Large model data enhanced short video published information detection method

Abstract

The invention provides a large model data enhanced short video release information detection method. Comprising the following steps: collecting short video multi-modal data and constructing an event time sequence context; generating a title and a picture description by using a large language model, and realizing completion from a text to an image through a vision-language generation model; extracting features of complemented data and original data, and performing feature space mapping and deep fusion modeling; quality perception evaluation and adaptive weight adjustment are constructed according to text-image semantic consistency, and unified representation is obtained; carrying out hierarchical fusion on the unified representation and the main feature of the basic detection model, and sending into a classifier to obtain a preliminary label; co-event multi-source information is subjected to collaborative reasoning and correction, and a final label and confidence are output; according to the method, the detection robustness and accuracy are improved under the condition of mode missing, plug-in and migratable engineering adaptability is achieved, and support can be provided for classification, anomaly recognition and mode discovery in big data mining.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and multimodal information processing; and more particularly to a method for detecting short video posting information using large model data augmentation. Background Technology

[0002] With the widespread adoption of mobile internet, short videos have become a primary medium for information dissemination. While enriching people's cultural and spiritual lives, they have also provided new channels for the rapid spread of misinformation. Misinformation on short video platforms often exhibits characteristics such as rapid dissemination, wide-ranging impact, and high difficulty in identification, posing a serious threat to social stability and public perception. Therefore, designing a misinformation detection method suitable for short video platforms to mitigate its negative impact is an urgent problem to be solved.

[0003] Existing methods for detecting information posted in short videos primarily rely on multimodal feature fusion technology. This involves analyzing text, images, audio, and other modal information within the video to determine its authenticity. This process typically requires big data processing tools to collect, store, and preprocess massive amounts of short videos, while simultaneously depending on big data analytics for statistical modeling and pattern exploration of multimodal features. Traditional methods often employ deep learning networks to extract features from each modality and then fuse them using attention mechanisms or feature concatenation. In recent years, large language generative models have made groundbreaking progress. Their powerful semantic understanding, information generation, and reasoning capabilities have provided new ideas for information detection tasks and offered higher-level support for pattern discovery, semantic association, and risk identification in big data mining.

[0004] However, existing methods still have shortcomings in practical applications: content posted on short video platforms often suffers from modal information loss, such as missing titles and descriptions, blurry or missing keyframes, etc. Traditional detection models are not robust enough in such scenarios, leading to a significant drop in detection accuracy. Therefore, designing a method that can effectively handle modal information loss and leverage the advantages of large-scale models to improve the detection performance of traditional models has become an urgent technical challenge and a key area that big data mining must focus on in the governance of misinformation in short videos. Summary of the Invention

[0005] The purpose of this invention is to address the problem that existing technologies have poor robustness in scenarios with missing modal information, leading to a significant drop in detection accuracy. This invention proposes a short video posting information detection method with large model data augmentation.

[0006] This invention is achieved through the following technical solution:

[0007] This invention relates to a method for detecting short video posting information using large-scale model data augmentation, comprising the following steps:

[0008] Collect multimodal data from short videos and construct a sample set of the same event context based on time series and temporal relationships; the multimodal data includes text data, image data, video data, and audio data;

[0009] For scenarios with missing text, missing images, and double missing text, a large language model is used to generate titles and image descriptions, and a visual-language generation model is used to achieve image generation and image feature enhancement.

[0010] The large language model can be any multimodal large language model. The multimodal large language model used in this invention is gpt-4o-2024-11-20.

[0011] The visual-language generation model can be any text-based image generation model, and the visual-language generation model used in this invention is BLIP3-o;

[0012] Features are extracted from the completed data and the original data, and feature space mapping and deep fusion modeling are performed. Based on the semantic consistency between text and image, a quality perception assessment and adaptive weight adjustment are constructed to obtain a unified representation of the completed feature fusion. The original video data is modeled using an existing short video information detection small model to obtain a unified representation of the original video features.

[0013] Existing short video information detection mini-models (this invention can provide plug-in enhancements for all existing short video information detection mini-models; the mini-models tested in this invention are BERT, VGG19, SVFEND, and FakingRecipe);

[0014] The original video features and the completed features are adaptively fused through a gating mechanism to obtain aggregated features, which are then fed into a classifier to obtain preliminary labels.

[0015] By introducing a large language model, and based on structured prompts, collaborative reasoning and correction are performed on multi-source information of the same event, and the final label and confidence score are output.

