Active learning text classification method, device and equipment for information redundancy suppression

Abstract

This application relates to an active learning text classification method, apparatus, and device for information redundancy suppression. The method uses the CLIPCMIC zero-shot model as a baseline, constructs a high-coverage initial annotation set through a representative initialization strategy, introduces an active learning mechanism, and employs a two-stage query strategy that balances informativeness and representativeness. It filters high-uncertainty samples based on prediction entropy to ensure sample informativeness. Subsequently, diversity constraints are introduced into the candidate sample set to reduce duplicate sampling within semantic clusters, thereby improving sample coverage and the quality of the supervision signal. This method combines cross-modal semantic completion capabilities with an informative-representative collaborative active query mechanism, achieving more stable and significant performance gains with lower annotation costs, and providing an effective technical path for further learnability of cross-modal zero-shot benchmarks.

Description

Technical Field

[0001] This invention belongs to the field of active learning text classification technology, and relates to an active learning text classification method, apparatus and device for information redundancy suppression. Background Technology

[0002] With the rapid development of the big data era, a large amount of unstructured text data is accumulating in various industries. Text classification, as a core technology of natural language processing, is widely used in multiple application scenarios such as sentiment analysis and medical diagnosis. However, in scenarios with labeled data, active learning methods have gradually become an important research direction in order to improve the utilization rate of labeled data and reduce labeling costs. However, text data usually exhibits a clear semantic cluster distribution, and traditional active learning strategies are prone to repeatedly selecting samples in local semantic regions, leading to information redundancy and wasted resources.

[0003] Text classification, a fundamental task in natural language processing, has wide applications in text retrieval, sentiment analysis, news text classification, and medical diagnosis. Traditional supervised learning methods heavily rely on large-scale, high-quality labeled data. However, in real-world applications, obtaining high-quality labeled data is limited, costly, and inefficient. This is particularly pronounced in professional fields such as law and medicine, or in emerging tasks, where insufficient and outdated labeled resources become critical bottlenecks hindering model deployment and continuous optimization. To alleviate the challenges posed by data scarcity, researchers have proposed zero-shot learning paradigms, providing a feasible path for text classification in scenarios without labeled data.

[0004] In recent years, with the development of multimodal pre-trained models, cross-modal zero-shot text classification methods based on CLIP models have made significant progress. These methods reconstruct the text classification task into a text-image matching approach, leveraging the contrastive learning capability of the CLIP model in a unified semantic space to achieve text classification without target domain labeled data. Existing research has proposed the CLIPTEXT framework, LABCLIP method, and CLIPMulti framework, which, through gradual optimization and improvement, have verified the effectiveness of visual image information in assisting zero-shot text classification tasks, but still have certain limitations. To address this, reference 1 (CN120144764A, Zero-shot Text Classification Method and Device Based on Cross-modal Information Completion) proposes a CLIPCMIC framework. Through a cross-modal information completion strategy, it effectively alleviates key problems such as label semantic bias and missing feature alignment information, surpassing existing zero-shot benchmark methods on multiple benchmark datasets and providing a superior zero-shot benchmark model for text classification tasks.

[0005] While zero-shot baseline models can achieve text classification in unlabeled data scenarios, there are still performance limitations. When the category semantics are fine-grained, the sample representations are diverse, and the distribution structure is complex, relying solely on static label semantics and prompt templates often fails to fully cover the true semantic patterns, resulting in room for further improvement in classification performance and stability. Furthermore, in practical applications, a small amount of labeled data is often available for model training to improve model performance. Therefore, how to efficiently utilize this labeled data to further improve model performance with the lowest possible labeling cost has become a core problem that urgently needs to be solved. Active learning, as an efficient sample selection strategy, iteratively selects the most informative samples for labeling, providing an effective solution for balancing labeling costs and model performance. Its core advantage lies in alleviating the need for large amounts of labeled data in traditional learning and reducing labeling costs.

[0006] However, existing active learning strategies often face a key problem in practice: information redundancy. In large-scale unlabeled data, similar samples often appear in clusters. If sorting and batch selection are based solely on a single indicator such as uncertainty, it is easy to repeatedly sample within local high-density semantic clusters, causing a high degree of overlap in the supervision signals of the same round of labeled samples, resulting in duplicate labeling and diminishing marginal returns. As shown in Table 1, taking text from the Yahoo! Answers dataset as an example.

[0007] Table 1 Sample Content Examples

[0008] Text 1 and Text 2 are both labeled "Business & Finance" and are almost semantically equivalent, both pointing to the same need: "resume writing." The differences lie only in capitalization, expression, and subtle wording (such as "do" vs. "can," and whether "good" is included). These nearly duplicated samples are extremely close in the representation space, making them prone to being selected simultaneously in uncertainty-driven queries, thus reducing the effective utilization of the annotation budget. This redundancy is prevalent in many types of public datasets, especially when the data's semantic structure exhibits clear clustering characteristics. Summary of the Invention

[0009] To address the challenges of efficiently utilizing a small amount of labeled data to further improve model classification performance on top of a zero-shot benchmark model, and the common problem of information redundancy in active learning, this invention proposes an active learning text classification method, apparatus, and device aimed at suppressing information redundancy. This method can achieve stable performance improvement with a low labeling ratio and approach or even reach the performance of full training on some datasets, significantly improving the efficiency of labeled data utilization.

[0010] To achieve the above objectives, the embodiments of the present invention adopt the following technical solutions: On the one hand, an active learning text classification method oriented towards information redundancy suppression is provided, including the following steps: Step 1: Obtain the unlabeled text set, extract cross-modal semantic features for each text using the pre-trained CLIPCMIC zero-shot benchmark framework, and construct the initial semantic feature space; Step 2: In the initialization phase, a greedy k-center selection strategy is adopted. Based on the distance between samples in the initial semantic feature space, the sample farthest from the current selected set is iteratively selected to construct the core set as the labeled sample pool. Step 3: Train the initial text classification model using the labeled sample pool; Step 4: A two-stage hybrid query strategy is adopted. First, a candidate set with high uncertainty is screened from the unlabeled sample pool based on the prediction entropy. Then, a diversity constraint is introduced to select samples with high information content from the candidate set to build a query pool. After manually labeling the samples in the query pool, the labeled sample pool and the pseudo-labeled sample pool are updated. The updated labeled sample pool is used to complete the iterative optimization of the text classification model parameters. The two-stage hybrid query is repeated until the preset number of iterations is reached to obtain the trained text classification model. Step 5: Analyze the text set to be classified using the trained text classification model to obtain the text classification results.

[0011] On the other hand, an active learning text classification device for information redundancy suppression is also provided, comprising: The initial semantic feature space construction module is used to obtain an unlabeled text set, extract cross-modal semantic features for each text using the pre-trained CLIPCMIC zero-shot benchmark framework, and construct the initial semantic feature space. The initialization module is used in the initialization phase. It adopts a greedy k-center selection strategy and iteratively selects the sample that is farthest away from the current selected set to construct the core set as the labeled sample pool based on the distance between samples in the initial semantic feature space. The initial training module is used to train the initial text classification model using the labeled sample pool. The two-stage hybrid query module employs a two-stage hybrid query strategy. First, it filters a candidate set with high uncertainty from the unlabeled sample pool based on the prediction entropy. Then, it introduces diversity constraints to select samples with high information content from the candidate set to construct a query pool. After manually labeling the samples in the query pool, it updates the labeled sample pool and the pseudo-labeled sample pool. The updated labeled sample pool is then used to iteratively optimize the parameters of the text classification model. The two-stage hybrid query is repeated until the preset number of iterations is reached to obtain the trained text classification model. The text classification module is used to analyze the set of texts to be classified using a trained text classification model, and obtain the text classification results.

