Multi-mode video public opinion intelligent monitoring method and system

Abstract

The invention relates to the technical field of monitoring, and discloses a multi-modal video public opinion intelligent monitoring method and system, and the system comprises a multi-modal data collection module, a data processing module, a coefficient generation module, an intelligent analysis module, a dynamic threshold module, and a public opinion visual platform. According to the method, a visual, voice and character cross-modal cooperative computing framework is established, and the one-sidedness problem of traditional single-modal analysis is solved; a recognition-tendency-propagation three-dimensional index system is constructed, and accurate quantitative evaluation of public opinion evolution is achieved; designing a static comprehensive coefficient threshold value and a dynamic growth threshold value of the dual-threshold grading system, and cooperating with an imaginary main body filtering algorithm and KOL polarization value calculation, so that the false alarm rate is remarkably reduced, and meanwhile, rapid early warning response is realized; a three-dimensional propagation map is generated through propagation path concentration analysis, content feature evolution tracking and influence radiation, and a complete public opinion evolution chain and hotspot traceability are provided.

Description

Technical Field

[0001] This invention relates to the field of monitoring technology, and more specifically to a multimodal video public opinion intelligent monitoring method and system. Background Technology

[0002] With the widespread adoption of 5G networks and short video platforms, video content has become a major carrier of online public opinion.

[0003] Traditional video-based public opinion monitoring primarily relies on keyword matching and basic text analysis. It involves manually setting keywords such as brand names and product names to target specific content on short video platforms, combining this with dissemination metrics like view counts and comment counts to determine public opinion intensity. Data processing uses relevant technologies to classify comments based on sentiment polarity. Early warning mechanisms typically employ fixed thresholds, triggering manual review when negative sentiment exceeds a preset value. However, this approach also has the following drawbacks:

[0004] Modal singularity defect: Existing methods mostly rely on text analysis, which is insufficient for analyzing the core visual information and speech features of videos, resulting in a generally low accuracy rate in public opinion judgment.

[0005] Limitations of static assessment: Traditional indicator systems lack dynamic response mechanisms, cannot capture key turning points in sudden public opinion events, and use fixed-weight arithmetic average algorithms, which are not sensitive enough to abnormal propagation behaviors.

[0006] Delayed response problem: Mainstream systems rely on manual threshold setting, resulting in delayed early warning response and a lack of ability to identify fictitious content, leading to a high false alarm rate.

[0007] Shortcomings of planar analysis: Existing technologies only provide a two-dimensional propagation path display, lacking dimensions such as KOL influence polarization analysis and cross-platform penetration tracking, resulting in insufficient completeness of key node identification in the source tracing report.

[0008] Therefore, multimodal fusion analysis capabilities, dynamic quantitative evaluation models, intelligent early warning and response mechanisms, and three-dimensional traceability visualization systems are needed to solve the above problems. Summary of the Invention

[0009] In order to overcome the above-mentioned defects of the prior art, the present invention provides a multimodal video public opinion intelligent monitoring method and system to solve the problems existing in the background art.

[0010] To achieve the above objectives, the present invention provides the following technical solution: a multimodal video public opinion intelligent monitoring method, characterized by comprising the following steps:

[0011] S1. Extract visual feature data, voice feature data, and text feature data from the video, and collect dynamic data of video transmission.

[0012] S2. Clean and process the collected data, remove special effects and non-semantic symbol interference, and generate public opinion identification indicators, public opinion trend indicators, and public opinion dissemination indicators.

[0013] S3. Generate a public opinion judgment coefficient based on public opinion identification indicators, a public opinion tendency coefficient based on public opinion tendency indicators, and a public opinion dissemination coefficient based on public opinion dissemination indicators.

[0014] S4. Generate a comprehensive public opinion coefficient from the public opinion judgment coefficient, the public opinion tendency coefficient, and the public opinion dissemination coefficient;

[0015] S5. Establish a dual threshold mechanism to set thresholds for the comprehensive public opinion coefficient, classify and issue warnings, and automatically jump to a higher level when the 24-hour growth rate of a certain dimension indicator exceeds 200%.

[0016] S6. Visualize the three-dimensional propagation map of public opinion trends and generate a propagation chain report to trace the source of hotspots.

[0017] A multimodal video public opinion intelligent monitoring system, characterized in that it specifically includes:

[0018] Multimodal data acquisition module: used to extract visual feature data, speech feature data and text feature data from videos, and to collect dynamic data of video propagation;

[0019] Data processing module: used to clean and process the collected data, remove special effects and non-semantic symbol interference, and generate public opinion identification indicators, public opinion trend indicators and public opinion dissemination indicators;

[0020] Coefficient generation module: used to generate public opinion judgment coefficients based on public opinion identification indicators, public opinion tendency coefficients based on public opinion tendency indicators, and public opinion dissemination coefficients based on public opinion dissemination indicators.

