A BPO customer consultation content analysis system for analyzing and managing a recruitment platform

By extracting recruitment entities from the recruitment platform analysis and management system, assessing acoustic stress and susceptibility to escaping, constructing escaping gradients, and applying particle swarm optimization algorithm, the semantic offset problem caused by environmental acoustic interference in open BPO contact centers was solved, enabling accurate analysis and risk identification of recruitment consultation content.

CN123070905APending Publication Date: 2026-11-10SHOUPIN (BEIJING) TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611141446.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-30
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

In open BPO contact center recruitment platforms, existing technologies struggle to effectively identify and manage semantic shifts caused by environmental acoustic interference, leading to misjudgments and underreporting, increasing manual review costs, and failing to track the true risks of job seekers' offline lead generation.

Method used

By extracting recruitment entities, assessing acoustic pressure and susceptibility to escaping, and combining semantic vectors and violation corpus vectors, an escaping gradient is constructed. Particle swarm optimization algorithm is then applied for iterative feedback to achieve accurate identification and monitoring of consultation risks.

Benefits of technology

It effectively eliminates the acoustic crosstalk blind spot in open office environments, achieves accurate separation and dynamic monitoring of hidden risks and job seekers' real needs, and improves the platform's ability to identify and manage fraudulent recruitment and abnormal promotional activities in compliance with regulations.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123070905A_ABST
    Figure CN123070905A_ABST
Patent Text Reader

Abstract

The application relates to the technical field of data analysis, and discloses a BPO customer consultation content analysis system for analysis and management of a recruitment platform, which comprises the following steps: analyzing consultation records to obtain semantic vectors and element word frequencies of transcribed segments, and simultaneously calculating environmental acoustic pressure according to equipment group normalized acoustic indexes; combining the historical appearance probability and category distribution entropy of recruitment entities to evaluate the easy-escape degree, and constructing an escape gradient representing an information evolution path according to the easy-escape degree; inputting the above parameters into a particle swarm optimization algorithm, introducing actual disposal order feedback for iterative optimization to obtain optimal parameters; finally obtaining consultation risk values and segment element contribution degrees to evaluate the overall risk of an account, and calculating divergence by using a closed-loop disposal time length to trigger the next round of iteration. The application effectively overcomes semantic misjudgment caused by open-seat background interference and hidden diversion, and improves the accuracy of platform compliance analysis and risk tracking.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and more specifically, to a BPO customer inquiry content analysis system for recruitment platform analysis and management. Background Technology

[0002] In the operational system of modern online recruitment platforms, these platforms typically outsource the handling of massive amounts of job seeker inquiries and complaints to large-scale business process outsourcing (BPO) contact centers. As recruitment processes become more decentralized, the entry points for business leads are increasingly fragmented. Currently, offline recruitment displays, posters, or labor agency flyers are often distributed using QR codes for multiple job postings or short link redirects. When job seekers scan the code to access human assistance, the original source traceability chain is lost. The agent can only capture the specific text of the job seeker's inquiry but cannot trace back to which replaced QR code or specific offline gathering point the job seeker entered from. This makes it difficult for the management system to fully reconstruct fraudulent chains such as inducing payment of training fees and redirecting traffic away from the platform to third-party private domains.

[0003] On the other hand, current BPO contact centers generally adopt an open-plan office layout. During peak recruitment seasons with high concurrency, adjacent agents often handle highly homogeneous job inquiries simultaneously. Combined with headset audio leakage, background noise from multiple speakers, and compression losses from recording equipment, the audio channel experiences severe sound field crosstalk pollution. In daily engineering analysis, it has been found that existing Automatic Speech Recognition (ASR) engines or conventional Natural Language Processing (NLP) models, when dealing with noisy audio, often focus on basic noise reduction or improving the overall word error rate, neglecting the business specificity of recruitment inquiries: job seekers' core concerns are usually highly concentrated on very short, key entities such as job titles, salary figures, location abbreviations, and training costs. Under the acoustic interference of an open environment, the decreased confidence of ASR can easily lead to the misclassification or replacement of these high-value short texts. This replacement is not just a simple literal error, but also causes semantic shifts at the entire business analysis level. For example, a routine inquiry intended to confirm the training period might be misidentified as a high-risk complaint about a training loan due to ambient noise, or the acoustic residue of a nearby intermediary charging fees might be mixed into a current compliant job inquiry. The existing analytics pipeline lacks a mechanism to dynamically decouple this semantic shift caused by ambient noise from subsequent compliance assessments. This results in significant false positives and false negatives when the platform assesses employer quality, tracks the true risks of job positions, and responds to job seeker complaints, substantially increasing the operational costs of manual review. Summary of the Invention

[0004] This invention provides a BPO customer consultation content analysis system for recruitment platform analysis and management, which solves the technical problems mentioned in the background.

[0005] This invention provides a BPO customer inquiry content analysis system for recruitment platform analysis and management, comprising: The consultation and analysis unit extracts recruitment entities from the transcribed text fragments, statistically analyzes the distribution of element frequency representations of the corresponding recruitment element categories, and generates semantic vectors to represent the text context. The acoustic pressure calculation unit uses a smoothing term to normalize the acoustic indices of the transcribed text fragment by device group to eliminate device difference interference and obtain the acoustic pressure characterizing environmental crosstalk. The escaping evaluation unit assesses the ease of escaping semantic distortion by combining the occurrence probability of the recruitment entity, the entity length, and the distribution entropy of each recruitment element category. The gradient construction unit, combining the acoustic pressure, the ease of escaping, and the similarity between the semantic vector and the job posting text vector and the violation corpus vector, constructs an escaping gradient with feature dimensions. The feature optimization unit inputs the basic feature set formed by the escaped gradient into the particle swarm optimization algorithm and sorts and optimizes it in combination with the actual processing results to obtain the optimal parameters containing the optimal selection weight and the optimal feature weight. The risk analysis unit uses the optimal parameters and the escape gradient to derive the consultation risk value and the fragment contribution of the character element weights, and aggregates them to generate the account risk value. The feedback and reinjection unit obtains the processing time reflecting the closed-loop cycle, calculates the distribution drift, and uses the distribution drift to correct the basic feature set to eliminate time delay interference and trigger iteration.

