Short video multi-dimensional analysis and credibility evaluation method and system

Abstract

The invention discloses a short video multi-dimensional analysis and credibility evaluation method and system, and relates to the technical field of short video recommendation, the short video multi-dimensional analysis and credibility evaluation method mainly comprises the following steps: analyzing a short video to be analyzed to obtain keywords, short video metadata and short video theme information, obtaining a similar theme version group, an initial version of the short video to be analyzed, a propagation path and a propagation node; and according to the propagation node, the short video theme information, the keyword and the short video metadata, performing calculation to obtain a comprehensive credibility score. By implementing the short video multi-dimensional analysis and credibility evaluation method and system provided by the invention, the understanding of the short video content and the accuracy of source tracing can be improved, and the user experience is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of short video recommendation, and more specifically, to a short video multi-dimensional analysis and credibility assessment method and system. Background Art

[0002] Currently, short videos spread rapidly and widely. Users usually receive a large amount of information in a very short period of time, but they often do not spend too much energy to verify the authenticity of the video or deeply understand the content. In addition, the secondary processing and cross-platform dissemination of the original video content by some video publishers will reduce the authenticity of the information and easily lead to incorrect public opinion. The existing short video platforms have a large number of videos, but they fail to effectively identify and trace the source of video information. The video content that is taken out of context or tampered with is somewhat misleading. At the same time, the short video platform has not provided authoritative popular science on processed or more complex video information. Users with different educational backgrounds have different levels of understanding of the video, which makes users with lower education levels very likely to believe the incorrect interpretation of the comment area or video publisher. Summary of the invention

[0003] The purpose of the present invention is to provide a short video multi-dimensional analysis and credibility assessment method and system, which can improve the understanding of short video content and the accuracy of source tracing, and enhance user experience.

[0004] The present invention provides a short video multi-dimensional analysis and credibility assessment method, comprising the following steps: S1: obtaining a short video to be analyzed, analyzing the short video to be analyzed, and obtaining keywords, short video metadata and short video subject information; S2: obtaining a similar subject version group, an initial version of the short video to be analyzed, a propagation path and a propagation node based on the short video subject information and a video database; S3: obtaining a comprehensive credibility score based on the propagation node, short video subject information, keywords and short video metadata.

[0005] Furthermore, step S1 of the above-mentioned short video multi-dimensional analysis and credibility assessment method specifically includes: S11: obtaining the short video to be analyzed, and obtaining short video metadata, voice data and image data based on the short video to be analyzed, and obtaining text data based on the voice data and image data; S12: obtaining keywords and emotions based on the text data and short video metadata using a pre-trained language model; S13: obtaining objects, scenes, behaviors and events based on the short video to be analyzed using a convolutional neural network and a video classification model; S14: obtaining short video topic information based on keywords, emotions, scenes, behaviors and events using a multimodal Transformer model and a Softmax classifier.

[0006] Furthermore, step S2 of the above-mentioned short video multi-dimensional analysis and credibility assessment method specifically includes: S21: based on the short video subject information and the video database, using the clustering method, obtain a similar subject version group; S22: based on the similar subject version group and the video database, obtain the initial version of the short video to be analyzed; S23: based on the initial version of the short video to be analyzed and the video database, using the AI ​​model, obtain the propagation path and propagation nodes.

[0007] Furthermore, step S3 of the above short video multi-dimensional analysis and credibility assessment method specifically includes: S31: obtaining a video source credibility score based on propagation nodes; S32: obtaining a video content consistency score based on short video subject information and keywords; S33: obtaining a propagation path score based on propagation nodes; S34: obtaining a timestamp verification score based on short video metadata; S35: obtaining a comprehensive credibility score based on the video source credibility score, video content consistency score, propagation path score and timestamp verification score.

[0008] Furthermore, the above-mentioned short video multi-dimensional analysis and credibility assessment method also includes: S4: obtaining comprehensive prompt content based on short video metadata, short video theme information and keywords; S5: obtaining real-time data stream, and updating the comprehensive credibility score based on the real-time data stream and propagation nodes; S6: obtaining user prompt content based on the comprehensive credibility score and comprehensive prompt content.

