A Method and Device for Short Video Content Risk Detection Based on Multimodal Fusion Model
By constructing a multimodal fusion model for short video content risk detection, this method solves the problems of noise interference and fuzzy sample discrimination in short video review, and achieves accurate identification of hidden risks and low false alarm rate.
Patent Information
- Application Number
- CN202611013315.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-08
- Publication Date
- 2026-11-17
AI Technical Summary
Existing short video review technologies suffer from problems such as high data noise interference, lack of differentiation between primary and secondary modes in multimodal fusion mechanisms, and weak ability to distinguish between ambiguous samples, making it difficult to balance the false positive and false negative rates.
A short video content risk detection method based on a multimodal fusion model is constructed. The video stream and audio stream are separated by an adaptive cleaning algorithm and an asymmetric cross-modal attention network. The semantic perplexity and character density are calculated using a pre-trained language model. A dynamic confidence gating is constructed to generate clean text sequences. Text enhancement features are generated through an asymmetric cross-attention module. Finally, a judgment is made based on dual threshold logic.
It effectively removes dynamic watermarks and bullet comments from short videos, improves the text signal-to-noise ratio, establishes the semantic alignment accuracy of image-text alignment, reduces the false positive rate of edge samples, and meets the actual business needs of content risk control.
Smart Images

Figure CN123119331A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical fields of network content security and artificial intelligence, and in particular to a method for detecting risks in short video content based on a multimodal fusion model, as well as a device for detecting risks in short video content based on a multimodal fusion model. Background Technology
[0002] With the explosive growth of short video platforms, the compliance review of massive amounts of video content faces enormous challenges. Existing short video review technologies mainly suffer from the following shortcomings:
[0003] 1. High levels of data noise and difficulty in cleaning: Short videos are often filled with irrelevant text such as platform watermarks, advertising stickers, and scrolling comments in live streams. Current technologies typically only perform simple filtering based on "text length" or "keyword matching." This crude approach leads to the following problems: if the threshold is too high, key short words (such as slang) will be missed; if the threshold is too low, a large amount of garbled text and irrelevant words will introduce noise, seriously interfering with subsequent semantic understanding.
[0004] 2. Lack of Priority in Multimodal Fusion Mechanisms: Most existing multimodal detection models employ "symmetrical fusion" (such as directly concatenating or simply weighting video, audio, and text features). However, in the field of content security, the text is often the core basis for defining the nature of the risk (for example, the same scene is compliant if the caption is "acting," but non-compliant if the caption is "real fight"). Existing symmetrical fusion ignores the dominant role of text, resulting in excessive weighting of background audio or irrelevant visuals, diluting the core risk information.
[0005] 3. Weak ability to distinguish "fuzzy" samples: Existing classification models usually output a single probability value (such as 0.6). For "fuzzy samples" at the classification boundary (such as soft pornography and subtle traffic generation), there is a lack of an effective secondary judgment mechanism, which makes it difficult to balance the false positive rate and the false negative rate. Summary of the Invention
[0006] To overcome the shortcomings of existing technologies, the technical problem to be solved by this invention is to provide a short video content risk detection method based on a multimodal fusion model. This method achieves accurate identification of hidden risks in short videos by constructing an adaptive cleaning algorithm based on semantic consistency and an asymmetric cross-modal attention network.
[0007] The technical solution of this invention is: a short video content risk detection method based on a multimodal fusion model, comprising: (1) Acquire the short video to be detected and separate the video stream and audio stream; use an adaptive step size strategy to extract key frames and resample the audio; at the same time, extract the original text stream through optical character recognition (OCR) and automatic speech recognition (ASR); (2) Calculate the semantic perplexity and effective character density of the original OCR text lines using a pre-trained language model; construct a dynamic confidence gating system to filter out noisy lines with discrete semantics and low character density; perform fuzzy matching recall on the filtered text to ensure that key words are not lost and generate a clean text sequence. (3) Extract text, visual and audio features through BERT, ResNet and Wav2Vec respectively; construct an asymmetric cross attention module to force text features as query vectors and visual and audio features as key-value pairs; introduce modal sparsity constraints to suppress redundant video frames and background noise that are irrelevant to the semantics of the text and generate text-enhanced visual and audio features. (4) Input the fused features into the classifier and make a judgment based on the dual threshold logic: if it is below the safety threshold, it is considered safe; if it is above the high-risk threshold, it is considered a violation. For ambiguous samples in the middle, calculate the modal conflict degree. If the conflict degree is too high, trigger a second deep reasoning or manual review.
