A sample intelligent auxiliary labeling method for complex business scenarios

Abstract

This application provides a sample intelligent auxiliary annotation method for complex business scenarios, relating to the field of scenario sample annotation technology. The method includes: S1, pre-annotating samples to obtain initial annotations and acquiring a pre-trained model; S2, correcting the pre-trained model through parameter transfer to obtain a domain-adapted model suitable for the current scenario; S3, generating a second annotated sample set through knowledge transfer and iterative optimization using a teacher-student model; S4, expanding the second annotated sample set by extending the sample distribution through diverse data augmentation strategies to obtain third annotated data; and S5, constructing a generative adversarial network to generate fourth annotated data. This application is based on a multi-stage intelligent auxiliary annotation system combining sample pre-annotation, data distillation, data augmentation, and balanced sample generation. Through phased iteration and closed-loop optimization, it achieves intelligent and standardized data annotation.

Description

Technical Field

[0001] This application relates to the field of scene sample annotation technology, and in particular to an intelligent auxiliary annotation method for samples in complex business scenarios. Background Technology

[0002] With the deep integration of artificial intelligence technology into various industries, high-quality corpus annotation has become the core support for the research and application of intelligent models. Currently, corpus annotation is facing a series of challenges. First, traditional manual annotation methods suffer from problems such as low efficiency and inconsistent standards; Secondly, in complex business scenarios, the diversity and complexity of data also bring challenges. Due to the different types of data such as text, images, signals, and spatiotemporal data, the labeling of samples has become more difficult. Third, in the actual annotation process, too much attention is paid to the correctness of individual samples, while the distribution of the overall sample set is ignored, resulting in an imbalance in the distribution of sample difficulty, an imbalance in the distribution of categories, and inconsistent sample quality.

[0003] Therefore, how to balance efficiency and quality and explore AI-assisted annotation is a key issue that urgently needs to be addressed in the current development of corpus annotation technology. Summary of the Invention

[0004] In view of this, the purpose of this application is to address the problems of low efficiency and uneven distribution in corpus annotation in complex business scenarios, and to propose a multi-stage intelligent auxiliary annotation system based on sample pre-annotation, data distillation, data augmentation, and balanced sample generation. This system takes human-machine collaboration as its core mechanism, and through phased iteration and closed-loop optimization, it gradually realizes the intelligentization and standardization of data annotation, providing support for building high-quality samples in fields such as education and healthcare.

[0005] Firstly, embodiments of this application provide a sample intelligent auxiliary annotation method for complex business scenarios, including: S1. Pre-label the samples to obtain initial sample labels, and obtain a pre-trained model based on the initial sample labels. ; S2. Apply parameter transfer to the pre-trained model. Make corrections to obtain a domain adaptation model suitable for the current scenario. Domain Adaptation Model Used to combine knowledge and rule constraints to generate the first labeled data; S3. Through knowledge transfer and iterative optimization using the teacher-student model, feature extraction is performed on the first labeled data to generate a second labeled sample set; S4. Expand the sample distribution by using diversified data augmentation strategies to expand the second labeled sample set and obtain the third labeled data; S5. Construct a generative adversarial network (GAN), which is used to perform sample balancing on the third labeled data to generate fourth labeled data.

[0006] In one possible technical solution, the pre-trained model is modified through parameter transfer in step S2. Make corrections to obtain a domain adaptation model suitable for the current scenario. step, ; in, This represents the parameters obtained after adjustment on a business-labeled dataset adapted to the current scenario; This indicates a compound operation of parameters.

[0007] In one possible technical solution, the domain adaptation model in S2 This step is used to combine knowledge and rule constraints to generate the first labeled data. The business knowledge base and scenario rules adapted to the current scenario are formalized into a set of constraint functions. Domain-adaptive models The output is subjected to a consistency check to obtain the corresponding constraint score. ; ; in, It is a domain adaptation model For the original text The predicted output; It is a rule validation function used to determine... Does it satisfy rule j? This is an indicator function that returns either 0 or 1. Based on the constraint score The pre-trained samples are filtered to obtain the first labeled data.

