A review expert library random extraction evaluation method

Abstract

The application discloses a kind of review expert database random extraction evaluation methods.The method relates to intelligent matching extraction technical field, including the following steps: multi-source data acquisition and vectorization processing, matching accuracy dynamic adjustment, data assetization closed loop transformation and regulation and expert random extraction execution.The application is by collecting review related data and document and carries out vectorization processing, it is to unstructured, semi-structured data into structured vector data, obtains expert and review matter matching accuracy, and judges whether to carry out dynamic adjustment, then carries out expert and review project correlation analysis.Based on correlation analysis acquisition data assetization closed loop transformation degree, and judges whether to carry out closed loop adjustment, finally completes expert random extraction, realizes review matching precision, data transformation efficiency, extraction process is stable and reliable, improve a kind of review expert database random extraction evaluation reliability, solve the low reliability of a kind of review expert database random extraction evaluation in prior art.

Description

Technical Field

[0001] This invention relates to the field of intelligent matching and extraction technology, and in particular to a method for randomly selecting and evaluating experts from a pool of reviewers. Background Technology

[0002] First, the purchaser initiates the extraction of requirements on the electronic trading platform, and completes strong identity authentication and encrypted transmission of requirement instructions through CA (Certificate Authority) digital certificates and electronic signature technology to ensure the legitimacy and non-repudiation of the initiating party. Subsequently, based on the project characteristics, the structured database query language is invoked to perform multi-dimensional conditional screening (such as professional category, professional qualification, avoidance situation, etc.) in the expert database to lock the candidate pool. Then, the core random selection process is launched. The system backend no longer uses the basic pseudo-random function, but adopts a high-intensity true random algorithm based on hardware noise or quantum random number generator, combined with distributed cluster computing technology, to complete the "lottery" selection from the candidate pool in milliseconds, ensuring that the probability of each qualified expert being selected is mathematically absolutely equal and the process is unpredictable. After the extraction results are generated, the intelligent voice outbound call and SMS push module is automatically triggered. The NLP (Natural Language Processing) technology is used to perform semantic recognition and confirmation of the expert's response. For experts who refuse or do not respond, the system immediately starts the supplementary extraction program according to the preset automatic iteration logic until the number of people required for review is met. Finally, during the strict confidentiality phase before the review began, the list of experts was sealed using the AES (Advanced Encryption Standard) until the bid opening deadline, when it was automatically decrypted and made public through a smart contract based on blockchain technology. This formed a standardized process for selecting review experts that was highly integrated with technology, from demand initiation, random algorithm selection, smart notification confirmation to encrypted storage and decryption.

[0003] However, in the process of implementing the inventive technical solution in the embodiments of this application, it was found that the above-mentioned technology has at least the following technical problems: In existing technologies, the extraction of review experts generally adopts a literal fuzzy matching method based on relational databases. This method can only perform simple textual matching of information such as expert qualifications, job descriptions, and review content. It cannot achieve accurate matching from the perspectives of professional dimensions, business scenarios, and ability levels. This results in a low accuracy rate in matching experts with review matters and easily leads to problems such as experts being selected with mismatched expertise or insufficient review experience. This directly affects the standardization of the review process and the scientific nature and authority of the review results.

[0004] Meanwhile, because the data on expert qualifications, job definitions and descriptions, and review items come from different organizations, departments and professional fields, there is a lack of unified data standards, attribute definitions and field specifications. There are problems such as inconsistent standards and large differences in descriptions among different data sources, which further aggravates the matching error. Even after fuzzy matching, a lot of manual verification and adjustment still need to be done by staff with rich review work experience, making the expert selection process cumbersome, time-consuming and inefficient.

[0005] Furthermore, expert-related and review-related information is scattered across structured data in various business systems and unstructured documents such as Word, Excel, PDF, and JPG. Existing technologies cannot effectively extract, compare, and correlate detailed information within unstructured documents. Staff members need to spend a lot of time and effort in sorting, confirming, and matching information. Moreover, the results after matching fail to form a reusable and iterative closed-loop data asset, and the data value cannot be effectively accumulated. There is a problem of low reliability in randomly selecting evaluations from the review expert database. Summary of the Invention

[0006] To address the low reliability of random selection evaluation from a review expert database in existing technologies, this invention provides a method for random selection evaluation from a review expert database. The technical solution is as follows: On the one hand, a method for randomly selecting and evaluating review experts is provided. This method includes: collecting various data and documents related to review work; vectorizing the collected data and documents to convert unstructured and semi-structured data into structured vector data that can be used for model analysis; obtaining matching parameters between experts and review items; quantitatively representing the matching accuracy between the extracted experts' professional capabilities and the actual needs of the review projects based on the matching parameters; and determining whether to dynamically adjust the matching accuracy based on the matching accuracy to improve the accuracy of review assignment, reduce the mismatch rate, and improve review efficiency and quality. If yes, then a correlation analysis of experts and review projects will be conducted after dynamic adjustment; otherwise, the correlation analysis of experts and review projects will proceed directly. Data assetization closed-loop parameters will be obtained from the correlation analysis of experts and review projects. Based on these parameters, the efficiency and completeness of transforming fragmented raw information into structured assets that can be iteratively utilized will be quantitatively characterized, resulting in the data assetization closed-loop conversion degree. Based on the data assetization closed-loop conversion degree, it will be determined whether to dynamically adjust the closed-loop conversion degree to achieve optimal data asset conversion efficiency and reduce resource consumption. If yes, then a random expert selection process will be conducted after dynamic adjustment; otherwise, the random expert selection process will proceed directly.

[0007] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: 1. By collecting multi-source data such as review projects, expert qualifications, and historical cases, deep learning is used to vectorize and label unstructured and semi-structured data, enabling a quantitative representation of expert professional capabilities and review requirements, and accurately obtaining the matching accuracy between experts and review items. When the matching accuracy is lower than the benchmark value, a dynamic adjustment mechanism is automatically triggered. This mechanism uses an extraction latency-mapping rate adjustment coefficient table to monitor the extraction latency of individual expert information in real time. When latency is low, the field normalization mapping rate is increased to enhance the standardization of expert information and matching accuracy; when latency exceeds the benchmark, the field normalization mapping rate is reduced to decrease verification time and shorten overall processing latency. Simultaneously, combined with batch document processing latency range judgment, the concurrent processing throughput is dynamically adjusted. When latency is low, throughput is increased to improve efficiency; when latency is high, throughput is reduced to alleviate system load; and when latency is normal, a stable balance is maintained. Therefore, while ensuring the accuracy of expert-review item matching, the overall efficiency of document parsing, information extraction, and batch processing is significantly improved, reducing mismatch rates and resource waste.

