Information transfer learning method, device and equipment based on deep conditional distribution alignment

Abstract

The present application relates to an information transfer learning method, device and equipment based on deep conditional distribution alignment. The method includes: constructing a basic model for news authenticity identification, using a source domain news data set to conduct supervised training on the basic model to obtain the predicted output of the source domain news data; determining the pseudo-labels and high-confidence samples of the target domain based on the target domain news data set and the basic model; calculating the conditional maximum mean difference loss, cross entropy loss and mutual information loss based on the predicted output of the source domain news data, the pseudo-labels of the target domain and the source domain and target domain news data sets, and determining the total learning loss of the basic model; training the basic model based on the total learning loss; using the trained basic model to process the target domain news data samples to obtain the identification results of the target domain news data. This method enhances the robustness and stability of the training, and improves the stability and effectiveness of the model on the task of authenticity identification.

Description

Technical Field

[0001] The present application relates to the field of information analysis technology, and in particular to an information transfer learning method, device and equipment based on deep conditional distribution alignment. Background Art

[0002] In the rapidly developing digital age, information transfer learning (commonly referred to as transfer learning) plays a vital role in bridging the knowledge gap between different fields. The significance of information transfer learning lies in its ability to leverage the knowledge gained in the source domain (source dataset) and apply it to the target domain (target domain dataset), especially when labeled data in the target domain is scarce. By transferring knowledge between different domains, transfer learning can effectively improve the generalization ability of the model, improve prediction accuracy, and reduce dependence on target domain data, making the learning process more efficient and resource-saving.

[0003] Traditional transfer learning methods usually assume that the feature distributions between the source and target domains are similar. These methods are usually based on a supervised learning paradigm and fine-tune the pre-trained model on the target domain. Although these methods have achieved some success in many applications, they often perform poorly when faced with distribution differences between domains. This problem is often referred to as "domain shift" or "domain difference", which prevents the knowledge transferred from the source domain from being effectively applied to the target domain, thereby affecting the performance of the model in the target domain.

[0004] To address these limitations, a variety of transfer learning methods have emerged in recent years, aiming to reduce domain differences by aligning the feature distributions of the source and target domains. These methods include field adaptation techniques such as maximum mean difference (MMD), adversarial training, and correlation alignment, which focus on achieving knowledge transfer by minimizing the distribution difference between the source and target domains. Although these methods have been successful in some application scenarios, they ignore the conditional dependency between labels and features, which may lead to mode collapse or training instability.

[0005] This problem is particularly critical for news authenticity identification. False news often has unique language styles, dissemination patterns, and social backgrounds, and these factors are not fully reflected in independent label distributions. Therefore, if the model cannot effectively handle the conditional dependency between labels and features, it may not be able to accurately identify new or variant false news, resulting in an increase in the misjudgment rate. Therefore, the conditional dependency between labels and features must be considered to improve the stability and effectiveness of the model for the task of authenticity identification. Summary of the invention

[0006] Based on this, it is necessary to provide an information transfer learning method, device and equipment based on deep conditional distribution alignment that can realize model knowledge transfer in cross-domain complex scenarios to address the above technical problems.

[0007] An information transfer learning method based on deep conditional distribution alignment, the method comprising:

[0008] Obtain source domain news dataset and target domain news dataset for the news authenticity identification task.

[0009] A basic model for news authenticity identification is constructed, and the source domain news dataset is used to perform supervised training on the basic model to obtain the predicted output of the source domain news data.

[0010] According to the target domain news dataset and basic model, the pseudo labels and high confidence samples of the target domain are determined.

[0011] According to the source domain news dataset, the target domain news dataset, the predicted output of the source domain news data, and the pseudo-label of the target domain, the conditional maximum mean difference loss, cross entropy loss, and mutual information loss are calculated.

[0012] The total learning loss of the base model is determined based on the conditional maximum mean difference loss, cross entropy loss, and mutual information loss.

[0013] The base model is trained according to the total learning loss and the parameters of the base model are updated.

[0014] The basic model with updated parameters is used to identify the authenticity of target domain news data samples and obtain the identification results of target domain news data.

