A semi-supervised text classification method and system based on multi-encoder generative adversarial learning

Abstract

The present invention relates to a semi-supervised text classification method and system based on multi-encoder generative adversarial learning, which belongs to the field of supervised learning technology. A multi-encoder structure is introduced to provide multiple output representations for the same text sample. The generator part converts the vector sampled from the random noise into a vector representation with the same dimension as the output of the encoder. At the same time, in order to realize semi-supervised text classification, an unsupervised and supervised loss combination module is constructed on the discriminator. Generative adversarial learning is used to continuously promote the generator to generate adversarial samples that are close to the real ones, and at the same time, the discriminator's ability to distinguish between real text samples and adversarial samples generated by the generator, as well as the ability to classify categories, is improved. In addition, through the joint optimization of the generator and the discriminator, the text semantic representation ability of the encoder is further improved.

Description

Technical Field

[0001] The present invention relates to a semi-supervised text classification method and system based on multi-encoder generative adversarial learning, belonging to the technical field of supervised learning. Background Art

[0002] Semi-supervised text classification is one of the key research areas in computer understanding of human language and an important branch of natural language processing (NLP). It aims to enable computers to classify text based on a small amount of labeled data and a large amount of unlabeled data. At present, academic research on semi-supervised text classification has made great achievements, but there are still many important problems that need to be solved in real-world applications, such as category labeling, text semantic representation, and text generation quality.

[0003] Generative adversarial learning, especially the introduction of generative adversarial networks (GANs), provides a new approach to solving the above problems. Through the mutual game between the generator and the discriminator, GAN significantly improves the performance of the model on unlabeled data, especially in generating latent distributions and improving the robustness of classifiers. However, most of the existing semi-supervised methods based on adversarial learning are concentrated in the image field, and there is relatively little research on text classification tasks.

[0004] With the development of deep learning and the implementation of artificial intelligence-related applications, semi-supervised text classification has broad application prospects in many real-world fields. For example, in medical text case analysis, there are usually few accurately labeled case texts for a particular disease, which leads to a decrease in the recognition accuracy in the case analysis and judgment stage, affecting the final result. In addition, the labeling of cases requires professional doctors to make judgments, and it also requires a lot of manpower, material resources and time, which is a huge waste of resources. Semi-supervised text classification uses a small amount of labeled medical text data to realize model training and reasoning, which can avoid the cost of manual labeling resources while ensuring the quality of medical text analysis. In addition, semi-supervised text classification also has great application potential in real-world fields such as small language analysis and specific subject text processing.

[0005] Category labeling is an important issue in semi-supervised text classification research. Currently, most studies usually add category labels to unlabeled text data through some methods, and then increase the number of labeled text data to improve the recognition and classification capabilities of the model. However, in the process of category labeling, it is easy to have labeling errors, introduce low-quality data samples into the training data set, and cause the model to misclassify during training, affecting the classification performance of the model.

[0006] The problem of text semantic representation is also an issue that cannot be ignored in semi-supervised text classification research. With limited labeled samples and a large number of unlabeled samples, the model is difficult to be fully trained due to the insufficient amount of labeled text sample data and the limited category information carried by the data samples. It is easy to overfit, resulting in poor classification and recognition capabilities of the model. In other words, with a limited number of labeled text samples, it is difficult for the model to simultaneously learn the semantic representation capabilities of a high-quality single sample and the representation information of the relevant category.

[0007] In summary, in order to promote the practical application of semi-supervised text classification technology and improve productivity, it is necessary to conduct in-depth research on the above problems and propose reasonable solutions. This invention focuses on solving the problem of category labeling and the problem of text semantic representation, which is of great significance in promoting the development and implementation of cutting-edge technologies in this field.

[0008] Existing semi-supervised text classification methods can be divided into traditional methods and deep learning methods. Traditional semi-supervised text classification methods mainly include generative methods based on EM (Expectation Maximum) ideas, methods based on low-density separation, graph-based methods, and methods based on collaborative training, etc. Semi-supervised text classification methods based on deep learning use limited annotated text data and a large amount of unlabeled text data to train deep neural networks, and learn the mapping function from text data features to category labels. According to the different ways of training data set expansion, semi-supervised text classification technology based on deep neural networks can be divided into two categories: one is to annotate real data based on annotated data enhancement, and gradually annotate unlabeled text data to continuously expand the labeled data set; the other is to generate new text data based on real text data.