[0008] 2. By introducing a data assetization closed-loop conversion rate indicator, and quantitatively coupling parameters such as the accuracy of matching to be evaluated, data iteration update frequency, and tag dynamic update latency, the system objectively measures the completeness and efficiency of transforming fragmented review business data into structured, iteratively usable data assets. When the closed-loop conversion rate does not meet the baseline requirements, dynamic adjustment of the closed-loop conversion rate is automatically initiated. Through an adaptive mapping relationship between incremental information extraction latency and parallel processing rate, the parallel processing rate of data parsing is adjusted in real time: when the latency is low, the parallel rate is increased to improve incremental update efficiency; when the latency exceeds the standard, the parallel rate is decreased to alleviate resource competition and reduce extraction latency, ensuring the real-time performance and stability of incremental data updates. Through the above mechanism, the system achieves optimal data asset conversion efficiency and minimizes resource consumption, enabling review data to continuously iterate and self-improve, forming a reusable and evolving data asset closed loop.

[0009] 3. For scenarios where expert information exists in multiple modal formats such as databases, PDFs, and scanned certificates, a multimodal fusion latency-parallel processing number mapping relationship is constructed. Using fusion latency deviation and closed-loop conversion degree deviation as inputs, the system adaptively outputs the gain or attenuation of the parallel processing number, achieving intelligent adjustment of the parallel processing number for a single task: when the fusion latency is low, the parallel processing number is increased to fully utilize idle resources and accelerate cross-modal reading, parsing, alignment, and merging; when the fusion latency is too high, the parallel processing number is reduced to prevent further latency deterioration, ensuring the integrity and accuracy of multimodal data fusion. After completing the above end-to-end optimization, a correlation analysis model is constructed based on labeled and vectorized data. Correlation scores are calculated from three dimensions: review content requirements, expert qualifications, and professional matching, forming a candidate expert pool. Through encrypted random sampling, conflict of interest avoidance, and compliance verification, the system ultimately generates expert random selection results with high professional matching, a fair and compliant extraction process, and stable and reliable review quality. Overall, the extraction of the review expert pool is intelligent, accurate, efficient, and standardized. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 A flowchart of a method for randomly selecting and evaluating review experts from a database, provided for embodiments of this application; Figure 2 A flowchart illustrating the intelligent adjustment of batch processing throughput for a random selection evaluation method from a review expert database, provided in this application embodiment; Figure 3 A flowchart illustrating the multimodal fusion parallelism adaptive adjustment of a random selection evaluation method for an expert database provided in this application embodiment. Detailed Implementation

[0012] The technical solution provided in this application will now be described with reference to the accompanying drawings.

[0013] To facilitate understanding of the embodiments of this application, the following points will be explained first: First, in this application, "at least one" means one or more, and "more than one" means two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A alone, A and B simultaneously, or B alone, where A and B can be singular or plural. The character " / " generally indicates an "or" relationship between the preceding and following related objects, but it does not exclude the possibility of indicating an "and" relationship; the specific meaning can be understood in context. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can mean: a, b, c; a and b; a and c; b and c; or a and b and c. Here, a, b, and c can be single or multiple.

[0014] Second, the use of prefixes such as "first" and "second" in this application is solely for the purpose of distinguishing and describing different things belonging to the same category, and does not constrain the order, size, or quantity of things. For example, "first message" and "second message" are simply different messages, and there is no chronological, size, or priority relationship between them.

[0015] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0016] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0017] like Figure 1 The diagram shown is a flowchart of a method for randomly selecting and evaluating review experts according to an embodiment of this application. The method includes the following steps: As the first step of this method, various data and documents related to the review work are collected. The collected data and documents are vectorized to convert unstructured and semi-structured data into structured vector data that can be used for model analysis. The matching parameters between experts and review items are obtained. Based on the matching parameters, the accuracy of the matching between the extracted expert professional capabilities and the actual needs of the review project is quantitatively represented, and the accuracy of the matching between experts and review items is obtained. The data includes, but is not limited to, review project information, review content details, basic expert information, expert professional qualification information, expert review experience information, and expert research field information. Documents include, but are not limited to, review standard documents, expert qualification certificates, and historical review case documents. Vectorization processing employs a deep learning-based text embedding algorithm to retain core feature information from the data and documents, generating corresponding data vectors and document vectors. The generated and stored structured vector data undergoes cleaning processing, including removing invalid data, correcting erroneous data, removing duplicate and redundant data, and supplementing missing data to ensure data accuracy, completeness, and consistency. Based on the core needs of the review work, several tags and tag definition standards are set. Tags include review content tags, expert qualification tags, expert level tags, and professional field tags. Review content tags correspond to the specific review points of various review projects; expert qualification tags correspond to the expert's education, professional title, professional qualifications, and other qualification information; expert level tags set grading standards based on the expert's review experience, professional ability, and industry recognition; and professional field tags correspond to the expert's research direction and areas of expertise in review. The cleaned vector data is then associated and bound with the set tags to complete the data tagging process.

[0018] It should be understood that determining whether to dynamically adjust matching accuracy previously included: The parameters for matching experts with review items include unstructured document parsing rate, key information extraction latency, and OCR recognition accuracy. Specifically, the unstructured document parsing rate is the ratio of the number of unstructured documents successfully parsed (based on backend log statistics) to the total number of unstructured documents received and processed. The key information extraction latency is the difference between the timestamp of the information returned by the performance monitoring platform and the timestamp of the interface request start. The OCR recognition accuracy is the ratio of the number of correctly recognized words obtained through the OCR engine's built-in evaluation metrics to the actual total number of words in the sample.

[0019] The document parsing rate correction value is obtained by interactively processing the ratio of the unstructured document parsing rate to the document parsing rate reference value using a document parsing rate compensation factor. The specific constraint expression for obtaining the document parsing rate correction value is as follows: ; In the formula, A represents the document parsing rate correction value; s1 represents the document parsing rate compensation factor obtained from the database by intelligent matching of the expert database; E represents the unstructured document parsing rate; and E0 represents the document parsing rate reference value obtained from the database by intelligent matching of the expert database.

[0020] The extraction delay correction value is obtained by interactively processing the ratio of the extraction delay reference value to the key information extraction delay using an extraction delay compensation factor; the specific constraint expression for obtaining the extraction delay correction value is as follows: ; In the formula, B represents the extraction delay correction value; s2 represents the extraction delay compensation factor obtained from the expert database intelligent matching extraction database; R represents the key information extraction delay; and R0 represents the extraction delay reference value obtained from the expert database intelligent matching extraction database.

[0021] The ratio of OCR recognition accuracy to a reference value is interactively processed by a recognition accuracy compensation factor to obtain a recognition accuracy correction value; the specific constraint expression for obtaining the recognition accuracy correction value is as follows: ; In the formula, C represents the recognition accuracy correction value; s3 represents the recognition accuracy compensation factor obtained from the database by intelligent matching of the expert database; Q represents the OCR recognition accuracy; and Q0 represents the recognition accuracy reference value obtained from the database by intelligent matching of the expert database.