[0021] Intelligent analysis module: used to generate a comprehensive public opinion coefficient from the public opinion judgment coefficient, public opinion tendency coefficient, and public opinion propagation coefficient;

[0022] Dynamic threshold module: Used to establish a dual threshold mechanism, set thresholds for the comprehensive public opinion coefficient, classify and issue warnings. When the 24-hour growth rate of a certain dimension indicator exceeds 200%, it will automatically jump to a higher level.

[0023] Public opinion visualization platform: used to visualize the three-dimensional propagation map of public opinion trends and generate propagation chain reports for hotspot tracing.

[0024] This multimodal video public opinion intelligent monitoring method integrates visual, voice, and text feature data. After data cleaning, it generates three core indicators: public opinion identification, trend, and dissemination. These indicators are then converted into quantitative coefficients and finally integrated into a comprehensive public opinion coefficient to achieve multidimensional evaluation. The system adopts a dual-threshold dynamic early warning mechanism, automatically upgrading the response level for indicators that increase by more than 200% in 24 hours. At the same time, it achieves visualized tracking of public opinion trends through three-dimensional dissemination maps and source tracing reports, forming a closed-loop management process from data collection to crisis response.

[0025] The technical effects and advantages of this invention are as follows:

[0026] 1. This invention establishes a cross-modal collaborative computing framework for vision, speech, and text, and achieves complementary verification of multi-source information through feature alignment and dynamic weight allocation, thereby solving the one-sidedness problem of traditional single-modal analysis.

[0027] 2. This invention constructs a three-dimensional indicator system of "identification-tendency-propagation", and innovatively adopts a geometric average algorithm and a hybrid normalization method, combined with a 24-hour growth rate trigger mechanism, to achieve accurate quantitative assessment of public opinion evolution.

[0028] 3. This invention designs a dual-threshold hierarchical system with a static comprehensive coefficient threshold and a dynamic growth threshold, combined with a fictitious subject filtering algorithm and KOL polarization value calculation, which significantly reduces the false alarm rate while achieving rapid early warning response.

[0029] 4. This invention provides a complete public opinion evolution chain and hotspot tracing capability by generating a three-dimensional dissemination map through dissemination path concentration analysis, content feature evolution tracking, and influence radiation. Attached Figure Description

[0030] Figure 1 This is a structural block diagram of the present invention.

[0031] Figure 2 This is a flowchart of the present invention. Detailed Implementation

[0032] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. In addition, the forms of the various structures described in the following embodiments are merely illustrative. The automatic unloading device for rotary kiln with self-cooling function involved in the present invention is not limited to the structures described in the following embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] Reference Figure 1This invention provides a multimodal video public opinion intelligent monitoring method and system, including a multimodal data acquisition module, a data processing module, a coefficient generation module, an intelligent analysis module, a dynamic threshold module, and a public opinion visualization platform.

[0034] Reference Figure 2 The specific implementation steps of the present invention include the following steps:

[0035] S1. Extract visual feature data, voice feature data, and text feature data from the video, and collect dynamic data of video transmission.

[0036] It should be specifically explained that visual feature extraction involves analyzing video footage using computer vision technology, including facial recognition to detect facial expressions such as tears and upturned corners of the mouth, action recognition such as frowning and sobbing, and scene analysis; it also includes the duration of a single person's emotion and the consistency of emotions in multiple people's footage.

[0037] The speech feature processing specifically involves a dual approach: acoustic analysis and semantic parsing. At the acoustic level, the fundamental frequency variance is extracted to determine the sharpness of the intonation and the speech rate to determine the degree of urgency. At the semantic level, specific sentence structures such as rhetorical questions, exclamations, and affirmative sentences are identified.

[0038] The text feature extraction specifically involves: combining existing technologies to identify video text, subtitles, and bullet comments, and using models to analyze the density of sensitive words and the statistics of extreme, negative, and positive words in terms of sentiment; and continuously learning similar semantic words, internet slang, and meme culture.

[0039] The dynamic data collection of dissemination specifically involves: capturing video dissemination data across all channels in real time, including cross-platform reposts, comments, and exposure, and summarizing the influence of key opinion leaders.

[0040] S2. Clean and process the collected data, remove special effects and non-semantic symbol interference, and generate public opinion identification indicators, public opinion trend indicators, and public opinion dissemination indicators.