[0015] Among them, according to the target domain news dataset and the basic model, the pseudo labels and high confidence samples of the target domain are determined, including:

[0016] According to the target domain news dataset and the basic model, the pseudo labels of the samples are determined; the set of pseudo labels is:

[0017] ;

[0018] in, represents the set of pseudo labels, The target domain news dataset c The target domain i samples, is the output of the basic model; is the total number of samples in the target domain news dataset, c is the target domain ID, C is the total number of target domains.

[0019] For samples whose built-in confidence in the target domain is higher than a preset value, a pseudo label corresponding to the target domain is assigned.

[0020] In one embodiment, a basic model for news authenticity identification is constructed, and supervised training is performed on the basic model using a source domain news data set to obtain a predicted output of the source domain news data, including:

[0021] Construct a basic model for news authenticity identification; the basic model includes a feature extractor and a classification head; the classification head includes two fully connected layers.

[0022] The basic model is supervisedly trained using cross entropy loss based on the source domain news dataset to obtain the predicted output of the source domain news data.

[0023] In one embodiment, the process of calculating the conditional maximum mean difference loss includes:

[0024] The samples and corresponding labels in the source domain news dataset are mapped to the Hilbert space using a nonlinear mapping function to obtain the source domain conditional features.

[0025] The samples in the target source news dataset and the pseudo labels of the target domain are mapped to the Hilbert space using a nonlinear mapping function to obtain the target domain conditional features.

[0026] According to the conditional features of the source domain and the target domain, the conditional embeddings of the source domain and the target domain are determined.

[0027] According to the conditional embedding of the source domain and the target domain, the conditional maximum distribution difference loss is determined as:

[0028] ;

[0029] ;

[0030] ;

[0031] in, is the conditional maximum mean distribution difference loss, and Respectively and The conditional embedding of is the Hilbert space mapping of the source domain features, is the Hilbert space mapping of the source domain sample labels, represents the kernel function, is the positive constraint factor, is the Frobenius norm.

[0032] In one embodiment, the cross entropy loss is calculated based on the samples and corresponding labels in the source domain news dataset as follows:

[0033] ;

[0034] in, is the cross entropy loss, Represents the source domain news dataset i samples, represents the total number of samples in the source domain news dataset, Represents the source domain news dataset i The true label corresponding to each sample is The representation model is about the source domain news dataset i The predicted output for each sample.

[0035] In one embodiment, the mutual information loss is:

[0036] ;

[0037] in, is the mutual information loss, Represents the predicted label distribution The entropy of Represents the conditional label distribution based on the sample The entropy of For the i Output prediction of target domain news samples, For the i The entropy of the target domain news sample output, n is the total number of samples in the target domain.

[0038] In one embodiment, based on the conditional maximum mean difference loss, the cross entropy loss, and the mutual information loss, the total learning loss of the basic model is determined as:

[0039] ;

[0040] in, is the total learning loss of the base model, is the mutual information loss, is the cross entropy loss, is the conditional maximum mean distribution difference loss, , All are preset constants.

[0041] An information transfer learning device based on deep conditional distribution alignment, the device comprising:

[0042] The source domain and target domain dataset construction module is used to obtain the source domain news dataset and the target domain news dataset for the news authenticity identification task.

[0043] The basic model construction and training module is used to build a basic model for news authenticity identification. The basic model is supervised trained using the source domain news data set to obtain the predicted output of the source domain news data.

[0044] The target domain pseudo-label determination module is used to determine the pseudo-labels and high-confidence samples of the target domain based on the target domain news dataset and the basic model.

[0045] The loss function determination module is used to calculate the conditional maximum mean difference loss, cross entropy loss and mutual information loss based on the source domain news dataset, the target domain news dataset, the predicted output of the source domain news data and the pseudo-label of the target domain; based on the conditional maximum mean difference loss, cross entropy loss and mutual information loss, determine the total learning loss of the basic model.

[0046] The basic model parameter update module is used to train the basic model according to the total learning loss and update the parameters of the basic model.

[0047] The target domain news authenticity identification module is used to use the basic model with updated parameters to identify the authenticity of the target domain news data samples and obtain the target domain news data identification results.

[0048] The target domain pseudo-label determination module is also used to determine the pseudo-label of the sample according to the target domain news dataset and the basic model; the set of pseudo-labels is:

[0049] ;

[0050] in, represents the set of pseudo labels, The target domain news dataset c The target domain i samples, is the output of the basic model; is the total number of samples in the target domain news dataset, c is the target domain ID, C is the total number of target domains.

