A small sample event false information detection method based on bias decoupling
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHONGQING UNIV OF POSTS & TELECOMM
- Filing Date
- 2024-01-29
- Publication Date
- 2026-07-24
AI Technical Summary
Existing technologies struggle to effectively identify and detect misinformation in small sample events on social media platforms, especially when data volume is limited and feature distribution varies greatly, making identification difficult and preventing the establishment of a general model.
A bias-decoupled few-sample event misinformation detection method is adopted. Through a two-stage training process, AdaIN perturbation and hierarchical discriminator design, multi-granular text feature extraction and decoupling of event content are performed. The BERT model and multilayer perceptron are used for feature processing. Combined with adaptive instance normalization and cross-perturbation mechanism, the adaptability and detection accuracy of the model are improved.
It improves the model's adaptability and detection accuracy in the face of emergencies, optimizes the false information detection task in small sample scenarios, and enhances the model's robustness and generalization ability.
Smart Images
Figure CN117951299B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of false information detection technology, specifically relating to a method for detecting false information in small sample events based on bias decoupling. Background Technology
[0002] False events are defined as events lacking substantial factual basis, deliberately fabricated, exaggerated, or distorted, encompassing multiple aspects of information. On social media platforms, some users may report unverified events or exaggerate or distort facts to create sensationalism, thereby misleading the public. The widespread dissemination of false events can confuse and manipulate public opinion, altering how people react to the truth and causing serious negative impacts on individuals and society.
[0003] To improve the credibility of information on online social networks and mitigate the negative impact of the spread of misinformation, experts from industry and academia have begun actively exploring effective strategies for combating misinformation on social media. For example, Sina Weibo users can report potentially misleading posts to the Weibo Community Management Center, where professional fact-checkers then verify the post's authenticity. Facebook encourages users to actively flag suspicious or unusual tweets containing misinformation. Furthermore, the Cyberspace Administration of China's Illegal and Harmful Information Reporting Center has established the China Internet Joint Rumor-Refuting Platform to verify misinformation circulating on social networks. Well-known international fact-checking websites (FactCheck.org and PolitiFact.com) rely on a team of professional fact-checkers for manual verification, thereby improving the accuracy of their findings.
[0004] The methods described above primarily rely on manual review to verify social media data. However, in today's information overload, academia is increasingly focusing on exploring automated methods for judging the authenticity of events. Early researchers in the field of misinformation detection mainly concentrated on manually designed features, such as statistical features, topic features, word features, and syntactic features. Subsequently, they used traditional machine learning algorithms, such as Naive Bayes, Support Vector Machines, and Decision Trees, to classify misinformation. With the rapid development of artificial intelligence technology in both depth and breadth, technologies such as natural language processing, machine learning, and deep learning have enabled the automatic analysis and processing of large amounts of text and images. In recent years, researchers have increasingly adopted deep learning techniques for feature extraction, marking continuous innovation and improvement in misinformation detection methods.
[0005] However, as the methods of fabricating misinformation become increasingly sophisticated, more and more misinformation circulating on social media is exhibiting characteristics of small-sample events. This presents new challenges to the field of misinformation detection: First, the number of descriptions of the same event on social media platforms is limited, lacking sufficient evidence to support its authenticity. This scarcity of information renders mutual verification baseless, making identification difficult. Second, the data feature distributions among different small-sample events vary significantly, making it impossible to establish a universal model to summarize the commonalities of most misinformation. Summary of the Invention
[0006] To address the shortcomings of existing technologies, this invention proposes a method for detecting misinformation in few-shot events based on bias decoupling. The method includes: acquiring the text describing the content of the event to be detected and inputting it into a trained model for detecting misinformation in few-shot events to obtain the detection result of misinformation in few-shot events.
[0007] The training process for the small sample event misinformation detection model includes:
[0008] S1: Obtain the labeled description text of the event to be detected and the description text of the random event;
[0009] S2: Perform multi-granular text feature extraction on the text describing the content of the event to be detected and the text describing the content of random events to obtain multi-view context representations of the two types of text;
[0010] S3: Decouple the two types of text multi-view context representations to obtain the uncoupling features to be detected and random uncoupling features;
[0011] S4: Adaptive instance normalization is used to perturb the decoupled features to be detected and the random decoupled features. A discriminator is used to process the perturbation results to obtain the false information detection results for small sample events.
