of multimodal information and standardized structured transformation, a unified and standardized data input format is provided for subsequent large language model analysis.

[0029] For example, converting job description text into a structured data format can yield... Figure 2 The structured job information shown; by converting enterprise entity information into a structured data format, we can obtain... Figure 3 The structured enterprise information shown.

[0030] According to an embodiment of this application, optionally, step 110, which converts the job description text and enterprise entity information into a predefined structured data format, may specifically include: parsing the business license image using optical character recognition (OCR) technology to extract enterprise qualification text information, wherein the enterprise qualification text information includes at least the registered address, business scope, establishment time, and enterprise size; parsing the job description text using natural language processing technology to extract job attribute information, wherein the job attribute information includes at least the work address, job description, job function, and salary range; and querying the enterprise's historical job posting records from the enterprise database, wherein the records include at least the number of times the enterprise's job postings were approved and rejected in the past.

[0031] Among them, enterprise qualification text information refers to key text data extracted from enterprise qualification documents such as business licenses; job attribute information refers to standardized job characteristic data parsed from job description text; job description may include, but is not limited to: job title, job content, responsibilities, qualifications, team introduction, company benefits, whether it is a remote job, etc.

[0032] Specifically, the system calls the OCR engine to recognize the business license image, extracts fields such as "registered capital: 5 million yuan" through regular expression matching, uses the named entity recognition model to extract entities such as salary range and work location from the job description text, and queries the company's historical job posting records from the enterprise database. For example, when the system recognizes the establishment date as "2023-01-01", it automatically calculates the company's establishment time and adds a query showing that the company has been approved 15 times and rejected 3 times in the past year.

[0033] In this embodiment, through the collaborative work of multimodal information extraction technology, heterogeneous raw data is converted into a unified structured data format, providing an accurate and standardized data foundation for subsequent correlation analysis.

[0034] Step 120 involves generating system prompts that match the job description text to be reviewed, based on structured enterprise information and structured job information.

[0035] In step 120, the system prompts include at least the role definition context, task objectives, review rules including review steps, and output format constraints. The role definition context is used to limit the large language model to the role of a compliance review expert for job information. The review steps are used to instruct the large language model to perform correlation analysis between structured job information and structured enterprise information in N target job review dimensions according to the review steps.

[0036] The system has a pre-built rule library for recruitment review scenarios. This library contains rule templates associated with N target job review dimensions. After reading the structured job information and structured company information generated in step 110, the system fills the placeholders in the rule templates with specific data values ​​and performs logical assembly to obtain system prompts. Specifically, this can be a combination of instantiated rules for each dimension, role definitions, output format requirements, etc. Then, the large language model can perform logical judgments on the association between the structured job information and structured company information based on the system prompts. For example, instead of solely considering salary level, it associates "salary range" with "job type," "company size," etc., to determine its reasonableness.

[0037] The N target job review dimensions include at least: consistency verification between structured job information and enterprise qualification text information, specifically cross-validation, such as comparing the work location in the structured job information with the registered address in the enterprise qualification text information to determine geographical rationality; enterprise credibility assessment based on the enterprise's historical job posting records, specifically historical behavior and statistical feature analysis, such as calculating the ratio of historical rejections to (historical approvals + historical rejections), and marking it as a high-risk history if the result exceeds a preset rejection rate threshold; and compliance identification of the job description text itself, specifically scanning the job description text to detect the presence of non-compliant keywords in the labor law keyword database. N is a positive integer.

[0038] For example, if the business scope in the structured enterprise information is "catering services", and the job description in the structured job information contains the words "Java development", the system will dynamically instantiate a rule template when generating the prompt words for the consistency verification dimension: determine whether the core work content in the job description is business-related to the company's business scope, and after instantiation: determine whether "Java development" is business-related to "catering services".

[0039] In step 130, the system prompt words are input into a pre-trained large language model so that the large language model can analyze the N target job review dimensions based on the system prompt words to obtain the review results for the job description text to be reviewed. The review results include at least the risk level.

[0040] In step 130, after receiving the system prompts, the large language model first loads its role settings, limiting its self-awareness to the role of a recruitment review expert. Then, following the review steps specified in the system prompts, the model sequentially performs correlation analysis across N target job review dimensions. During the analysis, an analysis result is obtained for each target job review dimension. Logical correlation and further analysis are then performed based on the analysis results of the N dimensions. After the analysis is complete, the large language model organizes the final review conclusion into a predefined structured data object according to the output format constraints specified in the system prompts, obtaining the final review result. In this way, the general capabilities of the large language model can be applied to specific tasks in the job review field, and the output review results can be directly used for business decisions, improving the practical value and processing efficiency of intelligent review and providing a technical foundation for replacing high-cost, low-efficiency manual review.

[0041] According to an embodiment of this application, optionally, the audit result is a structured data object, which includes: a risk level identifier (risk_level) field, a summary field (summary) field (text summary of the basis for risk judgment), a detailed_analysis field, and a recommendation field; The `detailed_analysis` field contains subfields corresponding to N target job review dimensions. Each subfield is used to populate a list of risk items under the corresponding dimension identified by the large language model. The `recommendation` field's value is a preset decision identifier associated with the risk level identifier. The `detailed_analysis` field contains the following three subfields: the `onsistency_risk` field associated with the consistency verification dimension, which indicates the specific points of inconsistency in the analysis information under the consistency verification dimension (left blank if none); the `company_credibility_risk` field associated with the credibility assessment dimension, which indicates the list of risk items analyzed based on the company's history and behavior under the credibility assessment dimension (left blank if none); and the `compliance_risk` field associated with the compliance identification dimension, which lists specific types of violations, such as recruitment discrimination and salary fraud (left blank if none).

[0042] For example, the audit result is: {risk_level:"high", detailed_analysis:{consistency_risk:["Business scope does not match"]}, recommendation:"reject"}.