[0051] For samples whose built-in confidence in the target domain is higher than a preset value, a pseudo label corresponding to the target domain is assigned.

[0052] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any of the above methods when executing the computer program.

[0053] The above-mentioned information transfer learning method, device and equipment based on deep conditional distribution alignment include: constructing a basic model for news authenticity identification, using a source domain news data set to perform supervised training on the basic model to obtain the predicted output of the source domain news data; determining the pseudo-labels and high-confidence samples of the target domain based on the target domain news data set and the basic model; calculating the conditional maximum mean difference loss, cross entropy loss and mutual information loss based on the source domain news data set, the target domain news data set, the predicted output of the source domain news data and the pseudo-labels of the target domain, and determining the total learning loss of the basic model; training the basic model based on the total learning loss; using the trained basic model to process the target domain news data samples to obtain the identification results of the target domain news data. This method enhances the robustness and stability of training, and improves the stability and effectiveness of the model on the task of authenticity identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 It is a flowchart of an information transfer learning method based on deep conditional distribution alignment in one embodiment;

[0055] Figure 2 A basic architecture diagram of an information transfer learning method based on deep conditional distribution alignment in one embodiment;

[0056] Figure 3 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0058] In one embodiment, Figure 1 As shown, an information transfer learning method based on deep conditional distribution alignment is provided, and the method includes the following steps:

[0059] Step 100: Obtain a source domain news dataset and a target domain news dataset for the news authenticity identification task.

[0060] Specifically, the news authenticity identification task in this method can also be any other task. The news authenticity identification task refers to inferring the authenticity of news given news and related evidence. This embodiment uses relevant open source news datasets as source domain and target domain news datasets, and represents the source domain news dataset as , the target domain news dataset is represented as ; Among them, the data format of the samples in the source domain and target domain news datasets is text data.

[0061] Step 102: Construct a basic model for news authenticity identification, use the source domain news data set to perform supervised training on the basic model, and obtain the predicted output of the source domain news data.

[0062] Specifically, this method is not limited by the basic model architecture and is applicable to all deep learning models. In a specific embodiment, the basic model uses BERT as the feature extractor f(.), and a two-layer fully connected network as the classification head h(.), and outputs binary inference for a given input, that is, 0 is false and 1 is true. Then the source domain data is input for supervised training.

[0063] Step 104: Determine the pseudo labels and high confidence samples of the target domain based on the target domain news dataset and the basic model.

[0064] Step 104 specifically includes: determining the pseudo labels of the samples according to the target domain news dataset and the basic model; the set of pseudo labels is:

[0065] ;

[0066] in, represents the set of pseudo labels, The target domain news dataset c The target domain i samples, is the output of the basic model; is the total number of samples in the target domain news dataset, c is the target domain ID, C is the total number of target domains.

[0067] For samples whose built-in confidence in the target domain is higher than a preset value, a pseudo label corresponding to the target domain is assigned.

[0068] Specifically, the basic model is called, the target domain data is input to obtain the pseudo label of the sample, and high-confidence samples are sampled according to the model output.

[0069] Step 106: Calculate the conditional maximum mean difference loss, cross entropy loss, and mutual information loss based on the source domain news dataset, the target domain news dataset, the predicted output of the source domain news data, and the pseudo-label of the target domain.

[0070] Specifically, we first perform conditional distribution alignment based on the conditional maximum mean difference (CMMD); at the same time, in order to ensure the basic capabilities of the basic model, we need to ensure that the basic model's ability to identify the source domain data set is not damaged, so we use the source domain data set and the corresponding labels to calculate the cross entropy loss; in addition, we also need to consider the discriminability of the features; in the field of probability theory and informatics, the mutual information of X and Y can be used to measure how much information Y contains in the X variable. Therefore, by maximizing the target domain features Its corresponding pseudo label The mutual information loss is proposed to enhance the discriminability of features.

[0071] Step 108: Determine the total learning loss of the basic model based on the conditional maximum mean difference loss, the cross entropy loss, and the mutual information loss.

[0072] Specifically, according to the above conditions, the maximum mean difference loss, cross entropy loss and mutual information loss, the total learning loss of the model can be obtained, and the model parameters can be trained according to the generated gradients.