[0008] In one possible technical solution, step S3, which involves extracting features from the first labeled data and generating a second labeled sample set through knowledge transfer and iterative optimization using a teacher-student model, includes: Construct a teacher model and a student model, wherein the teacher model is a domain-adapted model. ; The teacher model is used to process the first labeled sample set to obtain the teacher model output; Based on the manually corrected labels from the first labeled sample set and the output of the teacher model, the distillation loss function of the student model is calculated. ; Based on the distillation loss function The student model is trained to obtain the trained student model; The first labeled sample set is processed based on the trained student model to generate the second labeled sample set.

[0009] In one possible technical solution, the distillation loss function of the student model as follows: ; in, Cross-entropy loss; Let KL divergence be a metric. This is the output of the teacher model; This is the output of the student model; For the softmax function, For temperature parameters, This is the first balance coefficient; This is the second balance coefficient; Labels were manually corrected. This is the predicted output of the student model.

[0010] In one possible technical solution, the calculation of the distillation loss function of the student model is... Also includes: Calculate the prediction uncertainty of the student model for each sample in the first labeled sample set; Based on the aforementioned prediction uncertainty Determining the difficulty coefficient of each sample ; Based on the aforementioned difficulty coefficient The distillation loss function is weighted to obtain the weighted distillation loss function L. weighted ; Wherein, the prediction uncertainty , ; in The predicted probability of the student model in category c; In one possible technical solution, step S4 involves expanding the sample distribution of the second labeled sample set using diversified data augmentation strategies to obtain the third labeled data. Perform regular text enhancement operations on the text samples in the second labeled sample set to generate a first enhanced sample subset; Perform specialization enhancement operations on samples related to specific business scenarios in the second labeled sample set to generate a second enhanced sample subset; The conventional text enhancement operations include at least one of synonym replacement, sentence transformation, and context rearrangement; the specialized enhancement operations include at least one of scene information expansion and event perturbation generation.

[0011] In one possible technical solution, S5 constructs a generative adversarial network (GAN), which is used to perform sample equalization on the third labeled data to generate fourth labeled data, including: Construct a generative adversarial network, wherein the generative adversarial network includes a generator and a discriminator; The generator is used to generate candidate negative samples based on the input positive samples; The candidate negative samples are filtered based on feature difference constraints and semantic constraints to obtain an intermediate negative sample set; The intermediate negative sample set is quality-assessed, and samples that pass the assessment are added to the generated negative sample set.

[0012] In one possible technical solution, in the step of filtering the candidate negative samples based on feature difference constraints and semantic constraints to obtain an intermediate negative sample set, set up (x) is a semantic feature extraction function, requiring that the generated negative sample x′ and positive sample x satisfy the following: ; in This is the difference threshold.

[0013] Compared with the prior art, the embodiments of this application have the following advantages: By constructing an integrated intelligent assisted annotation process, a complete closed loop from pre-annotation to sample balancing is formed, solving the problems of low annotation efficiency and unbalanced sample distribution in complex business scenarios.

[0014] By combining multiple technologies such as pre-trained model transfer, knowledge distillation, data augmentation, and sample balancing, the entire annotation process is made intelligent, significantly reducing the cost of manual annotation.

[0015] The method is highly versatile and can be adapted to complex business scenarios in multiple vertical fields such as education and healthcare, thereby improving the portability and scalability of the annotation system.

[0016] Phased iterative optimization continuously improves annotation quality and model generalization ability, supporting the training of highly robust AI models.

[0017] Other features and advantages of this application will be set forth in the following description and will be apparent in part from the description or may be learned by practicing the application. The objectives and other advantages of this application are realized and obtained through the structures particularly pointed out in the description, claims and drawings.

[0018] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 This document provides a flowchart of a sample intelligent auxiliary annotation method for complex business scenarios, as illustrated in an embodiment of this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] To facilitate understanding of this embodiment, a detailed description of the intelligent auxiliary annotation method for complex business scenarios disclosed in this application will be provided first, such as... Figure 1 As shown; S1. For example, the open-source general-purpose large model Qwen-32B is used to pre-label the samples (original text data and original image data) to obtain the initial labels, and a pre-trained model is obtained based on the initial labels. ; S2. Apply parameter transfer to the pre-trained model. Make corrections to obtain a domain adaptation model suitable for the current scenario. Domain Adaptation Model Used to combine knowledge and rule constraints to generate the first labeled data; In sample pre-labeling, a base model is used to pre-label the data, and the results are manually reviewed and corrected to complete the labeling. The pre-labeling mechanism is the starting point of the entire intelligent assisted labeling system. Its main task is to achieve rapid labeling of large-scale data at a low cost, thereby providing high-quality input for subsequent manual correction, data synthesis and data augmentation.

