A network public opinion anomaly detection method, device and system

Abstract

The application discloses a network public opinion anomaly detection method, device and system. Dynamic triggering is combined with multi-dimensional quantitative indicators, negative sentiment growth rate and comment quantity growth rate are used for preliminary triggering and screening, four-dimensional features of sentiment fluctuation, heat spread, high-heat negative comment aggregation and cross-platform synchronous anomaly are fused for weighted scoring, non-anomaly fluctuation interference is reduced, the synergistic influence of sentiment diffusion and spread intensity is considered, the early warning accuracy and recall rate are improved, low-confidence secondary review and historical benchmark comparison are used to eliminate misjudgment caused by semantic ambiguity and benchmark drift, the reliability is enhanced, key words, representative comments and spread nodes are output through traceability explanation, verifiable evidence chain is provided, and weight and threshold are adjusted through artificial feedback to realize adaptive optimization. In conclusion, the scheme significantly improves the timeliness, accuracy and interpretability of public opinion anomaly detection, and has cross-platform adaptive capability.

Description

Technical Field

[0001] This application relates to the field of public opinion detection technology, and in particular, to a method, apparatus and system for detecting abnormal online public opinion. Background Technology

[0002] Online platforms, short video platforms, social media, news comment sections, and forums continuously generate data such as comment text, posting time, number of likes, number of replies, number of reposts, comment source, user identifiers, and hashtags. This data contains both public sentiment and the speed of dissemination and intensity of interaction, serving as crucial input for public opinion monitoring systems to assess risks.

[0003] The closest existing implementation can complete basic public opinion monitoring, but its processing flow is usually statically configured. For example, the system performs the same analysis steps regardless of the strength of the risk; when negative sentiment suddenly rises, the system cannot automatically call clustering and tracing tools to locate the triggering topic; when the popularity rises but there is a clear divergence in sentiment, the system cannot automatically arrange sentiment review or extract representative comments; when a warning is manually confirmed to be a false alarm, the system cannot automatically adjust the threshold for the same topic or similar platforms.

[0004] From a technical implementation perspective, the core flaws of the existing solutions are: they do not convert the dynamic changes in comment data into an executable analysis task plan; they do not record the observation results of each analysis tool and change the subsequent call chain based on the observation results; they do not bind sentiment anomalies, popularity anomalies, and source tracing explanation results to the same task status table; and they do not use human feedback to update the parameters of subsequent monitoring strategies for similar events, resulting in inaccurate public opinion detection. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this application provides a method, apparatus, and system for detecting abnormal online public opinion, in order to solve the problem of inaccurate results in existing online public opinion detection schemes.

[0006] The technical solution adopted by this application to solve its technical problem is: Firstly, a method for detecting abnormal online public opinion is provided, including: Obtain comment data for a target topic within a continuous time window. The comment data includes comment text, posting time, and popularity data, wherein the popularity data includes at least one of the following: number of likes, number of replies, and number of reposts. Calculate the growth rate of negative sentiment and the growth rate of the number of comments for each time window; When the growth rate of negative sentiment and the growth rate of the number of comments exceed their respective growth thresholds, sentiment fluctuation scores are calculated. Popularity dissemination score High score for gathering of negative comments And cross-platform synchronization of abnormal scores ; Calculate the risk score of abnormal public opinion , , , , , These are the weighting coefficients corresponding to each score; When the risk score is greater than a preset risk threshold, a source tracing and explanation task is performed to obtain evidence of the cause of the anomaly. An early warning result is generated based on the evidence of the cause of the anomaly.

[0007] As an optional implementation of this application, the negative sentiment growth rate is calculated as follows: in, Indicates the first The growth rate of negative sentiment within a specific time window. Indicates the first The proportion of negative emotions within a given time window. , Indicates the first The number of negative comments within a time window Indicates the first The total number of comments within a time window. Indicates the first The proportion of negative emotions within a given time window. This represents a very small constant to prevent division by zero; the growth rate of the number of comments is calculated as follows: in, This represents the growth rate of the number of comments in the t-th time window. This represents the total number of comments within the (t-1)th time window.

[0008] As an optional implementation of this application, the emotional fluctuation score is calculated based on the proportion of negative emotions, the growth rate of negative emotions, the magnitude of change in the mean emotional value, and the intensity of emotional fluctuation in the current time window. The popularity spread score is calculated based on the total popularity, average popularity, popularity growth rate, proportion of high-popularity comments, and proportion of high-popularity negative comments within the current time window. The cluster score of hot negative comments is calculated based on the proportion of hot negative comments, the proportion of negative sentiment, and the growth rate of negative sentiment in the current time window. The cross-platform synchronization anomaly score is calculated based on cross-platform keyword synchronization, cross-platform time synchronization, and the number of cross-platform dissemination platforms.

[0009] As an optional implementation of this application, the emotional fluctuation score The calculation formula is:

[0010] The heat dissipation score The calculation formula is:

[0011] The high-scoring negative comments cluster score The calculation formula is:

[0012] Cross-platform synchronization of abnormal scores The calculation formula is:

[0013] in, Represents the normalization function. Indicates the proportion of negative emotions. Indicates the growth rate of negative emotions. This indicates the magnitude of change in the mean sentiment. Indicates the intensity of emotional fluctuations. Indicates the total heat of the window. Indicates the growth rate of popularity. This indicates the percentage of highly popular negative comments. Indicates the degree of cross-platform keyword synchronization. Indicates cross-platform time synchronization. This represents the normalized value of the number of platforms used for cross-platform dissemination. to These are the weight coefficients corresponding to each sub-feature.

[0014] Each sub-feature is calculated in the following manner:

[0015]

[0016]

[0017]

[0018] in, Indicates the first A collection of comments within a time window. This represents the total number of comments within the set. Indicates the first The symbolic sentiment score of each comment. and These represent the positive probability and the negative probability, respectively. This represents the mean sentiment at time t. This represents the mean sentiment value at time t-1.

[0019]

[0020]

[0021] in, Indicates the first The popularity value of each comment This indicates the high heat threshold determined based on a historical baseline window. Indicates the emotion category, This indicates a function that takes the value 1 if the condition is true and 0 otherwise; the aforementioned scoring formula... , and Take the corresponding window respectively , and ;

[0022]

[0023]

[0024]

[0025]

[0026] in, Indicates the first The number of platforms participating in the target topic within a given time window. Indicates platform The set of keywords in this window, Indicates platform The moment when the abnormal trigger condition is first met. Indicates the time decay scale, Indicates the total number of monitoring platforms; when When less than 2, and All are set to 0;

[0027] in, and Representing features respectively Within the lower and upper bounds of the historical benchmark window, the normalization result is taken as 0 when it is less than 0 and as 1 when it is greater than 1; the aforementioned scoring formula... , and Take the corresponding window respectively , and .