[0073] Step 110: Train the basic model according to the total learning loss and update the parameters of the basic model.

[0074] Specifically, the basic model is trained according to the total learning loss, the parameters of the basic model are updated, and the data to be tested are input for model output. In this embodiment, since it is a task of distinguishing true from false news, the maximum value index of the output is taken as the final output, with 0 for false and 1 for true.

[0075] Step 112: Use the basic model with updated parameters to identify the target domain news data samples to obtain the target domain news data identification results.

[0076] In the implementation process, the open source rumor dataset Snopes is selected as the source domain news dataset, and the open source news dataset Politifact is selected as the target domain news dataset. The model architecture is designed with BERT as the feature extractor f(.), and a two-layer fully connected network is superimposed as the classification head h(.). First, supervised training is performed using the source domain news data and labels through cross entropy loss, and the model weights are saved. Then the model weights are read, and the conditional maximum mean difference loss, cross entropy loss, and mutual information loss are calculated using the target domain news data to update the model. Finally, according to the maximum index of the logical output, inference is made. If it is 0, the news is judged to be false, and if it is 1, the news is judged to be true. Through this implementation process, the corresponding labels of the target domain data are no longer required, thereby achieving the purpose of information migration.

[0077] In the above-mentioned information transfer learning method based on deep conditional distribution alignment, the method includes: constructing a basic model for news authenticity identification, using the source domain news data set to conduct supervised training on the basic model to obtain the predicted output of the source domain news data; determining the pseudo labels and high confidence samples of the target domain according to the target domain news data set and the basic model; calculating the conditional maximum mean difference loss, cross entropy loss and mutual information loss according to the source domain news data set, the target domain news data set, the predicted output of the source domain news data and the pseudo labels of the target domain, and determining the total learning loss of the basic model; training the basic model according to the total learning loss; using the trained basic model to process the target domain news data samples to obtain the identification results of the target domain news data. This method enhances the robustness and stability of training, and improves the stability and effectiveness of the model on the task of authenticity identification.

[0078] In one embodiment, step 102 includes: constructing a basic model for news authenticity identification; the basic model includes a feature extractor and a classification head; the classification head includes two fully connected layers; and supervised training is performed on the basic model using a cross entropy loss based on a source domain news data set to obtain a predicted output of the source domain news data.

[0079] In one embodiment, the process of calculating the conditional maximum mean difference loss in step 106 includes:

[0080] The samples and corresponding labels in the source domain news dataset are mapped to the Hilbert space using a nonlinear mapping function to obtain the source domain conditional features; the samples in the target source news dataset and the pseudo labels of the target domain are mapped to the Hilbert space using a nonlinear mapping function to obtain the target domain conditional features; according to the source domain conditional features and the target domain conditional features, the conditional embedding of the source domain and the target domain is determined; according to the conditional embedding of the source domain and the target domain, the conditional maximum distribution difference loss is determined as:

[0081] ;

[0082] ;

[0083] ;

[0084] in, is the conditional maximum mean distribution difference loss, and Respectively and The conditional embedding of is the Hilbert space mapping of the source domain features, is the Hilbert space mapping of the source domain sample labels, represents the kernel function, is the positive constraint factor, is the Frobenius norm.

[0085] In one embodiment, the cross entropy loss is calculated based on the samples and corresponding labels in the source domain news dataset as follows:

[0086] ;

[0087] in, is the cross entropy loss, Represents the source domain news dataset i samples, represents the total number of samples in the source domain news dataset, Represents the source domain news dataset i The true label corresponding to each sample is The representation model is about the source domain news dataset i The predicted output for each sample.

[0088] In one embodiment, the mutual information loss in step 106 is:

[0089] ;

[0090] in, is the mutual information loss, Represents the predicted label distribution The entropy of Represents the conditional label distribution based on the sample The entropy of For the i Output prediction of target domain news samples, For the i The entropy of the target domain news sample output, n is the total number of samples in the target domain.

[0091] Specifically, we first use the conditional maximum mean difference (CMMD) to perform conditional distribution alignment. To enhance the domain invariance of features, we first map the label and sample features to the Hilbert kernel space. This nonlinear mapping is used to and Representation. Given a sample feature Z and the corresponding label y, the conditional kernel mean embedding can be expressed as:

[0092] ;

[0093] in, express The conditional embedding of , and They represent the cross-covariance operators between the label and the representation and the label itself, respectively.