[0041] It should be noted that the indicators for public opinion identification include the public dissemination coefficient, the public interaction coefficient, and the subject orientation coefficient.

[0042] The public propagation coefficient is calculated as follows: base platform weight + cross-platform gain.

[0043] The logic for determining the basic platform coefficient is as follows: the content is disseminated on multiple platforms, including news websites, social media, forums, and encyclopedias, and is divided into 5 categories from largest to smallest based on its volume. The public dissemination coefficient is assigned a value from 0.9 to 0.5 based on its importance.

[0044] If the message is propagated across multiple platforms, the coefficient increases by 0.1 for each additional platform, up to a maximum of 0.3. It should be noted that if the base platform weight plus the cross-platform gain is greater than 1, the public propagation coefficient is set to 1.

[0045] The public interaction coefficient is the geometric mean of the comment density coefficient, the dissemination coefficient, and the participation user coefficient, which is simply the sum of these factors divided by 3.

[0046] Comment density = 100 * number of valid comments / content exposure. The specific rules for assigning the comment density coefficient are as follows: 0 points for comment density ≤ first threshold, 0 points for comment density between the first threshold and the second threshold, and 1 point for comment density exceeding the second threshold. The first threshold ranges from 0.3% to 0.4%, and the second threshold ranges from 0.8% to 0.9%.

[0047] The diffusion growth rate = the increase in forwarding within 24 hours / the initial forwarding base. The specific rules for assigning the propagation coefficient are as follows: if the growth rate is less than or equal to the first threshold, it is 0 points; if the growth rate is between the first threshold and the second threshold, it is scored linearly; if it exceeds the second threshold, it is scored directly as 1 point. The first threshold ranges from 10% to 12%, and the second threshold ranges from 30% to 35%.

[0048] The percentage of unique accounts = number of non-duplicate participating users / total number of interactive accounts. The specific rules for assigning the participation user coefficient are as follows: if the percentage of unique accounts is less than or equal to the first threshold, 0 points are awarded; if the percentage of unique accounts is between the first and second thresholds, points are awarded proportionally; if the percentage exceeds the second threshold, 1 point is awarded directly. The first threshold ranges from 40% to 50%, and the second threshold ranges from 70% to 80%.

[0049] The specific rules for assigning the subject-specific coefficient are as follows: Event-related: such as "tourist overcharging incidents at scenic spots" and "policy promulgation"; Brand / Institution-related: such as "enterprises" and "government departments"; Person-related: such as controversial behaviors of public figures and typical events of social figures; 0.8 points for meeting one of the criteria, 0.9 points for meeting two criteria, and 1 point for meeting all three criteria.

[0050] It is important to note that the existence of the subject needs to be verified by linking the knowledge graph; if the subject is fictitious, the coefficient is set to 0. A weak association compensation mechanism is added, which calculates semantic similarity through the model. If the similarity is greater than or equal to the threshold, the similarity is assigned proportionally between the threshold and 100%, with the threshold value ranging from 70% to 80%. If it is a vague reference, it is assigned a score of 40% to 50%.

[0051] It should be noted that the indicators related to public opinion trends include visual coefficients, voice coefficients, and text coefficients.

[0052] The visual coefficient is calculated based on visual feature data, identifies extreme emotions as core evidence of a crisis, and amplifies the visual effect according to special circumstances.

[0053] If the visual feature data contains images of facial tears, red eyes, or sobbing, the visual coefficient is -0.9 to -0.8; if the visual feature data contains expressions such as frowning, dilated pupils, or tense facial muscles, the visual coefficient is -0.8 to -0.7; if the visual feature data contains expressions such as upturned corners of the mouth and wrinkled eyes, the visual coefficient is +0.6 to +0.7.

[0054] It should be noted that if the same visual feature data lasts for more than 5 seconds, the visual coefficient is amplified by 12 to 1.3; if there are multiple people in the scene and the visual feature data is consistent, the visual coefficient is amplified by 1.4 to 1.5 to amplify the group's emotions, with a maximum of 1 and a minimum of -1.

[0055] The speech coefficients are calculated based on speech feature data, and the acoustic parameters capture stress emotions to compensate for the concealment of text expression. Semantic analysis identifies questioning sentences and strong emotions to reveal potential conflicts.