[0028] As an optional implementation of this application, the execution of the source tracing and interpretation task includes: For negative comments within a time window where the risk score exceeds a preset risk threshold, keyword extraction, topic clustering, representative comment selection, and propagation node location are performed. The output includes a list of main keywords that triggered the anomaly, representative high-profile negative comments, the earliest anomaly time window, the main propagation platforms, and the risk contribution factors of each factor, serving as evidence of the cause of the anomaly.

[0029] As an optional implementation of this application, before performing the source tracing and interpretation task, the following is also included: Estimate the average sentiment confidence level corresponding to the current risk score:

[0030]

[0031] in, , , They represent the first The probability that a comment belongs to the positive, negative, or neutral category. Indicates the first Sentiment confidence of the comments Indicates the first Average sentiment confidence within a time window Indicates the first The total number of comments within a time window; When the average sentiment confidence score is lower than the confidence threshold, a sentiment review analysis task is performed. The low-confidence comments are re-evaluated using a backup sentiment classification model, and the risk score is updated based on the re-evaluation results.

[0032] As an optional implementation of this application, it also includes: Determine whether the proportion of negative comments in the current time window exceeds the average of the historical baseline window; If the number of comments exceeds the threshold and the growth rate of the number of comments increases synchronously, then a sentiment review analysis task is executed to perform a second sentiment assessment on the negative comments within the current time window to confirm whether the negative sentiment has truly increased, and the number of negative comments and the growth rate of negative sentiment in the current time window are updated based on the results of the second assessment.

[0033] As an optional implementation of this application, it also includes: Receive manual feedback information on the warning results; The weighting coefficients α, β, γ, and δ may be adjusted based on the human feedback information, or the growth threshold may be adjusted, or the preset risk threshold may be adjusted.

[0034] Secondly, a device for detecting abnormal online public opinion is provided, comprising: The comment data acquisition module is used to acquire comment data of a target topic within a continuous time window. The comment data includes comment text, posting time, and popularity data. The popularity data includes at least one of the following: number of likes, number of replies, and number of reposts. The growth rate calculation module is used to calculate the negative sentiment growth rate and the comment quantity growth rate for comment data in each time window. The dimension score calculation module is used to calculate sentiment fluctuation scores when the negative sentiment growth rate and the comment quantity growth rate are respectively higher than the corresponding growth thresholds. Popularity dissemination score High score for gathering of negative comments And cross-platform synchronization of abnormal scores ; The risk scoring calculation module is used to calculate the risk score for abnormal public opinion events. , , , , , These are the weighting coefficients corresponding to each score; The abnormal evidence acquisition module is used to perform a source tracing and interpretation task and acquire evidence of the cause of the abnormality when the risk score is greater than a preset risk threshold. The early warning result generation module is used to generate early warning results based on the evidence of the cause of the anomaly.

[0035] Thirdly, a system for detecting abnormal online public opinion is provided, including: At least one processor and at least one memory; The memory stores the executable instructions of the processor; The processor is configured to perform any of the above-described methods for detecting network public opinion anomalies.

[0036] Beneficial effects: This application provides a method, device, and system for detecting anomalies in online public opinion. By combining a dynamic triggering mechanism with multi-dimensional quantitative indicators, it solves the problems of false alarms, missed alarms, and delayed warnings caused by fixed thresholds and single-dimensional judgments in existing solutions. Specifically, it first uses the growth rate of negative sentiment and the growth rate of the number of comments for initial trigger screening. Then, it integrates features from four dimensions—sentiment fluctuation, popularity spread, high-profile negative comment aggregation, and cross-platform synchronous anomalies—for weighted scoring. This reduces interference from non-abnormal fluctuations while taking into account the synergistic effect of sentiment diffusion and propagation intensity, improving the accuracy and recall rate of warnings. Simultaneously, through low-confidence secondary verification and historical benchmark comparison mechanisms, it further eliminates misjudgments caused by semantic ambiguity or benchmark drift, enhancing the credibility of warnings. By tracing and interpreting the source, it outputs keywords, representative comments, and propagation nodes, providing managers with a verifiable chain of evidence. Through manual feedback to adjust weights and thresholds, it achieves adaptive optimization of the monitoring strategy. In summary, this solution significantly improves the timeliness, accuracy, and interpretability of anomaly detection, and possesses cross-platform adaptability. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a flowchart of a method for detecting abnormal online public opinion provided in an embodiment of this application; Figure 2 This is a schematic diagram of the structure of an abnormal network public opinion detection device provided in an embodiment of this application; Figure 3 This is a system architecture diagram provided in an embodiment of this application; Figure 4 This is a flowchart of a multi-agent collaborative judgment process provided in an embodiment of this application; Figure 5 This is a schematic diagram of a task state update for a scheduling agent provided in an embodiment of this application; Figure 6 This is a schematic diagram of a risk scoring and early warning output provided in an embodiment of this application. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be described in detail below with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments of this application, all other implementation methods obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0040] Existing methods for detecting anomalies in online public opinion typically treat data collection, sentiment classification, popularity statistics, and early warning output as a fixed pipeline, which mainly suffers from the following technical problems: 1. Fixed analysis process. Existing systems generally calculate keyword hit rate, number of comments, or sentiment tendency of individual texts according to preset rules. They cannot automatically decide whether to add source tracing analysis, cluster analysis, secondary sentiment review, or manual verification tasks based on the current public opinion status, resulting in insufficient timeliness in the early identification of anomalies.

[0041] 2. Insufficient synergy among multi-source features. Sentiment analysis can reflect commentary trends, but it's difficult to determine whether negative sentiment is spreading; popularity statistics can reflect the intensity of dissemination, but it's difficult to determine whether increased discussion carries negative risks. When these two are separated, it's easy to misjudge as "hot discussion but low risk" or "lots of negative sentiment but weak dissemination."

[0042] 3. Lack of an interpretable task loop. Existing methods mostly output risk scores or levels directly, without explaining the time window that triggers anomalies, representative comments, key keywords, propagation nodes, and the judgment criteria of various analysis tools, which is not conducive to managers' review and handling.

[0043] 4. Thresholds and strategies are difficult to adapt. The baseline of comments varies significantly across different platforms, topics, and time periods, making it easy for fixed thresholds to cause false positives or false negatives; manual verification results are also difficult to be fed back into subsequent monitoring strategies in a timely manner.