[0006] The beneficial effects of this invention are as follows: Addressing the problem of misinterpretation of recruitment-related terms caused by environmental acoustic interference and fragmented offline lead generation in open business process outsourcing workstations, this invention extracts acoustic parameters from consultation records and assesses the ease of translatability by combining the rarity and distribution entropy of the recruitment entity itself. Then, it calculates the evolution gradient based on sound field pressure and semantic similarity, and applies a particle swarm optimization algorithm to introduce real-world processing results for iterative feedback. This effectively eliminates the acoustic crosstalk blind spot in open office environments, achieving accurate separation and dynamic monitoring of hidden risks and job seekers' genuine needs. This effectively improves the platform's systematic identification and compliance management efficiency against fraudulent recruitment and abnormal lead generation behaviors. Attached Figure Description

[0007] Figure 1 This is a flowchart of a BPO customer consultation content analysis system for recruitment platform analysis and management, as described in this invention. Detailed Implementation

[0008] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0009] like Figure 1 As shown, a BPO customer inquiry content analysis system for recruitment platform analysis and management includes: The consultation and analysis unit extracts recruitment entities from the transcribed text fragments, statistically analyzes the distribution of element frequency representations of the corresponding recruitment element categories, and generates semantic vectors to represent the text context. The acoustic pressure calculation unit uses a smoothing term to normalize the acoustic indices of the transcribed text fragment by device group to eliminate device difference interference and obtain the acoustic pressure characterizing environmental crosstalk. The escaping evaluation unit assesses the ease of escaping semantic distortion by combining the occurrence probability of the recruitment entity, the entity length, and the distribution entropy of each recruitment element category. The gradient construction unit, combining the acoustic pressure, the ease of escaping, and the similarity between the semantic vector and the job posting text vector and the violation corpus vector, constructs an escaping gradient with feature dimensions. The feature optimization unit inputs the basic feature set formed by the escaped gradient into the particle swarm optimization algorithm and sorts and optimizes it in combination with the actual processing results to obtain the optimal parameters containing the optimal selection weight and the optimal feature weight. The risk analysis unit uses the optimal parameters and the escape gradient to derive the consultation risk value and the fragment contribution of the character element weights, and aggregates them to generate the account risk value. The feedback and reinjection unit obtains the processing time reflecting the closed-loop cycle, calculates the distribution drift, and uses the distribution drift to correct the basic feature set to eliminate time delay interference and trigger iteration.

[0010] The BPO customer inquiry content analysis system for recruitment platform analysis and management provided in this embodiment is applied to a data processing device that includes a voice acquisition device, a text transcription server, and a data analysis server. The data processing device includes, but is not limited to, enterprise-level servers, cloud computing clusters, and industrial control computers. The voice acquisition device supports analog and digital audio input, with a sampling rate of 16kHz and a sampling precision of 16bit, used to synchronously acquire audio from conversations between agents and job seekers. The text transcription server is equipped with an automatic speech recognition engine with a word error rate of less than 5%, used to convert the acquired audio into transcribed text segments by time slices. The data analysis server is used to perform subsequent semantic parsing, acoustic calculations, risk assessment, and feedback iteration, with a processing latency of less than 100ms for a single inquiry record.

[0011] S201: Obtain the individual recruitment entities from the transcribed text fragment.

[0012] The transcribed text segments are text units obtained by the automatic speech recognition engine from call audio by dividing it into time slices. Each transcribed text segment is 2 seconds long, with a 0.5-second overlap between adjacent segments. Recruitment entities are key information units related to recruitment within the transcribed text segments, including seven categories: job title, salary, work location, training fees, start date, benefits, and contact information. Recruitment entities are obtained using a named entity recognition model based on a pre-trained language model. The base model is BERT-base-Chinese-v2, fine-tuned on a recruitment-specific dataset containing at least 100,000 manually annotated transcribed text segments. The annotation specifications clearly define the boundaries and annotation rules for each entity category. The model training uses a cross-entropy loss function with a learning rate of 2e-5, 10 training epochs, and a batch size of 32. For entities not logged in, a rule-matching mechanism is used as a fallback. The rule base includes regular expression templates for common job title prefixes, salary units, location administrative divisions, and contact information formats.

[0013] S202: Determine the category of each recruitment entity using matching conditions, and perform frequency aggregation based on the inverse frequency data of the recruitment entity in the corpus to obtain the element frequency reflecting the corresponding recruitment element category, which satisfies: ; Where i is the index of the consultation record corresponding to the transcribed text fragment, t is the index of the transcribed text fragment, and k is the index of the recruitment element category. To transcribe a text fragment, r represents the recruitment entity. To determine whether an entity belongs to category k, 1[·] is an indicator mapping that takes one if the condition is met and zero if the condition is not met. This represents summing over all entities within the fragment, where idf(r) is the inverse frequency data of the recruitment entity in the corpus. For element word frequency.