[0023] S3. Through knowledge transfer and iterative optimization using the teacher-student model, feature extraction is performed on the first labeled data to generate a second labeled sample set; In data distillation, key features are extracted from existing labeled data through knowledge transfer and iterative optimization of the teacher-student model. At the same time, a hard example mining mechanism is introduced in the distillation stage to solve the problem that the data distribution is complex and dynamic in complex business scenarios. Relying solely on pre-trained models and manual correction is still insufficient to guarantee the robustness of the model in long-tail categories, hidden targets, or high-noise environments, thus generating high-quality synthetic data.

[0024] S4. Expand the sample distribution by using diversified data augmentation strategies to expand the second labeled sample set and obtain the third labeled data; In terms of data augmentation, we expand the sample distribution through diverse data augmentation strategies, such as character replacement, syntactic transformation, context rearrangement, and specialized augmentation for specific business scenarios. These techniques rewrite and transform existing labeled samples to further increase the number of samples and alleviate the problem of insufficient sample data in actual business scenarios.

[0025] S5. Construct a generative adversarial network (GAN), which is used to perform sample balancing on the third labeled data to generate fourth labeled data.

[0026] In terms of balanced sample generation, a generative adversarial network based on feature differences and semantic constraints is used to automatically generate negative samples. At the same time, combined with the generative model and quality control mechanism, a sufficient number of negative samples with controllable quality and reasonable distribution are systematically generated, providing high-quality samples with a balanced positive and negative distribution, i.e., fourth-label data, for actual business.

[0027] First, initial sample labeling is performed based on pre-labeled data. Pre-labeling is the starting point of the entire intelligent assisted labeling system. Its main task is to achieve rapid labeling of large-scale data at a low cost, thereby providing high-quality input for subsequent manual correction, data synthesis, and data augmentation. The core of this mechanism includes pre-trained model transfer and knowledge and rule constraints. As a possible design, in step S2, parameter transfer is used to modify the pre-trained model. Make corrections to obtain a domain adaptation model suitable for the current scenario. step, ; in This represents the parameters obtained after adjustment on a business-labeled dataset adapted to the current scenario; The parameter composition operation (custom operator) is essentially a parameter transfer calculation method from a pre-trained model to a domain-adapted model. Common specific calculations include parameter weighted fusion, gradient descent update, LoRA, etc. In this method, the low-rank parameters are obtained through LoRA / QLoRA.

[0028] Instead of using all model parameters, this method allows for the rapid acquisition of preliminary annotation results with a certain degree of usability even with limited samples.

[0029] In this embodiment, LoRA / QLoRA low-rank parameter fine-tuning is used to achieve the transfer from the pre-trained model to the domain-adapted model, which has a small number of parameters, fast training, and low computing cost.

[0030] In this embodiment, rapid domain adaptation can be achieved by fine-tuning a small number of low-rank parameters, and stable and usable initial annotation capabilities can still be obtained under small sample business data.

[0031] This embodiment retains the general capabilities of the pre-trained model while accurately injecting business features, thus balancing generalization and domain specialization.

[0032] In this embodiment, low-rank fine-tuning facilitates model iteration and updates, adapts to dynamic changes in business scenarios, and improves the flexibility of the annotation system.

[0033] One possible design is the domain adaptation model in S2. This step is used to combine knowledge and rule constraints to generate the first labeled data. To avoid semantic deviations or logical errors in the model within complex business scenarios, the business knowledge base and scenario rules adapted to the current scenario are formalized into a set of constraint functions. Domain-adaptive models The output is subjected to a consistency check to obtain the corresponding constraint score. ; ; in, It is a domain adaptation model For the original text The predicted output; It is a rule validation function used to determine... If the j-th rule is satisfied, return True / False. This is an indicator function that returns either 0 or 1; if A lower risk level indicates a conflict between the pre-labeling results and business knowledge. This sample will be marked as "high-risk" and prioritized for manual review. Through this mechanism, pre-labeling not only relies on model inference results but also maintains logical consistency with business common sense and task rules.

[0034] Based on the constraint score The pre-trained samples are filtered to obtain the first labeled data.