[0044] To solve the above problems, refer to Figure 1 This application provides a method for detecting abnormal online public opinion, including: S11: Obtain comment data for the target topic within a continuous time window. The comment data includes comment text, posting time, and popularity data. The popularity data includes at least one of the following: number of likes, number of replies, and number of reposts. For example, comment data can be obtained from different platforms, and then it can be determined whether it is related to the target topic based on keywords. Alternatively, the cosine similarity between the vector of the comment data and the vector of the target topic can be calculated, and the comment data is determined to be related to the target topic if the cosine similarity is greater than a preset similarity.

[0045] In addition, when acquiring comment data, it is also necessary to mark its source platform for subsequent risk assessment.

[0046] S12: Calculate the negative sentiment growth rate and comment quantity growth rate for comment data in each time window; In a preferred implementation of this application, the negative sentiment growth rate is calculated as follows: in, Indicates the first The growth rate of negative sentiment within a specific time window. Indicates the first The proportion of negative emotions within a given time window. , Indicates the first The number of negative comments within a time window Indicates the first The total number of comments within a time window. Indicates the first The proportion of negative emotions within a given time window. This represents a very small constant to prevent division by zero; the growth rate of the number of comments is calculated as follows: in, This represents the growth rate of the number of comments in the t-th time window. This represents the total number of comments within the (t-1)th time window.

[0047] This application prioritizes calculating the growth rate of negative sentiment and the growth rate of the number of comments for each time window, rather than directly counting the absolute number or proportion of negative comments. The reason for this is that the essential characteristic of public opinion risk is mutation rather than existing stock. The growth rate can reflect the speed at which emotions deteriorate and the scale of discussion expands, capturing abnormal signals in the early stages of public opinion fermentation, thereby greatly improving the timeliness of early warning. At the same time, the growth rate calculation is a lightweight statistical operation. Using it as a pre-screening condition, subsequent high-cost multi-dimensional in-depth analysis is only triggered when both are simultaneously higher than the threshold. This can effectively avoid performing complex calculations on each comment during calm periods, significantly saving system resources. In addition, binding the growth rate of negative sentiment and the growth rate of the number of comments as a co-triggered event can accurately filter out interference scenarios of emotional venting but with a very small spread and heated discussions but with neutral sentiment, accurately identifying high-risk public opinion events where negative consensus is spreading.

[0048] S13: When the growth rate of negative sentiment and the growth rate of the number of comments are both higher than the corresponding growth thresholds, calculate the sentiment fluctuation score respectively. Popularity dissemination score High score for gathering of negative comments And cross-platform synchronization of abnormal scores ; The emotional fluctuation score is calculated based on the proportion of negative emotions, the growth rate of negative emotions, the magnitude of change in the mean emotional value, and the intensity of emotional fluctuations within the current time window. The popularity spread score is calculated based on the total popularity, average popularity, popularity growth rate, proportion of high-popularity comments, and proportion of high-popularity negative comments within the current time window. The cluster score of hot negative comments is calculated based on the proportion of hot negative comments, the proportion of negative sentiment, and the growth rate of negative sentiment in the current time window. The cross-platform synchronization anomaly score is calculated based on cross-platform keyword synchronization, cross-platform time synchronization, and the number of cross-platform dissemination platforms.

[0049] Preferably, the emotional fluctuation score The calculation formula is:

[0050] The heat dissipation score The calculation formula is:

[0051] The high-scoring negative comments cluster score The calculation formula is:

[0052] Cross-platform synchronization of abnormal scores The calculation formula is:

[0053] in, Represents the normalization function. Indicates the proportion of negative emotions. Indicates the growth rate of negative emotions. This indicates the magnitude of change in the mean sentiment. Indicates the intensity of emotional fluctuations. Indicates the total heat of the window. Indicates the growth rate of popularity. This indicates the percentage of highly popular negative comments. Indicates the degree of cross-platform keyword synchronization. Indicates cross-platform time synchronization. This represents the normalized value of the number of platforms used for cross-platform dissemination. to These are the weight coefficients corresponding to each sub-feature.

[0054] Each sub-feature is calculated in the following manner:

[0055]

[0056]

[0057]

[0058] in, Indicates the first A collection of comments within a time window. This represents the total number of comments within the set. Indicates the first The symbolic sentiment score of each comment. and These represent the positive probability and the negative probability, respectively. This represents the mean sentiment at time t. This represents the mean sentiment value at time t-1.

[0059]

[0060]

[0061] in, Indicates the first The popularity value of each comment This indicates the high heat threshold determined based on a historical baseline window. Indicates the emotion category, This indicates a function that takes the value 1 if the condition is true and 0 otherwise; the aforementioned scoring formula... , and Take the corresponding window respectively , and ;

[0062]

[0063]

[0064]

[0065]

[0066] in, Indicates the first The number of platforms participating in the target topic within a given time window. Indicates platform The set of keywords in this window, Indicates platform The moment when the abnormal trigger condition is first met (i.e., the moment when the growth rate of negative sentiment and the growth rate of the number of comments on the platform simultaneously exceed the corresponding growth threshold), Indicates the time decay scale, Indicates the total number of monitoring platforms; when When less than 2, and All are set to 0;

[0067] in, and Representing features respectively Within the lower and upper bounds of the historical benchmark window, the normalization result is taken as 0 when it is less than 0 and as 1 when it is greater than 1; the aforementioned scoring formula... , and Take the corresponding window respectively , and .

[0068] This approach quantifies abnormal public opinion from four independent dimensions: emotional fluctuations, widespread popularity, high-profile negative aggregation, and cross-platform synchronous anomalies. It can characterize risks from four dimensions: the speed of emotional deterioration, the intensity of dissemination, the concentration of negative impact, and the breadth of cross-platform spread. Sudden changes in any dimension can be detected in a timely manner. At the same time, a comprehensive judgment can be made through weighted fusion. This avoids misjudgments caused by relying on a single indicator (such as high popularity but neutral sentiment, or strong negativity but weak dissemination). It can also distinguish the risk transmission paths of different types of public opinion, making the scoring results more interpretable and providing a clear direction for subsequent source tracing, localization, and differentiated handling.

[0069] S14: Calculate the risk score for abnormal public opinion. , , , , , These are the weighting coefficients corresponding to each score.

[0070] S15: When the risk score is greater than the preset risk threshold, perform the source tracing and explanation task to obtain evidence of the cause of the anomaly; The task of performing source tracing and interpretation includes: For negative comments within a time window where the risk score exceeds a preset risk threshold, keyword extraction, topic clustering, representative comment selection, and propagation node location are performed. The output includes a list of main keywords that triggered the anomaly, representative high-profile negative comments, the earliest anomaly time window, the main propagation platforms, and the risk contribution factors of each factor, serving as evidence of the cause of the anomaly.

[0071] Representative comments can be selected by comprehensively sorting them according to average sentiment confidence, popularity value, comment posting time and distance from the topic center.