[0014] The matching criteria are based on a pre-defined entity category dictionary. Each recruitment element category in the dictionary contains at least 1000 standard entity examples and corresponding matching rules. Entity boundaries are defined as follows: Job title includes general job titles and specific job titles, such as software engineer and Java development engineer; salary includes the corresponding values ​​and units for basic salary, performance-based salary, subsidies, and year-end bonus; work location includes province, city, district, street, and specific address; training fees include the fees for various training programs; start date includes the specific date and relative time; benefits include social security, housing fund, accommodation, and bonuses; contact information includes mobile phone number, WeChat ID, and QQ number. The matching rules combine exact matching and fuzzy matching. Exact matching is used for standard entities, and fuzzy matching is used for variant entities. The similarity threshold for fuzzy matching is 0.85, and the similarity calculation uses the edit distance algorithm. When an entity matches multiple categories simultaneously, the category affiliation is determined according to the priority order: contact information > training fees > salary > job title > work location > start date > benefits. The inverse frequency data idf(r) is calculated as follows: Where N is the total number of documents in the corpus, and df(r) is the number of documents containing entity r. The corpus consists of transcribed text from all platforms over the past three months, and is automatically updated every Sunday at 2 AM. For entities not logged in, the default value of idf(r) is 3.0. The smoothing term ε is set to 1e-6 to avoid division by zero errors during calculation.

[0015] S203: Perform encoding processing on the transcribed text fragment to obtain a semantic vector, which satisfies: ; Where Enc(·) is the mapping function that performs the encoding process. It is a semantic vector.

[0016] The encoding process uses the same BERT-base-Chinese-v2 pre-trained language model as the named entity recognition model. The preprocessing rules for the input text are as follows: convert the transcribed text fragments to lowercase, remove special characters and extra spaces, add special markers at the beginning and end of the text, pad or truncate the text sequence to 128 characters (text exceeding the length is truncated in the middle, text insufficient is padded with zeros at the end). The output vector is extracted from the first token of the last layer of the model, namely the CLStoken, with a fixed semantic vector dimension of 768.

[0017] S301: Determine the median of acoustic indicators and the median deviation characterizing the dispersion within the same equipment group.

[0018] Equipment groups are collections of recording devices categorized by model and production batch used by agents. The acoustic characteristics of devices within the same group differ by less than 5%. When a device is replaced or its firmware is upgraded, it is automatically removed from its original group and added to a new group corresponding to its model and batch. Acoustic metrics include four categories: audio signal-to-noise ratio (SNR), speech energy, background noise intensity, and spectral flatness. Each metric corresponds to a category index b. Audio SNR is the average SNR of a speech segment. Speech segments are extracted using an energy-based speech activity detection algorithm, with a frame length of 25ms, a frame shift of 10ms, and a Hamming window as the window function. The energy threshold for speech activity detection is 1.5 times the average energy of the segment. Speech energy is the average short-time energy of all speech frames. The formula for calculating short-time energy is... , where x(n) is the amplitude of the nth sampling point, and N is the frame length. Background noise intensity is the average energy of all silent frames, and silent segments are those identified as non-speech segments by the speech activity detection algorithm. Spectral flatness is the spectral flatness within the 0-8kHz frequency band, calculated using the following formula: Where X(k) is the amplitude of the k-th frequency point, and K is the total number of frequency points. (Median of the index) It refers to all acoustic parameters within the same equipment group g. The median is calculated over a 24-hour period, recalculated daily at 3:00 AM. Median deviation. It refers to all acoustic parameters within the same equipment group g. The median of the absolute deviation from the median of the index.

[0019] For newly added equipment to a group, during the first 24 hours of use, the median and median deviation of the indicators calculated most recently by other equipment groups of the same model and batch are used as temporary values. After 24 hours, the corresponding median and median deviation of the indicators are recalculated based on the acoustic data generated by the equipment itself and updated in the statistical data of its respective equipment group. Data from other equipment groups are not included in the acoustic indicator statistics of other equipment groups; each equipment group only uses its own historical data for normalization calculations.

[0020] The speech activity detection algorithm employs a dual-threshold decision method, combining short-time energy and short-time zero-crossing rate to determine speech segments and silence segments. The formula for calculating the short-time zero-crossing rate is as follows: ,in For symbolic functions, Let N be the amplitude of the nth sampling point, and N be the frame length. The short-time energy threshold is 1.5 times the average energy of the segment, and the short-time zero-crossing rate threshold is 2.0 times the average zero-crossing rate of the segment. A frame is determined to be a speech frame when it simultaneously satisfies the conditions of having a short-time energy greater than the energy threshold and a short-time zero-crossing rate less than the zero-crossing rate threshold; otherwise, it is determined to be a silence frame.

[0021] S302: Extract the absolute deviation of the acoustic index from the median of the index. Perform relative deviation mapping using the absolute deviation, median deviation, and smoothing term to obtain the normalized index, which satisfies: ; Where b is the traversal index of the acoustic index category, and B is the total number of acoustic index categories, with a value of 4. For acoustic parameters, g represents the equipment group. The median of the index. The median deviation is ε, and the smoothing term has a value of 1e-6. This is a normalized indicator.

[0022] S303: Based on the normalized index of the current transcribed text fragment, and the global distribution status of the normalized index and smoothing term of all transcribed text fragments within the same consultation record, the acoustic pressure is quantified by performing proportion quantification, which satisfies: ; Where s is the traversal index of all transcribed text fragments within the same consultation record, and S is the total number of transcribed text fragments within the same consultation record. Here, ε is the normalization index for the traversal state, and ε is the smoothing term with a value of 1e-6. For acoustic pressure.

[0023] The acoustic pressure value ranges from 0 to 1. The larger the value, the more severe the environmental acoustic interference corresponding to the current transcribed text segment.

[0024] S401: Calculate the probability of occurrence and length of the recruitment entity.

[0025] probability of occurrence It is calculated by dividing the number of times the recruitment entity 'r' appears in the transcribed text fragments across all platforms in the past month by the total number of times all recruitment entities appear. A one-fold smoothing method is used to handle low-frequency entities. The calculation formula is as follows: Here, count(r) is the number of times entity r appears, N is the total number of times all recruitment entities appear, and V is the total number of recruitment entities. The probability of occurrence ranges from 0 to 1. The entity length len(r) is the number of characters contained in the recruitment entity r. For Chinese entities, the number of characters is the number of Chinese characters; for English entities, the number of characters is the number of words. The entity length ranges from 1 to 20, with entities exceeding 20 being treated as having 20. The smoothing term ε takes the value 1e-8.