[0012] On the other hand, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described active learning text classification method for information redundancy suppression.

[0013] One of the above technical solutions has the following advantages and beneficial effects: The aforementioned active learning text classification method, apparatus, and device for information redundancy suppression use the CLIPCMIC zero-shot model as a baseline. It constructs a high-coverage initial annotation set through a representative initialization strategy, introduces an active learning mechanism, and employs a two-stage query strategy that balances informativeness and representativeness. High-uncertainty samples are filtered based on prediction entropy to ensure sample informativeness. Subsequently, diversity constraints are introduced into the candidate sample set to reduce duplicate sampling within semantic clusters, thereby improving sample coverage and the quality of the supervision signal. This method combines cross-modal semantic completion capabilities with an informative-representative collaborative active query mechanism, achieving more stable and significant performance gains with lower annotation costs, and providing an effective technical path for further learnability of cross-modal zero-shot benchmarks. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the conventional technology, the drawings used in the description of the embodiments or the conventional technology will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a flowchart illustrating an active learning text classification method for information redundancy suppression in one embodiment. Figure 2 This is a schematic diagram of a pool-based active learning scenario in one embodiment; Figure 3 This is a schematic diagram of the overall framework of a text classification model in one embodiment; Figure 4 This is a schematic diagram illustrating the comparative experimental results of various models on the Yahoo! Answers dataset in one embodiment. Figure 4 (a) is a diagram showing the comparison of accuracy of various models on the Yahoo! Answers dataset. Figure 4 (b) is a comparison diagram of Macro_F1 scores for each model on the Yahoo! Answers dataset; Figure 5 This is a schematic diagram illustrating the comparative experimental results of various models on the Trec dataset in one embodiment. Figure 5 (a) shows a comparison of the accuracy of various models on the Trec dataset. Figure 5 (b) is a diagram showing the comparison of Macro_F1 scores for each model on the Trec dataset; Figure 6 This is a schematic diagram illustrating the comparative experimental results of various models on the AG's News dataset in one embodiment. Figure 6 (a) is a diagram showing the comparison of accuracy of various models on the AG's News dataset. Figure 6 (b) is a comparison diagram of the Macro_F1 scores of various models on the AG's News dataset; Figure 7 This is a schematic diagram illustrating the comparative experimental results of various models on the Subj dataset in one embodiment, where... Figure 7 (a) is a diagram showing the comparison of accuracy of various models on the Subj dataset. Figure 7 (b) is a diagram showing the comparison of Macro_F1 scores for each model on the Subj dataset; Figure 8 This is a schematic diagram illustrating the comparative experimental results of various models on the Snips dataset in one embodiment, where... Figure 8 (a) shows a comparison of the accuracy of various models on the Snips dataset. Figure 8 (b) is a diagram showing the comparison of Macro_F1 scores for each model on the Snips dataset; Figure 9 This is a schematic diagram illustrating the comparative experimental results of various models on the Atis dataset in one embodiment, where... Figure 9 (a) is a diagram showing the comparison of accuracy of various models on the Atis dataset. Figure 9 (b) is a comparison diagram of Macro_F1 scores for each model on the Atis dataset; Figure 10 This is a schematic diagram illustrating the experimental results of the Yahoo! Answers dataset in one embodiment; Figure 11 This is a schematic diagram of the experimental results for the Treec dataset in one embodiment; Figure 12 This is a schematic diagram of the experimental results of the AG's News dataset in one embodiment; Figure 13 This is a schematic diagram of the experimental results for the Subj dataset in one embodiment; Figure 14 This is a schematic diagram of the experimental results of the Snips dataset in one embodiment; Figure 15This is a schematic diagram of the experimental results for the Atis dataset in one embodiment; Figure 16 This is a schematic diagram of ablation experiment results for a Yahoo! Answers dataset with the same amount of annotations in one example. Figure 17 This is a schematic diagram of ablation experiment results for a Tree dataset with the same amount of annotations in one example. Figure 18 This is a schematic diagram of ablation experiment results for an example dataset of AG's News, with the same amount of annotations. Figure 19 This is a schematic diagram of ablation experiment results for the Subj dataset in one example, under the same amount of annotation. Figure 20 This is a schematic diagram of ablation experiment results for an example Snips dataset with the same amount of annotation. Figure 21 This is a schematic diagram of ablation experiment results for the Atis dataset in one example, under the same amount of annotation. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0017] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.

[0018] It should be noted that, in this document, the reference to "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The presentation of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. Those skilled in the art will understand that the embodiments described herein can be combined with other embodiments. The term "and / or" as used herein refers to any combination of one or more of the associated listed items, and all possible combinations, including such combinations.

[0019] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0020] The active learning text classification method for information redundancy suppression proposed in this application adopts a text classification model abbreviated as: CMIC-RAL model.

[0021] In-depth analysis of the Yahoo! Answers, Tree, AG's News, Subj, Snips, and Atis datasets reveals that most datasets suffer from information redundancy. To address this, this application, building upon the CLIPCMIC zero-shot benchmark framework, introduces an active learning mechanism to design an active learning text classification method aimed at suppressing information redundancy. This method employs a two-stage active learning query strategy, the CMIC-RAL model, which balances informativeness and representativeness. It aims to efficiently utilize limited labeled data to further improve the performance of zero-shot benchmark models, while maximizing the utilization efficiency of labeled resources and mitigating information redundancy. The specific process is as follows: To enhance the stability and representativeness of the initial labeled set, this method employs a greedy k-center selection strategy to construct a core set during the initialization phase. It iteratively selects samples furthest from the already selected set to enhance feature space coverage, providing a more representative seed set for subsequent iterations. In the iterative query phase: The first phase is based on the idea of ​​uncertainty sampling. It uses prediction entropy to sort unlabeled samples and select high uncertainty samples to form a candidate set, so that the query budget is prioritized to the region most sensitive to the decision boundary. The second phase imposes diversity constraints on the candidate set and performs distributed selection based on CLIP text feature representation to suppress cluster concentration and duplicate queries, reduce information redundancy and improve coverage of different semantic patterns, thereby achieving higher performance gains with limited labeling budget.

[0022] In one embodiment, such as Figure 1 As shown, an active learning text classification method for information redundancy suppression is provided, which may include the following processing steps 1 to 5: Step 1: Obtain the unlabeled text set, extract cross-modal semantic features for each text using the pre-trained CLIPCMIC zero-shot benchmark framework, and construct the initial semantic feature space.

[0023] Specifically, the specific components of the pre-trained CLIPCMIC zero-shot benchmark framework are as described in reference 1 (publication number CN120144764A Zero-shot text classification method and device based on cross-modal information completion), and will not be repeated here.

[0024] Step 2: In the initialization phase, a greedy k-center selection strategy is adopted. Based on the distance between samples in the initial semantic feature space, the sample farthest from the current selected set is iteratively selected to construct the core set as the labeled sample pool.

[0025] Step 3: Train the initial text classification model using the labeled sample pool.

[0026] Specifically, the initial text classification model is a zero-shot classification model based on a pre-trained CLIP dual encoder (parameter frozen) and learnable parameters initialized from the category semantic image, which has not yet been trained using any labeled samples.