[0012] S5: Calculate the first-stage loss based on the false information detection results of small sample events;
[0013] S6: Freeze the discriminator parameters, input the decoupling feature to be detected into the discriminator for processing, and calculate the effective separation loss based on the discriminator output;
[0014] S7: Adjust the parameters of the model other than the discriminator based on the effective separation loss to obtain the minimum effective separation loss; calculate the second-stage loss based on the effective separation loss and the minimum effective separation loss; adjust the model parameters based on the second-stage loss to obtain the trained small sample event false information detection model.
[0015] Preferably, the process of obtaining a multi-view context representation includes:
[0016] S21: Use the BERT model to extract the global and local representations of the descriptive text respectively;
[0017] S22: Compute the global-to-local consistent representation and the local-to-local consistent representation;
[0018] S23: The attention aggregation mechanism is used to process the global representation, the consistent representation from global to local, and the consistent representation from local to local to obtain the multi-view context representation.
[0019] Preferably, in step S3, the multi-view context representations of the two types of text are processed to obtain the embedded features to be detected and the random embedded features, specifically including:
[0020] Two independent multilayer perceptrons are used to process the multi-view context representations of two types of text to obtain the detection embedding features and random embedding features. The detection embedding features include the detection content-related embeddings and the detection event-related embeddings, and the random embedding features include random content-related embeddings and random event-related embeddings.
[0021] Furthermore, in step S4, the process of perturbing the embedded features to be detected and the random embedded features using adaptive instance normalization includes:
[0022] S41: Adaptive instance normalization is used to perturb the content-related embedding and random event-related embedding to obtain content feature perturbation;
[0023] S42: Adaptive instance normalization is used to perturb the random content-related embeddings and the event-related embeddings to be detected, resulting in event feature perturbations.
[0024] Furthermore, the formulas for content feature perturbation and event feature perturbation are obtained as follows:
[0025]
[0026]
[0027] in, This indicates a perturbation of content features. This represents the event feature perturbation; α and β represent the first and second weight parameters, respectively; and C represents the relevant embedding of the content to be detected. E represents the embedding related to random content, and E represents the embedding related to the event to be detected. This indicates a random event-related embedding, and AdaIN indicates adaptive instance normalization.
[0028] Preferably, the process of calculating the first-stage loss includes:
[0029] The loss for content authenticity classification and the loss for event classification are calculated based on the results of small sample event false information detection and the real labels; the loss for relevance is calculated based on the embedded features to be detected; and the loss for content authenticity classification, the loss for event classification, and the loss for relevance are weighted and summed to obtain the first stage loss.
[0030] Preferably, the formula for calculating the effective separation loss is:
[0031]
[0032] in, Indicates effective separation loss. D represents the i-th true / false label sampled from a uniform distribution. c (E i () represents the predicted value of the i-th category output by the content authenticity detector, K represents the number of authenticity tag categories, and D represents the number of event tag categories. D represents the label of the i-th event sampled from a uniform distribution. e (C i ) represents the predicted value of the i-th category output by the event discriminator.
[0033] Preferably, the formula for calculating the second-stage loss is:
[0034]
[0035] in, This represents the second-stage loss, where λ represents the adjustment coefficient. Indicates effective separation loss. This represents the minimum effective separation loss.
[0036] The beneficial effects of this invention are as follows: This invention proposes a small-sample event misinformation detection method based on bias decoupling. By employing techniques such as a two-stage training process, AdaIN perturbation, and hierarchical discriminator design to decouple event information, it improves the model's adaptability to different data distributions when dealing with sudden events. The model's learning process is optimized, thus more effectively handling misinformation detection tasks in small-sample scenarios, improving the model's detection accuracy, and demonstrating promising application prospects. Attached Figure Description
[0037] Figure 1 This is a schematic diagram of the training process of the small sample event false information detection model in this invention. Detailed Implementation
[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0039] This invention proposes a method for detecting misinformation in small sample events based on bias decoupling, such as... Figure 1 As shown, the method includes the following: obtaining the text describing the content of the event to be detected and inputting it into a trained small sample event misinformation detection model to obtain the misinformation detection result; wherein the small sample event misinformation detection result includes the judgment result of the authenticity of the content and the judgment result of the domain to which the event belongs.