[0094] Based on the above discussion, the conditional distribution embedding of the source domain and the target domain can be expressed as:

[0095] ;

[0096] ;

[0097] in, and Respectively and Conditional embedding of . In order to calculate and In the Hilbert space example, we can calculate by fixing y and The distance between the source domain and the target domain can be reduced by minimizing the CMMD loss.

[0098] In practice, it is necessary to evaluate the CMMD loss in each batch. In order to maintain the stability of the evaluation, we randomly select and Random sampling is performed from and Sampling is performed. The actual estimate is: .

[0099] about The actual estimate of The similarity of , then the actual estimation loss of CMMD can be rewritten as:

[0100] ;

[0101] in, , , , , , , , .

[0102] The label information in the target domain is necessary for estimating the CMMD difference. Similar to some domain adaptation methods, this method generates pseudo labels for samples with higher confidence in the target domain. In order to improve the representation ability of the model and the accuracy of pseudo labels, this embodiment uses BERT as a feature extractor and a fully connected network with a softmax activation function as a classifier.

[0103] At the same time, in order to ensure the basic capabilities of the basic model, it is necessary to ensure that the identification ability of the basic model in the source domain is not destroyed. Therefore, the cross entropy loss is calculated using the above cross entropy loss expression based on the source domain data and the corresponding labels.

[0104] In addition, the discriminability of features should also be considered. In the field of probability theory and informatics, the mutual information of X and Y can be used to measure how much information Y contains in the X variable. Therefore, the target domain features can be maximized by selecting high-confidence samples on the target domain, that is, samples with a maximum confidence category greater than 0.9. Its corresponding pseudo label The mutual information of is used to enhance the discriminability of features, so the proposed mutual information loss can be expressed as:

[0105] ;

[0106] in, Represents the predicted label distribution The entropy of Represents the conditional label distribution based on the sample In actual operation, because the data arrives in batches, that is, it is sampled from the target domain to obtain the corresponding classification probability , so the mutual information loss can be rewritten as shown in the above expression of mutual information loss.

[0107] In one embodiment, step 108 includes: determining the total learning loss of the basic model according to the conditional maximum mean difference loss, the cross entropy loss, and the mutual information loss as:

[0108] ;

[0109] in, is the total learning loss of the base model, is the mutual information loss, is the cross entropy loss, is the conditional maximum mean distribution difference loss, , All are preset constants, as the preferred choice, Set to 0.10, Set to 0.20.

[0110] In one embodiment, the target task scenario in step 100 is a news authenticity identification task scenario.

[0111] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1 At least part 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 part of the sub-steps or stages of other steps.

[0112] In one embodiment, an information transfer learning device based on deep conditional distribution alignment is provided, comprising: a source domain and a target domain dataset construction module, a target domain pseudo-label determination module, a loss function determination module, a basic model parameter update module, and a target domain news authenticity identification module, wherein:

[0113] Source domain and target domain dataset construction module, used to obtain source domain news dataset and target domain news dataset for news authenticity identification task;

[0114] The basic model construction and training module is used to construct a basic model for news authenticity identification, and use the source domain news data set to conduct supervised training on the basic model to obtain the predicted output of the source domain news data;

[0115] A target domain pseudo-label determination module is used to determine the pseudo-labels and high-confidence samples of the target domain based on the target domain news dataset and the basic model;

[0116] A loss function determination module is used to calculate the conditional maximum mean difference loss, cross entropy loss and mutual information loss based on the source domain news data set, the target domain news data set, the predicted output of the source domain news data and the pseudo-label of the target domain; and determine the total learning loss of the basic model based on the conditional maximum mean difference loss, cross entropy loss and mutual information loss;

[0117] The basic model parameter update module is used to train the basic model according to the total learning loss and update the parameters of the basic model.

[0118] The target domain news authenticity identification module is used to use the basic model with updated parameters to identify the authenticity of the target domain news data samples and obtain the target domain news data identification results.

[0119] Among them, the target domain pseudo-label determination module is also used to determine the pseudo-label of the sample according to the target domain news data set and the basic model; the set of pseudo-labels is shown in the expression of the above-mentioned set of pseudo-labels; and the sample whose built-in confidence in the target domain is higher than the preset value is assigned the pseudo-label corresponding to the target domain.

