Intelligent algorithm-based corpus automatic collection and quality control method and system

Abstract

The present application relates to the technical field of corpus collection control, and particularly relates to a corpus automatic collection and quality control method and system based on an intelligent algorithm, which constructs state representation by acquiring corpus attribute information and generates an adversarial collection strategy, extracts blind area corpus after execution, selects boundary corpus and collects, constructs feature mapping to determine associated corpus, transmits quality mark value to associated corpus to obtain a corrected quality value, transmits utility measurement value to boundary corpus to obtain a predicted utility value, and finally fuses quality and utility values to execute corpus quality control, thereby improving the comprehensiveness of corpus collection and the accuracy of corpus quality control.

Description

Technical Field

[0001] This invention relates to the field of corpus acquisition and control technology, and in particular to a method and system for automatic corpus acquisition and quality control based on intelligent algorithms. Background Technology

[0002] Existing corpus collection methods mostly employ random sampling or keyword-matching-based targeted crawling strategies. For large-scale corpus construction, the sources of the corpus are typically screened by manually pre-defined rules, combined with simple lexical statistical features for initial filtering. Post-collection quality control often relies on manual annotation, obtaining corpus quality labels through multiple rounds of cross-checking, or using statistical methods such as label consistency checks to remove outliers. Some systems introduce active learning frameworks, selecting samples requiring focused annotation through uncertainty sampling or diversity sampling to reduce the cost of manual annotation. However, these methods all face significant limitations when dealing with complex semantic scenarios.

[0003] Random sampling easily leads to uneven coverage of the collected data, with a large amount of redundant and low-value corpus being repeatedly captured, while key boundary corpus is ignored due to its low frequency of occurrence. While rule-based targeting strategies can improve the purity of corpus data in specific domains, rule design heavily relies on expert experience, making it difficult to adapt to the semantic variability of the corpus. Furthermore, conflicts and redundancies between rules further increase maintenance costs. Manual quality control is inefficient, and annotation consistency fluctuates significantly due to subjective factors, especially in scenarios with dense domain terminology or multilingual environments. Differences in the understanding of quality standards among different annotators can easily lead to accumulated label noise. Even when active learning methods are used to alleviate annotation pressure, their core sampling strategies are still based on single uncertainty or representativeness indicators, lacking deep modeling of the semantic spatial distribution of the corpus. This results in sample selection favoring known high-density areas, while potentially high-quality corpus data in semantic blind spots is often systematically missed, ultimately affecting the completeness and generalization ability of the corpus. These shortcomings are particularly prominent in practical applications such as cross-domain knowledge fusion and low-frequency semantic parsing, urgently requiring a solution that balances collection efficiency with robust quality control. Summary of the Invention

[0004] This invention provides a method and system for automatic corpus acquisition and quality control based on intelligent algorithms, which can solve the problems in the prior art.

[0005] A first aspect of this invention provides a method for automatic corpus acquisition and quality control based on intelligent algorithms, comprising:

[0006] Obtain the attribute information of the corpus to be collected, construct the corpus state representation, generate the current collection strategy based on the corpus state representation, and invert the objective function of the current collection strategy to construct an adversarial collection strategy.

[0007] The target corpus set and the adversarial corpus set are obtained by executing the current acquisition strategy and the adversarial acquisition strategy respectively. The corpus set that does not belong to the adversarial corpus set in the target corpus set is extracted to form the blind zone corpus set.

[0008] Calculate the distance between each corpus in the blind zone corpus and the center point of the collected corpus, select the corpus with the largest distance as the boundary corpus and perform collection;

[0009] Construct a feature mapping relationship from the boundary corpus to the collected corpus set, and determine the associated corpus set through the feature mapping relationship;

[0010] Obtain the quality label value of the boundary corpus, propagate the quality label value to each related corpus in the related corpus set to obtain the corrected quality value, obtain the utility measurement value of each related corpus in the related corpus set, propagate the utility measurement value to the boundary corpus to obtain the predicted utility value;

[0011] The quality label value and the predicted utility value are combined to obtain a comprehensive evaluation value, and the corpus quality control is performed based on the comprehensive evaluation value and the corrected quality value.

[0012] The current acquisition strategy is generated based on the corpus state representation. The adversarial acquisition strategy is constructed by negating the objective function of the current acquisition strategy, including:

[0013] Extract the feature values ​​of each dimension from the state representation of the corpus, use the feature values ​​of each dimension as independent variables to construct a coverage function with the optimization objective of maximizing corpus coverage, and solve the coverage function to obtain the policy vector as the current collection policy;

[0014] Add a negative sign to the function value in the coverage function to obtain the reverse coverage function. Input the feature values ​​of each dimension in the corpus state representation into the reverse coverage function, solve the reverse coverage function to obtain the adversarial strategy vector as the adversarial acquisition strategy.

[0015] The strategy vector is matched with the attribute information of each corpus to be collected to obtain the target matching value. The corpus to be collected is then filtered according to the target matching value to form the target corpus set.

[0016] The adversarial strategy vector is matched with the attribute information of each corpus to be collected to obtain the adversarial matching value. The corpus to be collected is then filtered according to the adversarial matching value to form an adversarial corpus set.

[0017] The target corpus is used to extract data that is not present in the adversarial corpus to form a blind zone corpus.

[0018] The target corpus and adversarial corpus are obtained by executing the current acquisition strategy and the adversarial acquisition strategy respectively. The corpus that does not belong to the adversarial corpus in the target corpus is extracted to form the blind zone corpus, which includes:

[0019] The target projection intensity is obtained by calculating the projection intensity of the attribute information of the corpus to be collected in the feature space defined by the policy vector. The corpus to be collected is selected to form a target corpus set and the target projection intensity corresponding to each corpus is recorded.

[0020] The adversarial projection intensity is obtained by calculating the projection intensity of the attribute information of the corpus to be collected in the feature space defined by the adversarial strategy vector. The corpus to be collected is selected to form an adversarial corpus set and the adversarial projection intensity corresponding to each corpus is recorded.

[0021] Extract the target projection intensity and adversarial projection intensity for each corpus in the target corpus set, and calculate the ratio of the target projection intensity to the adversarial projection intensity to obtain the projection intensity ratio;

[0022] The projection intensity ratio is used as a weight to weight each corpus in the target corpus set, and the deviation of the weighted corpus in the feature space of the collected corpus set is calculated to obtain the boundary deviation.

[0023] The target corpus is filtered based on the degree of boundary deviation, and the filtered corpus is determined to not belong to the adversarial corpus and constitutes the blind zone corpus.

[0024] Calculate the distance between each corpus in the blind zone and the center point of the collected corpus, select the corpus with the largest distance as the boundary corpus, and perform the collection process including:

[0025] Extract the attribute information of each corpus in the collected corpus set, calculate the distribution density of the attribute information, and obtain the center point of the collected corpus set by offsetting the geometric center in the feature space based on the distribution density;

[0026] Extract the attribute information of each corpus in the blind zone corpus set, and calculate the spatial distance between the attribute information of each corpus in the blind zone corpus set and the center point of the collected corpus set to obtain the baseline distance;

[0027] Extract the corresponding projection intensity ratio for each corpus in the blind zone corpus set, and use the projection intensity ratio as a direction vector to construct a directional path from the center point of the collected corpus set to each corpus in the blind zone corpus set;

[0028] Detection points are set along the directional path. The number of corpora in the collected corpus set within the preset range around each detection point is counted. The number of corpora at each detection point is summed to obtain the number of corpora covered by the path.

[0029] The distance between each corpus in the blind zone and the center point of the collected corpus is obtained by multiplying the inverse of the number of corpus points covered by the path and the baseline distance.

[0030] Sort the data in the blind zone corpus by the distance between each data point and the center point of the collected data set, and select the data point with the largest distance as the boundary data point and perform the collection.

[0031] Construct a feature mapping relationship from the boundary corpus to the collected corpus set, and determine the associated corpus set through the feature mapping relationship, including:

[0032] Extract the attribute information of the boundary corpus, and vectorize the attribute information of the boundary corpus to obtain the feature vector of the boundary corpus;

[0033] Extract the attribute information of each corpus in the collected corpus set, and vectorize the attribute information of each corpus in the collected corpus set to obtain the feature vector of the collected corpus;

[0034] Construct a propagation path from the boundary corpus to each corpus in the collected corpus set, extract the number of corpus covered by the path in the collected corpus set region traversed by the propagation path, and set a feature decay function based on the number of corpus covered by the path.

