Intelligent traffic management public opinion analysis and early warning system based on VIT multi-mode

Abstract

The invention belongs to the technical field of intelligent analysis and early warning, and discloses a traffic management public opinion intelligent analysis and early warning system based on VIT multi-mode, which comprises a platform adaptive data acquisition unit, a VIT cross-mode public opinion situation awareness module, a dynamic weight public opinion risk assessment module, a knowledge mapping public opinion guide aid decision module and a grading early warning push unit. According to the platform adaptive data acquisition unit, adaptive strategies are designed for various short video platforms, timed incremental pulling and high-heat real-time triggering acquisition are combined, videos, metadata and text information are covered at the same time, and the problems that existing public opinion monitoring is limited in coverage range and low in efficiency are solved; multi-modal analysis and association fusion are achieved through the VIT cross-modal public opinion situation awareness module, a three-level index system and real-time weight adjustment of the dynamic weight public opinion risk assessment module are matched, artificial experience judgment is replaced with quantitative calculation, and the problems that public opinion risk assessment is high in subjectivity and non-uniform in standard are solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent analysis and early warning, and particularly relates to an intelligent analysis and early warning system for traffic management public opinion based on VIT multi-modal. BACKGROUND

[0002] Traffic management public opinion is public opinion, event feedback and emotional expression related to traffic management work in the Internet, covering traffic accidents, traffic police on-duty and law enforcement, interpretation of traffic management policies, traffic facility settings, team management and other fields. With the rapid development of short video platforms, traffic management related videos are prone to form a communication hotspot in a short time due to their intuitive pictures, rapid information transmission and strong user interaction. If negative public opinion cannot be discovered and effectively controlled in time, it may lead to public doubts about traffic management work, and even affect public order and social stability. Therefore, it has become an urgent need for current traffic management work to realize real-time monitoring, in-depth analysis and early warning of short video traffic management public opinion with the help of intelligent technology.

[0003] In the prior art, the patent with publication number CN117332029A "Public opinion collection platform and method based on artificial intelligence and cloud computing technology" proposes a collection and analysis scheme for traffic management public opinion. It collects public transportation related public opinion data, analyzes the type, location and time of public opinion by combining artificial intelligence technology, and pushes the results marked on the map to the relevant departments, which improves the timeliness of public opinion collection to a certain extent.

[0004] However, the scheme and similar existing technologies still have the following problems: First, the coverage of public opinion monitoring is limited and the efficiency is low. It mainly relies on manual screening or single platform data grabbing, which cannot cover the full content of multiple short video platforms simultaneously, and it is also difficult to effectively capture key sensitive information in comments, bullet screen and topic tags. When facing massive short video data, the screening efficiency is extremely low, and the situation that public opinion has spread but has not been discovered often occurs. Second, the public opinion risk assessment is highly subjective and the standards are not unified. It mainly relies on manual experience judgment, and there are differences in the understanding of traffic management business and the grasp of public opinion transmission law among different personnel, which leads to deviations in risk grading of the same public opinion event, and cannot accurately measure the severity of public opinion. Finally, the public opinion disposal lacks standardized process and speech support. When facing sudden public opinion, the standard response speech needs to be written temporarily, and the coping measures need to be determined through a meeting, which not only consumes a lot of time, but also easily causes problems such as mismatch between disposal measures and actual situation and inconsistency in response caliber, making it difficult to effectively control the spread of public opinion. SUMMARY

[0005] In order to solve the problems of limited monitoring range and low efficiency, subjective risk assessment and non-standardized disposal process in the prior art, the present application provides an intelligent analysis and early warning system for traffic management public opinion based on VIT multi-modal, which improves the comprehensiveness of traffic management public opinion monitoring, the accuracy of risk assessment and the standardization of disposal, and provides effective support for traffic management public opinion control.

[0006] In order to achieve the above-mentioned purpose, the present application provides the following scheme: The intelligent analysis and early warning system for traffic management public opinion based on VIT multi-modal comprises a platform adaptive data acquisition unit, a VIT cross-modal public opinion situation awareness module, a dynamic weight public opinion risk assessment module, a knowledge graph public opinion guidance auxiliary decision module and a hierarchical early warning pushing unit. The platform adaptive data acquisition unit is used to collect short video files, video metadata and text data according to different short video platform characteristics. The VIT cross-modal public opinion situation awareness module is used to perform multi-modal analysis and associated fusion on the collected short video files, video metadata and text data, and generate video topic labels and public emotion classification. The dynamic weight public opinion risk assessment module is used to perform risk assessment on public opinion according to the video topic labels, the public opinion emotion score and the preset evaluation index system, and dynamically adjust the weight coefficients of the evaluation indexes according to the real-time situation, so as to obtain a dynamic weight total score. The knowledge graph public opinion guidance auxiliary decision module is used to construct a ternary knowledge graph among public opinion, historical cases and preplans, and recommend standardized disposal preplans and response standards for public opinion through a rule engine and semantic matching. The hierarchical early warning pushing unit is used to push early warning information to different pushing objects according to the risk level determined by the standardized disposal preplans, the response standards and the dynamic weight total score, and attach time limit reminders.

