Content auditing method and system
By extracting and fusion of multimodal information and using the multimodal Transformer model for content auditing, the problems of inefficient content auditing and difficulty in understanding deep meanings in the existing technology are solved, and efficient and accurate content audit is achieved.
Patent Information
- Application Number
- CN202411537773.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-30
- Publication Date
- 2025-02-18
AI Technical Summary
Existing content review techniques are inefficient and difficult to understand the deep meaning of multimedia content, especially rhetorical techniques such as metaphors and satire.
By extracting multimodal information of the content to be reviewed, including acoustic features, word embedding vectors of text sequences, professional vocabulary feature information in the field of rhetoric, video images and subtitle text, etc., information fusion is carried out, and the fusion information is input into the multimodal Transformer Model to obtain content review results.
It improves the efficiency and accuracy of content review and realizes the effect of automatically detecting and filtering inappropriate content.
Smart Images
Figure CN119475229A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of computer science and artificial intelligence technology, and more specifically to a content review method and system. Background Art
[0002] With the rapid development of the Internet, a large amount of user-generated content (UGC) such as text, pictures, videos, etc. has spread rapidly on social platforms, news websites and forums. These contents may contain illegal, sensitive or inappropriate information. Current content review mainly relies on manual review, which is inefficient and costly. At the same time, when the review system understands multimedia information such as text, pictures, and videos, it is difficult to grasp its deep meaning due to the complexity of semantic understanding, especially for rhetoric such as metaphor and irony, which increases the difficulty of understanding.
[0003] Therefore, how to propose a content review method and system that can automatically detect and filter inappropriate content to improve review efficiency and accuracy is an urgent problem that technical personnel in this field need to solve. Summary of the invention
[0004] In view of this, the present invention provides a content review method and system, which can automatically detect and filter inappropriate content to improve review efficiency and accuracy. In order to achieve the above purpose, the present invention adopts the following technical solutions:
[0005] A content audit method comprising:
[0006] Obtain multimodal information based on the content to be reviewed;
[0007] The extracted acoustic features are input into the encoder model to obtain the deep information of the acoustic features;
[0008] Input the word embedding vector of the text sequence into the prediction model to obtain the deep information of the word embedding vector;
[0009] Inputting the feature information of professional vocabulary in the field of rhetoric into the bias encoding model, the deep information of professional vocabulary in the field of rhetoric is obtained;
[0010] Input the video image into the face recognition model to obtain the deep information of the face;
[0011] Fuse multiple deep information to obtain multimodal fusion information corresponding to the audit content;
[0012] Input the multimodal fusion information into the multimodal Transformer model to obtain the content review results.
[0013] Optionally, the method further includes extracting frames from the video image, identifying subtitle text in the extracted frame video image, inputting the subtitle text in the video image into an encoder model for encoding, and obtaining deep information of the subtitle text in the video image.
[0014] Optionally, inputting the video image into a face recognition model to obtain deep facial information includes: dividing the video image into a number of image sub-regions, inputting the several image sub-regions into the face recognition model to perform face detection, and obtaining deep facial information of the video image.
[0015] Optionally, the multimodal information includes: acoustic information, text information, professional vocabulary information, video image information, and image subtitle text information of the content to be reviewed.
[0016] Optionally, before inputting the professional vocabulary feature information of the rhetoric field into the bias encoding model, the following steps are further performed:
[0017] Obtaining rhetorical sentence feature combinations;
[0018] Performing rhetorical sentence pattern expansion processing based on the rhetorical sentence pattern feature combination to obtain a new sentence containing the rhetorical sentence pattern feature combination;
[0019] A label is set for the newly added sentence, and the label is used to indicate that the newly added sentence contains a rhetorical sentence feature combination.
[0020] Optionally, also include:
[0021] Using the newly added sentences with the label as new training samples for the rhetorical sentence pattern feature extraction model to iteratively train the rhetorical sentence pattern feature extraction model;
[0022] Obtain the trained rhetorical sentence feature extraction model;
[0023] The rhetorical sentence pattern feature extraction model completed through the training extracts the rhetorical sentence pattern feature combination from the sentence to be processed.