[0016] Preferably, the construction of the same event context sample set is based on timestamp similarity. From the candidate set belonging to the same event as the target short video, several samples with the smallest time difference are selected as supplementary information sources to form the same event context sample set. The selection process is as follows:

[0017]

[0018] in, For the same event context sample set, t q The timestamp of the target sample. This is the timestamp of the candidate samples. This mechanism ensures that the selected samples are highly correlated with the target video in both time and semantics, providing reliable semantic support for subsequent modality completion and feature enhancement.

[0019] Preferably, the specific process of generating the title and screen description is as follows: for samples with missing text, combine them with a sample set of contextual data from the same event. Title information With target video frame features I q Generate titles using large language models For missing image samples, combine them with a sample set from the same event context. Image information With target text information T q Generate scene descriptions using a large language model In cases where both text and images are missing, a title and screen description are generated simultaneously based on the event context sample set information to achieve multimodal completion.

[0020] Preferably, the specific content of the image generation and image feature enhancement includes: describing the image... A large model is generated from a multimodal image as input, producing a complete image that matches the semantic description. Will and Simultaneously input the multimodal large model fusion encoder to obtain the multimodal fusion hidden representation of the specified layer:

[0021]

[0022] in, It consists of completed image features enhanced with semantic supplementation of the image description. It is a multi-modal encoder at the lth * The method employs a fused hidden representation output from the layer. While maintaining semantic consistency, it incorporates spatial and visual detail information to enhance multimodal consistency and discriminative ability.

[0023] Preferably, the specific content of the unified representation of the completion feature fusion is: [The text abruptly ends here, likely due to an incomplete sentence or a formatting error.] With the description of the picture The semantic fusion feature F is obtained by fusing through a Transformer encoder. sem :

[0024]

[0025] Where Mean represents average pooling, and Trm represents the Transformer encoder layer;

[0026] Calculate image features With the description of the picture cosine similarity s q As a quality indicator:

[0027]

[0028] semantic fusion feature F sem and Input gating mechanism, based on s q Adaptive adjustment The fusion weights are used to obtain a unified representation of the completed features. Used for subsequent feature fusion and classification:

[0029]

[0030] Where σ is the Sigmoid function, W g β is the gated linear mapping weight, and β is the amplification factor when fusing augmented image enhancement features.

[0031] Preferably, the specific content of the unified representation of the original video features is as follows: the input video sample is fed into an existing fake news short video detection mini-model, and the original video features F output by the previous layer of its classifier are extracted. v This process is independent of specific model structures and can directly interface with various existing detection models to achieve a plug-and-play feature acquisition method, which is used for subsequent fusion and classification with the completed features.

[0032] F v =φ pre (V q ),

[0033] Among them, V q It is the input raw video sample, φ pre It is the feature extraction mapping of the layer before the classifier in the existing short video detection model.

[0034] Preferably, the adaptive fusion of the original video features and the completed features through a gating mechanism specifically involves: fusing the original video features F... v Unified representation with completion features After concatenation, the input is given to a gating network, which adaptively obtains the fused weight vector G:

[0035]

[0036] Where σ is the Sigmoid function, W g These are gating network parameters;

[0037] α is derived from the quality index s q The control bias term uses quality metrics to adjust the gating bias in order to suppress interference from low-quality completion features:

[0038] α=(1-s q )·γ,

[0039] Among them, s q It is a quality indicator, and γ is an adjustment coefficient;

[0040] The aggregated feature F is obtained by weighted summation of the two types of features based on their weights. fused Used for subsequent classification and discrimination:

[0041]

[0042] Preferably, the method for obtaining the initial label is as follows: the adaptively fused aggregated features F... fused The input is fed into a multilayer perceptron classifier (MLP) and mapped to predicted probabilities via a sigmoid function.

[0043]

[0044] The cross-entropy loss function is used to measure the difference between the probability distributions of the predicted labels and the actual labels. Its loss function is expressed as:

[0045]

[0046] Where y is the actual tag of the short video posting information; Let y be the predicted probability of short video posts containing false information, where 1 represents a fake news video and 0 represents a real news video. This gives the initial label y for predicting the authenticity of short video posts. init :

[0047]

[0048] in, τ is an indicator function, which takes the value 1 when the condition is true and 0 otherwise; τ is the classification threshold, which is usually 0.5.