[0035] In this embodiment, a business knowledge base and rule constraints are introduced to perform consistency verification on the model annotation results, avoiding semantic deviations and logical errors, and improving annotation accuracy.

[0036] In this embodiment, the risk of labeling is quantified by constrained scores, high-risk samples are automatically identified and prioritized for manual review, thereby reducing the risk of erroneous labels flowing into subsequent stages.

[0037] In this embodiment, the rules can be flexibly configured through formalization to adapt to the professional specifications of different business scenarios, thereby enhancing the standardization and controllability of annotation.

[0038] In this embodiment, efficient verification is achieved through indicator functions, which are computationally lightweight, do not increase the system burden, and ensure the efficiency of the pre-labeling process.

[0039] After pre-annotation and manual correction, the first labeled data has been obtained. However, in complex business scenarios, the data distribution is complex and dynamic, and relying solely on pre-trained models and manual correction is still insufficient to guarantee the robustness of the model in long-tail categories, hidden targets, or high-noise environments. This application introduces a data distillation mechanism into the overall process to improve the generalization ability of the annotation system. As a possible design, step S3 involves extracting features from the first labeled data to generate the second labeled sample set through knowledge transfer and iterative optimization using a teacher-student model, including: Construct a teacher model and a student model, wherein the teacher model is a domain-adapted model. ; The teacher model is used to process the first labeled sample set to obtain the teacher model output; Based on the manually corrected labels from the first labeled sample set and the output of the teacher model, the distillation loss function of the student model is calculated. ; Based on the distillation loss function The student model is trained to obtain the trained student model; The first labeled sample set is processed based on the trained student model to generate the second labeled sample set.

[0040] In this embodiment, key features are extracted from high-quality labeled data by using a teacher-student model for knowledge transfer, generating a purer and more valuable second labeled sample set.

[0041] In this embodiment, by combining manually corrected labels with teacher model output, the student model learns both truth supervision and probability distribution knowledge simultaneously, resulting in stronger generalization ability.

[0042] In this embodiment, new labeled data can be continuously incorporated through an iterative optimization mechanism to continuously improve the feature extraction effect and adapt to the dynamic changes of complex scenarios.

[0043] In this embodiment, by making the student model lightweight, inference is faster and large-scale unlabeled data can be processed more efficiently, thereby improving the overall labeling productivity.

[0044] One possible design is the distillation loss function of the student model. as follows: ; in, Cross-entropy loss; Let KL divergence be a metric. This is the output of the teacher model; This is the output of the student model; For the softmax function, For temperature parameters, This is the first balance coefficient; This is the second balance coefficient; Labels were manually corrected. This is the predicted output of the student model.

[0045] In this embodiment, the cross-entropy loss and KL divergence are fused by distillation loss to balance manual label supervision and teacher model knowledge transfer, resulting in more stable training and better performance.

[0046] In this embodiment, by introducing temperature parameters and balance coefficients, the weight of the monitoring signal can be flexibly adjusted to adapt to different business scenarios and data distributions.

[0047] In this embodiment, the loss function is designed to be rigorous and interpretable, making it easy to reproduce and optimize, thereby improving the reliability and practicality of the technical solution.

[0048] In this embodiment, by simultaneously absorbing information from both hard and soft labels, the model's discrimination accuracy and robustness in complex scenarios are significantly improved.

[0049] In complex business scenarios, typical challenging samples often come from long-tail categories (such as rare equipment targets), complex backgrounds (such as camouflage scenarios), or extreme conditions (such as low light or interference environments). To enhance the model's robustness to these samples, a hard example mining mechanism will be introduced during the distillation stage.

[0050] One possible design is to calculate the distillation loss function of the student model. Also includes: Calculate the prediction uncertainty of the student model for each sample in the first labeled sample set; Based on the aforementioned prediction uncertainty Determining the difficulty coefficient of each sample ; Based on the aforementioned difficulty coefficient The distillation loss function is weighted to obtain the weighted distillation loss function L. weighted ; Wherein, the prediction uncertainty , ; in The predicted probability of the student model in class c; when When the value is high, the sample is labeled as a "hard sample" and given a higher weight in the loss function. ; ; in, The difficulty coefficient is used to optimize the model for difficult samples during training, thereby improving its discrimination ability in complex business scenarios.