[0027] Step 4: A two-stage hybrid query strategy is adopted. First, a candidate set with high uncertainty is selected from the unlabeled sample pool based on the prediction entropy. Then, a diversity constraint is introduced to select samples with high information content from the candidate set to build a query pool. After manually labeling the samples in the query pool, the labeled sample pool and the pseudo-labeled sample pool are updated. The updated labeled sample pool is used to complete the iterative optimization of the text classification model parameters. The two-stage hybrid query is repeated until the preset number of iterations is reached to obtain the trained text classification model.

[0028] Specifically, in supervised learning tasks where annotation resources are extremely limited, how to obtain training samples that are both representative and contain the best amount of information with the lowest manual cost has always been one of the core research problems in the field of machine learning. Active Learning (AL) provides an effective way to alleviate this core bottleneck by allowing the model to actively select samples to be labeled during the training process. Its core idea is that if the model can prioritize the samples that are most valuable for improving its own performance, it can achieve a high efficiency improvement in model prediction performance while significantly reducing the size of the annotation.

[0029] From an application paradigm perspective, active learning mainly includes three typical scenarios: member query synthesis, stream-based selective sampling, and pool-based sampling. Among them, pool-based active learning methods (such as...) Figure 2 (As shown) This paradigm, due to its better alignment with real-world data acquisition patterns, has become the most widely used setting in natural language processing tasks such as text classification. Under this paradigm, the model iteratively filters samples to be labeled from a large pool of unlabeled samples using a specific query strategy, expands this pool into a small labeled training set, and then iteratively trains based on the updated labeled set to continuously improve model performance. This method also adopts this approach to construct its framework.

[0030] While active learning has demonstrated significant advantages in reducing annotation costs, existing methods in text classification tasks still largely rely on a single modality or representation space for sample value evaluation. Furthermore, active learning strategies often face several key challenges in practice: firstly, information content measurement typically depends solely on the uncertainty of text representation or model output, easily leading to duplicate queries and information redundancy within semantic clusters, resulting in diminishing marginal returns for supervisory signals; secondly, when category semantics are fine-grained, diverse in expression, and complex in distribution structure, the discriminative criteria formed solely by static label semantics and prompt templates often fail to adequately cover real semantic patterns, limiting performance and stability. With the development of cross-modal pre-trained models, visual-language alignment provides an "external prior" that can be used for semantic enhancement, particularly in zero-shot text classification, where it has proven effective in improving semantic separability and robustness. However, how to explicitly introduce cross-modal alignment and semantic completion mechanisms during the active learning query process, while ensuring information content while suppressing redundancy and improving representativeness and coverage, still lacks systematic strategy design and sufficient validation.

[0031] Therefore, this application proposes a two-stage active learning query strategy. The two-stage active learning query strategy mainly includes three parts: the establishment and initialization phase of the data pool, and the two-stage active learning sample query strategy. The most crucial phase is the iterative optimization phase of active learning, i.e., the two-stage sample query phase. This phase adopts a cyclic iterative architecture, achieving sustainable performance gains for the model within a limited labeling budget through multiple rounds of "sample selection → data annotation → model update" interaction. The text classification framework CMIC-RAL, which integrates cross-modal information completion and active learning, is described below. Figure 3 As shown, this paper aims to efficiently utilize labeled information under limited labeling budget constraints, alleviate information redundancy, and achieve progressive performance improvement from zero-sample baseline to supervised model, providing an effective solution for text classification in data-scarce scenarios. The framework mainly consists of three stages: (1) Zero-sample baseline construction: First, a zero-sample text classifier is built based on CLIPCMIC's cross-modal information completion mechanism to obtain the initial prediction ability under unlabeled conditions, which serves as the baseline for subsequent active learning optimization. (2) Two-stage active learning query strategy: Then, the active learning process is executed on a large-scale unlabeled sample pool. To address the problem that existing strategies are prone to repeated sampling within semantic clusters, leading to redundancy of supervised signals and diminishing marginal returns, a two-stage query mechanism that balances "information and representativeness" is designed. (3) Model inference and evaluation: Finally, under budget constraints, the supervised model obtained through iteration is used for inference and system evaluation to verify the effectiveness, stability, and generalization ability of the proposed framework in low-labeling-cost scenarios.

[0032] The two-stage active learning query strategy effectively alleviates the two core problems of high annotation cost and information redundancy through the collaborative mechanism of "high uncertainty screening and diversity perception selection", improves sample query efficiency and annotation resource utilization, and verifies its effectiveness and stability on six public datasets through ablation and comparison experiments.

[0033] Specifically, each iteration follows a closed-loop process of "model training → sample query → annotation augmentation → performance evaluation": First, the model is trained based on the initialized labeled sample pool, and a candidate set is constructed in the unlabeled pool to estimate sample uncertainty. Second, a two-stage hybrid query strategy is adopted: first, a candidate set with high uncertainty is selected based on prediction entropy; then, diversity constraints are introduced to select samples with high information content to construct a query pool, thereby reducing redundant annotations and enhancing sample coverage. Next, samples in the query pool are manually labeled and incrementally added to the labeled sample pool. Finally, the updated labeled sample pool is used to iteratively optimize the model parameters, and the model performance is evaluated on an independent test set. The entire process is repeated until the preset number of iterations is reached. Through this iterative mechanism, the model can gradually focus on samples with greater discriminative value and a wider representation distribution. Under budget constraints, it achieves priority acquisition of high-information samples and explicit suppression of redundant queries, thereby effectively improving learning efficiency and performance stability.

[0034] This method is based on the zero-sample baseline CLIPCMIC and constructs a sample selection strategy that balances informativeness and diversity. By combining uncertainty measurement and diversity constraints, it effectively suppresses redundant information while ensuring the value of sample information.

[0035] Step 5: Analyze the text set to be classified using the trained text classification model to obtain the text classification results.

[0036] Specifically, a well-trained text classification model refers to a well-trained CLIP dual encoder.

[0037] The aforementioned active learning text classification method for information redundancy suppression uses the CLIPCMIC zero-shot model as a baseline. It constructs a high-coverage initial annotation set through a representative initialization strategy, introduces an active learning mechanism, and employs a two-stage query strategy that balances informativeness and representativeness. High-uncertainty samples are filtered based on prediction entropy to ensure sample informativeness. Subsequently, diversity constraints are introduced into the candidate sample set to reduce duplicate sampling within semantic clusters, thereby improving sample coverage and the quality of the supervision signal. This method combines cross-modal semantic completion capabilities with an informative-representative collaborative active query mechanism, achieving more stable and significant performance gains with lower annotation costs, and providing an effective technical path for further learnable cross-modal zero-shot benchmarks.

[0038] In one embodiment, step 1 includes: setting the category set of the target dataset; for each label category, constructing the corresponding semantically enhanced category image; using the image encoder and text encoder in the pre-trained CLIP model to extract the embedding representations of the label category and the text sample respectively; defining the matching score between the text and the label category by scaling cosine similarity based on the extracted embedding representations of the label category and the text sample; for a single-label text classification task, the zero-sample prediction result is determined by the maximum matching score. (1) in, For zero-sample prediction results, The set of label categories for the target dataset. For tag categories; For text and label categories Match score, For text samples; The initial semantic feature space is constructed based on the zero-sample prediction results.

[0039] In one embodiment, the matching score between the text and the tag category is: (2) in, This represents the temperature scaling factor, used to adjust the smoothness of the similarity distribution. , This represents a learnable temperature scalar parameter in the CLIP model, used to scale the cosine similarity between the image and text embeddings. An embedded representation of the tag category. An embedded representation of a text sample.