[0024] Optionally, obtaining a new sentence containing the rhetorical sentence feature combination includes:
[0025] According to the rhetorical sentence pattern feature word combination, sentence matching is performed in the database to determine the matching sentence in the database that contains the rhetorical sentence pattern feature combination;
[0026] A first arrangement sentence arranged according to a first rule is selected from the matching sentences by a rhetoric discriminator;
[0027] filtering out second arranged sentences arranged according to a second rule from the first arranged sentences by using a rhetorical implication classifier;
[0028] The second arrangement statement is combined with the first arrangement statement as a newly added statement.
[0029] Optionally, the information fusion of multiple deep information to obtain multimodal fusion information corresponding to the audit content includes:
[0030] The deep information of professional vocabulary and subtitle text is cross-processed to obtain the attention features of professional vocabulary-subtitle text; the deep information of acoustic features and the deep information of faces are cross-processed to obtain the attention features of acoustic features-faces; the deep information of word embedding vectors and the deep information of professional vocabulary are cross-processed to obtain the attention features of word embedding vectors-professional vocabulary;
[0031] According to the attention features of professional vocabulary-caption text, the attention features of acoustic features-faces, the attention features of word embedding vectors-professional vocabulary and the fusion algorithm, a first multimodal fusion representation is obtained; the first multimodal fusion representation is input into a feedforward neural network, and the multimodal fusion information is obtained through an updating algorithm.
[0032] Optionally, a content review system comprising:
[0033] Acquisition module: used to obtain multimodal information based on the content to be reviewed;
[0034] Acoustic feature processing module: used to input the extracted acoustic features into the encoder model to obtain deep information of acoustic features;
[0035] Text sequence processing module: used to input the word embedding vector of the text sequence into the prediction model to obtain the deep information of the word embedding vector;
[0036] Professional vocabulary processing module: used to input the professional vocabulary feature information in the rhetoric field into the bias encoding model to obtain the deep information of professional vocabulary in the rhetoric field;
[0037] Face recognition module: used to input video images into the face recognition model to obtain deep facial information;
[0038] Fusion module: used to fuse multiple deep information to obtain multimodal fusion information corresponding to the audit content;
[0039] Audit module: used to input multimodal fusion information into the multimodal Transformer model to obtain content audit results.
[0040] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a content review method and system, which has the following beneficial effects:
[0041] The present invention proposes a content review method and system, including: obtaining multimodal information according to the content to be reviewed; inputting the extracted acoustic features into the encoder model to obtain deep information of the acoustic features; inputting the word embedding vector of the text sequence into the prediction model to obtain deep information of the word embedding vector; inputting the professional vocabulary feature information in the rhetoric field into the bias encoding model to obtain the professional vocabulary deep information in the rhetoric field; inputting the video image into the face recognition model to obtain the deep information of the face; fusing multiple deep information to obtain multimodal fusion information corresponding to the review content; inputting the multimodal fusion information into the multimodal Transformer model to obtain the content review result. The present invention considers that it is difficult to grasp the deep meaning when reviewing the content and understanding multimedia information such as text, pictures, and videos, especially for rhetorical techniques such as metaphors and irony, which increases the difficulty of understanding. By extracting the multimodal information of the content to be reviewed, including but not limited to the professional vocabulary feature information in the rhetoric field, multiple deep information are fused to obtain the multimodal fusion information corresponding to the review content; the multimodal fusion information is input into the multimodal Transformer model to obtain the content review result. It achieves the effect of improving review efficiency and accuracy, and automatically detecting and filtering inappropriate content. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.