[0049] Preferably, the specific content of the collaborative reasoning and correction is as follows: selecting input combinations based on the missing modality type; when no modality is missing, the input... When only the text modality is missing, input When only the image modality is missing, input When both text and image modalities are missing, input

[0050] From the same event context sample set Extract title information With the predicted initial label set Predicted initial tags based on the current video Structured prompts Π qThe data is fed into a large language model (LLM) for inference correction, outputting corrected labels and confidence scores. Taking into account (whether to use the large model for corrected labels is determined by whether the large model's confidence score exceeds a set decision threshold, here set to 0.85), the large model's corrected labels, confidence scores, and the small model's initial predicted labels, a final prediction result for the authenticity of the short video's published information is obtained.

[0051]

[0052] in, For labels predicted by large model inference, θ q The confidence score given by the large model is δ, which is the set decision threshold, set to 0.85.

[0053] The present invention has the following advantages:

[0054] (1) The method of this invention achieves efficient detection of short video posting information under the condition of missing modal information. Through the collaborative work of a large language model and a visual-language generation model, it can intelligently complete the missing text description and image information. It constructs an information association mechanism based on the event temporal context, fully explores the temporal correlation and semantic consistency of multi-source information of the same event, and provides reliable semantic support for modal completion and feature enhancement in big data mining. Through quality-aware evaluation and adaptive weight adjustment mechanism, it ensures high-quality fusion of completed features and avoids the negative impact of low-quality generated content on detection performance. It introduces a large language model to perform collaborative reasoning and correction on multi-source information of the same event based on structured prompts, and outputs the final label and confidence score, realizing pattern discovery and risk identification of large-scale short video data. It solves the problem of poor detection performance of existing technologies under the condition of missing modal information and improves detection robustness and accuracy.

[0055] (2) The method of the present invention has good plug-in characteristics, can be flexibly connected with a variety of existing detection models, has strong engineering adaptability and portability, and can provide a new technical path for the governance of false information in short videos driven by big data mining. Attached Figure Description

[0056] Figure 1 This is a flowchart of a short video posting information detection method involving large model data augmentation, as described in this invention.

[0057] Figure 2 This is a framework diagram of the fake news short video detection in an embodiment of the present invention. Detailed Implementation

[0058] The present invention will now be described in detail with reference to specific embodiments. It should be noted that the following embodiments are merely further illustrations of the present invention, but the scope of protection of the present invention is not limited to the following embodiments.

[0059] Example 1

[0060] This embodiment relates to a method for detecting short video posting information using large model data augmentation, such as... Figure 1 As shown, the specific steps include:

[0061] S1. Input video multimodal data and collect multimodal information such as text, images, video and audio from short videos;

[0062] S2. Construct a sample set of contexts for the same event, and select several samples with the smallest time difference from the candidate set of the same event based on time series and time relationship as supplementary information sources;

[0063] S3. Complete text information and image descriptions. For scenarios with missing text, missing images, and double missing text, use a large language model combined with event context information to generate titles and image descriptions.

[0064] S4. Image generation and image feature enhancement: The supplementary image description is converted into a complete image through a visual-language generation model, and semantically enhanced image features are obtained using a multimodal fusion encoder.

[0065] S5. Complete feature fusion and unified representation: The supplementary title features and the screen description features are fused through the Transformer encoder. Combined with quality perception evaluation and adaptive weight adjustment mechanism, a complete feature fusion and unified representation is obtained.

[0066] S6. Unified representation of original video features: Use existing short video posting information detection models to extract and fuse features from the original video data to obtain the original feature representation;

[0067] S7. Adaptive fusion of original features and completed features: The original features and completed features of the video are adaptively fused through a gating mechanism to obtain aggregated features;

[0068] S8. Based on the fusion features, the predicted initial label is obtained, and the aggregated features are input into the classifier for preliminary classification judgment;

[0069] S9. Large-scale collaborative reasoning correction results: The large language model is introduced to perform collaborative reasoning and correction of multi-source information of the same event based on structured prompts.

[0070] S10. Output the final classification result of the short video, taking into account the corrected labels of the large model, the confidence score and the initial labels predicted by the small model, and output the final detection result.

[0071] Example 2

[0072] This embodiment is based on the large model data augmentation short video information detection method involved in Embodiment 1 above, and is used to detect fake news published in short videos. Specifically, it includes the following steps:

[0073] S1: Input news videos, and create a news video dataset D = {D1, D2, ..., D...} n Each news video instance D in} i Format as D i = (T,A,I,V,E,M), where T, A,I,V,E,M represent text content, audio, keyframe, video clip, associated event, and publication time, respectively.