[0022] The document parsing rate correction value, extraction latency correction value, and recognition accuracy correction value are coupled to obtain the expert-review item matching accuracy. The specific constraint expression for obtaining the expert-review item matching accuracy is as follows: ; In the formula, D represents the accuracy of matching experts with review items.

[0023] It's important to explain that a higher OCR (Optical Character Recognition) accuracy means more precise recognition of text, table boundaries, paragraph structures, and key characters (such as amounts, dates, and ID numbers) in documents, with fewer garbled characters, missing characters, or recognition errors. It also results in a higher unstructured document parsing rate. Higher OCR accuracy means more accurate key characters and text fragments output by the OCR. The extraction module can directly match key information according to preset rules (such as regular expressions and entity recognition models) without additional error correction (such as character validation or fuzzy matching), resulting in a smooth extraction process and shorter key information extraction latency. Similarly, a higher unstructured document parsing rate results in complete and clearly structured content (with clearly defined text segments, tables, and key fields). The extraction module doesn't need to handle "parsing failure" exceptions (such as skipping failed documents or retrying parsing), and can directly extract key information from the parsed content, leading to a smoother process and shorter key information extraction latency. Meanwhile, there is a positive correlation between the unstructured document parsing rate and the accuracy of matching experts with review items. A higher unstructured document parsing rate means that it can be successfully parsed, the structure is correctly split (such as distinguishing review item descriptions, technical indicators, qualification requirements, etc.), the content is complete and there are no garbled characters, and the accuracy of matching experts with review items is higher. There is a negative correlation between the key information extraction latency and the accuracy of matching review items. The longer the key information extraction latency, the more likely problems such as "feature extraction timeout, missing features in some documents, and algorithm processing congestion" will occur in high-concurrency scenarios. In order to complete the matching within the specified time, the system will simplify the matching rules (such as matching only based on 1-2 core features and ignoring secondary features) or skip the accurate matching of some timeout documents and directly assign general experts, resulting in a decrease in matching accuracy and a lower accuracy of matching experts with review items. There is a positive correlation between the OCR recognition accuracy and the accuracy of matching experts with review items. A higher OCR recognition accuracy means that the core text, professional terms, and key indicators in the review document can be accurately identified, with no character errors, no missing information, and no misjudgment of professional terms, and the accuracy of matching experts with review items is higher.

[0024] As the second step of this method, it is determined whether to dynamically adjust the matching accuracy based on the accuracy of the matching between experts and review items, in order to improve the accuracy of review assignment, reduce the mismatch rate, and improve the efficiency and quality of review. If so, the correlation analysis between experts and review items is carried out after dynamic adjustment; otherwise, the correlation analysis between experts and review items is carried out directly.

[0025] Furthermore, the specific steps for determining whether to dynamically adjust the matching accuracy are as follows: If the accuracy of matching experts with review items is not less than the baseline value of matching accuracy, no dynamic adjustment of matching accuracy will be performed. If the accuracy of matching experts with review items is less than the baseline value of matching accuracy, adaptive control of field normalization mapping rate latency and intelligent adjustment of batch processing throughput will be performed based on the accuracy deviation. The accuracy deviation represents the negative difference between the accuracy of matching experts with review items and the baseline value of matching accuracy.

[0026] It should be further explained that the specific steps for performing latency-aware adaptive adjustment of the field normalization mapping rate are as follows: An extraction delay-mapping rate adjustment coefficient mapping table is constructed. This table is used to dynamically and adaptively calculate the set of quantitative mapping relationships of the normalized mapping rate adjustment coefficient based on the real-time extraction delay deviation and the matching accuracy deviation. The table takes the delay deviation parameters and the matching accuracy deviation as input dimensions, and the mapping rate gain coefficient and the mapping rate reduction coefficient as output dimensions, so as to achieve a precise, quantifiable and reproducible mapping between the delay state and the adjustment intensity. The system collects the extraction latency of a single expert information entry from the start of parsing to the completion of key field output in real time. It compares the extraction latency of a single expert information entry with a preset latency threshold and dynamically adjusts the field normalization mapping rate based on the comparison result.

[0027] The specific steps for dynamically adjusting the field normalization mapping rate based on the comparison results are as follows: If the extraction delay of a single piece of expert information is less than or equal to the preset delay threshold, the matching accuracy deviation and extraction delay offset are input into the constructed extraction delay-mapping rate adjustment coefficient mapping table. The mapping rate gain coefficient is output, and the preset benchmark field normalized mapping rate and mapping rate gain coefficient are interactively processed to obtain the target field normalized mapping rate. This improves the field normalized mapping rate, thereby enhancing the standardization and matching accuracy of expert information fields. The extraction delay offset represents the negative difference between the extraction delay of a single piece of expert information and the preset delay threshold. This strengthens the format uniformity, content completeness, and field matching accuracy of expert information fields, reducing the mismatch between experts and review items caused by non-standard fields. It improves matching accuracy from the data source. At the same time, since the current extraction delay is within a controllable range, improving the mapping rate will not significantly increase the overall processing time, achieving a positive balance between accuracy and efficiency.

[0028] If the extraction delay of a single expert information entry exceeds a preset delay threshold, the matching accuracy deviation and extraction delay correction are input into the constructed extraction delay-mapping rate adjustment coefficient mapping table. The mapping rate reduction coefficient is output, and the preset baseline field normalization mapping rate and mapping rate reduction coefficient are interactively processed to obtain the target field normalization mapping rate. This reduces the field normalization mapping rate, thereby reducing normalization verification time and shortening the overall information processing delay. The extraction delay correction represents the positive difference between the extraction delay of a single expert information entry and the preset delay threshold. This reduces the multi-level normalization verification steps and computational load for expert information fields, directly shortening the normalization processing time of a single piece of information, thus reducing the overall information processing delay. It avoids problems such as batch expert information processing bottlenecks and delays in the expert extraction process caused by excessive time consumption in a single step. While ensuring basic matching requirements, it prioritizes ensuring the timeliness of the extraction process and maintains the stability of the overall system processing efficiency.

[0029] In this embodiment, the extraction latency of a single piece of expert information from the start of parsing to the completion of key field output is collected in real time and compared with a preset latency threshold. Based on the comparison result and the pre-constructed extraction latency-mapping rate adjustment coefficient mapping table, the mapping rate gain coefficient or mapping rate reduction coefficient is adaptively output by combining the matching accuracy deviation, extraction latency offset, or extraction latency correction. The normalized mapping rate of the benchmark field is interactively processed to obtain the normalized mapping rate of the target field. This can improve the field normalized mapping rate when the extraction latency is within a reasonable range, thereby enhancing the standardization degree and matching accuracy of expert information fields and improving the matching accuracy between experts and review matters. When the extraction latency exceeds the threshold, the field normalized mapping rate is reduced, thereby reducing the time spent on normalization verification, shortening the overall information processing latency, and avoiding a decrease in system processing efficiency and process delays. Ultimately, in the process of expert information extraction, a dynamic balance between matching accuracy and processing latency is achieved, and the synergistic optimization of system resource utilization efficiency and data normalization quality is realized, providing stable and reliable data support for the accuracy and timeliness of subsequent expert extraction.