[0051] Distillation Feature Iterative Optimization: Data distillation is not a single operation, but a continuous iterative process. New results are manually corrected and then re-enter the distillation cycle. The distillation iteration process automatically terminates when any of the following conditions are met: 1) Model performance convergence: After two consecutive iterations, the improvement of key indicators (precision, recall, F1 score) of the student model on the business test set is less than the preset threshold, indicating that the model has stabilized and further iterations will yield minimal benefits. 2) Pre-annotation accuracy meets standards: The pre-annotation results automatically generated by the model have been validated by rules and evaluated by sampling, and their accuracy meets the requirements of business applications, requiring no further optimization; 3) Sufficient learning of difficult samples: The average loss of the difficult samples selected based on uncertainty measurement continues to decrease and tends to stabilize. The model's ability to distinguish long-tail, hidden, and complex scene samples has met the system requirements. 4) Maximum number of iterations: To control computational resources and annotation costs, a maximum number of iterations is set, and the loop automatically terminates once the maximum is reached. Through these multi-dimensional termination markers, an optimal balance can be achieved between model performance, iteration efficiency, and engineering costs, ensuring a stable, reliable, and feasible distillation process.

[0052] In this embodiment, difficult sample mining is achieved based on prediction uncertainty. It automatically identifies difficult samples such as long-tailed samples, complex backgrounds, and high noise, thereby improving the model's attention to difficult samples.

[0053] In this embodiment, a difficulty coefficient-weighted distillation loss is used to reinforce learning on difficult samples, effectively solving the problems of poor long-tail category recognition and insufficient model generalization.

[0054] In this embodiment, uncertainty calculation is used to adapt to large-scale data scenarios.

[0055] This embodiment significantly improves the model's robustness in complex business scenarios such as concealed targets, extreme conditions, and rare categories.

[0056] In complex business domains, the limited quantity and source of business data often result in insufficient labeled samples. Therefore, the data augmentation mechanism proposed in this project aims to expand the text sample space through automated and diversified generation. One possible design is that in step S4, the sample distribution is expanded using diversified data augmentation strategies to extend the second labeled sample set, resulting in the third labeled data step. Perform regular text enhancement operations on the text samples in the second labeled sample set to generate a first enhanced sample subset; Perform specialization enhancement operations on samples related to specific business scenarios in the second labeled sample set to generate a second enhanced sample subset; The conventional text enhancement operations include at least one of synonym replacement, sentence transformation, and context rearrangement; the specialized enhancement operations include at least one of scene information expansion and event perturbation generation.

[0057] Common text augmentation strategies: In text data, the most common augmentation methods are synonym substitution, phrase expansion, and sentence transformation. Additionally, sentence transformation can be used for augmentation, such as converting active sentences to passive sentences. This augmentation method helps improve the model's robustness to different grammatical expressions.

[0058] Specialized enhancements for specific business scenarios: Unlike general natural language, business scenario documents often contain highly specialized terminology and formatted expressions. This project introduces domain-specific enhancements into its enhancement strategy, including: Scene information expansion: Add auxiliary descriptions to the original text, such as enhancing "ships enter the port" to "ships enter the port in the morning" to simulate scenarios under multiple conditions; Event perturbation generation: A slight perturbation is applied to the event matrix S to generate a new description. For example: ; in This approach represents minute variations in event elements (time, location, task), thereby generating similar but not identical samples. These enhancement methods both expand the number of samples and improve the model's coverage of task-specific representations.

[0059] This embodiment employs a dual-layer enhancement strategy of conventional and business-specific approaches to balance the robustness of general text with the coverage of domain-specific expressions.

[0060] In this embodiment, the sample diversity is expanded at low cost through conventional enhancements (synonym substitution, sentence transformation, etc.) to alleviate the problem of insufficient data.

[0061] In this embodiment, specialization enhancements (scene expansion, event perturbation) are used to accurately match the business scenario, generating highly relevant and high-value enhanced samples.

[0062] This embodiment effectively expands the sample space to improve the model's adaptability to business scenarios and its generalization performance.

