Credit evaluation system for outsourcing service providers based on ai large model multi-source data fusion

Abstract

This invention relates to the field of credit assessment technology, and more particularly to a credit assessment system for outsourced service providers based on AI large-scale model multi-source data fusion. The system includes: a data crawling module for crawling data from various target websites for each service provider; a queue processing and analysis module for performing queue processing analysis, determining the similarity balance of related items based on load reference values ​​or similarity reference differences as constraint processing conditions; a text selection module for determining the coverage of domain-shifted content based on historical data support and data volume, using historical keywords or general sentiment words, and determining whether to adjust the data analysis method from general sentiment analysis to adaptive representation analysis based on the domain-shifted content coverage of the analysis queue to obtain several selection texts; and a credit assessment module. This invention can improve the quality and efficiency of credit assessment for outsourced service providers.

Description

Technical Field

[0001] This invention relates to the field of credit assessment technology, and in particular to a credit assessment system for outsourced service providers based on AI large-scale model multi-source data fusion. Background Technology

[0002] In outsourced service provider credit assessment, accurate, comprehensive, and timely data collection and in-depth analysis are crucial prerequisites for ensuring the reliability of assessment results. However, the current surge in internet data volume, coupled with the prevalence of multi-source heterogeneous data, has led to prominent issues such as data redundancy, information fragmentation, and difficulty in identifying effective information. These problems severely reduce data utilization efficiency and affect the accuracy and timeliness of credit assessments. Therefore, how to efficiently extract relevant information for credit assessment from massive amounts of data to improve the quality and efficiency of outsourced service provider credit assessments has become a pressing issue for those skilled in the art.

[0003] Chinese patent publication CN110602147A discloses an AI-driven method for extracting and evaluating multi-dimensional credit features of collaborative enterprises. This method includes employing more financial data indicators and incorporating textual data indicators of enterprise credit reviews to characterize the enterprise image from multiple dimensions, and using multiple models to extract features. However, this technical solution suffers from the following problems: it lacks dynamic crawling control, intelligent queue partitioning, and adaptive text filtering based on actual application scenarios, resulting in poor data collection completeness, task processing efficiency, and text filtering accuracy. Consequently, the accuracy and reliability of credit assessment results for outsourced service providers are poor. Summary of the Invention

[0004] To address this, the present invention provides an outsourcing service provider credit assessment system based on AI large-scale model multi-source data fusion, which overcomes the problems in existing technologies that cannot dynamically crawl and control data according to actual application scenarios, intelligently divide queues, and adaptively filter text, resulting in poor data collection integrity, task processing efficiency, and text filtering accuracy, and consequently, poor accuracy and reliability of outsourcing service provider credit assessment results.

[0005] To achieve the above objectives, this invention provides a credit assessment system for outsourcing service providers based on AI large-scale model multi-source data fusion, comprising: The data crawling module is used to crawl data from various service providers at various target URLs to obtain crawled data. Specifically, for any crawling task of any service provider at any target URL, the module determines whether to perform crawling adjustment based on the cycle duration and whether identity verification is triggered, and controls the crawling depth under dual-objective constraints based on the data acquisition loss and data acquisition satisfaction. The queue processing and analysis module is connected to the data crawling module and is used to perform queue processing and analysis, obtain the similarity balance of related items among the service providers, and determine the load reference value or similarity reference difference as the constraint processing condition based on the similarity balance of related items to obtain several analysis queues. The text selection module is connected to the data crawling module and the queue processing and analysis module respectively. It is used to determine the coverage of domain drift content based on historical data support and data volume, and to determine whether to change the data analysis method from general sentiment analysis to adaptive representation analysis based on the coverage of domain drift content in the analysis queue, so as to obtain a number of texts to be selected. The credit assessment module, which is connected to the filter text selection module, is used to conduct credit assessments based on the filter texts and assessment items of each service provider.

[0006] Furthermore, if the data crawling module's response cycle working time exceeds the preset cycle working time or if an authentication condition is triggered, it will determine to adjust the crawling process. The crawling adjustment includes randomly inserting simulated events, adjusting the click frequency in reverse preset mode, and intermittent stop control.

[0007] Furthermore, the data crawling module responds to the condition that the data acquisition loss is greater than the preset data acquisition loss or the data acquisition satisfaction is greater than the preset data acquisition satisfaction, determines that the crawling depth has reached the boundary value under the dual-objective constraint, and triggers the crawling termination condition.

[0008] Furthermore, the data acquisition loss is determined based on the number of consecutive failures during the data crawling process; The data acquisition satisfaction level is determined based on the amount of data acquired and the required amount of data acquired.

[0009] Furthermore, the queue processing analysis module determines the analysis queue based on the load reference value as a constraint processing condition when the similarity balance of related items is greater than the preset similarity balance of related items. The queue processing and analysis module determines the queue based on the similarity reference difference as a constraint condition when the similarity balance of related items is less than or equal to the preset similarity balance of related items.

[0010] Furthermore, the queue processing analysis module uses the similarity reference difference as a constraint processing condition; In this analysis, a single analysis queue contains several service providers, and the similarity reference difference between the service providers in a single analysis queue is greater than the preset similarity reference difference. The queue processing and analysis module determines the priority coefficient of each analysis queue based on the estimated compensation coefficient, and determines the filter text of the service provider in each analysis queue in descending order of priority coefficient. After the filter text of a single analysis queue is determined, keyword compensation is performed on that analysis queue.