[0030] It should be understood that, such as Figure 2 The diagram shows a flowchart of the intelligent adjustment process for batch processing throughput of a random selection evaluation method for a review expert database provided in this application embodiment. The specific process is as follows: the obtained batch document processing latency is used as the judgment basis and compared with a preset latency range: if the latency is less than the preset lower latency limit, the concurrent processing throughput is set to the benchmark upper limit value to improve the overall processing efficiency; if the latency is greater than the preset lower latency limit, the concurrent processing throughput is set to the benchmark lower limit value to reduce system load and avoid continuous latency deterioration; if the latency is within the preset range, the current throughput is maintained to maintain a stable balance between processing efficiency and load, and finally the dynamic adjustment of batch document concurrent processing throughput is completed.

[0031] It should be further explained that the specific process for intelligent adjustment of batch processing throughput is as follows: The system obtains the batch document processing latency from task issuance to completion in real time, compares the batch document processing latency with the preset latency interval, and dynamically adjusts the batch document concurrent processing throughput based on the comparison result. The preset latency interval represents the closed interval formed by the preset latency lower limit and the preset latency upper limit. If the batch document processing latency is less than the preset latency lower limit, the current batch document concurrent processing throughput will be set to the baseline upper limit of batch document concurrent processing throughput to improve the overall batch document processing efficiency. If the batch document processing latency exceeds the preset latency lower limit, the current batch document concurrent processing throughput will be set to the baseline lower limit value of batch document concurrent processing throughput to reduce the batch document concurrent processing throughput, thereby alleviating the load and preventing the processing latency from continuing to deteriorate. If the batch document processing latency is within the preset latency range, the current batch document concurrent processing throughput will be maintained to ensure a stable balance between processing efficiency and system load.

[0032] In this embodiment, the batch document processing latency from task issuance to completion is obtained in real time and compared with a preset latency range consisting of a preset lower latency limit and a preset upper latency limit. The concurrent processing throughput of batch documents is dynamically adjusted according to different latency range states. When the batch document processing latency is less than the preset lower latency limit, the concurrent processing throughput is increased to the baseline upper limit, fully utilizing idle system resources and significantly improving the overall processing efficiency of batch documents. When the batch document processing latency is greater than the preset lower latency limit, the concurrent processing throughput is reduced to the baseline lower limit, effectively reducing the system's computational load and preventing the processing latency from continuously deteriorating due to excessive concurrency, thus ensuring process stability. When the batch document processing latency is within the preset latency range, the current concurrent processing throughput is maintained, ensuring a stable balance between processing efficiency and system load. Ultimately, adaptive collaborative optimization of efficiency, latency, and system load in batch document processing is achieved, providing efficient, stable, and reliable operational guarantees for the batch parsing, extraction, and standardization of review expert information.

[0033] As the third step of this method, the data assetization closed-loop parameters in the correlation analysis between experts and review projects are obtained. Based on the data assetization closed-loop parameters, the efficiency and completeness of the transformation of expert review business data from fragmented raw information into structured assets that can be used iteratively are quantitatively characterized, and the data assetization closed-loop conversion degree is obtained.

[0034] It should be understood that the closed-loop parameters for data assetization include the accuracy of matching to be evaluated, the frequency of data iteration updates, and the latency of dynamic tag updates. Specifically, the accuracy of matching to be evaluated refers to the accuracy of the newly acquired expert-review item matching if dynamic adjustments to the matching accuracy have been made; otherwise, the current accuracy of the expert-review item matching is recorded as the accuracy of matching to be evaluated. The frequency of data iteration updates refers to the number of times an automatic update task is executed within a unit of time, recorded by the background task scheduling log, including the start time, execution time, and completion time. The latency of dynamic tag updates refers to the difference between the time when tag extraction, verification, fusion, and storage are completed and become effective, and the time when the information in the business / document / database is actually modified.

[0035] The matching accuracy correction value is obtained by interactively processing the ratio of the matching accuracy to be evaluated to the matching accuracy reference value through a matching accuracy compensation factor; the specific constraint expression for obtaining the matching accuracy correction value is as follows: ; In the formula, D represents the matching accuracy correction value; k1 represents the matching accuracy compensation factor obtained from the expert database intelligent matching extraction database; T represents the matching accuracy to be evaluated; and T0 represents the matching accuracy reference value obtained from the expert database intelligent matching extraction database.

[0036] The update frequency correction value is obtained by interactively processing the ratio of the data iteration update frequency to the update frequency reference value through an update frequency compensation factor; the specific constraint expression for obtaining the update frequency correction value is as follows: ; In the formula, G represents the update frequency correction value; k2 represents the update frequency compensation factor obtained from the database extracted by intelligent matching from the expert database; P represents the data iteration update frequency; and P0 represents the update frequency reference value obtained from the database extracted by intelligent matching from the expert database.

[0037] The update delay correction value is obtained by interactively processing the ratio of the update delay reference value to the tag dynamic update delay using an update delay compensation factor; the specific constraint expression for obtaining the update delay correction value is as follows: ; In the formula, F represents the update delay correction value; k3 represents the update delay compensation factor obtained from the database by intelligent matching in the expert database; Y represents the dynamic update delay of the tag; and Y0 represents the update delay reference value obtained from the database by intelligent matching in the expert database.

[0038] By coupling the matching accuracy correction value, update frequency correction value, and update latency correction value, the data assetization closed-loop conversion degree is obtained. The specific constraint expression for obtaining the data assetization closed-loop conversion degree is as follows: ; In the formula, H represents the degree of data assetization closed-loop conversion.

[0039] It should be understood that a higher data iteration and update frequency means that information such as experts' titles, institutions, research directions, qualifications, and review experience can be corrected, supplemented, and improved in a timely manner, resulting in higher accuracy of the matching to be evaluated. The longer the dynamic update latency of tags, the more likely the expert information has changed but the tags have not been updated, and the system will still use expired tags for matching, leading to mismatch, mismatch, and failure to meet the requirements of recommended experts and review matters, resulting in lower accuracy of the matching to be evaluated. A higher data iteration and update frequency indicates that there is a high-frequency, real-time, and automated update mechanism, and that the entire process of change perception, incremental extraction, multimodal fusion, and tag writing is efficient, with shorter dynamic update latency of tags. Meanwhile, there is a positive correlation between the accuracy of the assessed matching and the conversion rate of the data assetization closed loop. The higher the accuracy of the assessed matching, the higher the quality of each stage of the data assetization closed loop (parsing, extraction, standardization, fusion, and updating). Complete and accurate expert profiles, a tag system that fits the business, conflict-free multimodal data fusion, and timely data updates are necessary to support accurate matching and a higher conversion rate of the data assetization closed loop. There is also a positive correlation between the data iteration and update frequency and the conversion rate of the data assetization closed loop. The higher the data iteration and update frequency, the more complete the system's automated update mechanism, enabling it to quickly perceive expert information. Changes in review information (such as adding expert qualification certificates or updating review areas) automatically complete incremental extraction, multimodal fusion, tag updates, and data entry iterations, resulting in a higher degree of data assetization closed-loop conversion. However, there is a negative correlation between the dynamic update latency of tags and the degree of data assetization closed-loop conversion. A longer dynamic update latency means that after changes in expert information, tags cannot be updated in a timely manner, the expert profile becomes disconnected from the actual situation, and the data asset is in a "lagging state." Even with a high data iteration update frequency, the latency issue leads to "untimely updates," preventing efficient reuse and resulting in a lower degree of data assetization closed-loop conversion.