[0072] Topic clustering can employ keyword vector clustering, text vector clustering, or semantic clustering based on pre-trained language models; The location of propagation nodes can be statistically analyzed based on the comment source platform, parent comment, forwarding link, or highly interactive users.

[0073] In addition, prior to performing the source tracing and interpretation task, the following steps are also included: Estimate the average sentiment confidence level corresponding to the current risk score:

[0074]

[0075] in, , , They represent the first The probability that a comment belongs to the positive, negative, or neutral category. Indicates the first Sentiment confidence of the comments Indicates the first Average sentiment confidence within a time window Indicates the first The total number of comments within a time window; When the average sentiment confidence score is lower than the confidence threshold, a sentiment review analysis task is performed. The low-confidence comments are re-evaluated using a backup sentiment classification model, and the risk score is updated based on the re-evaluation results.

[0076] It should be noted that when sentiment classification models handle sarcastic, rhetorical, pun-based, or context-dependent comments, their output probability distribution often exhibits a state of "near approximate probabilities across the three categories," meaning the confidence level is low. Directly including such unreliable labels in negative comment statistics and risk scoring will lead to distorted scoring. By estimating the average sentiment confidence level before source interpretation and triggering a backup model for secondary discrimination when it falls below a threshold, we can conduct targeted verification of samples with questionable initial classification, ensuring that the data sources for the final risk scoring and evidence output are accurate and reliable.

[0077] S16: Generate an early warning result based on the evidence of the cause of the anomaly.

[0078] Specifically, alert results are generated based on the anomaly level, source tracing explanations, and business rules. These alert results include the topic name, risk level, risk score, triggering reason, key time window, representative comments, relevant keywords, suggested actions, and a manual review entry point. For high-risk events (events with extremely high risk scores), work orders can be automatically generated, designated personnel notified, or the frequency of subsequent sampling increased.

[0079] It should be noted that this also includes: determining whether the proportion of negative comments in the current time window exceeds the average of the historical baseline window; If the number of comments exceeds the threshold and the growth rate of the number of comments increases synchronously, then a sentiment review analysis task is executed to perform a second sentiment assessment on the negative comments within the current time window to confirm whether the negative sentiment has truly increased, and the number of negative comments and the growth rate of negative sentiment in the current time window are updated based on the results of the second assessment.

[0080] The core function of this step is to eliminate misjudgments caused by baseline drift and semantic ambiguity. The average of the historical baseline window reflects the public sentiment baseline of the topic under normal conditions. Only when the negative proportion in the current window exceeds this baseline and comments surge simultaneously is it judged as a real anomaly worthy of attention, rather than normal daily fluctuations. At the same time, triggering a secondary sentiment review under this condition can re-verify the low-confidence negative labels caused by semantic ambiguity such as sarcasm, rhetorical questions, and abbreviations in the initial classification, avoiding misjudging neutral banter as real negative sentiment, and ensuring that the numerical sources of the subsequent negative sentiment growth rate and comment number growth rate participating in risk scoring are authentic and reliable.

[0081] As a preferred implementation of the embodiments of this application, it further includes: Receive manual feedback information on the warning results; The weighting coefficients α, β, γ, and δ may be adjusted based on the human feedback information, or the growth threshold may be adjusted, or the preset risk threshold may be adjusted.

[0082] Because the baseline of comments varies across different platforms and topics, fixed thresholds and weights are difficult to adapt to all scenarios. By introducing human feedback to dynamically adjust weight coefficients, growth thresholds, and risk thresholds, the system can continuously learn the actual risk performance of specific topic types or platforms, quantifying expert experience into iterative parameter updates. This allows subsequent early warning strategies to gradually align with real business scenarios, effectively reducing false positives and false negatives caused by static configurations.

[0083] It should be noted that any process or method description in the flowchart or otherwise described herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order according to the functions involved, as should be understood by those skilled in the art to which the embodiments of this application pertain.

[0084] Furthermore, in the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0085] Based on the same inventive concept, such as Figure 2 As shown, this application provides a network public opinion anomaly detection device 20, which applies the network public opinion anomaly detection method provided in the above embodiments. The device includes: The comment data acquisition module 21 is used to acquire comment data of a target topic within a continuous time window. The comment data includes comment text, posting time, and popularity data. The popularity data includes at least one of the following: number of likes, number of replies, and number of reposts. The growth rate calculation module 22 is used to calculate the negative sentiment growth rate and the comment quantity growth rate of the comment data for each time window; Dimensional score calculation module 23 is used to calculate sentiment fluctuation scores when the negative sentiment growth rate and the comment quantity growth rate are respectively higher than the corresponding growth thresholds. Popularity dissemination score High score for gathering of negative comments And cross-platform synchronization of abnormal scores ; Risk scoring calculation module 24 is used to calculate the risk score for abnormal public opinion. , , , , , These are the weighting coefficients corresponding to each score; The abnormal evidence acquisition module 25 is used to perform a source tracing and interpretation task and acquire evidence of the cause of the abnormality when the risk score is greater than a preset risk threshold. The early warning result generation module 26 is used to generate an early warning result based on the evidence of the cause of the anomaly.

[0086] Based on the same inventive concept, embodiments of this application provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the network public opinion anomaly detection method provided in any of the above embodiments.

[0087] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0088] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0089] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0090] The storage media mentioned above can be read-only memory, disk, or optical disk, etc.

[0091] Based on the same inventive concept, this application also provides a network public opinion anomaly detection system, including: At least one processor and at least one memory; The memory stores the executable instructions of the processor; The processor is configured to execute the network public opinion anomaly detection method provided in the above embodiments.

[0092] To more clearly illustrate the proposed solution, a network public opinion anomaly detection and early warning system based on multi-agent collaborative analysis is provided below. Targeting online platform comment data, a scheduling agent serves as the task hub, incorporating sentiment fluctuations, trending topics, aggregation of highly popular negative comments, and source tracing as callable analysis tools or sub-agent capabilities, forming a closed loop of "monitoring and perception, task planning, tool invocation, result observation, risk decision-making, and feedback optimization."

[0093] like Figure 3 As shown, the system includes a data access layer 101, an intelligent agent collaboration layer 102, a feature calculation layer 103, a risk decision layer 104, and a result output layer 105. The data access layer 101 is used to access comment text, interaction popularity, and time window data; the intelligent agent collaboration layer 102 is used to organize multiple intelligent agents to complete public opinion analysis tasks; the feature calculation layer 103 is used to calculate features such as sentiment fluctuations, popularity propagation, topic clustering, and risk scores; the risk decision layer 104 is used to generate risk levels, triggering reasons, handling suggestions, and manual review tasks; and the result output layer 105 is used to output processing results to the front-end page, message notification system, work order system, or strategy update module.