[0008] The beneficial technical effects of the present invention are as follows:
[0009] Strong anti-interference capability: Through semantic perplexity filtering, it effectively removes common dynamic watermarks and bullet screen interference in short videos, significantly improving the text signal-to-noise ratio.
[0010] Precise semantic alignment: The asymmetric fusion mechanism establishes the logic of "searching for images with text and listening to audio with text", enabling the model to understand advanced and subtle risks such as "discrepancy between text and image" or "indirect criticism".
[0011] Flexible decision-making mechanism: The dual threshold and modal conflict detection mechanism effectively reduces the false alarm rate of edge samples, which is more in line with the actual business needs of content risk control.
[0012] It also provides a short video content risk detection device based on a multimodal fusion model, including: The multi-source data acquisition and preprocessing module is configured to acquire the short video to be detected, separate the video stream and audio stream; extract key frames using an adaptive step size strategy and resample the audio; and extract the original text stream through optical character recognition (OCR) and automatic speech recognition (ASR). The semantic consistency-based adaptive text cleaning module is configured to use a pre-trained language model to calculate the semantic perplexity and effective character density of the original OCR text lines; construct a dynamic confidence gating to filter out noisy lines with discrete semantics and low character density; and perform fuzzy matching recall on the filtered text to ensure that key words are not lost and generate a clean text sequence. The asymmetric feature extraction and fusion module guided by text anchors is configured to extract text, visual, and audio features through BERT, ResNet, and Wav2Vec respectively; an asymmetric cross-attention module is constructed to force text features as query vectors and visual and audio features as key-value pairs; modal sparsity constraints are introduced to suppress redundant video frames and background noise that are irrelevant to the semantics of the text, and to generate text-enhanced visual and audio features; The risk-sensitive dynamic decision-making module is configured to input fused features into a classifier and make judgments based on dual threshold logic: samples below the safety threshold are judged as safe, and samples above the high-risk threshold are judged as violations; for ambiguous samples in the middle, the modal conflict degree is calculated, and if the conflict degree is too high, a second deep reasoning or manual review is triggered. Attached Figure Description
[0013] Figure 1 A flowchart illustrating the overall process of the short video content risk detection method based on a multimodal fusion model according to the present invention is shown. The entire process is demonstrated from [Video Input] -> [S1 Preprocessing] -> [S2 Cleaning] -> [S3 Encoding and Fusion] -> [S4 Decision] -> [Result Output].
[0014] Figure 2 A flowchart of the semantic consistency-based adaptive text cleaning method according to the present invention is shown. It illustrates the logic of calculating [PPL score] and [character density] in two paths after OCR text input, merging them into [dynamic gating], and outputting [clean text].
[0015] Figure 3 A structural diagram of the asymmetric cross-modal cross-attention fusion module according to the present invention is shown. The diagram uses text features as the query and visual and audio features as the key / value pairs. After dimension alignment, cross-attention, sparsity masking, and sigmoid gating, the fused features are output.
[0016] Figure 4 A logic diagram for dual-threshold dynamic risk decision-making according to the present invention is shown. It illustrates the different flow paths of the probability value P in the [safe zone], [fuzzy zone], and [high-risk zone], as well as the logic for calculating the "modal conflict degree" in the [fuzzy zone].