[0026] S402: Calculate the distribution entropy of each recruitment element category.

[0027] Distribution Entropy The calculation formula is: ,in This is the set of all recruitment entities contained in the k-th recruitment element category. Distribution entropy measures the uniformity of entity distribution within this category; a higher entropy value indicates a more uniform entity distribution and a higher probability of semantic translation. The value of distribution entropy ranges from 0 to ln(N_k), where N_k is the total number of entities contained in the k-th recruitment element category. Zero-probability entities are ignored when calculating distribution entropy. The smoothing term ε takes the value 1e-8.

[0028] S403: Perform logarithmic decay mapping using the occurrence probability and smoothing term, and perform length penalty adjustment in combination with the logarithmic transformation result of the entity length to obtain the first weight characterizing the entity's anti-interference ability.

[0029] The formula for calculating the first weight is: Entities with lower occurrence probabilities are considered rarer and more susceptible to escaping due to acoustic interference; therefore, logarithmic decay mapping assigns them higher weights. Shorter entities are more prone to misidentification, and length penalty adjustment assigns them higher weights. The first weight ranges from 0 to positive infinity. The smoothing term ε takes the value 1e-8.

[0030] S404: Perform entropy mapping by utilizing the proportional relationship between the distribution entropy of the corresponding recruitment element category and the overall distribution entropy of all recruitment element categories and the smoothing term to obtain the second weight.

[0031] The formula for calculating the second weight is as follows: Where q is the traversal index of the recruitment element category, and Q is the total number of recruitment element categories, with a value of 7. The distribution entropy is for each recruitment element category. The second weight reflects the overall semantic stability of that recruitment element category; the higher the distribution entropy, the higher the probability of semantic escaping, and the larger the second weight. The value of the second weight ranges from 0 to 1. The smoothing term ε takes the value 1e-6.

[0032] S405: Perform a joint weighting operation based on the first and second weights of each recruitment entity to obtain the escapability degree, which satisfies: ; in, For the set of entities that are associated with the frequency of elements, It is easy to change the meaning.

[0033] The higher the susceptibility to semantic distortion, the more easily the entities corresponding to the recruitment element category in the transcribed text fragment are to undergo semantic distortion.

[0034] S501: Obtain the job posting text vector and violation corpus vector of the corresponding consultation record.

[0035] Job posting text vector This is a semantic vector obtained by inputting the job posting text associated with the corresponding consultation record into the same encoding processing mapping function Enc(·) as in S203. The association method between consultation records and job posting texts is as follows: when an agent logs into the system, they bind the currently processed job ID. When a job seeker calls, the system automatically associates the agent's currently bound job ID with the consultation record, thereby obtaining the corresponding job posting text. If the association fails, i.e., the agent has not bound a job ID or the job posting text does not exist, the average vector of all job posting texts from that business line is used as the general job vector. (Violation corpus vector) The semantic vector is the center vector obtained by inputting pre-collected violation corpora into the same encoding and processing mapping function. The violation corpora include 12 categories: inducing payment of training fees, redirecting traffic away from the platform to third-party private domains, false salary promises, false work locations, forced overtime, collecting deposits, withholding documents, false start dates, false benefits, harassing job seekers, inappropriate language, and other violations. Each category of violation corpus contains at least 5000 manually annotated transcribed text fragments. The violation corpus is automatically updated every Sunday at 4 AM, and the center vector is recalculated after each update.

[0036] When an agent is handling multiple job postings simultaneously, the system automatically matches the cosine similarity between the transcribed text of the consultation record from the first 3 seconds and the text posted by each job posting to be processed. The system then associates the job ID with the highest similarity with the consultation record. The similarity threshold is 0.6. If the similarity of all job postings is below 0.6, the average vector of all job postings from all job postings within that business line is used as the general job vector. The average vector of all job postings from all job postings within that business line is calculated as follows: All valid job postings from that business line are input into an encoding processing mapping function to obtain the corresponding semantic vectors; then, the arithmetic mean of each element of all semantic vectors is taken.

[0037] The labeling standards for violations clearly define the specific boundaries for each type of violation: Inducing training fees refers to requiring job seekers to pay any form of training fee, deposit, or guarantee in advance; Directing traffic away from the platform to a third-party private domain refers to providing contact information outside the platform or guiding job seekers to add friends outside the platform; False salary promises refer to a discrepancy of more than 20% between the actual salary and the posted salary; False work location refers to a distance of more than 5 kilometers between the actual work location and the posted location; Forced overtime refers to requiring more than 44 hours of work per week without overtime pay; Collecting deposits refers to requiring job seekers to pay any form of money as a condition for employment; Withholding documents refers to requiring job seekers to hand over their ID cards, graduation certificates, etc., to the employer for safekeeping; False start date refers to a discrepancy of more than 7 days between the actual start date and the promised start date; False benefits refer to benefits provided that do not match the job posting; Harassing job seekers refers to sending vulgar, threatening, or harassing messages unrelated to recruitment; Inappropriate remarks refer to discriminatory content related to gender, age, race, religion, etc.; Other violations refer to other behaviors that violate the platform's recruitment rules.

[0038] S502: Extract the positive similarity between the semantic vector and the job posting text vector, perform a reverse metric mapping based on the positive similarity, and obtain the deviation coefficient representing the degree of deviation from the job posting facts.

[0039] Forward similarity is calculated using cosine similarity, and the formula is: ; in, The dot product of the semantic vector and the job posting text vector. The L2 norm of the semantic vector. The L2 norm of the text vector used to post job openings. Cosine similarity values ​​range from -1 to 1. The formula for calculating the deviation coefficient is... The value ranges from 0 to 2. The larger the value, the greater the degree of semantic deviation between the transcribed text fragment and the semantics of the job posting text.