[0120] In one of the embodiments, the basic model construction and training module is also used to construct a basic model for news authenticity identification; the basic model includes a feature extractor and a classification head; the classification head includes two fully connected layers; the basic model is supervised trained using cross entropy loss based on the source domain news data set to obtain the predicted output of the source domain news data.

[0121] In one of the embodiments, the process of calculating the conditional maximum mean difference loss in the loss function determination module includes: mapping the samples and corresponding labels in the source domain news dataset to the Hilbert space using a nonlinear mapping function to obtain the source domain conditional features; mapping the samples in the target source news dataset and the pseudo labels of the target domain to the Hilbert space using a nonlinear mapping function to obtain the target domain conditional features; determining the conditional embedding of the source domain and the target domain based on the source domain conditional features and the target domain conditional features; determining the conditional maximum distribution difference loss based on the conditional embedding of the source domain and the target domain; the conditional maximum distribution difference loss is as shown in the expression of the conditional maximum distribution difference loss mentioned above.

[0122] In one of the embodiments, the loss function determination module is further used to calculate the cross entropy loss using the above-mentioned cross entropy loss expression according to the samples and corresponding labels in the source domain news dataset.

[0123] In one of the embodiments, the mutual information loss in the loss function determination module is as shown in the above expression of mutual information loss.

[0124] In one embodiment, the loss function determination module is further used to determine the total learning loss as shown in the expression of the total learning loss of the above basic model based on the conditional maximum mean difference loss, the cross entropy loss and the mutual information loss.

[0125] For the specific definition of the information transfer learning device based on deep conditional distribution alignment, please refer to the definition of the information transfer learning method based on deep conditional distribution alignment above, which will not be repeated here. Each module in the above-mentioned information transfer learning device based on deep conditional distribution alignment can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0126] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 3 As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, an information transfer learning method based on deep conditional distribution alignment is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covered on the display screen, or a button, trackball or touchpad set on the computer device housing, or an external keyboard, touchpad or mouse, etc.

[0127] Those skilled in the art will understand that Figure 3 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0128] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps in the above method embodiment when executing the computer program.

[0129] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, 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.

[0130] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. An information transfer learning method based on deep conditional distribution alignment, characterized in that: The method comprises: Obtain a source domain news dataset and a target domain news dataset for the news authenticity identification task; the data format of the samples in the source domain and target domain news datasets are both text data; Constructing a basic model for news authenticity identification, and using the source domain news data set to perform supervised training on the basic model to obtain a predicted output of the source domain news data; Determine pseudo labels and high-confidence samples of the target domain according to the target domain news dataset and the basic model; Calculate the conditional maximum mean difference loss, cross entropy loss and mutual information loss according to the source domain news dataset, the target domain news dataset, the predicted output of the source domain news data and the pseudo label of the target domain; Determining a total learning loss of a basic model according to the conditional maximum mean difference loss, the cross entropy loss, and the mutual information loss; Training the basic model according to the total learning loss and updating the parameters of the basic model; The basic model with updated parameters is used to identify the target domain news data samples to obtain the target domain news data identification results; Wherein, determining the pseudo labels and high-confidence samples of the target domain according to the target domain news dataset and the basic model includes: According to the target domain news dataset and the basic model, the pseudo labels of the samples are determined; the set of pseudo labels is: in, represents the set of pseudo labels, The target domain news dataset c The target domain i samples, is the output of the basic model; is the total number of samples in the target domain news dataset, c is the target domain ID, C is the total number of target domains; For samples whose built-in confidence in the target domain is higher than the preset value, a pseudo label corresponding to the target domain is assigned; Among them, the process of calculating the conditional maximum mean difference loss includes: The samples and corresponding labels in the source domain news dataset are mapped to the Hilbert space using a nonlinear mapping function to obtain the source domain conditional features; The samples in the target source news dataset and the pseudo labels of the target domain are mapped to the Hilbert space using a nonlinear mapping function to obtain the target domain conditional features; Determining conditional embedding of the source domain and the target domain according to the source domain conditional features and the target domain conditional features; According to the conditional embedding of the source domain and the target domain, the conditional maximum distribution difference loss is determined as: in, is the conditional maximum mean distribution difference loss, and Respectively and The conditional embedding of is the Hilbert space mapping of the source domain features, is the Hilbert space mapping of the source domain sample labels, represents the kernel function, is the positive constraint factor, is the Frobenius norm.