[0007] Preferably, the process of multi-modal analysis and associated fusion on the collected short video files, video metadata and text data, and generating video topic labels and public emotion classification comprises: The short video is frame-extracted at a preset frequency, input into a pre-trained VIT model, and the traffic management related elements in the picture are identified, including traffic law enforcement scenes and traffic accident scenes. The audio modal analysis is used to transcribe voice content, extract emotional tone features, and associate the actions in the vision. The text modal analysis is used to extract keywords in the title, comments and subtitles, and associate and match the places and character identities mentioned in the text with the geographical identifiers and personnel costumes in the vision.

[0008] Preferably, the process of identifying traffic management-related elements in the footage based on public sentiment classification and generating a public sentiment score includes: ; in, Scoring visual emotion features, Scoring based on audio sentiment and intonation features. Score the sentiment keywords in the text; , , These are the weighting coefficients for visual, audio, and text modalities, respectively.

[0009] Preferably, the process of conducting a risk assessment of public opinion based on video topic tags, public sentiment scores, and a pre-set evaluation indicator system, and dynamically adjusting the weight coefficients of the evaluation indicators according to real-time conditions to obtain a dynamic weighted total score includes: ; in, For the first The benchmark score of each basic characteristic indicator. For the first Spatiotemporal correlation weight coefficients of basic feature indicators; For the first The baseline score for each propagation and diffusion indicator, For the first Platform characteristic weighting coefficients for each propagation and diffusion indicator; The total score for public sentiment; and These represent the number of basic characteristic indicators and the number of propagation and diffusion indicators, respectively.

[0010] Preferably, the process of constructing a ternary knowledge graph connecting public opinion, historical cases, and contingency plans, and recommending standardized handling plans and response statements for public opinion through rule engines and semantic matching, includes: The node layer of the ternary knowledge graph includes public opinion feature nodes, historical case nodes, and contingency plan nodes; The association layer establishes a connection through semantic similarity and feedback on the handling effect. If the feature similarity between a new public opinion and a certain historical case reaches a preset threshold, and the public opinion heat drops more than a preset threshold after the case is handled by the contingency plan, then the public opinion feature node, historical case node and contingency plan node are strongly associated and marked with high fit. Build a contingency plan library and a policy definition library. The contingency plan library contains tiered response procedures, and the policy definition library provides standardized templates for various scenarios. Pre-set business rules and threshold conditions, and automatically activate the corresponding contingency plans and statements when public opinion meets the conditions; By mining similar case handling experience through knowledge graphs, personalized handling suggestions can be recommended.

[0011] Preferably, the calculation process of the similarity of the new public opinion and the characteristics of a historical case comprises: ; wherein, is the feature vector of the new public opinion query, is the feature vector of the historical case; is the dot product operator, is the Euclidean norm of the vector.

[0012] Preferably, according to the risk level determined by the standardized disposal plan, the response range and the total score of the dynamic weight, the process of grading and pushing the early warning information to different push objects and attaching time limit reminders comprises: low-risk public opinion is only pushed to the public opinion monitoring post; medium-risk public opinion is pushed to the public opinion monitoring post and the command center dispatch post; high-risk public opinion is pushed to the public opinion monitoring post, the command center and the decision-making layer, and mobile phone short message reminders are sent synchronously through real-time push protocols; the early warning information is displayed in a structured manner according to the core characteristics of public opinion, the core factors of risk and the preliminary disposal suggestions.

[0013] Compared with the prior art, the present application has the following advantages: The present application solves the problems of limited coverage and low efficiency of existing public opinion monitoring by designing an adaptive strategy for multiple short video platforms through a platform-adaptive data acquisition unit, combining timed incremental pulling and high-heat real-time triggering acquisition, and covering video, metadata and text information. The present application solves the problems of strong subjectivity and non-uniform standards of public opinion risk assessment by realizing multi-modal analysis and correlation fusion through a VIT (Vision Transformer, Vision Transformer) cross-modal public opinion situation awareness module, and by using a three-level index system and real-time weight adjustment of a dynamic weight public opinion risk assessment module to replace manual experience judgment with quantitative calculation. The present application solves the problems of lack of standardized process and speech support in public opinion disposal by constructing a triple knowledge graph through a knowledge graph-based public opinion guidance and auxiliary decision-making module, and by relying on semantic matching to recommend standardized plans and ranges, thereby improving the comprehensiveness of traffic management public opinion monitoring, the accuracy of risk assessment and the standardization of disposal, and providing effective support for traffic management public opinion control. BRIEF DESCRIPTION OF DRAWINGS