[0040] The training process for the small sample event misinformation detection model includes:
[0041] S1: Obtain the labeled description text of the event to be detected and the description text of the random event.
[0042] Obtain labeled event content description text, and randomly select one sample from a large number of event content description text samples as random event content description text.
[0043] S2: Perform multi-granular text feature extraction on the text describing the content of the event to be detected and the text describing the content of random events to obtain multi-view context representations of the two types of text.
[0044] S21: Use the BERT model to extract the global and local representations of the descriptive text.
[0045] To capture the rich information in event descriptions, the BERT model is used to extract multi-granularity text features from the event content description text to obtain multi-level semantic representations. Both the event content description text to be detected and the random event content description text are extracted to obtain a global representation T for each type of description text. g and local representation T l .
[0046] S22: Compute the global-to-local consistent representation and the local-to-local consistent representation.
[0047] Introducing global to local (T) g→l ) and local to local (T) l→l This method measures the consistency between representations to enhance the analytical capabilities for interactions between semantic information at different granularities of text encoding. The formulas for calculating global-to-local and local-to-local consistent representations are as follows:
[0048]
[0049]
[0050] Among them, T g→l T represents the consistent global-to-local representation under H multi-head attention. l→l This represents a local-to-local consistent representation, where L represents the maximum sentence length, H represents the number of multi-head attention, and T represents the number of attention points. g Indicates global representation; T l The local representation is an L-row matrix; d r σ represents the scaling factor, and σ represents the activation function.
[0051] S23: The attention aggregation mechanism is used to process the global representation, the consistent representation from global to local, and the consistent representation from local to local to obtain the multi-view context representation.
[0052] Three representations of T g T g→l and T l→l All of these aim to capture comprehensive knowledge at multiple granularities. An attention aggregation mechanism is employed to compute three different vectors T. g T g→l and T l→l The L1 norm is used to determine the distribution of the attention weights {a} through a softmax operation. g a g→l a l→l The cumulative multi-view context representation O is obtained by weighting these weights.
[0053] O = a g T g +a g→l T g→l +a l→l T l→l
[0054] Through the above process, two multi-view context representations of descriptive text are obtained.
[0055] S3: Decouple the two types of text multi-view context representations to obtain the uncoupling features to be detected and random uncoupling features.
[0056] Two independent multilayer perceptrons (MLPs) are used to process the multi-view context representations of the two types of text to obtain the embedded features to be detected and the random embedded features. The processing can be represented as follows:
[0057] C = FC(O)
[0058] E = FE(O)
[0059] The detected embedding features include content-related embeddings C and event-related embeddings E, and the random embedding features include random content-related embeddings. Embedding related to random events
[0060] S4: Adaptive instance normalization is used to perturb the embedded features to be detected and the random embedded features. A discriminator is used to process the perturbation results to obtain the false information detection results.
[0061] A cross-perturbation mechanism is introduced, employing Adaptive Instance Normalization (AdaIN) to perturb the features, thereby decoupling the initial features; specifically:
[0062] S41: Adaptive instance normalization is used to perturb the embeddings of the content to be detected and the embeddings of random events to obtain content feature perturbations; specifically: AdaIN is used to fuse the embeddings of the content to be detected and the embeddings of random events, and then combined with the embeddings of the content to be detected through weighted summation, which can be expressed as:
[0063]
[0064] S42: Adaptive instance normalization is used to perturb the random content-related embeddings and the event-related embeddings to be detected, resulting in event feature perturbations; this can be expressed as:
[0065]
[0066] in, This indicates a perturbation of content features. The event characteristic perturbation is represented by α and β, which represent the first and second weighting parameters, respectively, both derived from a uniform distribution. Obtained by sampling.
[0067] To enhance the decoupling between content and events, this invention introduces a CrossPerturbation (CP) mechanism. This mechanism facilitates the joint learning of features from both by introducing event perturbations into the content representation and content perturbations into the event representation. Event perturbations allow the content authenticity classifier to focus more on content attributes, isolating them from event-biased information. Simultaneously, content perturbations enhance the event classifier's ability to capture event features, resulting in more accurate domain-specific predictions.