[0094] For example, the data access layer connects to data sources such as Weibo, Bilibili, short video platforms, news comment sections, and forums to obtain comment text, posting time, number of likes, number of replies, number of reposts, comment ID, topic ID, user ID, and source platform. The intelligent agent collaboration layer includes scheduling agents, monitoring agents, sentiment analysis agents, trending topics agents, anomaly detection agents, source tracing and explanation agents, early warning decision-making agents, and feedback optimization agents. The feature calculation layer includes text preprocessing tools, sentiment analysis tools, trending topics analysis tools, time window statistics tools, keyword extraction tools, clustering and source tracing tools, and risk scoring tools. The risk decision-making layer generates risk levels, early warning reasons, handling suggestions, and manual review tasks. The result output layer outputs early warning results to front-end pages, messaging systems, or ticketing systems.

[0095] To avoid interpreting agents as abstract functional names, each agent can be deployed as a software service with a unified interface. This unified interface includes input fields, output fields, status codes, and confidence scores. Input fields include topic identifier, platform identifier, time window identifier, comment set, historical baseline window, and task context; output fields include feature results, observation results, confidence score, evidence list, and next action suggestions. Agents do not directly modify the original comment database; instead, they exchange intermediate results through a task status table.

[0096] Multi-agent collaborative judgment process as follows Figure 4The diagram illustrates the task flow from comment collection to early warning output. After comment data is collected and preprocessed, the scheduling agent generates an analysis task plan. The scheduling agent then calls upon the sentiment analysis agent, the trending topics agent, the anomaly analysis agent, and the source tracing and explanation agent according to the task plan. The anomaly analysis results are further input into the early warning decision agent to generate early warning results. Human feedback results are input into the feedback optimization agent, which returns the updated strategy to the scheduling agent for subsequent task planning.

[0097] The monitoring agent collects comment data according to a preset list of platforms, topic keywords, or event identifiers. Each comment data entry includes at least the comment text, posting time, and popularity data. Popularity data includes one or more of the following: number of likes, replies, shares, and comment growth rate. The system cleans the comment text, removes duplicate data, filters invalid characters, normalizes sensitive characters, and divides the time window. The time window can be set to 10 minutes, 30 minutes, 1 hour, or dynamically adjusted according to the comment growth rate. When the number of comments increases rapidly within a unit of time, the monitoring agent sends a "surge in popularity observation result" to the scheduling agent.

[0098] Comment data records can adopt the following field structure: Comment ID to identify a single comment; Topic ID to identify the monitored topic; Platform ID to identify the source platform; Publication Time to record the comment's publication time; Comment Body to record the comment text; Number of Likes, Replies, and Shares to record interaction data; Parent Comment ID to record the comment hierarchy; Sentiment Category and Sentiment Confidence to record the sentiment analysis results; and Comment Popularity Value to record the popularity result calculated by the popularity propagation agent. This field structure allows subsequent agents to perform continuous processing on the same data object.

[0099] like Figure 5 The diagram illustrates how the scheduling agent performs closed-loop processing around the task state table. Observational inputs include trigger information such as sentiment fluctuations and sudden increases in popularity; the task plan generation process determines the list of agents to be invoked and their invocation order; the tool invocation process executes sentiment, popularity, source tracing, and scoring tasks; and the result feedback process writes feature results and interpretive evidence into the task state table. The task state table records the task identifier, topic identifier, current window, execution status, observation results, and termination conditions, and is updated cyclically until a warning output or task termination condition is met.

[0100] Specifically, the scheduling agent maintains a task status table, which includes topic identifier, current time window, executed tools, tool return results, risk candidate scores, pending tasks, and termination conditions. After receiving signals from the monitoring agent regarding comment increments, sentiment fluctuations, or abnormal popularity, the scheduling agent generates a public opinion analysis task plan. The task plan includes at least the agents to be invoked, the invocation order, the input data range, output fields, and task termination conditions.

[0101] The task status table may include the following fields: task identifier, topic identifier, time window identifier, trigger type, historical baseline window, list of executed agents, list of pending agents, list of observation results, candidate risk score, evidence pointer, manual feedback status, and termination condition. Trigger types include one or more of the following: surge in sentiment, surge in popularity, insufficient confidence, clustering of highly negative sentiment, and cross-platform synchronization anomalies. Termination conditions include: risk score below the attention threshold, completion of source tracing and explanation, issuance of a warning, manual termination, or reaching the maximum number of call rounds.

[0102] Comment growth rate The change in the number of comments in the current time window relative to the previous time window can be represented as:

[0103] in, Indicates the first The total number of comments within a time window. Indicates the first The total number of comments within a time window. This represents a very small constant that prevents division by zero.

[0104] The task planning rules for scheduling agents include: if the negative sentiment growth rate... The growth rate of the number of comments is higher than the first growth threshold. If the sentiment score exceeds the first popularity threshold, it indicates a simultaneous increase in negative sentiment and discussion scale. The scheduling agent adds the sentiment analysis agent, popularity propagation agent, anomaly judgment agent, and source tracing and explanation agent to the pending tasks. If the sentiment analysis confidence score is lower than the confidence threshold, it indicates that the single sentiment classification result is unstable. The scheduling agent adds a sentiment review analysis task to re-evaluate low-confidence comments using backup sentiment models, keyword rules, or manual review entry points. If the same or similar keywords appear on multiple platforms within adjacent time windows, it indicates that the topic may have cross-platform spread. The scheduling agent adds a cross-platform synchronization anomaly calculation task to statistically analyze the degree of keyword synchronization, temporal relationship, and risk contribution between different platforms. If the risk score is lower than the attention threshold, subsequent source tracing tasks are stopped to reduce computational resource consumption.

[0105] For example, when the proportion of negative comments exceeds the average of the historical baseline window and the growth rate of the number of comments increases simultaneously, the sentiment analysis task is used to confirm whether negative sentiment has truly increased, the heat disseminated task is used to calculate the intensity of interactive dissemination, the high-heat negative comment identification task is used to filter negative comments with dissemination influence, and the source explanation task is used to output triggering keywords, representative comments, and abnormal time windows; when the heat is rising but the sentiment confidence is low, the sentiment review analysis task is used to reduce misjudgments caused by sarcasm, rhetorical questions, abbreviations, or semantic ambiguity; when the same keyword appears simultaneously on multiple platforms, the cross-platform aggregation task is used to determine whether the topic has developed from a single-platform discussion into a cross-platform public opinion event.