[0063] In complex business scenarios, there is often an imbalance in sample distribution, such as sufficient data for common categories but scarce samples for rare categories, leading to model training bias towards high-frequency categories. This application constructs a systematic automatic negative sample generation method in the annotation and training system to achieve balanced and optimized sample distribution. One possible design is that S5 constructs a generative adversarial network (GAN), which is used to balance the third labeled data and generate the fourth labeled data, including: Construct a generative adversarial network, wherein the generative adversarial network includes a generator and a discriminator; The generator is used to generate candidate negative samples based on the input positive samples; The candidate negative samples are filtered based on feature difference constraints and semantic constraints to obtain an intermediate negative sample set; The intermediate negative sample set is quality-assessed, and samples that pass the assessment are added to the generated negative sample set.

[0064] Specifically, negative sample generation based on feature differences and semantic constraints in generative adversarial networks includes: A Generative Adversarial Network (GAN) framework is constructed, consisting of a generator and a discriminator. The generator aims to produce candidate negative samples that are semantically similar to positive samples but differ in key features. The discriminator aims to distinguish between real negative samples and generated negative samples. In this game-like process, the generator continuously optimizes its generation strategy, making its output data gradually approach the distribution of real negative samples.

[0065] However, relying solely on GANs may generate semantically ambiguous or even invalid samples; therefore, feature differences and semantic constraints must be introduced to improve the generation quality. Let... (x) is a semantic feature extraction function, requiring that the generated negative sample x′ and positive sample x satisfy the following: ; in The difference threshold, (x) Select a mature pre-trained model based on the sample type. Common choices are: for text samples, use pre-trained models in the NLP field, such as BERT and RoBERTa, which are used to extract the semantic embedding vectors of the text; for image samples, use pre-trained models in the CV field, such as ResNet, VGG, and ViT (Visual Transformer), which are used to extract the semantic-level feature maps of the image.

[0066] This avoids generating "pseudo-negative examples" that are too close to positive samples.

[0067] Quality assessment of generated samples: The generated samples must undergo rigorous quality assessment and screening before they can be included in the training set. The assessment includes three levels.

[0068] One is the consistency check, which calculates the semantic similarity with positive samples and removes samples that are too close to positive examples to avoid confusion between positive and negative samples.

[0069] Secondly, there is the distribution test, which uses statistical analysis methods to compare the distribution characteristics of the generated samples with those of the real negative samples to ensure that the generated data does not lead to an imbalance in the overall sample distribution.

[0070] Thirdly, task relevance testing is performed. Combining business knowledge base and expert experience, the samples are checked for logical rationality and application relevance to ensure they reflect actual business scenarios rather than meaningless sentence combinations. Through this multi-layered screening, the generated negative samples not only increase in quantity but also become more reliable in quality, making them directly usable for model training.

[0071] In this embodiment, negative samples are automatically generated based on GAN, which solves the problems of imbalanced positive and negative sample distribution and few rare class samples, and optimizes the overall sample distribution.

[0072] In this embodiment, negative samples are screened by combining feature differences and semantic constraints to avoid generating false negative examples and semantically invalid samples, thus ensuring the quality of generation.

[0073] In this embodiment, multi-layered quality assessment (consistency, distribution, task relevance) is used to ensure that the generated samples are compliant, usable, and relevant to business needs.

[0074] This embodiment reduces the cost of manual synthesis and screening by automating the balancing of the sample set, thereby improving the scalability and standardization of the annotation system.

[0075] In this embodiment, a semantic feature distance threshold is used as a constraint to ensure that negative samples and positive samples are significantly different, thus avoiding confusion between positive and negative samples.

[0076] In this embodiment, the constraints are clear and quantifiable, which facilitates implementation and optimization, and ensures that the generation of negative samples is stable and controllable.

[0077] In this embodiment, the business relevance of negative and positive samples is preserved while ensuring discriminability, thereby improving the model's discrimination boundary and classification performance.

[0078] This embodiment ensures the validity of negative samples from a semantic level, avoids meaningless generation, and improves the overall quality of the sample set.

[0079] The device provided in this application embodiment has the same implementation principle and technical effect as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the device embodiment can be referred to the corresponding content in the aforementioned method embodiment.

[0080] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0081] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0082] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0083] The foregoing has described specific embodiments of this application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0084] In the description of the embodiments of this application, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the embodiments of this application. In the embodiments of this application, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in the embodiments of this application, as well as the features of different embodiments or examples.

[0085] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of embodiments of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0086] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order according to the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0087] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0088] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present application should be included within the scope of protection of the present application.