[0040] As the fourth step of this method, the decision to dynamically adjust the closed-loop conversion rate based on the data assetization closed-loop conversion rate aims to optimize data asset conversion efficiency and reduce resource consumption. If so, a random expert selection process is initiated after dynamic adjustment; otherwise, the random expert selection process proceeds directly. This process is executed through a correlation model construction module. Based on the tagged data, a correlation analysis model between experts and review projects is constructed. This model employs machine learning algorithms and focuses on the correlation analysis of information across three core dimensions: review content requirements, expert qualifications, and expert professional matching. First, starting from the review content requirements dimension, the specific requirements for each review item are broken down, clarifying the necessary qualification conditions and qualification level requirements for each item, forming a review content-qualification level correspondence table. Second, based on the pre-defined review content-qualification level correspondence table, the correlation model analyzes all expert data in the expert database, extracting the matching between expert qualification tags, level tags, professional field tags, and review content tags. The system calculates the relevance score between experts and each review item, with the score ranging from 0 to 100. A higher score indicates a higher degree of matching between the expert and the review content. Matching levels are defined: a relevance score ≥ 80 indicates a high match, 60 ≤ relevance score < 80 indicates a medium match, and a relevance score < 60 indicates a low match, thus completing the initial relevance matching between experts and review content. Based on preset expert selection rules and the obtained relevance scores, experts are selected according to the expert selection ratio required by review management. Experts whose relevance scores meet a preset threshold (i.e., moderate matching or above) are selected from the pool to form a candidate expert pool. An encrypted random sampling algorithm is used to randomly select experts from the candidate expert pool according to a preset ratio. During the selection process, it is ensured that each candidate expert has an equal probability of being selected, while avoiding conflicts of interest and eliminating experts with conflicts of interest with the review projects, thus generating preliminary selection results. The preliminary selection results are then verified for compliance to confirm that the number of experts, their qualification levels, and their professional distribution all meet the review management requirements, forming the final selection results of review experts.

[0041] Furthermore, the specific process for determining whether to dynamically adjust the closed-loop conversion rate is as follows: If the closed-loop conversion degree of data assetization is not less than the benchmark value of closed-loop conversion degree, no dynamic adjustment of closed-loop conversion degree will be performed. If the closed-loop conversion degree of data assetization is less than the benchmark value of closed-loop conversion degree, the incremental parsing parallel rate latency adaptive adjustment and multi-modal fusion parallelism adaptive adjustment will be performed based on the closed-loop conversion degree deviation. The closed-loop conversion degree deviation represents the negative difference between the closed-loop conversion degree of data assetization and the benchmark value of closed-loop conversion degree.

[0042] It should be further explained that the specific process for adaptive adjustment of incremental parsing parallel rate is as follows: A mapping relationship between extraction latency and parallel processing rate is constructed. This mapping relationship is used to quantify the bidirectional mapping between the incremental information extraction latency deviation, the data assetization closed-loop transformation degree deviation, and the data parsing parallel processing rate adjustment amount. This enables adaptive matching between incremental information extraction latency and data parsing parallel processing rate, taking into account incremental update efficiency, resource utilization, and data assetization closed-loop transformation quality. The incremental information extraction latency corresponding to the incremental information extraction of changed data is obtained in real time. The incremental information extraction latency is compared with the preset extraction latency threshold, and the data parsing parallel processing rate is dynamically adjusted according to the comparison result.

[0043] The specific process for dynamically adjusting the parallel processing rate of data parsing based on the comparison results is as follows: If the incremental information extraction delay is less than or equal to the preset extraction delay threshold, the closed-loop conversion degree deviation and the extraction delay correction are input into the established extraction delay-parallel processing rate mapping relationship, and the parallel processing rate adjustment is output. The preset baseline data parsing parallel processing rate and the parallel processing rate adjustment are superimposed and rounded down to obtain the target data parsing parallel processing rate, thereby improving the data parsing parallel processing rate and improving the efficiency of incremental data parsing and information extraction. The extraction delay correction represents the negative difference between the incremental information extraction delay and the preset extraction delay threshold. If the incremental information extraction latency is greater than the preset extraction latency threshold, the closed-loop conversion degree deviation and the extraction latency correction amount are input into the established extraction latency-parallel processing rate mapping relationship. The parallel processing rate reduction amount is output. The preset baseline data parsing parallel processing rate and the parallel processing rate reduction amount are then processed by difference and rounded up to obtain the target data parsing parallel processing rate. This reduces the data parsing parallel processing rate to reduce resource contention caused by parallel parsing, reduce the incremental information extraction latency, and ensure the real-time performance of incremental updates. The extraction latency correction amount represents the positive difference between the incremental information extraction latency and the preset extraction latency threshold.

[0044] In this embodiment, by constructing a mapping relationship between the extraction latency and parallel processing rate of quantitatively correlated incremental information extraction latency, the data assetization closed-loop conversion degree deviation, and the data parsing parallel processing rate adjustment, adaptive matching between incremental information extraction latency and data parsing parallel processing rate is achieved. Furthermore, by combining the incremental information extraction latency of real-time acquired change data with a preset extraction latency threshold, when the incremental information extraction latency is less than or equal to the threshold, the parallel processing rate is increased by using the closed-loop conversion degree deviation and the extraction latency correction amount through the mapping relationship. This is then applied to the baseline data parsing parallel processing rate and rounded down to improve the target data parsing parallel processing rate, thereby enhancing the incremental data parsing and parallel processing rate. Information extraction efficiency: When the incremental information extraction latency exceeds a threshold, the parallel processing rate is reduced by mapping the closed-loop transformation degree deviation and the extraction latency correction. The difference in the parallel processing rate of the benchmark data parsing is rounded up to reduce the parallel processing rate of the target data parsing. This reduces resource contention caused by parallel parsing, lowers the incremental information extraction latency, and ensures the real-time nature of incremental updates. Ultimately, throughout the entire process of incremental information extraction and data parsing, incremental update efficiency, system resource utilization, and the quality of closed-loop data assetization are all considered. This achieves synergistic optimization of latency control, parallel processing, and data value transformation, providing stable and efficient technical support for the dynamic updating and iterative governance of expert information.