[0106] Sentiment Analysis Agent The sentiment analysis agent identifies the sentiment category of the preprocessed comment text and outputs the positive probability, negative probability, neutral probability, and confidence score. For the first... For each comment, the sentiment category is chosen from the category with the highest probability among positive, negative, and neutral probabilities. The sentiment confidence score for each comment is... It can be represented as:

[0107] No. Average sentiment confidence level for each time window It can be represented as:

[0108] in, , , They represent the first The probability that a comment belongs to the positive, negative, or neutral category. Indicates the first Sentiment confidence of the comments Indicates the first Average sentiment confidence within a time window Indicates the first The total number of comments within a time window. When When the confidence level is below the threshold, or when the proportion of low-confidence comments is higher than the preset ratio, the scheduling agent triggers the sentiment review analysis task.

[0109] For the The system calculates the proportion of negative emotions within a specific time window. Negative emotional growth rate , Mean change in sentiment and intensity of emotional fluctuations The proportion of negative emotions can be expressed as:

[0110] The parameters are explained below: Indicates the first The proportion of negative emotions within a given time window. Indicates the first The number of negative comments within a time window Indicates the first The total number of comments within a time window.

[0111] The growth rate of negative emotions can be expressed as:

[0112] The parameters are explained below: Indicates the first The growth rate of negative sentiment within a specific time window. Indicates the first The proportion of negative emotions within a given time window. Indicates the first The proportion of negative emotions within a given time window. This represents a very small constant that prevents division by zero.

[0113] Heat Dissemination Intelligent Agent The trend propagation agent calculates trend characteristics based on the number of likes, replies, shares, and the growth rate of comments. For the first... 100 comments, comment popularity score It can be represented as:

[0114] The parameters are explained below: Indicates the first The popularity value of each comment This indicates the number of likes. Indicates the number of replies. Indicates the number of reposts. This indicates the contribution of the topic to which this comment belongs to the growth of comments in the current window. , , , These are the configurable weighting coefficients corresponding to the popularity feature.

[0115] The contribution to the growth of comments is calculated in the following way:

[0116]

[0117] in, Indicates the first The non-negative comment growth rate of a time window relative to the previous window. This means that the growth rate will be evenly distributed across the current window. The growth contribution from each comment; when the number of comments does not increase, Set it to 0. The popularity propagation agent will assign each comment a value of 0. Write to the task status table, and with It performs window aggregation, high-profile comment identification, and high-profile negative comment identification based on the foundation.

[0118] The system further calculates the total popularity of the window, the average popularity, the popularity growth rate, the proportion of highly popular comments, and the proportion of highly popular negative comments.

[0119] The system uses a set of historical baseline windows prior to the current time window. Extract comment popularity values ​​and determine high popularity thresholds using historical mean and standard deviation:

[0120]

[0121]

[0122]

[0123]

[0124]

[0125] in, and These represent the mean and standard deviation of the comment popularity values ​​within the historical baseline window, respectively. This is the threshold sensitivity coefficient. Indicates the average heat of the window. This indicates the percentage of highly popular comments. This represents the proportion of comments that simultaneously meet both high popularity and negative sentiment criteria out of the total number of comments in the current window; thus, the popularity value of a single comment is calculated. The calculations are performed sequentially on the window's total popularity, average popularity, popularity growth rate, percentage of highly popular comments, and percentage of highly popular negative comments.

[0126] Cross-platform keyword synchronization, time synchronization, and the normalized values ​​of the number of dissemination platforms are calculated as follows:

[0127]

[0128]

[0129]

[0130]

[0131] in, For the platform With the platform Jaccard similarity of keyword sets The time proximity of the anomaly trigger times on the two platforms is considered; the remaining parameters are consistent with the definitions in claim 4. The anomaly assessment agent substitutes the above-mentioned current window indicators into the calculation formulas for sentiment fluctuation score, popularity spread score, high-popularity negative comment aggregation score, and cross-platform synchronization anomaly score, respectively.

[0132] Risk scoring and early warning output, such as Figure 6 The diagram illustrates the scoring input, fusion calculation, and early warning output process of the anomaly assessment intelligent agent. The emotional fluctuation score E, the popularity spread score H, the high-profile negative comment aggregation score C, and the cross-platform synchronous anomaly score S are input to the feature fusion unit; the feature fusion unit generates a risk score R; the risk score is input to the level judgment unit to obtain a normal, attention, warning, or high-risk level; the early warning result unit outputs the risk level, triggering reason, and handling suggestions.

[0133] The anomaly detection agent normalizes the characteristics of emotional fluctuations, the characteristics of trending spread, the characteristics of clustered negative comments, and the characteristics of cross-platform synchronization anomalies, and calculates the emotional fluctuation score for each. Popularity dissemination score High score for gathering of negative comments and cross-platform synchronization of abnormal scores Among them, the emotional fluctuation score It can be represented as:

[0134] Popularity Spread Score It can be represented as:

[0135] High score for gathering negative comments It can be represented as:

[0136] Cross-platform synchronization of abnormal scores It can be represented as:

[0137] in, Represents the normalization function. Indicates the proportion of negative emotions. Indicates the growth rate of negative emotions. This indicates the magnitude of change in the mean sentiment. Indicates the intensity of emotional fluctuations. Indicates the total heat of the window. Indicates the growth rate of popularity. This indicates the percentage of highly popular negative comments. Indicates the degree of cross-platform keyword synchronization. Indicates cross-platform time synchronization. This represents the normalized value of the number of platforms used for cross-platform dissemination. to These are the weight coefficients corresponding to each sub-feature.

[0138] The anomaly analysis agent weights and fuses the scores from the four categories mentioned above to generate a public opinion anomaly risk score. The risk score can be represented as:

[0139] The parameters are explained below: Indicates the risk score of abnormal public opinion. Indicates the score of emotional fluctuation. Indicates the score for the spread of popularity. The score indicates a high concentration of negative comments. This indicates an abnormal score during cross-platform synchronization. , , , These are the weighting coefficients corresponding to the scores of each feature.

[0140] Anomaly detection agent based on The value classifies public opinion status into normal, attention, warning, and high risk.

[0141] Tracing the origins of intelligent agents When the risk score reaches the attention level or above, the scheduling agent invokes the source tracing and explanation agent. The source tracing and explanation agent extracts keywords, clusters topics, selects representative comments, and locates propagation nodes for negative comments within the high-risk time window. It outputs the main keywords triggering the anomaly, representative high-profile negative comments, the earliest anomaly time window, major propagation platforms, comment growth curves, and risk contribution factors. This output serves as the explanation for the warning result, rather than simply providing an unexplainable score.

[0142] Representative comments can be selected by comprehensively ranking them according to sentiment confidence, popularity, comment posting time, and distance from the topic center. Topic clustering can be achieved using keyword vector clustering, text vector clustering, or semantic clustering based on a pre-trained language model; propagation node localization can be based on statistics from the comment source platform, parent comments, forwarding links, or highly interactive users.