[0035] The feature vectors of the boundary corpus are attenuated by applying a feature attenuation function along the propagation path to obtain the attenuated feature vectors of each corpus position in the collected corpus set.

[0036] The matching strength is obtained by matching the decay feature vector with the feature vector of the collected corpus. The matching strength and the propagation path length are then fused to obtain the feature mapping relationship from the boundary corpus to each corpus in the collected corpus set.

[0037] The collected corpora are sorted according to their feature mapping relationships, and corpora whose feature mapping relationships exceed a preset mapping threshold are selected to form an associated corpus set.

[0038] Obtain the quality label values ​​of the boundary corpus, propagate these quality label values ​​to each related corpus in the associated corpus set to obtain corrected quality values, obtain the utility measurement values ​​of each related corpus in the associated corpus set, and propagate these utility measurement values ​​to the boundary corpus to obtain predicted utility values, including:

[0039] Extract the attribute information of the boundary corpus, calculate the spatial distance between the boundary corpus and the center point of the collected corpus set, and perform quality assessment on the attribute information of the boundary corpus based on the spatial distance to obtain the quality label value of the boundary corpus;

[0040] Extract the feature mapping relationship between the boundary corpus and each associated corpus in the associated corpus set, construct a quality attenuation coefficient based on the feature mapping relationship, apply the quality attenuation coefficient to the quality label value of the boundary corpus to attenuate and propagate to obtain the propagation quality value of each associated corpus in the associated corpus set;

[0041] Extract the initial quality value of each related corpus in the related corpus set, and perform a fusion operation between the propagation quality value and the initial quality value to obtain the corrected quality value of each related corpus in the related corpus set;

[0042] Based on the corrected quality value, the number of corpora covered by each associated corpus in the collected corpus is statistically analyzed, and the number of corpora covered is used as the utility measurement value of each associated corpus.

[0043] Extract the utility measurement value of each associated corpus and the corresponding feature mapping relationship. Backpropagate the utility measurement value to the boundary corpus along the reverse path of the feature mapping relationship. Accumulate the utility measurement values ​​backpropagated from each associated corpus in the associated corpus set to obtain the predicted utility value of the boundary corpus.

[0044] A comprehensive evaluation value is obtained by fusing the quality label value and the predicted utility value. Corpus quality control is then performed based on the comprehensive evaluation value and the corrected quality value, including:

[0045] The quality utility ratio is obtained by calculating the ratio of the quality-labeled value to the predicted utility value. The quality utility ratio is then used as a dynamic weight to weight and fuse the quality-labeled value and the predicted utility value to obtain the comprehensive evaluation value.

[0046] Extract the corrected quality value of each related corpus from the related corpus set, and calculate the deviation between the corrected quality value and the comprehensive evaluation value;

[0047] The associated corpus with a deviation exceeding the preset fluctuation threshold is marked as abnormal corpus. The feature mapping relationship between the abnormal corpus and the boundary corpus is extracted, and the path length of the feature mapping relationship is calculated to obtain the propagation distance.

[0048] The propagation attenuation coefficient is constructed by applying the propagation distance to the corrected quality value of the abnormal corpus. Based on the propagation attenuation coefficient, the corrected quality value is propagated back to the boundary corpus along the reverse path of the feature mapping relationship to obtain the backtracked quality value.

[0049] The quality value backtracking is compared with the quality label value. When the difference exceeds the preset correction threshold, the quality label value is iteratively corrected and the comprehensive evaluation value is recalculated to obtain the corrected comprehensive evaluation value.

[0050] Based on the corrected comprehensive evaluation value, the corrected quality value of each associated corpus in the associated corpus set is used to screen and implement corpus quality control.

[0051] A second aspect of this invention provides an automatic corpus acquisition and quality control system based on intelligent algorithms, comprising:

[0052] The strategy generation unit is used to obtain the attribute information of the corpus to be collected, construct the corpus state representation, generate the current collection strategy based on the corpus state representation, and invert the objective function of the current collection strategy to construct an adversarial collection strategy.

[0053] The blind spot extraction unit is used to execute the current acquisition strategy and the adversarial acquisition strategy respectively to obtain the target corpus set and the adversarial corpus set, and extract the corpus in the target corpus set that does not belong to the adversarial corpus set to form the blind spot corpus set;

[0054] The boundary selection unit is used to calculate the distance between each corpus in the blind zone corpus and the center point of the collected corpus, select the corpus with the largest distance as the boundary corpus and perform collection.

[0055] The association determination unit is used to construct the feature mapping relationship from the boundary corpus to the collected corpus set, and to determine the associated corpus set through the feature mapping relationship.

[0056] The propagation correction unit is used to obtain the quality label value of the boundary corpus, propagate the quality label value to each related corpus in the related corpus set to obtain the corrected quality value, obtain the utility measurement value of each related corpus in the related corpus set, and propagate the utility measurement value to the boundary corpus to obtain the predicted utility value.

[0057] The fusion control unit is used to fuse the quality label value and the predicted utility value to obtain a comprehensive evaluation value, and to perform corpus quality control based on the comprehensive evaluation value and the corrected quality value.

[0058] A third aspect of the present invention provides an electronic device, comprising:

[0059] processor;

[0060] Memory used to store processor-executable instructions;

[0061] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0062] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0063] This invention automatically locates uncovered blind spots by identifying differences between target and adversarial corpora. It also filters boundary corpora based on their distance from the center point of already collected corpora, ensuring maximum corpus diversity in each collection and avoiding redundant collection and resource waste. This is suitable for unstructured, highly variable corpus scenarios. Feature mapping transfers boundary corpus features to related corpora, enabling automatic extraction of semantic relationships between corpora. Quality annotation values ​​propagate along the mapping path, assigning corrected quality values ​​to related corpora, reducing manual annotation costs and ensuring consistency. Utility measurement values ​​propagate backward, providing predicted utility values ​​for boundary corpora, accurately assessing potential application value without requiring individual corpus evaluation, thus improving the timeliness of collection decisions. A comprehensive evaluation value integrates quality annotation and predicted utility, fully reflecting corpus quality and practicality, and dynamically filters low-quality or low-utility corpora through linked analysis. High-precision quality control significantly reduces manual intervention, achieving simultaneous optimization of collection efficiency and data quality. Attached Figure Description

[0064] Figure 1 This is a flowchart illustrating a method for automatic corpus acquisition and quality control based on intelligent algorithms.

[0065] Figure 2 Flowchart for calculating the predicted utility value and quality propagation of boundary corpus. Detailed Implementation

[0066] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0067] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0068] Figure 1 This is a flowchart illustrating the automatic corpus acquisition and quality control method based on intelligent algorithms according to an embodiment of the present invention.

[0069] The automatic data collection and quality control methods for corpora based on intelligent algorithms include:

[0070] Obtain the attribute information of the corpus to be collected, construct the corpus state representation, generate the current collection strategy based on the corpus state representation, and invert the objective function of the current collection strategy to construct an adversarial collection strategy.

[0071] The target corpus set and the adversarial corpus set are obtained by executing the current acquisition strategy and the adversarial acquisition strategy respectively. The corpus set that does not belong to the adversarial corpus set in the target corpus set is extracted to form the blind zone corpus set.

[0072] Calculate the distance between each corpus in the blind zone corpus and the center point of the collected corpus, select the corpus with the largest distance as the boundary corpus and perform collection;

[0073] Construct a feature mapping relationship from the boundary corpus to the collected corpus set, and determine the associated corpus set through the feature mapping relationship;

[0074] Obtain the quality label value of the boundary corpus, propagate the quality label value to each related corpus in the related corpus set to obtain the corrected quality value, obtain the utility measurement value of each related corpus in the related corpus set, propagate the utility measurement value to the boundary corpus to obtain the predicted utility value;

[0075] The quality label value and the predicted utility value are combined to obtain a comprehensive evaluation value, and the corpus quality control is performed based on the comprehensive evaluation value and the corrected quality value.

[0076] The current acquisition strategy is generated based on the corpus state representation. The adversarial acquisition strategy is constructed by negating the objective function of the current acquisition strategy, including:

[0077] Extract the feature values ​​of each dimension from the state representation of the corpus, use the feature values ​​of each dimension as independent variables to construct a coverage function with the optimization objective of maximizing corpus coverage, and solve the coverage function to obtain the policy vector as the current collection policy;

[0078] Add a negative sign to the function value in the coverage function to obtain the reverse coverage function. Input the feature values ​​of each dimension in the corpus state representation into the reverse coverage function, solve the reverse coverage function to obtain the adversarial strategy vector as the adversarial acquisition strategy.