[0009] The prior art has the following technical problems:

[0010] 1. Category labeling problem. When labeling unlabeled text samples, existing semi-supervised text classification methods do not have good classification capabilities due to the small amount of original labeled text data, which can easily lead to category labeling errors. Then, amplifying the incorrectly labeled text data into the original labeled data set will introduce errors during model training, leading to incorrect classification of the model.

[0011] 2. Text semantic representation problem. With a small number of labeled text samples and a large number of unlabeled samples, since the original data set carries little classification information, most semi-supervised text classification algorithms based on deep neural networks find it difficult to have both excellent semantic representation capabilities for text sample data and good classification tasks.

[0012] 3. Serious data homogeneity problem. Most semi-supervised text methods perform data amplification on the basis of the original data, which leads to a high degree of homogeneity between samples in the amplified data set, and the feature information contained in the amplified data set is limited. Ultimately, it is difficult to effectively improve the classification ability of the model. Summary of the invention

[0013] In view of the shortcomings of the prior art, in order to avoid the problem of category labeling and improve the text semantic representation ability of the model, the present invention proposes a semi-supervised text classification method based on a multi-encoder generative adversarial network (Semi-supervised text classification based on Multi-encoder Generative Adversarial Network, SMGAN). On the basis of the generative adversarial network, SMGAN introduces a multi-encoder structure, and at the same time, in order to realize semi-supervised text classification, an unsupervised and supervised loss combination module is constructed on the discriminator. Specifically, the multi-encoder structure is constructed by stacking multiple BERTs (Bidirectional Encoder Representations from Transformers) with the same structure in parallel, providing multiple output representations for the same text sample. The generator part converts the vector sampled from random noise into a vector representation with the same dimension as the output of the encoder. The discriminator mainly includes two aspects of loss: one is the unsupervised loss, which is calculated as the relative entropy between the multiple output representations of each sample in all samples; the other is the supervised loss, which is calculated as the cross entropy calculated for all labeled samples for their respective multiple output representations and their labels. Generative adversarial learning is used to continuously promote the generator to generate adversarial samples that are close to reality, while improving the discriminator's ability to distinguish between real text samples and adversarial samples generated by the generator, as well as its ability to classify categories. In addition, through the joint optimization of the generator and the discriminator, the encoder's text semantic representation ability is further improved.

[0014] The technical solution of the present invention is as follows:

[0015] A semi-supervised text classification method based on multi-encoder generative adversarial learning includes the following steps:

[0016] (1) Constructing a multi-encoder structure

[0017] Construct a generative adversarial learning architecture based on multiple encoders. In terms of the construction of multiple encoders, first copy multiple BERT encoders with the same structure, the purpose is to use multiple encoders to obtain multiple output representations of the same sample; set M encoders to be constructed, and the labeled data and unlabeled data are respectively multi-encoded to obtain their respective output representations;

[0018] For labeled sample data (x i ,y i ), the output after the multi-encoder is expressed as:

[0019]

[0020] For unlabeled sample data (x j ), the output after the multi-encoder is expressed as:

[0021]

[0022] (2) Construct a generative adversarial network, including a generator and a discriminator

[0023] The present invention adopts a multi-layer perceptron with one layer as the basic architecture when constructing the generator and the discriminator; wherein the input of the generator is a 100-dimensional noise vector randomly sampled from a normal distribution, and the output is a vector with the same dimension as the encoder, wherein the dimension of the vector is 768, and the BERT encoder is adopted, so the dimension is 768; the input of the discriminator is a multi-dimensional representation vector of the real data and the generated data, wherein the real data is the output of the encoder in step (1), and the generated data is the output of the generator; the output of the discriminator is a k+1-dimensional logit representation, wherein k is the original number of categories, i.e., the number of categories contained in the text data set in the text classification task, and the k+1th dimension is the category of the sample generated by the generator;

[0024] For the input of the generator, the formula is as follows:

[0025]

[0026] In the above formula, It is represented as a normal distribution with a mean of 0 and a variance of 1;

[0027] Next, the output of random noise after passing through the generator is expressed as:

[0028] P G = noise ~G (4)

[0029] In the above formula, G represents the generator;

[0030] (3) Training the generator and discriminator

[0031] In the training of generative adversarial learning, the generator and discriminator are updated and iterated alternately; for the generator, we want it to generate new text samples that are more similar to the real ones; for the discriminator, the purpose is to distinguish between real samples and generated samples, and use labeled data for text classification tasks.

