A classification method and system based on label propagation semi-supervised learning
By using label augmentation methods in semi-supervised learning to obtain a public dataset with features consistent with the labeled dataset, and then using a deep learning model to generate pseudo-labels and optimize the pseudo-label dataset, the problem of insufficient label data augmentation is solved, and the performance of the classification model is improved efficiently.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- XIAMEN KUAISHANGTONG TECH CORP LTD
- Filing Date
- 2022-10-12
- Publication Date
- 2026-06-19
Smart Images
Figure CN115618275B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of classification algorithms, and in particular to a classification method and system based on label-enhanced semi-supervised learning. Background Technology
[0002] For classification problems, a simple and effective method is data augmentation. Data augmentation increases the size of the training set by synthesizing data. Time-series-based data augmentation methods often employ techniques such as flipping, window slicing, and window distortion, which can lead to inaccurate data. Furthermore, a common practice in traditional data augmentation is to assign the same original label to new samples obtained through data augmentation. If the new samples differ significantly from the original samples, it can lead to an expanded data distribution, making it difficult to determine the classification boundary.
[0003] In the field of semi-supervised learning, in order to predict pseudo-labels for unlabeled data, most existing augmentation schemes only augment the training data, and there are very few algorithms that use label information for augmentation.
[0004] Unlabeled data is virtually endless, and how to effectively and even efficiently select the most valuable data for learning has always been a major challenge.
[0005] The evaluation metric for classification algorithm models is the label, but current semi-supervised learning schemes rarely model the label or optimize the model based on the label. Summary of the Invention
[0006] The main objective of this invention is to overcome the aforementioned deficiencies in the prior art and propose a classification method based on label enhancement semi-supervised learning. By using the label information learned by the model, retrieving public datasets, and implementing a pseudo-label semi-supervised learning scheme, as well as an adaptive learning scheme that modifies the pseudo-label dataset according to the model's learning process, the performance of the classification model is enhanced. Under the premise of achieving the same performance, the requirement for the number of labels can be greatly reduced.
[0007] The present invention adopts the following technical solution:
[0008] A classification method based on label enhancement semi-supervised learning includes:
[0009] S1: Obtain a public dataset with a feature distribution consistent with the labeled dataset;
[0010] S2: Train the deep learning model using labeled datasets and public datasets, specifically:
[0011] S21: Divide the labeled dataset into a training set and a validation set;
[0012] S22: Each time without replacement, n training examples are drawn from the training set to perform initial training on the deep learning model;
[0013] S23: After every m initial training cycles, the public dataset is converted into vectors using the initially trained deep learning model and stored in the vector dataset. Using the label information learned by the initially trained deep learning model, the vector dataset is retrieved and pseudo-labels are assigned to the selected vectors. The data corresponding to the selected vectors and the pseudo-labels are then input into the initially trained deep learning model for model optimization.
[0014] Specifically, in step S23, the publicly available dataset is converted into vectors using the initially trained deep learning model and stored in the vector dataset. The label information learned by the initially trained deep learning model is used to retrieve data from the vector dataset and assign pseudo-labels to the selected vectors. The data corresponding to the selected vectors and the pseudo-labels are then input into the initially trained deep learning model for model optimization. Specifically:
[0015] The encoder in the initially trained deep learning model is used to convert the public dataset into vectors and store them in a vector database.
[0016] Then, the classifier in the initially trained deep learning model is used to recall vectors in the vector dataset according to the recall strategy, and pseudo-labels are applied.
[0017] Specifically, after recalling vectors in the vector dataset according to the recall strategy and assigning pseudo-labels, it also includes:
[0018] The recalled vectors are filtered according to the filtering strategy;
[0019] The text corresponding to the filtered vectors, along with the corresponding pseudo-labels, are input into the initially trained deep learning model for optimization.
[0020] Specifically, the recall strategy includes, but is not limited to:
[0021] The first recall strategy is to recall the top n vectors in the vector database that are most relevant to a certain label.