[0079] The strategy vector is matched with the attribute information of each corpus to be collected to obtain the target matching value. The corpus to be collected is then filtered according to the target matching value to form the target corpus set.

[0080] The adversarial strategy vector is matched with the attribute information of each corpus to be collected to obtain the adversarial matching value. The corpus to be collected is then filtered according to the adversarial matching value to form an adversarial corpus set.

[0081] The target corpus is used to extract data that is not present in the adversarial corpus to form a blind zone corpus.

[0082] After obtaining the state representation of the corpus, it is necessary to extract feature values ​​for each dimension. These feature values ​​encompass quantitative descriptions of multiple dimensions of the corpus, such as domain distribution, semantic density, topic diversity, and vocabulary coverage. Each dimensional feature value corresponds to a measurable attribute in the corpus space. All dimensional feature values ​​are organized into a vector form, denoted as [vector]. ,in Indicates the first Feature values ​​in each dimension This represents the total number of feature dimensions. As the independent variable, a coverage function is constructed with the optimization objective of maximizing corpus coverage. The core idea of ​​the coverage function is to measure the range of corpus space that the collection strategy can reach under the current corpus state. The larger the value, the wider the corpus distribution area that the strategy can cover, thereby reducing the collection blind spot.

[0083] Overriding function The construction of this approach typically incorporates submodular function theory, modeling the corpus coverage problem as maximizing a measure of diversity between the collected corpus and the corpus to be collected in the feature space. For the coverage function... To solve this problem, we need to find a solution that satisfies the resource acquisition constraints. The strategy configuration that achieves the maximum value is solved as a strategy vector. The output is in the form of a policy vector, which serves as the current acquisition strategy. Each component in the algorithm corresponds to a weight or priority of a data collection operation, guiding the subsequent screening and sorting of the data to be collected. The solution process can employ greedy optimization or gradient ascent methods, converging to a local optimum within a finite number of iterations to ensure the policy vector... An effective estimate can be obtained within the acceptable range of computing resources.

[0084] After obtaining the covering function and policy vector Then, by adding a negative sign to the function values ​​in the overriding function, an inverse overriding function is constructed. Reverse overlay function The optimization objective is exactly the opposite of the coverage function, namely, minimizing corpus coverage, so that the solution tends to focus on regions in the corpus space that the coverage function considers "unimportant" or "low-value". This involves representing the feature values ​​of each dimension in the corpus state representation. Input reverse overwrite function Solving for it yields the adversarial strategy vector. Adversarial strategy vector This represents a collection tendency that contradicts the current collection strategy. Its purpose is to actively explore corpus regions that are easily overlooked by the current strategy, thereby providing a comparative benchmark for subsequent blind spot identification. This design draws on the idea of ​​the game between the generator and the discriminator in adversarial learning, exposing potential coverage gaps in the collection process through the differences between the two strategies.

[0085] Obtain the policy vector Next, it is matched with the attribute information of each piece of corpus to be collected. Each piece of corpus to be collected has a corresponding attribute vector. ,in This refers to the index number of the corpus to be collected. The matching operation calculates the strategy vector. With attribute vectors The similarity or inner product between them yields the target matching value. Target matching value Reflects the first The degree of fit between the data to be collected and the current collection strategy; a higher value indicates that the data better matches the collection goal of maximizing coverage. (Based on target matching value) All data to be collected are sorted in descending order, and data with matching values ​​exceeding a preset threshold or those ranked highly are selected to form the target data set. Target corpus set This represents the subset of corpus that the current collection strategy deems most worthy of priority collection, and the corpus within it has high distribution diversity and coverage value in the feature space.

[0086] Similarly, the adversarial strategy vector Attribute vectors of each corpus to be collected Perform matching operations to obtain the adversarial matching value. Adversarial match value Measure the first The degree of fit between the corpus to be collected and the adversarial collection strategy. Based on adversarial matching value. The collected corpus is filtered, and corpus with matching values ​​exceeding the corresponding thresholds is selected to form an adversarial corpus set. Confrontational Corpus The corpus in the target set is the portion that the adversarial strategy considers valuable for acquisition. This corpus often concentrates in certain local regions within the feature space, unlike the target corpus. There is some overlap in their distributions, but also some differences. It's important to note the target matching value. Matching value with adversaries Different strategy vectors are used for calculation; they have different meanings and should not be used interchangeably.

[0087] After obtaining the target corpus set adversarial corpus Then, the target corpus is extracted from the adversarial corpus by using set difference operations, forming the blind zone corpus. Blind spot corpus collection The construction logic lies in: target corpus set The corpus data in question has been identified as valuable by the current acquisition strategy, but adversarial acquisition strategies have failed to include it in their acquisition scope. This indicates that this portion of the corpus lies in a "disagreement zone" between the two strategies—that is, an area that the current acquisition strategy can discover but the adversarial strategy cannot cover. This discrepancy precisely reveals a potential blind spot in the acquisition process—these corpora are located at the boundary in the feature space, possessing certain coverage value but easily overlooked by conventional strategies. This can be addressed by explicitly constructing a set of blind spot corpora. This allows subsequent data collection resources to be concentrated on these highly uncertain areas, improving the completeness and balance of the overall corpus collection.

[0088] The size of the blind zone corpus directly affects the candidate range for subsequent boundary corpus selection. A large number of entries in the blind zone indicates a significant divergence between the current acquisition strategy and the adversarial acquisition strategy, suggesting numerous unexplored boundary regions within the corpus space. Conversely, a smaller blind zone indicates that the coverage of both strategies is converging, indicating a relatively comprehensive acquisition process. In practical applications, the iteration rounds of the acquisition strategy can be dynamically adjusted based on the size of the blind zone corpus. When the blind zone corpus shrinks to below a preset proportion, a termination condition is triggered, achieving a balance between acquisition efficiency and coverage completeness. The entire process, through the synergy of strategy vectors and adversarial strategy vectors, achieves adaptive detection of the corpus space boundaries, providing high-value candidate corpus input for subsequent quality control processes.

[0089] The target corpus and adversarial corpus are obtained by executing the current acquisition strategy and the adversarial acquisition strategy respectively. The corpus that does not belong to the adversarial corpus in the target corpus is extracted to form the blind zone corpus, which includes:

[0090] The target projection intensity is obtained by calculating the projection intensity of the attribute information of the corpus to be collected in the feature space defined by the policy vector. The corpus to be collected is selected to form a target corpus set and the target projection intensity corresponding to each corpus is recorded.

[0091] The adversarial projection intensity is obtained by calculating the projection intensity of the attribute information of the corpus to be collected in the feature space defined by the adversarial strategy vector. The corpus to be collected is selected to form an adversarial corpus set and the adversarial projection intensity corresponding to each corpus is recorded.

[0092] Extract the target projection intensity and adversarial projection intensity for each corpus in the target corpus set, and calculate the ratio of the target projection intensity to the adversarial projection intensity to obtain the projection intensity ratio;

[0093] The projection intensity ratio is used as a weight to weight each corpus in the target corpus set, and the deviation of the weighted corpus in the feature space of the collected corpus set is calculated to obtain the boundary deviation.

[0094] The target corpus is filtered based on the degree of boundary deviation, and the filtered corpus is determined to not belong to the adversarial corpus and constitutes the blind zone corpus.

[0095] After acquiring the attribute information of the corpus to be collected, completing the corpus state representation, and generating the current collection strategy and the adversarial collection strategy, it is necessary to execute the two types of strategies to filter the corpus separately, and identify the blind spot corpus that can be covered by the current collection strategy but cannot be covered by the adversarial collection strategy, so as to provide a basis for the selection of the subsequent boundary corpus.

[0096] For the current data collection strategy, the attribute information of the corpus to be collected is projected onto the strategy vector. In the defined feature space, the projection intensity of each piece of corpus to be collected is calculated in the direction of that feature space to obtain the target projection intensity. Specifically, for the ... The corpus to be collected, its attribute vector In the policy vector Target projection intensity in the direction This reflects the degree of matching between the corpus and the current acquisition strategy. A higher projection intensity indicates that the corpus better aligns with the acquisition direction defined by the current acquisition strategy, and therefore should be included in the target corpus set. A screening threshold is set based on the target projection intensity, and corpora that meet the threshold conditions are included in the target corpus set. Simultaneously, the target projection intensity corresponding to each corpus is fully recorded. This is for use in subsequent calculations.