[0009] The present invention also provides a short video multi-dimensional analysis and credibility assessment system, comprising the following modules: a short video content analysis module, configured to obtain a short video to be analyzed, analyze the short video to be analyzed, and obtain keywords, short video metadata and short video theme information; a source tracing module, configured to obtain a similar theme version group, an initial version of the short video to be analyzed, a propagation path and a propagation node based on the short video theme information and a video database; a credibility scoring module, configured to obtain a comprehensive credibility score based on the propagation node, short video theme information, keywords and short video metadata.

[0010] Furthermore, the short video content analysis module of the above-mentioned short video multi-dimensional analysis and credibility assessment system is specifically configured as follows: obtain the short video to be analyzed, obtain short video metadata, voice data and image data based on the short video to be analyzed, and obtain text data based on the voice data and image data; obtain keywords and emotions based on the text data and short video metadata using a pre-trained language model; obtain objects, scenes, behaviors and events based on the short video to be analyzed using a convolutional neural network and a video classification model; obtain short video topic information based on keywords, emotions, scenes, behaviors and events using a multimodal Transformer model and a Softmax classifier.

[0011] Furthermore, the source tracing module of the above-mentioned short video multi-dimensional analysis and credibility assessment system is specifically configured as follows: based on the short video subject information and the video database, a clustering method is used to obtain a similar subject version group; based on the similar subject version group and the video database, an initial version of the short video to be analyzed is obtained; based on the initial version of the short video to be analyzed and the video database, an AI model is used to obtain the propagation path and propagation nodes.

[0012] Furthermore, the credibility scoring module of the above-mentioned short video multi-dimensional analysis and credibility assessment system is specifically configured as follows: according to the propagation nodes, a video source credibility score is obtained; according to the short video theme information and keywords, a video content consistency score is obtained; according to the propagation nodes, a propagation path score is obtained; according to the short video metadata, a timestamp verification score is obtained; according to the video source credibility score, the video content consistency score, the propagation path score and the timestamp verification score, a comprehensive credibility score is obtained.

[0013] Furthermore, the above-mentioned short video multi-dimensional analysis and credibility assessment system also includes: an authoritative science popularization module, configured to obtain comprehensive prompt content based on short video metadata, short video theme information and keywords; a real-time monitoring and update module, configured to obtain real-time data streams, and update the comprehensive credibility score based on real-time data streams and propagation nodes; a user prompt module, configured to obtain user prompt content based on the comprehensive credibility score and comprehensive prompt content.

[0014] The implementation of the short video multi-dimensional analysis and credibility assessment method and system provided by the present invention has the following beneficial effects: The present invention analyzes the short video to be analyzed through multimodal data fusion technology and AI model, obtains keywords, short video metadata, short video theme information, similar theme version group, initial version of the short video to be analyzed, propagation path and propagation nodes and other information, and calculates the comprehensive credibility score of the short video, thereby improving the accuracy of understanding the content of the short video and tracing the source; the system provides the user with the video credibility score and popularizes science in the form of semi-transparent card prompts, helping them to better understand the video content while trying not to be disturbed by the prompt module, comprehensively improving the quality of the current short video and user experience, increasing the depth of user understanding while also reducing the risk of this group being misled. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which: Figure 1 It is a flow chart of the short video multi-dimensional analysis and credibility assessment method provided by the present invention; Figure 2 It is a flow chart of the short video multi-dimensional analysis, credibility evaluation and popular science method provided by the present invention. DETAILED DESCRIPTION

[0016] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described in detail with reference to the accompanying drawings.