[0011] Furthermore, the text selection module responds to the condition that the historical data support is greater than the preset historical data support and the data volume is greater than the preset data volume, and the domain drift content coverage is determined based on historical keywords; The text selection module responds to conditions where the historical data support is less than or equal to a preset historical data support or the data volume is less than or equal to a preset data volume. The domain drift content coverage is determined based on common sentiment words.

[0012] Furthermore, the text selection module responds to the condition that the domain drift content coverage is greater than the preset domain drift content coverage, and determines to adjust the data analysis method from general sentiment analysis screening to adaptive representation analysis screening.

[0013] Furthermore, the text selection module performs adaptive representation analysis filtering; The text selection module determines the text representation difference degree based on the map difference reference value, and selects texts for each service provider in the analysis queue based on the text representation difference degree.

[0014] Furthermore, the text selection module performs general sentiment analysis filtering; Among these methods, a pre-trained AI model is used to select and filter text based on the sentiment scores of keywords in the crawled data.

[0015] Compared with the prior art, the beneficial effects of the present invention are that, in the technical solution of the present invention, the crawling adjustment is determined by the collaborative determination of the periodic working time and the identity verification trigger state, and a crawling depth control mechanism with dual objective constraints is constructed based on the data acquisition loss degree and the data acquisition satisfaction degree, so that the crawling strategy can dynamically adapt to the anti-crawling strength and data collection progress of different target websites, avoid the continuous blocking caused by fixed frequency collection or the inefficiency caused by excessive conservatism, and achieve a precise balance between collection efficiency and resource consumption.

[0016] Furthermore, this invention adaptively switches between load reference value and similarity reference difference as queue constraint processing conditions by comparing the similarity balance of related items with a preset threshold. In the similarity reference difference mode, the priority coefficient and keyword compensation strategy of the analysis queue are determined according to the estimated compensation coefficient. This enables the service provider grouping method to be dynamically adjusted according to the homogeneity and heterogeneity of the group. While ensuring the system load balance, it improves the completeness of credit feature representation and effectively suppresses the negative impact of information redundancy or feature loss on the accuracy of evaluation.

[0017] Furthermore, this invention uses a comprehensive approach, combining historical data support and data volume, to determine the calculation benchmark for domain-drift content coverage. Based on the comparison between domain-drift content coverage and a preset threshold, the data analysis method is adaptively adjusted from general sentiment analysis filtering to adaptive representation analysis filtering based on graph difference reference values. This allows the system to reuse highly targeted historical keywords when historical data is abundant, and to achieve unsupervised filtering using general sentiment words or knowledge graphs when historical data is sparse. This effectively overcomes the problem of missing evaluation content caused by domain-drift data and significantly improves the robustness of credit assessment in cross-domain scenarios.

[0018] Furthermore, in the adaptive representation analysis screening mode of this invention, the graph difference reference value corresponding to each keyword node in the crawled data is calculated by pre-trained knowledge graph, thereby determining the text representation difference degree. Based on the text representation difference degree, text paragraphs that are closely related to the core system of credit assessment are screened. This makes the screening process no longer rely on the ability of general sentiment models to distinguish non-evaluative content, but rather measures the degree of matching between the text and the credit domain from the knowledge structure level, significantly reducing the risk of invalid data mixing due to domain drift and improving the credit feature representativeness of the screened text.

[0019] Furthermore, this invention employs a collaborative mechanism of incremental updates and keyword compensation. When grouping based on load reference values, high-frequency keywords are uniformly extracted and stored across all analysis queues. When grouping based on similarity reference differences, historical keywords are compensated in real-time for each queue according to priority coefficients. This enables the construction and utilization of the historical knowledge base to match the distribution characteristics of service provider groups under different grouping modes. It avoids resource waste caused by a single compensation strategy when the groups are highly homogeneous or data barriers caused by highly heterogeneous groups, achieving a dual optimization of knowledge transfer efficiency and system processing fairness. Attached Figure Description

[0020] Figure 1 This is a module connection diagram of the outsourcing service provider credit assessment system based on AI large model multi-source data fusion, as an embodiment of the present invention. Figure 2 This is a flowchart illustrating how to determine whether to perform crawling adjustments based on the cycle duration and whether identity verification is triggered, according to an embodiment of the present invention. Figure 3 This is a flowchart illustrating the process of determining the similarity balance of related items based on load reference values ​​or similarity reference differences as constraint conditions in an embodiment of the present invention. Figure 4 This is a flowchart illustrating how, according to an embodiment of the present invention, a method for determining whether to change the data analysis method from general sentiment analysis filtering to adaptive representation analysis filtering is adjusted based on the domain drift content coverage of the analysis queue. Detailed Implementation