[0056] If the fundamental frequency variance in the acoustic feature data of the speech feature data is greater than the threshold, it indicates a sharp tone. The threshold value ranges from 50Hz to 60Hz, and the speech coefficient is -0.5 to -0.4. If the speaking speed is greater than the threshold, and the threshold value ranges from 5 words / second to 6 words / second, it indicates a rapid speaking speed, and the speech coefficient is -0.4 to -0.3. If the semantic feature data of the speech feature data contains rhetorical questions, such as "Is it reasonable?", the speech coefficient is -0.6 to -0.5; if it contains exclamatory sentences, such as "That's awful!", the speech coefficient is -0.7 to -0.6; if it contains affirmative sentences, such as "Very satisfied.", the speech coefficient is +0.5 to +0.6. The sum of the two values ​​is a maximum of 1 and a minimum of -1.

[0057] The text coefficient is calculated based on text feature data, and public opinion is quickly identified based on the part of speech and density of sensitive words.

[0058] If extremely negative words appear, such as "garbage," "complaint," "fake," or "deceptive," the word coefficient is -1 to -0.9. If generally negative words appear, such as "disappointed" or "not good," the word coefficient is -0.6 to -0.5. If positive words appear, such as "like" or "satisfied," the word coefficient is +0.5 to +0.6.

[0059] It should be noted that if a sensitive word appears more than 3 times within 10 seconds, the coefficient will be amplified by 1.6-1.8, with a maximum of 1 and a minimum of -1.

[0060] It should be noted that the indicators related to public opinion dissemination include diffusion intensity coefficient, key node coefficient, and cross-platform penetration coefficient.

[0061] Where the diffusion intensity coefficient = total information base value * interaction stability coefficient * dynamic gain coefficient.

[0062] The base value of total information is calculated as ln(total information within the statistical period + 1) * 10, where the total information includes the number of relevant content items from all channels, including text, video, and bullet comments. The base value is normalized as follows: normalized base value = (base value - historical minimum base value) / (historical maximum base value - historical minimum base value). The maximum and minimum base values ​​are based on the most recent 30 days, and the base value is mapped to the interval [0,1].

[0063] The interaction stability coefficient is calculated using the variance of likes / shares / comments. The variance is set with a threshold based on historical experience. When the variance is less than or equal to the first threshold, the interaction stability coefficient is 1. When the variance exceeds the first threshold but does not exceed the second threshold, the interaction stability coefficient decreases proportionally. When the variance exceeds the second threshold, the coefficient is 0. The first threshold ranges from 20% to 30%, and the second threshold ranges from 40% to 50%.

[0064] The dynamic gain coefficient is based on 0.7-0.8. If the 24-hour information growth rate is greater than or equal to the threshold, and the threshold range is 50%-60%, the dynamic gain coefficient is amplified by 1.2-1.3. If the interaction volume increase is greater than the threshold per hour for 2 consecutive hours, and the threshold range is 15%-20%, the dynamic gain coefficient is multiplied by 1.1-1.2. If the drop after the peak is greater than the threshold for 4 hours, and the threshold range is 40%-50%, the dynamic gain coefficient is reduced by 0.5-0.6, with a maximum of 1.

[0065] Key node coefficient = KOL stance polarization value * propagation path concentration.

[0066] The KOL polarization value is calculated as: |Total influence of supporting KOLs - Total influence of opposing KOLs| / Total influence of the top ten KOLs. KOLs are key opinion leaders. The absolute value of the difference between the influence of supporting and opposing KOLs reflects the intensity of the positional conflict. The total influence of the top ten KOLs is used to standardize the value. Influence is calculated based on the number of followers / interactions. When the polarization value is greater than the threshold, the threshold range is 0.3-0.4, and the coefficient is amplified to 1.5-1.6.

[0067] The concentration of the propagation path = (propagation volume of the initial account + propagation volume of secondary nodes) / total propagation volume. When the ratio is greater than the first threshold, which is 60%-70%, the coefficient is magnified by 1.3-1.4 times, reflecting the dominance of key nodes. If the ratio is less than the second threshold, which is 30%-40%, the coefficient is reduced by 0.6-0.7 times.

[0068] The platform penetration coefficient is calculated based on the number of platform types: 0.8 for 1-2 platforms, 1.0 for 3-4 platforms, and 1.2 for ≥5 platforms.

[0069] S3. Generate a public opinion judgment coefficient based on public opinion identification indicators, a public opinion tendency coefficient based on public opinion tendency indicators, and a public opinion dissemination coefficient based on public opinion dissemination indicators.

[0070] It should be noted that the public opinion identification coefficient is as follows:

[0071] B = G * H * Z;

[0072] Where B is the public opinion identification coefficient, which is associated with three necessary conditions: the information must be disseminated through public channels, it must trigger group discussions and have a clear associated object. The multiplicative structure can ensure that if any dimension is missing, the comprehensive coefficient will be zero, highlighting the shortcoming effect, and accurately identifying public opinion through data-driven approaches.