2. The information transfer learning method based on deep conditional distribution alignment according to claim 1 is characterized in that: Constructing a basic model for news authenticity identification, using the source domain news data set to perform supervised training on the basic model, and obtaining a predicted output of the source domain news data, including: Constructing a basic model for news authenticity identification; the basic model includes a feature extractor and a classification head; the classification head includes two fully connected layers; The basic model is supervisedly trained using cross entropy loss according to the source domain news data set to obtain a predicted output of the source domain news data.

3. The information transfer learning method based on deep conditional distribution alignment according to claim 1 is characterized in that: The cross entropy loss is calculated based on the samples and corresponding labels in the source domain news dataset: in, is the cross entropy loss, Represents the source domain news dataset i samples, represents the total number of samples in the source domain news dataset, Represents the source domain news dataset i The true label corresponding to each sample is The representation model is about the source domain news dataset i The predicted output for each sample.

4. The information transfer learning method based on deep conditional distribution alignment according to claim 1, characterized in that: The mutual information loss is: in, is the mutual information loss, Represents the predicted label distribution The entropy of Represents the conditional label distribution based on the sample The entropy of For the i Output prediction of target domain news samples, For the i The entropy of the target domain news sample output, n is the total number of samples in the target domain.

5. The information transfer learning method based on deep conditional distribution alignment according to claim 1, characterized in that: According to the conditional maximum mean difference loss, the cross entropy loss and the mutual information loss, the total learning loss of the basic model is determined as: in, is the total learning loss of the base model, is the mutual information loss, is the cross entropy loss, is the conditional maximum mean distribution difference loss, , All are preset constants.

6. An information transfer learning device based on deep conditional distribution alignment, characterized in that: The device comprises: The source domain and target domain dataset construction module is used to obtain the source domain news dataset and the target domain news dataset for the news authenticity identification task; the data format of the samples in the source domain and the target domain news dataset are both text data; A basic model construction and training module is used to construct a basic model for news authenticity identification, and to use the source domain news data set to perform supervised training on the basic model to obtain a predicted output of the source domain news data; A target domain pseudo-label determination module, used to determine the pseudo-labels and high-confidence samples of the target domain according to the target domain news dataset and the basic model; A loss function determination module, used to calculate the conditional maximum mean difference loss, the cross entropy loss and the mutual information loss according to the source domain news data set, the target domain news data set, the predicted output of the source domain news data and the pseudo-label of the target domain; determine the total learning loss of the basic model according to the conditional maximum mean difference loss, the cross entropy loss and the mutual information loss; A basic model parameter updating module, used to train the basic model according to the total learning loss and update the parameters of the basic model; The target domain news authenticity identification module is used to identify the authenticity of target domain news data samples using the basic model with updated parameters to obtain the target domain news data identification results; The target domain pseudo-label determination module is further used to determine the pseudo-label of the sample according to the target domain news dataset and the basic model; the set of pseudo-labels is: in, represents the set of pseudo labels, The target domain news dataset c The target domain i samples, is the output of the basic model; is the total number of samples in the target domain news dataset, c is the target domain ID, C is the total number of target domains; For samples whose built-in confidence in the target domain is higher than the preset value, a pseudo label corresponding to the target domain is assigned; Among them, the process of calculating the conditional maximum mean difference loss in the loss function determination module includes: mapping the samples and corresponding labels in the source domain news data set to the Hilbert space using a nonlinear mapping function to obtain the source domain conditional features; mapping the samples in the target source news data set and the pseudo labels of the target domain to the Hilbert space using a nonlinear mapping function to obtain the target domain conditional features; determining the conditional embedding of the source domain and the target domain according to the source domain conditional features and the target domain conditional features; determining the conditional maximum distribution difference loss according to the conditional embedding of the source domain and the target domain as follows: in, is the conditional maximum mean distribution difference loss, and Respectively and The conditional embedding of is the Hilbert space mapping of the source domain features, is the Hilbert space mapping of the source domain sample labels, represents the kernel function, is the positive constraint factor, is the Frobenius norm.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the information transfer learning method based on deep conditional distribution alignment described in any one of claims 1 to 5 is implemented.

Images