[0043] Figure 1 A flow chart of a content review method provided by the present invention. DETAILED DESCRIPTION
[0044] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0045] The embodiment of the present invention discloses a content review method, such as Figure 1 As shown, including:
[0046] Obtain multimodal information based on the content to be reviewed;
[0047] The extracted acoustic features are input into the encoder model to obtain the deep information of the acoustic features;
[0048] Input the word embedding vector of the text sequence into the prediction model to obtain the deep information of the word embedding vector;
[0049] Inputting the feature information of professional vocabulary in the field of rhetoric into the bias encoding model, the deep information of professional vocabulary in the field of rhetoric is obtained;
[0050] Input the video image into the face recognition model to obtain the deep information of the face;
[0051] Fuse multiple deep information to obtain multimodal fusion information corresponding to the audit content;
[0052] Input the multimodal fusion information into the multimodal Transformer model to obtain the content review results.
[0053] Furthermore, the method also includes extracting frames from video images, identifying subtitle text in the extracted frame video images, inputting the subtitle text in the video images into an encoder model for encoding, and obtaining deep information of the subtitle text in the video images.
[0054] In a specific embodiment, the encoder model is a long short-term memory network, h t =LSTM(h t-1 , x t ) Among them, LSTM is the unit function of the long short-term memory network, h t is the deep information corresponding to the subtitle text at time t, h t-1 is the deep information corresponding to the subtitle text at the t-1th moment, x t is the subtitle text at time t.
[0055] In a specific embodiment, the encoder model is a long short-term memory network. p =LSTM(o p-1 ,h p ) Among them, LSTM is the unit function of the long short-term memory network, o p is the deep information corresponding to the acoustic feature at time p, o p-1 is the deep information corresponding to the acoustic feature at the p-1th moment, h p is the acoustic feature at time p.
[0056] In a specific implementation, the bias encoding model is a long short-term memory network, which obtains deep information corresponding to the professional vocabulary sequence in the field of rhetoric: k =LSTM(b k-1 , z k ), where LSTM is the unit function of the long short-term memory network, b k-1 is the deep information corresponding to the word in the professional vocabulary sequence in the rhetoric field at the k-1th position, zk The embedding vector of the word at position k in the professional vocabulary sequence in the field of rhetoric.
[0057] In a specific embodiment, the prediction model is a long short-term memory network, c t =LSTM(c j-1 ,y j ), where LSTM is the unit function of the long short-term memory network, c j-1 is the deep information corresponding to the word at position j-1, y j is the embedding vector of the word at position j.
[0058] Furthermore, the step of inputting the video image into a face recognition model to obtain deep face information includes: dividing the video image into a number of image sub-regions, inputting the several image sub-regions into the face recognition model to perform face detection, and obtaining deep face information of the video image.
[0059] In a specific implementation, the face recognition model includes:
[0060] S1: Detect and select facial key points for each video frame, including:
[0061] The key points of the face include eyes, nose, mouth, eyebrows, jaw line and face contour features. The key points of the face are detected for each video frame, and the FaceMesh algorithm in mediapipe is used to detect the key points of the face. In a specific embodiment, video frames are extracted based on the medication video recorded by the patient, and the key points of the face are detected for each video frame. The FaceMesh algorithm in mediapipe is used to detect the key points of the face. The FaceMesh algorithm detects 468 key points of the face, which contain rich facial features. The matching of the mask and the face does not require such fine features. The selection of key points includes facial key features: eyes, nose, mouth, eyebrows, jaw line, face contour, etc. 75 of the 468 key points are selected as the final facial key points to be obtained. 75 key points, 0-16 are jaw contours. 17-21 are the left eyebrow, 22-26 are the right eyebrow, 27-35 are the nose bridge and nose, 36-41 are the left eye, 42-47 are the right eye, 48-59 are the outer contour of the mouth, 60-67 are the inner contour of the lips, and 68-74 are the forehead contour; it also includes obtaining a facial image of a human body when performing rhetorical language description, and obtaining a rhetorical mask according to the facial image and key points of the human body;
[0062] S2: constructing a rhetorical mask and marking the key points of the mask, selecting a rhetorical mask that is suitable for the human face, wherein the key points of the rhetorical mask are fixed, and after the rhetorical mask is selected, marking the key points, wherein the key point marking matches the selected human face key points, specifically, matches the 75 key points determined above;
[0063] S3: training a face recognition model based on a deep neural network according to the key points of the rhetorical mask, the key points of the face, and the key points of normal facial expressions;
[0064] S4: Acquire a real-time frame image, acquire key points of the real-time frame image, identify the key points of the real-time frame image through a face recognition model, and obtain deep face information of the video image.