[0014] In order to more clearly illustrate the technical solutions of the present application, the following briefly introduces the drawings needed in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0015] Figure 1 This is a flowchart of a platform-adaptive data acquisition process according to an embodiment of the present invention; Figure 2 This is a flowchart of the VIT cross-modal analysis according to an embodiment of the present invention; Figure 3 This is a flowchart illustrating the dynamic risk assessment process according to an embodiment of the present invention. Figure 4 This is a flowchart illustrating the knowledge graph-assisted decision-making process according to an embodiment of the present invention. Figure 5 This is a flowchart illustrating the hierarchical early warning push process according to an embodiment of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] Example 1 Please see Figures 1-5 This invention provides a traffic management public opinion intelligent analysis and early warning system based on VIT multimodal, including: a platform-adaptive data acquisition unit, a VIT (Vision Transformer) cross-modal public opinion situation perception module, a dynamic weighted public opinion risk assessment module, a knowledge graph-based public opinion guidance and auxiliary decision-making module, and a hierarchical early warning push unit; The platform-adaptive data acquisition unit is used to collect short video files, video metadata, and text data for different short video platforms. The VIT cross-modal public opinion situation awareness module is used to perform multimodal analysis and correlation fusion of collected short video files, video metadata, and text data, and generate video topic tags and public sentiment classifications. The generated video topic tags are used to supplement the feature dimensions of the subsequent dynamic weighted public opinion risk assessment module, and the public sentiment classification provides the basic classification basis for calculating the public opinion sentiment score. The VIT cross-modal public opinion situation awareness module is also used to identify traffic management-related elements in the footage and generate a public opinion sentiment score using the following formula. : ; in, Scoring visual emotion features, Scoring based on audio sentiment and intonation features. Score the sentiment keywords in the text; , , These are the weighting coefficients for visual, audio, and text modalities, respectively; this public sentiment score is the total public sentiment score in the dynamic weighted public sentiment risk assessment module. It is directly used as one of the core indicators for risk assessment; Based on the multimodal feature fusion results, the video topic tag relevance score is generated using the following formula, and the top 3 with the highest relevance are selected as the final topic tags: ,in Calculate the confidence level (0-10 points) for visual traffic management element recognition. Audio semantic relevance score (0-10 points). Text keyword matching score (0-10). , , The modal weights are all initially set to 0.33 and can be dynamically adjusted according to the characteristics of the platform's content to ensure that the tags are highly consistent with the core content of public opinion.

[0019] The dynamic weighted public opinion risk assessment module is used to assess public opinion risks based on a preset assessment indicator system, and dynamically adjusts the weight coefficients of the assessment indicators according to real-time conditions to obtain a dynamic weighted total score. The dynamic weighted public opinion risk assessment module calculates the dynamic weighted total score using the following formula. : ; in, For the first The benchmark score of each basic characteristic indicator. For the first Spatiotemporal correlation weight coefficients of basic feature indicators; For the first The baseline score for each propagation and diffusion indicator, For the first Platform characteristic weighting coefficients for each propagation and diffusion indicator; The total score for public sentiment; and These are the quantities of basic characteristic indicators and propagation and diffusion indicators, respectively. The knowledge graph-based public opinion guidance and decision-making assistance module constructs a ternary knowledge graph connecting public opinion, historical cases, and contingency plans. Through a rule engine and semantic matching, it recommends standardized handling plans and response statements for public opinion. These recommended standardized handling plans and response statements are synchronized to the tiered early warning push unit and pushed to the corresponding recipients along with the early warning information, providing immediate reference for public opinion handling. The knowledge graph-based public opinion guidance and decision-making assistance module calculates the semantic similarity between new public opinion and historical cases using the following formula. : ; in, This is the feature vector for new public opinion queries. The feature vectors of historical cases; This is the dot product operator. Let be the Euclidean norm of the vector; To enhance the reliability of the association, a formula for calculating the association strength is introduced: ,in The effectiveness coefficient of handling historical cases (when the rate of decrease in public opinion intensity after handling is ≥50%) When it is 30%-50% When <30% ), Matching coefficient for applicable scenarios of the contingency plan (scenarios are completely consistent) Partially consistent ), It is determined to be a strong association at that time.

[0020] The tiered early warning push unit is used to push early warning information to different push objects according to the risk level determined by the dynamic weighted total score, and to attach time limit reminders.

[0021] Specifically, this addresses the problems of limited coverage and low efficiency in existing technologies for public opinion monitoring, strong subjectivity and inconsistent standards in risk assessment, and a lack of standardized procedures and rhetoric for handling public opinion.

[0022] First, a platform-adaptive data collection unit was built. Taking into account the differences in interface specifications and data structures of different short video platforms, a differentiated collection strategy was designed to simultaneously collect video metadata such as short video files, play counts, and like counts, as well as related text data.