[0040] Specifically, based on the CLIPCMIC framework, to quantify the performance gain of active learning by introducing labeled samples and to ensure a consistent reference for comparisons between different query strategies, this method first constructs a zero-shot text classification benchmark under the condition that no task-labeled samples are used for training. The target dataset's category set is set as follows: For each tag category Construct corresponding semantically enhanced category images This information is then used as cross-modal completion information for category semantics to enhance category representation. Based on the cross-modal alignment capability of the pre-trained CLIP model, the embedding representations of label categories and text samples are respectively generated by the image encoder. With text encoder Extraction, i.e. and Next, the matching score between the text and each label category is defined by scaling cosine similarity, as shown in formula (2). For single-label text classification tasks, the zero-shot prediction result is determined by the maximum matching score, as shown in formula (1).

[0041] Based on this, using the test set Perform zero-shot inference and evaluation to obtain zero-shot benchmark performance. It is used to measure the performance improvement of the model after active learning iterations.

[0042] In one embodiment, step 2 includes: an initialization phase in which a corresponding number of samples are drawn from the initial unlabeled sample pool to construct an initialization candidate set; the semantic representation of the samples is extracted using the CLIP text encoder, and the extracted features are L2 normalized to obtain normalized features; based on the normalized features, a greedy k-center selection strategy is used to gradually construct a representative core set, and the core set is mapped to the original training set to simulate manual annotation; when the initialization budget is reached, the obtained core set is used as the labeled sample pool, and the unlabeled sample pool is updated.

[0043] In one embodiment, a representative core set is gradually constructed based on normalized features using a greedy k-center selection strategy. The core set is then mapped to the original training set to simulate manual annotation, including: Initialize the core set to an empty set, in the... The step to select the sample with the largest nearest distance to the current set is: (3) in, In the first The step selects the sample with the largest nearest distance to the current set. , Let be the cosine distance between the candidate sample and the selected sample in the normalized feature space, which is defined as the complement of the cosine similarity between feature vectors. These are the candidate samples and selected samples in the candidate pool, respectively. The first i The first sample and the first j Normalized features of each sample; To initialize the candidate set, , For the initial unlabeled sample pool, For the first The core set of steps; In the The sample with the largest nearest distance to the current set is selected in the next step. The core set of the step is updated and mapped to the original training set to simulate manual annotation.

[0044] In one embodiment, filtering a candidate set with high uncertainty from the unlabeled sample pool based on the prediction entropy includes: selecting from the currently unlabeled sample pool... The process involves sampling and selecting a candidate pool; calculating uncertainty within the candidate pool; specifically, this includes: for any... The model outputs the matching scores between the text and each label category, and obtains the predicted probability distribution: (4) in, To predict the probability distribution, For text and label categories c Match score, .

[0045] Based on the predicted probability distribution, the entropy value used to measure uncertainty is calculated as follows: (5) in, Entropy is the value used to measure uncertainty.

[0046] From the candidate pool Select the one with the highest entropy value A high-uncertainty candidate set is constructed from 1 sample: (6) in, For candidate set, The size of the candidate set in each iteration. To check the budget, The size of the candidate set.

[0047] In one embodiment, a diversity constraint is introduced to select highly informative samples from the candidate set to construct a query pool. After manually labeling the samples in the query pool, the labeled sample pool and the pseudo-labeled sample pool are updated. The updated labeled sample pool is then used to iteratively optimize the parameters of the text classification model. This includes: when the size of the candidate set is not less than twice the query budget, a cluster-based diversity selection is adopted; specifically, representative samples are selected from each cluster to construct a query pool of a certain size. query pool When the candidate set size is less than twice the query budget or the clustering is unstable, a distance-based greedy diversity selection strategy is adopted to gradually expand the query pool based on the farthest point criterion. ; Query pool All samples are manually labeled, and after labeling, they are input into the labeled sample pool for data update. The updated labeled sample pool is then used to update the parameters of the text classification model. Each active learning iteration takes place on an independent test set. Evaluation model performance And combined with zero-sample benchmark performance As a reference, the final evaluation result is obtained after completing all iterations.

[0048] Specifically, the overall design primarily adopts a pooled active learning paradigm, performing incremental model adaptation within a limited annotation budget. The training set is then defined. test set At the start of active learning, the training data is first divided into data pools: the initial unlabeled pool is defined as... The pool was initially labeled as an empty set. To ensure the traceability of subsequent sampling and training, a non-overlapping constraint was maintained in each iteration of the experiment, i.e. as well as .

[0049] Under budget constraints, random initialization may result in insufficient semantic coverage of the initial labeled set, leading to high noise in the early supervision signals obtained by the model, instability in the training process, and affecting the subsequent query strategy's ability to distinguish "informative" samples. Therefore, this method employs a representative initialization strategy (greedy k-center) to construct the seed labeled set: first, starting from the initial unlabeled sample pool... An initial candidate set is constructed by extracting a corresponding number of samples from the sample. The upper limit of the scale is set to This is used to control computational overhead, i.e. Then, the CLIP text encoder is used to extract the semantic representation of the samples. And perform L2 normalization on the features. This simplifies cosine similarity to vector dot product. Based on normalized features, a greedy k-center selection strategy is used to gradually construct a representative core set. .initialization In the step Select the sample with the largest nearest distance to the current set. As shown in formula (3).

[0050] renew Finally, the dataset is mapped back to the original training set to simulate manual annotation. This process continues until the initial budget is reached. At that time, the final core set is used as the initial annotation set, that is And update the unlabeled pool. ( Indicates from Remove from The initialization phase aims to improve the coverage and representativeness of the initial annotation set in the feature space, thereby providing a more stable starting point for subsequent information sampling.

[0051] After initialization is complete, it will enter the next step. Round of active learning iteration Each iteration first uses the currently labeled sample pool as a basis. Supervised updates were performed on the model, and subsequently, an unlabeled sample pool was used. Select the most valuable samples to construct the query pool And update the data pool according to the rules shown in formula (6): (7) Phase 1: Candidate Sample Screening Phase Based on Uncertainty. Let the query budget for each round be... (that is, each category) One sample, i.e. To reduce the computational cost of evaluating large-scale unlabeled sample pools sample by sample, we first start from... Sampling and candidate pool selection Its upper limit is ,Right now Next, the uncertainty is calculated on the candidate pool: for any The model outputs the matching scores between the text and each label category, and obtains the predicted probability distribution as shown in formula (4). Then, the uncertainty (information) is measured by the prediction entropy as shown in formula (5).

[0052] Then from Select the one with the highest entropy value Each sample constructs a high-uncertainty sub-candidate set as shown in Equation (6).

[0053] Phase Two: Diversity-Aware Batch Sample Selection Phase. To alleviate the problem of query information redundancy caused by sampling based solely on uncertainty, in A diversity constraint strategy is introduced. This is achieved by using CLIP text features. Measure sample similarity and from The selected size is query pool This ensures that the selected samples cover as many different semantic regions as possible while maintaining high uncertainty, thus ensuring that the selected samples are both informative and diverse in the feature space. In specific implementation, according to... Scale-adaptive diversity selection strategy: when the candidate size is large ( When using clustering-based diversity selection, representative samples are selected from each cluster to construct a sample of size 1. query pool When the candidate size is small ( When clustering is unstable, a distance-based greedy diversity selection strategy is adopted, gradually expanding based on the farthest point criterion. This reduces information redundancy. Next, all samples in the label pool are manually labeled. After labeling, they are input into the labeled sample pool for data updates, and the updated labeled sample pool is used to update the base classifier model clip. Each active learning iteration is performed on an independent test set. Evaluation model performance And combined with zero-sample benchmark performance For reference, the final evaluation result will be output after all iterations are completed.