[0065] Furthermore, the multimodal information includes: acoustic information, text information, professional vocabulary information, video image information, and image subtitle text information of the content to be reviewed.
[0066] Furthermore, before inputting the professional vocabulary feature information in the rhetoric field into the bias coding model, the following steps are also performed:
[0067] Obtaining rhetorical sentence feature combinations;
[0068] Performing rhetorical sentence pattern expansion processing based on the rhetorical sentence pattern feature combination to obtain a new sentence containing the rhetorical sentence pattern feature combination;
[0069] A label is set for the newly added sentence, and the label is used to indicate that the newly added sentence contains a rhetorical sentence feature combination.
[0070] Furthermore, it also includes:
[0071] Using the newly added sentences with the label as new training samples for the rhetorical sentence pattern feature extraction model to iteratively train the rhetorical sentence pattern feature extraction model;
[0072] Obtain the trained rhetorical sentence feature extraction model;
[0073] The rhetorical sentence pattern feature extraction model completed through the training extracts the rhetorical sentence pattern feature combination from the sentence to be processed.
[0074] Furthermore, the obtaining of the newly added sentence containing the rhetorical sentence pattern feature combination includes:
[0075] According to the rhetorical sentence pattern feature word combination, sentence matching is performed in the database to determine the matching sentence in the database that contains the rhetorical sentence pattern feature combination;
[0076] A first arrangement sentence arranged according to a first rule is selected from the matching sentences by a rhetoric discriminator;
[0077] filtering out second arranged sentences arranged according to a second rule from the first arranged sentences by using a rhetorical implication classifier;
[0078] The second arrangement statement is combined with the first arrangement statement as a newly added statement.
[0079] In a specific implementation, the newly added statement is also judged:
[0080] The output results of the four networks, namely, the encoder composed of the long short-term memory network, the prediction network composed of the long short-term memory network, the encoder composed of the long short-term memory network, and the bias encoding network composed of the long short-term memory network, are input through a fusion network composed of a multi-layer perceptron to predict the next word of the text sequence;
[0081] The fusion network composed of the multi-layer perceptron fuses the output results of the four networks, namely, the encoder composed of the long short-term memory network, the prediction network composed of the second long short-term memory network, the encoder composed of the long short-term memory network, and the bias encoding network composed of the third long short-term memory network, to predict the next word of the text sequence according to the following formula:
[0082] P(y j+1 )=MLP([c t ,b k ,h t ,o p ]), where MLP is the function of the multi-layer perceptron, where h t is the deep information corresponding to the acoustic feature sequence at the tth moment extracted by the encoder composed of the long short-term memory network, o p is the deep information corresponding to the acoustic feature at time p, b k is the deep information corresponding to the word of the professional vocabulary sequence in the rhetoric field at the kth position of the bias encoding network composed of long short-term memory networks, c t The deep information corresponding to the word at the tth position of the prediction network composed of the long short-term memory network;
[0083] According to the Euclidean distance between the deep information of the newly added sentence and the deep information corresponding to the predicted word, when a preset threshold is met, it is saved as the final newly added sentence.
[0084] Furthermore, the information fusion of multiple deep information to obtain multimodal fusion information corresponding to the audit content includes:
[0085] The deep information of professional vocabulary and subtitle text is cross-processed to obtain the attention features of professional vocabulary-subtitle text; the deep information of acoustic features and the deep information of faces are cross-processed to obtain the attention features of acoustic features-faces; the deep information of word embedding vectors and the deep information of professional vocabulary are cross-processed to obtain the attention features of word embedding vectors-professional vocabulary;
[0086] According to the attention features of professional vocabulary-caption text, the attention features of acoustic features-faces, the attention features of word embedding vectors-professional vocabulary and the fusion algorithm, a first multimodal fusion representation is obtained; the first multimodal fusion representation is input into a feedforward neural network, and the multimodal fusion information is obtained through an updating algorithm.