[0040] S503: Extract the negative similarity between the semantic vector and the violation corpus vector, perform a same-direction metric translation mapping based on the negative similarity, and obtain the convergence coefficient representing the tendency towards the violation pattern.

[0041] Negative similarity is also calculated using cosine similarity, with the following formula: ; in, The dot product of the semantic vector and the vector of the violation corpus. The L2 norm of the semantic vector. The L2 norm of the violation corpus vectors. The cosine similarity ranges from -1 to 1. The formula for calculating the convergence coefficient is... The value ranges from 0 to 1. The larger the value, the closer the semantics of the transcribed text fragment are to the semantics of the illegal corpus.

[0042] S504: Based on acoustic pressure, ease of escaping, deviation coefficient, and convergence coefficient, feature fusion processing is performed to obtain the escaping gradient, which satisfies the following: ; in, This is for escaping gradients.

[0043] The escaping gradient integrates features from four dimensions: environmental acoustic interference, the ease of escaping of the entity itself, the degree of semantic deviation from the facts of the job, and the degree of semantic tendency towards the violation pattern. It is used to characterize the possibility of semantic escaping in transcribed text fragments and the resulting risks.

[0044] S601: Merge the escaped gradients of each feature dimension to form a basic feature set.

[0045] Basic feature set Each component corresponds to a feature dimension d. The total number of feature dimensions D is equal to the total number of recruitment element categories Q multiplied by the maximum number of transcribed text fragments S_max for each consultation record, where Q is 7, S_max is 200, and therefore D is 1400. For consultation records with fewer than 200 transcribed text fragments, the escape gradient values ​​of the missing parts are padded with zeros; for consultation records with more than 200 transcribed text fragments, the escape gradient values ​​of the first 200 fragments are truncated. Each basic feature set component... The escape gradient value corresponding to the i-th consultation record on the d-th feature dimension.

[0046] S602: In the particle swarm optimization algorithm, optimize particles are set. Based on the feature selection weights and feature weights contained in the optimize particles, feature weighted fusion is performed on the basic feature set to obtain the risk prediction value of the optimize particles, which satisfies: ; Where z represents the optimized particle, d represents the traversal index of the feature dimension, and D represents the total number of feature dimensions, with a value of 1400. For feature selection rights, For feature weights, Basic feature set components, This is the predicted risk value.

[0047] The optimized particle encoding structure is as follows: the first D bits are binary feature selection weights, and the last D bits are continuous feature weights. The total dimension of the particle is 2D, or 2800. (Feature selection weights) This is a binary variable, taking the value 0 or 1. A value of 1 indicates that the feature dimension is retained, and a value of 0 indicates that the feature dimension is removed. Feature weight This is a continuous variable, ranging from 0 to 1, used to adjust the contribution of this feature dimension. The initial particle swarm size is 50, and the initial positions of the particles are randomly generated. The initial value of the feature selection weight is either 0 or 1 with a probability of 0.5, and the initial value of the feature weight is a uniformly distributed random number between 0 and 1. The initial velocity of the particles is set to 0.

[0048] The range of particle velocity values ​​is set to ,in When the updated velocity of a particle exceeds this range, the velocity is truncated to the corresponding boundary value. The value range of the feature weights always remains within a certain range. When the updated feature weights of the particles exceed this range, they are truncated to 0 or 1. The feature selection weights remain binary values ​​throughout the particle update process; when the updated feature selection weight value is greater than 0.5, it is set to 1; otherwise, it is set to 0.

[0049] S603: Assess the statistical correlation coefficient between the predicted risk values ​​and the actual disposal results.

[0050] The statistical correlation coefficient used is the Spearman rank correlation coefficient, and the formula is: ; Where N is the total number of consultation records, The risk prediction values ​​are sorted according to the rule that the larger the value, the higher the ranking. The ranking of actual handling results is based on the rule that higher actual risk ranks higher. The Spearman rank correlation coefficient measures the monotonic correlation between two ranked sequences, ranging from -1 to 1. A value closer to 1 indicates a higher consistency between the risk prediction and the actual handling results. The labeling rules for the actual handling results are as follows: risk levels are divided into 5 levels, from low to high (1 to 5), with the priority being handling intensity > refund amount > handling duration. Labeling personnel must have at least 3 years of experience in compliance review work on recruitment platforms. Labeling results must undergo cross-review by two people, with an accuracy rate of over 95%. Labeling results that fail the review must be re-labeled.

[0051] The specific criteria for determining the 5 risk levels in the actual handling results ranking are as follows: Level 1 risk corresponds to a handling intensity of 0, a refund amount of 0, and a handling time of less than 300 seconds; Level 2 risk corresponds to a handling intensity of 1, a refund amount of 0, and a handling time between 300 and 1800 seconds; Level 3 risk corresponds to a handling intensity of 1, a refund amount between 0 and 1000 yuan, or a handling intensity of 2 and a refund amount of 0; Level 4 risk corresponds to a handling intensity of 2, a refund amount between 1000 and 5000 yuan; and Level 5 risk corresponds to a handling intensity of 3, or a refund amount exceeding 5000 yuan.

[0052] S604: Perform sparse metricization based on the feature selection weights of all feature dimensions and the total number of feature dimensions to obtain the dimension proportions.

[0053] The formula for calculating the dimensional proportion is: The value ranges from 0 to 1. The larger the value, the more feature dimensions are selected, and the higher the complexity of the model.

[0054] S605: Determine the between-group variance of the predicted mean of risk predictions within the same equipment group, and the individual variance of the risk predictions.

[0055] Between-group variance It is the variance of the mean of the risk prediction values ​​among different equipment groups, calculated using the following formula: Where G is the total number of equipment groups. This represents the average of the risk prediction values ​​for all consultation records within the g-th equipment group. This represents the overall mean of the risk predictions for all consultation records. Individual variance. The overall variance of risk predictions across all consultation records is calculated using the following formula: .

