Intelligent identification system of public opinion means based on artificial intelligence large model

Abstract

The present invention relates to the field of intelligent identification of public opinion means, and specifically to an intelligent identification system for public opinion means based on a large artificial intelligence model, which improves the reliability and accuracy of the identification results. The public opinion data storage module in the system uses a distributed database to store original public opinion data, the knowledge base storage module uses a graph database to store the public opinion tactics knowledge system, and the result cache module uses a high-speed cache to store analysis results; the public opinion event data is obtained through the data acquisition and preprocessing module, and is cleaned and standardized. After the data is standardized, the core element information of the public opinion event is identified and extracted through the event element extraction module, and then the real development context of the event is reviewed and confirmed by the fact verification module, and then the public opinion tactics identification module is used to perform intelligent identification of public opinion means, and finally the identification result is confirmed and corrected by the human-computer collaborative verification module. The present invention is suitable for public opinion identification.

Description

Technical Field

[0001] The present invention relates to the field of intelligent recognition of public opinion means, and in particular to an intelligent recognition system for public opinion means based on an artificial intelligence large model. Background Art

[0002] Currently, traditional public opinion analysis systems primarily focus on surface-level features such as sentiment, dissemination trends, and hot topics, and have developed relatively mature analytical frameworks and technical methods. These systems typically employ techniques such as machine learning and natural language processing to automate the processing and analysis of massive amounts of public opinion data.

[0003] However, in a complex public opinion environment, relying solely on basic functions like sentiment analysis and hotspot detection is no longer sufficient to meet the needs of in-depth public opinion analysis. Traditional technologies, in particular, have significant limitations when it comes to identifying and responding to various public opinion tactics.

[0004] Existing public opinion analysis technologies have the following main technical defects:

[0005] Single Dimension of Analysis: Existing public opinion analysis systems primarily focus on surface-level features such as sentiment polarity and dissemination volume, making it difficult to deeply identify and understand complex public opinion strategies and discourse patterns. Specifically, they can only identify basic sentimental tendencies such as positive, negative, and neutral, but are unable to understand the underlying dissemination intentions and strategies within the text, lacking the ability to analyze the dissemination pathways and influence mechanisms of public opinion.

[0006] Lack of contextual understanding: Existing machine learning models perform poorly when handling tasks that require deep contextual understanding, making it difficult to accurately identify propaganda tactics such as frame-ups and exaggerations that require contextual understanding. Key issues include an inability to accurately understand contextual relationships, limited ability to understand long texts, and difficulty capturing subtle semantic nuances.

[0007] Difficulty in fact-verification: Traditional systems lack the ability to compare and verify public opinion with the facts, making them unable to effectively identify false information dissemination methods such as "making something out of nothing." This is specifically manifested in the lack of automated fact-verification mechanisms, the inability to establish a mapping between facts and public opinion, and low accuracy in authenticity judgments.

[0008] High labor costs: Existing methods for identifying public opinion tactics rely primarily on manual analysis and judgment, requiring a large number of professionals. This is inefficient and difficult to scale. Problems include: manual review is time-consuming, with the average review time for a single piece of content exceeding 5 minutes. The high cost of training professionals makes it difficult to quickly respond to sudden public opinion incidents.

[0009] Poor explainability: Traditional machine learning models often operate as black boxes when identifying propaganda tactics, making it difficult to provide clear judgment basis and explanation. This is mainly manifested in: an inability to explain the model's decision-making process, a lack of visual analytical evidence, and difficulty quantifying the credibility of the results.

[0010] Insufficient Adaptability: Existing systems struggle to quickly adapt to emerging public opinion tactics and communication strategies, lacking sufficient flexibility and scalability. Specific issues include long model update cycles, difficulty expanding the knowledge base, and poor ability to identify new public opinion tactics. Summary of the Invention

[0011] The purpose of the present invention is to overcome the shortcomings of the prior art and provide an intelligent recognition system for public opinion means based on an artificial intelligence large model, thereby improving the reliability and accuracy of the recognition results.