[0023] Next, the VIT (Vision Transformer, a visual recognition model based on the Transformer architecture, capable of accurate image content parsing and feature extraction) cross-modal public opinion situation awareness module is activated. Frames are extracted from the short video at a rate of 2 frames per second. These frames are then input into the pre-trained VIT (Vision Transformer model, a visual recognition model based on the Transformer architecture, achieving accurate recognition of complex scene elements through image patch embedding and attention mechanisms) model to identify traffic management-related elements such as law enforcement and traffic accidents. Existing speech-to-text algorithms are used to transcribe the audio content and extract emotional tone features, which are then associated with visual actions of people. Keywords are extracted from the text data, and locations mentioned in the text are matched with geographical markers such as road signs, and personal identities are matched with clothing. Finally, a public opinion sentiment score is calculated using a formula. ; In the formula, Visual emotion feature scores are obtained by analyzing facial expressions and scene atmosphere in frame images. The audio emotion and intonation feature score is based on speech rate, volume, and pitch variation. Scoring is assigned to text sentiment keywords, quantified based on the frequency of negative keywords. , , These are the weighting coefficients for visual, audio, and text modalities, respectively, with an initial value of 0.33 for each, which can be adjusted according to platform characteristics.

[0024] Next, a three-tiered assessment indicator system is constructed using a dynamic weighted public opinion risk assessment module. Basic characteristic indicators include the severity of the event and its scope; dissemination and diffusion indicators include the speed of dissemination and the number of users reached; and public sentiment indicators are as described above. Corresponding total score The dynamic weighted total score is calculated using the following formula: ; In the formula, The baseline score for the i-th basic feature indicator ranges from 0 to 10. The spatiotemporal correlation weight coefficient for the i-th basic feature indicator is set to 1.5 for sensitive periods and key areas, and 1.0 for the rest. The base score for the j-th propagation and diffusion index is 0-10. The platform characteristic weighting coefficient is set to 1.4 for the j-th propagation and diffusion index, and 1.0 for the rest. and The number of basic characteristic indicators and propagation and diffusion indicators are both set to 5. The total score for public sentiment is calculated, ranging from 0 to 10. A ternary knowledge graph is then constructed using a knowledge graph-based public opinion guidance and decision-making support module. The node layer includes three types of nodes: public opinion characteristics, historical cases, and contingency plans. The semantic similarity between new public opinion and historical cases is calculated using a formula: ; In the formula, The feature vector for new public opinion queries is composed of dimensions such as event type and sentiment tendency. The feature vectors of historical cases have a structure consistent with new public opinion. This is the dot product operator. The similarity is the Euclidean norm of the vectors; a similarity greater than 0.8 indicates a successful match and a corresponding recommendation is given. Finally, the tiered early warning push unit pushes information based on... The risk level is determined, and early warning information is pushed to the corresponding targets. This implementation method achieves comprehensive coverage of public opinion across multiple platforms, improves the objectivity of risk assessment, provides standardized handling support, and effectively solves existing technical problems.

[0025] In this embodiment, the platform-adaptive data acquisition unit is specifically used for: The system employs a combination of daily timed incremental retrieval and real-time triggering based on high popularity to collect short video files, video metadata, and text data. The text data includes video titles, subtitles, comments, and hashtags; Adaptive data collection strategies are designed for different short video platforms, including adding dedicated data collection channels for comment sections on platforms with high comment interaction density and linking them to the corresponding video content.

[0026] Specifically, based on the platform-adaptive data acquisition unit of the previous embodiment, a collection mode combining daily timed incremental retrieval and real-time triggering based on high popularity is adopted. During the low system load period from 2 AM to 4 AM daily, incremental retrieval of newly added traffic management-related short video data from the previous day ensures data integrity. Text data specifically covers video titles, automatically generated and manually added subtitles, user comments in the comment section, and topic tags, comprehensively capturing user feedback information. For platforms with intensive comment interaction, a dedicated comment section collection channel is added, directly connecting to the platform's comment database via API interface. Each comment is associated with its corresponding video subject through a video ID, preventing comments from being disconnected from videos. This implementation method improves the targeting and comprehensiveness of data collection, ensuring no sensitive information in the comment section is missed, and further improving the efficiency of public opinion monitoring.

[0027] In this embodiment, the platform-adaptive data acquisition unit is further used for: The privacy content filtering mechanism automatically removes non-public content involving personal privacy, retaining only publicly accessible information; When the number of views or comments on a short video exceeds a preset threshold within a preset time, real-time data collection is automatically triggered.