[0017] Figure 1 A schematic diagram of a short video multi-dimensional analysis and credibility assessment method of this embodiment is shown. In this embodiment, the short video multi-dimensional analysis and credibility assessment method includes the following steps: S1: Obtain the short video to be analyzed, analyze the short video to be analyzed, and obtain keywords, short video metadata and short video theme information; Specifically, step S1 includes: S11: Obtain a short video to be analyzed, obtain short video metadata, voice data and image data according to the short video to be analyzed, and obtain text data according to the voice data and image data; Specifically, short video metadata includes subtitles, titles, and tags; It should be noted that in step S11, speech-to-text is used to identify text information from the image through OCR technology and extract metadata such as subtitles, titles, and tags from the video with subtitles; S12: Obtain keywords and emotions based on text data and short video metadata using a pre-trained language model; Specifically, the pre-trained language model is a BERT pre-trained language model; It should be noted that in step S12, the BERT pre-trained language model is used to extract accurate keywords in combination with contextual information such as text data and short video metadata, and the emotions expressed in the video are classified; S13: Based on the short video to be analyzed, the convolutional neural network and video classification model are used to obtain objects, scenes, behaviors and events; Specifically, the video classification model is the TimeSformer model; It should be noted that in step S13, the main objects and scenes in the video are automatically detected by the convolutional neural network, and the TimeSformer model is used to capture complex behaviors and events to extract the scenes, tasks, and events in the video; S14: Based on keywords, emotions, scenes, behaviors and events, the multimodal Transformer model and Softmax classifier are used to obtain the short video topic information; Specifically, step S14 includes: S141: Assign initial weights and fuse keywords, emotions, scenes, behaviors, and events to obtain preliminary fused features; Specifically, the weight of the scene is 0.2, the weight of the behavior is 0.3, the weight of the event is 0.2, the weight of the keyword is 0.2, and the weight of the emotion is 0.1; S142: Based on the initially fused features, the short video topic information is obtained using a multimodal Transformer model and a Softmax classifier; It should be noted that, in step S142, the initially fused features are input into the multimodal Transformer model, and the complex associations among the elements are captured through the self-attention mechanism. Then, the multi-layer Transformer model outputs the deeply fused feature vector to improve the accuracy of topic recognition, that is, the feature vector is input into the Softmax classifier to output the short video topic information; wherein, the short video topic information includes topic tags; Specifically, short video metadata also includes author information, time, location, and device information; It should be noted that, in step S142, the metadata of the video is analyzed, and the author information, time, location, device information, etc. are used as a reference for tracing the source of the video; wherein the time can be reflected in the form of a timestamp; S2: Based on the short video theme information and video database, obtain the similar theme version group, the initial version of the short video to be analyzed, the propagation path and the propagation node; It should be noted that in step S2, the time and location clues and multimodal features in the short video are processed and cross-compared with the database, and the video propagation path and the earliest release node are identified in combination with the prediction model; Specifically, step S2 includes: S21: Based on the short video theme information and the video database, a clustering method is used to obtain a similar theme version group; Specifically, the video database includes metadata and subject information of historical short videos; It should be noted that in step S21, the short video theme information is compared with theme information of other videos in a video database, wherein the database stores metadata of videos uploaded to the online short video platform and theme information obtained through multimodal information fusion technology; then cluster analysis is used to identify version groups with similar themes to locate other versions related to the original version; S22: Obtaining an initial version of the short video to be analyzed based on the similar topic version group and the video database; It should be noted that in step S22, by comparing the differences between the various versions of the video, the modification traces of the video at different dissemination nodes are determined, and the time and location information of the video are extracted and compared with the metadata of each version of the video stored in the database, and the non-original versions are gradually excluded, tracing back to the initial version of the video; S23: Based on the initial version of the short video to be analyzed and the video database, the propagation path and propagation nodes are obtained using the AI ​​model; It should be noted that in step S23, the propagation path and propagation nodes of the short video to be analyzed are obtained by