[0017] Figure 5 The physical architecture diagram of the system deployment according to the present invention is shown. Frontend (Streamlit) <-> API Gateway <-> Algorithm Inference Service (ONNX) <-> Database (MySQL / Redis). Detailed Implementation
[0018] like Figure 1As shown, this short video content risk detection method based on a multimodal fusion model includes: (1) Acquire the short video to be detected and separate the video stream and audio stream; use an adaptive step size strategy to extract key frames and resample the audio; at the same time, extract the original text stream through optical character recognition (OCR) and automatic speech recognition (ASR); (2) Calculate the semantic perplexity and effective character density of the original OCR text lines using a pre-trained language model; construct a dynamic confidence gating to filter out noisy lines with discrete semantics and low character density; perform fuzzy matching recall on the filtered text to ensure that key words are not lost and generate a clean text sequence. (3) Extract text, visual and audio features through BERT, ResNet and Wav2Vec respectively; construct an asymmetric cross-attention module to force text features to be used as query vectors and visual and audio features as key-value pairs; introduce modal sparsity constraints to suppress redundant video frames and background noise that are not related to the semantics of the text and generate text-enhanced visual and audio features; (4) Input the fused features into the classifier and make a judgment based on the dual threshold logic: if it is below the safety threshold, it is considered safe; if it is above the high-risk threshold, it is considered a violation. For ambiguous samples in the middle, calculate the modal conflict degree. If the conflict degree is too high, trigger a second deep reasoning or manual review.
[0019] The beneficial technical effects of the present invention are as follows:
[0020] Strong anti-interference capability: Through semantic perplexity filtering, it effectively removes common dynamic watermarks and bullet screen interference in short videos, significantly improving the text signal-to-noise ratio.
[0021] Precise semantic alignment: The asymmetric fusion mechanism establishes the logic of "searching for images with text and listening to audio with text", enabling the model to understand advanced and subtle risks such as "discrepancy between text and image" or "indirect criticism".
[0022] Flexible decision-making mechanism: The dual threshold and modal conflict detection mechanism effectively reduces the false alarm rate of edge samples, which is more in line with the actual business needs of content risk control.
[0023] Preferably, step (2) includes the following steps: (2.1) Construct a blacklist regular expression library containing common interference words to remove explicit noise; (2.2) Input the text line T_i into the lightweight language model and calculate its perplexity PPL_i. If PPL_i is too high, it is marked as low confidence. Where T_i represents the i-th line of OCR / ASR text, PPL_i represents the semantic perplexity of the text line. The larger PPL_i is, the lower the semantic coherence and the higher the possibility of it being garbled or noisy text. (2.3) Calculate the proportion of valid Chinese characters in the line Density_i, where Density_i represents the proportion of valid characters such as Chinese characters, numbers and business keywords in the i-th line of text to the total number of characters in the line, which is used to measure the information density of the text line; (2.4) Set the fusion scoring function Score_i = α·(1 / PPL_i) + β·Density_i. When Score_i < Threshold, the row is discarded. Where Score_i represents the comprehensive confidence score of the i-th row of text, α and β represent the weight coefficients of the semantic perplexity item and the character density item and α+β=1, and Threshold represents the text cleaning threshold, which can be preset or dynamically updated according to the statistical distribution of the sample set.
[0024] Preferably, in step (3), a parallel feature extraction channel is constructed: The text channel uses the BERT-Base-Chinese model. The input is a cleaned token sequence, and the [CLS] vector is extracted as the global semantic representation V_text∈R^768. Here, V_text represents the global semantic feature vector of the text modality, and 768 represents the output dimension of the hidden layer of the BERT-Base-Chinese model, which is used to represent the contextual semantic information of the short video text content. The visual channel uses ResNet-50 to encode the sampled keyframes, extracts Layer4 output, and performs global average pooling to obtain the serialized visual features V_vis∈R^(T×1024); where V_vis represents the visual modality feature sequence, T represents the number of sampled keyframes, and 1024 represents the visual feature dimension of each keyframe after ResNet-50 encoding. The audio channel resamples the audio to 16kHz, extracts the Log-Mel spectrogram, and extracts the acoustic features V_aud∈R^(S×128) through the VGGish network; where V_aud represents the audio modal feature sequence, S represents the number of time segments after audio segmentation, and 128 represents the acoustic feature dimension corresponding to each time segment.
[0025] Preferably, in step (3), V_vis and V_aud are mapped to 768 dimensions through a fully connected layer to obtain V'_vis and V'_aud respectively; the text feature V_text is used as the query vector Query, and the visual feature V'_vis and / or audio feature V'_aud are used as the key and value, respectively, to calculate the attention weight matrix A=Softmax((V_text·(V'_vis)^T) / √d); where d represents the common feature dimension of Query and Key, and in this embodiment d=768, and A represents the attention weight of text semantics on each visual frame or audio segment.