[0021] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0022] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0023] Please see Figure 1 The diagram shown is a module connection diagram of an outsourcing service provider credit assessment system based on AI large-scale model multi-source data fusion according to an embodiment of the present invention. This embodiment of the present invention provides an outsourcing service provider credit assessment system based on AI large-scale model multi-source data fusion, comprising: The data crawling module is used to crawl data from various service providers at various target URLs to obtain crawled data. Specifically, for any crawling task of any service provider at any target URL, the module determines whether to perform crawling adjustment based on the cycle duration and whether identity verification is triggered, and controls the crawling depth under dual-objective constraints based on the data acquisition loss and data acquisition satisfaction. The queue processing and analysis module is connected to the data crawling module and is used to perform queue processing and analysis, obtain the similarity balance of related items among the service providers, and determine the load reference value or similarity reference difference as the constraint processing condition based on the similarity balance of related items to obtain several analysis queues. The text selection module is connected to the data crawling module and the queue processing and analysis module respectively. It is used to determine the coverage of domain drift content based on historical data support and data volume, and to determine whether to change the data analysis method from general sentiment analysis to adaptive representation analysis based on the coverage of domain drift content in the analysis queue, so as to obtain a number of texts to be selected. The credit assessment module, which is connected to the filter text selection module, is used to conduct credit assessments based on the filter texts and assessment items of each service provider.

[0024] In this embodiment of the invention, the service provider, which is the object of credit assessment, includes several types of outsourcing companies, including but not limited to software development service providers, human resources outsourcing service providers, logistics and distribution service providers, and data labeling service providers. The target URL refers to the address of the internet data source used to collect relevant information from service providers. Target URLs include, but are not limited to, enterprise credit information disclosure systems, bidding information websites, industry forums, social media platforms, news websites, and government regulatory announcement pages. A single crawled data item is the text consisting of all paragraphs containing the service provider's name that appear on the page corresponding to a single network link collected by the data crawling module from each target URL; A web crawling task refers to a complete data crawling operation performed on a single target URL for a single service provider. This complete data crawling operation includes: first, initializing the task by loading pre-configured control parameters such as the baseline interval and click threshold, and resetting and initializing the built-in counter; then, initiating a compliant network access request to search for the webpages of the service provider appearing on the target URL. This is easily understood by those skilled in the art and will not be elaborated further. In this embodiment, the baseline interval is set to 2.0 seconds, serving as the basic waiting time between two adjacent operations; the click threshold is set to 10 times, serving as the upper limit for the number of consecutive operations that trigger intermittent stop control; the initial value of the continuous operation counter is set to 0, and it increments by 1 after each successful click.

[0025] Please see Figure 2 As shown, it is a flowchart of an embodiment of the present invention for determining whether to perform crawling adjustment based on the periodic working time and whether identity verification is triggered.

[0026] Specifically, if the data crawling module's response cycle working time exceeds the preset cycle working time or if the authentication condition is triggered, it will determine to adjust the crawling process. The crawling adjustment includes randomly inserting simulated events, adjusting the click frequency in reverse preset mode, and intermittent stop control.

[0027] In this embodiment of the invention, the periodic working time is the total time for crawling data from a single target URL for a single service provider; Triggering authentication means that the target website needs to further verify the identity of the current access behavior. Specifically, it includes one or more of the following situations: (1) The response status code is 401 Unauthorized or 403 Forbidden, indicating that the current request lacks valid identity credentials; (2) A verification code input box appears on the response page, including graphic verification code, sliding verification code, click verification code or SMS verification code; (3) The response page contains keywords such as continue access after login, authentication failure, session expired, etc.; (4) The response page is redirected to the login page or identity authentication page, and its URL contains characteristic fields such as login, auth, verify, etc.; (5) The expected data field is missing in the response content, and anti-crawling prompt information appears. The anti-crawling prompt information includes, but is not limited to, please complete the verification and abnormal access detected.

[0028] In this embodiment of the invention, the periodic working time reflects the total time cost required for a single service provider to complete a full data crawling operation at a single target URL. For application scenarios with higher requirements for data collection timeliness, the greater the triggering demand for crawling adjustment, the smaller the value of the preset periodic working time, so as to intervene in adjustment and control task time consumption earlier. Preferably, historical credit assessment logs that meet the assessment requirements are obtained, and the average value of the periodic working time corresponding to each historical credit assessment log is calculated as 180s, and this average value is recorded as the preset periodic working time.

[0029] In this embodiment, meeting the assessment requirements means that the marking rate is greater than a preset marking rate. The marking rate is based on the ratio of the number of marked service providers in a single historical credit assessment log to the total number of service providers used for credit assessment in that historical credit assessment log. The number of marked service providers refers to the number of service providers whose scores, given by experts, are higher than a preset score for the completeness and accuracy of the selected text used in the credit assessment. In this embodiment, the preset marking rate and preset score can be configured by the user; the preset marking rate is set to 95%, and the preset score is set to 90 points. The historical credit assessment log is a complete record of the entire process of data crawling, queue processing and analysis, text selection, and credit assessment for several service providers.

[0030] In this embodiment of the invention, the periodic working time and whether authentication is triggered reflect the intensity of anti-crawling pressure faced by the current crawling task and the blocking tendency of the target website. Therefore, the data crawling module compares the periodic working time with the corresponding preset threshold and combines it with whether authentication is triggered to determine whether to adjust the crawling, thus avoiding the problems of being continuously blocked due to a fixed crawling frequency or low collection efficiency due to overly conservative adjustment.