[0068] The perturbation results are then processed by a discriminator to obtain the authenticity detection results of the content. Specifically, the perturbation of the content features injected with event noise is input into the content authenticity discriminator to obtain the detection results of whether the small sample event is true or not. The perturbation of the event features injected with content noise is input into the event discriminator to obtain the probability of the small sample event belonging to each domain. The content authenticity discriminator and the event discriminator are both simple two-layer multilayer perceptrons (MLP).
[0069] S5: Calculate the first-stage loss based on the false information detection results of small sample events.
[0070] Based on the results of small-sample event misinformation detection and the real labels, calculate the content authenticity classification loss and the event classification loss; the calculation formula is expressed as follows:
[0071]
[0072]
[0073] in, This indicates the loss in content authenticity classification. Let K represent the event classification loss, D represent the number of true / false labels, and y represent the number of event classification labels. i Let d represent the i-th true / false label. i This represents the i-th event label. This represents the i-th predicted value output by the content authenticity discriminator, representing the perturbation of the input content features. This represents the predicted value output by the event discriminator, which reflects the perturbation of the input event features.
[0074] Calculate the correlation loss based on the embedded features to be detected. Specifically: The Pearson correlation between the content-related embedding C and the event-related embedding E is calculated as an auxiliary indicator to measure their independence. The convergence of the correlation loss indicates that the content features and event features are statistically independent.
[0075]
[0076] The first-stage loss is obtained by weighted summing of the content authenticity classification loss, event classification loss, and relevance loss; expressed as:
[0077]
[0078] in, γ represents the first-stage loss; γ represents the relevance weight coefficient, which is a hyperparameter used to control the contribution of the independence between content representation and event representation during training.
[0079] S6: Freeze the discriminator parameters, input the embedded features to be detected into the discriminator for processing, and calculate the effective separation loss based on the discriminator output.
[0080] To reduce the interrelationship between content features and event features, the content and event discriminators are frozen, and the content features and event features are fed into opposite discriminators (the content-related embedding C to be detected is fed into the event discriminator, and the event-related embedding E to be detected is fed into the content authenticity discriminator), so as to achieve effective separation of content and event representation.
[0081] Calculate the effective separation loss based on the discriminator output:
[0082]
[0083] in, Indicates effective separation loss. D represents the i-th true / false label sampled from a uniform distribution. c (E i () represents the input event-related embedding, the predicted value of the i-th category output by the content authenticity discriminator, K represents the number of authenticity label categories, and D represents the number of event label categories. D represents the label of the i-th event sampled from a uniform distribution. e (C i ) represents the predicted value of the i-th category output by the event discriminator, which is the relevant embedding of the input content to be detected.
[0084] This invention optimizes the decoupled feature representation by adjusting the model architecture, ensuring that the content authenticity detector focuses more on distinguishing between true and false information, while the event detector focuses more on accurately determining the event domain. This separation helps strengthen the model's independent understanding of content and events, improving the model's performance robustness across both tasks.
[0085] S7: Adjust the parameters of the model other than the discriminator based on the effective separation loss to obtain the minimum effective separation loss; calculate the second-stage loss based on the effective separation loss and the minimum effective separation loss; adjust the model parameters based on the second-stage loss to obtain the trained small sample event false information detection model.
[0086] Fine-tune the rest of the model to minimize in The cross-entropy between task prediction (fake information detection / domain detection) and uniform distribution is determined, specifically expressed as:
[0087]
[0088] The second-stage loss is calculated based on the effective separation loss and the minimum effective separation loss, and is expressed as follows:
[0089]
[0090] in, This indicates the losses in the second phase. Indicates effective separation loss. λ represents the minimum effective separation loss; λ represents the adjustment coefficient, which is a hyperparameter for the trade-off of the adjustment loss term.
[0091] The goal of the second stage of training is to refine the decoupled representation of features by filtering out potential overlap between content features and event features; by adjusting the model parameters according to the second stage loss, a well-trained small-sample event misinformation detection model can be obtained.
[0092] The text describing the content of the event to be detected is obtained and input into a trained few-sample event misinformation detection model. After obtaining the embedded features to be detected in the model, the model is processed by a discriminator and outputs the content authenticity detection and domain detection results of the few-sample event.