[0022] The second recall strategy is to recall n vectors in the vector database that are more than a certain threshold p in relevance to a certain label.
[0023] The third recall strategy is to recall n vectors in the vector database that are more relevant to a certain label than a threshold p and less relevant to any other label except that label than q.
[0024] The correlation calculation methods include, but are not limited to, calculating cosine similarity, vector inner product, or lp distance.
[0025] Specifically, the filtering strategy corresponding to the third recall strategy is as follows:
[0026] p = F(step)
[0027] q = H(step)
[0028] Where step is the number of training steps, f and h represent the function mapping relationship, including but not limited to linear functions and exponential functions, respectively, threshold p is inversely correlated with the number of training steps, and threshold q is positively correlated with the number of training steps.
[0029] Another embodiment of the present invention provides a classification system based on label enhancement semi-supervised learning, comprising:
[0030] Public dataset acquisition unit: Acquire public datasets with feature distributions consistent with the labeled datasets;
[0031] Model training unit: The deep learning model is trained using labeled datasets and public datasets, specifically as follows:
[0032] Dataset classification subunit: The labeled dataset is divided into a training set and a validation set;
[0033] Initial training subunit: n training examples are extracted from the training set without replacement each time to perform initial training on the deep learning model;
[0034] Model optimization subunit: After every m initial training cycles, the public dataset is converted into vectors using the initially trained deep learning model and stored in the vector dataset. Using the label information learned by the initially trained deep learning model, the vector dataset is retrieved and pseudo-labels are assigned to the selected vectors. The data corresponding to the selected vectors and the pseudo-labels are then input into the initially trained deep learning model for model optimization.
[0035] Specifically, in the model optimization subunit, the publicly available dataset is converted into vectors using the initially trained deep learning model and stored in the vector dataset. The label information learned by the initially trained deep learning model is used to retrieve data from the vector dataset. The model then assigns pseudo-labels to the selected data and trains the model. Specifically:
[0036] The encoder in the initially trained deep learning model is used to convert the public dataset into vectors and store them in a vector database.
[0037] Then, the classifier in the initially trained deep learning model is used to recall vectors in the vector dataset according to the recall strategy, and pseudo-labels are applied.
[0038] Specifically, after recalling vectors in the vector dataset according to the recall strategy and assigning pseudo-labels, it also includes:
[0039] The recalled vectors are filtered according to the filtering strategy;
[0040] The text corresponding to the filtered vectors, along with the corresponding pseudo-labels, are input into the initially trained deep learning model for optimization.
[0041] Specifically, the recall strategy includes, but is not limited to:
[0042] The first recall strategy is to recall the top n vectors in the vector database that are most relevant to a certain label.
[0043] The second recall strategy is to recall n vectors in the vector database that are more than a certain threshold p in relevance to a certain label.
[0044] The third recall strategy is to recall n vectors in the vector database that are more relevant to a certain label than a threshold p and less relevant to any other label except that label than q.
[0045] The correlation calculation methods include, but are not limited to, calculating cosine similarity, vector inner product, or lp distance.
[0046] Specifically, the filtering strategy corresponding to the third recall strategy is as follows:
[0047] p = F(step)
[0048] q = H(step)
[0049] Where step is the number of training steps, f and h represent the function mapping relationship, including but not limited to linear functions and exponential functions, respectively, threshold p is inversely correlated with the number of training steps, and threshold q is positively correlated with the number of training steps.
[0050] In another aspect, the present invention provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described label-enhanced semi-supervised learning classification method.
[0051] In another aspect, this invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the aforementioned label-enhanced semi-supervised learning classification method.