[0032] For labeled sample data (x i ,y i ), its output after the multi-encoder is expressed as follows, and then the logit obtained after the discriminator is expressed as:

[0033]

[0034] For the unlabeled sample data (x j ), its output after the multi-encoder is expressed as follows, and then the logit obtained after the discriminator is expressed as:

[0035]

[0036] When training the generator in the model, it is necessary to avoid the generated data samples being successfully identified as generated data samples by the discriminator. The loss is calculated as shown in the following formula:

[0037]

[0038] in, is the generator loss, θ represents the parameters of the model, represents the expected calculation; x is the sample generated by the random noise generator, y is the generated sample label, P G is the representation of the data sample after passing through the generator, is the label obtained by the discriminator, and p is the category probability representation;

[0039] Construct a module combining unsupervised and supervised losses. The supervised loss in this module is mainly for labeled text sample data and sample data generated by the generator. The discriminator needs to correctly classify the above text samples and calculate the cross entropy loss. When training the discriminator in the model, on the one hand, it is necessary to correctly classify the labeled text samples, that is, assign each text sample to one of the k categories, and on the other hand, it is necessary to identify the generated samples as the k+1th category. The loss calculation formula is as follows:

[0040]

[0041] P BERT It is the representation of the data sample after the encoder BERT;

[0042] (4) Optimize training of the entire model

[0043] While the generator and the discriminator are alternately trained, the multiple encoders are adjusted, that is, the model parameters of multiple BERT encoders are updated to improve the model's text representation ability and text classification performance. In the unsupervised and supervised loss combination module, the unsupervised loss in this module is mainly for labeled sample data and unlabeled sample data. Each sample data obtains multiple different output representations after passing through multiple encoders and discriminators, and the relative entropy is calculated on this basis. Specifically, for multiple output representations of the same sample, the bidirectional Kullback-Leibler divergence is used to constrain multiple output representations so that these outputs tend to be consistent. The loss calculation formula is as follows:

[0044]

[0045] Where c and m represent the c-th and m-th encoders in M ​​BERT encoders, D KL [·||·] indicates the calculation of KL (Kullback-Leibler) divergence; is the representation of the data after passing through the cth encoder and discriminator, It is the representation of the data after passing through the mth encoder and discriminator.

[0046] During the overall training, we need to calculate and These three losses iteratively update the parameters of the generator, discriminator, and multiple encoders.

[0047] Preferably, the iteration is terminated when the set number of training iterations is reached or the three losses become stable and no longer change.

[0048] Further preferably, the maximum number of iteration rounds is set to 50. If the number of rounds is less than 50, and the changes of the three losses tend to be stable and no longer change in 10 consecutive rounds, the iteration is terminated early.

[0049] (5) After the update is completed, a trained discriminator and multiple encoders are obtained; the data to be classified is input into the discriminator, and its multiple category probability representations are obtained through the discriminator, and finally its category is determined according to the majority principle.

[0050] The present invention focuses on constructing a generative adversarial learning architecture based on multiple encoders. In the case of limited annotated text data and a large amount of unannotated text data, multiple encoders are used to obtain multiple output representations of the same text sample, and combined with a generative adversarial learning network, the model learns higher-quality semantic representation capabilities and improves the overall classification performance of the model. The entire method mechanism of semi-supervised text classification based on a multi-encoder generative adversarial network is protected.

[0051] On the one hand, the present invention constructs a learning framework for interactive generation and confrontation between a single generator and multiple encoders. Specifically, multiple encoders are used to obtain multiple output representations of the same text sample, providing richer feature information. On the other hand, on the discriminator, an unsupervised and supervised loss combination module corresponding to the multiple encoders is constructed. The module uses supervised loss to promote the model to better classify labeled samples and generated data samples; at the same time, unsupervised loss is used to encourage multiple encoders to improve their representation capabilities.

[0052] A computer-readable storage medium stores a program thereon, which, when executed by a processor, implements the steps in the above-mentioned semi-supervised text classification method based on multi-encoder generative adversarial learning.