[0097] To counter adversarial data collection strategies, the attribute information of the corpus to be collected is projected onto the adversarial strategy vector. In the defined feature space, the projection intensity of each piece of corpus to be collected in the adversarial direction is calculated to obtain the adversarial projection intensity. For the ... The corpus to be collected, and its corresponding adversarial projection strength This reflects the degree of matching of the corpus from the perspective of adversarial data collection strategies. A higher adversarial projection strength indicates that the corpus is more easily identified and covered by adversarial data collection strategies, and its value in the blind spot corpus set is relatively low. A screening threshold is set based on the adversarial projection strength, and corpora meeting the threshold conditions are included in the adversarial corpus set. Similarly, the adversarial projection strength corresponding to each corpus is fully recorded. .

[0098] After completing the target corpus sets respectively adversarial corpus After construction, targeting the target corpus set For each piece of data, the corresponding target projection intensity is extracted simultaneously. and the intensity of the opposing projection And calculate the ratio of the two to obtain the projection intensity ratio. ,Right now ,in A minimal positive smoothing term is introduced to prevent the denominator from being zero. Projection intensity ratio. The physical meaning is that when the projection intensity of a corpus in the current acquisition strategy direction is much greater than its projection intensity in the adversarial strategy direction, the corpus is more likely to be in the blind spot of the adversarial strategy and has a higher potential to be included in the blind spot corpus set; conversely, if If the value is close to 1 or even less than 1, it indicates that the corpus also has a strong response in the adversarial direction and is difficult to identify as blind zone corpus.

[0099] Projection intensity ratio As a weight, for the target corpus set The attribute vectors of each corpus are weighted to obtain the weighted corpus representation. The purpose of weighting is to amplify the feature responses of corpora that have high potential in the blind spots of adversarial strategies, while suppressing the feature contributions of corpora that have been fully covered by adversarial strategies, so that the subsequent deviation calculation can more accurately reflect the true distribution of corpora in the feature space of the collected corpus set.

[0100] After weighting, the weighted corpus representation is calculated. The degree of deviation in the feature space of the collected corpus is used to obtain the boundary deviation. Specifically, using the distribution center of the collected corpus in the feature space as a reference, the calculation is performed. The greater the distance from the center, the greater the boundary deviation. The higher the value, the more marginal the corpus is in the feature distribution of the collected corpus set. Boundary Deviation This comprehensive assessment of the corpus's novelty and its adversarial blind spot attributes is the core criterion for determining whether a corpus should be included in the blind spot corpus set. A boundary deviation threshold is set. , will satisfy The corpus is from the target corpus set The selected data were used as candidate blind spot corpora.

[0101] The candidate corpora selected above are further subjected to attribution determination to verify whether they belong to the adversarial corpus set. The determination method is as follows: if a candidate corpus does not appear in the adversarial corpus set, it is confirmed as a blind zone corpus and included in the blind zone corpus set. If the corpus also appears in the adversarial corpus set, it means that the adversarial acquisition strategy has covered the corpus, and it will not be included in the blind zone corpus set. This determination step ensures that the corpus in the blind zone corpus set has a true "blind zone" attribute, that is, it is only identified by the current acquisition strategy and not covered by the adversarial acquisition strategy, thereby ensuring the accuracy and representativeness of the subsequent selection of boundary corpus.

[0102] Through the above process, the corpus in the final blind zone collection possesses two key attributes: first, it exhibits a strong matching response in the direction of the current acquisition strategy, falling within the effective coverage of the current strategy; second, its response in the direction of adversarial acquisition strategies is relatively weak, failing to be recognized by adversarial strategies and existing in the blind zone of adversarial strategy coverage. These two attributes together ensure the blind zone corpus... This effectively reveals the coverage differences between the current data collection strategy and adversarial data collection strategies, providing a reliable data foundation for subsequently selecting boundary data from the blind zone corpus and expanding the coverage boundaries of the collected corpus. In practical applications, the threshold... The settings can be dynamically adjusted according to the size and distribution density of the collected corpus set to ensure that the size of the blind zone corpus set is moderate, neither too sparse to make it difficult to select boundary corpus, nor too redundant to waste computing resources.

[0103] Calculate the distance between each corpus in the blind zone and the center point of the collected corpus, select the corpus with the largest distance as the boundary corpus, and perform the collection process including:

[0104] Extract the attribute information of each corpus in the collected corpus set, calculate the distribution density of the attribute information, and obtain the center point of the collected corpus set by offsetting the geometric center in the feature space based on the distribution density;

[0105] Extract the attribute information of each corpus in the blind zone corpus set, and calculate the spatial distance between the attribute information of each corpus in the blind zone corpus set and the center point of the collected corpus set to obtain the baseline distance;

[0106] Extract the corresponding projection intensity ratio for each corpus in the blind zone corpus set, and use the projection intensity ratio as a direction vector to construct a directional path from the center point of the collected corpus set to each corpus in the blind zone corpus set;

[0107] Detection points are set along the directional path. The number of corpora in the collected corpus set within the preset range around each detection point is counted. The number of corpora at each detection point is summed to obtain the number of corpora covered by the path.

[0108] The distance between each corpus in the blind zone and the center point of the collected corpus is obtained by multiplying the inverse of the number of corpus points covered by the path and the baseline distance.

[0109] Sort the data in the blind zone corpus by the distance between each data point and the center point of the collected data set, and select the data point with the largest distance as the boundary data point and perform the collection.

[0110] In determining the boundary corpus, the calculation of the center point of the collected corpus set is not simply the geometric mean of the attribute vectors of each corpus, but requires positional correction based on the distribution density of the corpus in the feature space. After extracting the attribute information of each corpus in the collected corpus set, the density of the entire feature space is estimated to obtain the distribution density value of each region. Regions with higher density indicate that a large amount of corpus has been collected in that direction, and have a stronger "attraction" to the center point; regions with lower density mean relatively sparse coverage. Based on this density distribution, an offset is applied to the geometric center of the feature space, causing the center point to move closer to the direction with higher density, thus obtaining the center point of the collected corpus set that truly reflects the center of gravity of the collected corpus distribution. This approach avoids the problem of distorted distance calculations caused by a small number of outliers shifting the geometric center.

[0111] Suppose that the collected corpus contains a total of The corpus contains 10 data points, and their attribute vectors are 100. ( ), in the feature space, for position The distribution density at location is estimated as follows: The density-weighted center point is calculated as follows: ,in For the first The local density estimates of the data at the attribute vector positions of the collected data are as follows: This represents the total number of corpora in the collected dataset. The center point is determined through density weighting. It can more accurately reflect the distribution center of the collected corpus, rather than a pure geometric center disturbed by sparse edge points.

[0112] Obtain the center point Then, the attribute information of each corpus in the blind zone corpus is extracted, and the attribute vector and center point of each blind zone corpus are calculated one by one. The Euclidean distance between them is denoted as the reference distance. subscript The first in the corresponding blind spot corpus The baseline distance reflects the original distance of the blind zone corpus relative to the centroid of the collected corpus in the feature space. However, the baseline distance alone cannot accurately depict the actual location of the blind zone, because there may already be a large amount of collected corpus along the path from the center point to a certain blind zone corpus. This path direction is actually fully covered and does not constitute a true blank area in the collection.

[0113] To solve this problem, the projection intensity ratio is introduced. As a direction vector, construct the system from the center point of the collected corpus. The first in the blind spot corpus The directional path of the corpus. Projection intensity ratio. The data, calculated for each piece of corpus to be collected in the previous steps, encodes the directional deviation information between the current collection strategy and the adversarial collection strategy. This effectively indicates which direction to extend from the center point to reach the true collection blind spot. As a direction vector, for from To the Blind zone corpus attribute vector The connection path is parameterized, and several detection points are evenly set along the path.

[0114] The number of detection points set along each directional path can be determined based on the reference distance. The size of the detection points is adaptively determined; the longer the path, the denser the detection points, ensuring the accuracy of path coverage evaluation. For each detection point on the path, the number of corpora in the collected corpus set within its preset range is counted. The preset range can be set as a sphere in the feature space with the detection point as the center and a fixed radius. The number of corpora within a hypersphere region of radius is calculated by summing the corpora around all detection points along a path in a given direction. This gives the total number of corpora covered by that path. . The larger the value, the greater the distance from the center point to the th. The more data has been collected along the path of a blind zone, the higher the coverage of that direction, and the corresponding blind zone data is not actually located in the true data collection blank area.