[0045] It should be understood that, such as Figure 3 The diagram shows a flowchart of the adaptive adjustment of multimodal fusion parallelism in a random sampling evaluation method for an expert database provided in this application. The specific process is as follows: the fusion delay of the collected multimodal data is used as the judgment basis and compared with the fusion delay reference interval. If the fusion delay is less than the lower limit of the fusion delay reference, the closed-loop conversion degree deviation and the fusion delay offset are input into the mapping relationship, and the parallel processing number gain is output. If the fusion delay is greater than the upper limit of the fusion delay reference, the closed-loop conversion degree deviation and the fusion delay offset are input into the mapping relationship, and the parallel processing number attenuation is output. If the fusion delay is within the fusion delay reference interval, the current single-task parallel processing number is maintained, and finally the dynamic adjustment of the single-task parallel processing number in the multimodal data fusion scenario is completed.

[0046] It should be further explained that the specific steps for adaptive adjustment of multimodal fusion parallelism are as follows: A pre-constructed mapping relationship between fusion latency and parallel processing number is used to adaptively and quantitatively output a set of quantitative decision mapping rules for adjusting the number of parallel processing numbers in a single task, based on the deviation of multimodal data fusion latency and the deviation of data assetization closed-loop conversion degree. The set takes latency deviation and closed-loop quality deviation as inputs and the gain and attenuation of parallel processing number as outputs. Real-time acquisition of multimodal data fusion latency refers to the total time required to complete cross-modal data reading, parsing, alignment, association, and merging when expert information exists simultaneously in multiple modalities (including structured databases, PDF documents, images (certificate scans, etc.). The multimodal data fusion latency is compared with the fusion latency reference interval, and the number of parallel processing tasks per task is dynamically adjusted based on the comparison results. The number of parallel processing tasks per task refers to the maximum number of multimodal data parsing, cross-modal alignment, and fusion processing tasks that can be executed simultaneously at the same time. The fusion latency reference interval represents the closed interval formed by the lower limit and upper limit of the fusion latency reference.

[0047] If the multimodal data fusion delay is less than the reference lower limit of fusion delay, the closed-loop conversion degree deviation and fusion delay offset are input into the established fusion delay-parallel processing number mapping relationship, and the parallel processing number gain is output. The preset benchmark single-task parallel processing number and the parallel processing number gain are superimposed and rounded down to obtain the target single-task parallel processing number, thereby increasing the single-task parallel processing number to make full use of the system's idle resources, improve the efficiency of multimodal data fusion, and shorten the overall fusion time. The fusion delay offset represents the negative difference between the multimodal data fusion delay and the reference lower limit of fusion delay. If the multimodal data fusion delay is within the fusion delay reference interval, then the current number of parallel processing tasks per task will be maintained. If the multimodal data fusion delay is greater than the upper limit of the fusion delay reference, the closed-loop conversion degree deviation and the fusion delay reference value are input into the established fusion delay-parallel processing number mapping relationship. The parallel processing number attenuation is output. The preset benchmark single-task parallel processing number and the parallel processing number attenuation are processed by difference and rounded up to obtain the target single-task parallel processing number. This reduces the multimodal data fusion delay, avoids the delay from continuing to deteriorate, and ensures the integrity and accuracy of multimodal data fusion. The fusion delay reference value represents the positive difference between the multimodal data fusion delay and the lower limit of the fusion delay reference.

[0048] In this embodiment, by pre-constructing a fusion delay-parallel processing number mapping relationship with multimodal data fusion delay deviation and data assetization closed-loop conversion degree deviation as inputs and parallel processing number gain and attenuation as outputs, a decision rule for adjusting the single-task parallel processing number can be adaptively and quantitatively output, achieving precise control of the multimodal fusion process. After real-time collection of the total time spent on cross-modal reading, parsing, alignment, association, and merging of expert information in multimodal data such as structured databases, PDF documents, and certificate scans, it is compared with a fusion delay reference interval consisting of a lower and upper limit of the fusion delay reference. When the fusion delay is less than the lower limit of the reference, the parallel processing number gain is output through the mapping relationship, and the baseline single-task parallel processing number is superimposed and rounded down to improve the target. The system optimizes the number of parallel processing tasks per task to fully utilize idle system resources, improve the efficiency of multimodal data fusion, and shorten the overall fusion time. When the fusion latency is within the reference range, the current number of parallel processing tasks is maintained to ensure stable system operation. When the fusion latency exceeds the reference upper limit, the system outputs the attenuation of the number of parallel processing tasks through a mapping relationship, rounds up the difference between the baseline number of parallel processing tasks per task and the target number of parallel processing tasks per task, thereby effectively reducing the latency of multimodal data fusion and preventing the latency from deteriorating further. At the same time, it ensures the integrity and accuracy of the cross-modal data fusion process, and ultimately achieves a dynamic balance between latency, resource utilization, processing efficiency, and data quality of multimodal expert information. This provides a stable, efficient, and reliable technical guarantee for the data aggregation, standardized governance, and assetization closed loop of the review expert database.

[0049] The various features and processes described above can be used independently of each other or can be combined in various ways. All possible combinations and sub-combinations are intended to fall within the scope of this disclosure. Furthermore, certain method or process blocks may be omitted in some embodiments. The methods and processes described herein are not limited to any particular order, and the blocks or states associated with them may be performed in other suitable orders. For example, the described blocks or states may be performed in an order different from the order specifically disclosed, or multiple blocks or states may be combined in a single block or state. Example blocks or states may be performed serially, in parallel, or in some other manner. Blocks or states may be added to or removed from the disclosed example embodiments. The exemplary systems and components described herein may be configured differently from those described. For example, elements may be added to, removed from, or rearranged compared to the disclosed example embodiments.

[0050] The various operations of the example methods described herein can be performed at least in part by an algorithm. This algorithm can be contained in program code or instructions stored in memory (e.g., the aforementioned non-transitory computer-readable storage medium). Such an algorithm may include a machine learning algorithm. In some embodiments, the machine learning algorithm may not be explicitly programmed into the computer to perform the function, but can learn from training data to create a predictive model that performs the function.

[0051] The various operations of the example methods described herein can be performed, at least in part, by one or more processors that are temporarily configured (e.g., by software) or permanently configured to perform the relevant operations. Whether temporarily or permanently configured, such processors can constitute the engine of a processor implementation that operates to perform one or more of the operations or functions described herein.

[0052] Similarly, the methods described herein can be implemented at least in part by a processor, where one or more specific processors are examples of hardware. For example, at least some operations of a method can be performed by one or more processors or an engine implemented by a processor. Furthermore, one or more processors can also be operated to support the performance of related operations in a “cloud computing” environment or as “Software as a Service” (SaaS). For example, at least some operations can be performed by a set of computers (as an example of a machine including processors), where these operations are accessible via a network (e.g., the Internet) and via one or more suitable interfaces (e.g., application programming interfaces (APIs)).