using the AI ​​model according to the initial version of the short video to be analyzed and the video database; first, the propagation path is analyzed by using the AI ​​prediction model: by predicting and tracking the propagation path of the current video through the AI ​​model, a propagation diagram of information in the social network can be constructed, and each propagation node can be identified; then, the AI ​​model is combined to detect abnormal propagation nodes: through the classification model, the characteristics of normal user behavior are learned, and nodes that are significantly different from them are detected; that is, the behavior pattern of each node is analyzed, and nodes that are significantly different from normal user behavior are identified; for example, nodes that frequently forward a large amount of information in a short period of time may be robots or automated scripts; Specifically, the dissemination nodes include publishers and users who disseminate videos and their behavior patterns; S3: Obtain a comprehensive credibility score based on the dissemination nodes, short video topic information, keywords and short video metadata; Specifically, step S3 includes: S31: Obtaining a video source credibility score according to the propagation node; Specifically, in step S31, the credibility of the initial publisher is evaluated: the credibility of the publisher of the initial version of the video is evaluated, and an automated scoring mechanism is used. When the publisher is an authoritative organization, a certified user or an expert, with a good historical record and widely recognized content, the video source credibility score is 10 points; when the publisher's identity is basically credible and there is occasional low-quality content, the video source credibility score is 7-9 points; when the publisher's identity is uncertain and some content is unreliable, the video source credibility score is 4-6 points; when the publisher's identity contains a lot of false information, the video source credibility score is 0-3 points; the specific scoring can be accurately scored using a questionnaire scale; S32: Obtaining a video content consistency score based on the short video theme information and keywords; Specifically, in step S32, the consistency of multimodal data of visual, audio, text and other videos is evaluated: the consistency of the current video content and the initial version of the video content is evaluated, using an automated scoring mechanism. When the subject is completely consistent and the keywords and subject tags are highly overlapped, the video content consistency score is 10 points; when the subject is basically consistent and some keywords and tags are similar, the video content consistency score is 7-9 points; when the subject is somewhat different and the overlap of keywords and tags is low, the video content consistency score is 4-6 points; when the subject is completely inconsistent and the keywords and tags are irrelevant, the video content consistency score is 0-3 points; the specific scoring can be accurately scored using a questionnaire scale; the specific scoring can be accurately scored using a questionnaire scale; S33: Obtain a propagation path score according to the propagation node; Specifically, in step S33, the transparency and traceability of the video transmission path and the credibility of the transmission nodes are evaluated using an automated scoring mechanism; when all transmission nodes are trusted users and the behavior patterns are normal and stable, the transmission path score is 10 points; when most nodes are trusted and there are a few abnormal behaviors, the transmission path score is 7-9 points; when some nodes exhibit abnormal behaviors and the credibility is reduced, the transmission path score is 4-6 points; when most nodes exhibit obvious abnormalities and may be robots or malicious accounts, the transmission path score is 0-3 points; the specific scoring can be accurately scored using a questionnaire scale; S34: Obtain a timestamp verification score based on the short video metadata; Specifically, in step S34, it is determined whether the timestamp of the video matches the timeline of the event, and the timestamp of the video is compared with the time information of the related event to determine whether the timestamp reflects the actual time of the event, and an automated scoring mechanism is used; when the timestamp is completely accurate and highly consistent with the timeline of the event, the timestamp verification score is 10 points; when the timestamp is basically accurate with a few small time differences, the timestamp verification score is 7-9 points; when there are obvious differences in the timestamps, which may cause difficulties in understanding the content, the timestamp verification score is 4-6 points; when the timestamp is seriously wrong or missing, inconsistent with the event timeline, resulting in highly misleading content, the timestamp verification score is 0-3 points; the specific scoring can be accurately scored using a questionnaire scale; S35: Obtain a comprehensive credibility score according to the video source credibility score, the video content consistency score, the propagation path score, and the timestamp verification score; Specifically, step S35 includes: obtaining a comprehensive credibility score according to the video source credibility score, the video content consistency score, the propagation path score and the timestamp verification score, such as the formula:

[0018] in, For the overall credibility score, is the credibility weight of the video source, is the credibility score of the video source, is the video content consistency weight, is the video content consistency score, is the propagation path weight, Score the propagation path. timestamp validation weight, Score timestamp verification; Optionally, in step S35, a weighted comprehensive evaluation is performed on the above factors based on the scoring model, and each factor is scored separately, and each factor is scored from 0 to 10 points. After calculation, a credibility score is generated. The specific calculation formula is: comprehensive credibility score = 0.4×video source credibility score + 0.4×video content consistency score + 0.1×propagation path score + 0.1×timestamp verification score; Optionally, the short video multi-dimensional analysis and evaluation method further includes: S4: Obtain comprehensive prompt content based on short video metadata, short video theme information and keywords; Specifically, in step S4, the short video theme information is extracted, and popular science information from multiple sources, such as popular science content from authoritative institutions, experts and scholars, or credible third-party platforms, is integrated to generate concise prompt content to help users understand the background knowledge and scientific information related to the video; It should be noted that in step S4, based on the video keywords, themes and semantic analysis information extracted by natural language processing technology, relevant popular science information is found; GPT-4 is used to simplify and integrate popular science information from multiple sources to generate comprehensive prompt content; S5: Obtain real-time data streams and update comprehensive credibility scores based on the real-time data streams and propagation nodes; Specifically, in step S5, the real-time data stream of social media is combined to analyze the short video interaction and dissemination path, and the credibility score is dynamically updated to ensure that users obtain the latest evaluation results; that is, the system dynamically monitors the credibility of the video throughout its life cycle after it is released; It should be noted that in step S5, in combination with the real-time data stream of social media, a machine learning algorithm is used to classify and cluster user behaviors, identify abnormal behavior patterns, such as abnormally high forwarding volume and sudden emergence of comments, continuously analyze the newly emerged short video interactions and propagation paths, and adjust the credibility score accordingly to ensure that users obtain the latest evaluation results; S6: Obtaining user prompt content according to the comprehensive credibility score and the comprehensive prompt content; Specifically, in step S6, the credibility and popular science information are displayed through floating cards, and the user can click to view the detailed content, that is, the user prompt content; that is, the original content and related popular science information are displayed to the user, and their credibility scores are marked; It should be noted that in step S6, the credibility score and authoritative popular science tips are placed on the edge of the short video playback interface in the form of a semi-transparent floating card, and detailed information is displayed after clicking to activate the card; In some embodiments, the above-mentioned short video multi-dimensional analysis and credibility assessment method can also be implemented in the following way.