[0031] In this embodiment of the invention, the random insertion of simulated events includes: during the continuous crawling process, the system randomly intersperses simulated user interaction events such as page scrolling, cursor hovering, and page jumping. The timing and duration of the event triggering are randomly generated to mimic real human browsing habits and weaken the characteristics of automated programs. The reverse preset adjustment of click frequency includes: real-time collection and statistics of the number of clicks per unit time; when the actual click frequency is not within the preset frequency range, the operation frequency parameter is adjusted in reverse so that the overall click frequency returns to the preset frequency range; in this embodiment of the invention, the system is configured with a unit time of 10 seconds and the preset frequency range is set to [0.5, 1.5] times / second, which can be adaptively adjusted according to the anti-crawling strength of the target website.

[0032] Please continue reading. Figure 1 and Figure 2As shown in this embodiment of the invention, the data crawling module responds to the condition that the data acquisition loss is greater than the preset data acquisition loss or the data acquisition satisfaction is greater than the preset data acquisition satisfaction, determines that the crawling depth has reached the boundary value under the dual-objective constraint, and triggers the crawling termination condition.

[0033] Specifically, the data acquisition loss is determined based on the number of consecutive failures during the data crawling process; The data acquisition satisfaction level is determined based on the amount of data acquired and the required amount of data acquired.

[0034] In this embodiment of the invention, the data acquisition loss degree, for a single target URL and a single service provider, is determined based on the ratio of the number of consecutive failures to a preset number of consecutive failures, reflecting the degree of data loss caused by anti-crawling interception in the current crawling task; The consecutive failure count, which is the number of times identity verification is triggered within a unit of time after a compliant network access request is initiated to the target website to search for the webpages of the service provider, reflects the frequency with which the current crawling task is blocked by the target website's anti-crawling mechanism. In this embodiment of the invention, the preset consecutive failure count is 3 times, and the preset data acquisition loss degree is 1, reflecting that the degree of data loss caused by anti-crawling interception in the current crawling task has reached a critical level.

[0035] The data acquisition satisfaction is determined based on the ratio of the amount of data acquired to the amount of data required to be acquired. It is used to reflect the completion ratio and progress status of the amount of data completed in the current crawling task relative to the expected target amount of data. In this embodiment of the invention, the amount of data acquired is the number of pages crawled for the service provider to search for the target URL, and the amount of data required to be acquired is the average number of filtered pages corresponding to each marked service provider in each historical credit assessment log. In this embodiment of the invention, a higher data acquisition satisfaction level indicates that the amount of data completed in the current crawling task is closer to or exceeds the expected target data amount, the crawling progress is more ideal, and the remaining data to be crawled is less. Therefore, the greater the demand for application scenarios with high data acquisition completion rates, the smaller the preset data acquisition satisfaction level should be, so that the crawling deadline condition is triggered earlier when the crawling task is nearing completion, avoiding excessive crawling and resource waste. Specifically, historical crawling records from outsourced service providers with different data acquisition satisfaction levels are collected, with preset data acquisition satisfaction levels of 3.5, 3.0, 2.5, 2.0, and 1.5, respectively. The effective utilization rate of the collected data and the accuracy of credit assessment after triggering the crawling deadline condition under different preset values ​​are evaluated. Simulation results show that when the preset data acquisition satisfaction level is 2.0, the data acquisition integrity and system resource consumption reach the optimal balance. Therefore, the preset data acquisition satisfaction level is set to 2.0.

[0036] Please see Figure 3As shown, it is a flowchart of the process of determining the similarity balance of related items based on the load reference value or the similarity reference difference as a constraint condition in an embodiment of the present invention.

[0037] Specifically, the queue processing analysis module determines the queue based on the load reference value as a constraint processing condition when the similarity balance of related items is greater than the preset similarity balance of related items. The queue processing and analysis module determines the queue based on the similarity reference difference as a constraint condition when the similarity balance of related items is less than or equal to the preset similarity balance of related items.

[0038] In this embodiment of the invention, the similarity balance of related items is determined based on the average of the similarity reference values ​​corresponding to each service provider, reflecting the overall similarity among all service providers; in this embodiment of the invention, for a single service provider, its corresponding similarity reference value is the average of the sub-similarity reference values ​​corresponding to that service provider and those of other service providers. For any two service providers, their corresponding sub-similarity reference values ​​are determined based on relevant item scores to reflect the degree of similarity between the two service providers. In this embodiment of the invention, the sub-similarity reference value is determined based on the difference between a benchmark coefficient and a difference ratio. The benchmark coefficient is 1, and the difference ratio is determined based on the ratio of the similarity reference difference to the larger relevant item score of the two service providers. The similarity reference difference is determined based on the absolute value of the difference between the relevant item scores of the two service providers. For a single service provider, its relevant item score is determined based on the weighted sum of the scores corresponding to each relevant item, with each relevant item having a weight coefficient of 0.25. The relevant items include basic qualifications, credit record, service quality, and financial status. The score for each relevant item is obtained by experts scoring the initial information submitted by the service provider. The highest score for a single relevant item is 100 points. In this embodiment of the invention, each service provider has an initial submission information, which is the material submitted by the service provider itself. The initial submission information includes: basic qualification information, including basic enterprise information, business license and qualification certificate; credit record information, including performance record, default record, complaint record and legal proceedings; service quality information, including project performance, customer feedback and after-sales record; and financial status information, including financial statements, revenue data and liabilities.