[0053] The performance of certain operations can be distributed across processors, residing not only within a single machine but also deployed across multiple machines. In some example embodiments, the processor or processor-implemented engine may reside in a single geographic location (e.g., within a home environment, office environment, or server cluster). In other example embodiments, the processor or processor-implemented engine may be distributed across multiple geographic locations.

[0054] In this specification, multiple instances may implement components, operations, or structures described as single instances. Although individual operations of one or more methods are shown and described as separate operations, one or more of the separate operations may be performed simultaneously and do not need to be performed in the order shown. Structures and functions presented as separate components in the example configuration may be implemented as composite structures or components. Similarly, structures and functions presented as single components may be implemented as separate components. These and other variations, modifications, additions, and improvements fall within the scope of this document.

[0055] While an overview of the subject matter has been described with reference to specific example embodiments, various modifications and changes can be made to these embodiments without departing from the broader scope of embodiments of this disclosure. Such embodiments of the subject matter are referred to herein, individually or collectively, by the term "invention," and are used for convenience only and are not intended to limit the scope of this application to any single disclosure or concept, should more than one disclosure or concept be disclosed in fact.

[0056] The embodiments described herein have been described in sufficient detail to enable those skilled in the art to practice the disclosed teachings. Other embodiments may be used and derived therefrom, such that structural and logical substitutions and changes may be made without departing from the scope of this disclosure. Therefore, the detailed description should not be construed as limiting, and the scope of the various embodiments is defined only by the appended claims and the full scope of their equivalents.

Claims

1. A method for randomly selecting evaluation experts from a pool of reviewers, characterized in that, Includes the following steps: Collect various data and documents related to the review work, and perform vectorization processing on the collected data and documents to convert unstructured and semi-structured data into structured vector data that can be used for model analysis. Obtain matching parameters between experts and review items, and quantitatively characterize the matching accuracy between the extracted expert professional capabilities and the actual needs of the review project based on the matching parameters between experts and review items to obtain the matching accuracy between experts and review items. Whether to dynamically adjust the matching accuracy based on the accuracy of the matching between experts and review items, in order to improve the accuracy of review assignment, reduce the mismatch rate, and improve the efficiency and quality of review. If yes, then conduct the correlation analysis between experts and review items after dynamic adjustment; otherwise, directly conduct the correlation analysis between experts and review items. The data assetization closed-loop parameters are obtained in the correlation analysis between experts and review projects. Based on the data assetization closed-loop parameters, the efficiency and completeness of the transformation of expert review business data from fragmented raw information into structured assets that can be used iteratively are quantitatively characterized, and the data assetization closed-loop conversion degree is obtained. The decision to dynamically adjust the closed-loop conversion rate based on the data assetization closed-loop conversion rate is to achieve optimal data asset conversion efficiency and reduce resource consumption. If so, the expert random selection process will be carried out after dynamic adjustment; otherwise, the expert random selection process will be carried out directly. If the accuracy of matching experts with review items is less than the matching accuracy benchmark, then the field normalization mapping rate latency perception adaptive adjustment and batch processing throughput intelligent adjustment are performed based on the matching accuracy deviation. The matching accuracy deviation is used to characterize the degree of negative deviation between the accuracy of matching experts with review items and the matching accuracy benchmark. The specific steps for performing field normalization mapping rate latency-aware adaptive adjustment are as follows: An extraction delay-mapping rate adjustment coefficient mapping table is constructed. The extraction delay-mapping rate adjustment coefficient mapping table is used to dynamically and adaptively calculate the set of quantitative mapping relationships of the field normalized mapping rate adjustment coefficient based on the real-time extraction delay deviation and the matching accuracy deviation. The table takes the delay deviation class parameters and the matching accuracy deviation as input dimensions, and the mapping rate gain coefficient and the mapping rate reduction coefficient as output dimensions, so as to realize the accurate, quantifiable and reproducible mapping between the delay state and the adjustment intensity. The extraction delay of a single piece of expert information from the start of parsing to the completion of key field output is collected in real time. The extraction delay of a single piece of expert information is compared with a preset delay threshold, and the field normalization mapping rate is dynamically adjusted according to the comparison result. The closed-loop parameters for data assetization include the accuracy of matching to be evaluated, the frequency of data iteration and update, and the latency of dynamic tag update. The matching accuracy correction value is obtained by interactively processing the ratio of the matching accuracy to be evaluated to the matching accuracy reference value through the matching accuracy compensation factor. The update frequency correction value is obtained by interactively processing the ratio of the data iteration update frequency to the update frequency reference value through the update frequency compensation factor. The update delay correction value is obtained by interactively processing the ratio of the update delay reference value to the tag dynamic update delay through the update delay compensation factor. By coupling the matching accuracy correction value, the update frequency correction value, and the update latency correction value, the data assetization closed-loop conversion degree is obtained. The specific process for determining whether to perform dynamic adjustment of the closed-loop conversion rate is as follows: If the data assetization closed-loop conversion degree is not less than the closed-loop conversion degree benchmark value, no dynamic adjustment of the closed-loop conversion degree will be performed. If the data assetization closed-loop conversion degree is less than the closed-loop conversion degree benchmark value, the incremental parsing parallel rate latency adaptive adjustment and the multimodal fusion parallelism adaptive adjustment will be performed based on the closed-loop conversion degree deviation. The closed-loop conversion degree deviation represents the degree of negative deviation between the data assetization closed-loop conversion degree and the closed-loop conversion degree benchmark value.

2. The method for randomly selecting evaluation experts from a database as described in claim 1, characterized in that, The determination of whether to dynamically adjust the matching accuracy also includes: The matching parameters between experts and review items include unstructured document parsing rate, key information extraction latency, and OCR recognition accuracy. The document parsing rate correction value is obtained by interactively processing the ratio of the unstructured document parsing rate to the document parsing rate reference value through a document parsing rate compensation factor. The extraction delay correction value is obtained by interactively processing the ratio of the extraction delay reference value to the key information extraction delay by extracting a delay compensation factor. By interactively processing the ratio of OCR recognition accuracy to recognition accuracy reference value through recognition accuracy compensation factor, a recognition accuracy correction value is obtained. By coupling the document parsing rate correction value, the extraction latency correction value, and the recognition accuracy correction value, the accuracy of matching experts with review items is obtained. If the accuracy of matching experts with review items is not less than the baseline value of matching accuracy, then no dynamic adjustment of matching accuracy will be made.