[0056] S606: The variance ratio is obtained based on the correlation between inter-group variance and individual variance with the smoothing term.

[0057] The formula for calculating the variance ratio is: The value ranges from 0 to positive infinity. A larger value indicates a greater difference in the predicted risk values ​​between different equipment groups, and a worse equipment robustness of the model. The smoothing term ε takes the value 1e-6.

[0058] S607: Fitness is obtained based on the distribution relationship between the relative dimensional proportions and variance ratios of the statistical correlation coefficient, which satisfies: ; Where fit(z) is the fitness.

[0059] The fitness function takes into account the model's prediction accuracy, sparsity, and device robustness. A higher fitness value indicates a better combination of parameters for the optimized particles.

[0060] S608: Iterative updates are performed on the optimized particles using inertia weights, a first learning factor, and a second learning factor. The feature selection weights and feature weights contained in the optimized particle with the highest fitness are taken as the optimal selection weights and optimal feature weights to form the optimal parameters; the particle update relationship satisfies: ; Where j is the internal dimension index of the optimized particle, and r is the iteration round. and To update the particle velocity to the current velocity, For inertial weights, and Let α and β be the first and second learning factors, respectively, and α and β be random perturbation terms. For the individual's historical best position, Current position This is the globally optimal position.

[0061] The inertia weight adopts a linear decay strategy, and the calculation formula is as follows: ,in , , The maximum number of iterations is set to 200. First learning factor. With the second learning factor Both are fixed at a value of 2.0. The random perturbation terms α and β are uniformly distributed random numbers between 0 and 1, and are regenerated in each iteration. Individual historical best position. The update rule is: when the current particle's fitness is strictly greater than its historical best fitness, the current position is updated to its historical best position. Global best position. The update rule is as follows: when the fitness of any particle in the current particle swarm is strictly greater than the global historical best fitness, the current position of that particle is updated to the global best position. The iteration termination condition is: the number of iterations reaches the maximum number of iterations, 200, or the global best fitness has not improved for 20 consecutive iterations. After the iteration terminates, the feature selection weight and feature weight corresponding to the global best position are used as the optimal selection weight and optimal feature weight to form the optimal parameters.

[0062] S701: By substituting the optimal choice weight and the optimal feature weight into the weighted fusion process of the risk prediction value, the consultation risk value corresponding to the consultation record that generates the transcribed text fragment is obtained, which satisfies: ; in, For the optimized particles corresponding to the optimal parameters. To substitute the optimal parameters into the weighted fusion relation, This is for consulting risk values.

[0063] The consultation risk value ranges from 0 to 1, with a higher value indicating a higher risk for the consultation record.

[0064] S702: The fusion attribute of the optimal selection weight, optimal feature weight, and escape gradient associated with the current transcribed text fragment and its corresponding recruitment element category is quantified relative to the fusion attribute and smoothing term of all text fragments and all categories to obtain the fragment contribution, which satisfies: ; in, This represents the optimal choice at the current time. The current optimal feature weight. Let be the corresponding escape gradient, s be the traversal index of all text segments, S be their total number, q be the traversal index of all categories, Q be their total number, and take a value of 7. and Optimal selection weight and optimal feature weight in the traversal state. To iterate through the corresponding escape gradients, ε is the smoothing term, taking a value of 1e-6. Contribution to the segment.

[0065] The value of the fragment contribution ranges from 0 to 1. The larger the value, the greater the contribution of the content of the corresponding recruitment element category in the transcribed text fragment to the consultation risk value.

[0066] S703: Obtain the set of consultation records associated with the corresponding account, perform aggregation processing based on all consultation risk values ​​within the consultation record set, and obtain the account risk value of the corresponding account, which satisfies: ; Where 'o' represents the account number. This is a collection of consultation records. The total number of records in the consultation record set. For consulting risk values, For time decay weight, Account risk value.

[0067] The formula for calculating the time decay weight is as follows: ,in , For the current time, This refers to the generation time of the consultation record. The time window is the past 90 days; consultation records generated more than 90 days ago are not included in the aggregation. For accounts with no consultation records, the default account risk value is 0.1. The account risk value ranges from 0 to 1; a higher value indicates a higher overall risk for the account.

[0068] Parameters of time decay weight This means that the weight of consultation records decreases by 1% every day, that is, for every day that has passed since the record was created, the contribution of the consultation record to the account risk value decreases by 1%. Consultation records created more than 90 days ago have their weight reduced to approximately 0.407 and will no longer be included in the aggregate calculation of the account risk value.

[0069] S801: Obtain the processing time, refund amount, and processing intensity that represent the depth of management intervention corresponding to the ranking of actual processing results.

[0070] Processing time This is the time difference between the time the consultation record was generated and the time the processing was completed, expressed in seconds. Refund amount. This is the total refund amount generated from this consultation record, in yuan. Processing Intensity It is a quantitative value representing the depth of management intervention, ranging from 0 to 3. 0 indicates that no human intervention is required and the system completes the handling automatically; 1 indicates that a first-level agent intervenes, only answering questions or conducting preliminary information verification; 2 indicates that a second-level supervisor intervenes, initiating a refund operation or temporarily freezing the account; and 3 indicates that the compliance department intervenes, permanently freezing the account or transferring it to judicial authorities for processing.

[0071] S802: Based on the processing time and refund amount, perform compression mapping, combine processing intensity with joint metric fusion and ranking normalization to obtain the feedback ranking value, which satisfies: ; in, This is the normalized processing time. This is the normalized refund amount. To determine the strength of the sorting, ln(·) is the natural logarithm of the compression mapping, and ranknorm(·) is the mapping function that performs sorting normalization. This is the feedback sorting value.