[0053] An electronic device comprises a memory, a processor and a program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the above-mentioned semi-supervised text classification method based on multi-encoder generative adversarial learning are implemented.

[0054] The present invention can be used in text information understanding tasks in specific fields. For example, in sentiment analysis of minority languages, sentiment recognition and analysis can be performed on regional dialects or minority languages; in intelligent human-computer interaction in medical fields, auxiliary judgment of disease conditions can be provided in the case of a small number of known cases; in foreign language translation and communication, translation between uncommon foreign languages ​​can be provided to facilitate communication between people speaking different languages.

[0055] The beneficial effects of the present invention are:

[0056] This invention solves three problems that cannot be solved in most existing semi-supervised text classification algorithms:

[0057] 1. Category labeling problem. The present invention uses a multi-coding structure to obtain multiple output representations of the same sample, and improves the quality and discrimination ability of the overall generated samples of the model by generating interactions between adversarial models. In the above process, no operation of labeling unlabeled data is added, which fundamentally avoids the problem of incorrect labels caused by category labeling, and also reduces the introduction of incorrect sample data in model training.

[0058] 2. Text semantic representation problem. The present invention introduces a multi-encoder structure into the generative adversarial network, and uses the fact that the distribution of adversarial samples generated by the generator tends to be similar to that of real samples, prompting the multi-encoder to improve its own representation ability. Similarly, unsupervised loss is added to the discriminator to prompt the multiple output representations of the same sample after passing through multiple encoders to be consistent, further helping the encoder to learn more feature information. The above modules can all achieve a high-quality improvement in the overall semantic representation ability of the model.

[0059] 3. Serious problem of data homogeneity. The present invention utilizes interactive learning between generative adversarial networks. The input of the generator is a vector sampled from random noise, and the condition that limits the generator's generation is the distribution of the entire data set, and the discriminator is used to promote the generator's generation quality. Throughout the process, the generated data not only depends on the original data samples, but also on the overall data distribution, and continuously generates data representations that are more similar to real samples based on the noise data. The sample data generated in this way is more differentiated, and the feature information it carries is richer, alleviating the problem of data homogeneity. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 A schematic diagram of a semi-supervised text classification method based on a multi-encoder generative adversarial network according to the present invention;

[0061] Figure 2 This is a schematic diagram of translation and communication between foreign languages ​​of multiple countries.

[0062] Figure 3 Schematic diagram of the existing generation-based data augmentation method for semi-supervised text classification. DETAILED DESCRIPTION

[0063] The present invention will be further described below by way of embodiments in conjunction with the accompanying drawings, but is not limited thereto.

[0064] Embodiment 1:

[0065] A semi-supervised text classification method based on multi-encoder generative adversarial learning, such as Figure 1 As shown, the steps include:

[0066] (1) Constructing a multi-encoder structure

[0067] Construct a generative adversarial learning architecture based on multiple encoders. In terms of the construction of multiple encoders, first copy multiple BERT encoders with the same structure, the purpose is to use multiple encoders to obtain multiple output representations of the same sample; set M encoders to be constructed, and the labeled data and unlabeled data are respectively multi-encoded to obtain their respective output representations;

[0068] For labeled sample data (x i ,y i ), the output after the multi-encoder is expressed as:

[0069]

[0070] For unlabeled sample data (x j ), the output after the multi-encoder is expressed as:

[0071]

[0072] (2) Construct a generative adversarial network, including a generator and a discriminator

[0073] The present invention adopts a multi-layer perceptron with one layer as the basic architecture when constructing the generator and the discriminator; wherein the input of the generator is a 100-dimensional noise vector randomly sampled from a normal distribution, and the output is a vector with the same dimension as the encoder, wherein the dimension of the vector is 768, and the BERT encoder is adopted, so the dimension is 768; the input of the discriminator is a multi-dimensional representation vector of the real data and the generated data, wherein the real data is the output of the encoder in step (1), and the generated data is the output of the generator; the output of the discriminator is a k+1-dimensional logit representation, wherein k is the original number of categories, i.e., the number of categories contained in the text data set in the text classification task, and the k+1th dimension is the category of the sample generated by the generator;

[0074] For the input of the generator, the formula is as follows:

[0075]

[0076] In the above formula, It is represented as a normal distribution with a mean of 0 and a variance of 1;

[0077] Next, the output of random noise after passing through the generator is expressed as:

[0078] P G = noise ~G (4)

[0079] In the above formula, G represents the generator;

[0080] (3) Training the generator and discriminator

[0081] In the training of generative adversarial learning, the generator and discriminator are updated and iterated alternately; for the generator, we want it to generate new text samples that are more similar to the real ones; for the discriminator, the purpose is to distinguish between real samples and generated samples, and use labeled data for text classification tasks.