[0143] The source tracing and explanation results should include at least the evidence identifier, keyword list, representative comment identifier, earliest anomaly time window, main dissemination platforms, risk contribution factors, and explanatory text.

[0144] Early warning decision-making intelligent agent The early warning decision-making intelligent agent generates early warning results based on the anomaly level, source tracing and explanation results, and business rules. The early warning results include the topic name, risk level, risk score, triggering reason, key time window, representative comments, relevant keywords, suggested actions, and a manual review entry point. For high-risk events, the early warning decision-making intelligent agent can automatically generate work orders, notify designated personnel, or increase the subsequent sampling frequency.

[0145] Feedback Optimization Agent After managers confirm, reject, or adjust the warning level, the feedback optimization agent records human feedback and adjusts the thresholds, weights, sampling frequencies, and task planning rules for similar topics or platforms based on the feedback. For example, when a certain type of entertainment topic frequently generates high-profile but low-risk discussions, the system reduces the weight of the popularity spread score in that type of topic; when a certain type of public safety topic experiences a small-scale negative gathering that requires attention, the system increases the weight of the emotional fluctuation score and shortens the time window.

[0146] Taking the comment section of a sudden event on a certain online platform as an example, the system establishes a time window every 10 minutes. In the t-th window, there are a total of 2000 comments, with 900 negative comments. The proportion of negative comments in the previous window was 0.20, so the proportion of negative sentiment in the current window is 0.45, and the growth rate of negative sentiment is 1.25. At the same time, the number of likes and replies to many negative comments in this window is significantly higher than the historical benchmark, and the proportion of highly popular negative comments exceeds a preset threshold. The monitoring agent reports this observation to the scheduling agent; the scheduling agent generates a task plan, calls the sentiment analysis agent to verify the comment sentiment, calls the popularity propagation agent to calculate the propagation intensity, calls the anomaly judgment agent to generate a risk score, calls the source tracing and explanation agent to extract representative comments and triggering keywords, and finally, the early warning decision agent outputs the "early warning" level and handling suggestions. If the early warning is confirmed to be effective by humans, the feedback optimization agent increases the sampling frequency of similar events and updates the weight parameters.

[0147] Furthermore, to reduce computational resource consumption, the scheduling agent only invokes high-cost source tracing or secondary verification tasks when trigger conditions are met; when the risk is below the attention threshold, only basic statistical results are retained. For data that may involve personal information, such as user identifiers, the system can use hashing, anonymization, or only store anonymous identifiers within the platform; when displaying representative comments on the alert output side, user accounts can be hidden according to business needs. This processing does not change the technical process of public opinion anomaly detection, but helps the system comply with legal compliance requirements in actual deployment.

[0148] The embodiments of this application verify the technical effect through the following indicators: on the same historical dataset, compared with the fixed threshold pipeline scheme, the anomaly detection lead time, false alarm rate, false negative rate, average calculation time and manual review pass rate are compared; on the same topic type, the threshold adaptation effect before and after adding feedback optimization is compared; on the same risk event, whether keywords, representative comments, trigger windows and risk contribution factors can be output are compared.

[0149] It should be noted that the number of agents can be adjusted. For example, the sentiment analysis agent and the trend propagation agent can be merged into a feature analysis agent, or fact-checking agents, rumor identification agents, cross-platform propagation path agents, or multilingual analysis agents can be added.

[0150] The scheduling method can be replaced. The scheduling agent can use rule engines, finite state machines, reinforcement learning strategies, task planners based on large language models, or a combination of the above methods to generate task plans.

[0151] Sentiment analysis models can be used as alternatives. Traditional machine learning models, deep learning models, pre-trained Chinese language models, large language models with prompts for classification, or multi-model voting methods can be employed to output sentiment categories and confidence scores.

[0152] Popularity metrics can be substituted. In addition to the growth rates of likes, replies, reposts, and comments, metrics such as the number of favorites, shares, exposure, number of user followers, comment level, interaction speed, and propagation path length can also be added.

[0153] Risk scoring methods can be replaced. Weighted summation, logistic regression, gradient boosting trees, neural network scoring models, Bayesian updates, or multi-indicator ranking models can be used to generate risk scores.

[0154] Feedback optimization methods can be used as alternatives. Human feedback can be used to adjust thresholds and weights, as well as to update the sample library, trigger model retraining, change the sampling frequency, or update task planning rules.

[0155] Output methods can be alternatives. Alert results can be output to the front-end page, SMS, email, WeChat Work, DingTalk, ticketing system, or emergency command platform.

[0156] Compared with existing fixed-line public opinion monitoring solutions, the present invention has at least the following advantages: 1. More timely detection. The scheduling agent can automatically shorten the analysis chain and add key analysis tasks when negative sentiment rises rapidly or popularity surges, thereby outputting attention or early warning signals in the early stages of public opinion.

[0157] 2. More accurate judgment. The system simultaneously considers sentiment fluctuations, popularity, clusters of highly popular negative comments, and cross-platform synchronization anomalies, which can reduce false positives and false negatives caused by a single keyword, a single sentiment category, or a single popularity statistic.

[0158] 3. Results are more interpretable. The system traces the keywords, representative comments, time windows, and propagation platforms that triggered the anomaly in the agent's output, enabling managers to understand the reasons for the alerts and quickly review them.

[0159] 4. The strategy is adaptive. The feedback optimization agent uses human confirmation results to adjust thresholds, weights, and sampling frequencies, enabling the system to adapt to data distributions across different platforms, topic types, and time periods.

[0160] 5. More efficient use of computing resources. The scheduling agent decides whether to invoke high-cost analysis tasks based on triggering conditions, reducing unnecessary clustering tracing and secondary verification calculations during low-risk windows.

[0161] 6. The agent collaboration process is more specific. This scheme defines the calling relationships and data processing processes between agents through task status tables, unified interfaces, triggering rules, scoring formulas, and feedback update mechanisms, thus avoiding the problem of summarizing the system structure solely by function names.

[0162] 7. The technical solution is more complete. This solution simultaneously covers data collection, task planning, feature calculation, anomaly analysis, source tracing and interpretation, early warning output, and feedback optimization, forming a complete closed loop for network public opinion anomaly detection and early warning.

[0163] It is understood that the same or similar parts in the above embodiments can be referred to each other, and the contents not described in detail in some embodiments can be referred to the same or similar contents in other embodiments.

[0164] It should be noted that in the description of this application, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Furthermore, in the description of this application, unless otherwise stated, "a plurality of" means at least two.