[0073] G is the public dissemination coefficient, which monitors the spread of public opinion across multiple platforms, identifies the scope of dissemination, and quantifies the breadth and intensity of public opinion's spread.

[0074] H represents the public interaction coefficient, which reflects the degree of public participation through effective comments and reposts from non-online promoters.

[0075] Z is the subject orientation coefficient, which identifies the specific subject mentioned in public opinion to pinpoint the correlation and focus of conflict. It should be noted that the public opinion tendency coefficient is specifically:

[0076] Q = ω s *S+ω y *Y+ω w *W;

[0077] Q is the public opinion tendency coefficient, which objectively measures the positive, neutral or negative tendency of public opinion by quantitatively analyzing multimodal data, and judges the emotional tendency of public opinion. Subjective emotions are converted into objective values ​​in the range of [-1,1]. A negative value indicates negative public opinion, and a positive value indicates positive public opinion. The larger the absolute value, the stronger the emotional tendency.

[0078] S is the visual coefficient, which captures non-verbal emotional signals such as facial expressions and body movements, reflecting the true emotional intensity of the individual or group; Y is the speech coefficient, which combines acoustic features such as tone of voice, speech rate, and semantic features such as sentence structure to analyze the implied emotions in language; W is the textual coefficient, which quantifies the emotions expressed in text and reflects the immediate attitude.

[0079] It should be explained that the coefficient weight allocation needs to be automatically adjusted according to different scenarios to avoid misjudgment caused by modal interference. For visually dominant videos such as shooting footage, the visual weight is the highest, with the visual weight ωs ranging from 0.6 to 0.7, the voice weight ωy ranging from 0.2 to 0.3, and the text weight ωw ranging from 0.1 to 0.2.

[0080] For audio recordings or spoken videos, the voice is the most important element. The visual weight ωs ranges from 0.1 to 0.2, the voice weight ωy ranges from 0.8 to 0.9, and the text weight ωw ranges from 0.1 to 0.2.

[0081] For live interactive videos or long text-based videos, the text has the highest weight, the visual weight ωs ranges from 0.2 to 0.3, the voice weight ωy ranges from 0.1 to 0.2, and the text weight ωw ranges from 0.6 to 0.7.

[0082] For multimodal videos, the visual weight ωs ranges from 0.3 to 0.4, the speech weight ωy ranges from 0.3 to 0.4, and the text weight ωw ranges from 0.2 to 0.3.

[0083] It should be noted that the public opinion dissemination coefficient is as follows:

[0084] C = K * J * P * η;

[0085] Where C is the public opinion propagation coefficient, which assesses the intensity of public opinion propagation by quantifying the relevant dynamics of public opinion spread.

[0086] K is the diffusion intensity coefficient, which quantifies the scale and growth rate of information by logarithmic normalization of the total amount of information, fluctuations in interaction, and growth.

[0087] J is the key node coefficient, which assesses the opinion leaders' control over the path by measuring the difference in influence between the support and opposition camps and the concentration and depth of path propagation.

[0088] P is the platform penetration coefficient, which measures the overall coverage by the number of platforms.

[0089] η is the credibility coefficient, which is determined by the initial source of information. If the initial source is an authoritative media, the coefficient is 1.2-1.3; if the initial source is an anonymous account, the coefficient decreases by 0.9-0.95 times with each level of dissemination factor.

[0090] S4. Generate a comprehensive public opinion index by combining the public opinion judgment coefficient, the public opinion tendency coefficient, and the public opinion dissemination coefficient;

[0091] It should be noted that the comprehensive public opinion index is as follows:

[0092] A=α*B′+β*|Q′|*X+γ*C′;

[0093] A represents the comprehensive public opinion index, which assesses the development of public opinion by confirming public opinion, identifying its sentiment trends and diffusion intensity.

[0094] B is the public opinion identification coefficient, which quantifies the authenticity of public opinion monitored by the monitoring system. The higher the value, the stronger the authenticity and importance of the event. B′ is the standardized public opinion identification coefficient, and the value range after standardization is [0,10].

[0095] Q is the public opinion tendency coefficient. The absolute value is taken to remove the influence of negative numbers and to measure the intensity of emotions. The larger the value, the stronger the emotion. Q′ is the standardized public opinion tendency coefficient. The value range after standardization is [0,10].