[0082] For labeled sample data (x i ,y i ), its output after the multi-encoder is expressed as follows, and then the logit obtained after the discriminator is expressed as:

[0083]

[0084] For the unlabeled sample data (x j ), its output after the multi-encoder is expressed as follows, and then the logit obtained after the discriminator is expressed as:

[0085]

[0086] When training the generator in the model, it is necessary to avoid the generated data samples being successfully identified as generated data samples by the discriminator. The loss is calculated as shown in the following formula:

[0087]

[0088] in, is the generator loss, θ represents the parameters of the model, represents the expected calculation; x is the sample generated by the random noise generator, y is the generated sample label, P G is the representation of the data sample after passing through the generator, is the label obtained by the discriminator, and p is the category probability representation;

[0089] Construct a module combining unsupervised and supervised losses. The supervised loss in this module is mainly for labeled text sample data and sample data generated by the generator. The discriminator needs to correctly classify the above text samples and calculate the cross entropy loss. When training the discriminator in the model, on the one hand, it is necessary to correctly classify the labeled text samples, that is, assign each text sample to one of the k categories, and on the other hand, it is necessary to identify the generated samples as the k+1th category. The loss calculation formula is as follows:

[0090]

[0091] P BERT It is the representation of the data sample after the encoder BERT;

[0092] (4) Optimize training of the entire model

[0093] While the generator and the discriminator are alternately trained, the multiple encoders are adjusted, that is, the model parameters of multiple BERT encoders are updated to improve the model's text representation ability and text classification performance. In the unsupervised and supervised loss combination module, the unsupervised loss in this module is mainly for labeled sample data and unlabeled sample data. Each sample data obtains multiple different output representations after passing through multiple encoders and discriminators, and the relative entropy is calculated on this basis. Specifically, for multiple output representations of the same sample, the bidirectional Kullback-Leibler divergence is used to constrain multiple output representations so that these outputs tend to be consistent. The loss calculation formula is as follows:

[0094]

[0095] Where c and m represent the c-th and m-th encoders in M ​​BERT encoders, D KL[·||·] indicates the calculation of KL (Kullback-Leibler) divergence; is the representation of the data after passing through the cth encoder and discriminator, It is the representation of the data after passing through the mth encoder and discriminator.

[0096] During the overall training, we need to calculate and These three losses iteratively update the parameters of the generator, discriminator, and multiple encoders.

[0097] The maximum number of iteration rounds is set to 50. If the number of rounds is less than 50 and the changes in the three losses tend to be stable and no longer change in 10 consecutive rounds, the iteration is terminated early.

[0098] (5) After the update is completed, a trained discriminator and multiple encoders are obtained; the data to be classified is input into the discriminator, and its multiple category probability representations are obtained through the discriminator, and finally its category is determined according to the majority principle.

[0099] Embodiment 2:

[0100] A computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps in a semi-supervised text classification method based on multi-encoder generative adversarial learning as described in Example 1.

[0101] Embodiment 3:

[0102] An electronic device comprises a memory, a processor and a program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in a semi-supervised text classification method based on multi-encoder generative adversarial learning as described in Example 1 are implemented.