[0052] As can be seen from the above description of the present invention, compared with the prior art, the present invention has the following beneficial effects:
[0053] This invention provides a classification method based on label-enhanced semi-supervised learning, comprising: S1: obtaining a public dataset with a feature distribution consistent with the labeled dataset; S2: training a deep learning model using the labeled dataset and the public dataset, specifically: S21: dividing the labeled dataset into a training set and a validation set; S22: extracting n training examples from the training set without replacement each time to perform initial training on the deep learning model; S23: after every m initial training iterations, using the initially trained deep learning model to convert the public dataset into vectors, storing them in a vector dataset, using the label information learned by the initially trained deep learning model to retrieve data from the vector dataset and assign pseudo-labels to the selected vectors, and inputting the data corresponding to the selected vectors and the pseudo-labels into the initially trained deep learning model for model optimization. The method provided by this invention enhances the performance of the classification model by using the label information learned by the model to retrieve data from the public dataset, employing a semi-supervised learning scheme of pseudo-labeling, and adaptively modifying the pseudo-label dataset according to the model learning process. While achieving the same performance, it can significantly reduce the requirement for the number of labels. Attached Figure Description
[0054] Figure 1 A flowchart of a classification method based on label enhancement semi-supervised learning is provided for an embodiment of the present invention;
[0055] Figure 2 A detailed flowchart of the model training steps provided in the embodiments of the present invention;
[0056] Figure 3 A classification system architecture diagram based on label enhancement semi-supervised learning is provided for an embodiment of the present invention;
[0057] Figure 4 A schematic diagram of an electronic device provided in an embodiment of the present invention;
[0058] Figure 5 This is a schematic diagram of an embodiment of a computer-readable storage medium provided in this invention.
[0059] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Detailed Implementation
[0060] This invention provides a classification method based on label enhancement semi-supervised learning. By using the label information learned by the model to retrieve public datasets and apply pseudo-labels in a semi-supervised learning scheme, and by adaptively modifying the pseudo-label dataset according to the model's learning process, the performance of the classification model is enhanced. Under the premise of achieving the same performance, the requirement for the number of labels can be greatly reduced.
[0061] like Figure 1The flowchart below illustrates a classification method based on label enhancement semi-supervised learning, as provided in this embodiment of the invention. The method includes the following steps:
[0062] S1: Obtain a public dataset with a feature distribution consistent with the labeled dataset;
[0063] Public datasets can be obtained by using publicly available data on GitHub for training pre-trained models, or by using web crawlers to scrape data from websites such as Baidu Baike or Tieba. This process can yield a large amount of publicly available natural language corpus. Consistent feature distribution means that the sources are the same. For example, if the labeled dataset comes from natural text in a medical forum, then the public dataset should also come from that source; if the labeled dataset is image data of animal classification, then the public dataset should also be images of animals.
[0064] S2: Train the deep learning model using labeled datasets and public datasets, specifically: Figure 2 A detailed flowchart of the model training steps provided in the embodiments of the present invention;
[0065] S21: Divide the labeled dataset into a training set and a validation set;
[0066] S22: Each time without replacement, n training examples are drawn from the training set to perform initial training on the deep learning model;
[0067] S23: After every m initial training cycles, the public dataset is converted into vectors using the initially trained deep learning model and stored in the vector dataset. Using the label information learned by the initially trained deep learning model, the vector dataset is retrieved and pseudo-labels are assigned to the selected vectors. The data corresponding to the selected vectors and the pseudo-labels are then input into the initially trained deep learning model for model optimization.
[0068] A classification model can be viewed as having two modules: an encoder and a classifier. The encoder is responsible for extracting training features and converting them into vectors that the machine can compute, while the classifier typically consists of a fully connected layer and a softmax layer, and is responsible for classifying the features learned by the encoder.
[0069] Assume the features learned by the encoder are x, the parameters of the fully connected layer are W, and ignore the bias parameters of the fully connected layer. The dimension of x is 1*dim, and the dimension of the fully connected layer is num_label *dim. dim is the dimension of the feature vector learned by the encoder, and numlabel is the number of categories. The calculation of the fully connected layer is W. T X can be understood as the inner product of the features learned by the encoding layer and the label category features learned by the model. The size of the inner product also reflects the correlation between the feature and a certain label category.