[0072] The normalized formula for processing time is: ,in , ,when hour, The value is 0; when hour, The value is 1. The normalization formula for the refund amount is: ,in , ,when hour, The value is 0; when hour, The value is 1. The sorting normalization uses rank transformation normalization, and the formula is: Where rank(x) is the ranking of x among all samples, with the ranking rule being that the larger the value, the higher the ranking. N is the total number of samples. For samples with the same value, the average ranking is taken. The feedback ranking value ranges from 0 to 1, with a larger value indicating a higher actual risk for the consultation record.

[0073] S803: Obtain the distribution divergence of the corresponding recruitment element category between the newly effective distribution and the historical effective distribution, as the distribution drift of the corresponding recruitment element category, which satisfies: ; in, For the newly effective distribution, For the historical effective distribution, JS(·,·) is a mapping function for calculating the distribution divergence. This represents the distribution drift.

[0074] The formula for calculating JS divergence is: ,in KL represents the KL divergence, calculated using the following formula: ε is the smoothing term, with a value of 1e-8. Newly effective distribution. This is the probability distribution of the occurrence of all recruitment entities under the k-th recruitment element category within the last 3 days, and the historical effective distribution. This represents the probability distribution of all job entities appearing under the k-th job category within the previous 3 days. The probability distribution uses a frequency estimation method combined with one-fold smoothing. The distribution drift ranges from 0 to 1; a larger value indicates a greater difference between the two distributions, meaning that the distribution of this job category has experienced a significant drift.

[0075] The new effective distribution and the historical effective distribution are updated daily at 5:00 AM. Each update uses the transcribed text data from the most recent 3 days as the statistical basis for the new effective distribution, and the transcribed text data from the previous 3 days as the statistical basis for the historical effective distribution. For recruitment entities that did not appear within the statistical period, their probability of occurrence is calculated using the minimum value after smoothing by adding one.

[0076] S804: Utilize the distribution drift of each recruitment element category to perform bias compensation on the basic feature set corresponding to the previous iteration, obtaining a corrected feature set. The compensation relationship of this feature set satisfies: ; in, Let be the components of the basic feature set along the feature dimension d, and k(d) be the correspondence function that strictly maps the feature dimension d to the corresponding recruitment element category k. This represents the distribution drift for the corresponding category. To correct the feature set.

[0077] The mapping rule from feature dimensions to recruitment element categories is as follows: For feature dimension d, Where Q is the total number of recruitment element categories, with a value of 7. That is, every 7 consecutive feature dimensions correspond to 7 recruitment element categories of a transcribed text fragment. The 1st to 7th feature dimensions correspond to the 1st transcribed text fragment, the 8th to 14th feature dimensions correspond to the 2nd transcribed text fragment, and so on.

[0078] S805: Update the actual disposal result ranking using the feedback ranking value, and trigger the next iteration of the particle swarm optimization algorithm using the modified feature set.

[0079] When updating the actual handling results ranking, the feedback ranking value is used as the new actual handling results ranking, replacing the original ranking. Iteration triggering uses a combination of fixed-cycle and threshold triggering. The fixed iteration cycle is 3 days, and the distribution drift threshold is set to 0.3. When the distribution drift of any recruitment element category... When an iteration occurs, it is immediately triggered. During the iteration, ongoing risk assessment tasks continue to use the optimal parameters from the previous round, while newly submitted risk assessment tasks use the updated optimal parameters.

[0080] When an emergency iteration is triggered (i.e., the distribution drift exceeds 0.3), the system prioritizes using the latest available data for model training and updates the optimal parameters immediately after training. During the iteration process, the system retains the optimal parameters from the last three iterations. If the fitness of the newly trained model is lower than the optimal fitness of the previous three iterations, it automatically rolls back to the optimal parameters from the previous iteration.

[0081] Specifically, when the voice acquisition device fails to acquire the voice data or the audio file is corrupted, the processing of that consultation record is skipped, and an error log is recorded. When the transcribed text fragment returned by the text transcription server is empty, the consultation risk value of that consultation record is set to 0. When the named entity recognition model fails to recognize any recruitment entity, the word frequencies of all elements corresponding to that transcribed text fragment are set to 0. When numerical overflow, division by zero, or other abnormal situations occur during the calculation process, the default value of this parameter is used as a substitute, and an exception log is recorded. When a downstream system interface call fails, it is retried 3 times, with a 1-second interval between each attempt. If it still fails, an error log is recorded, and the processing of that record is skipped.

[0082] Specifically, the criteria for determining numerical anomalies include: a calculation result of infinity, a calculation result that is not numerical, and a calculation result that exceeds a preset reasonable range. For calculation results exceeding the reasonable range, if the result is greater than the preset maximum value, the maximum value is used; if the result is less than the preset minimum value, the minimum value is used. The error log records the time of the anomaly, the anomaly type, the step in which the anomaly occurred, the relevant consultation record ID, and detailed information about the anomaly. The error log is stored in a designated directory on the data analysis server and retained for 90 days.

[0083] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A BPO customer inquiry content analysis system for recruitment platform analysis and management, characterized in that, include: The consultation and analysis unit extracts recruitment entities from the transcribed text fragments, statistically analyzes the distribution of element frequency representations of the corresponding recruitment element categories, and generates semantic vectors to represent the text context. The acoustic pressure calculation unit uses a smoothing term to normalize the acoustic indices of the transcribed text fragments by device group to eliminate device difference interference and obtain the acoustic pressure characterizing environmental crosstalk. The escaping evaluation unit assesses the ease of escaping semantic distortion by combining the probability of occurrence of the recruitment entity, the entity length, and the distribution entropy of each recruitment element category. The gradient construction unit combines acoustic pressure, ease of escaping, and the similarity between the semantic vector and the job posting text vector and the violation corpus vector to construct an escaping gradient with feature dimensions. The feature optimization unit inputs the basic feature set formed by the escaped gradient into the particle swarm optimization algorithm and sorts and optimizes it according to the actual processing results to obtain the optimal parameters including the optimal selection weight and the optimal feature weight. The risk analysis unit uses optimal parameters and escape gradients to derive the consultation risk value and the fragment contribution of the weights of the characterizing elements, and aggregates them to generate the account risk value. The feedback and reinjection unit obtains the processing time reflecting the closed-loop cycle, calculates the distribution drift, and uses the distribution drift to correct the basic feature set to eliminate time delay interference and trigger iteration.