[0074] S2: From the candidate set belonging to the same event as the target short video, select several samples with the smallest time difference as supplementary information sources to form a context sample set of the same event. The selection process is as follows:

[0075]

[0076] in, For the same event context sample set, t q The timestamp of the target sample. The timestamps of the candidate samples are used to ensure that the selected samples are highly correlated with the target video in both time and semantics, providing reliable semantic support for subsequent modality completion and feature enhancement.

[0077] S3: For samples with missing text, combine them with a sample set from the same event context. Title information With target video frame features I q Generate titles using large language models For missing image samples, combine them with a sample set from the same event context. Image information With target text information T q Generate scene descriptions using a large language model In cases where both text and images are missing, a title and screen description are generated simultaneously based on the event context sample set information to achieve multimodal completion.

[0078] S4: Describe the scene A large model is generated from a multimodal image as input, producing a complete image that matches the semantic description. Will and Simultaneously input the multimodal large model fusion encoder to obtain the multimodal fusion hidden representation of the specified layer:

[0079]

[0080] in, It consists of completed image features enhanced with semantic supplementation of the image description. It is a multi-modal encoder at the lth * The method uses a fusion hidden representation of the layer output. While maintaining semantic consistency, it introduces spatial and visual detail information to improve multimodal consistency and discriminative ability.

[0081] S5: Title Features Features of the image description The semantic fusion feature F is obtained by fusing through a Transformer encoder. sem :

[0082]

[0083] Here, Mean represents average pooling, and Trm represents the Transformer encoder layer.

[0084] Calculate image enhancement features Features of the image description cosine similarity s q As a quality indicator:

[0085]

[0086] semantic fusion feature F sem and Input gating mechanism, based on s q Adaptive adjustment The fusion weights are used to obtain a unified representation of the completed features. Used for subsequent feature fusion and classification:

[0087]

[0088] Where σ is the Sigmoid function, W g β is the gated linear mapping weight, and β is the amplification factor when fusing augmented image enhancement features.

[0089] S6: Feed the input video sample into the existing fake news short video detection model, and extract the original video features F from the output of the previous layer of its classifier. v This process does not depend on a specific model structure and can be directly interfaced with various existing detection models to achieve a plug-and-play feature acquisition method, which is used for subsequent fusion and classification with the completed features.

[0090] F v =φ pre (V q ),

[0091] Among them, V q It is the input raw video sample, φ pre It is the feature extraction mapping of the layer before the classifier in the existing short video detection model.

[0092] S7: Extract the original video features F v Unified representation with completion features After concatenation, the input is given to a gating network, which adaptively obtains the fused weight vector G:

[0093]

[0094] Where σ is the Sigmoid function, W g These are gating network parameters.

[0095] α is derived from the quality index s q The control bias term uses quality metrics to adjust the gating bias in order to suppress interference from low-quality completion features:

[0096] α=(1-s q )·γ,

[0097] Among them, s q It is a quality indicator, and γ is an adjustment coefficient.

[0098] Finally, the two types of features are weighted and summed to obtain the fused feature F. fused Used for subsequent classification and discrimination:

[0099]

[0100] S8: The adaptively fused aggregated features F fused The input is fed into a multilayer perceptron classifier (MLP) and mapped to predicted probabilities via a sigmoid function.

[0101]

[0102] The cross-entropy loss function is used to measure the difference between the probability distributions of the predicted labels and the actual labels. Its loss function is expressed as:

[0103]

[0104] Where y is the actual tag of the short video posting information; The probability of short video posts being false is predicted, with 1 for fake news videos and 0 for real news videos. This gives the initial label y for predicting the authenticity of short video posts. init :

[0105]

[0106] in, τ is an indicator function, which takes the value 1 when the condition is true and 0 otherwise; τ is the classification threshold, which is usually 0.5.

[0107] S9: Select the input combination based on the missing mode type. When no mode is missing, input... When only the text modality is missing, input When only the image modality is missing, input When both text and image modalities are missing, input From the same event context sample set Extract the title set With the predicted initial label set And combined with the predicted initial tags of the current video Structured prompts Π q The data is fed into a large language model (LLM) for inference and correction, outputting corrected labels and confidence scores. Taking into account the corrected labels from the large model, the confidence scores, and the initial labels predicted by the small model, a final prediction result for the authenticity of the information posted in the short video is obtained.