[0012] The present invention adopts the following technical solutions to achieve the above-mentioned purpose. The present invention provides an intelligent identification system for public opinion means based on an artificial intelligence large model, including a data layer and a service layer. The data layer includes a public opinion data storage module, a knowledge base storage module, and a result cache module. The public opinion data storage module uses a distributed database to store original public opinion data, the knowledge base storage module uses a graph database to store the knowledge system of public opinion tactics, and the result cache module uses a high-speed cache to store analysis results.

[0013] The service layer includes a data collection and preprocessing module, an event element extraction module, a fact verification module, a public opinion tactics identification module and a human-machine collaborative verification module. The data collection and preprocessing module is used to obtain public opinion event data, and then cleans and standardizes it. After the data is standardized, the event element extraction module is used to identify and extract the core element information of the public opinion event. The fact verification module is used to review and confirm the true development context of the event. The public opinion tactics identification module is then used to perform intelligent identification of public opinion means. Finally, the human-machine collaborative verification module is used to confirm and correct the identification results.

[0014] Furthermore, the system also includes an interface layer, which includes a standard service interface, a real-time push interface and an asynchronous task interface. The standard service interface is used to provide standardized service calls, the real-time push interface is used to support real-time data push, and the asynchronous task interface is used to process asynchronous tasks.

[0015] Furthermore, the event element extraction module identifies and extracts the core element information of the public opinion event, including:

[0016] The event element extraction module is used to determine the basic theme and scope of the public opinion event to be analyzed, and to extract the key elements of the public opinion event, including time, place, people, and the course of events.

[0017] Furthermore, the public opinion tactics identification module is also used to build a public opinion tactics knowledge base. Building the public opinion tactics knowledge base specifically includes:

[0018] S201. Conduct formal definition of public opinion warfare tactics;

[0019] The fueling type is formally defined as follows:

[0020] Basic characteristics: exaggeration, embellishment, or over-interpretation during the dissemination of public opinion events;

[0021] Core elements: secondary dissemination, emotional rendering, and influence expansion;

[0022] Judgment dimensions: temporal characteristics, content characteristics, and dissemination characteristics;

[0023] Quantitative indicator system:

[0024] Text deviation: similarity threshold with the original facts;

[0025] Sentiment intensity: emotional word ratio and polarity value threshold;

[0026] Dissemination impact: topic growth rate and dissemination range threshold;

[0027] Keyword features: frequency threshold of inflammatory words;

[0028] Comprehensive score: multi-dimensional weighted calculation method;

[0029] The types of frame-up are formally defined as follows:

[0030] Basic characteristics: the act of fabricating or distorting facts to attribute blame to a specific person;

[0031] Core elements: false information, causal distortion, and targeting;

[0032] Judgment dimensions: fact deviation, target orientation, and causal relationship;

[0033] Quantitative indicator system:

[0034] Factual consistency: the matching threshold with the verified facts;

[0035] Target-directedness: threshold for frequency of mention of a specific object;

[0036] Causal credibility: threshold for semantic relationship analysis;

[0037] Comprehensive score: multi-dimensional weighted calculation method;

[0038] S202. Build a knowledge base of public opinion tactics;

[0039] Construct a basic knowledge system, including a classification system for public opinion tactics, formal definition standards, quantitative indicator calculation methods, and evidence chain construction rules;

[0040] Case database construction, including typical case annotation specifications, case feature extraction methods, case classification indexing mechanism, and similar case matching rules;

[0041] Dynamic update mechanism, including extraction of new tactics features, dynamic adjustment of indicator system, adaptive optimization of thresholds, and continuous iteration of rules.

[0042] Furthermore, the intelligent identification of public opinion tactics through the public opinion tactics identification module specifically includes:

[0043] S301, scene classification mechanism;

[0044] Scene classification includes scene feature quantification, which includes the following indicators:

[0045] Reach indicators:

[0046] Platform coverage: the number of platforms involved or the total number of monitored platforms;

[0047] Geographical distribution: the number of regions involved or the total number of monitored regions;

[0048] User engagement: number of interactive users or total number of exposed users;

[0049] Feature significance index:

[0050] Keyword matching degree: the number of hit keywords or the total number of keywords;

[0051] Topic relevance: cosine similarity with the core topic;

[0052] Sentiment polarity: standard deviation of sentiment values;

[0053] Impact indicators:

[0054] Propagation speed: the growth rate per unit time;

[0055] KOL participation: the weighted sum of the influence of participating KOLs;

[0056] Interaction intensity: the time-decayed weighted sum of comments and reposts;

[0057] Scene judgment rules:

[0058] Common scenarios:

[0059] Platform coverage < 0.3;

[0060] Feature significance > 0.7;

[0061] Influence index <0.4;

[0062] Processing strategy: routine monitoring, response within 24 hours;

[0063] Complex scenarios:

[0064] 0.3≤ platform coverage<0.6;

[0065] 0.4≤feature significance≤0.7;

[0066] 0.4≤Influence Index<0.7;

[0067] Handling strategy: Focus on it and respond within 12 hours;

[0068] Emergency scenarios:

[0069] Platform coverage ≥ 0.6;

[0070] Feature significance < 0.4;

[0071] Influence index ≥ 0.7;

[0072] Treatment strategy: Immediate treatment, response within 2 hours;

[0073] Dynamic adjustment mechanism:

[0074] Real-time update: indicator values ​​are updated every 5 minutes;

[0075] Threshold optimization: adjust the threshold once a week based on historical data;

[0076] Manual intervention: supports manual adjustment of grading results;

[0077] Emergency response: Rapid escalation mechanism under set circumstances;

[0078] S302, identification specification definition;

[0079] Identification specifications include:

[0080] Input specifications, including basic information about the event, verified facts, and public opinion to be analyzed;

[0081] Analysis requirements, including feature matching analysis, fact comparison and verification, impact assessment, and comprehensive research and judgment;

[0082] Output specifications, including identification results, evidence description, and disposal recommendations;

[0083] The identification results include the type and confidence level of the public opinion tactics; the evidence description includes characteristics and factual deviations; and the handling recommendations include impact assessment and response plans;

[0084] Feature matching analysis includes:

[0085] Public opinion tactics feature matching:

[0086] Matching with the promotion type, including calculating text deviation, evaluating sentiment intensity, analyzing communication impact, and counting keyword features;

[0087] Matching the frame-up type, including assessing factual consistency, calculating target orientation, analyzing causal credibility, and calculating a comprehensive score;

[0088] Formal definition matching:

[0089] Basic feature comparison, including calculating text semantic similarity, calculating core element coverage, and determining dimension conformity;

[0090] Quantitative indicator matching, including indicator value calculation, threshold judgment, weight allocation, and comprehensive score calculation;

[0091] Fact comparison verification includes:

[0092] Basic fact verification, which includes time dimension, space dimension, and subject dimension;

[0093] The time dimension includes:

[0094] Accuracy of event time, completeness of time series, temporal relationship of key nodes, and rationality of time span;

[0095] The spatial dimensions include:

[0096] Accuracy of location information, rationality of spatial scope, verification of geographical relevance, and authenticity of scene description;

[0097] The subject dimensions include:

[0098] Verify identity of individuals, verify organizational structure, sort out relationship networks, and confirm role positioning;

[0099] Content consistency verification, including factual element comparison and logical relationship verification;

[0100] The comparison of factual elements includes the matching degree of core information, accuracy of detailed description, reliability of data citation, and source credibility assessment;

[0101] The verification of logical relationships includes the rationality of causal relationships, the integrity of the reasoning process, the logical rigor of the argument, and the reliability evaluation of the conclusions;

[0102] Impact assessment includes: communication dimension assessment, public opinion impact assessment, and social impact assessment;

[0103] The evaluation of communication dimensions includes: platform influence and user engagement;

[0104] Platform influence includes mainstream media coverage, social media activity, self-media communication power, and comprehensive influence score calculation;

[0105] User engagement includes interaction level assessment, user portrait analysis, engagement depth analysis, and KOL influence assessment;

[0106] Public opinion impact assessment includes: sentiment tendency analysis, issue development assessment;

[0107] Sentiment tendency analysis includes sentiment polarity distribution, sentiment intensity changes, group sentiment evolution, and polarization degree assessment;

[0108] Topic development assessment includes topic popularity trends, topic diffusion scope, related topic linkage, and development trend prediction;

[0109] Social impact assessment includes: cognitive impact, behavioral impact;