[0028] Specifically, based on the platform-adaptive data collection unit of the previous embodiment, a privacy content filtering mechanism is established. Through a combination of keyword matching and semantic recognition, non-public content involving personal privacy such as ID numbers, mobile phone numbers, and home addresses is automatically removed, retaining only publicly accessible public opinion-related information to protect user privacy and security. In implementation, a preset timeframe of 24 hours can be set. When a short video's views exceed 100,000 or comments exceed 5,000 within 24 hours, a real-time collection mechanism is automatically triggered. The system immediately retrieves the complete data and related interactive information of the video, quickly capturing high-profile public opinion dynamics and preventing the spread of public opinion before it is discovered.

[0029] In this embodiment, the VIT cross-modal public opinion situation awareness module is specifically used for: The short video is framed at a preset frequency and input into a pre-trained VIT model to identify traffic management-related elements in the video, including duty enforcement scenes and traffic accident scenes. By transcribing speech content through audio modality analysis, emotional tone features are extracted and correlated with visual actions; Keywords were extracted from titles, comments, and captions through text modal analysis, and the locations and identities of people mentioned in the text were matched with visual geographical markers and clothing.

[0030] Specifically, based on the VIT cross-modal public opinion situation awareness module of the above embodiment, frames are extracted from short videos at a preset frequency of 3 frames per second. The extracted frame images are input into a pre-trained VIT model. The model learns traffic management scene features and accurately identifies traffic management-related elements in the video, such as traffic police stopping vehicles for inspection and traffic accident scenes such as vehicle collisions. Audio content is transcribed using existing speech recognition algorithms to extract emotional tone features such as accelerated speech and intense tone, and these are associated with visual body movements such as arguments and onlookers, enhancing the accuracy of emotion judgment. Keywords in titles, comments, and subtitles are extracted using text mining algorithms. Locations mentioned in the text, such as an intersection, are associated with visual geographical markers such as road signs and landmarks, and the identities of people, such as traffic police and drivers, are associated with visual uniforms and clothing features, achieving deep fusion of multimodal data. This implementation improves the accuracy of public opinion situation awareness, making topic tags and emotion classifications more consistent with reality.

[0031] In this embodiment, the VIT cross-modal public opinion situation awareness module is further used for: An automatic mechanism for mining new words in public opinion is established. When a word that is not included in the index appears more frequently than a preset threshold in traffic management public opinion on multiple platforms and the proportion of associated negative emotions exceeds a preset threshold, it is automatically added to the keyword library. The duty enforcement scenario model and the traffic accident scenario model are periodically iterated, and newly emerging scenario features are added to the scenario library.

[0032] Specifically, based on the VIT cross-modal public opinion situation awareness module in the previous embodiment, an automatic new word mining mechanism for public opinion is constructed. A preset threshold for word frequency is set at 50 times, and a preset threshold for the proportion of associated negative sentiment is set at 70%. For example, when a word not included in the database appears more than 50 times in traffic management public opinion across multiple short video platforms, and the proportion of associated negative sentiment exceeds 70%, the system automatically adds the word to the keyword library and updates the sensitive word library in real time. The duty enforcement scenario model and the traffic accident scenario model are iterated monthly, collecting newly emerging scenario features such as new enforcement equipment and special accident types to supplement the scenario library, ensuring that the model can recognize constantly changing traffic management public opinion scenarios. This implementation method allows the system to adapt to the dynamic changes in public opinion content and continuously improve its scenario recognition and keyword capture capabilities.

[0033] In this embodiment, the dynamic weighted public opinion risk assessment module is specifically used for: Construct a three-tiered evaluation indicator system that includes basic characteristic indicators, dissemination and diffusion indicators, and public sentiment indicators; The baseline score of the basic characteristic index is set as the baseline score multiplied by the spatiotemporal correlation weight coefficient, which is adjusted according to the correlation between real-time sensitive time and sensitive location. The baseline score for the propagation and diffusion index is set as the baseline score multiplied by the platform characteristic weight coefficient, which is adjusted according to the propagation speed characteristics of different short video platforms.

[0034] Specifically, based on the dynamic weighted public opinion risk assessment module in the above embodiments, a three-level assessment indicator system is constructed, comprising basic characteristic indicators, dissemination and diffusion indicators, and public sentiment indicators. Basic characteristic indicators include the nature of the event and its scope of impact; dissemination and diffusion indicators include the number of reposts and the scope of dissemination; and the public sentiment indicator is the total public sentiment score. The baseline score of the basic characteristic indicators is multiplied by a spatiotemporal correlation weight coefficient, which is adjusted according to real-time conditions. During major events and sensitive times and locations such as around schools, the coefficient is set to 1.5, while for ordinary times and areas, the coefficient is set to 1.0. The baseline score of the dissemination and diffusion indicators is multiplied by a platform characteristic weight coefficient, which is adjusted according to the dissemination speed characteristics of different short video platforms. For platforms with fast dissemination speeds, the coefficient is set to 1.4, and for platforms with slower dissemination speeds, the coefficient is set to 1.0. This implementation method makes the risk assessment more closely aligned with real-world scenarios, improving the accuracy and objectivity of the assessment results.