[0043] Correspondingly, the output format constraints in the system prompts can also include the field names and descriptions of the structured data objects mentioned above, specifically including the following information: For example, please output your review results strictly according to the following JSON format: json { "risk_level": "High risk | Medium risk | Low risk", "summary": A short summary that clearly identifies the main risks or provides a safety conclusion. "detailed_analysis": { "consistency_risk": ["Analyze the specific points of inconsistency in the information; leave blank if none are found"], "company_credibility_risk": ["Risk identified based on company history and behavior analysis; leave blank if none exists"], "compliance_risk": ["List specific types of violations, such as 'recruitment discrimination', 'job requirements do not meet relevant regulations', etc. Leave blank if none are found"], "fraud_risk": ["Analyze whether there are any fraud traps, such as 'fee traps' or 'job fraud'. Leave blank if none are found"] }, "recommendation": "Clearly provide the review suggestion: 'Approved,' 'Rejected,' or 'Transfer to manual review.'" } The aforementioned fraud_risk field is also associated with the compliance identification dimension.

[0044] In this embodiment, a standardized output format ensures that the review results can be accurately parsed and processed by downstream systems, thus automating the review process. Simultaneously, the review results include detailed risk information and recommendations related to the job review. The `detailed_analysis` field requires the model to categorize and list specific risk points according to preset dimensions (such as consistency risk, reputation risk, etc.), making the source of risk clear and traceable, avoiding the ambiguity of free text summaries. Combined with explicit `recommendation` fields (such as "reject" or "transfer to manual review"), it provides direct and accurate action instructions for subsequent processing, improving the accuracy and efficiency of review decisions.

[0045] According to an embodiment of this application, optionally, the enterprise qualification text information may include the registered address and business scope, and the structured job information may include the work address, job description, job type, and salary range extracted from the job description text. The analysis based on system prompts on the consistency verification dimension in step 130 above may specifically include the following steps: The first verification result is obtained by determining the similarity between the keyword sequence of the job description and the text vector of the business scope in the semantic space, and determining whether the business relevance meets the requirements based on a preset relevance threshold. The second verification result is obtained by converting the work address and registration address into geographic coordinates based on geocoding services and determining whether the actual geographic distance between them exceeds the preset distance threshold set for remote positions. The third verification result is obtained by mapping the job type to the target category in the preset job classification system, querying the associated market salary distribution data based on the target category, obtaining the market salary range corresponding to the target category, and determining whether the salary range is within the market salary range corresponding to the target category. The third verification result is obtained by combining the first, second, and third verification results.

[0046] Specifically, the similarity in the semantic space can be the cosine similarity calculated after mapping the text and keyword sequences to a high-dimensional vector space using text embedding technology. The preset distance threshold is a spatial distance parameter dynamically configured based on the geographic information system and job characteristics. Both the preset distance threshold and the preset relevance threshold can be set according to specific needs, and this application does not impose specific limitations on them. The target category is the functional category to which the job description text's functional type belongs in a preset functional classification system. The preset functional classification system can be a pre-established standardized job classification framework, dividing jobs into major categories such as technical research and development, marketing, and administration according to functional attributes, with each major category further subdivided into smaller categories. The information consistency verification result can be obtained by combining the first verification result, the second verification result, and the third verification result.

[0047] For example, the Sentence-BERT model can be used to encode the business scope and job description into 384-dimensional vectors, and then the cosine similarity between them can be calculated. For instance, if the similarity score between "Java development" and "catering services" is 0.15, which is far below the preset relevance threshold of 0.6, then the business can be judged as irrelevant. Geographic rationality verification can be performed by converting the work address and registration address into latitude and longitude coordinates through a map geocoding interface, calculating the spherical distance using the Haversine formula, and comparing it with a preset distance threshold. For example, if the distance threshold for remote positions is set to 50 kilometers, and a distance of 120 kilometers between the work address and registration address is detected, it is marked as an anomaly. Salary matching verification first maps the job function to the target category in the job classification system, then queries the associated market salary distribution data, and finally determines whether the salary range is within a reasonable range. For example, "Java engineer" is mapped to the technical R&D category - backend development subcategory, and the market salary range is [8000, 15000]. If a salary of 20000 is detected, which exceeds the upper limit of the range, then a third verification result indicating that the salary is at risk is output.

[0048] In this application embodiment, a quantitative assessment of business relevance, geographical rationality, and salary matching degree is achieved through the fusion of multiple technologies, including semantic analysis, spatial computing, and data comparison. Addressing the limitation of existing technologies that can only perform surface text matching, this application employs a multi-dimensional cross-validation mechanism to identify potential risks that are difficult to detect from a single dimension, such as concealed fraud where the business scope does not match the job title but the text contains no sensitive words. This effectively improves the depth and accuracy of the review process and overcomes the deficiency in in-depth risk insight during job review in related technologies.

[0049] According to an embodiment of this application, optionally, the enterprise's historical job posting records include the number of times the enterprise's job postings have been approved and the number of times the job postings have been rejected. The analysis based on system prompts in step 130 can specifically include: determining the rejection rate based on the number of times the job postings have been approved and the number of times the job postings have been rejected, and determining the enterprise's behavioral risk score based on the rejection rate to obtain the credibility assessment result.

[0050] Among these, the enterprise behavior risk score is positively correlated with the review rejection rate. The review rejection rate can be the percentage of historical review rejections in the total number of reviews, serving as an important quantitative indicator of enterprise behavior risk. Optionally, the review rejection rate can be directly determined as the enterprise behavior risk score, or the enterprise behavior risk score can be determined by combining the review rejection rate with a preset mapping relationship.

[0051] For example, if a company has a history of 15 rejections and 3 rejections, with a rejection rate of 3 / (15+3)=16.7%, then according to the preset mapping relationship {rejection rate <5%: low risk, 5%-15%: medium risk, >15%: high risk}, it is judged to be at the medium risk level, and its associated score is 45 points.