2. The BPO customer inquiry content analysis system for recruitment platform analysis and management according to claim 1, characterized in that, The consultation and analysis unit is specifically used for: Obtain each of the recruitment entities from the transcribed text fragment; The category of each recruitment entity is determined by matching conditions, and frequency aggregation is performed based on the inverse frequency data of the recruitment entity in the corpus to obtain the element word frequency that reflects the category of the corresponding recruitment element. The transcribed text fragment is encoded to obtain the semantic vector.

3. The BPO customer inquiry content analysis system for recruitment platform analysis and management according to claim 2, characterized in that, The acoustic pressure calculation unit is specifically used for: Determine the median of the acoustic indicators and the median deviation characterizing the dispersion within the same equipment group; Extract the absolute deviation of the acoustic index relative to the median of the index, and perform relative deviation mapping using the absolute deviation, the median deviation, and the smoothing term to obtain the normalized index; The acoustic pressure is obtained by performing a percentage quantization of the normalized index of the current transcribed text fragment, the normalized index of all transcribed text fragments within the same consultation record, and the global distribution state of the smoothing term.

4. The BPO customer inquiry content analysis system for recruitment platform analysis and management according to claim 3, characterized in that, Escape evaluation unit, specifically used for: The probability of occurrence of the recruitment entity and the length of the entity are statistically analyzed. Calculate the distribution entropy for each of the aforementioned recruitment element categories; Logarithmic decay mapping is performed using the occurrence probability and the smoothing term, and length penalty adjustment is performed in combination with the logarithmic transformation result of the entity length to obtain the first weight characterizing the entity's anti-interference ability; By performing entropy mapping based on the proportional relationship between the distribution entropy of the corresponding recruitment element category and the overall distribution entropy of all recruitment element categories and the smoothing term, a second weight is obtained; The ease of escaping is obtained by performing a joint weighting operation based on the first weight and the second weight of each of the recruitment entities.

5. The BPO customer inquiry content analysis system for recruitment platform analysis and management according to claim 4, characterized in that, Gradient construction unit, specifically used for: Obtain the job posting text vector and the violation corpus vector corresponding to the consultation record; Extract the positive similarity between the semantic vector and the job posting text vector, and perform a reverse metric mapping based on the positive similarity to obtain a deviation coefficient that represents the degree of deviation from the job posting facts; Extract the negative similarity between the semantic vector and the violation corpus vector, and perform a same-direction metric translation mapping based on the negative similarity to obtain the convergence coefficient representing the tendency towards the violation pattern; The escape gradient is obtained by performing feature fusion processing based on the acoustic pressure, the ease of escape, the deviation coefficient, and the convergence coefficient.

6. The BPO customer inquiry content analysis system for recruitment platform analysis and management according to claim 5, characterized in that, Feature optimization unit, specifically used for: The escaped gradients of each of the aforementioned feature dimensions are combined to form the basic feature set; In the particle swarm optimization algorithm, optimization particles are set, and based on the feature selection weights and feature weights contained in the optimization particles, feature weighted fusion is performed on the basic feature set to obtain the risk prediction value of the optimization particles. Evaluate the statistical correlation coefficient between the predicted risk values ​​and the ranking of the actual disposal results; Sparse metricization is performed based on the feature selection weights of all the aforementioned feature dimensions and the total number of feature dimensions to obtain the dimension proportions; Determine the between-group variance of the predicted mean of the risk predictions within the same group of equipment, and the individual variance of the risk predictions. The variance ratio is obtained based on the correlation between the inter-group variance and the individual variance with the smoothing term; The fitness is obtained based on the distribution relationship between the statistical correlation coefficient, the proportion of the dimension, and the variance ratio. The optimized particles are iteratively updated using inertia weights, a first learning factor, and a second learning factor. The feature selection weights and feature weights contained in the optimized particles with the highest fitness are used as the optimal selection weights and the optimal feature weights to form the optimal parameters.

7. The BPO customer inquiry content analysis system for recruitment platform analysis and management according to claim 6, characterized in that, The risk analysis unit is specifically used for: By substituting the optimal selection weight and the optimal feature weight into the weighted fusion process of the risk prediction value, the consultation risk value corresponding to the consultation record that generates the transcribed text fragment is obtained; The contribution of the current transcribed text fragment and the fusion attribute of the optimal selection weight, optimal feature weight and escape gradient associated with the corresponding recruitment element category are quantified relative to the fusion attribute of all text fragments and all categories and the smoothing term to obtain the fragment contribution. Obtain the set of consultation records associated with the corresponding account, and perform aggregation processing based on all the consultation risk values ​​in the set of consultation records to obtain the account risk value corresponding to the account.

8. The BPO customer inquiry content analysis system for recruitment platform analysis and management according to claim 7, characterized in that, The feedback and recharge unit is specifically used for: Obtain the processing time, refund amount, and processing intensity representing the depth of management intervention corresponding to the ranking of the actual processing results; Based on the processing time and the refund amount, a compression mapping is performed, and combined with the processing intensity, a joint metric fusion and sorting normalization are performed to obtain a feedback sorting value; Obtain the distribution divergence between the newly effective distribution and the historical effective distribution for the corresponding recruitment element category, and use it as the distribution drift for the corresponding recruitment element category; By using the distribution drift of each of the recruitment element categories to perform deviation compensation on the basic feature set corresponding to the previous iteration, a corrected feature set is obtained; The actual processing result ranking is updated using the feedback ranking value, and the next iteration of the particle swarm optimization algorithm is triggered using the corrected feature set.