[0039] In this embodiment of the invention, a higher correlation similarity balance indicates a better overall similarity among service providers in terms of correlation dimensions, a more concentrated group distribution, and stronger homogeneity among service providers. When the service provider group is highly homogeneous, each service provider tends to perform consistently in terms of basic qualifications, credit records, service quality, and financial status, resulting in high information redundancy between queues. In this case, keyword compensation has limited effect on improving the credit feature coverage of subsequent screening texts. Therefore, the greater the demand for keyword compensation to improve the credit feature coverage of subsequent screening texts, the larger the value of the preset correlation similarity balance. Preferably, historical credit assessment logs that meet the assessment requirements and are based on similarity reference differences as constraints are obtained, and the maximum value of the correlation similarity balance among the historical credit assessment logs is 0.67. This maximum value is recorded as the preset correlation similarity balance.

[0040] In this embodiment of the invention, when the analysis queue is determined to be based on the load reference value as a constraint processing condition, the queue processing analysis module sorts each service provider in descending order of load, and sequentially selects and accumulates the load of the selected service providers as the load reference value until the load reference value is greater than the preset load reference value. Then, all service providers that were not recorded in the analysis queue before that service provider are recorded in the analysis queue, and the above process is repeated for the remaining service providers until all service providers are assigned. For a single service provider, the load capacity is the total number of web pages crawled for that service provider at each target URL. In this embodiment of the invention, the load reference value, for a group of service providers, is the sum of the loads of these service providers, reflecting the total data collection volume of the group of service providers. The larger the load reference value, the higher the system resource consumption. Therefore, the greater the demand for system resource load balancing, the smaller the preset load reference value should be. Preferably, historical credit assessment logs that meet the assessment requirements and are processed based on the load reference value as a constraint condition are obtained. The average load reference value of each analysis queue in each historical credit assessment log is calculated to be 247, and this average value is recorded as the preset load reference value.

[0041] Specifically, the queue processing analysis module uses the similarity reference difference as a constraint processing condition; In this analysis, a single analysis queue contains several service providers, and the similarity reference difference between the service providers in a single analysis queue is greater than the preset similarity reference difference. The queue processing and analysis module determines the priority coefficient of each analysis queue based on the estimated compensation coefficient, and determines the filter text of the service provider in each analysis queue in descending order of priority coefficient. After the filter text of a single analysis queue is determined, keyword compensation is performed on that analysis queue.

[0042] In this embodiment of the invention, for a single analysis queue, the similarity reference difference between the service provider outside the analysis queue and at least one service provider in the analysis queue is less than or equal to a preset similarity reference difference; In this embodiment of the invention, a larger similarity reference difference indicates a more significant difference between the two service providers in the relevant item dimension, weaker homogeneity, and stronger complementarity of the compensation dimensions between the analysis queue and subsequent analysis queues. Therefore, the greater the demand for enriching credit feature dimensions through keyword compensation between queues, the larger the preset similarity reference difference value should be. Preferably, historical credit assessment logs that meet the assessment requirements and are processed based on the similarity reference difference as a constraint condition are obtained. The average value of the similarity reference difference between any two service providers in each analysis queue of each historical credit assessment log is calculated to be 34, and this average value is recorded as the preset similarity reference difference value.

[0043] The estimated compensation coefficient, for a single analysis cohort, is determined based on the ratio of the average number of differentiated service providers to the average total number of service providers. It is used to reflect the degree of difference between this analysis cohort and other cohorts and the potential value of obtaining incremental information through keyword compensation. In this embodiment of the invention, the average number of service providers is the average number of service providers corresponding to each reference queue, the average number of different service providers is the average number of different service providers corresponding to the analysis queue and other reference queues, and the reference analysis queues are the other analysis queues besides the single analysis queue. Specifically, for a single reference queue, the number of corresponding difference service providers is the number of service providers in the reference queue whose similarity reference difference with at least one service provider in the single analysis queue is greater than a preset similarity reference difference.

[0044] In this embodiment of the invention, the queue processing and analysis module determines the priority coefficient of each analysis queue based on the estimated compensation coefficient. The priority coefficient of a single analysis queue is equal to the estimated compensation coefficient of that analysis queue. The larger the priority coefficient, the higher the priority of the analysis queue in selecting the text to be filtered. When performing keyword compensation for a single analysis queue, keywords in that analysis queue that appear more frequently than a preset frequency in all filtered texts corresponding to a single service provider are stored as historical keywords in the system database.

[0045] It should be noted that if the queue processing analysis module determines the processing condition based on the load reference value, the priority coefficient of each analysis queue is not limited, and the processing order of the analysis queues can be randomly set. Furthermore, in this mode, there is no successive compensation for the filtered text of a single analysis queue. Instead, after all analysis queues have completed the selection of filtered text, a unified incremental update operation is performed. This incremental update involves storing keywords that appear more frequently than a preset frequency in all analysis queues as historical keywords in the system database.