[0026] Preferably, in step (3), a sparsity mask is introduced before Softmax to suppress visual frames with weights lower than 0.1, forcing the model to focus on images that are strongly related to the text.
[0027] Preferably, in step (3), a text-guided visual enhancement feature V'_vis is generated and fused with the original text feature through Sigmoid gating: F = V_text + σ(W·V'_vis)⊙V'_vis; where F represents the fused multimodal representation, W represents the learnable gating weight matrix, σ(·) represents the Sigmoid activation function, ⊙ represents element-wise multiplication, and V'_vis represents the visual enhancement feature after cross-attention weighting.
[0028] Preferably, step (3) further includes model training and loss function: In the training engine, design a composite loss function: In Focal Loss, L_focal = -α_t(1-p_t)^γ log(p_t), γ=2.0; where p_t represents the model's predicted probability of the true class, α_t represents the class balance factor, and γ represents the focus factor, which is used to reduce the loss weight of easily classified samples and enhance the contribution of difficult-to-classify and minority class risk samples to the model training. In the multimodal contrast loss, the distance between matching illegal text and images is reduced in the feature space, while the distance between illegal text and normal images is increased, thereby enhancing the model's ability to distinguish between text and image mismatch samples.
[0029] Preferably, in step (4), Quantization acceleration: The model is quantized to INT8 format using ONNXExporter and deployed on a T4 GPU server; Dynamic decision logic: set a safety threshold T1 and a high-risk threshold T2, where T1<T2; if the prediction probability P<T1, it is determined as safe; if P>T2, it is determined as high-risk violation; if P∈[T1,T2], enter the fuzzy state processing flow, calculate the difference Δ=|P_text-P_fusion| between the text branch prediction result P_text and the fusion result P_fusion. If Δ>δ, mark the request as REVIEW_REQUIRED and route it to the manual review queue, otherwise output according to the risk level of P_fusion.
[0030] Those of ordinary skill in the art may understand that all or part of the steps of the method in the above embodiments may be completed by a program instructing relevant hardware. The program may be stored in a computer-readable storage medium. When the program is executed, the steps of the method in the above embodiments are included, and the storage medium may be: ROM / RAM, magnetic disk, optical disk, memory card, etc. Therefore, corresponding to the method of the present invention, the present invention also relates to a short video content risk detection device based on a multimodal fusion model, which is usually represented in the form of functional modules corresponding to each step of the method. The device includes: A multi-source data acquisition and preprocessing module, configured to acquire a to-be-detected short video, separate a video stream and an audio stream; extract key frames using an adaptive step size strategy, and resample the audio; and extract an original text stream through Optical Character Recognition (OCR) and Automatic Speech Recognition (ASR). An adaptive text cleaning module based on semantic consistency, configured to calculate the semantic perplexity and effective character density of original OCR text lines using a pre-trained language model; construct a dynamic confidence gating to filter out noise lines with discrete semantics and low character density; perform fuzzy matching recall on the filtered text to ensure that key short words are not lost, and generate a pure text sequence. A text anchor-guided asymmetric feature extraction and fusion module, configured to extract text, visual and audio features through BERT, ResNet and Wav2Vec respectively; construct an asymmetric cross-attention module, forcing text features to be used as query vectors, and visual and audio features to be used as key-value pairs; introduce modal sparsity constraints to suppress redundant video frames and background noise irrelevant to copy semantics, and generate text-enhanced visual features and audio features. A dynamic decision module based on risk sensitivity, configured to input the fused features into a classifier, and perform determination based on dual-threshold logic: a value lower than the safety threshold is determined as safe, and a value higher than the high-risk threshold is determined as violation; for intermediate fuzzy state samples, the modal conflict degree is calculated, and if the conflict degree is too high, secondary deep inference or manual review is triggered.
[0031] The content of the present invention will be described in more detail below. Example 1: Text data cleaning based on semantic consistency
[0032] To address the issue of OCR recognition results containing a large number of meaningless characters (such as "TikTok ID" and "UI button") in short video scenarios, this embodiment designs the OCRTextCleaner module:
[0033] Preliminary screening using regular expressions: Construct a blacklist regular expression library containing common interference words (such as r"ID[:]" and r"follow") to remove explicit noise.