[0115] Based on path coverage corpus Distance from the reference point The first word in the calculation blind zone corpus The final distance between the corpus and the center point of the collected corpus set The calculation method is as follows:

[0116] The denominator is incremented by 1 to prevent a division-by-zero exception when the number of path-covered corpora is zero. The physical meaning can be understood as: taking into account the coverage of the collected corpus along the path direction, the first The "effective distance" of a blind zone corpus relative to the center point. If the baseline distance of a blind zone corpus is large, but the corpus collected along its path direction is very dense ( (larger), then The data will be significantly compressed, indicating that the corpus is not a true boundary of the data collection blind spot; conversely, if the data collected along the path direction is sparse, near This indicates that there is indeed a significant gap in data collection in this direction, and the corresponding corpus is more likely to be a true boundary corpus.

[0117] All data in the blind spot corpus are processed according to... Sort from largest to smallest, and select... The largest corpus is designated as the boundary corpus. The boundary corpus represents the corpus furthest from the currently collected corpus in the feature space along the direction most in need of supplementary collection. Collecting from this boundary corpus maximizes the expansion of the coverage boundary of the collected corpus, improving corpus diversity and coverage balance. Once the boundary corpus is determined, it is incorporated into the collected corpus, triggering subsequent feature mapping relationship construction and quality propagation processes, thus forming a dynamic iterative boundary exploration mechanism throughout the collection process.

[0118] This distance metric combines geometric distance in feature space with actual coverage density along the path, avoiding the drawbacks of pure geometric distance ignoring path occlusion effects and the limitation of relying solely on density estimation while neglecting directional information. It can more accurately locate the corpus that is truly at the boundary of the blank collection area, providing high-value labeled samples for subsequent quality control.

[0119] Construct a feature mapping relationship from the boundary corpus to the collected corpus set, and determine the associated corpus set through the feature mapping relationship, including:

[0120] Extract the attribute information of the boundary corpus, and vectorize the attribute information of the boundary corpus to obtain the feature vector of the boundary corpus;

[0121] Extract the attribute information of each corpus in the collected corpus set, and vectorize the attribute information of each corpus in the collected corpus set to obtain the feature vector of the collected corpus;

[0122] Construct a propagation path from the boundary corpus to each corpus in the collected corpus set, extract the number of corpus covered by the path in the collected corpus set region traversed by the propagation path, and set a feature decay function based on the number of corpus covered by the path.

[0123] The feature vectors of the boundary corpus are attenuated by applying a feature attenuation function along the propagation path to obtain the attenuated feature vectors of each corpus position in the collected corpus set.

[0124] The matching strength is obtained by matching the decay feature vector with the feature vector of the collected corpus. The matching strength and the propagation path length are then fused to obtain the feature mapping relationship from the boundary corpus to each corpus in the collected corpus set.

[0125] The collected corpora are sorted according to their feature mapping relationships, and corpora whose feature mapping relationships exceed a preset mapping threshold are selected to form an associated corpus set.

[0126] When extracting attribute information from boundary corpora, it is necessary to convert their multi-dimensional raw attributes, such as language form, semantic category, source domain, and syntactic structure, into a unified vector representation. Specifically, the text content of the boundary corpus is segmented and embedded, and word-level semantic information is aggregated into sentence-level dense vectors. At the same time, the metadata attributes of the corpus (such as collection source, corpus length, domain labels, etc.) are encoded as additional feature dimensions, which are then concatenated with the semantic vectors to form a complete boundary corpus feature vector. For the first digit in the collected corpus... Each corpus undergoes the same vectorization process to obtain the feature vectors of the collected corpus. During the vectorization process, the feature vectors of the boundary corpus With the feature vectors of the collected corpus Using the same feature extraction model and normalization method ensures that both are in the same feature space, providing a consistent metric basis for the construction and matching of subsequent propagation paths.

[0127] The propagation path is constructed starting from the feature vector position of the boundary corpus, and then from the 1st feature vector in the already collected corpus. The feature vector position of each corpus is the endpoint, and a directed path is formed by connecting the two points in the feature space. .path Traversing the region covered by the collected corpus in the feature space, several detection points are set at equal intervals along the path, and the radius of a preset range around each detection point is calculated. The number of collected corpora within the path is calculated by summing the statistical results of all detection points along the path. Path coverage corpus number Number of corpora covered by the path This reflects the density of the corpus in the area traversed by the propagation path: when the path passes through a dense area, A larger value indicates that the path direction is already covered by a large amount of corpus, and the feature information will experience strong attenuation during propagation; when the path traverses a sparse region, The smaller the size, the weaker the attenuation effect, and the more complete the feature information of the boundary corpus can be propagated to the target location.

[0128] Based on path coverage corpus Set the characteristic decay function Used to describe the feature vectors of the boundary corpus along the propagation path The attenuation pattern during propagation. The design of the feature attenuation function follows these principles: the larger the corpus covered by the path, the larger the attenuation coefficient, and the more significant the difference between the feature vector at the propagation endpoint and the feature vector of the original boundary corpus; the smaller the corpus covered by the path, the smaller the attenuation coefficient, and the higher the degree of feature preservation. The path length... (i.e., the feature vector of the boundary corpus and the first) Euclidean distance between feature vectors of collected corpora and number of corpora covered by the path Substituting all the values ​​into the decay function, we obtain the path-specific values. Comprehensive attenuation coefficient The calculation method is as follows ,in For attenuation intensity adjustment parameters, To prevent the smoothing of extremely small positive terms with a denominator of zero, an attenuation coefficient is used. The range of values ​​is When the path covers a very small amount of corpus and the path is long, A value approaching 1 indicates almost no decay; when the path covers a very large number of corpora, A value close to 0 indicates that the feature information has almost completely decayed.

[0129] Feature vectors of boundary corpus Along the propagation path The feature decay function is applied for decay propagation to obtain the first digit of the collected corpus. Decaying feature vector at the position of the corpus The calculation method is as follows Decaying eigenvector The original feature orientation of the boundary corpus is preserved, but the amplitude is scaled according to the density characteristics of the propagation path. This results in a smaller amplitude of the decayed feature vector obtained in dense feature space regions, and a larger amplitude of the decayed feature vector obtained in sparse regions. The physical significance of this design is that in densely populated regions of the collected corpus, the contribution of new information carried by the boundary corpus is relatively limited, and the degree of correlation with the collected corpus in that region is suppressed; while in sparse regions, the feature information of the boundary corpus can more effectively establish a correlation with the target corpus, demonstrating the stronger information extension capability of the boundary corpus in the sparse direction.

[0130] Matching decay feature vectors With the feature vectors of the collected corpus At that time, cosine similarity is used to measure the directional consistency between the two in the feature space, and the matching strength is obtained. The calculation method is as follows Match strength This reflects the degree of semantic agreement between the boundary corpus features after attenuation propagation and the features of the collected corpus, with a value range of [value range missing]. A larger value indicates a more consistent semantic direction. This relates to the matching strength. With propagation path length Perform fusion operations to obtain the boundary corpus up to the th Feature mapping relationship values ​​of the collected corpus The calculation method is as follows ,in is the path length penalty coefficient, used to control the suppression strength of path distance on the mapping relationship. The core logic of the fusion operation is that even if two corpora are highly consistent in the semantic direction, if the propagation path is too long, the mapping relationship value will be reduced due to the distance penalty, thereby avoiding the inclusion of corpora that are far apart in the feature space into the association set, and ensuring the locality and compactness of the association corpus set in the feature space.

[0131] Based on feature mapping relationship values All of the collected corpus Sort the corpus in descending order, and Exceeding the preset mapping threshold The corpus was filtered out to form a related corpus set. Mapping threshold The setting needs to comprehensively consider the size of the collected corpus and the distribution of the feature space: when the size of the collected corpus is small, the setting can be appropriately reduced. To include more potentially related corpora and ensure the coverage of subsequent quality dissemination; when the collected corpus is large in scale and has dense feature distribution, the accuracy can be appropriately increased. This ensures that the associated corpus set contains only corpora highly related to the boundary corpora, avoiding the introduction of noise propagation. (Associated corpus set) The construction result directly determines the range of subsequent quality labeling value propagation and utility measurement value backpropagation, and is the core data structure connecting the boundary corpus and the collected corpus set in the corpus quality control process. Through the complete construction process of the above feature mapping relationship, it is possible to accurately and reliably determine the associated corpus set while fully considering the density and distance characteristics of the propagation path, while preserving the original feature information of the boundary corpus.