[0087] In a specific implementation, according to the professional vocabulary deep information, the subtitle text deep information and the first cross-attention feature extraction algorithm, the output features of several first cross-attention heads are obtained; according to the output features of several first cross-attention heads and the first multi-layer perceptron algorithm, the attention features of professional vocabulary-subtitle text are obtained;
[0088] According to the acoustic feature deep information, the face deep information and the second cross attention feature extraction algorithm, the output features of several second cross attention heads are obtained; according to the output features of several second cross attention heads and the second multi-layer perceptron algorithm, the attention features of acoustic features-faces are obtained;
[0089] According to the deep information of the word embedding vector, the deep information of the professional vocabulary and the third cross-attention feature extraction algorithm, the output features of several third cross-attention heads are obtained; according to the output features of several third cross-attention heads and the third multi-layer perceptron algorithm, the attention features of the word embedding vector-professional vocabulary are obtained;
[0090] According to the attention features of professional vocabulary-caption text, the attention features of acoustic features-faces, the attention features of word embedding vectors-professional vocabulary and a preset fusion algorithm, a first multimodal fusion representation is obtained; the first multimodal fusion representation is input into a feedforward neural network, and multimodal fusion information is obtained through an updating algorithm.
[0091] Specifically, the cross-attention feature extraction algorithm is the CLIP cross-attention feature extraction algorithm, which combines the features of images and texts and realizes multimodal feature extraction through the cross-attention mechanism; the fusion algorithm is the MVAE variational autoencoder fusion algorithm.
[0092] In a specific implementation, a content review system includes:
[0093] Acquisition module: used to obtain multimodal information based on the content to be reviewed;
[0094] Acoustic feature processing module: used to input the extracted acoustic features into the encoder model to obtain deep information of acoustic features;
[0095] Text sequence processing module: used to input the word embedding vector of the text sequence into the prediction model to obtain the deep information of the word embedding vector;
[0096] Professional vocabulary processing module: used to input the professional vocabulary feature information in the rhetoric field into the bias encoding model to obtain the deep information of professional vocabulary in the rhetoric field;
[0097] Face recognition module: used to input video images into the face recognition model to obtain deep facial information;
[0098] Fusion module: used to fuse multiple deep information to obtain multimodal fusion information corresponding to the audit content;
[0099] Audit module: used to input multimodal fusion information into the multimodal Transformer model to obtain content audit results.
[0100] The present invention considers that it is difficult to grasp the deep meaning when reviewing content and understanding multimedia information such as text, pictures, and videos, especially for metaphors, irony, etc., which increases the difficulty of understanding. By extracting multimodal information of the content to be reviewed, including but not limited to professional vocabulary feature information in the field of rhetoric, the present invention fuses multiple deep information to obtain multimodal fusion information corresponding to the review content; the multimodal fusion information is input into the multimodal Transformer model to obtain the content review result. The review efficiency and accuracy are improved, and inappropriate content is automatically detected and filtered.
[0101] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same and similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts refer to the method part description. The above description of the disclosed embodiments enables professionals and technicians in this field to implement or use the present invention. Various modifications to these embodiments will be obvious to professionals and technicians in this field, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown in this article, but will comply with the widest range consistent with the principles and novel features disclosed herein.
Claims
1. A content review method, characterized in that: include: Obtain multimodal information based on the content to be reviewed; The extracted acoustic features are input into the encoder model to obtain the deep information of the acoustic features; Input the word embedding vector of the text sequence into the prediction model to obtain the deep information of the word embedding vector; Inputting the feature information of professional vocabulary in the field of rhetoric into the bias encoding model, the deep information of professional vocabulary in the field of rhetoric is obtained; Input the video image into the face recognition model to obtain the deep information of the face; Fuse multiple deep information to obtain multimodal fusion information corresponding to the audit content; Input the multimodal fusion information into the multimodal Transformer model to obtain the content review results.