[0096] X is the emotional reinforcement coefficient. When Q < 0, it is a negative public opinion, so Q is amplified: X = 1.5 + tanh(3*|Q|) to enhance the crisis's accelerated spread effect. When Q > 0, it is a positive public opinion, because it proves that public opinion does not need intervention, so Q is attenuated: X = 1 - 0.5 * [1 - tanh(3*|Q|)] to reduce the impact on the overall public opinion coefficient.

[0097] C is the public opinion propagation coefficient, which quantifies the intensity of information diffusion, the influence of key nodes, and the ability to penetrate across platforms. C′ is the standardized public opinion propagation coefficient, with a value range of [0, 80] after standardization.

[0098] α represents the weight of the public opinion identification coefficient, reflecting the importance of the event authenticity judgment in the comprehensive coefficient; β represents the weight of the public opinion tendency coefficient and the emotional reinforcement coefficient, reflecting the decision value of the intensity of public opinion emotions; γ represents the weight of the public opinion dissemination coefficient, measuring the proportion of information diffusion in the comprehensive evaluation.

[0099] It should be noted that the weight allocation is different for different periods. In the early stage of public opinion development, that is, when B≥b, the value of b is between 0.6 and 0.65, the identification coefficient has the highest weight, the value of α is between 0.5 and 0.6, and the values ​​of the tendency weight β and the propagation weight γ are between 0.2 and 0.25.

[0100] During the period of public opinion outbreak, i.e. when C≥c, the value of c ranges from 0.6 to 0.65. The public opinion propagation coefficient has the highest weight, ranging from 0.5 to 0.6. The identification weight is between 0.15 and 0.2, and the tendency weight is between 0.25 and 0.3.

[0101] During the conversion period, the public opinion tendency coefficient has the highest weight, ranging from 0.5 to 0.6, the identification weight is between 0.25 and 0.3, and the dissemination weight is between 0.15 and 0.2.

[0102] S5. Establish a dual threshold mechanism to set thresholds for the comprehensive public opinion coefficient, classify and issue warnings, and automatically jump to a higher level when the 24-hour growth rate of a certain dimension indicator exceeds 200%.

[0103] It should be specifically explained that the threshold setting for the comprehensive public opinion coefficient, the classification and early warning system are as follows:

[0104] When 0≤A≤x, it is a low-level alert, and a blue warning is issued; daily monitoring and weekly reporting are conducted.

[0105] When x < A ≤ y, it is classified as intermediate level, and a yellow alert is issued; a temporary emergency response team is established, a 12-hour emergency duty system is activated, and a response plan is drafted.

[0106] When y < A ≤ z, it is considered a high-level alert, and an orange alert is issued; a crisis team is established, and the first response is released within 4 hours.

[0107] When z < A ≤ 100, it is classified as a critical level and a red alert is issued; the highest level of emergency response is activated; senior management takes the lead in handling the situation, and legal / regulatory agencies intervene in coordination.

[0108] The value of x ranges from 30 to 35, the value of y ranges from 50 to 55, and the value of z ranges from 70 to 75.

[0109] The specific measures to handle abnormal growth are as follows:

[0110] Establish a real-time data comparison model. When the 24-hour growth rate of any dimension indicator is ≥200%, the system will automatically trigger the early warning level upgrade mechanism, immediately push the early warning upgrade notification to all responsible personnel terminals, and simultaneously start the corresponding higher-level response measures. The manual intervention interface is reserved for special circumstances.

[0111] S6. Visualize the three-dimensional propagation map of public opinion trends and generate a propagation chain report to trace the source of hotspots.

[0112] The generation of the 3D map involves constructing a propagation network model based on a database, which includes node dimensions, time dimensions, and spatial dimensions.

[0113] The node dimension needs to distinguish between the initial account, key opinion leaders and ordinary users. The node size is dynamically rendered according to influence, which is calculated based on the number of followers, likes, comments and reposts.

[0114] The time dimension displays the path of public opinion dissemination on a timeline and marks the outbreak points where the growth rate exceeds 200% within 24 hours.

[0115] The spatial dimension maps the cross-platform penetration path, and the penetration intensity is intuitively displayed by using different colors and shades of color according to the platform type.

[0116] The propagation chain analysis specifically involves: using a model to trace the propagation path, identifying the initial propagation source (i.e., the first account) and nodes where the secondary spread and forwarding volume exceeds the threshold, and automatically generating a chain report containing the influence polarization value of key nodes (|influence of supporting KOLs - influence of opposing KOLs| / total influence).

[0117] The hotspot tracing mechanism works by focusing on the time period when high-frequency sensitive words first appear, and combining visual / voice feature mutation points to locate the source video of the public opinion outbreak.