[0019] In this embodiment, the short video multi-dimensional analysis and credibility assessment method includes the following steps: Step S01: Use natural language processing technology to extract and analyze video text data: first use speech-to-text technology to convert the audio in the video into text, use OCR technology to identify text information from pictures and extract metadata such as subtitles, titles, and tags from videos with subtitles, and then use the BERT pre-trained language model combined with contextual information to extract accurate keywords and classify the emotions expressed in the video; extract scenes, behaviors, and events in the video: automatically detect the main objects and scenes in the video through convolutional neural networks, and use the TimeSformer model to capture complex behaviors and events; multi-modal information fusion outputs topic information: first perform preliminary fusion through weighted summation: Initial weights are assigned to keywords, emotions, scenes, behaviors, and events extracted from the short video text, audio, and visual analysis: 0.2 for scenes, 0.3 for behaviors, 0.2 for events, 0.2 for keywords, and 0.1 for emotions. The initially fused features are then input into the multimodal Transformer model to capture the complex associations between the elements through the self-attention mechanism. The multi-layer Transformer model then outputs a deeply fused feature vector to improve the accuracy of topic recognition. The feature vector is then input into the Softmax classifier to output the short video topic information. The metadata of the video (author information, time, location, device information, etc.) is analyzed to provide a reference for tracing the source of the video. Step S02: Video source tracing: Compare the short video subject information extracted in step S1 with the data subject information features of other videos in the database, which stores the metadata of videos uploaded to the online short video platform and the exact subject information obtained through multimodal information fusion technology. Then, cluster analysis is used to identify version groups with similar themes to locate other versions related to the original version. Then, by comparing the differences between the versions of the video, the modification traces of the video at different propagation nodes can be determined, and the time and location information of the video can be extracted and compared in combination with the metadata of each version of the video stored in the database, and non-original versions can be gradually excluded. Finally, the initial version of the video can be successfully traced back. Use AI prediction model to analyze the propagation path: By predicting and tracking the propagation path of the current video through AI model, a propagation diagram of information in social networks can be constructed to identify each propagation node. Combine AI model to detect abnormal propagation nodes: Through the classification model, learn the characteristics of normal user behavior and detect nodes that are significantly different from it; that is, analyze the behavior pattern of each node and identify nodes that are significantly different from normal user behavior; for example, detect nodes that frequently forward a large amount of information in a short period of time, which may be robots or automated scripts; Step S03: Based on the calculation of multiple key factors, a comprehensive evaluation is performed, and the following are the main evaluation factors: Video source credibility: Evaluate the credibility of the publisher of the initial version of the video, using an automated scoring mechanism. For example, if the publisher is an authoritative organization, a certified user or an expert, has a good historical record, and the content is widely recognized, it is scored as 10 points; the identity is basically credible, and occasionally there is low-quality content, it is scored as 7-9 points; the identity is uncertain, and some content is unreliable, it is scored as 4-6 points; the identity contains a lot of false information, it is scored as 0-3 points; Video content consistency: Evaluate the consistency of the current video content with the initial version of the video, using an automated scoring mechanism, the theme is completely consistent, and the keywords and theme tags are highly overlapped, which can be scored as 10 points; the theme is basically consistent, and some keywords and tags are similar, which can be scored as 7-9 points; there are certain differences in the theme, and the overlap of keywords and tags is low, which can be scored as 4-6 points; the theme is completely inconsistent, and the keywords and tags are irrelevant, which can be scored as 0-3 points; Propagation path: Evaluate the credibility of the propagation nodes of the video, using an automated scoring mechanism, all nodes are credible users, and the behavior pattern is normal and stable, which can be scored as 10 points; most nodes can be The credibility of the video is high, but there may be a few abnormal behaviors, with a score of 7-9 points; some nodes show abnormal behaviors, and the credibility is reduced, with a score of 4-6 points; most nodes show obvious abnormalities, which may be robots or malicious accounts, with a score of 0-3 points; timestamp verification: compare the timestamp of the video with the time information of the relevant event to determine whether the timestamp reflects the actual time of the event, and use an automated scoring mechanism; the timestamp is completely accurate and highly consistent with the event timeline, which can be scored as 10 points; the timestamp is basically accurate, but there are a few small time differences, with a score of 7-9 points; the obvious difference in timestamps may cause difficulties in understanding the content, with a score of 4-6 points; the timestamp is seriously wrong or missing, inconsistent with the event timeline, resulting in highly misleading content, with a score of 0-3 points; then a weighted comprehensive evaluation of the above four factors is performed based on the scoring model, and each factor is scored separately (0 to 10 points), and the credibility score is generated after calculation. The specific calculation formula is: credibility score = 0.4×video source credibility score + 0.4×video content consistency score + 0.1×propagation path score + 0.1×timestamp verification score; Step S04: Find relevant popular science information based on the video keywords, themes and semantic analysis information extracted by natural language processing technology; use GPT-4 to simplify and integrate popular science information from multiple sources to generate a comprehensive prompt content; Step S05: Combined with the real-time data stream of social media, the machine learning algorithm is used to classify and cluster user behaviors, identify abnormal behavior patterns (such as abnormally high forwarding volume, sudden emergence of comments), continuously analyze the newly emerging short video interactions and dissemination paths, and adjust the credibility score accordingly to ensure that users obtain the latest evaluation results; Step S06: The credibility score and authoritative popular science tips are placed on the edge of the short video playback interface in the form of a semi-transparent floating card, and detailed information is displayed after clicking to activate the card.