[0110] Cognitive influence includes the degree of public cognitive bias, the trend of group attitude change, the degree of value conflict, and the influence of social consensus;

[0111] Behavioral impacts include online behavior changes, offline action conversions, group behavior trends, and social governance impacts;

[0112] Comprehensive assessment includes: multi-dimensional scoring and risk warning;

[0113] Multi-dimensional scoring includes calculation of feature matching score, fact verification score, impact assessment score, comprehensive score calculation and grading;

[0114] Risk early warning includes risk level determination, development trend prediction, intervention suggestion generation, and effect evaluation mechanism;

[0115] S303, result processing mechanism;

[0116] The result processing mechanism includes confidence assessment and processing strategy;

[0117] The confidence assessment includes feature matching degree assessment, evidence integrity assessment, impact assessment, and comprehensive confidence calculation;

[0118] The feature matching degree evaluation includes semantic feature matching degree, emotional feature matching degree, and factual feature matching degree;

[0119] The semantic feature matching degree includes cosine similarity with the standard pattern, key concept coverage, and semantic coherence score;

[0120] The emotional feature matching degree includes emotional polarity consistency and emotional intensity similarity;

[0121] The fact feature matching degree includes entity alignment rate and event element matching degree;

[0122] The evidence integrity assessment includes the integrity of the evidence chain and the quality of evidence;

[0123] The integrity of the evidence chain includes key node coverage, logical link integrity, and temporal coherence;

[0124] The quality of evidence includes source reliability score, freshness of evidence, and degree of mutual confirmation of evidence;

[0125] The impact assessment includes communication influence and timeliness impact;

[0126] The communication influence includes platform weight score, user influence, and communication scope;

[0127] The timeliness impact includes real-time popularity and development trends;

[0128] The comprehensive confidence calculation includes a weighted average calculation, which is calculated as follows: Confidence = 0.4 * feature matching + 0.35 * evidence integrity + 0.25 * impact level, where Confidence represents confidence, and the confidence levels are graded as follows: Confidence greater than or equal to 0.8 is high confidence, Confidence between 0.6 and 0.8 is medium confidence, and Confidence less than 0.6 is low confidence;

[0129] The processing strategy specifically includes: directly accepting high confidence results, manually reviewing medium confidence results, and reanalyzing low confidence results.

[0130] S304, optimization feedback;

[0131] The optimization feedback includes result accumulation and continuous optimization;

[0132] The accumulation of results includes the storage of typical cases, updating of feature database, and summary of handling experience;

[0133] The continuous optimization includes adjustment of grading standards, improvement of identification specifications, and optimization of processing strategies.

[0134] Furthermore, the recognition results are confirmed and corrected by the human-machine collaborative verification module, specifically including:

[0135] The review process includes a three-level review mechanism: preliminary review, re-review, and final review. The average review time for a single piece of content does not exceed 2 minutes. It supports real-time updating of prompt word templates, regular weekly updates to the knowledge base, and an emergency update mechanism. The preliminary review focuses on basic feature matching, the re-review emphasizes the integrity of the evidence chain, and the final review ensures the reliability of the judgment results;

[0136] Feedback mechanism, which includes real-time feedback of audit results, continuous optimization and updating of models, dynamic expansion of knowledge base, and continuous improvement of system performance.

[0137] The beneficial effects of the present invention are:

[0138] This invention greatly improves the recognition accuracy through the formalized definition of public opinion tactics and quantitative indicator system. With the help of a multi-dimensional fact comparison mechanism, it reduces the error rate for reliably identifying public opinion content that does not conform to the actual situation. Relying on the human-computer collaborative verification mechanism, it ensures the reliability of the recognition results and improves the overall accuracy.

[0139] The present invention reduces the analysis time of a single piece of content through a standardized public opinion tactics identification process, supports the processing of more than multiple concurrent requests per second through an optimized resource scheduling mechanism, and improves analysis efficiency through a scenario-based template reuse mechanism.

[0140] The present invention improves the traceability of the analysis process through a complete evidence chain construction mechanism, facilitates professional understanding and verification through a standardized result display scheme, and improves the credibility of the system through a visual display of the analysis process.