[0108]

[0109]

[0110] in, For labels predicted by large model inference, θ q δ represents the confidence score given by the large model, and δ is the set decision threshold.

[0111] S10: Output the classification results of fake news short videos.

[0112] See Figure 2The diagram shows the framework for short video fake news detection using large-scale model data augmentation, which consists of three stages: (a) Modal completion generation: For different types of modal missing scenarios, supplementary information sources are constructed using a sample set of the same event context, which is the execution module of step S2 above. Missing titles and image descriptions are generated through a large language model, which is the execution module of step S3 above. Text-to-image completion is achieved using a visual-language generation model, which is the execution module of step S4 above. (b) Feature fusion enhancement: The completed features are adaptively fused with the original features. The unified representation of the completed features corresponds to the execution module of step S5, the small model perceptron corresponds to the execution module of step S6, and the adaptive gating fusion corresponds to the execution module of step S7. (c) Large model inference correction: The small model classifier corresponds to the execution module of step S8. Based on structured prompts, multi-source information of the same event is input into the large language model for collaborative inference, which corresponds to the execution module of step S9. The prediction results are corrected by combining the event context information, and the final output of the classification result of real or fake corresponds to the execution module of step S10. The entire framework achieves efficient detection of fake news in short videos under modal missing conditions through the coordinated work of three stages: modality completion, feature fusion, and inference correction.

[0113] Tables 1 and 2 present the experimental results of this invention on the FakeSV and FakeTT datasets under different modal missing rates, respectively. Five missing rate levels from 0 to 0.7 were set, comparing the original performance of single-modal methods (BERT text model, VGG19 image model), multimodal baseline methods (SVFEND, FakingRecipe), and the performance after enhancement using the method of this invention (w / Ours). The experiments used two core evaluation metrics, accuracy (ACC) and F1 score, for performance evaluation.

[0114] Table 1

[0115]

[0116]

[0117] Table 2

[0118]

[0119] The results show that the performance of all original methods decreases significantly with the increase of modality missing rate. However, after adopting the large model data augmentation strategy of the present invention, all baseline methods have achieved significant performance improvement at various missing rate levels, especially in high missing rate scenarios. This verifies that the method of the present invention can effectively enhance the robustness and accuracy of existing detection models under modality missing information conditions.

[0120] The method involved in this invention constructs an event temporal context, utilizes a large language model and a visual-language generation model to intelligently complete missing modal information such as text and images, and performs feature fusion modeling with the completed data and the original data, thereby providing a more complete and usable feature space for big data mining. Simultaneously, a quality-aware evaluation mechanism is constructed based on semantic consistency to achieve adaptive fusion weight adjustment. Finally, a large language model is introduced to infer and correct the preliminary prediction results, outputting the final detection results, achieving pattern discovery and risk identification on large-scale short video data. This method effectively solves the problem of poor performance of traditional detection techniques under modal missing conditions, significantly improves detection robustness and accuracy, and has good engineering adaptability and transferability, providing a new technical solution for big data mining-driven detection of false information in short videos.

[0121] The specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various modifications or variations within the scope of the claims, which do not affect the essence of the present invention.

Claims

1. A method for detecting short video posting information using large-scale model data augmentation, characterized in that, Includes the following steps: Input short video multimodal data and construct a sample set of the same event context based on time series and time relationship; the multimodal data includes text data, image data, video data and audio data; For scenarios with missing text, missing images, and double missing text, a large language model is used to generate titles and image descriptions, and a visual-language generation model is used to achieve image generation and image feature enhancement. The large language model is gpt-4o-2024-11-20; The visual-language generation model is BLIP3-o; Features are extracted from the completed data and the original data, and feature space mapping and deep fusion modeling are performed. Based on the semantic consistency between text and image, a quality perception assessment and adaptive weight adjustment are constructed to obtain a unified representation of the completed feature fusion. The original video data is modeled using an existing short video information detection small model to obtain a unified representation of the original video features. The existing short video information detection models are BERT, VGG19, SVFEND, and FakingRecipe. The original video features and the completed features are adaptively fused through a gating mechanism to obtain aggregated features, which are then fed into a classifier to obtain preliminary labels. By introducing a large language model, and based on structured prompts, collaborative reasoning and correction are performed on multi-source information of the same event, and the final label and confidence score are output.

2. The method for detecting short video posting information using large model data augmentation as described in claim 1, characterized in that, The construction of the same event context sample set is based on timestamp similarity. From the candidate set belonging to the same event as the target short video, several samples with the smallest time difference are selected as supplementary information sources to form the same event context sample set. The selection process is as follows: in, For the same event context sample set, t q The timestamp of the target sample. The timestamp of the candidate sample.