[0035] In this embodiment, the dynamic weighted public opinion risk assessment module is further used for: The risk level of public opinion is determined based on the dynamic weighted total score, including low risk, medium risk, and high risk. Generate a risk contribution report to identify the core indicators that drive risk escalation.

[0036] Specifically, based on the dynamic weighted public opinion risk assessment module of the previous embodiment, the dynamic weighted total score is used to assess public opinion risk. Determine the risk level of public opinion. A score below 10 indicates low risk, 10-20 indicates medium risk, and above 20 indicates high risk, clearly defining the risk classification standards. A risk contribution report is generated, detailing the scores and weightings of each basic characteristic indicator and dissemination indicator, identifying core indicators driving risk escalation such as high dissemination speed and strong negative sentiment. This implementation method allows staff to clearly understand the degree of public opinion risk and key triggers, providing a basis for targeted handling and improving the efficiency of public opinion management.

[0037] In this embodiment, the knowledge graph-based public opinion guidance and decision-making assistance module is specifically used for: The node layer of the ternary knowledge graph includes public opinion feature nodes, historical case nodes, and contingency plan nodes; The association layer establishes a connection through semantic similarity and feedback on the handling effect. If the feature similarity between a new public opinion and a certain historical case reaches a preset threshold, and the public opinion heat drops more than a preset threshold after the case is handled by the contingency plan, then the public opinion feature node, historical case node and contingency plan node are strongly associated and marked with high fit. Build a contingency plan library and a policy definition library. The contingency plan library contains tiered response procedures, and the policy definition library provides standardized templates for various scenarios.

[0038] Specifically, based on the knowledge graph-based public opinion guidance and decision-making assistance module of the above embodiments, the node layer of the ternary knowledge graph includes public opinion feature nodes such as event type and sentiment tendency, historical case nodes such as similar past traffic management public opinion events, and contingency plan nodes such as corresponding handling plans. The association layer establishes associations through semantic similarity and handling effect feedback. If the feature similarity between a new public opinion and a historical case reaches a preset threshold of 0.8, and the public opinion heat drops by more than a preset threshold of 50% after the contingency plan used in that case is handled, then the three are strongly associated and marked as highly adaptable. A contingency plan library and a policy document library are constructed. The contingency plan library includes graded response processes such as monitoring processes for low-risk public opinion and emergency handling processes for high-risk public opinion. The policy document library provides standardized templates for various scenarios such as traffic accident responses and law enforcement dispute responses. This implementation method enables rapid matching of handling contingency plans and response policies, reduces the time for ad-hoc writing and discussion, and improves the standardization of public opinion handling.

[0039] In this embodiment, the knowledge graph-based public opinion guidance and decision-making assistance module is further used for: Pre-set business rules and threshold conditions, and automatically activate the corresponding contingency plans and statements when public opinion meets the conditions; By mining similar case handling experience through knowledge graphs, personalized handling suggestions can be recommended.

[0040] Specifically, based on the knowledge graph-based public opinion guidance and decision-making assistance module of the previous embodiment, preset business rules and threshold conditions, such as automatically activating the Level 1 response plan for high-risk public opinion and automatically matching specific response guidelines for public opinion involving sensitive areas. When public opinion meets the corresponding conditions, the corresponding plan and guidelines are automatically activated. By mining the handling experience of similar cases through knowledge graph, key links and time nodes of the handling measures in the cases are extracted. Combined with the unique characteristics of new public opinion, personalized handling suggestions are recommended, such as adding the linkage of local traffic management departments for public opinion in specific areas.

[0041] The suitability of personalized treatment recommendations is calculated using the following formula: ,in Number of key steps in handling similar cases (default) ), For the first The importance weight of each key link ( ), For new public opinion and historical cases Similarity of features between links (0-1). For historical cases The effectiveness of each step is scored (0-10 points) for screening. The top 3 suggestions are used as personalized recommendations.

[0042] This implementation method combines automation and personalization in public opinion management, further improving the effectiveness and efficiency of the management.