[0046] In this embodiment of the invention, constraint processing conditions are determined by comparing the similarity balance of related items with a preset similarity balance of related items. When the similarity balance of related items is greater than the preset similarity balance of related items, it indicates that the service providers are highly homogeneous, and the service provider group cannot be effectively distinguished based on the similarity reference difference. Grouping based on load can achieve a balanced allocation of system resources. When the similarity balance of related items is less than or equal to the preset similarity balance of related items, it indicates that the service providers are highly heterogeneous and the group distribution is scattered. Fine-grained grouping based on the similarity reference difference and subsequent keyword compensation are more necessary, thereby achieving adaptive switching of constraint processing conditions. It can adaptively select grouping strategies according to the inherent similarity of the service provider group, improving the completeness of credit feature representation while ensuring processing efficiency, and thus achieving a dynamic balance between system resource utilization and evaluation accuracy.

[0047] Please continue reading. Figures 1 to 3 As shown in this embodiment of the invention, the text selection module responds to the condition that the historical data support is greater than the preset historical data support and the data volume is greater than the preset data volume, and the domain drift content coverage is determined based on historical keywords. The text selection module responds to conditions where the historical data support is less than or equal to a preset historical data support or the data volume is less than or equal to a preset data volume. The domain drift content coverage is determined based on common sentiment words.

[0048] In this embodiment of the invention, for a single service provider, its corresponding historical data support level is determined based on the ratio of the number of service providers of the same outsourcing enterprise type as the service provider in the historical credit assessment log to the total number of service providers in the historical credit assessment log. This reflects the richness of the sample of the enterprise type to which the service provider belongs in the historical data. The higher the historical data support level, the more sufficient the sample is, and the higher the reliability of the analysis based on historical keywords. Therefore, the greater the demand for sufficient coverage of historical data samples, the larger the value of the preset historical data support level. Preferably, historical credit assessment logs that meet the assessment requirements are obtained, and the average value of the historical data support level corresponding to each service provider, which is determined based on historical keywords for the domain drift content coverage in each historical credit assessment log, is 0.34. This average value is recorded as the preset historical data support level.

[0049] In this embodiment of the invention, for a single service provider, the corresponding data volume is the total number of service providers in the historical credit assessment log, which is used to reflect the overall sample size and data coverage of the historical credit assessment log. Therefore, the greater the demand for sufficient historical data sample size, the larger the value of the preset data volume. Preferably, historical credit assessment logs that meet the assessment requirements are obtained, and the average data volume of each service provider corresponding to the domain drift content coverage determined based on historical keywords in each historical credit assessment log is 156. This average value is recorded as the preset data volume.

[0050] Domain drift content coverage, which is determined based on historical keywords or general sentiment keywords, is used to reflect the proportion of non-credit assessment core domain content in all page text crawled for a single service provider, and is used to quantify the severity of data domain drift. In this embodiment of the invention, when the domain-shifted content coverage is determined based on historical keywords, the domain-shifted content coverage is determined based on the difference between the baseline coefficient and the proportion of historical keywords; wherein, for a single service provider, the proportion of its corresponding historical keywords is determined based on the ratio of the number of historical keywords appearing in all pages crawled for that service provider to the total number of keywords in all pages crawled for that service provider; it should be noted that if a single keyword appears multiple times, it is only counted once; The historical keywords are: high-frequency keywords corresponding to each service provider in each historical credit assessment log, and high-frequency keywords corresponding to a single service provider, which are keywords that appear more frequently than a preset frequency in all the filtered texts corresponding to that service provider; the frequency of occurrence is the total number of times a single keyword appears in all the filtered texts corresponding to a single service provider. In this embodiment of the invention, the preset frequency of occurrence is set to 5 times. In this embodiment of the invention, the historical keywords are stored in the system database; When the domain-shifted content coverage is determined based on general sentiment words, the domain-shifted content coverage is determined based on the difference between the baseline coefficient and the proportion of general sentiment words; wherein, for a single service provider, the proportion of general sentiment words is determined based on the ratio of the number of general sentiment words appearing in all pages crawled for that service provider to the total number of keywords in all pages crawled for that service provider. The general sentiment words are determined based on a general domain pre-trained language model. For a single keyword, if the model outputs a sentiment confidence score greater than a preset threshold, then the keyword is a general sentiment word. In this embodiment of the invention, the preset threshold is preferably 0.7. The general domain pre-trained language model is trained by several keywords with sentiment confidence scores. The sentiment confidence scores are determined by expert scoring. The training of the model is a common technique used by those skilled in the art, and will not be elaborated on in detail.

[0051] In this embodiment of the invention, historical data support and data volume are used to reflect the richness of samples of similar service providers and the overall sample size in historical data. Therefore, the text selection module compares historical data support and data volume with corresponding preset values ​​to determine which parameters to use to determine the coverage of domain drift content. When the historical data support level is greater than the preset historical data support level and the data volume is greater than the preset data volume, it indicates that the historical data is sufficient and reliable. Determining based on historical keywords has high accuracy and domain specificity, so it is determined based on historical keywords. When the historical data support level is less than or equal to the preset historical data support level or the data volume is less than or equal to the preset data volume, it indicates that the historical data is sparse or insufficient, lacking sufficient historical keyword support. Therefore, it is determined based on general sentiment words. Determining domain drift content coverage based on general sentiment words can infer the severity of the text's deviation from the core domain of credit assessment by measuring the sparsity of sentiment expression in the text without relying on historical data. The lower the proportion of general sentiment words, the more objective and non-evaluative the text is, and the more severe the domain drift. This allows for a comprehensive evaluation of the usability of historical data from two dimensions: historical data support level and data volume. It enables adaptive switching of keyword selection strategies and avoids the problem of being unable to calculate domain drift content coverage or having unreliable calculation results due to insufficient historical data.