[0054] In one embodiment, each active learning loop iteration takes place on an independent test set. Evaluation model performance And combined with zero-sample benchmark performance As a reference, the final evaluation result is obtained after completing all iterations, including: for any test sample The text classification model outputs the matching score between the text and each label category, and the category with the highest score is taken as the prediction result; based on the prediction results, the accuracy is calculated, and the accuracy is used as the evaluation metric; the accuracy is: (8) Where acc is the accuracy rate. For the predicted results, x For any test sample, For real labels, This is represented as an indicator function, which takes a value of 1 when the predicted category matches the true label, and a value of 0 otherwise. Zero-sample benchmark performance For reference, define the first The performance improvement after the first iteration is the [number]th. The model's evaluation results on the test set and its zero-shot baseline performance after all rounds of iterations. The difference; the minimum amount of annotation required to exceed the zero-sample baseline is defined as: (9) in, Minimum annotation size, For the labeled sample pool, For the first The evaluation results of the model on the test set after the first round of iterations.

[0055] Specifically, after completing the active learning iteration, the model is in the labeled sample pool. Training is performed to obtain an adapted classification model. To verify the effectiveness and practicality of the text classification model under limited annotation budget conditions, an independent test set was used. A unified reasoning and evaluation process is performed, and the results are compared with those of a zero-sample baseline. Let the category set be... For any test sample The model outputs the matching scores (logits) between the text and each label category, resulting in a vector. In this method, the category with the highest score is used as the prediction result. (10) To quantify the classification performance of the model, the accuracy, as shown in formula (8), is mainly used as the evaluation index.

[0056] Furthermore, to better characterize the gains brought by active learning, this method uses zero-shot benchmark performance. For reference, define the first The performance improvement after round iteration is as follows: (11) in Indicates the first The evaluation results of the model on the test set after the round of iterations. In addition, in order to characterize the sample efficiency of the annotation budget, the minimum number of annotations required to exceed the zero-sample baseline is defined as shown in Equation (9). Equation (9) means the set of all labeled samples that satisfy the condition that "the model performance exceeds the zero-sample baseline performance".

[0057] This metric can intuitively reflect the minimum supervision cost required for different query strategies to achieve effective performance improvement under a limited annotation budget, thus providing a reference for strategy selection and deployment in resource-constrained scenarios.

[0058] In a confirmatory embodiment, multiple sets of comparative and ablation experiments were designed to verify the effectiveness of the proposed strategy. The comparative experiments selected five classic active learning baseline methods and compared their performance on six public datasets. The experimental results show that the text classification module (CMIC-RAL model) proposed in this application outperforms the five baseline methods on most datasets. The ablation experiments verified the necessity of the two stages by removing the first and second stages of the two-stage query strategy. The results show that the two-stage strategy complements each other and is indispensable. Uncertainty filtering provides "informative" assurance for the query, while diversity selection effectively suppresses redundancy and enhances coverage; their synergy is key to achieving efficient learning.

[0059] (1) Experimental dataset As shown in Table 2, this embodiment evaluates the proposed method on six publicly available text classification datasets, including Yahoo! Answers, Trec, AG's News, Subj, Snips, and Atis datasets. These datasets differ significantly in size and category granularity: the number of training samples ranges from 81 (Subj dataset) to 120,000 (AG's News), and the number of categories covers 2-15, which can well reflect text classification scenarios of different complexities.

[0060] To further analyze the potential information redundancy and distribution characteristics in the unlabeled sample pool, this embodiment statistically analyzed the number of similar sample pairs and class imbalance in each dataset. Table 2 shows that the Trec, Snips, and Atis datasets have a high number of similar sample pairs (1909, 2055, and 1689 respectively), indicating that these datasets are more likely to form semantic cluster structures. When relying solely on uncertain ranking for sample queries, duplicate sampling may occur within local semantic clusters, resulting in redundant labeling. Simultaneously, the Atis dataset also exhibits a significant class imbalance, further increasing the complexity of query strategy design. In contrast, the AG's News dataset has a more balanced class distribution, while the Subj dataset has almost no similar sample pairs, serving as a low-redundancy control scenario.

[0061] The above data characteristics illustrate the necessity of considering both informativeness and diversity in the active learning process. Therefore, this embodiment designs a "two-stage strategy" in the iterative query stage, namely, firstly screening high-information samples based on uncertainty, and then reducing redundant sampling through diversity constraints, so as to improve the efficiency of annotation utilization and enhance the stability of model training.

[0062] Table 2. Relevant information about the dataset

[0063] (2) Experimental baseline method The CMIC-RAL model is compared with the following benchmark methods: 1) Random: This method selects samples from unlabeled data in a uniformly random manner for labeling without utilizing any model predictions or data feature information. Because it does not introduce any heuristic or information-driven selection criteria, it is often used as a baseline method in active learning to measure the effectiveness of more complex strategies in terms of labeling efficiency and performance improvement.

[0064] 2) Cvirs: By combining unified uncertainty with inconsistency in the label set, a general and effective active learning strategy is achieved.

[0065] 3) Cbmal: A two-stage active learning strategy that improves learning performance by evaluating the information content and removing redundancy from sample-label pairs respectively, thereby reducing query and computation costs.

[0066] 4) Entropy: An active learning query strategy that takes advantage of the highest uncertainty.

[0067] 5) Mcvie: A batch pattern active learning method for text classification. It calculates the basic information score of example-label pairs by combining prediction uncertainty and class vector inconsistency, and uses the Euclidean distance of text feature vectors to iteratively select a batch of sample label pairs with rich information and low redundancy for labeling.

[0068] (3) Experimental parameter settings In the experiments, the clip model was used as the base classifier, which consisted of a text encoder and an image encoder. The maximum context of the input text was fixed at 77. In the zero-shot benchmark stage, the model directly performed inference based on cross-modal matching scores to obtain zero-shot benchmark performance. During the active learning phase, the model training uses the Adam optimizer with a learning rate of [missing information]. The default number of training epochs is 5, and the batch size is 128. The maximum number of active learning epochs is 20. When the unlabeled sample pool is large enough, the candidate pool... The maximum size of the dataset was set to 7000, and the maximum size of the candidate set during the initialization phase was also set to 7000. All experiments were run on NVIDIA RTX 3080 Ti GPUs with a fixed random seed of 10 to ensure reproducibility of the results.

[0069] (4) Experimental results and analysis like Figure 4 As shown, this embodiment selects five typical active learning methods—Random, Cvirs, Mcvie, Entropy, and Cbmal—as comparative baselines, and introduces the Macro-F1 metric to supplement the evaluation of class-level performance balance. To ensure experimental fairness, all methods are compared under the same annotation budget and consistent experimental settings, with the same query size and number of iterations per round. CMIC-RAL is the method proposed in this application.

[0070] From an overall iterative trend perspective, on the Yahoo! Answers, Snips, AG's News, and Atis datasets, this method shows a more stable and significant improvement in both Acc and Macro-F1 metrics, outperforming the five baseline methods overall. This indicates that on datasets with complex semantic structures and significant semantic clustering or redundancy risks in sample distribution, the two-stage query strategy proposed in this application can effectively reduce duplicate sampling while ensuring sample informativeness, thereby improving the utilization efficiency of labeled data and bringing more stable performance improvements.