[0020] In some embodiments, the above short video multi-dimensional analysis and credibility assessment system can also be implemented in the following way. The system process is as follows: Figure 2 As shown; in this embodiment, the system includes the following modules: Short video content analysis module: Use natural language processing technology to extract and analyze video text data: First, use speech-to-text technology to convert the audio in the video into text, use OCR technology to identify text information from pictures and extract metadata such as subtitles, titles, and tags from videos with subtitles. Then use the BERT pre-trained language model combined with contextual information to extract accurate keywords and classify the emotions expressed in the video; extract scenes, behaviors, and events in the video: automatically detect the main objects and scenes in the video through convolutional neural networks, and use the TimeSformer model to capture complex behaviors and events; multi-modal information fusion outputs topic information: first perform preliminary fusion through weighted summation : Assign initial weights to keywords, emotions, scenes, behaviors, and events extracted from the short video text, audio, and visual analysis: 0.2 for scenes, 0.3 for behaviors, 0.2 for events, 0.2 for keywords, and 0.1 for emotions. Then input the initially fused features into the multimodal Transformer model, and capture the complex associations between the elements through the self-attention mechanism. Then, the multi-layer Transformer model outputs the deeply fused feature vector to improve the accuracy of topic recognition, that is, input the feature vector into the Softmax classifier to output the short video topic information; analyze the metadata of the video (author information, time, location, device information, etc.) to provide a reference for tracing the source of the video; Source tracing module: Video source tracing: Compare the short video subject information extracted in step S1 with the data subject information features of other videos in the database, which stores the metadata of videos uploaded to the online short video platform and the exact subject information obtained through multimodal information fusion technology. Then, cluster analysis is used to identify version groups with similar themes to locate other versions related to the original version. Then, by comparing the differences between the versions of the video, the modification traces of the video at different propagation nodes can be determined, and the time and location information of the video can be extracted and compared in combination with the metadata of each version of the video stored in the database, and non-original versions can be gradually excluded. Finally, the initial version of the video can be successfully traced back. Use AI prediction model to analyze the propagation path: By predicting and tracking the propagation path of the current video through AI model, a propagation diagram of information in social networks can be constructed to identify each propagation node. Combine AI model to detect abnormal propagation nodes: Through the classification model, learn the characteristics of normal user behavior and detect nodes that are significantly different from it; that is, analyze the behavior pattern of each node and identify nodes that are significantly different from normal user behavior; for example, detect nodes that frequently forward a large amount of information in a short period of time, which may be robots or automated scripts; Credibility scoring module: Based on the calculation of multiple key factors, a comprehensive evaluation is conducted. The following are the main evaluation factors: Video source credibility: Evaluate the credibility of the publisher of the initial version of the video, using an automated scoring mechanism. For example, if the publisher is an authoritative organization, a certified user or an expert, has a good historical record, and the content is widely recognized, it is scored as 10 points; the identity is basically credible, and occasionally there is low-quality content, it is scored as 7-9 points; the identity is uncertain, and some content is unreliable, it is scored as 4-6 points; the identity contains a lot of false information, it is scored as 0-3 points; Video content consistency: Evaluate the consistency of the current video content with the initial version of the video, using an automated scoring mechanism, the theme is completely consistent, and the keywords and theme tags are highly overlapped, which can be scored as 10 points; the theme is basically consistent, and some keywords and tags are similar, which can be scored as 7-9 points; there are certain differences in the theme, and the overlap of keywords and tags is low, which can be scored as 4-6 points; the theme is completely inconsistent, and the keywords and tags are irrelevant, which can be scored as 0-3 points; Propagation path: Evaluate the credibility of the video's propagation nodes, using an automated scoring mechanism, all nodes are trusted users, and the behavior pattern is normal and stable, which can be scored as 10 points; most nodes Credible, but there may be a few abnormal behaviors, with a score of 7-9 points; some nodes show abnormal behaviors, and the credibility is reduced, with a score of 4-6 points; most nodes show obvious abnormalities, which may be robots or malicious accounts, with a score of 0-3 points; Timestamp verification: compare the timestamp of the video with the time information of the relevant event to determine whether the timestamp reflects the actual time of the event, and use an automated scoring mechanism; The timestamp is completely accurate and highly consistent with the event timeline, which can be scored as 10 points; The timestamp is basically accurate, but there are a few small time differences, which are scored as 7-9 points; There are obvious differences in timestamps, which may cause difficulties in understanding the content, with a score of 4-6 points; The timestamp is seriously wrong or missing, inconsistent with the event timeline, resulting in highly misleading content, with a score of 0-3 points; Then, based on the scoring model, a weighted comprehensive evaluation is performed on the above four factors, and each factor is scored separately (0 to 10 points). After calculation, a credibility score is generated. The specific calculation formula is: Credibility score = 0.4×video source credibility score + 0.4×video content consistency score + 0.1×propagation path score + 0.1×timestamp verification score; Authoritative science module: Based on the video keywords, themes and semantic analysis information extracted by natural language processing technology, relevant science information is found; GPT-4 is used to simplify and integrate science information from multiple sources to generate a comprehensive prompt content; Real-time monitoring and updating module: Combined with the real-time data stream of social media, it uses machine learning algorithms to classify and cluster user behaviors, identify abnormal behavior patterns (such as abnormally high forwarding volume, sudden emergence of comments), continuously analyze the newly emerging short video interactions and dissemination paths, and adjust the credibility score accordingly to ensure that users obtain the latest evaluation results; User prompt module: The credibility score and authoritative popular science tips are placed on the edge of the short video playback interface in the form of a semi-transparent floating card. Click to activate the card to display detailed information.