[0034] Semantic perplexity calculation: Input text line T_i into the lightweight language model and calculate its perplexity PPL_i. If PPL_i is too high (indicating that the sentence is not fluent and may be garbled), it is marked as low confidence.
[0035] Character density weighting: Calculates the proportion of valid Chinese characters in a line, Density_i, to measure the information density of a text line.
[0036] Dynamic gating decision: A fusion scoring function is defined as Score_i = α·(1 / PPL_i) + β·Density_i. When Score_i < Threshold, the line is discarded. Here, Score_i represents the overall confidence score of the text line, α and β are weighting coefficients, and Threshold is the text cleaning threshold. This step effectively solves the problem of traditional methods relying solely on length filtering, which leads to the accidental deletion of short key sentences (such as "add WeChat"). Example 2: Multimodal Feature Extraction Network
[0037] Constructing parallel feature extraction channels:
[0038] Text channel: The BERT-Base-Chinese model is used as input. The cleaned token sequence is input and the [CLS] vector is extracted as the global semantic representation V_text∈R^768; where V_text represents the global semantic feature vector of the text modality and 768 represents the output dimension of the hidden layer of the BERT-Base-Chinese model.
[0039] Visual channel: ResNet-50 is used to encode the sampled keyframes, the Layer4 output is extracted and global average pooling is used to obtain the serialized visual features V_vis∈R^(T×1024); where T is the number of keyframes and 1024 is the dimension of visual features per frame.
[0040] Audio channel: The audio is resampled to 16kHz, the Log-Mel spectrogram is extracted, and the acoustic features V_aud∈R^(S×128) are extracted through the VGGish network; where S is the number of audio time segments, and 128 is the acoustic feature dimension of each segment. Example 3: Asymmetric Cross-Attention Fusion (Core Algorithm)
[0041] Unlike traditional stitching and fusion, this embodiment constructs an AsymmetricFusionModule: Dimension alignment: V_vis and V_aud are mapped to 768 dimensions through a fully connected layer, resulting in V'_vis and V'_aud respectively, where 768 is the common feature dimension of Query, Key, and Value in the cross-modal attention module.
[0042] Cross-attention calculation: Use the text feature V_text as the query.
[0043] Use visual feature V'_vis and / or audio feature V'_aud as the key and value.
[0044] Calculate the attention weight matrix A=Softmax((V_text·(V'_vis)^T) / √d), where d represents the common feature dimension and A represents the attention weight of text semantics on visual frames or audio segments.
[0045] Innovation achieved by introducing a sparsity mask before Softmax to suppress visual frames with weights below 0.1, forcing the model to focus on images strongly related to the text (e.g., when the text mentions "guns," the gun image is highlighted while the background is ignored).
[0046] Gated residual connection: Generate text-guided visual enhancement features V'_vis, and fuse them with the original text features through Sigmoid gating: F = V_text + σ(W·V'_vis)⊙V'_vis; where W represents the learnable gating weight matrix, σ(·) represents the Sigmoid activation function, and ⊙ represents element-wise multiplication. Example 4: Model Training and Loss Function
[0047] In TrainerEngine, design a composite loss function: Focal Loss: This loss function addresses the class imbalance problem caused by the extreme scarcity of black market and gray market samples. Its loss function is L_focal = -α_t(1-p_t)^γ log(p_t), where γ=2.0. Here, p_t represents the model's predicted probability of the true class, α_t represents the class balance factor, and γ represents the focus factor, which is used to reduce the loss weight of easily classified samples and enhance the contribution of difficult-to-classify and minority-class risk samples to the model training.
[0048] Contrastive Loss: In the feature space, it brings the matching (illegal text, illegal image) closer together and pushes the (illegal text, normal image) further apart, thereby strengthening the model's ability to distinguish "image-text mismatch" samples. Example 5: Dual-threshold dynamic decision-making and deployment
[0049] During the inference phase (inference_server.py): Quantization acceleration: The model is quantized to INT8 format using ONNXExporter and deployed on a T4 GPU server.