[0070] Specifically, in step S23, the publicly available dataset is converted into vectors using the initially trained deep learning model and stored in the vector dataset. The label information learned by the initially trained deep learning model is used to retrieve data from the vector dataset and assign pseudo-labels to the selected vectors. The data corresponding to the selected vectors and the pseudo-labels are then input into the initially trained deep learning model for model optimization. Specifically:
[0071] The encoder in the initially trained deep learning model is used to convert the public dataset into vectors and store them in a vector database.
[0072] Then, the classifier in the initially trained deep learning model is used to recall vectors in the vector dataset according to the recall strategy, and pseudo-labels are applied.
[0073] Specifically, the process involves first inputting labeled data to allow the model to learn a better encoder and classifier. Then, the encoder vectorizes the publicly available dataset collected in the first step and stores it in a vector database. Next, the vectors representing each label learned by the classifier are retrieved from the vector database using a recall strategy, and pseudo-labels are added. The quality of the retrieved vectors varies in this step, so a filtering strategy is needed to further refine them. Finally, the corresponding text and pseudo-labels are input into the model to optimize it.
[0074] More specifically:
[0075] Traverse the public dataset, convert it into a vector using an encoder, and add an index. For example, assuming the first sentence of the dataset is "hello, world", and the vector is [0.1, 0.3, 0.5], then we can get the following correspondence: [0, "hello, world", [0.1, 0.3, 0.5]).
[0076] The recall strategy has two components: recall metrics and recall plan.
[0077] In this embodiment of the invention, the recall index refers to the method of calculating the correlation between two vectors. The available options include cosine similarity, vector inner product, lp distance, etc.
[0078] In this embodiment of the invention, the recall scheme refers to using tag vectors, and the following schemes are available:
[0079] First recall strategy: Retrieve the top n vectors in the database that are most relevant to a certain tag;
[0080] The second recall strategy is to recall n vectors from the database that are more than a certain threshold p in relevance to a certain tag.
[0081] The third recall scheme is to recall n vectors in the database that are more relevant to a certain label than a threshold p and less relevant to any other label except that label than q.
[0082] Filtering strategies refer to methods for selecting recall vectors. We need to use deep learning models to label the recalled data, and ideally, during model optimization, the error rate will continuously decrease and gradually converge to a small value. The purpose of filtering strategies is to mitigate the inevitable errors that occur during model optimization.
[0083] Taking the third recall scheme as an example, as the number of training steps increases, the threshold p should be inversely correlated with it, and the threshold q should be positively correlated with it.
[0084] p = F(step)
[0085] q = H(step)
[0086] Here, f and h represent a function mapping relationship, which can be a linear function, an exponential function, and so on.
[0087] Next, pseudo-labels are applied to the data that passes the filtering strategy.
[0088] For example, if p is set to 0.9 and q to 0.1, it means that the recalled vector must satisfy the following conditions: its correlation with label A is greater than 0.9 and its correlation with any other label is less than 0.1. This ensures that the obtained data does not belong to two categories at the same time. Finally, a pseudo-label A is added.
[0089] As shown above, there is a unique index corresponding to each sentence and its transformed vector. Once the vectors are selected, the corresponding text can be obtained through the unique index. The text and pseudo-labels are then input into the initially trained model for optimization.
[0090] like Figure 3 This is a diagram illustrating the architecture of a classification system based on label-enhanced semi-supervised learning, as provided in an embodiment of the present invention.
[0091] Public dataset acquisition unit 31: Acquire a public dataset with a feature distribution consistent with the labeled dataset;
[0092] Public datasets can be obtained by using publicly available data on GitHub for training pre-trained models, or by using web crawlers to scrape data from websites such as Baidu Baike or Tieba. This process can yield a large amount of publicly available natural language corpus. Consistent feature distribution means that the sources are the same. For example, if the labeled dataset comes from natural text in a medical forum, then the public dataset should also come from that source; if the labeled dataset is image data of animal classification, then the public dataset should also be images of animals.