[0165] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A method for detecting abnormal online public opinion, characterized in that, include: Obtain comment data for a target topic within a continuous time window. The comment data includes comment text, posting time, and popularity data, wherein the popularity data includes at least one of the following: number of likes, number of replies, and number of reposts. Calculate the growth rate of negative sentiment and the growth rate of the number of comments for each time window; When the growth rate of negative sentiment and the growth rate of the number of comments exceed their respective growth thresholds, sentiment fluctuation scores are calculated. Popularity dissemination score High score for gathering of negative comments and cross-platform synchronization of abnormal scores ; Calculate the risk score of abnormal public opinion , , , , , These are the weighting coefficients corresponding to each score; When the risk score is greater than a preset risk threshold, a source tracing and explanation task is performed to obtain evidence of the cause of the anomaly. An early warning result is generated based on the evidence of the cause of the anomaly.

2. The method according to claim 1, characterized in that: The negative sentiment growth rate is calculated as follows: in, Indicates the first The growth rate of negative sentiment within a specific time window. Indicates the first The proportion of negative emotions within a given time window. , Indicates the first The number of negative comments within a time window Indicates the first The total number of comments within a time window. Indicates the first The proportion of negative emotions within a given time window. This represents a very small constant to prevent division by zero; the growth rate of the number of comments is calculated as follows: in, This represents the growth rate of the number of comments in the t-th time window. This represents the total number of comments within the (t-1)th time window.

3. The method according to claim 1, characterized in that: The emotional fluctuation score is calculated based on the proportion of negative emotions, the growth rate of negative emotions, the magnitude of change in the mean emotional value, and the intensity of emotional fluctuations within the current time window. The popularity spread score is calculated based on the total popularity, average popularity, popularity growth rate, proportion of high-popularity comments, and proportion of high-popularity negative comments within the current time window. The cluster score of hot negative comments is calculated based on the proportion of hot negative comments, the proportion of negative sentiment, and the growth rate of negative sentiment in the current time window. The cross-platform synchronization anomaly score is calculated based on cross-platform keyword synchronization, cross-platform time synchronization, and the number of cross-platform dissemination platforms.

4. The method according to claim 3, characterized in that: The emotional fluctuation score The calculation formula is: The heat dissipation score The calculation formula is: The high-scoring negative comments cluster score The calculation formula is: Cross-platform synchronization of abnormal scores The calculation formula is: in, Represents the normalization function. Indicates the proportion of negative emotions. Indicates the growth rate of negative emotions. This indicates the magnitude of change in the mean sentiment. Indicates the intensity of emotional fluctuations. Indicates the total heat of the window. Indicates the growth rate of popularity. This indicates the percentage of highly popular negative comments. Indicates the degree of cross-platform keyword synchronization. Indicates cross-platform time synchronization. This represents the normalized value of the number of platforms used for cross-platform dissemination. to These are the weight coefficients corresponding to each sub-feature.

5. Each of the sub-features is calculated in the following manner: in, Indicates the first A collection of comments within a time window. This represents the total number of comments within the set. Indicates the first The symbolic sentiment score of each comment. and These represent the positive probability and the negative probability, respectively. This represents the mean sentiment value at time t. This represents the mean sentiment value at time t-1. in, Indicates the first The popularity value of each comment This indicates the high heat threshold determined based on a historical baseline window. Indicates the emotion category, This indicates a function that takes the value 1 if the condition is true and 0 otherwise; the aforementioned scoring formula... , and Take the corresponding window respectively , and ; in, Indicates the first The number of platforms participating in the target topic within a given time window. Indicates platform The set of keywords in this window, Indicates platform The moment when the abnormal trigger condition is first met. Indicates the time decay scale, Indicates the total number of monitoring platforms; when When less than 2, and All are set to 0; in, and Representing features respectively Within the lower and upper bounds of the historical benchmark window, the normalization result is taken as 0 when it is less than 0 and as 1 when it is greater than 1; the aforementioned scoring formula... , and Take the corresponding window respectively , and .

6. The method according to claim 1, characterized in that: The task of performing source tracing and interpretation includes: For negative comments within a time window where the risk score exceeds a preset risk threshold, keyword extraction, topic clustering, representative comment selection, and propagation node location are performed. The output includes a list of main keywords that triggered the anomaly, representative high-profile negative comments, the earliest anomaly time window, the main propagation platforms, and the risk contribution factors of each factor, serving as evidence of the cause of the anomaly.

7. The method according to claim 1, characterized in that: Before performing the source tracing and interpretation task, the following is also included: Estimate the average sentiment confidence level corresponding to the current risk score: in, , , They represent the first The probability that a comment belongs to the positive, negative, or neutral category. Indicates the first Sentiment confidence of the comments Indicates the first Average sentiment confidence within a time window Indicates the first The total number of comments within a time window; When the average sentiment confidence score is lower than the confidence threshold, a sentiment review analysis task is performed. The low-confidence comments are re-evaluated using a backup sentiment classification model, and the risk score is updated based on the re-evaluation results.

8. The method according to claim 1, characterized in that, Also includes: Determine whether the proportion of negative comments in the current time window exceeds the average of the historical baseline window; If the number of comments exceeds the threshold and the growth rate of the number of comments increases synchronously, then a sentiment review analysis task is executed to perform a second sentiment assessment on the negative comments within the current time window to confirm whether the negative sentiment has truly increased, and the number of negative comments and the growth rate of negative sentiment in the current time window are updated based on the results of the second assessment.

9. The method according to claim 1, characterized in that, Also includes: Receive manual feedback information on the warning results; The weighting coefficients α, β, γ, and δ may be adjusted based on the human feedback information, or the growth threshold may be adjusted, or the preset risk threshold may be adjusted.

10. A device for detecting abnormal online public opinion, characterized in that, include: The comment data acquisition module is used to acquire comment data of a target topic within a continuous time window. The comment data includes comment text, posting time, and popularity data. The popularity data includes at least one of the following: number of likes, number of replies, and number of reposts. The growth rate calculation module is used to calculate the negative sentiment growth rate and the comment quantity growth rate for comment data in each time window. The dimension score calculation module is used to calculate sentiment fluctuation scores when the negative sentiment growth rate and the comment quantity growth rate are respectively higher than the corresponding growth thresholds. Popularity dissemination score High score for gathering of negative comments and cross-platform synchronization of abnormal scores ; The risk scoring calculation module is used to calculate the risk score for abnormal public opinion events. , , , , , These are the weighting coefficients corresponding to each score; The abnormal evidence acquisition module is used to perform a source tracing and interpretation task and acquire evidence of the cause of the abnormality when the risk score is greater than a preset risk threshold. The early warning result generation module is used to generate early warning results based on the evidence of the cause of the anomaly.

11. A system for detecting abnormal online public opinion, characterized in that, include: At least one processor and at least one memory; The memory stores the executable instructions of the processor; The processor is configured to perform the method according to any one of claims 1-8.

Images