[0052] In this embodiment of the application, in terms of credibility assessment, the enterprise's behavioral risk score is reasonably quantified based on the enterprise's historical review rejection rate. The historical behavior is quantified into an objective risk score. Compared with the existing method of judging enterprise credibility based on human experience, this enterprise credibility assessment method has consistency and repeatability, avoids judgment bias caused by differences in the experience of reviewers, and improves the quality and efficiency of job information review from the perspective of objective assessment.

[0053] According to an embodiment of this application, optionally, the enterprise's historical job posting records include the number of times the enterprise's job postings were approved in the past, the number of times the job postings were rejected in the past, and the timestamp of each review. The analysis based on system prompts in step 130 on the credibility assessment dimension may further include: assigning a time decay weight to each historical rejection event based on the timestamp, wherein the closer the historical rejection event is to the current time, the higher its weight value; determining the weighted sum of the number of historical rejections based on the time decay weight, and combining it with the sum of the number of historical approvals to obtain the dynamic rejection rate; acquiring and analyzing the time distribution characteristics of historical rejection events, and if the time distribution characteristics are used to characterize the concentrated distribution of historical rejection events within a preset time window, obtaining the behavior pattern recognition result used to characterize high-risk behavior patterns; inputting the dynamic rejection rate and the behavior pattern recognition result into a preset credibility assessment function, and outputting the credibility assessment result, wherein the credibility assessment result is used to quantify the enterprise's reputation score.

[0054] The time decay weight can be a weighting coefficient assigned to historical review rejection events based on a time decay function. This coefficient is negatively correlated with the interval between the event's occurrence and the current time. The event decay function is, for example, an exponential decay model: weight = e^(-λ×Δt), where λ is the decay factor, typically taken as 0.01, and Δt is the number of days since the current time. The dynamic rejection rate refers to the weighted rejection rate calculated after considering the time decay weight, which more accurately reflects the company's recent compliance performance. High-risk behavior patterns refer to the statistically significant clustering distribution characteristics of historical review rejection events over the time dimension.

[0055] For example, the system obtains each review event and its corresponding timestamp from the company's historical job posting records. For instance, a company has 10 historical review events, including 3 rejection events, occurring in January, March, and May 2024. A time decay weight is assigned to each rejection event based on a time decay function. The current time is June 1, 2024. The rejection event that occurred in January has a Δt=150-day weight, e^(-0.01×150)=0.22; the rejection event that occurred in March has a Δt=90-day weight, e^(-0.01×30)=0.74; and the rejection event that occurred in May has a Δt=30-day weight, e^(-0.01×30)=0.74.

[0056] The weights of the three rejection events were 0.22, 0.45, and 0.74, respectively, with a weighted sum of 1.41; the total number of reviews was 10, and the dynamic rejection rate was 1.41 / 10 = 14.1%.

[0057] Meanwhile, the system analyzes the temporal distribution characteristics of rejection events to detect whether there are high-risk behavioral patterns. For example, by using a sliding window algorithm, the system counts the distribution density of rejection events on a monthly basis. It detects that two rejection events occurred consecutively in March, and the distribution density was significantly higher than the historical average, which was marked as a high-risk behavioral pattern.

[0058] Finally, the dynamic rejection rate and the results of high-risk behavior pattern identification are input into a preset credibility assessment function. The assessment function applies additional risk weights to enterprises with high-risk behavior patterns. For example, a dynamic rejection rate of 14.1% corresponds to a base score of 65 points, and the presence of high-risk behavior patterns deducts 20 points, resulting in a final enterprise credibility score of 45 points.

[0059] In this application embodiment, by introducing a time-weighted mechanism and pattern recognition, risk assessment can dynamically reflect the latest situation of the enterprise. Compared with the problem that the existing technology simply statistically analyzes historical data and cannot reflect the changes in risk trends, this application can more sensitively identify the recent deterioration of risk conditions through time decay weight and distribution pattern analysis, thereby improving the timeliness and accuracy of risk assessment and providing technical support for timely detection of risk changes.

[0060] According to an embodiment of this application, optionally, the analysis based on system prompts in step 130 on the credibility assessment dimension may further include: determining the company's establishment duration based on the establishment time in the structured company information, and finding a first job posting quantity range associated with the establishment duration range based on the company's establishment duration range; querying the number of valid job postings from the company's historical job posting records, and determining a second job posting quantity range where the number of job postings falls, wherein the valid status indicates that the company is currently recruiting for the job postings; determining a first company stability score based on the difference between the upper limit of the first job posting quantity range and the upper limit of the second job posting quantity range; and obtaining the credibility assessment result by combining the company behavior risk score and the first company stability score.

[0061] Specifically, the establishment duration range refers to the development stage range divided according to the company's establishment time, such as 0-1 years as the startup stage, 1-3 years as the growth stage, and more than 3 years as the mature stage. The first job posting quantity range refers to the reasonable range of recruitment positions corresponding to a specific establishment duration range, which can be determined based on industry statistical data.

[0062] The aforementioned difference (referred to as the first difference for clarity) is obtained by subtracting the upper limit of the second job posting range from the upper limit of the first job posting range. The stability score of the first company is positively correlated with the first difference. A larger (positive) first difference indicates that the actual number of hires is far less than the reasonable upper limit, suggesting a conservative and stable hiring strategy, thus warranting a higher stability score. Conversely, a smaller or negative first difference indicates that the actual number of hires is close to or even exceeds the reasonable upper limit, suggesting the company may be engaging in aggressive or abnormal hiring practices, thus warranting a lower stability score. In other words, the stability score of the first company is a positively correlated score with the first difference. Specifically, it can be calculated based on the first difference using a preset positive correlation function or obtained by querying a first preset mapping table, which includes the mapping relationship between the first difference and the stability score.

[0063] Optionally, the preset positive correlation function can be a linear positive correlation function, a piecewise linear positive correlation function (setting multiple segmented intervals based on the difference, with different linear increment rates in each interval), a logic function, a threshold jump function (setting multiple difference thresholds, and the score jumps directly to the specified value when the difference exceeds the threshold), etc.