[0071] On the Tree dataset, with increasing iterations, our method gradually surpasses other baseline methods in the Acc metric, indicating that this strategy effectively improves the overall classification accuracy. However, in the Macro-F1 metric, the random strategy shows an advantage in some stages. This may be because the Tree dataset has a relatively small number of categories, and the text expressions have certain templated characteristics, resulting in significant differences in separability and intra-class clustering between different categories. In this case, relying solely on uncertainty for sample selection may concentrate the query process on a few categories with ambiguous decision boundaries or dense sample sizes, thus improving overall accuracy but insufficient coverage of some smaller categories, thereby limiting the improvement in Macro-F1. In contrast, the Random strategy has stronger randomness in category sampling, and therefore may show a stage-wise advantage in class balance.

[0072] On the Subj dataset, due to the small training set size and the presence of only two classes, the overall learning curve exhibits relatively large fluctuations. In such small-scale, low-class-complexity scenarios, the differences between different active learning strategies are often difficult to fully reflect, and model performance is more susceptible to the influence of initial samples and random factors. Therefore, except for the Cvirs method, which has weak comparability on this dataset, the performance gap between the other methods and our proposed method is generally small, with the baseline method even slightly outperforming in some epochs. This phenomenon is common in small-scale data tasks and does not necessarily indicate that the proposed method is ineffective.

[0073] Overall, the comparative experimental results show that on most datasets with complex semantic structures or potential redundancy risks, the proposed method achieves better or more stable performance on both the Acc and Macro-F1 metrics, verifying the effectiveness and applicability of the proposed two-stage active learning strategy under limited annotation budgets.

[0074] As shown in Table 3, this embodiment statistically analyzes the average runtime cost (based on the average time taken per complete experiment) of the five baseline methods and the proposed two-stage method on six datasets. Overall, the Random strategy has the lowest time cost, averaging only 4 minutes and 17 seconds per experiment. This is mainly because random sampling does not require model prediction, uncertainty calculation, or sample similarity measurement during the query phase, resulting in minimal additional computational overhead. The overall time is primarily consumed in the model training and testing phases.

[0075] Table 3. Average time cost of each method across the six datasets.

[0076] In non-random strategies, the two-stage method proposed in this application has a lower time cost, with an average running time of approximately 5 minutes and 30 seconds, which is roughly on the same order of magnitude as the Entropy strategy. The query overhead of the Entropy method mainly comes from performing a batch forward inference on unlabeled samples and calculating the prediction entropy value, without involving pairwise comparisons between samples, thus resulting in high computational efficiency. Although this method introduces a two-stage mechanism of "candidate screening and diversity constraints" in the query phase, the additional computation is mainly focused on the uncertainty assessment of candidate samples and the lightweight sample selection process, so the overall running time remains at a low level.

[0077] In comparison, Cvirs and Mcvie had average runtimes of 7 minutes and 13 seconds and 7 minutes and 30 seconds, respectively, which is about 2 minutes longer than our method. This is mainly because both methods require additional similarity or inconsistency calculations between candidate samples and labeled samples during the query phase, thus increasing the overall computational cost.

[0078] Furthermore, Cbmal has the highest time cost, averaging 46 minutes and 6 seconds per experiment, significantly longer than other methods. This is mainly because the method requires extracting feature representations of candidate and labeled samples during the query phase and performing numerous pairwise cosine similarity calculations to assess diversity. As the number of labeled samples increases, the complexity of this calculation process continuously rises, leading to a significant increase in overall runtime.

[0079] In summary, while ensuring improved classification performance, the two-stage method proposed in this application still maintains good computational efficiency. Its running time is significantly lower than that of the Cbmal method, which relies on large-scale similarity calculations, and is comparable to that of the lightweight uncertainty strategy Entropy, indicating that the method has good engineering feasibility in practical applications.

[0080] (5) Performance change analysis during active learning like Figures 10 to 15As shown in the figure, this embodiment illustrates the performance variation curves of six datasets during the active learning iteration process. The number of iteration rounds is set to 30 for Yahoo! Answers, Trec, and AG's News, 20 for Snips and Atis, and 9 for Subj. "Sample" represents the number of samples selected and labeled in each category per round, i.e., the category-level query budget. The blue horizontal line in the figure represents the zero-sample baseline performance based on CLIPCMIC, and the red horizontal line represents the upper limit of performance obtained using supervised training with the full training data.

[0081] Overall, the performance of models on all datasets shows a continuous improvement as the number of labeled samples gradually increases, and stabilizes above the zero-sample baseline after reaching a certain labeled scale. Among them, Trec, AG's News, Snips, and Atis show the most significant improvements, requiring only a few labeled samples to surpass the zero-sample baseline. Subsequent iterations further enhance performance with minimal fluctuations, demonstrating good learning efficiency and stability. In contrast, Yahoo! Answers requires a relatively larger number of labeled samples to surpass the zero-sample baseline, but it remains within an acceptable range overall.

[0082] For the Subj dataset, due to the small training set size, model performance fluctuates significantly when only one sample is selected for each class. As the number of samples increases, model performance gradually stabilizes. It should be noted that, due to dataset size limitations, a relatively high percentage of samples is required for proportional statistics, but overall, performance still shows a steady improvement trend with increasing labeled data.

[0083] Further analysis of the quantitative statistics in Table 4 leads to a clearer conclusion: on most datasets, the models require only a very small amount of labeled data to surpass the zero-sample baseline. For example, Trec, AG's News, and Atis require only 0.19%, 0.02%, and 0.61% of training data, respectively, to achieve performance superiority, with Atis achieving an 8.9% improvement over the zero-sample baseline when surpassing it. While Yahoo! Answers, Subj, and Snips require relatively higher percentages of labeled data, these percentages are still generally kept below 10%.

[0084] Furthermore, on some datasets, the models achieve or approach the performance of fully supervised training with a relatively small percentage of labeled data. For example, Subj, Snips, and Atis achieve or exceed the full-training results using only 9.88%, 5.91%, and 15.33% of the training data, respectively; AG's News achieves 88.5% performance with 0.39% of the training data, close to 90.9% with full training; Yahoo! Answers achieves 63.9% with 35.62% of the data, only 0.1% different from the full-training result. For the Trec dataset, the model achieves 72.4% performance with a labeled percentage of 10.44%, which, while still slightly lower than the optimal result of full training, is a competitive performance achieved while significantly reducing labeling costs.

[0085] Table 4 shows the accuracy values ​​for the first time exceeding the zero-shot baseline performance during active learning, the highest accuracy value, and the corresponding proportions of used data and complete data.

[0086] In summary, Figures 4 to 9 Consistent with the results in Table 4, this method can achieve high annotation utilization efficiency under a limited annotation budget, obtain stable performance improvement with less annotation data, thereby effectively reducing the cost of manual annotation and alleviating the problem of scarce annotation data.

[0087] (6) Ablation test like Figures 16 to 21 As shown, in this embodiment, the number of active learning iterations is uniformly set to 20 on six datasets, and ablation experiments are conducted under the same annotation budget to ensure the comparability of different strategies. The ablation settings include: (1) removing the uncertainty candidate screening in stage one (only retaining the subsequent batch selection mechanism); (2) removing the diversity-aware batch selection in stage two (only querying based on uncertainty); (3) the complete two-stage CMIC-RAL.