[0132] like Figure 2 As shown, Figure 2This is a flowchart illustrating the calculation of the predicted utility value of the boundary corpus and the quality propagation operation in an embodiment of the present invention.

[0133] Obtain the quality label values ​​of the boundary corpus, propagate these quality label values ​​to each related corpus in the associated corpus set to obtain corrected quality values, obtain the utility measurement values ​​of each related corpus in the associated corpus set, and propagate these utility measurement values ​​to the boundary corpus to obtain predicted utility values, including:

[0134] Extract the attribute information of the boundary corpus, calculate the spatial distance between the boundary corpus and the center point of the collected corpus set, and perform quality assessment on the attribute information of the boundary corpus based on the spatial distance to obtain the quality label value of the boundary corpus;

[0135] Extract the feature mapping relationship between the boundary corpus and each associated corpus in the associated corpus set, construct a quality attenuation coefficient based on the feature mapping relationship, apply the quality attenuation coefficient to the quality label value of the boundary corpus to attenuate and propagate to obtain the propagation quality value of each associated corpus in the associated corpus set;

[0136] Extract the initial quality value of each related corpus in the related corpus set, and perform a fusion operation between the propagation quality value and the initial quality value to obtain the corrected quality value of each related corpus in the related corpus set;

[0137] Based on the corrected quality value, the number of corpora covered by each associated corpus in the collected corpus is statistically analyzed, and the number of corpora covered is used as the utility measurement value of each associated corpus.

[0138] Extract the utility measurement value of each associated corpus and the corresponding feature mapping relationship. Backpropagate the utility measurement value to the boundary corpus along the reverse path of the feature mapping relationship. Accumulate the utility measurement values ​​backpropagated from each associated corpus in the associated corpus set to obtain the predicted utility value of the boundary corpus.

[0139] After extracting the attribute information of the boundary corpus, the centroid of the boundary corpus and the collected corpus set is calculated. The spatial distance between the boundary corpora reflects the degree of deviation of the boundary corpora from the existing collection distribution in the feature space: the greater the distance, the sparser the feature region where the corpus is located, and the more difficult it is to assess its quality, but it also means that the corpus has a higher potential value for expanding the collection coverage. Based on this spatial distance, the quality of the attribute information of the boundary corpora is assessed by comprehensively weighting the distance factor with dimensions such as semantic completeness and format regularity in the attribute information to obtain the quality label value of the boundary corpora. Specifically, the greater the spatial distance, the higher the component related to scarcity in the corresponding quality assessment weights, so that the quality label value can simultaneously reflect the comprehensive evaluation of both the content quality and the scarcity of the corpus.

[0140] After obtaining the quality label values, extract the boundary corpus and related corpus sets. Feature mapping relationship values ​​between various related corpora .because The value has already been calculated in the feature mapping stage; it can be directly reused here to construct the mass attenuation coefficient. Mass attenuation coefficient The design principle is: feature mapping relationship value The higher the value, the better. The stronger the correlation between the collected corpus and the boundary corpus in the feature space, the smaller the attenuation of quality information during the propagation process should be. Should follow The quality label value of the boundary corpus increases with the increase of [something]. Application of mass attenuation coefficient Perform attenuation propagation to obtain the first [word] in the associated corpus. The propagation quality score of related corpora ,satisfy This propagation mechanism ensures that quality information can be effectively diffused along the feature mapping relationship, while avoiding excessive interference from quality information on paths with weak feature correlations.

[0141] In obtaining the transmission quality value Next, it is necessary to extract the initial quality values ​​of each related corpus in the related corpus set. The initial quality score is derived from the historical quality assessment results recorded when the corpus entered the collected corpus set, reflecting the inherent quality level of the corpus before it was affected by boundary corpora. The quality score will be propagated. Compared with the initial mass value Perform fusion calculations to obtain the corrected quality value. The fusion calculation uses a weighted average method, introducing fusion weight parameters. , making ,in The proportion of the contribution of the propagation quality value and the initial quality value to the correction result is controlled. When When the value is large, the corrected quality value is more affected by the quality annotation value of the boundary corpus; when When the value is small, the corrected quality value of the associated corpus retains more of its historical quality assessment results. This fusion mechanism, while introducing quality information from newly collected corpora, preserves the quality history of existing corpora, preventing the quality annotation of a single boundary corpus from excessively interfering with the entire associated corpus set.

[0142] Complete the corrected quality value After calculation, based on the corrected quality value, the number of corpora covered by each associated corpus in the associated corpus set in the collected corpus set is counted. The specific method of coverage statistics is as follows: taking the first... Centering on the related corpus, within the feature space of the collected corpus, statistical analysis and its modified quality values ​​are performed. Similar features with a feature distance within a preset coverage radius The number of corpora within the scope is denoted as the number of covered corpora. . This reflects the actual coverage capability of the associated corpus to the collected corpus set under the revised quality assessment framework: The larger the value, the more effectively the associated corpus, after quality correction, can represent the quality status of more collected corpora, and the higher its utility in the collection system. This indicates the number of corpora covered. This indicator, which can be directly used as a utility measure for each associated corpus, is both simple and intuitive, and can quantify the actual contribution of associated corpora to the collected corpus from a quantitative perspective.

[0143] Obtain utility measurements for each related corpus. Corresponding feature mapping value Then, the utility measurement value is backpropagated to the boundary corpus along the reverse path of the feature mapping relationship. The direction of backpropagation is opposite to the direction of forward propagation of the quality label value: forward propagation starts from the boundary corpus and proceeds along the propagation path. The quality information is disseminated to all related corpora; backpropagation, on the other hand, starts from each related corpus and proceeds along... The reverse path gathers utility information back into the boundary corpus. The utility backtracking value for each reverse path... From utility measurement value Features mapping relationship value Joint decision, to meet In other words, the stronger the feature mapping relationship, the greater the contribution of the associated corpus to the boundary corpus. Accumulate the associated corpus set. The backpropagation utility values ​​of all related corpora are used to obtain the predicted utility values ​​of the boundary corpus. ,satisfy This accumulation mechanism ensures that the predicted utility value of the boundary corpus can comprehensively reflect its potential impact range and intensity on the entire associated corpus set, providing a reliable quantitative basis for subsequent comprehensive evaluation.

[0144] Predicted utility value The physical significance lies in the fact that after estimating the boundary corpus, it can effectively correct and improve the quality status of how many corpora in the collected corpus set through feature mapping relationship transmission. A higher quality rating indicates a higher utility return for the decision to collect this boundary corpus. In resource-constrained collection scenarios, this indicator can serve as one of the important bases for prioritizing collection decisions. This is achieved by using quality-labeled values... Compared with predicted utility value Integration can comprehensively evaluate boundary corpora from two complementary dimensions: quality assurance and collection effectiveness, thereby driving the refined execution of subsequent corpus quality control processes.

[0145] A comprehensive evaluation value is obtained by fusing the quality label value and the predicted utility value. Corpus quality control is then performed based on the comprehensive evaluation value and the corrected quality value, including:

[0146] The quality utility ratio is obtained by calculating the ratio of the quality-labeled value to the predicted utility value. The quality utility ratio is then used as a dynamic weight to weight and fuse the quality-labeled value and the predicted utility value to obtain the comprehensive evaluation value.

[0147] Extract the corrected quality value of each related corpus from the related corpus set, and calculate the deviation between the corrected quality value and the comprehensive evaluation value;

[0148] The associated corpus with a deviation exceeding the preset fluctuation threshold is marked as abnormal corpus. The feature mapping relationship between the abnormal corpus and the boundary corpus is extracted, and the path length of the feature mapping relationship is calculated to obtain the propagation distance.

[0149] The propagation attenuation coefficient is constructed by applying the propagation distance to the corrected quality value of the abnormal corpus. Based on the propagation attenuation coefficient, the corrected quality value is propagated back to the boundary corpus along the reverse path of the feature mapping relationship to obtain the backtracked quality value.

[0150] The quality value backtracking is compared with the quality label value. When the difference exceeds the preset correction threshold, the quality label value is iteratively corrected and the comprehensive evaluation value is recalculated to obtain the corrected comprehensive evaluation value.