[0064] For example, a company established on January 1, 2023, will have been established for one year by January 1, 2024, falling into the startup phase. By querying a pre-defined table of the correspondence between establishment duration ranges and reasonable recruitment scales, the corresponding first job posting range is obtained. The reasonable recruitment scale range for a startup is determined to be [1, 5] positions, and for a growth phase, [5, 15] positions. The number of currently active job postings (8) is retrieved from the company's historical job posting records, determining its second job posting range to be [6, 10]. The difference between the upper limit of the first and second job posting ranges is then calculated. The upper limit for the startup phase is 5, and the upper limit for the actual recruitment range is 8. Based on this difference, the search query... Figure 4 The first preset mapping table shown determines that when the difference Δ=3, the first stability score is 50+5×3=65, corresponding to a medium stability risk.

[0065] In this embodiment of the application, by establishing a correlation model between the enterprise development stage and the reasonable recruitment scale, it is possible to identify abnormal recruitment behaviors that exceed the normal development pattern, such as the risk characteristics of large-scale recruitment by newly established companies. This enhances the audit system's ability to identify abnormal patterns, provides a quantitative basis for enterprise stability assessment, and enhances the scientific nature of risk assessment.

[0066] According to an embodiment of this application, optionally, the enterprise's historical job posting records are used to record the target enterprise's historical job posting records. The above step 130, which analyzes the credibility assessment dimension based on system prompts, may further include: mapping the enterprise size in the structured enterprise information to a preset enterprise size level, and obtaining a baseline job quantity range associated with the preset enterprise size level; querying the number of jobs posted by enterprises in a valid state from the enterprise's historical job posting records, and the job level corresponding to each job posted by an enterprise, wherein the valid state is used to indicate that the enterprise is recruiting for the job; obtaining the job weight value corresponding to each job posted by an enterprise by querying a preset job level weight table; determining the equivalent number of job openings corresponding to the target enterprise and the equivalent recruitment quantity range based on the number of job openings posted by the enterprise and the job weight value corresponding to each job posted by an enterprise; determining a second enterprise stability score based on the difference between the upper limit of the baseline job quantity range and the upper limit of the equivalent recruitment quantity range; and obtaining the credibility assessment result by combining the enterprise behavior risk score and the second enterprise stability score.

[0067] Specifically, the preset enterprise size level can be a tiered system based on the number of employees, such as small enterprises (1-50 people), medium-sized enterprises (51-200 people), and large enterprises (201 people or more). The preset job level weight table is used to assign corresponding job weight values ​​to different levels and types of positions within the enterprise. The equivalent number of recruitment positions is a standardized recruitment scale indicator calculated by weighting the job level weights.

[0068] The aforementioned difference (referred to as the second difference for clarity) is obtained by subtracting the smaller value from the larger of the upper limit of the baseline job quantity range and the upper limit of the equivalent recruitment quantity range. This second difference characterizes the match between the target company's current job postings and its size. The second company stability score is negatively correlated with the second difference. A smaller second difference indicates a higher match between the current job posting and the company size, thus requiring a higher stability score; conversely, a larger second difference indicates a lower match, thus requiring a lower stability score. In other words, the second company stability score is a negatively correlated score with the second difference. Specifically, it can be calculated based on the second difference using a preset negative correlation function or obtained by querying a second preset mapping table. This preset mapping table includes the mapping relationship between the second difference and the stability score. The preset negative correlation function can be, for example, a linear negative correlation function, a piecewise linear negative correlation function (setting multiple segmented intervals based on the difference, with different linear deceleration rates in each interval), or an exponentially decaying negative correlation function (the larger the second difference, the faster the score decays), etc.

[0069] For example, Company 1 publishes a job description text awaiting review. Company 1 is a medium-sized company with 150 employees, and its associated baseline number of positions ranges from [10, 30]. Query the number of currently valid job postings for Company 1 and the level of each position, and then... Figure 5 The preset job level weight table shown calculates the equivalent number of job openings. Recruiting 2 senior management positions (weight 3.0) and 5 general staff positions (weight 1.0) results in an equivalent number of 11 positions (2 × 3.0 + 5 × 1.0 = 11.0). The equivalent job opening range [10, 15] is determined, and the difference of 15 between the upper limit of the baseline job opening range and the upper limit of the equivalent job opening range is calculated. The second enterprise stability score is determined based on the magnitude of this difference.

[0070] In this embodiment, by introducing job level weights, both the quantity and quality of positions are considered when assessing recruitment scale. A quantitative comparison is made between the benchmark recruitment range corresponding to the company size and the actual recruitment scale weighted by job level. This establishes an analytical dimension combining the quantity and quality of positions in company stability assessment, overcoming the limitations of traditional methods that only examine the quantity of recruited positions while ignoring differences in job levels. For example, it can accurately identify anomalies such as over-recruiting senior positions, making recruitment scale assessment more scientific and reasonable, improving the accuracy of judging the rationality of corporate recruitment behavior, and enhancing the ability to identify complex risk patterns, thus providing more comprehensive data support for credibility assessment results.

[0071] According to an embodiment of this application, optionally, the analysis based on system prompts in step 130 above on the compliance identification dimension may include: scanning the job description text using keyword matching and semantic analysis technology to identify whether there is sensitive information related to recruitment discrimination and statements that violate labor laws, thereby obtaining a result of violation identification; obtaining the keyword vector of the job title in the job description text and the semantic vector of the job content in the job description text, and determining the similarity between the two in the embedding space; when the similarity is lower than a preset similarity threshold, determining that there is a risk of job fraud, thereby obtaining a result of fraud risk identification; and combining the result of violation identification and the result of fraud risk identification to obtain a result of compliance identification.

[0072] The preset similarity threshold can be associated with the function type. Matching similarity thresholds can be set for different function types in advance. The compliance identification results can include the identification results of illegal content and the identification results of fraud risks.