[0093] In summary, this invention improves the robustness and generalization of the method by employing a two-stage training process, AdaIN perturbation, and a hierarchical discriminator design. During training, the AdaIN perturbation is introduced to enhance the model's adaptability to different data distributions. Simultaneously, the hierarchical discriminator design refines the capture of content bias and domain bias, effectively decoupling information bias in small-sample events. Because this invention focuses on decoupling event bias, the model is not limited to data from a specific domain during training, thus expanding the model's application scenarios.
[0094] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for detecting false information in small sample events based on bias decoupling, characterized in that, include: Obtain the description text of the event to be detected and input it into the trained few-shot event misinformation detection model to obtain the few-shot event misinformation detection results; The training process for the small sample event misinformation detection model includes: S1: Obtain the labeled description text of the event to be detected and the description text of the random event; S2: Perform multi-granular text feature extraction on the text describing the content of the event to be detected and the text describing the content of random events to obtain multi-view context representations of the two types of text; S3: Process the multi-view context representations of the two types of text to obtain the embedding features to be detected and the random embedding features; specifically, processing the multi-view context representations of the two types of text to obtain the embedding features to be detected and the random embedding features includes: Two independent multilayer perceptrons are used to process the multi-view context representations of two types of text to obtain the detection embedding features and random embedding features. The detection embedding features include the detection content-related embeddings and the detection event-related embeddings, and the random embedding features include random content-related embeddings and random event-related embeddings. S4: Adaptive instance normalization is used to perturb the embedded features to be detected and the random embedded features. A discriminator is used to process the perturbation results to obtain the false information detection results for small sample events. The process of perturbing the embedded features to be detected and the random embedded features using adaptive instance normalization includes: S41: Adaptive instance normalization is used to perturb the content-related embedding and random event-related embedding to obtain content feature perturbation; S42: Adaptive instance normalization is used to perturb the random content-related embeddings and the event-related embeddings to be detected, resulting in event feature perturbations; The formulas for obtaining the content feature perturbation and the event feature perturbation are as follows: ； ； in, This indicates a perturbation of content features. Indicates event characteristic perturbation; , These represent the first and second weight parameters, respectively. This indicates embeddings related to the content to be detected. This indicates a random content-related embedding. Indicates the embedding related to the event to be detected. Represents embeddings related to random events. This indicates adaptive instance normalization processing; S5: Calculate the first-stage loss based on the false information detection results of small sample events; S6: Freeze the discriminator parameters, input the embedded features to be detected into the discriminator for processing, and calculate the effective separation loss based on the discriminator output; S7: Adjust the parameters of the model other than the discriminator based on the effective separation loss to obtain the minimum effective separation loss; calculate the second-stage loss based on the effective separation loss and the minimum effective separation loss; adjust the model parameters based on the second-stage loss to obtain the trained small sample event false information detection model.
2. The method for detecting false information in small sample events based on bias decoupling according to claim 1, characterized in that, The process of obtaining a multi-view context representation includes: S21: Use the BERT model to extract the global and local representations of the descriptive text respectively; S22: Compute the global-to-local consistent representation and the local-to-local consistent representation; S23: The attention aggregation mechanism is used to process the global representation, the consistent representation from global to local, and the consistent representation from local to local to obtain the multi-view context representation.
3. The method for detecting false information in small sample events based on bias decoupling according to claim 1, characterized in that, The process of calculating the first-stage loss includes: The loss for content authenticity classification and the loss for event classification are calculated based on the results of small sample event false information detection and the real labels; the loss for relevance is calculated based on the embedded features to be detected; and the loss for content authenticity classification, the loss for event classification, and the loss for relevance are weighted and summed to obtain the first stage loss.
4. The method for detecting false information in small sample events based on bias decoupling according to claim 1, characterized in that, The formula for calculating the effective separation loss is: ； in, Indicates effective separation loss. This represents the i-th true / false label sampled from a uniform distribution. This represents the predicted value for the i-th category output by the content authenticity detector. Indicates the number of genuine and counterfeit labels. Indicates the number of event tag categories. This represents the label of the i-th event sampled from a uniform distribution. This represents the predicted value for the i-th category output by the event discriminator.
5. The method for detecting false information in small sample events based on bias decoupling according to claim 1, characterized in that, The formula for calculating the second-stage loss is: ； in, This indicates the losses in the second phase. Indicates the adjustment factor. Indicates effective separation loss. This represents the minimum effective separation loss.