3. The short video posting information detection method with large model data augmentation as described in claim 1, characterized in that, The specific process for generating the title and screen description is as follows: For samples with missing text, combine them with a sample set of samples from the same event context. Title information With target video frame features I q Generate titles using large language models For missing image samples, combine them with a sample set from the same event context. Image information With target text information T q Generate scene descriptions using a large language model 4. The method for detecting short video posting information using large model data augmentation as described in claim 1, characterized in that, The specific content of the image generation and image feature enhancement is as follows: describing the image... A large model is generated from a multimodal image as input, producing a complete image that matches the semantic description. Will and Simultaneously input the multimodal large model fusion encoder to obtain the multimodal fusion hidden representation of the specified layer: in, It consists of completed image features enhanced with semantic supplementation of the image description. It is a multi-mode encoder in the first The fused hidden representation of the layer output.

5. The method for detecting short video posting information using large model data augmentation as described in claim 1, characterized in that, The specific content of the unified representation of the completion feature fusion is as follows: [The title is missing from the original text.] With the description of the picture The semantic fusion feature F is obtained by fusing through a Transformer encoder. sem : Where Mean represents average pooling, and Trm represents the Transformer encoder layer; Calculate image features With the description of the picture cosine similarity s q As a quality indicator: semantic fusion feature F sem and Input gating mechanism, based on s q Adaptive adjustment The fusion weights are used to obtain a unified representation of the completed features. Used for subsequent feature fusion and classification: Where σ is the Sigmoid function, W g β is the gated linear mapping weight, and β is the amplification factor when fusing augmented image enhancement features.

6. The method for detecting short video posting information using large model data augmentation as described in claim 1, characterized in that, The specific content of the unified representation of the original video features is as follows: Input video samples are fed into an existing fake news short video detection model, and the original video features F output from the previous layer of its classifier are extracted. v ; F v =φ pre (V q ), Among them, V q It is the input raw video sample, φ pre It is the feature extraction mapping of the layer before the classifier in the existing short video detection model.

7. The method for detecting short video posting information using large model data augmentation as described in claim 1, characterized in that, The specific content of the adaptive fusion of the original video features and the completed features through a gating mechanism is as follows: The original video features F... v Unified representation with completion features After concatenation, the input is given to a gating network, which adaptively obtains the fused weight vector G: Where σ is the Sigmoid function, W g These are gating network parameters; α is derived from the quality index s q The bias term of the control, α=(1-s q )·c Among them, s q It is a quality indicator, and γ is an adjustment coefficient; The aggregated feature F is obtained by weighted summation of the two types of features based on their weights. fused Used for subsequent classification and discrimination:

8. The method for detecting short video posting information using large model data augmentation as described in claim 1, characterized in that, The method for obtaining the initial label is as follows: The adaptively fused aggregated features F... fused The input is fed into a multilayer perceptron classifier (MLP) and mapped to predicted probabilities via a sigmoid function. The cross-entropy loss function is used to measure the difference between the probability distributions of the predicted labels and the actual labels. Its loss function is expressed as: Where y is the actual tag of the short video posting information; Let y be the predicted probability of short video posts containing false information, where 1 represents a fake news video and 0 represents a real news video. This gives the initial label y for predicting the authenticity of short video posts. init : in, τ is an indicator function, which takes the value 1 when the condition is true and 0 otherwise; τ is the classification threshold, which is usually 0.

5.

9. The method for detecting short video posting information using large model data augmentation as described in claim 1, characterized in that, The specific content of the collaborative reasoning and correction is as follows: Input combinations are selected based on the missing modality type; when no modality is missing, the input... When only the text modality is missing, input When only the image modality is missing, input When both text and image modalities are missing, input From the same event context sample set Extract title information With the predicted initial label set Predicted initial tags based on the current video Structured prompts Π q The data is fed into a large language model (LLM) for inference and correction, outputting corrected labels and confidence scores. Whether to use the large model to correct labels is determined by whether the large model's confidence score exceeds a set decision threshold (set to 0.85). The large model's corrected labels, confidence scores, and the small model's predicted initial labels are combined to obtain the final prediction result regarding the authenticity of the short video's published information. in, For labels predicted by large model inference, θ q δ represents the confidence score given by the large model, and δ is the set decision threshold.

Images