[0073] For example, by scanning job description text using keyword matching and semantic analysis techniques, expressions such as "male only" are detected and marked as a risk of recruitment discrimination. The BERT model is used to convert both the job title and job description into 768-dimensional vectors, and their cosine similarity in the embedding space is calculated. For instance, the similarity between the "job title" for a data clerk and the job content "business promotion" is 0.3. If the similarity threshold for clerical positions is set to 0.6, and the actual similarity of 0.3 is lower than the threshold, it is considered a risk of job fraud.

[0074] Compared to traditional shallow analysis schemes that mainly rely on keyword matching, this application uses a deep learning model to capture deep semantic features to identify compliance risks. By calculating semantic similarity, it can discover fraudulent behaviors that are different on the surface but are essentially the same, solving the adversarial problem of black market operators using evasive rhetoric and improving the ability to identify hidden violations.

[0075] As a concrete example, such as Figure 6 As shown, the job information review process based on multimodal information is as follows: 1. The process is triggered when an HR posts a new job; 2. The system initiates parallel processing to extract the information; 3. OCR identifies the HR company's business license information; 4. The information on the job posting by the HR is extracted; 5. The registered address, company size, business scope, and establishment date of the HR company identified in step 3 are converted into structured information, and the number of times the company's job postings were approved and failed in the past are added to obtain structured company information; 6. The job description, job type, salary range, work address, and whether it is a remote job extracted in step 4 are converted into structured information to obtain structured job information.

[0076] 7. Based on the structured information from #5 and #6, generate AI prompt words. For example: Role: You are a professional job posting compliance review expert, specializing in identifying false, irregular, and fraudulent information in job postings. You will need to conduct correlation analysis and in-depth reasoning based on multi-dimensional information.

[0077] Task: Please carefully review the following [Company Information] and [Job Information] to comprehensively assess whether the job information contains any potential risks.

[0078] Review dimensions and reasoning requirements: Information consistency verification: Compare the job description with the company's business scope to determine its relevance. A company whose business scope is unrelated to technology that is hiring a senior technical expert may pose a risk.

[0079] Verify that the "work location" matches the "company registration location".

[0080] Assess the match between "salary range" and "job type".

[0081] Company credibility assessment: Historical Behavior Analysis: Pay close attention to the company's "historical number of rejections." If the historical rejection rate is too high, it indicates that the company has a habit of posting non-compliant job openings, and this particular job posting requires extreme caution.

[0082] Company stability analysis: Analyze the "establishment time". Newly established companies (e.g., less than six months old) posting a large number of high-paying positions may pose a fraud risk. Analyze whether the "company size" matches the number and level of job openings.

[0083] Specific violation identification: Recruitment discrimination: making direct or implicit references to illegal restrictions such as gender, age, region, and ethnicity in the job description.

[0084] Violation of labor laws: The description implies or requires "voluntary waiver of social insurance", "alternating weeks of alternating overtime without overtime pay", "unpaid trial period", etc.

[0085] Fake / fraud traps: Salary fraud: Salaries are offered far above the market average with extremely low requirements, or "high rebates" are used as bait.

[0086] Fee traps: Some positions require training, uniforms, and medical examinations, but the fees are vague in the job description. However, before the job is hired, fees such as "training fees, deposits, uniform fees, and medical examination fees" are required.

[0087] Job fraud: The job title is seriously inconsistent with the job content (for example, recruiting "data clerk" but actually requiring "business promotion").

[0088] Risks of "pig butchering scams": The job descriptions are vague, emphasizing "simple operation, high returns, and can be done at home," which may lead to illegal activities such as online gambling and order brushing.

[0089] Input information: [Company Information] #5 Structured Information [Job Information] #6 Structured Information Output format requirements: Please strictly follow the following JSON format to output your review conclusions.

[0090] json { "risk_level": "High risk | Medium risk | Low risk", "summary": A short summary that clearly identifies the main risks or provides a safety conclusion. "detailed_analysis": { "consistency_risk": ["Analyze the specific points of inconsistency in the information; leave blank if none are found"], "company_credibility_risk": ["Risk identified based on company history and behavior analysis; leave blank if none exists"], "compliance_risk": ["List specific types of violations, such as 'recruitment discrimination', 'job requirements do not meet relevant regulations', etc. Leave blank if none are found"], "fraud_risk": ["Analyze whether there are any fraud traps, such as 'fee traps' or 'job fraud'. Leave blank if none are found"] }, "recommendation": "Clearly provide the review suggestion: 'Approved,' 'Rejected,' or 'Transfer to manual review.'" } Now, please begin reviewing the company information and job information provided above.

[0091] 8. Input the generated prompts into the large model; 9. The large model outputs the review conclusion according to the format; 10. For positions output as high-risk / medium-risk by the large model, manual review and verification are required. HR can be asked to provide more information to prove that the position is genuinely open for recruitment, such as proof of company job requirements, proof of work environment, etc.; 11. Based on the information in step 10, approve / reject the high-risk position information.

[0092] This embodiment utilizes the aforementioned systematic job verification solution for the recruitment industry to enhance the AI ​​big data model's knowledge in the recruitment vertical, thereby improving the model's reasoning, judgment, and verification capabilities to achieve automated verification. Specifically, based on the characteristics of the recruitment industry and the job information content, it extracts two structured information components: basic company characteristics and informational characteristics. The first component includes core information about the recruiting company, such as company size, registered address, market salary trends, and historical behavior records. The second component includes job description, job type, salary range, work address, and whether it is a remote position. Based on these two structured information components, AI prompts are dynamically generated. The overall solution outputs results with clear logical reasoning, making the verification results interpretable, especially for positions suspected of being high-risk, facilitating manual verification.

[0093] Corresponding to the method embodiments of this application, this application also provides a job information verification system based on multimodal information.

[0094] Figure 7 This is a schematic diagram of the structure of a job information verification system based on multimodal information provided in an embodiment of this application. For example... Figure 7 As shown, the job information review system 700 based on multimodal information may include: a conversion module 710, a generation module 720, and a review module 730.