[0088] The iteration curves show that in the first 10 rounds or so, the performance of the two-stage strategy was not significant on the Yahoo! Answers, Trec, AG's News, and Snips datasets. This may be because there were very few labeled samples available in the early stages, the model had not yet formed a stable discrimination boundary, and both sample representation and uncertainty estimation contained significant noise, making it difficult to fully realize the advantages of the query strategy (this can be considered a "cold start" phase). However, as iterations progressed and labeled samples gradually accumulated, the differences between the three strategies began to widen significantly: the complete two-stage CMIC-RAL showed more consistent and stable performance gains on most datasets, gradually outperforming the variations of the two removal stages.

[0089] Further analysis of the maximum accuracy statistics in Table 5 reveals that, except for Yahoo! Answers, the highest accuracy of the complete two-stage strategy is higher than that of the two ablation strategies on all other datasets. For Yahoo! Answers, although the peak accuracy of the two-stage strategy is 0.2% lower than the optimal ablation variant, the overall iterative trend shows that as the number of iterations increases, the performance improvement of the two ablation strategies slows down or even declines, while the two-stage strategy maintains a more stable increase and higher later-stage performance, indicating better robustness and stability over long-term iterations.

[0090] Table 5 Comparison of relevant results from ablation experiments

[0091] In summary, the ablation experiments validated the effectiveness of the two-stage design from both overall trend and statistical results: the uncertainty screening in stage one provided "informative" assurance for the query process, prioritizing the allocation of budget to sample regions more sensitive to decision boundaries; the diversity selection in stage two effectively suppressed duplicate sampling within semantic clusters, enhanced sample coverage, and reduced information redundancy. The synergistic effect of these two approaches is key to achieving efficient and stable learning under a limited annotation budget.

[0092] It should be understood that, although the above Figure 1 The steps are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise explicitly stated in this document, there is no strict order in which these steps are executed; they can be performed in other orders. Furthermore, the above... Figure 1 At least some of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0093] In one embodiment, an active learning text classification device for information redundancy suppression is also provided, comprising: The initial semantic feature space construction module is used to obtain an unlabeled text set, extract cross-modal semantic features for each text using the pre-trained CLIPCMIC zero-shot benchmark framework, and construct the initial semantic feature space. The initialization module is used in the initialization phase. It adopts a greedy k-center selection strategy and iteratively selects the sample that is farthest away from the current selected set to construct the core set as the labeled sample pool based on the distance between samples in the initial semantic feature space. The initial training module is used to train the initial text classification model using the labeled sample pool. The two-stage hybrid query module employs a two-stage hybrid query strategy. First, it filters a candidate set with high uncertainty from the unlabeled sample pool based on the prediction entropy. Then, it introduces diversity constraints to select samples with high information content from the candidate set to construct a query pool. After manually labeling the samples in the query pool, it updates the labeled sample pool and the pseudo-labeled sample pool. The updated labeled sample pool is then used to iteratively optimize the parameters of the text classification model. The two-stage hybrid query is repeated until the preset number of iterations is reached to obtain the trained text classification model. The text classification module is used to analyze the set of texts to be classified using a trained text classification model, and obtain the text classification results.

[0094] In one embodiment, the initial semantic feature space construction module is also used to set the category set of the target dataset; for each label category, construct the corresponding semantically enhanced category image; use the image encoder and text encoder in the pre-trained CLIP model to extract the embedding representation of the label category and the text sample respectively, and define the matching score between the text and the label category by scaling cosine similarity based on the extracted embedding representation of the label category and the text sample; for the single-label text classification task, the zero-sample prediction result is determined by the maximum matching score as shown in formula (1); construct the initial semantic feature space based on the zero-sample prediction result.

[0095] In one embodiment, the matching score between text and label category in the initial semantic feature space construction module is shown in formula (2).

[0096] In one embodiment, the initialization module is further configured to: extract a corresponding number of samples from the initial unlabeled sample pool to construct an initialization candidate set; extract the semantic representation of the samples using the CLIP text encoder and perform L2 normalization on the extracted features to obtain normalized features; based on the normalized features, adopt a greedy k-center selection strategy to gradually construct a representative core set, and map the core set to the original training set to simulate manual annotation; when the initialization budget is reached, use the obtained core set as the labeled sample pool and update the unlabeled sample pool.

[0097] In one embodiment, the initialization module is further configured to initialize the core set to an empty set, in the... The step selects the sample with the largest nearest distance to the current set as shown in formula (3); in the 1st step... The sample with the largest nearest distance to the current set is selected in the next step. The core set of the step is updated and mapped to the original training set to simulate manual annotation.

[0098] In one embodiment, the two-stage hybrid query module is also used to retrieve data from the currently unlabeled sample pool. The process involves sampling and selecting a candidate pool; calculating uncertainty within the candidate pool; specifically, this includes: for any... The matching scores of each label category in the model are calculated, and the predicted probability distribution is obtained as shown in formula (4). Based on the predicted probability distribution, the entropy value used to measure the uncertainty is calculated using formula (5). From the candidate pool Select the one with the highest entropy value A high-uncertainty candidate set is constructed from the samples as shown in Equation (6).

[0099] In one embodiment, the two-stage hybrid query module is further configured to employ cluster-based diversity selection when the candidate set size is not less than twice the query budget; specifically, this includes selecting representative samples from each cluster to construct a sample set of a size of [missing information]. query pool When the candidate set size is less than twice the query budget or the clustering is unstable, a distance-based greedy diversity selection strategy is adopted to gradually expand the query pool based on the farthest point criterion. ; Query pool All samples are manually labeled, and after labeling, they are input into the labeled sample pool for data update. The updated labeled sample pool is then used to update the parameters of the text classification model. Each active learning iteration takes place on an independent test set. Evaluation model performance And combined with zero-sample benchmark performance As a reference, the final evaluation result is obtained after completing all iterations.

[0100] In one embodiment, the two-stage hybrid query module is also used for any test sample The text classification model outputs the matching score between the text and each label category, and takes the category with the highest score as the prediction result; based on the prediction result, the accuracy is calculated, and the accuracy shown in formula (8) is used as the evaluation index; the zero-sample benchmark performance is used as the evaluation index. For reference, define the first The performance improvement after the first iteration is the [number]th. The model's evaluation results on the test set and its zero-shot baseline performance after all rounds of iterations. The difference; the minimum amount of annotation required to exceed the zero sample baseline is defined as shown in Equation (9).

[0101] It is understood that for a detailed explanation of the active learning text classification device for information redundancy suppression, please refer to the corresponding explanations of the various embodiments of the active learning text classification method for information redundancy suppression described above, and will not be repeated here. Each module in the above-described active learning text classification device for information redundancy suppression can be implemented entirely or partially through software, hardware, or a combination thereof. Each module can be embedded in hardware or independently of a device with data processing capabilities, or stored in software in the memory of the aforementioned device, so that the processor can call and execute the operations corresponding to each module. The aforementioned device can be, but is not limited to, various types of data processing computer devices already existing in the art.

[0102] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0103] It is understood that, in addition to the memory and processor mentioned above, the computer equipment described above also includes other hardware and software components not listed in this specification. The specific components can be determined according to the model of the image processing computer in different application scenarios, and will not be listed and described in detail in this specification.

[0104] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0105] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of protection of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and all such modifications and improvements fall within the scope of protection of this application.