[0093] Model Training Unit 32: Train the deep learning model using labeled datasets and public datasets, specifically:
[0094] Dataset classification subunit 321: Divides the labeled dataset into a training set and a validation set;
[0095] Model initial training subunit 322: Each time without replacement, n training examples are drawn from the training set to perform initial training on the deep learning model;
[0096] Model optimization subunit 323: After every m initial training cycles, the public dataset is converted into vectors using the initially trained deep learning model and stored in the vector dataset. Using the label information learned by the initially trained deep learning model, the vector dataset is retrieved and pseudo-labels are assigned to the selected vectors. The data corresponding to the selected vectors and the pseudo-labels are then input into the initially trained deep learning model for model optimization.
[0097] A classification model can be viewed as having two modules: an encoder and a classifier. The encoder is responsible for extracting training features and converting them into vectors that the machine can compute, while the classifier typically consists of a fully connected layer and a softmax layer, and is responsible for classifying the features learned by the encoder.
[0098] Assume the features learned by the encoder are x, the parameters of the fully connected layer are W, and ignore the bias parameters of the fully connected layer. The dimension of x is 1*dim, and the dimension of the fully connected layer is num_label *dim. dim is the dimension of the feature vector learned by the encoder, and numlabel is the number of categories. The calculation of the fully connected layer is W. T X can be understood as the inner product of the features learned by the encoding layer and the label category features learned by the model. The size of the inner product also reflects the correlation between the feature and a certain label category.
[0099] Specifically, in model optimization subunit 323, the publicly available dataset is converted into vectors using the initially trained deep learning model and stored in the vector dataset. Using the label information learned by the initially trained deep learning model, the vector dataset is retrieved, and pseudo-labels are assigned to the selected vectors. The data corresponding to the selected vectors and the pseudo-labels are then input into the initially trained deep learning model for model optimization. Specifically:
[0100] The encoder in the initially trained deep learning model is used to convert the public dataset into vectors and store them in a vector database.
[0101] Then, the classifier in the initially trained deep learning model is used to recall vectors in the vector dataset according to the recall strategy, and pseudo-labels are applied.
[0102] Specifically, the process involves first inputting labeled data to allow the model to learn a better encoder and classifier. Then, the encoder vectorizes the publicly available dataset collected in the first step and stores it in a vector database. Next, the vectors representing each label learned by the classifier are retrieved from the vector database using a recall strategy, and pseudo-labels are added. The quality of the retrieved vectors varies in this step, so a filtering strategy is needed to further refine them. Finally, the corresponding text and pseudo-labels are input into the model to optimize it.
[0103] More specifically:
[0104] Traverse the public dataset, convert it into a vector using an encoder, and add an index. For example, assuming the first sentence of the dataset is "hello, world", and the vector is [0.1, 0.3, 0.5], then we can get the following correspondence: [0, "hello, world", [0.1, 0.3, 0.5]).
[0105] The recall strategy has two components: recall metrics and recall plan.
[0106] In this embodiment of the invention, the recall index refers to the method of calculating the correlation between two vectors. The available options include cosine similarity, vector inner product, lp distance, etc.
[0107] In this embodiment of the invention, the recall scheme refers to using tag vectors, and the following schemes are available:
[0108] First recall strategy: Retrieve the top n vectors in the database that are most relevant to a certain tag;
[0109] The second recall strategy is to recall n vectors from the database that are more than a certain threshold p in relevance to a certain tag.
[0110] The third recall scheme is to recall n vectors in the database that are more relevant to a certain label than a threshold p and less relevant to any other label except that label than q.
[0111] Filtering strategies refer to methods for selecting recall vectors. We need to use deep learning models to label the recalled data, and ideally, during model optimization, the error rate will continuously decrease and gradually converge to a small value. The purpose of filtering strategies is to mitigate the inevitable errors that occur during model optimization.