[0095] The conversion module 710, upon obtaining the job description text to be reviewed and its associated enterprise entity information, converts the job description text and enterprise entity information into a predefined structured data format to obtain structured job information and structured enterprise information. The enterprise entity information includes at least enterprise qualification text information and historical job posting records. The generation module 720, based on the structured enterprise information and structured job information, generates system prompts that match the job description text to be reviewed. These system prompts include review rules for performing correlation analysis between the structured job information and structured enterprise information across N target job review dimensions. These N target job review dimensions include at least consistency verification between the structured job information and enterprise qualification text information, enterprise credibility assessment based on historical job posting records, and compliance identification of the job description text itself. The review module 730 inputs the system prompts into a pre-trained large language model, enabling the model to analyze the N target job review dimensions based on the system prompts to obtain review results for the job description text to be reviewed. These review results include at least a risk level.

[0096] The job information review system based on multimodal information provided in this application converts the job description text to be reviewed and its associated enterprise entity information into predefined structured data, and generates system prompts containing specific review rules to guide a pre-trained large language model for analysis. This introduces multi-dimensional information and performs structured processing, providing a reliable data foundation for in-depth analysis. Based on this, the association analysis rules set in the system prompts effectively guide the large language model to perform cross-validation and association reasoning in multiple dimensions such as consistency verification, credibility assessment, and compliance identification. This structured analysis path effectively simulates the deep logical judgment process in manual review, possessing deep reasoning capabilities and achieving standardized and intelligent job information review. Furthermore, the large language model can effectively integrate multi-source information for multi-dimensional rule analysis, uncovering the implicit risks of authenticity, rationality, and background relevance in job information. It effectively identifies complex and hidden job information risks, improves the accuracy of identifying hidden fraudulent job postings, and ultimately outputs review results including risk levels.

[0097] In some embodiments, the conversion module is specifically used to: parse the business license image using optical character recognition (OCR) technology to extract enterprise qualification text information, wherein the enterprise qualification text information includes at least the registered address, business scope, establishment time, and enterprise size; parse the job description text using natural language processing technology to extract job attribute information, wherein the job attribute information includes at least the work address, job description, job function type, and salary range; and query the enterprise's historical job posting records from the enterprise database, wherein the records include at least the historical number of times the enterprise's job postings were approved and the historical number of times they were rejected.

[0098] In some embodiments, enterprise qualification text information includes registered address and business scope, and structured job information includes work address, job description, job type, and salary range extracted from job description text. The review module is specifically used to: determine the similarity between the keyword sequence of the job description and the text vector of the business scope in the semantic space, and determine whether the business relevance meets the requirements based on a preset relevance threshold, to obtain a first verification result; convert the work address and registered address into geographic coordinates based on geocoding services, and determine whether the actual geographic distance between them exceeds a preset distance threshold set for remote positions, to obtain a second verification result; map the job type to the target category in a preset job classification system, and query the associated market salary distribution data based on the target category to obtain the market salary range corresponding to the target category, and determine whether the salary range is within the market salary range corresponding to the target category, to obtain a third verification result; combine the first verification result, the second verification result, and the third verification result to obtain an information consistency verification result.

[0099] In some embodiments, the enterprise's historical job posting records include the number of times the enterprise's job postings have been approved and the number of times the enterprise's job postings have been rejected. The review module is specifically used to: determine the rejection rate based on the number of times the enterprise's job postings have been approved and the number of times the enterprise's job postings have been rejected, and determine the enterprise's behavioral risk score based on the rejection rate to obtain a credibility assessment result. The enterprise's behavioral risk score is positively correlated with the rejection rate.

[0100] In some embodiments, the enterprise's historical job posting records include the number of times the enterprise's job postings have been approved and rejected in the past, as well as the timestamp of each review. The review module is specifically used to: assign a time decay weight to each historical rejection event based on the timestamp, wherein the weight of a historical rejection event more recent than the current time is higher; determine the weighted sum of the number of historical rejections based on the time decay weight, and combine it with the sum of the number of historical approvals to obtain the dynamic rejection rate; acquire and analyze the time distribution characteristics of historical rejection events, and if the time distribution characteristics are used to characterize the concentrated distribution of historical rejection events within a preset time window, obtain the behavior pattern recognition result used to characterize high-risk behavior patterns; input the dynamic rejection rate and the behavior pattern recognition result into a preset credibility evaluation function, and output the credibility evaluation result, wherein the credibility evaluation result is used to quantify the enterprise's reputation score.

[0101] In some embodiments, the audit module is further configured to: determine the company's establishment duration based on the establishment time in the structured company information, and find a first job posting quantity range associated with the establishment duration range based on the company's establishment duration range; query the number of job postings that are in a valid state from the company's historical job posting records, and determine a second job posting quantity range in which the number of job postings is located, wherein the valid state is used to indicate that the company's job postings are currently being recruited; determine a first company stability score based on the difference between the upper limit of the first job posting quantity range and the upper limit of the second job posting quantity range; and obtain a credibility assessment result by combining the company's behavioral risk score and the first company stability score.

[0102] In some embodiments, the enterprise historical job posting record is used to record the target enterprise's historical job posting records. The audit module is also used to: map the enterprise size in the structured enterprise information to a preset enterprise size level, and obtain the baseline job quantity range associated with the preset enterprise size level; query the number of enterprise job postings in a valid state from the enterprise historical job posting record, and the job level corresponding to each enterprise job posting, wherein the valid state is used to indicate that the enterprise job posting is currently recruiting; obtain the job weight value corresponding to each enterprise job posting by querying a preset job level weight table; determine the equivalent number of recruitment positions corresponding to the target enterprise and the equivalent recruitment quantity range based on the number of enterprise job postings and the job weight value corresponding to each enterprise job posting; determine the second enterprise stability score based on the difference between the upper limit of the baseline job quantity range and the upper limit of the equivalent recruitment quantity range; and obtain the credibility assessment result by combining the enterprise behavior risk score and the second enterprise stability score.