[0151] Based on the corrected comprehensive evaluation value, the corrected quality value of each associated corpus in the associated corpus set is used to screen and implement corpus quality control.

[0152] Obtaining quality labeling values ​​for boundary corpora Compared with predicted utility value Next, the two need to be integrated into a single evaluation index that comprehensively reflects the value of the corpus. The integration process is not a simple weighted average with a fixed ratio, but rather uses the ratio of the two as a dynamic weight, thus adaptively biasing the integration result towards the more discriminative one. Specifically, the quality-utility ratio is calculated. ,in To prevent the introduction of a very small positive smoothing term when the denominator is zero. As dynamic weights, and A weighted fusion is performed to obtain a comprehensive evaluation value. , expressed as When the quality-labeled value is higher than the predicted utility value, As the weights increase, the overall evaluation value tilts towards the quality label value; conversely, it reflects the contribution of the predicted utility value more. This adaptive fusion mechanism avoids the mismatch problem of fixed weights in different corpus distribution scenarios, and maintains a dynamic balance between corpus quality and coverage utility in the overall evaluation value.

[0153] After obtaining the comprehensive evaluation value Next, extract the associated corpus set. Corrected quality values ​​of various related corpora Calculate the deviation between the corrected quality value and the overall evaluation value for each related corpus. Defined as Deviation This reflects the gap between the quality of the related corpus and the comprehensive evaluation of the boundary corpus. When the deviation of a certain piece of related corpus exceeds the preset fluctuation threshold... At this point, the corpus was marked as anomalous, indicating a significant discrepancy between its corrected quality value and overall evaluation value, requiring further source tracing analysis. Preset fluctuation threshold. The threshold can be adaptively set according to the overall quality distribution statistical characteristics of the corpus, for example, by taking a multiple of the standard deviation of the quality value distribution to ensure the universality of the threshold for corpus sets of different sizes.

[0154] For each piece of text marked as anomalous, extract the established feature mapping relationship between it and the boundary text. And calculate the path length corresponding to this mapping relationship, i.e., the propagation distance. , is the feature vector of the boundary corpus With the The Euclidean distance between the feature vectors of the anomalous corpus. Propagation distance. It characterizes the distance of anomalous corpora from the boundary corpora in the feature space. The greater the distance, the more the quality signal has been attenuated during the propagation process, and the credibility of the anomalous deviation also needs to be discounted accordingly.

[0155] Based on propagation distance Construct a propagation attenuation coefficient for each anomalous corpus. The calculation method is as follows ,in This is the attenuation intensity adjustment parameter. It adjusts the corrected quality value of the anomalous corpus. The backtracking propagation is performed along the reverse path of the feature mapping relationship to the boundary corpus, yielding the backtracking quality value contribution for this anomalous corpus. When multiple anomalous corpora exist in the associated corpus, the backtracking quality values ​​of all anomalous corpora are weighted and aggregated to obtain the final backtracking quality value. , expressed as ,in This represents the set of all data points marked as anomalous. Backtracking quality value. By inferring the quality level of boundary corpora from the perspective of anomalous corpora, a basis can be provided for the iterative correction of subsequent quality labeling values.

[0156] Backtrack quality value Compared with the original quality label value Perform a difference comparison and calculate the difference amount. .when Exceeding the preset correction threshold At that time, it was determined that the original quality label value had a deviation and needed to be iteratively corrected. The correction method was to... Updated to ,in To correct the step size parameter, the value range is: This is used to control the correction magnitude of each iteration, avoiding over-correction that could lead to oscillations. After completing the iterative correction of the quality label values, the updated values ​​are used... replace Substitute the values ​​back into the overall evaluation calculation process and recalculate the quality-utility ratio. The overall evaluation value is then updated to obtain the revised overall evaluation value. The above iterative process continues until the interval between two adjacent iterations. Below the preset correction threshold The process continues until the number of iterations reaches a preset upper limit, thus ensuring that the correction process converges.

[0157] After obtaining the revised comprehensive evaluation value Subsequently, this was used as the benchmark for quality control of the associated corpus. Corrected quality values ​​of various related corpora Filtering is performed. The filtering rule is: if the corrected quality value of a certain related corpus is... Not less than Compared with preset screening tolerance If the quality difference is within a certain range, the corpus passes quality control and is included in the final high-quality corpus; otherwise, it is marked as low-quality corpus and removed or manually reviewed. Preset screening tolerance. It can be flexibly configured according to the sensitivity of downstream tasks to corpus quality; for tasks with stringent quality requirements, it can be... Setting the value to a small value or even zero allows for more lenient requirements to be set for tasks with less stringent quality standards. Through this screening mechanism, low-quality data in the associated corpus is effectively filtered out while ensuring the accuracy of the screening criteria. This achieves an effective balance between the breadth of corpus coverage and the reliability of quality, completing the entire corpus quality control process.

[0158] A second aspect of this invention provides an automatic corpus acquisition and quality control system based on intelligent algorithms, comprising:

[0159] The strategy generation unit is used to obtain the attribute information of the corpus to be collected, construct the corpus state representation, generate the current collection strategy based on the corpus state representation, and invert the objective function of the current collection strategy to construct an adversarial collection strategy.

[0160] The blind spot extraction unit is used to execute the current acquisition strategy and the adversarial acquisition strategy respectively to obtain the target corpus set and the adversarial corpus set, and extract the corpus in the target corpus set that does not belong to the adversarial corpus set to form the blind spot corpus set;

[0161] The boundary selection unit is used to calculate the distance between each corpus in the blind zone corpus and the center point of the collected corpus, select the corpus with the largest distance as the boundary corpus and perform collection.

[0162] The association determination unit is used to construct the feature mapping relationship from the boundary corpus to the collected corpus set, and to determine the associated corpus set through the feature mapping relationship.

[0163] The propagation correction unit is used to obtain the quality label value of the boundary corpus, propagate the quality label value to each related corpus in the related corpus set to obtain the corrected quality value, obtain the utility measurement value of each related corpus in the related corpus set, and propagate the utility measurement value to the boundary corpus to obtain the predicted utility value.

[0164] The fusion control unit is used to fuse the quality label value and the predicted utility value to obtain a comprehensive evaluation value, and to perform corpus quality control based on the comprehensive evaluation value and the corrected quality value.

[0165] A third aspect of the present invention provides an electronic device, comprising:

[0166] processor;

[0167] Memory used to store processor-executable instructions;

[0168] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0169] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0170] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0171] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for automatic corpus acquisition and quality control based on intelligent algorithms, characterized in that, include: Obtain the attribute information of the corpus to be collected, construct the corpus state representation, generate the current collection strategy based on the corpus state representation, and invert the objective function of the current collection strategy to construct an adversarial collection strategy. The target corpus set and the adversarial corpus set are obtained by executing the current acquisition strategy and the adversarial acquisition strategy respectively. The corpus set that does not belong to the adversarial corpus set in the target corpus set is extracted to form the blind zone corpus set. Calculate the distance between each corpus in the blind zone corpus and the center point of the collected corpus, select the corpus with the largest distance as the boundary corpus and perform collection; Construct a feature mapping relationship from the boundary corpus to the collected corpus set, and determine the associated corpus set through the feature mapping relationship; Obtain the quality label value of the boundary corpus, propagate the quality label value to each related corpus in the related corpus set to obtain the corrected quality value, obtain the utility measurement value of each related corpus in the related corpus set, propagate the utility measurement value to the boundary corpus to obtain the predicted utility value; The quality label value and the predicted utility value are combined to obtain a comprehensive evaluation value, and the corpus quality control is performed based on the comprehensive evaluation value and the corrected quality value.

2. The method according to claim 1, characterized in that, The current acquisition strategy is generated based on the corpus state representation. The adversarial acquisition strategy is constructed by negating the objective function of the current acquisition strategy, including: Extract the feature values ​​of each dimension from the state representation of the corpus, use the feature values ​​of each dimension as independent variables to construct a coverage function with the optimization objective of maximizing corpus coverage, and solve the coverage function to obtain the policy vector as the current collection policy; Add a negative sign to the function value in the coverage function to obtain the reverse coverage function. Input the feature values ​​of each dimension in the corpus state representation into the reverse coverage function, solve the reverse coverage function to obtain the adversarial strategy vector as the adversarial acquisition strategy. The strategy vector is matched with the attribute information of each corpus to be collected to obtain the target matching value. The corpus to be collected is then filtered according to the target matching value to form the target corpus set. The adversarial strategy vector is matched with the attribute information of each corpus to be collected to obtain the adversarial matching value. The corpus to be collected is then filtered according to the adversarial matching value to form an adversarial corpus set. The target corpus is used to extract data that is not present in the adversarial corpus to form a blind zone corpus.