2. A content review method according to claim 1, characterized in that: The method also includes extracting frames from video images, identifying subtitle text in the extracted frame video images, inputting the subtitle text in the video images into an encoder model for encoding, and obtaining deep information of the subtitle text in the video images.
3. A content review method according to claim 1, characterized in that: The step of inputting the video image into a face recognition model to obtain deep face information includes: dividing the video image into a plurality of image sub-regions, inputting the plurality of image sub-regions into a face recognition model to perform face detection, and obtaining deep face information of the video image.
4. A content review method according to claim 1, characterized in that: The multimodal information includes: acoustic information, text information, professional vocabulary information, video image information, and image subtitle text information of the content to be reviewed.
5. A content review method according to claim 1, characterized in that: Before inputting the professional vocabulary feature information of the rhetoric field into the bias encoding model, the method further includes: Obtaining rhetorical sentence feature combinations; Performing rhetorical sentence pattern expansion processing based on the rhetorical sentence pattern feature combination to obtain a new sentence containing the rhetorical sentence pattern feature combination; A label is set for the newly added sentence, and the label is used to indicate that the newly added sentence contains a rhetorical sentence feature combination.
6. A content review method according to claim 5, characterized in that: Also includes: Using the newly added sentences with the label as new training samples for the rhetorical sentence pattern feature extraction model to iteratively train the rhetorical sentence pattern feature extraction model; Obtain the trained rhetorical sentence feature extraction model; The rhetorical sentence pattern feature extraction model completed through the training extracts the rhetorical sentence pattern feature combination from the sentence to be processed.
7. A content review method according to claim 5, characterized in that: The obtaining of the newly added sentence containing the rhetorical sentence pattern feature combination includes: According to the rhetorical sentence pattern feature word combination, sentence matching is performed in the database to determine the matching sentence in the database that contains the rhetorical sentence pattern feature combination; A first arrangement sentence arranged according to a first rule is selected from the matching sentences by a rhetoric discriminator; filtering out second arranged sentences arranged according to a second rule from the first arranged sentences by using a rhetorical implication classifier; The second arrangement statement is combined with the first arrangement statement as a newly added statement.
8. A content review method according to claim 1, characterized in that: The multi-modal fusion information corresponding to the audit content obtained by fusing multiple deep information includes: The deep information of professional vocabulary and subtitle text is cross-processed to obtain the attention features of professional vocabulary-subtitle text; the deep information of acoustic features and the deep information of faces are cross-processed to obtain the attention features of acoustic features-faces; the deep information of word embedding vectors and the deep information of professional vocabulary are cross-processed to obtain the attention features of word embedding vectors-professional vocabulary; According to the attention features of professional vocabulary-caption text, the attention features of acoustic features-faces, the attention features of word embedding vectors-professional vocabulary and the fusion algorithm, a first multimodal fusion representation is obtained; the first multimodal fusion representation is input into a feedforward neural network, and the multimodal fusion information is obtained through an updating algorithm.
9. A content review system, characterized in that: include: Acquisition module: used to obtain multimodal information based on the content to be reviewed; Acoustic feature processing module: used to input the extracted acoustic features into the encoder model to obtain deep information of acoustic features; Text sequence processing module: used to input the word embedding vector of the text sequence into the prediction model to obtain the deep information of the word embedding vector; Professional vocabulary processing module: used to input the professional vocabulary feature information in the rhetoric field into the bias encoding model to obtain the deep information of professional vocabulary in the rhetoric field; Face recognition module: used to input video images into the face recognition model to obtain deep facial information; Fusion module: used to fuse multiple deep information to obtain multimodal fusion information corresponding to the audit content; Audit module: used to input multimodal fusion information into the multimodal Transformer model to obtain content audit results.