[0043] To verify the system's effectiveness, a full-process calculation example is provided based on a real-world scenario: Assume a short video content is "a traffic police enforcement conflict at an intersection." After being collected by the platform-adaptive data collection unit, the VIT cross-modal public opinion situation perception module analyzes the data at a rate of 3 frames per second, scoring the visual emotion features. (Significant negative tendency), audio emotional tone feature score (Intense tone), text emotion keyword score (Contains a large number of negative words), and sets the weight coefficients for visual, audio, and text as follows: , , Through formula Calculate the public sentiment score ,Right now The dynamic weighted public opinion risk assessment module includes five basic characteristic indicators, among which the event severity benchmark score is... Spatiotemporal correlation weight coefficient (Sensitive intersection), the baseline score for the other four basic indicators is 6, the weighting coefficient is 1.0, and the total score for basic features is... There are a total of 5 indicators for the spread and diffusion, with the spread speed as the baseline score. Platform characteristic weighting coefficient (High-profile platform), the baseline score for the other four dissemination indicators is 7, with a weighting coefficient of 1.0, resulting in a total dissemination and diffusion score. ; through formula Calculate the dynamic weighted total score The situation was identified as a high-risk public opinion event. The knowledge graph-based public opinion guidance and decision-making assistance module calculated the similarity between this public opinion event and historical cases. (Above the threshold of 0.8), historical case handling effectiveness coefficient (Popularity decreased by 60%), Scene Matching Coefficient (Completely identical), through The calculated value of {R}_{Link} = 0.85 × 1.2 × 1.0 = 1.02 ≥ 0.7 is considered a strong association, matching the corresponding high-fitness plan. In the personalized suggestion fitness calculation, {W}_1 = 0.3, {W}_2 = 0.25, {W}_3 = 0.2, {W}_4 = 0.15, and {W}_5 = 0.1. , , , , , , , , , ,pass The calculated value is {F}_{Adapt}=0.3×0.9×9.5 + 0.25×0.88×9.2 + 0.2×0.85×8.8 + 0.15×0.82×8.5 + 0.1×0.8×8.0=8.63}, and the top 3 personalized suggestions are recommended.

[0044] The tiered early warning push unit sends early warning information to monitoring posts, command centers, and decision-making levels, simultaneously attaching standardized contingency plans, response guidelines, and personalized suggestions, and sending SMS reminders. Practical application verification shows that the system takes only 15 minutes from data collection to early warning push, improving efficiency by 80% compared to traditional manual methods. The risk assessment accuracy rate reaches 92%, and the public opinion heat decreased by 65% ​​within 24 hours after the response, fully validating the effectiveness of the model algorithm and system functions.

[0045] In this embodiment, the tiered early warning push unit is specifically used for: Low-risk public opinion is only pushed to the public opinion monitoring team; Medium-risk public opinion information is pushed to the public opinion monitoring post and the command center dispatch post; High-risk public opinion is pushed to public opinion monitoring posts, command centers and decision-making levels, and mobile phone text message reminders are sent simultaneously through real-time push protocol; The warning information is presented in a structured manner according to the core characteristics of public opinion, the core risk factors, and preliminary handling suggestions, and is accompanied by standardized handling plans, response guidelines, and personalized handling suggestions matched with knowledge graphs.

[0046] Specifically, based on the tiered early warning push unit of the above embodiments, tiered push is performed according to the risk level determined by the dynamic weighted total score. Low-risk public opinion is only pushed to the public opinion monitoring post, where monitoring personnel continuously monitor the dynamics; medium-risk public opinion is pushed to the public opinion monitoring post and the command center dispatch post, where the dispatch post prepares for handling; high-risk public opinion is pushed to the public opinion monitoring post, the command center, and the decision-making level, and simultaneously sent to mobile phone SMS reminders through a real-time push protocol to ensure that the decision-making level is informed in a timely manner. The early warning information is displayed in a structured manner according to the core characteristics of public opinion, such as event type, location of occurrence, core risk factors such as high dissemination, strong negative emotions, and preliminary handling suggestions, allowing the recipient to quickly grasp key information. This implementation method achieves accurate push of early warning information, ensuring that staff at different levels obtain the corresponding information in a timely manner, and buying time for rapid handling of public opinion.

[0047] In summary, this invention addresses the limitations of existing public opinion monitoring methods by designing adaptation strategies for various short video platforms through a platform-adaptive data acquisition unit. It combines timed incremental data retrieval with real-time triggering of high-trend data collection, covering video, metadata, and text information. Furthermore, it utilizes a VIT cross-modal public opinion situation awareness module to achieve multimodal analysis and correlation fusion, coupled with a dynamic weighted public opinion risk assessment module's three-level indicator system and real-time weight adjustment. This quantitative calculation replaces manual experience-based judgment, resolving the issues of strong subjectivity and inconsistent standards in public opinion risk assessment. Finally, it constructs a ternary knowledge graph through a knowledge graph-based public opinion guidance and decision-making support module, relying on semantic matching to recommend standardized plans and statements. This addresses the lack of standardized procedures and rhetoric support for public opinion handling, comprehensively improving the comprehensiveness of traffic management public opinion monitoring, the accuracy of risk assessment, and the standardization of handling, providing effective support for traffic management public opinion control.