[0118] Through the above description of the embodiments, those skilled in the art can clearly understand that the various embodiments of this application can be implemented by means of software or software combined with necessary general-purpose hardware platforms, and of course, they can also be implemented by hardware functions; based on this understanding, the technical solution of this application, in essence or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions to cause a computer device, such as including but not limited to a personal computer, server, or network device, to execute all or part of the steps of the method described in any embodiment of this application.

[0119] The foregoing describes exemplary embodiments of this application. It should be understood that the above exemplary embodiments are not restrictive but illustrative, and the scope of protection of this application is not limited thereto. It should be understood that those skilled in the art can make modifications and variations to the embodiments of this application without departing from the spirit and scope of this application, and such modifications and variations should be within the scope of protection of this application.

Claims

1. A multimodal video public opinion intelligent monitoring method, characterized in that, Specifically, the following steps are included: S1. Extract visual feature data, voice feature data, and text feature data from the video, and collect dynamic data of video transmission. S2. Clean and process the collected data, remove special effects and non-semantic symbol interference, and generate public opinion identification indicators, public opinion trend indicators, and public opinion dissemination indicators. S3. Generate a public opinion judgment coefficient based on public opinion identification indicators, a public opinion tendency coefficient based on public opinion tendency indicators, and a public opinion dissemination coefficient based on public opinion dissemination indicators. S4. Generate a comprehensive public opinion coefficient from the public opinion judgment coefficient, the public opinion tendency coefficient, and the public opinion dissemination coefficient; S5. Establish a dual threshold mechanism to set thresholds for the comprehensive public opinion coefficient, classify and issue warnings, and automatically jump to a higher level when the 24-hour growth rate of a certain dimension indicator exceeds 200%. S6. Visualize the three-dimensional propagation map of public opinion trends and generate a propagation chain report to trace the source of hotspots.

2. The multimodal video public opinion intelligent monitoring method according to claim 1, characterized in that: The indicators related to public opinion identification include the public dissemination coefficient, the public interaction coefficient, and the subject orientation coefficient; the indicators related to public opinion tendency include the visual coefficient, the voice coefficient, and the text coefficient; and the indicators related to public opinion dissemination include the diffusion intensity coefficient, the key node coefficient, and the cross-platform penetration coefficient.

3. The multimodal video public opinion intelligent monitoring method according to claim 2, characterized in that: The calculation of the public propagation coefficient is as follows: base platform weight + cross-platform gain; the base platform coefficient is assigned a value based on the size of the platform. If it propagates on multiple platforms, the coefficient increases for each additional platform. If the base platform weight + cross-platform gain is greater than 1, the public propagation coefficient is 1. The public interaction coefficient is the geometric mean of the comment density coefficient, the propagation coefficient, and the participating user coefficient; comment density = 100 * number of valid comments / content exposure, and the comment density coefficient is assigned a value based on the comment density threshold; diffusion growth rate = 24-hour forwarding increment / initial forwarding base, and the propagation coefficient is assigned a value based on the diffusion growth rate threshold; unique account ratio = number of non-duplicate participating users / total number of interactive accounts, and the participating user coefficient is assigned a value based on the unique account ratio threshold. The specific rules for assigning the subject-oriented coefficient are as follows: for event-type, brand / organization-type, and person-type, a value is assigned based on how many criteria are met; if all criteria are met, the coefficient is 1; and if a fictitious subject is identified, the coefficient is set to 0.

4. The multimodal video public opinion intelligent monitoring method according to claim 2, characterized in that: The visual coefficient is calculated based on visual feature data. Different coefficient values ​​are assigned to visual feature data that include facial tears, red eyes, sobbing, frowning, dilated pupils, tense facial muscles, upturned corners of the mouth, and wrinkled eyes. If the duration of the same visual feature data exceeds a threshold or if the visual feature data of multiple people are consistent, the visual coefficient needs to be amplified. The speech coefficients are calculated based on speech feature data. If the fundamental frequency variance in the acoustic feature data exceeds a threshold, it indicates a sharp tone; if the speaking speed exceeds a threshold, it indicates a rapid speaking speed. Different values ​​are assigned to the speech coefficients accordingly. If the semantic feature data contains rhetorical questions, exclamations, and affirmative sentences, different values ​​are assigned to the speech coefficients accordingly. The two are then added together to obtain the speech coefficients. The text coefficient is calculated based on text feature data, and public opinion is quickly identified based on the part of speech and density of sensitive words; different values ​​are assigned to the text coefficients of generally negative and positive words based on the occurrence of extremely negative words; if a sensitive word appears more than a certain number of times within a certain period of time, the coefficient needs to be amplified.