[0103] In some embodiments, the review module is specifically used to: scan job description text using keyword matching and semantic analysis technology to identify whether there is sensitive information related to recruitment discrimination and statements that violate labor laws, thereby obtaining a result for identifying illegal content; obtain keyword vectors of job titles and semantic vectors of job content in job description text, and determine the similarity between the two in the embedding space; when the similarity is lower than a preset similarity threshold, it is determined that there is a risk of job fraud, thereby obtaining a result for identifying fraud risk; and combine the result for identifying illegal content and the result for identifying fraud risk to obtain a result for identifying compliance.

[0104] In some embodiments, the audit result is a structured data object, which includes: a risk level identifier (risk_level) field, a summary field (summary) field (text summary of the risk judgment basis) field, a detailed_analysis field, and a recommendation field; wherein, the detailed_analysis field contains subfields corresponding to N target job audit dimensions, and each subfield is used to populate the risk item list under the corresponding dimension identified by the large language model; the field value of the recommendation field is a preset decision identifier associated with the risk level identifier.

[0105] The job information verification system based on multimodal information provided in this application embodiment can achieve... Figures 1-6 The various processes implemented in the method embodiments can achieve the same technical effect, and will not be described again here to avoid repetition.

[0106] Figure 8 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of this application.

[0107] like Figure 8 As shown, the electronic device 800 includes a memory 801, a processor 802, and a computer program stored in the memory 801 and executable on the processor 802.

[0108] In one example, the processor 802 described above may include a central processing unit (CPU), or an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0109] Memory 801 may include read-only memory (ROM), random access memory (RAM), disk storage media devices, optical storage media devices, flash memory devices, and electrical, optical, or other physical / tangible memory storage devices. Therefore, typically, memory includes one or more tangible (non-transitory) computer-readable storage media (e.g., memory devices) encoded with software including computer-executable instructions, and when the software is executed (e.g., by one or more processors), it is operable to perform the operations described in reference to the job information verification method based on multimodal information according to the embodiments of the first aspect of this application.

[0110] The processor 802 runs a computer program corresponding to the executable program code by reading the executable program code stored in the memory 801, in order to implement the job information review method based on multimodal information in the first aspect embodiment above.

[0111] In some examples, the electronic device 800 may also include a communication interface 803 and a bus 810. For example, Figure 8 As shown, the memory 801, processor 802, and communication interface 803 are connected through bus 810 and complete communication with each other.

[0112] The communication interface 803 is mainly used to enable communication between various modules, systems, units, and / or devices in the embodiments of this application. Input devices and / or output devices can also be connected through the communication interface 803.

[0113] Bus 810 includes hardware, software, or both, that couples components of electronic device 800 together. For example, and not as a limitation, bus 810 may include an Accelerated Graphics Port (AGP) or other graphics bus, an Enhanced Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), a Hyper Transport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a Micro Channel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-E) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local Bus (VLB) bus, or other suitable buses, or a combination of two or more of these. Where appropriate, bus 810 may include one or more buses. Although specific buses are described and illustrated in the embodiments of this application, this application considers any suitable bus or interconnection.

[0114] The electronic device provided in this application embodiment is capable of achieving Figures 1-6 The various processes implemented in the method embodiments can achieve the same technical effect, and will not be described again here to avoid repetition.

[0115] In conjunction with the job information verification method based on multimodal information in the above embodiments, this application embodiment can provide a computer storage medium for implementation. The computer storage medium stores computer program instructions; when these computer program instructions are executed by a processor, they implement any step in the above method embodiments.

[0116] In conjunction with the job information verification method based on multimodal information in the above embodiments, this application embodiment can provide a computer program product to implement it. This (computer) program product is stored in a non-volatile storage medium, and when executed by at least one processor, it implements any step in the above method embodiments.

[0117] This application also provides a chip, which includes a processor and a communication interface. The communication interface and the processor are coupled. The processor is used to run programs or instructions to implement the various processes of the above method embodiments and achieve the same technical effect. To avoid repetition, it will not be described again here.

[0118] It should be understood that the chip mentioned in the embodiments of this application may also be referred to as a system-on-a-chip, system chip, chip system, or system-on-a-chip, etc.

[0119] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0120] The functional blocks shown in the above-described block diagram 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, the elements of this application are programs or code segments used to perform the required tasks. Programs or code segments can be stored on a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried on a carrier wave. "Machine-readable medium" can include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, ROM, flash memory, erasable ROM (EROM), floppy disks, CD-ROMs, optical disks, hard disks, fiber optic media, radio frequency (RF) links, etc. Code segments can be downloaded via computer networks such as the Internet, intranets, etc.

[0121] It should also be noted that the exemplary embodiments mentioned in this application describe some methods or systems based on a series of steps or systems. However, this application is not limited to the order of the above steps; that is, the steps can be performed in the order mentioned in the embodiments, or in a different order, or several steps can be performed simultaneously.

[0122] The aspects of this disclosure have been described above with reference to flowchart illustrations and / or block diagrams of methods, systems (systems), and computer program products according to embodiments of this disclosure. It should be understood that each block in 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, a special-purpose computer, or other programmable data processing system to produce a machine such that these instructions, executable via the processor of the computer or other programmable data processing system, enable the implementation of the functions / actions specified in one or more blocks of the flowchart illustrations and / or block diagrams. Such a processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor, or a field-programmable logic circuit. It is also understood that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can also be implemented by special-purpose hardware performing the specified functions or actions, or can be implemented by a combination of special-purpose hardware and computer instructions.

[0123] The above description is merely a specific implementation 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 systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the protection scope of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the protection scope of this application.