[0052] Please see Figure 4 As shown, it is a flowchart of an embodiment of the present invention for determining whether to adjust the data analysis method from general sentiment analysis screening to adaptive representation analysis screening based on the domain drift content coverage of the analysis queue.

[0053] Specifically, the text selection module responds to the condition that the coverage of domain drift content is greater than the preset coverage of domain drift content, and determines to adjust the data analysis method from general sentiment analysis screening to adaptive representation analysis screening.

[0054] Specifically, the text selection module determines the data analysis method as general sentiment analysis filtering when the domain drift content coverage is less than or equal to the preset domain drift content coverage.

[0055] In this embodiment of the invention, a larger domain drift content coverage indicates a smaller proportion of evaluative content related to credit assessment in the crawled text, and a more severe data domain drift phenomenon. This leads to the inability of the general sentiment analysis screening text to effectively represent the service provider's credit characteristics. Therefore, the greater the requirement for representativeness of the screening text, the smaller the preset domain drift content coverage value. Preferably, historical credit assessment logs that meet the assessment requirements and whose data analysis method has been adjusted from general sentiment analysis screening to adaptive representation analysis screening are obtained. The average value of the domain drift content coverage corresponding to each historical credit assessment log is 0.52, and this average value is recorded as the preset domain drift content coverage.

[0056] Specifically, the text selection module performs adaptive representation analysis filtering; The text selection module determines the text representation difference degree based on the map difference reference value, and selects texts for each service provider in the analysis queue based on the text representation difference degree.

[0057] In this embodiment of the invention, when selecting filter text for a single service provider in the analysis queue based on text representation difference, the crawled data of that service provider with a text representation difference less than a preset text representation difference is used as the filter text. The text representation difference degree, which is determined based on the graph difference reference value, is used to reflect the degree of deviation between the crawled text expression and the credit assessment system, and to characterize the effective usability of the evaluative content information in the text. In this embodiment of the invention, the text representation difference degree is determined based on the ratio of the graph difference reference value to the preset graph difference reference value. Specifically, for a single crawled data point, the corresponding graph difference reference value is the ratio of the total number of nodes contained in the minimum path that can contain each node in the pre-trained knowledge graph for each keyword in the crawled data to the number of nodes corresponding to each keyword in the crawled data. This reflects the degree of semantic association dispersion and domain drift level between the crawled data keywords and the pre-trained knowledge graph. The larger the graph difference reference value, the longer the association link of the keyword node in the graph and the worse the semantic matching degree. Therefore, the greater the demand for improving the quality of effective evaluation data collection, the smaller the preset graph difference reference value should be. Preferably, historical credit evaluation logs that meet the evaluation requirements are obtained, and the average value of the graph difference reference value of each selected text corresponding to each service provider that has undergone adaptive representation analysis in each historical credit evaluation log is 0.56, which is recorded as the preset graph difference reference value.

[0058] Among them, the pre-trained knowledge graph is obtained by collecting compliant and publicly available credit industry labeled texts, corporate performance public opinion, and user credit evaluation samples, extracting entity keywords, attributes and relationships between entities, and relying on graph neural networks to complete entity linking, relationship alignment and graph iterative training. This is a commonly used technical method in this field and will not be elaborated on in detail.

[0059] In this embodiment of the invention, the greater the text representation difference, the more significantly the crawled text sentiment content deviates from the credit assessment benchmark, the lower the proportion of effective evaluation sentiment data, and the more prominent the data domain drift problem. Therefore, the greater the need to optimize the crawling strategy and increase the amount of effective evaluation data collected, the smaller the value of the preset text representation difference. Preferably, historical credit assessment logs that meet the assessment requirements are obtained, and the average value of the text representation difference of each text corresponding to each service provider that performs adaptive representation analysis in each historical credit assessment log is 0.66, which is recorded as the preset text representation difference.

[0060] Specifically, the text selection module performs general sentiment analysis filtering; Among these methods, a pre-trained AI model is used to select and filter text based on the sentiment scores of keywords in the crawled data.

[0061] In this embodiment of the invention, when selecting and filtering texts based on the sentiment scores of keywords in crawled data using a pre-trained AI model, for a single service provider, each crawled data of that service provider with a sentiment score greater than a preset sentiment score is used as the filtered text. The pre-trained AI model is a general-domain pre-trained language model based on the Transformer architecture, used to determine the sentiment tendency of input text and output a sentiment score. The sentiment score is a quantitative value of sentiment output by the model, ranging from [0,1], used to represent the positive or negative polarity of the text's sentiment. The preset sentiment score is a pre-set sentiment filtering threshold by the system; in this embodiment, it is preferably set to 0.6, meaning only crawled data with a sentiment score greater than 0.6 is retained as filtered text. The training of the pre-trained AI model is based on... The pre-trained AI model is trained based on a general-domain sentiment annotation dataset, including but not limited to SST-2 and IMDB. Supervised fine-tuning optimizes the model parameters to enable it to output text sentiment scores. The specific training process is a common technique used by those skilled in the art and will not be elaborated upon here.