[0048] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A traffic management public opinion intelligent analysis and early warning system based on VIT multimodal communication, characterized in that, The system includes: a platform-adaptive data acquisition unit, a VIT cross-modal public opinion situation perception module, a dynamic weighted public opinion risk assessment module, a knowledge graph-based public opinion guidance and decision-making assistance module, and a hierarchical early warning push unit. The platform-adaptive data acquisition unit is used to collect short video files, video metadata, and text data for different short video platforms. The VIT cross-modal public opinion situation perception module is used to perform multimodal analysis and correlation fusion of the collected short video files, video metadata and text data, and generate video theme tags and public sentiment classifications; the VIT cross-modal public opinion situation perception module is also used to identify traffic management-related elements in the video according to the public sentiment classifications and generate public opinion sentiment scores. The dynamic weighted public opinion risk assessment module is used to assess the risk of public opinion based on video topic tags, public opinion sentiment scores and a preset assessment indicator system, and dynamically adjust the weight coefficients of the assessment indicators according to the real-time situation to obtain a dynamic weighted total score. The knowledge graph-based public opinion guidance and decision-making assistance module is used to construct a ternary knowledge graph between public opinion, historical cases, and contingency plans, and recommends standardized handling contingency plans and response guidelines for public opinion through rule engines and semantic matching. The tiered early warning push unit is used to push early warning information to different push targets based on the risk level determined by the standardized response plan, response criteria, and dynamic weighted total score, and to include time limit reminders.

2. The system according to claim 1, characterized in that, The process of performing multimodal analysis and correlation fusion on the collected short video files, video metadata, and text data to generate video topic tags and public sentiment classifications includes: The short video is framed at a preset frequency and input into a pre-trained VIT model to identify traffic management-related elements in the video, including duty enforcement scenes and traffic accident scenes. By transcribing speech content through audio modality analysis, emotional tone features are extracted and correlated with visual actions; Keywords were extracted from titles, comments, and captions through text modal analysis, and the locations and identities of people mentioned in the text were matched with visual geographical markers and clothing.

3. The system according to claim 1, characterized in that, The process of identifying traffic management-related elements in the footage and generating a public sentiment score based on public sentiment classification includes: ； in, Scoring visual emotion features, Scoring based on audio sentiment and intonation features. Score the sentiment keywords in the text; , , These are the weighting coefficients for visual, audio, and text modalities, respectively.

4. The system according to claim 1, characterized in that, The process of conducting a risk assessment of public opinion based on video topic tags, sentiment scores, and a pre-set evaluation indicator system, and dynamically adjusting the weight coefficients of the evaluation indicators according to real-time conditions to obtain a dynamic weighted total score, includes: ； in, For the first The benchmark score of each basic characteristic indicator. For the first Spatiotemporal correlation weight coefficients of basic feature indicators; For the first The baseline score for each propagation and diffusion indicator, For the first Platform characteristic weighting coefficients for each propagation and diffusion indicator; The total score for public sentiment; and These represent the number of basic characteristic indicators and the number of propagation and diffusion indicators, respectively.

5. The system according to claim 1, characterized in that, The process of constructing a ternary knowledge graph connecting public opinion, historical cases, and contingency plans, and recommending standardized handling plans and response statements for public opinion through rule engines and semantic matching, includes: The node layer of the ternary knowledge graph includes public opinion feature nodes, historical case nodes, and contingency plan nodes; The association layer establishes a connection through semantic similarity and feedback on the handling effect. If the feature similarity between a new public opinion and a certain historical case reaches a preset threshold, and the public opinion heat drops more than a preset threshold after the case is handled by the contingency plan, then the public opinion feature node, historical case node and contingency plan node are strongly associated and marked with high fit. Build a contingency plan library and a policy definition library. The contingency plan library contains tiered response procedures, and the policy definition library provides standardized templates for various scenarios. Pre-set business rules and threshold conditions, and automatically activate the corresponding contingency plans and statements when public opinion meets the conditions; By mining similar case handling experience through knowledge graphs, personalized handling suggestions can be recommended.

6. The system according to claim 5, characterized in that, The calculation process for the similarity of new public opinion trends with a certain historical case includes: ； in, This is the feature vector for new public opinion queries. The feature vectors of historical cases; This is the dot product operator. Let be the Euclidean norm of the vector.

7. The system according to claim 1, characterized in that, Based on standardized emergency response plans, response guidelines, and risk levels determined by dynamic weighted total scores, the process of sending tiered warning information to different recipients, along with time-limited reminders, includes: Low-risk public opinion is only pushed to the public opinion monitoring team; Medium-risk public opinion information is pushed to the public opinion monitoring post and the command center dispatch post; High-risk public opinion is pushed to public opinion monitoring posts, command centers and decision-making levels, and mobile phone text message reminders are sent simultaneously through real-time push protocol; The early warning information is presented in a structured manner, based on the core characteristics of public opinion, the core risk factors, and preliminary handling suggestions.

Images