3. The method for randomly selecting evaluation experts from a pool as described in claim 1, characterized in that, The specific steps for dynamically adjusting the field normalization mapping rate based on the comparison results are as follows: If the extraction delay of a single piece of expert information is less than or equal to a preset delay threshold, the matching accuracy deviation and extraction delay offset are input into the constructed extraction delay-mapping rate adjustment coefficient mapping table, and the mapping rate gain coefficient is output. The preset benchmark field normalized mapping rate and the mapping rate gain coefficient are interactively processed to obtain the target field normalized mapping rate, thereby improving the field normalized mapping rate and enhancing the standardization degree and matching accuracy of the expert information field. The extraction delay offset represents the degree of negative deviation between the extraction delay of a single piece of expert information and the preset delay threshold. If the extraction delay of a single piece of expert information exceeds a preset delay threshold, the matching accuracy deviation and extraction delay correction are input into the constructed extraction delay-mapping rate adjustment coefficient mapping table. The mapping rate reduction coefficient is output, and the preset benchmark field normalization mapping rate and mapping rate reduction coefficient are interactively processed to obtain the target field normalization mapping rate. The field normalization mapping rate is reduced to reduce the normalization verification time and shorten the overall information processing delay. The extraction delay correction represents the degree of positive deviation between the extraction delay of a single piece of expert information and the preset delay threshold.

4. The method for randomly selecting evaluation experts from a pool as described in claim 1, characterized in that, The specific process for intelligent adjustment of batch processing throughput is as follows: The batch document processing latency from task issuance to completion is obtained in real time. The batch document processing latency is compared with a preset latency interval. The batch document concurrent processing throughput is dynamically adjusted based on the comparison result. The preset latency interval represents a closed interval formed by a preset lower latency limit and a preset upper latency limit. If the batch document processing latency is less than the preset latency lower limit, the current batch document concurrent processing throughput is set to the baseline upper limit of the batch document concurrent processing throughput to increase the batch document concurrent processing throughput and improve the overall batch document processing efficiency. If the batch document processing latency exceeds the preset latency lower limit, the current batch document concurrent processing throughput is set to the baseline lower limit value of batch document concurrent processing throughput to reduce the batch document concurrent processing throughput, thereby alleviating the load and preventing the processing latency from continuing to deteriorate. If the batch document processing latency is within the preset latency range, the current batch document concurrent processing throughput will be maintained to ensure a stable balance between processing efficiency and system load.

5. The method for randomly selecting evaluation experts from a pool as described in claim 1, characterized in that, The specific process for adaptive adjustment of the parallel rate and latency of incremental parsing is as follows: An extraction latency-parallel processing rate mapping relationship is constructed. This mapping relationship is used to quantify the bidirectional mapping of the incremental information extraction latency deviation, the data assetization closed-loop transformation degree deviation, and the data parsing parallel processing rate adjustment amount. This enables adaptive matching between incremental information extraction latency and data parsing parallel processing rate, taking into account incremental update efficiency, resource utilization, and data assetization closed-loop transformation quality. The incremental information extraction delay corresponding to the incremental information extraction of changed data is obtained in real time. The incremental information extraction delay is compared with a preset extraction delay threshold, and the data parsing parallel processing rate is dynamically adjusted according to the comparison result.

6. The method for randomly selecting evaluation experts from a pool as described in claim 5, characterized in that, The specific process for dynamically adjusting the parallel processing rate of data parsing based on the comparison results is as follows: If the incremental information extraction delay is less than or equal to the preset extraction delay threshold, the closed-loop conversion degree deviation and the extraction delay correction are input into the established extraction delay-parallel processing rate mapping relationship, and the parallel processing rate adjustment is output. The preset baseline data parsing parallel processing rate and the parallel processing rate adjustment are superimposed and rounded down to obtain the target data parsing parallel processing rate, thereby improving the data parsing parallel processing rate and improving the efficiency of incremental data parsing and information extraction. The extraction delay correction represents the degree of negative deviation between the incremental information extraction delay and the preset extraction delay threshold. If the incremental information extraction delay is greater than the preset extraction delay threshold, the closed-loop conversion degree deviation and the extraction delay correction amount are input into the established extraction delay-parallel processing rate mapping relationship. The parallel processing rate reduction amount is output. The preset baseline data parsing parallel processing rate and the parallel processing rate reduction amount are processed by difference and rounded up to obtain the target data parsing parallel processing rate. The data parsing parallel processing rate is reduced to reduce resource contention caused by parallel parsing, reduce incremental information extraction delay, and ensure the real-time performance of incremental updates. The extraction delay correction amount represents the degree of positive deviation between the incremental information extraction delay and the preset extraction delay threshold.

7. The method for randomly selecting evaluation experts from a pool as described in claim 1, characterized in that, The specific steps for adaptive adjustment of the parallelism of multimodal fusion are as follows: A pre-constructed fusion latency-parallel processing number mapping relationship is used to adaptively and quantitatively output a set of quantitative decision mapping rules for adjusting the single-task parallel processing number based on the deviation degree of multimodal data fusion latency and the deviation degree of data assetization closed-loop transformation. The set takes latency deviation and closed-loop quality deviation as inputs and parallel processing number gain and attenuation as outputs. The system collects multimodal data fusion latency in real time. Multimodal data fusion latency refers to the total time required to complete cross-modal data reading, parsing, alignment, association, and merging when expert information exists simultaneously in multiple modalities. The system compares the multimodal data fusion latency with a fusion latency reference interval and dynamically adjusts the number of parallel processing tasks per task based on the comparison results. The number of parallel processing tasks per task refers to the maximum number of multimodal data parsing, cross-modal alignment, and fusion processing tasks that can be executed simultaneously at the same time. The fusion latency reference interval represents the closed interval formed by the lower limit and upper limit of the fusion latency reference.

8. The method for randomly selecting evaluation experts from a pool as described in claim 7, characterized in that, The adaptive adjustment of the parallelism of the multimodal fusion also includes: If the multimodal data fusion delay is less than the reference lower limit of fusion delay, the closed-loop conversion degree deviation and fusion delay offset are input into the constructed fusion delay-parallel processing number mapping relationship, and the parallel processing number gain is output. The preset benchmark single-task parallel processing number and the parallel processing number gain are superimposed and rounded down to obtain the target single-task parallel processing number, thereby increasing the single-task parallel processing number to make full use of the system's idle resources, improve the efficiency of multimodal data fusion, and shorten the overall fusion time. The fusion delay offset represents the degree of negative deviation between the multimodal data fusion delay and the reference lower limit of fusion delay. If the multimodal data fusion delay is within the fusion delay reference interval, then the current number of parallel processing tasks per task will be maintained. If the multimodal data fusion delay is greater than the upper limit of the fusion delay reference, the closed-loop conversion degree deviation and the fusion delay reference value are input into the established fusion delay-parallel processing number mapping relationship, and the parallel processing number attenuation is output. The preset benchmark single-task parallel processing number and the parallel processing number attenuation value are processed by difference and rounded up to obtain the target single-task parallel processing number. This reduces the multimodal data fusion delay, avoids the delay from continuing to deteriorate, and ensures the integrity and accuracy of multimodal data fusion. The fusion delay reference value represents the degree of positive deviation between the multimodal data fusion delay and the lower limit of the fusion delay reference.

Images