Claims

1. An active learning text classification method for information redundancy suppression, characterized in that, Including the following steps: Step 1: Obtain the unlabeled text set, extract cross-modal semantic features for each text using the pre-trained CLIPCMIC zero-shot benchmark framework, and construct the initial semantic feature space; Step 2: In the initialization phase, a greedy k-center selection strategy is adopted. Based on the distance between samples in the initial semantic feature space, the sample farthest from the current selected set is iteratively selected to construct the core set as the labeled sample pool. Step 3: Train an initial text classification model using the labeled sample pool; Step 4: A two-stage hybrid query strategy is adopted. First, a candidate set with high uncertainty is screened from the unlabeled sample pool based on the prediction entropy. Then, a diversity constraint is introduced to select samples with high information content from the candidate set to build a query pool. After manually labeling the samples in the query pool, the labeled sample pool and the pseudo-labeled sample pool are updated. The updated labeled sample pool is used to complete the iterative optimization of the text classification model parameters. The two-stage hybrid query is repeated until the preset number of iterations is reached to obtain the trained text classification model. Step 5: Analyze the text set to be classified using the trained text classification model to obtain the text classification results.

2. The active learning text classification method for information redundancy suppression according to claim 1, characterized in that, Step 1 includes: Set the category set for the target dataset; For each label category, construct the corresponding semantically enhanced category image; The image encoder and text encoder in the pre-trained CLIP model are used to extract the embedding representations of label categories and text samples, respectively. Based on the embedded representations of extracted label categories and text samples, the matching score between text and label categories is defined by scaling cosine similarity. For single-label text classification tasks, the zero-sample prediction result is determined by the maximum matching score as follows: in, For zero-sample prediction results, The set of label categories for the target dataset. For tag categories; For text and label categories Match score, For text samples; The initial semantic feature space is constructed based on the zero-sample prediction results.

3. The active learning text classification method for information redundancy suppression according to claim 2, characterized in that, The matching score between the text and the tag category is: in, This represents the temperature scaling factor, used to adjust the smoothness of the similarity distribution. , This represents a learnable temperature scalar parameter in the CLIP model, used to scale the cosine similarity between the image and text embeddings. An embedded representation of the tag category. An embedded representation of a text sample.

4. The active learning text classification method for information redundancy suppression according to claim 1, characterized in that, Step 2 includes: During the initialization phase, a corresponding number of samples are drawn from the initial unlabeled sample pool to construct an initialization candidate set; The CLIP text encoder is used to extract the semantic representation of the samples, and the extracted features are L2 normalized to obtain normalized features. Based on the normalized features, a greedy k-center selection strategy is used to gradually construct a representative core set. The core set is then mapped to the original training set to simulate manual annotation. When the initial budget is reached, the obtained core set is used as the labeled sample pool, and the unlabeled sample pool is updated.

5. The active learning text classification method for information redundancy suppression according to claim 4, characterized in that, Based on the normalized features, a greedy k-center selection strategy is used to progressively construct a representative core set. The core set is then mapped to the original training set to simulate manual annotation, including: Initialize the core set to an empty set, at the... The step to select the sample with the largest nearest distance to the current set is: in, In the first The step selects the sample with the largest nearest distance to the current set. , Let be the cosine distance between the candidate sample and the selected sample in the normalized feature space, which is defined as the complement of the cosine similarity between feature vectors. These are the candidate samples and selected samples in the candidate pool, respectively. The first i The first sample and the first j Normalized features of each sample; To initialize the candidate set, , For the initial unlabeled sample pool, For the first The core set of steps; In the The sample with the largest nearest distance to the current set is selected in the next step. The core set of the step is updated and mapped to the original training set to simulate manual annotation.

6. The active learning text classification method for information redundancy suppression according to claim 1, characterized in that, Based on the prediction entropy, a candidate set with high uncertainty is selected from the unlabeled sample pool, including: Candidate pools are selected by sampling from the current unlabeled sample pool; Calculating uncertainty on the candidate pool; specifically including: for any The model outputs the matching scores between the text and each label category, and obtains the predicted probability distribution: in, To predict the probability distribution, For text and label categories c Match score, ; Based on the predicted probability distribution, the entropy value used to measure the uncertainty is calculated as follows: in, Entropy is the value used to measure uncertainty. From the candidate pool Select the one with the highest entropy value A high-uncertainty candidate set is constructed from 1 sample: in, For candidate set, The size of the candidate set in each iteration. To check the budget, The size of the candidate set.

7. The active learning text classification method for information redundancy suppression according to claim 1, characterized in that, A diversity constraint is introduced to select samples with high information content from the candidate set to construct a query pool. After manually labeling the samples in the query pool, the labeled sample pool and the pseudo-labeled sample pool are updated. The updated labeled sample pool is then used to iteratively optimize the parameters of the text classification model, including: When the candidate set size is at least twice the query budget, a cluster-based diversity selection is employed; specifically, this includes selecting representative samples from each cluster to construct a sample set of size... query pool ;in To check the budget; When the candidate set size is less than twice the query budget or the clustering is unstable, a distance-based greedy diversity selection strategy is adopted to gradually expand the query pool based on the farthest point criterion. ; query pool All samples were manually labeled, and after labeling, they were input into the labeled sample pool for data update. The updated labeled sample pool was then used to update the parameters of the text classification model. Each iteration of the active learning loop takes place on an independent test set. Evaluation model performance And combined with zero-sample benchmark performance As a reference, the final evaluation result is obtained after completing all iterations.

8. The active learning text classification method for information redundancy suppression according to claim 7, characterized in that, Each iteration of the active learning loop takes place on an independent test set. Evaluation model performance And combined with zero-sample benchmark performance As a reference, the final evaluation result obtained after completing all iterations includes: For any test sample The text classification model outputs the matching score between the text and each label category, and takes the category with the highest score as the prediction result; Based on the prediction results, the accuracy rate is calculated and used as an evaluation metric; the accuracy rate is: Where acc is the accuracy rate. For the predicted results, x For any test sample, For real labels, This is represented as an indicator function, which takes a value of 1 when the predicted category matches the true label, and a value of 0 otherwise. Zero-sample benchmark performance For reference, define the first The performance improvement after the first iteration is the [number]th. The model's evaluation results on the test set and its zero-shot baseline performance after all rounds of iterations. The difference is defined as the minimum amount of annotation required to exceed the zero-sample baseline: in, Minimum annotation size, For the labeled sample pool, For the first The evaluation results of the model on the test set after the first round of iterations.

9. An active learning text classification device for suppressing information redundancy, characterized in that, include: The initial semantic feature space construction module is used to obtain an unlabeled text set, extract cross-modal semantic features for each text using the pre-trained CLIPCMIC zero-shot benchmark framework, and construct the initial semantic feature space. The initialization module is used in the initialization phase. It adopts a greedy k-center selection strategy and iteratively selects the sample farthest from the current selected set to construct the core set as the labeled sample pool based on the sample distance in the initial semantic feature space. The initial training module is used to train an initial text classification model using the labeled sample pool. The two-stage hybrid query module employs a two-stage hybrid query strategy. First, it filters a candidate set with high uncertainty from the unlabeled sample pool based on the prediction entropy. Then, it introduces diversity constraints to select samples with high information content from the candidate set to construct a query pool. After manually labeling the samples in the query pool, it updates the labeled sample pool and the pseudo-labeled sample pool. The updated labeled sample pool is then used to iteratively optimize the parameters of the text classification model. The two-stage hybrid query is repeated until the preset number of iterations is reached to obtain the trained text classification model. The text classification module is used to analyze the set of texts to be classified using a trained text classification model, and obtain the text classification results.

10. A computer device, comprising a memory and a processor, characterized in that, The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the active learning text classification method for information redundancy suppression as described in any one of claims 1 to 8.

Images