[0050] Dynamic decision-making logic: Set a safety threshold T1 and a high-risk threshold T2, and T1 <T2。
[0051] If the predicted probability P∈[T1,T2], proceed to the fuzzy state processing flow: calculate the difference Δ=|P_text-P_fusion| between the text branch prediction result P_text and the fusion result P_fusion (i.e., modal conflict degree). If Δ>δ, mark the request as REVIEW_REQUIRED and route it to the manual review queue; otherwise, output according to the P_fusion value.
[0052] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.
Claims
1. A short video content risk detection method based on a multimodal fusion model, characterized in that: The method includes the following steps: (1) Acquire the short video to be detected, and separate the video stream and the audio stream; adopt an adaptive step size strategy to extract key frames, and resample the audio; at the same time, extract the original text stream through Optical Character Recognition (OCR) and Automatic Speech Recognition (ASR); (2) Use a pre-trained language model to calculate the semantic perplexity and effective character density of the original OCR text lines; construct a dynamic confidence gate to filter out noise lines with discrete semantics and low character density; perform fuzzy matching recall on the filtered text to ensure that key short words are not lost, and generate a pure text sequence; (3) Extract text, visual and audio features through BERT, ResNet and Wav2Vec respectively; construct an asymmetric cross-attention module, which forcibly takes text features as query vectors and visual and audio features as key-value pairs; introduce modality sparsity constraints to suppress redundant video frames and background noise irrelevant to the copy semantics, and generate text-enhanced visual features and audio features; (4) Input the fused features into the classifier, and make a determination based on the dual-threshold logic: a result lower than the safety threshold is judged as safe, and a result higher than the high-risk threshold is judged as non-compliant; for ambiguous samples in the middle, calculate the modality conflict degree, and if the conflict degree is too high, trigger secondary deep inference or manual review.
2. The short video content risk detection method based on a multimodal fusion model according to claim 1, characterized in that: The step (2) includes the following steps: (2.1) Construct a blacklist regular expression library containing common interference words to eliminate explicit noise; (2.2) Input the text line T_i into a lightweight language model, and calculate its perplexity PPL_i. If PPL_i is too high, it is marked as low confidence; where T_i represents the i-th line of OCR / ASR text, PPL_i represents the semantic perplexity of this text line, and a larger PPL_i indicates lower semantic coherence and higher possibility of being suspected garbled or noise text; (2.3) Calculate the proportion Density_i of valid Chinese characters in the line, where Density_i represents the proportion of the number of valid characters including Chinese characters, numbers and business keywords in the i-th line of text to the total number of characters in this line, which is used to measure the information density of the text line; (2.4) Set the fusion scoring function Score_i = α·(1 / PPL_i) + β·Density_i, and discard the line when Score_i < Threshold; where Score_i represents the comprehensive confidence score of the i-th line of text, α and β represent the weight coefficients of the semantic perplexity term and the character density term, and α+β=1, Threshold represents the text cleaning threshold, which is preset or dynamically updated according to the statistical distribution of the sample set.
3. The short video content risk detection method based on a multimodal fusion model according to claim 2, characterized in that: In the step (3), construct parallel feature extraction channels: The text channel uses the BERT-Base-Chinese model. The input is a cleaned token sequence, and the [CLS] vector is extracted as the global semantic representation V_text∈R^768. Here, V_text represents the global semantic feature vector of the text modality, and 768 represents the output dimension of the hidden layer of the BERT-Base-Chinese model, which is used to represent the contextual semantic information of the short video text content. The visual channel uses ResNet-50 to encode the sampled keyframes, extracts Layer4 output, and performs global average pooling to obtain the serialized visual features V_vis∈R^(T×1024); where V_vis represents the visual modality feature sequence, T represents the number of sampled keyframes, and 1024 represents the visual feature dimension of each keyframe after ResNet-50 encoding. The audio channel resamples the audio to 16kHz, extracts the Log-Mel spectrogram, and extracts the acoustic features V_aud∈R^(S×128) through the VGGish network; where V_aud represents the audio modal feature sequence, S represents the number of time segments after audio segmentation, and 128 represents the acoustic feature dimension corresponding to each time segment.