[0021] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation modes, which are merely illustrative rather than restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, all of which are within the protection of the present invention.

Claims

1. A short video multi-dimensional analysis and credibility assessment method, characterized in that: The following steps are involved: S1: Obtain a short video to be analyzed, analyze the short video to be analyzed, and obtain keywords, short video metadata, and short video theme information; S2: According to the short video theme information and the video database, a similar theme version group, an initial version of the short video to be analyzed, a propagation path and a propagation node are obtained; S3: Obtain a comprehensive credibility score based on the propagation node, the short video subject information, the keywords and the short video metadata.

2. The short video multi-dimensional analysis and credibility assessment method according to claim 1 is characterized in that: Step S1 specifically includes: S11: Obtain a short video to be analyzed, obtain short video metadata, voice data and picture data according to the short video to be analyzed, and obtain text data according to the voice data and the picture data; S12: Obtain keywords and emotions using a pre-trained language model according to the text data and the short video metadata; S13: According to the short video to be analyzed, using a convolutional neural network and a video classification model, obtaining objects, scenes, behaviors and events; S14: According to the keywords, the emotions, the scenes, the behaviors and the events, the short video topic information is obtained by using a multimodal Transformer model and a Softmax classifier.

3. The short video multi-dimensional analysis and credibility assessment method according to claim 1, characterized in that: Step S2 specifically includes: S21: Obtaining a similar theme version group using a clustering method according to the short video theme information and the video database; S22: Obtaining an initial version of the short video to be analyzed according to the similar theme version group and the video database; S23: According to the initial version of the short video to be analyzed and the video database, the propagation path and propagation nodes are obtained using the AI ​​model.

4. The short video multi-dimensional analysis and credibility assessment method according to claim 1, characterized in that: Step S3 specifically includes: S31: Obtaining a video source credibility score according to the propagation node; S32: Obtaining a video content consistency score according to the short video theme information and keywords; S33: Obtain a propagation path score according to the propagation node; S34: Obtaining a timestamp verification score according to the short video metadata; S35: Obtain a comprehensive credibility score according to the video source credibility score, the video content consistency score, the propagation path score and the timestamp verification score.

5. The short video multi-dimensional analysis and credibility assessment method according to claim 1, characterized in that: The short video multi-dimensional analysis and credibility assessment method also includes: S4: Obtaining comprehensive prompt content according to the short video metadata, short video theme information and keywords; S5: Acquire real-time data streams, and update comprehensive credibility scores according to the real-time data streams and propagation nodes; S6: Obtain user prompt content according to the comprehensive credibility score and the comprehensive prompt content.

6. A short video multi-dimensional analysis and credibility assessment system, characterized in that: The short video multi-dimensional analysis and credibility assessment system includes the following modules: The short video content analysis module is configured to: obtain a short video to be analyzed, analyze the short video to be analyzed, and obtain keywords, short video metadata, and short video theme information; The source tracing module is configured to: obtain a similar theme version group, an initial version of the short video to be analyzed, a propagation path, and a propagation node according to the short video theme information and the video database; The credibility scoring module is configured to obtain a comprehensive credibility score based on the propagation node, the short video theme information, the keywords and the short video metadata.

7. The short video multi-dimensional analysis and credibility assessment system according to claim 6, characterized in that: The short video content analysis module is specifically configured as follows: Acquire a short video to be analyzed, obtain short video metadata, voice data and picture data according to the short video to be analyzed, and obtain text data according to the voice data and the picture data; According to the text data and the short video metadata, using a pre-trained language model, obtaining keywords and emotions; According to the short video to be analyzed, using a convolutional neural network and a video classification model, objects, scenes, behaviors and events are obtained; According to the keywords, the emotions, the scenes, the behaviors and the events, the short video topic information is obtained by using a multimodal Transformer model and a Softmax classifier.

8. The short video multi-dimensional analysis and credibility assessment system according to claim 6, characterized in that: The source tracing module is specifically configured as follows: According to the short video theme information and the video database, a clustering method is used to obtain a similar theme version group; Obtaining an initial version of the short video to be analyzed according to the similar theme version group and the video database; According to the initial version of the short video to be analyzed and the video database, the propagation path and propagation nodes are obtained using the AI ​​model.

9. The short video multi-dimensional analysis and credibility assessment system according to claim 6, characterized in that: The credibility scoring module is specifically configured as follows: Obtaining a video source credibility score according to the propagation node; Obtaining a video content consistency score based on the short video theme information and keywords; According to the propagation node, a propagation path score is obtained; Obtaining a timestamp verification score according to the short video metadata; A comprehensive credibility score is obtained according to the video source credibility score, the video content consistency score, the propagation path score and the timestamp verification score.

10. The short video multi-dimensional analysis and credibility assessment system according to claim 6, characterized in that: The short video multi-dimensional analysis and credibility assessment system also includes: The authoritative science popularization module is configured to: obtain comprehensive prompt content according to the short video metadata, short video theme information and keywords; A real-time monitoring and updating module is configured to: obtain a real-time data stream, and update a comprehensive credibility score according to the real-time data stream and the propagation node; The user prompt module is configured to obtain user prompt content according to the comprehensive credibility score and the comprehensive prompt content.

Images