[0141] The present invention supports rapid adaptation to new public opinion tactics through a formalized knowledge base construction method, a template dynamic optimization mechanism to achieve flexible strategy adjustment, and a layered system architecture design to support horizontal expansion and rapid capacity expansion.

[0142] This invention reduces manual review costs through standardized processes, reduces system operation and maintenance costs through optimized resource scheduling, and improves the efficiency of public opinion handling through efficient identification mechanisms. BRIEF DESCRIPTION OF THE DRAWINGS

[0143] Figure 1 This is a diagram showing the overall architecture of an intelligent recognition system for public opinion means based on an artificial intelligence large model provided by an embodiment of the present invention;

[0144] Figure 2 This is a flowchart of public opinion tactics identification provided by an embodiment of the present invention;

[0145] Figure 3 This is a flowchart of scene classification recognition provided by an embodiment of the present invention;

[0146] Figure 4 This is a structural diagram of a prompt word template provided by an embodiment of the present invention;

[0147] Figure 5 This is a flowchart of human-machine collaborative verification provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0148] To make the objectives, technical solutions and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0149] The present invention provides an intelligent recognition system for public opinion means based on an artificial intelligence large model, such as Figure 1 As shown, including:

[0150] The data layer includes the public opinion data storage module, the knowledge base storage module and the result cache module. The public opinion data storage module uses a distributed database to store the original public opinion data, the knowledge base storage module uses a graph database to store the public opinion tactics knowledge system, and the result cache module uses a high-speed cache to store the analysis results.

[0151] The service layer includes a data collection and preprocessing module, an event element extraction module, a fact verification module, a public opinion tactics identification module, and a human-machine collaborative verification module. The data collection and preprocessing module is used to acquire, clean, and standardize public opinion event data. The event element extraction module is used to identify and extract the core element information of public opinion events. The fact verification module is used to confirm the true development context of the event through manual review. The public opinion tactics identification module is used for intelligent identification of public opinion means based on large models. The human-machine collaborative verification module is used to support professionals in confirming and correcting the identification results.

[0152] Interface layer, the interface layer includes standard service interface, real-time push interface and asynchronous task interface. The standard service interface is used to provide standardized service calls, the real-time push interface is used to support real-time data push, and the asynchronous task interface is used to process asynchronous tasks.

[0153] The specific implementation process of the present invention includes the following steps:

[0154] Step 1: Initialize public opinion events;

[0155] Determine the basic theme and scope of the incident to be analyzed;

[0156] Extract key elements of the event, including but not limited to time, place, people, and course of events;

[0157] Through system extraction and manual review, the true situation and development context of the incident are determined.

[0158] Step 2: Build a knowledge base of public opinion tactics;

[0159] S201, Formal definition of public opinion warfare tactics;

[0160] The fueling type is formally defined as follows:

[0161] Basic characteristics: exaggeration, embellishment, or over-interpretation during the dissemination of public opinion events;

[0162] Core elements: secondary dissemination, emotional rendering, and influence expansion;

[0163] Judgment dimensions: temporal characteristics, content characteristics, and dissemination characteristics;

[0164] Quantitative indicator system:

[0165] Text deviation: similarity threshold with the original facts;

[0166] Sentiment intensity: emotional word ratio and polarity value threshold;

[0167] Dissemination impact: topic growth rate and dissemination range threshold;

[0168] Keyword features: frequency threshold of inflammatory words;

[0169] Comprehensive score: multi-dimensional weighted calculation method;

[0170] The types of frame-up are formally defined as follows:

[0171] Basic characteristics: the act of fabricating or distorting facts to attribute blame to a specific person;

[0172] Core elements: false information, causal distortion, and targeting;

[0173] Judgment dimensions: fact deviation, target orientation, and causal relationship;

[0174] Quantitative indicator system:

[0175] Factual consistency: the matching threshold with the verified facts;

[0176] Target-directedness: threshold for frequency of mention of a specific object;

[0177] Causal credibility: threshold for semantic relationship analysis;

[0178] Comprehensive score: multi-dimensional weighted calculation method;

[0179] S202. Build a knowledge base of public opinion tactics;

[0180] Construct a basic knowledge system, including a classification system for public opinion tactics, formal definition standards, quantitative indicator calculation methods, and evidence chain construction rules;

[0181] Case database construction, including typical case annotation specifications, case feature extraction methods, case classification indexing mechanism, and similar case matching rules;

[0182] Dynamic update mechanism, including extraction of new tactics features, dynamic adjustment of indicator system, adaptive optimization of thresholds, and continuous iteration of rules.