Claims

1. A sample intelligent auxiliary annotation method for complex business scenarios, characterized in that, include: S1. Pre-label the samples to obtain initial sample labels, and obtain a pre-trained model based on the initial sample labels. ; S2. Apply parameter transfer to the pre-trained model. Make corrections to obtain a domain adaptation model suitable for the current scenario. The domain adaptation model Used to generate first labeled data based on knowledge and rule constraints; S3. Through knowledge transfer and iterative optimization using the teacher-student model, feature extraction is performed on the first labeled data to generate a second labeled sample set; S4. Expand the sample distribution by using diversified data augmentation strategies to obtain the third labeled data; S5. Construct a generative adversarial network (GAN), which is used to perform sample balancing on the third labeled data to generate fourth labeled data.

2. The method of claim 1, wherein, In S2, the pre-trained model is modified through parameter transfer. Make corrections to obtain a domain adaptation model suitable for the current scenario. step, ； in This represents the parameters obtained after adjustment on a business-labeled dataset adapted to the current scenario; This indicates a compound operation of parameters.

3. The method of claim 1, wherein, The domain adaptation model in S2 This step is used to combine knowledge and rule constraints to generate the first labeled data. The business knowledge base and scenario rules adapted to the current scenario are formalized into a set of constraint functions. Domain-adaptive models The output is subjected to a consistency check to obtain the corresponding constraint score. ; ； in, It is a domain adaptation model For the original text The predicted output; It is a rule validation function used to determine... Does it satisfy rule j? This is an indicator function that returns either 0 or 1. Based on the constraint score The pre-trained samples are filtered to obtain the first labeled data.

4. The method according to claim 2, characterized in that, In step S3, the second labeled sample set is generated by extracting features from the first labeled data through knowledge transfer and iterative optimization using a teacher-student model, including: Construct a teacher model and a student model, wherein the teacher model is a domain-adapted model. ; The teacher model is used to process the first labeled sample set to obtain the teacher model output; Based on the manually corrected labels from the first labeled sample set and the output of the teacher model, calculate the distillation loss function of the student model. ; Based on the distillation loss function The student model is trained to obtain the trained student model; The first labeled sample set is processed based on the trained student model to generate the second labeled sample set.

5. The method of claim 4, wherein, The distillation loss function of the student model as follows: ； in, Cross-entropy loss; Let KL divergence be a metric. This is the output of the teacher model; This is the output of the student model; For the softmax function, For temperature parameters, This is the first balance coefficient; This is the second balance coefficient; Labels were manually corrected. This is the predicted output of the student model.

6. The method of claim 4, wherein, The distillation loss function of the student model is calculated. Also includes: Calculate the prediction uncertainty of the student model for each sample in the first labeled sample set; Based on the aforementioned prediction uncertainty Determining the difficulty coefficient of each sample ; Based on the aforementioned difficulty coefficient The distillation loss function is weighted to obtain the weighted distillation loss function L. weighted ; Wherein, the prediction uncertainty , ; in This represents the predicted probability of the student model for category c.

7. The method according to claim 1, characterized in that, In step S4, the sample distribution is expanded using diverse data augmentation strategies to extend the second labeled sample set, resulting in the third labeled data step. Perform regular text enhancement operations on the text samples in the second labeled sample set to generate a first enhanced sample subset; Perform specialization enhancement operations on samples related to specific business scenarios in the second labeled sample set to generate a second enhanced sample subset; The conventional text enhancement operations include at least one of synonym replacement, sentence transformation, and context rearrangement; the specialized enhancement operations include at least one of scene information expansion and event perturbation generation.

8. The method according to claim 1, characterized in that, S5 constructs a generative adversarial network (GAN), which is used to perform sample equalization on the third labeled data and generate fourth labeled data, including: Construct a generative adversarial network, wherein the generative adversarial network includes a generator and a discriminator; The generator is used to generate candidate negative samples based on the input positive samples; The candidate negative samples are filtered based on feature difference constraints and semantic constraints to obtain an intermediate negative sample set; The intermediate negative sample set is quality-assessed, and samples that pass the assessment are added to the generated negative sample set.

9. The method according to claim 8, characterized in that, In the step of filtering the candidate negative samples based on feature difference constraints and semantic constraints to obtain an intermediate negative sample set, set up (x) is a semantic feature extraction function, requiring that the generated negative sample x′ and positive sample x satisfy the following: ； in, This is the difference threshold.

Images