Claims

1. A semi-supervised text classification method based on multi-encoder generative adversarial learning, characterized in that: The steps include: (1) Constructing a multi-encoder structure Construct a generative adversarial learning architecture based on multiple encoders. In terms of the construction of multiple encoders, first copy multiple BERT encoders with the same structure and use multiple encoders to obtain multiple output representations of the same sample; set M encoders to be constructed, and obtain their own output representations after multi-encoding of labeled data and unlabeled data; For labeled sample data (x i ,y i ), the output after the multi-encoder is expressed as: For unlabeled sample data (x j ), the output after the multi-encoder is expressed as: (2) Construct a generative adversarial network, including a generator and a discriminator When constructing the generator and the discriminator, a one-layer perceptron is used as the basic architecture. The input of the generator is a noise vector randomly sampled from a normal distribution, and the output is a vector with the same dimension as the encoder. The input of the discriminator is a multi-dimensional representation vector of real data and generated data, where the real data is the output of the encoder in step (1) and the generated data is the output of the generator. The output of the discriminator is a k+1-dimensional logit representation, where k is the number of original categories, that is, the number of categories contained in the text data set in the text classification task, and the k+1th dimension is the category of the sample generated by the generator. For the input of the generator, the formula is as follows: In the above formula, It is represented as a normal distribution with a mean of 0 and a variance of 1; Next, the output of random noise after passing through the generator is expressed as: P G = noise ~G (4) In the above formula, G represents the generator; (3) Training the generator and discriminator In the training of generative adversarial learning, the generator and discriminator are updated and iterated alternately; For labeled sample data (x i ,y i ), its output after the multi-encoder is expressed as, and then the logit obtained after the discriminator is expressed as: For the unlabeled sample data (x j ), its output after the multi-encoder is expressed as, and then the logit obtained after the discriminator is expressed as: When training the generator in the model, its loss is calculated as shown in the following formula: in, is the generator loss, θ represents the parameters of the model, represents the expected calculation; x is the sample generated by the random noise generator, y is the generated sample label, P G is the representation of the data sample after passing through the generator, is the label obtained by the discriminator, and p is the category probability representation; Construct a module combining unsupervised and supervised losses. The supervised loss in this module targets the labeled text sample data and the sample data generated by the generator. The discriminator needs to correctly classify the above text samples and calculate the cross entropy loss. When training the discriminator in the model, on the one hand, it is necessary to correctly classify the labeled text samples, that is, assign each text sample to one of the k categories, and on the other hand, it is necessary to identify the generated samples as the k+1th category. The loss calculation formula is as follows: P BERT It is the representation of the data sample after the encoder BERT; (4) Optimize training of the entire model While the generator and the discriminator are alternately iteratively trained, the multi-encoder is adjusted, that is, the parameters of multiple BERT encoders are updated. In the unsupervised and supervised loss combination module, the unsupervised loss in this module is for labeled sample data and unlabeled sample data. Each sample data obtains multiple different output representations after passing through multiple encoders and discriminators. On this basis, the relative entropy is calculated. For multiple output representations of the same sample, the bidirectional Kullback-Leibler divergence is used to constrain multiple output representations so that these outputs tend to be consistent. The loss calculation formula is as follows: Where c and m represent the c-th and m-th encoders in M ​​BERT encoders, D KL [·||·] indicates the calculation of Kullback-Leibler divergence; is the representation of the data after passing through the cth encoder and discriminator, The data is represented after passing through the mth encoder and discriminator; During the overall training, calculate and These three losses iteratively update the parameters of the generator, discriminator, and multiple encoders; (5) After the update is completed, a trained discriminator and multiple encoders are obtained; the data to be classified is input into the discriminator, and its multiple category probability representations are obtained through the discriminator, and finally its category is determined according to the majority principle.

2. The semi-supervised text classification method based on multi-encoder generative adversarial learning according to claim 1 is characterized in that: In step (2), the input of the generator is a 100-dimensional noise vector randomly sampled from a normal distribution, and the output is a vector with the same dimension as the encoder, which is 768 here.

3. The semi-supervised text classification method based on multi-encoder generative adversarial learning according to claim 1 is characterized in that: In step (4), the iteration is terminated when the set number of training iterations is reached or the three losses become stable and no longer change.

4. The semi-supervised text classification method based on multi-encoder generative adversarial learning according to claim 3 is characterized in that: The maximum number of iteration rounds is set to 50. If the number of rounds is less than 50 and the changes of the three losses are stable and no longer change in 10 consecutive rounds, the iteration is terminated early.

5. A computer-readable storage medium, characterized in that: A program is stored thereon, and when the program is executed by a processor, the steps in the semi-supervised text classification method based on multi-encoder generative adversarial learning as described in claim 1 are implemented.

6. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the semi-supervised text classification method based on multi-encoder generative adversarial learning as described in claim 1 are implemented.

Images