4. The short video content risk detection method based on a multimodal fusion model according to claim 3, characterized in that: In step (3), V_vis and V_aud are mapped to 768 dimensions through a fully connected layer to obtain V'_vis and V'_aud respectively; the text feature V_text is used as the query vector Query, and the visual feature V'_vis and / or audio feature V'_aud are used as the key and value, respectively, to calculate the attention weight matrix A=Softmax((V_text·(V'_vis)^T) / √d); where d represents the common feature dimension of Query and Key, d=768, and A represents the attention weight of text semantics on each visual frame or audio segment.
5. The short video content risk detection method based on a multimodal fusion model according to claim 4, characterized in that: In step (3), a sparsity mask is introduced before Softmax to suppress visual frames with weights lower than 0.1, forcing the model to focus on images that are strongly related to the text.
6. The short video content risk detection method based on a multimodal fusion model according to claim 5, characterized in that: In step (3), a text-guided visual enhancement feature V'_vis is generated and fused with the original text feature through a Sigmoid gate: F = V_text + σ(W·V'_vis)⊙V'_vis; where F represents the fused multimodal representation, W represents the learnable gate weight matrix, σ(·) represents the Sigmoid activation function, ⊙ represents element-wise multiplication, and V'_vis represents the visual enhancement feature after cross-attention weighting.
7. The short video content risk detection method based on a multimodal fusion model according to claim 6, characterized in that: Step (3) also includes model training and loss function: In the training engine, design a composite loss function: In Focal Loss, L_focal = -α_t(1-p_t)^γ log(p_t), where γ=2.0; p_t represents the predicted probability of the model for the true category, α_t represents the category balance factor, and γ represents the focusing factor, which is used to reduce the loss weight of easily classified samples and enhance the contribution of hard-to-classify and minority-class risk samples to model training; In the multimodal contrastive loss, the distance between the matched non-compliant text and non-compliant images is shortened in the feature space, and the distance between the non-compliant text and normal images is pushed away, so as to strengthen the model's discriminative power for samples with mismatched text and images.
8. The short video content risk detection method based on a multimodal fusion model according to claim 7, characterized in that: In the step (4), Quantization acceleration: Use ONNXExporter to quantize the model into INT8 format, and deploy it on a T4 GPU server; Dynamic decision logic: Set a safety threshold T1 and a high-risk threshold T2, where T1<T2; if the predicted probability P<T1, it is determined as safe; if P>T2, it is determined as high-risk non-compliant; if P∈[T1,T2], enter the ambiguous state processing flow, calculate the difference Δ=|P_text-P_fusion| between the text branch prediction result P_text and the fusion result P_fusion, if Δ>δ, mark the request as REVIEW_REQUIRED and route it to the manual review queue, otherwise output according to the risk level of P_fusion.
9. A short video content risk detection device based on a multimodal fusion model, characterized in that: includes: A multi-source data acquisition and preprocessing module, which is configured to acquire the short video to be detected, and separate the video stream and the audio stream; An adaptive step size strategy is adopted to extract key frames, and the audio is resampled; Simultaneously extract the original text stream through Optical Character Recognition (OCR) and Automatic Speech Recognition (ASR); A semantic consistency-based adaptive text cleaning module, which is configured to calculate the semantic perplexity and effective character density of the original OCR text lines using a pre-trained language model; Construct a dynamic confidence gate to filter out noise lines with discrete semantics and low character density; perform fuzzy matching and recall on the filtered text to ensure that key short words are not lost, and generate a pure text sequence; A text anchor-guided asymmetric feature extraction and fusion module, which is configured to extract text, visual, and audio features through BERT, ResNet, and Wav2Vec respectively; construct an asymmetric cross-attention module, forcing the text feature to be the query vector, and the visual and audio features to be the key-value pairs; introduce modality sparsity constraints to suppress redundant video frames and background noise that are irrelevant to the semantics of the text, and generate text-enhanced visual features and audio features; A dynamic decision module based on risk sensitivity, which is configured to input the fused features into the classifier, and make a determination based on the dual-threshold logic: a sample below the safety threshold is determined to be safe, and a sample above the high-risk threshold is determined to be non-compliant; for ambiguous samples in the middle, the modality conflict degree is calculated, and if the conflict degree is too high, secondary deep inference or manual review is triggered.