[0183] Step 3: Intelligent identification of public opinion tactics;

[0184] Intelligent identification of public opinion tactics Figure 2 As shown, the following steps are included:

[0185] S301, scene classification mechanism;

[0186] Scene classification includes scene feature quantification, scene judgment rules, and dynamic adjustment mechanism;

[0187] The quantification of scene features includes the following indicators;

[0188] Reach indicators:

[0189] Platform coverage: the number of platforms involved or the total number of monitored platforms;

[0190] Geographical distribution: the number of regions involved or the total number of monitored regions;

[0191] User engagement: number of interactive users or total number of exposed users;

[0192] Feature significance index:

[0193] Keyword matching degree: the number of hit keywords or the total number of keywords;

[0194] Topic relevance: cosine similarity with the core topic;

[0195] Sentiment polarity: standard deviation of sentiment values;

[0196] Impact indicators:

[0197] Propagation speed: the growth rate per unit time;

[0198] KOL participation: the weighted sum of the influence of participating KOLs;

[0199] Interaction intensity: the time-decayed weighted sum of comments and reposts;

[0200] Scene judgment rules, such as Figure 3 As shown:

[0201] Conventional scenario (low risk):

[0202] Platform coverage < 0.3;

[0203] Feature significance > 0.7;

[0204] Influence index <0.4;

[0205] Processing strategy: routine monitoring, response within 24 hours;

[0206] Complex scenarios (medium risk):

[0207] 0.3≤ platform coverage<0.6;

[0208] 0.4≤feature significance≤0.7;

[0209] 0.4≤Influence Index<0.7;

[0210] Handling strategy: Focus on it and respond within 12 hours;

[0211] Emergency scenario (high risk):

[0212] Platform coverage ≥ 0.6;

[0213] Feature significance < 0.4;

[0214] Influence index ≥ 0.7;

[0215] Treatment strategy: Immediate treatment, response within 2 hours;

[0216] Dynamic adjustment mechanism:

[0217] Real-time update: indicator values ​​are updated every 5 minutes;

[0218] Threshold optimization: adjust the threshold once a week based on historical data;

[0219] Manual intervention: supports manual adjustment of grading results;

[0220] Emergency response: Rapid escalation mechanism under set circumstances;

[0221] S302, identification specification definition;

[0222] Identification specifications include:

[0223] Input specifications, including basic information about the event, verified facts, and public opinion to be analyzed;

[0224] Analysis requirements, including feature matching analysis, fact comparison and verification, impact assessment, and comprehensive research and judgment;

[0225] Output specifications, including identification results, evidence description, and disposal recommendations;

[0226] The identification results include the type and confidence level of the public opinion tactics; the evidence description includes characteristics and factual deviations; and the handling recommendations include impact assessment and response plans;

[0227] Feature matching analysis includes:

[0228] Public opinion tactics feature matching:

[0229] Matching the type of incitement, including calculating text deviation (similarity to the original facts), evaluating sentiment intensity (percentage of sentimental words and polarity), analyzing dissemination impact (topic growth rate and scope), and statistically analyzing keyword features (frequency of inflammatory words);

[0230] Matching the frame-up type includes assessing factual consistency (matching with verified facts), calculating target orientation (frequency of mention of specific objects), analyzing causal credibility (semantic relationship analysis), and calculating a comprehensive score (multi-dimensional weighting);

[0231] Formal definition matching:

[0232] Basic feature comparison, including calculating text semantic similarity (using the BERT model), calculating core element coverage (greater than or equal to 85% is a valid match), and determining dimension conformity (greater than or equal to 80% is a valid match);

[0233] Quantitative indicator matching, including indicator value calculation (based on the specific type of public opinion warfare tactics), threshold judgment (comparison with preset standards), weight allocation (importance of different dimensions), and comprehensive score calculation (weighted average method);