[0112] Taking the third recall scheme as an example, as the number of training steps increases, the threshold p should be inversely correlated with it, and the threshold q should be positively correlated with it.
[0113] p = F(step)
[0114] q = H(step)
[0115] Here, f and h represent a function mapping relationship, which can be a linear function, an exponential function, and so on.
[0116] Next, pseudo-labels are applied to the data that passes the filtering strategy.
[0117] For example, if p is set to 0.9 and q to 0.1, it means that the recalled vector must satisfy the following conditions: its correlation with label A is greater than 0.9 and its correlation with any other label is less than 0.1. This ensures that the obtained data does not belong to two categories at the same time. Finally, a pseudo-label A is added.
[0118] As shown above, there is a unique index corresponding to each sentence and its transformed vector. Once the vectors are selected, the corresponding text can be obtained through the unique index. The text and pseudo-labels are then input into the initially trained model for optimization.
[0119] Figure 4 As shown, an embodiment of the present invention provides an electronic device 400, including a memory 410, a processor 420, and a computer program 411 stored in the memory 420 and executable on the processor 420. When the processor 420 executes the computer program 411, it implements a classification method based on label enhancement semi-supervised learning provided by the embodiment of the present invention.
[0120] Since the electronic device described in this embodiment is the device used to implement the embodiments of the present invention, those skilled in the art can understand the specific implementation methods and various variations of the electronic device in this embodiment based on the methods described in the embodiments of the present invention. Therefore, how the electronic device implements the methods in the embodiments of the present invention will not be described in detail here. Any device used by those skilled in the art to implement the methods in the embodiments of the present invention is within the scope of protection of the present invention.
[0121] Please see Figure 5 , Figure 5 This is a schematic diagram of an embodiment of a computer-readable storage medium provided in this invention.
[0122] like Figure 5 As shown, this embodiment provides a computer-readable storage medium 500, on which a computer program 511 is stored. When the computer program 511 is executed by a processor, it implements a classification method based on label enhancement semi-supervised learning provided in this embodiment of the invention.
[0123] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.
[0124] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0125] This invention provides a classification method based on label-enhanced semi-supervised learning, comprising: S1: obtaining a public dataset with a feature distribution consistent with the labeled dataset; S2: training a deep learning model using the labeled dataset and the public dataset, specifically: S21: dividing the labeled dataset into a training set and a validation set; S22: extracting n training examples from the training set without replacement each time to perform initial training on the deep learning model; S23: after every m initial training iterations, using the initially trained deep learning model to convert the public dataset into vectors, storing them in a vector dataset, using the label information learned by the initially trained deep learning model to retrieve data from the vector dataset and assign pseudo-labels to the selected vectors, and inputting the data corresponding to the selected vectors and the pseudo-labels into the initially trained deep learning model for model optimization. The method provided by this invention enhances the performance of the classification model by using the label information learned by the model to retrieve data from the public dataset, employing a semi-supervised learning scheme of pseudo-labeling, and adaptively modifying the pseudo-label dataset according to the model learning process. While achieving the same performance, it can significantly reduce the requirement for the number of labels.
[0126] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. The above descriptions are merely specific embodiments of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.
[0127] The above are merely specific embodiments of the present invention, but the design concept of the present invention is not limited thereto. Any non-substantial modifications made to the present invention using this concept shall be considered as infringing upon the protection scope of the present invention.