[0062] In this embodiment of the invention, the credit assessment module performs credit assessments based on the screening texts and assessment items of each service provider. The credit assessment module provides a manual assessment interface, allowing expert users to log in to the system and view the screening texts and detailed data of each assessment item for each service provider. Experts then determine the credit score of each service provider based on their professional judgment. The system does not impose any restrictions on the specific scoring standards and rules.

[0063] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A credit assessment system for outsourcing service providers based on AI large-scale model and multi-source data fusion, characterized in that, include: The data crawling module is used to crawl data from various service providers at various target URLs to obtain crawled data. Specifically, for any crawling task of any service provider at any target URL, the module determines whether to perform crawling adjustment based on the cycle duration and whether identity verification is triggered, and controls the crawling depth under dual-objective constraints based on the data acquisition loss and data acquisition satisfaction. The queue processing and analysis module is connected to the data crawling module and is used to perform queue processing and analysis, obtain the similarity balance of related items among the service providers, and determine the load reference value or similarity reference difference as the constraint processing condition based on the similarity balance of related items to obtain several analysis queues. The text selection module is connected to the data crawling module and the queue processing and analysis module, respectively. It is used to determine the coverage of domain drift content based on historical data support and data volume, and to determine whether to change the data analysis method from general sentiment analysis to adaptive representation analysis based on the coverage of domain drift content in the analysis queue, so as to obtain a number of texts to be selected. The credit assessment module, which is connected to the filter text selection module, is used to conduct credit assessments based on the filter texts and assessment items of each service provider.

2. The outsourcing service provider credit assessment system based on AI large-scale model multi-source data fusion as described in claim 1, characterized in that, If the data crawling module's response cycle working time exceeds the preset cycle working time or if the authentication condition is triggered, crawling adjustment will be performed. The crawling adjustment includes randomly inserting simulated events, adjusting the click frequency in reverse preset mode, and intermittent stop control.

3. The outsourcing service provider credit assessment system based on AI large-scale model multi-source data fusion as described in claim 2, characterized in that, The data crawling module responds to the condition that the data acquisition loss is greater than the preset data acquisition loss or the data acquisition satisfaction is greater than the preset data acquisition satisfaction, determines that the crawling depth has reached the boundary value under the dual-objective constraint, and triggers the crawling termination condition.

4. The outsourcing service provider credit assessment system based on AI large-scale model multi-source data fusion as described in claim 3, characterized in that, The data acquisition loss rate is determined based on the number of consecutive failures during the data crawling process; The data acquisition satisfaction level is determined based on the amount of data acquired and the required amount of data acquired.

5. The outsourcing service provider credit assessment system based on AI large-scale model multi-source data fusion as described in claim 1, characterized in that, The queue processing and analysis module determines the queue based on the load reference value as a constraint processing condition when the similarity balance of related items is greater than the preset similarity balance of related items. The queue processing and analysis module determines the queue based on the similarity reference difference as a constraint condition when the similarity balance of related items is less than or equal to the preset similarity balance of related items.

6. The outsourcing service provider credit assessment system based on AI large-scale model multi-source data fusion as described in claim 5, characterized in that, The queue processing and analysis module uses the similarity reference difference as a constraint processing condition; In this analysis, a single analysis queue contains several service providers, and the similarity reference difference between the service providers in a single analysis queue is greater than the preset similarity reference difference. The queue processing and analysis module determines the priority coefficient of each analysis queue based on the estimated compensation coefficient, and determines the filter text of the service provider in each analysis queue in descending order of priority coefficient. After the filter text of a single analysis queue is determined, keyword compensation is performed on that analysis queue.

7. The outsourcing service provider credit assessment system based on AI large-scale model multi-source data fusion as described in claim 1, characterized in that, The text selection module responds to the condition that the historical data support is greater than the preset historical data support and the data volume is greater than the preset data volume, and the domain drift content coverage is determined based on historical keywords. The text selection module responds to conditions where the historical data support is less than or equal to a preset historical data support or the data volume is less than or equal to a preset data volume. The domain drift content coverage is determined based on common sentiment words.

8. The outsourcing service provider credit assessment system based on AI large-scale model multi-source data fusion according to claim 7, characterized in that, The text selection module responds to the condition that the coverage of domain drift content is greater than the preset coverage of domain drift content, and determines to adjust the data analysis method from general sentiment analysis filtering to adaptive representation analysis filtering.

9. The outsourcing service provider credit assessment system based on AI large-scale model multi-source data fusion as described in claim 8, characterized in that, The text selection module performs adaptive representation analysis filtering. The text selection module determines the text representation difference degree based on the map difference reference value, and selects texts for each service provider in the analysis queue based on the text representation difference degree.

10. The outsourcing service provider credit assessment system based on AI large-scale model multi-source data fusion according to claim 9, characterized in that, The text selection module performs general sentiment analysis filtering. Among these methods, a pre-trained AI model is used to select and filter text based on the sentiment scores of keywords in the crawled data.

Images