[0234] Fact comparison verification includes:

[0235] Basic fact verification, which includes time dimension, space dimension, and subject dimension;

[0236] The time dimension includes:

[0237] Accuracy of event time, completeness of time series, temporal relationship of key nodes, and rationality of time span;

[0238] The spatial dimensions include:

[0239] Accuracy of location information, rationality of spatial scope, verification of geographical relevance, and authenticity of scene description;

[0240] The subject dimensions include:

[0241] Verify identity of individuals, verify organizational structure, sort out relationship networks, and confirm role positioning;

[0242] Content consistency verification, including factual element comparison and logical relationship verification;

[0243] The factual element comparison includes core information matching (greater than or equal to 90%), accuracy of detailed description (greater than or equal to 85%), reliability of data citation (greater than or equal to 95%), and source credibility assessment (greater than or equal to 80%);

[0244] Logical relationship verification includes the rationality of causal relationships, the integrity of the reasoning process, the logical rigor of the argument, and the reliability of the conclusions;

[0245] Impact assessment includes: communication dimension assessment, public opinion impact assessment, and social impact assessment;

[0246] The evaluation of communication dimensions includes: platform influence and user engagement;

[0247] Platform influence includes mainstream media coverage (weight: 0.4), social media activity (weight: 0.3), self-media communication (weight: 0.3), and comprehensive influence score calculation;

[0248] User engagement includes interaction level assessment, user portrait analysis, engagement depth analysis, and KOL influence assessment;

[0249] Public opinion impact assessment includes: sentiment tendency analysis, issue development assessment;

[0250] Sentiment tendency analysis includes sentiment polarity distribution (positive, negative, neutral), sentiment intensity changes (time series trend), group sentiment evolution (propagation path), and polarization degree assessment (opinion distribution);

[0251] Topic development assessment includes topic popularity trends (growth rate, sustainability), topic diffusion scope (region, population), related topic linkage (topic cluster analysis), and development trend prediction (trend model);

[0252] Social impact assessment includes: cognitive impact, behavioral impact;

[0253] Cognitive influence includes the degree of public cognitive bias, the trend of group attitude change, the degree of value conflict, and the influence of social consensus;

[0254] Behavioral impacts include online behavioral changes (interaction methods, participation), offline action conversion (field activities, social mobilization), group behavior trends (bandwagon effect, polarization phenomenon), and social governance impacts (policy response, institutional adjustment);

[0255] Comprehensive assessment includes: multi-dimensional scoring and risk warning;

[0256] The multi-dimensional scoring includes calculation of feature matching score (weight: 0.4), fact verification score (weight: 0.35), impact assessment score (weight: 0.25), comprehensive score calculation and grading;

[0257] Risk early warning includes risk level determination (high, medium, low), development trend prediction (24 hours, 3 days, 7 days), intervention suggestion generation (based on historical cases), and effect evaluation mechanism (post-intervention feedback);

[0258] S303, result processing mechanism;

[0259] The result processing mechanism includes confidence assessment and processing strategy;

[0260] Confidence assessment includes feature matching degree assessment (weight: 0.4), evidence integrity assessment (weight: 0.35), impact assessment (weight: 0.25), and comprehensive confidence calculation;

[0261] The feature matching degree evaluation includes semantic feature matching, emotional feature matching, and factual feature matching;

[0262] Semantic feature matching includes cosine similarity with the standard pattern (threshold > 0.8), key concept coverage (threshold > 0.7), and semantic coherence score (threshold > 0.75);

[0263] Emotional feature matching includes emotion polarity consistency (threshold > 0.8) and emotion intensity similarity (threshold > 0.7);

[0264] Fact feature matching includes entity alignment rate (threshold > 0.9) and event element matching (threshold > 0.85);

[0265] The evidence integrity assessment includes the completeness of the evidence chain and the quality of evidence;

[0266] The integrity of the chain of evidence includes key node coverage (required to be >90%), logical link integrity (required to be >85%), and temporal coherence (required to be >95%);

[0267] The quality of evidence includes the source reliability score (maximum score 100, requirement >80), freshness of evidence (within 7 days: 1.0, within 14 days: 0.8, within 30 days: 0.6), and the degree of mutual confirmation of evidence (required to have >3 independent sources);