Claims

1. A method for verifying job information based on multimodal information, characterized in that, include: Upon obtaining the job description text to be reviewed and its associated enterprise entity information, the job description text and enterprise entity information are converted into a predefined structured data format to obtain structured job information and structured enterprise information. The enterprise entity information includes at least enterprise qualification text information and enterprise historical job posting records. Based on structured enterprise information and structured job information, system prompt words are generated that match the job description text to be reviewed. The system prompt words contain review rules for performing correlation analysis between the structured job information and the structured enterprise information across N target job review dimensions. The N target job review dimensions include at least consistency verification between structured job information and enterprise qualification text information, enterprise credibility assessment based on the enterprise's historical job posting records, and compliance identification of the job description text itself. The system prompts are input into a pre-trained large language model, which then analyzes the N target job descriptions based on the system prompts to obtain the review results for the job description text to be reviewed. The review results include at least the risk level.

2. The method according to claim 1, characterized in that, Convert job description text and company entity information into a predefined structured data format, including: The business license image is analyzed using optical character recognition (OCR) technology to extract enterprise qualification text information, which includes at least the registered address, business scope, establishment time, and enterprise size. The job description text is parsed using natural language processing technology to extract job attribute information, which includes at least the work address, job description, job type, and salary range. The database retrieves historical job posting records from the enterprise database, including at least the number of times the job postings were approved and rejected.

3. The method according to claim 1 or 2, characterized in that, The enterprise qualification text information includes registered address and business scope; the structured job information includes work address, job description, job function type, and salary range extracted from the job description text; and analysis is performed on the consistency verification dimension based on system prompts, including: Determine the similarity in semantic space between the keyword sequence of the job description and the text vector of the business scope, and determine whether the business relevance meets the requirements based on a preset relevance threshold to obtain the first verification result; Based on the geocoding service, the work address and registration address are converted into geographic coordinates respectively, and it is determined whether the actual geographic distance between the two exceeds the preset distance threshold set for remote positions, thus obtaining the second verification result; The function type is mapped to the target category in the preset function classification system, and the market salary distribution data associated with the target category is queried to obtain the market salary range corresponding to the target category. It is then determined whether the salary range is within the market salary range corresponding to the target category, and a third verification result is obtained. By combining the first, second, and third verification results, the information consistency verification result is obtained.

4. The method according to claim 1 or 2, characterized in that, The company's historical job posting records include the number of times job applications were approved and rejected. Analysis is performed based on system prompts and on a credibility assessment basis, including: The audit rejection rate is determined based on the number of historical approvals and the number of historical rejections, and the corporate behavior risk score is determined based on the audit rejection rate to obtain the credibility assessment result. The corporate behavior risk score is positively correlated with the audit rejection rate.

5. The method according to claim 1 or 2, characterized in that, The company's historical job posting records include the number of times each job posting was approved and rejected, as well as the timestamp of each approval. Analysis is performed based on system prompts to assess credibility, including: Based on the timestamp, a time decay weight is assigned to each historical review rejection event, wherein the weight value is higher for historical review rejection events that are closer to the current time. Based on the time decay weight, the weighted sum of the historical rejection counts is determined, and combined with the sum of the historical approval counts, the dynamic rejection rate is obtained. Acquire and analyze the temporal distribution characteristics of the historical review rejection events. If the temporal distribution characteristics are used to characterize the concentrated distribution of historical review rejection events within a preset time window, obtain the behavior pattern recognition result used to characterize high-risk behavior patterns. The dynamic rejection rate and the behavioral pattern recognition result are input into a preset credibility evaluation function, and the credibility evaluation result is output. The credibility evaluation result is used to quantify the enterprise reputation score.

6. The method according to claim 4, characterized in that, The analysis based on system prompts in the credibility assessment dimension also includes: The establishment time of an enterprise is determined based on the establishment time in the structured enterprise information, and the first job posting quantity range associated with the establishment time range is found based on the establishment time range in which the establishment time of the enterprise is located. The number of job postings that are currently active in the company's historical job posting records is retrieved from the records, and the second job posting range in which the number of job postings falls is determined. The active status indicates that the company is currently recruiting for the job postings. The stability score of the first enterprise is determined based on the difference between the upper limit of the first job posting range and the upper limit of the second job posting range. The credibility assessment results are obtained by combining the corporate behavior risk score and the first corporate stability score.

7. The method according to claim 4, characterized in that, The enterprise's historical job posting records are used to record the target company's historical job postings. This data is analyzed based on system-suggested keywords in the credibility assessment dimension, and also includes: Map the enterprise size in the structured enterprise information to a preset enterprise size level, and obtain the baseline job quantity range associated with the preset enterprise size level; From the company's historical job posting records, query the number of jobs posted by the company that are in a valid state, and the job level corresponding to each job posted by the company. The valid state is used to indicate that the company is currently recruiting for the job. By querying the preset job level weight table, the job weight value corresponding to each job posted by a company can be obtained; Based on the number of job postings by a company and the job weight value corresponding to each job posting by a company, determine the equivalent number of job postings for the target company and the range of equivalent job postings it falls within. The second enterprise stability score is determined based on the difference between the upper limit of the baseline job quantity range and the upper limit of the equivalent recruitment quantity range. The credibility assessment result is obtained by combining the corporate behavior risk score and the second corporate stability score.

8. The method according to claim 1 or 2, characterized in that, Analysis based on system prompts is conducted on compliance identification, including: By scanning the job description text using keyword matching and semantic analysis techniques, the system identifies whether there is sensitive information related to recruitment discrimination or statements that violate labor laws, thus obtaining the results of the violation identification. Obtain the keyword vector of the job name in the job description text and the semantic vector of the job content in the job description text, and determine the similarity between the two in the embedding space. When the similarity is lower than a preset similarity threshold, it is determined that there is a risk of job fraud, and the fraud risk identification result is obtained. By combining the results of identifying illegal content and fraud risk, a compliance identification result is obtained.

9. The method according to claim 1, characterized in that, The audit result is a structured data object, which includes: a risk level identifier (risk_level field), a summary field (text summary field), a detailed_analysis field, and a recommendation field; The detailed_analysis field contains subfields corresponding to the N target job review dimensions, and each subfield is used to populate the risk item list under the corresponding dimension identified by the large language model; the recommendation field has a field value that is a preset decision identifier associated with the risk level identifier.

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

Images