5. The multimodal video public opinion intelligent monitoring method according to claim 2, characterized in that: The diffusion intensity coefficient = total information base value * interaction stability coefficient * dynamic gain coefficient; The information total base value = ln(total information in the statistical period + 1) * 10, where the information total includes the number of relevant content items from all channels, including text, video, and bullet comments, and is normalized. The normalized base value = (base value - historical minimum base value) / (historical maximum base value - historical minimum base value), where the maximum and minimum base values ​​are based on the most recent 30 days, and the base value is mapped to the interval [0,1]. The interaction stability coefficient is calculated using the variance of likes / shares / comments. The variance is set with a threshold based on historical experience values ​​and assigned a value according to the threshold. The dynamic gain coefficient is amplified when the 24-hour information volume growth rate or the interaction volume increase exceeds the corresponding threshold for two consecutive hours. The dynamic gain coefficient decreases when the interaction volume falls back to the threshold after 4 hours after the peak. The key node coefficient = KOL stance polarization value * dissemination path concentration; where KOL stance polarization value = |total influence of supporting KOLs - total influence of opposing KOLs| / total influence of the top ten KOLs, where KOLs are key opinion leaders, and their influence is calculated based on the number of followers / interactions; the coefficient is amplified when the polarization value is greater than the threshold; dissemination path concentration = (dissemination volume of the initial account + dissemination volume of secondary nodes) / total dissemination volume, the coefficient is amplified when the ratio is greater than the first threshold, and the coefficient is reduced when the ratio is less than the second threshold; The platform penetration coefficient is assigned a value directly based on the number of mainstream platform types.

6. The multimodal video public opinion intelligent monitoring method according to claim 1, characterized in that: The public opinion identification coefficient is as follows: B = G * H * Z; Where B is the public opinion identification coefficient, G is the public dissemination coefficient, H is the public interaction coefficient, and Z is the subject orientation coefficient; The public opinion tendency coefficient is as follows: Q=ω s *S+ω y *Y+ω w *W; Where Q is the public opinion tendency coefficient, S is the visual coefficient, Y is the voice coefficient, and W is the text coefficient. The weighting of the coefficients needs to be automatically adjusted according to different scenarios. The public opinion dissemination coefficient is as follows: C = K * J * P * η; Where C is the public opinion propagation coefficient, K is the diffusion intensity coefficient, J is the key node coefficient, P is the platform penetration coefficient, and η is the credibility coefficient, which is determined based on the initial information source.

7. The multimodal video public opinion intelligent monitoring method according to claim 1, characterized in that: The comprehensive public opinion index is as follows: A=α*B′+β*|Q′|*X+γ*C′; Where A is the comprehensive public opinion index, B is the public opinion identification coefficient, and B′ is the standardized public opinion identification coefficient; Q is the public opinion tendency coefficient, and Q′ is the standardized public opinion tendency coefficient; X is the emotional reinforcement coefficient, which amplifies Q when Q < 0, and X = 1.5 + tanh(3*|Q|); and attenuates Q when Q > 0, and C is the public opinion propagation coefficient, and C′ is the standardized public opinion propagation coefficient; α is the weight of the public opinion identification coefficient, β is the weight of the public opinion tendency coefficient and the emotional reinforcement coefficient, and γ is the weight of the public opinion propagation coefficient. The weight allocation varies for different periods.

8. A multimodal video public opinion intelligent monitoring system, characterized in that, Specifically, it includes: Multimodal data acquisition module: used to extract visual feature data, speech feature data and text feature data from videos, and to collect dynamic data of video propagation; Data processing module: used to clean and process the collected data, remove special effects and non-semantic symbol interference, and generate public opinion identification indicators, public opinion trend indicators and public opinion dissemination indicators; Coefficient generation module: used to generate public opinion judgment coefficients based on public opinion identification indicators, public opinion tendency coefficients based on public opinion tendency indicators, and public opinion dissemination coefficients based on public opinion dissemination indicators. Intelligent analysis module: used to generate a comprehensive public opinion coefficient from the public opinion judgment coefficient, public opinion tendency coefficient, and public opinion propagation coefficient; Dynamic threshold module: Used to establish a dual threshold mechanism, set thresholds for the comprehensive public opinion coefficient, classify and issue warnings. When the 24-hour growth rate of a certain dimension indicator exceeds 200%, it will automatically jump to a higher level. Public opinion visualization platform: used to visualize the three-dimensional propagation map of public opinion trends and generate propagation chain reports for hotspot tracing.

Images