Claims
1. A classification method based on label reinforcement semi-supervised learning, characterized in that, include: S1: Obtain a public dataset with a feature distribution consistent with the labeled dataset, wherein the labeled dataset and the public dataset contain text; S2: Train the deep learning model using labeled datasets and public datasets, specifically: S21: Divide the labeled dataset into a training set and a validation set; S22: Each time without replacement, n training examples are drawn from the training set to perform initial training on the deep learning model; S23: After every m initial training cycles, the public dataset is converted into vectors using the initially trained deep learning model and stored in the vector dataset. The vector dataset is then retrieved using the label information learned by the initially trained deep learning model, and pseudo-labels are applied to the selected vectors. Specifically, this includes: converting the public dataset into vectors using the encoder in the initially trained deep learning model and storing them in the vector database. Then, the classifier in the initially trained deep learning model is used to recall vectors in the vector dataset according to the recall strategy and assign pseudo-labels; the recalled vectors are filtered according to the filtering strategy; there is a unique index corresponding to each sentence and its transformed vector. After the vectors are filtered, the corresponding text is obtained through the unique index; the text corresponding to the filtered vectors and the corresponding pseudo-labels are input into the initially trained deep learning model for model optimization; the optimized deep learning model is used to classify the text.
2. The classification method based on label enhancement semi-supervised learning according to claim 1, characterized in that, The recall strategy specifically includes, but is not limited to: The first recall strategy is to recall the top n vectors in the vector database that are most relevant to a certain label. The second recall strategy is to recall n vectors in the vector database that are more than a certain threshold p in relevance to a certain label. The third recall strategy is to recall n vectors in the vector database that are more relevant to a certain label than a threshold p and less relevant to any other label except that label than q. The correlation calculation methods include, but are not limited to, calculating cosine similarity, vector inner product, or lp distance.
3. The classification method based on label enhancement semi-supervised learning according to claim 2, characterized in that, The filtering strategy corresponding to the third recall strategy is as follows: p = F(step) q = H(step) Where step is the number of training steps, F and H represent the function mapping relationship, including but not limited to linear functions and exponential functions, respectively, threshold p is inversely correlated with the number of training steps, and threshold q is positively correlated with the number of training steps.
4. A classification system based on label reinforcement semi-supervised learning, characterized in that, include: Public dataset acquisition unit: Acquires a public dataset with a feature distribution consistent with the labeled dataset, wherein both the labeled dataset and the public dataset contain text; Model training unit: The deep learning model is trained using labeled datasets and public datasets, specifically as follows: Dataset classification subunit: The labeled dataset is divided into a training set and a validation set; Initial training subunit: n training examples are extracted from the training set without replacement each time to perform initial training on the deep learning model; Model optimization subunit: After every m initial training cycles, the public dataset is converted into vectors using the initially trained deep learning model and stored in the vector dataset. The label information learned by the initially trained deep learning model is used to retrieve the vector dataset and assign pseudo-labels to the selected vectors. Specifically, this includes: using the encoder in the initially trained deep learning model to convert the public dataset into vectors and store them in the vector database. Then, the classifier in the initially trained deep learning model is used to recall vectors in the vector dataset according to the recall strategy and assign pseudo-labels; the recalled vectors are filtered according to the filtering strategy; there is a unique index corresponding to each sentence and its transformed vector. After the vectors are filtered, the corresponding text is obtained through the unique index; the text corresponding to the filtered vectors and the corresponding pseudo-labels are input into the initially trained deep learning model for model optimization; the optimized deep learning model is used to classify the text.
5. A classification system based on label enhancement semi-supervised learning according to claim 4, characterized in that, The recall strategy specifically includes, but is not limited to: The first recall strategy is to recall the top n vectors in the vector database that are most relevant to a certain label. The second recall strategy is to recall n vectors in the vector database that are more than a certain threshold p in relevance to a certain label. The third recall strategy is to recall n vectors in the vector database that are more relevant to a certain label than a threshold p and less relevant to any other label except that label than q. The correlation calculation methods include, but are not limited to, calculating cosine similarity, vector inner product, or lp distance.
6. A classification system based on label enhancement semi-supervised learning according to claim 5, characterized in that, The filtering strategy corresponding to the third recall strategy is as follows: p = F(step) q = H(step) Where step is the number of training steps, F and H represent the function mapping relationship, including but not limited to linear functions and exponential functions, respectively, threshold p is inversely correlated with the number of training steps, and threshold q is positively correlated with the number of training steps.