3. The method according to claim 1, characterized in that, The target corpus and adversarial corpus are obtained by executing the current acquisition strategy and the adversarial acquisition strategy respectively. The corpus that does not belong to the adversarial corpus in the target corpus is extracted to form the blind zone corpus, which includes: The target projection intensity is obtained by calculating the projection intensity of the attribute information of the corpus to be collected in the feature space defined by the policy vector. The corpus to be collected is selected to form a target corpus set and the target projection intensity corresponding to each corpus is recorded. The adversarial projection intensity is obtained by calculating the projection intensity of the attribute information of the corpus to be collected in the feature space defined by the adversarial strategy vector. The corpus to be collected is selected to form an adversarial corpus set and the adversarial projection intensity corresponding to each corpus is recorded. Extract the target projection intensity and adversarial projection intensity for each corpus in the target corpus set, and calculate the ratio of the target projection intensity to the adversarial projection intensity to obtain the projection intensity ratio; The projection intensity ratio is used as a weight to weight each corpus in the target corpus set, and the deviation of the weighted corpus in the feature space of the collected corpus set is calculated to obtain the boundary deviation. The target corpus is filtered based on the degree of boundary deviation, and the filtered corpus is determined to not belong to the adversarial corpus and constitutes the blind zone corpus.

4. The method according to claim 1, characterized in that, Calculate the distance between each corpus in the blind zone and the center point of the collected corpus, select the corpus with the largest distance as the boundary corpus, and perform the collection process including: Extract the attribute information of each corpus in the collected corpus set, calculate the distribution density of the attribute information, and obtain the center point of the collected corpus set by offsetting the geometric center in the feature space based on the distribution density; Extract the attribute information of each corpus in the blind zone corpus set, and calculate the spatial distance between the attribute information of each corpus in the blind zone corpus set and the center point of the collected corpus set to obtain the baseline distance; Extract the corresponding projection intensity ratio for each corpus in the blind zone corpus set, and use the projection intensity ratio as a direction vector to construct a directional path from the center point of the collected corpus set to each corpus in the blind zone corpus set; Detection points are set along the directional path. The number of corpora in the collected corpus set within the preset range around each detection point is counted. The number of corpora at each detection point is summed to obtain the number of corpora covered by the path. The distance between each corpus in the blind zone and the center point of the collected corpus is obtained by multiplying the inverse of the number of corpus points covered by the path and the baseline distance. Sort the data in the blind zone corpus by the distance between each data point and the center point of the collected data set, and select the data point with the largest distance as the boundary data point and perform the collection.

5. The method according to claim 1, characterized in that, Construct a feature mapping relationship from the boundary corpus to the collected corpus set, and determine the associated corpus set through the feature mapping relationship, including: Extract the attribute information of the boundary corpus, and vectorize the attribute information of the boundary corpus to obtain the feature vector of the boundary corpus; Extract the attribute information of each corpus in the collected corpus set, and vectorize the attribute information of each corpus in the collected corpus set to obtain the feature vector of the collected corpus; Construct a propagation path from the boundary corpus to each corpus in the collected corpus set, extract the number of corpus covered by the path in the collected corpus set region traversed by the propagation path, and set a feature decay function based on the number of corpus covered by the path. The feature vectors of the boundary corpus are attenuated by applying a feature attenuation function along the propagation path to obtain the attenuated feature vectors of each corpus position in the collected corpus set. The matching strength is obtained by matching the decay feature vector with the feature vector of the collected corpus. The matching strength and the propagation path length are then fused to obtain the feature mapping relationship from the boundary corpus to each corpus in the collected corpus set. The collected corpora are sorted according to their feature mapping relationships, and corpora whose feature mapping relationships exceed a preset mapping threshold are selected to form an associated corpus set.

6. The method according to claim 1, characterized in that, Obtain the quality label values ​​of the boundary corpus, propagate these quality label values ​​to each related corpus in the associated corpus set to obtain corrected quality values, obtain the utility measurement values ​​of each related corpus in the associated corpus set, and propagate these utility measurement values ​​to the boundary corpus to obtain predicted utility values, including: Extract the attribute information of the boundary corpus, calculate the spatial distance between the boundary corpus and the center point of the collected corpus set, and perform quality assessment on the attribute information of the boundary corpus based on the spatial distance to obtain the quality label value of the boundary corpus; Extract the feature mapping relationship between the boundary corpus and each associated corpus in the associated corpus set, construct a quality attenuation coefficient based on the feature mapping relationship, apply the quality attenuation coefficient to the quality label value of the boundary corpus to attenuate and propagate to obtain the propagation quality value of each associated corpus in the associated corpus set; Extract the initial quality value of each related corpus in the related corpus set, and perform a fusion operation between the propagation quality value and the initial quality value to obtain the corrected quality value of each related corpus in the related corpus set; Based on the corrected quality value, the number of corpora covered by each associated corpus in the collected corpus is statistically analyzed, and the number of corpora covered is used as the utility measurement value of each associated corpus. Extract the utility measurement value of each associated corpus and the corresponding feature mapping relationship. Backpropagate the utility measurement value to the boundary corpus along the reverse path of the feature mapping relationship. Accumulate the utility measurement values ​​backpropagated from each associated corpus in the associated corpus set to obtain the predicted utility value of the boundary corpus.

7. The method according to claim 1, characterized in that, A comprehensive evaluation value is obtained by fusing the quality label value and the predicted utility value. Corpus quality control is then performed based on the comprehensive evaluation value and the corrected quality value, including: The quality utility ratio is obtained by calculating the ratio of the quality-labeled value to the predicted utility value. The quality utility ratio is then used as a dynamic weight to weight and fuse the quality-labeled value and the predicted utility value to obtain the comprehensive evaluation value. Extract the corrected quality value of each related corpus from the related corpus set, and calculate the deviation between the corrected quality value and the comprehensive evaluation value; The associated corpus with a deviation exceeding the preset fluctuation threshold is marked as abnormal corpus. The feature mapping relationship between the abnormal corpus and the boundary corpus is extracted, and the path length of the feature mapping relationship is calculated to obtain the propagation distance. The propagation attenuation coefficient is constructed by applying the propagation distance to the corrected quality value of the abnormal corpus. Based on the propagation attenuation coefficient, the corrected quality value is propagated back to the boundary corpus along the reverse path of the feature mapping relationship to obtain the backtracked quality value. The quality value backtracking is compared with the quality label value. When the difference exceeds the preset correction threshold, the quality label value is iteratively corrected and the comprehensive evaluation value is recalculated to obtain the corrected comprehensive evaluation value. Based on the corrected comprehensive evaluation value, the corrected quality value of each associated corpus in the associated corpus set is used to screen and perform corpus quality control.

8. An automatic corpus acquisition and quality control system based on intelligent algorithms, used to implement the method as described in any one of claims 1-7, characterized in that, include: The strategy generation unit is used to obtain the attribute information of the corpus to be collected, construct the corpus state representation, generate the current collection strategy based on the corpus state representation, and invert the objective function of the current collection strategy to construct an adversarial collection strategy. The blind spot extraction unit is used to execute the current acquisition strategy and the adversarial acquisition strategy respectively to obtain the target corpus set and the adversarial corpus set, and extract the corpus in the target corpus set that does not belong to the adversarial corpus set to form the blind spot corpus set; The boundary selection unit is used to calculate the distance between each corpus in the blind zone corpus and the center point of the collected corpus, select the corpus with the largest distance as the boundary corpus and perform collection. The association determination unit is used to construct the feature mapping relationship from the boundary corpus to the collected corpus set, and to determine the associated corpus set through the feature mapping relationship. The propagation correction unit is used to obtain the quality label value of the boundary corpus, propagate the quality label value to each related corpus in the related corpus set to obtain the corrected quality value, obtain the utility measurement value of each related corpus in the related corpus set, and propagate the utility measurement value to the boundary corpus to obtain the predicted utility value. The fusion control unit is used to fuse the quality label value and the predicted utility value to obtain a comprehensive evaluation value, and to perform corpus quality control based on the comprehensive evaluation value and the corrected quality value.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images