[0268] Impact assessment includes communication influence and timeliness impact;

[0269] Communication influence includes platform weight score (mainstream media: 1.0, social media: 0.8, other: 0.6), user influence (KOL: 1.0, ordinary user: 0.6), and communication scope (national: 1.0, regional: 0.7, local: 0.4);

[0270] Timeliness impact includes real-time popularity (within 24 hours: 1.0, within 3 days: 0.8, within 7 days: 0.6) and development trend (increasing: 1.0, stable: 0.7, decreasing: 0.4);

[0271] The comprehensive confidence calculation includes weighted average calculation, which is calculated as follows: Confidence = 0.4 * feature matching + 0.35 * evidence completeness + 0.25 * impact level. Confidence represents the confidence level, and the confidence level is graded as follows: Confidence greater than or equal to 0.8 is high confidence, Confidence between 0.6 and 0.8 is medium confidence, and Confidence less than 0.6 is low confidence.

[0272] The processing strategy specifically includes: directly accepting high confidence results, manually reviewing medium confidence results, and reanalyzing low confidence results.

[0273] S304, optimization feedback;

[0274] Optimization feedback includes result accumulation and continuous optimization;

[0275] The accumulation of results includes the storage of typical cases, updating of feature database, and summary of handling experience; continuous optimization includes adjustment of classification standards, improvement of identification specifications, and optimization of processing strategies.

[0276] Step 4: Human-machine collaborative verification;

[0277] Human-machine collaborative verification includes:

[0278] (1) Review process:

[0279] A three-level review mechanism: preliminary review -> re-review -> final review;

[0280] The average review time for a single piece of content does not exceed 2 minutes;

[0281] Support real-time update of prompt word templates;

[0282] The knowledge base is updated regularly every week and supports emergency update mechanism.

[0283] (2) Audit standards:

[0284] Initial review: focus on basic feature matching;

[0285] Review: Focus on the integrity of the chain of evidence;

[0286] Final review: ensure the reliability of the judgment results.

[0287] (3) Feedback mechanism:

[0288] Real-time feedback on audit results;

[0289] The model is continuously optimized and updated;

[0290] Dynamic expansion of knowledge base;

[0291] System performance continues to improve.

[0292] The key technical points of the present invention include:

[0293] 1. Prompt word project:

[0294] Prompt word template such as Figure 4 As shown:

[0295] (1) Formal description framework:

[0296] a structured representation of the characteristics of public opinion warfare tactics;

[0297] Quantitative indicator system for judging standards;

[0298] normative definition of evidentiary requirements;

[0299] Standardized description of analytical procedures;

[0300] (2) Scenario-based template design:

[0301] Special templates for different types of public opinion warfare tactics;

[0302] Case-based template verification mechanism;

[0303] Formwork performance evaluation system;

[0304] Dynamic optimization feedback mechanism;

[0305] (3) Evidence chain construction standards:

[0306] Key feature extraction methods;

[0307] Evidence integrity requirements;

[0308] Credibility assessment criteria;

[0309] Result traceability mechanism.

[0310] 2. Fact comparison mechanism:

[0311] (1) Text similarity calculation:

[0312] a. Semantic vector model calculation:

[0313] Text encoding:

[0314] Use the BERT model to encode text into a 768-dimensional vector;

[0315] Use hierarchical pooling to obtain sentence representation;

[0316] Use attention mechanism to weight key information;

[0317] Output normalization processing;

[0318] Similarity calculation:

[0319] Cosine similarity: cos(θ) = (A·B) / (||A||·||B||);

[0320] Euclidean distance: d = sqrt(Σ(ai-bi)2);

[0321] Manhattan distance: d = Σ|ai-bi|;

[0322] Weighted combination score: Score = w1*cos+w2*(1-deuc)+w3*(1-dman); b. Edit distance algorithm:

[0323] Levenshtein distance calculation:

[0324] Insert operation cost: 1.0;

[0325] Delete operation cost: 1.0;

[0326] Replace operation cost: 2.0;

[0327] Normalization: score = 1-dist / max(len1,len2);

[0328] Longest Common Subsequence (LCS):

[0329] Dynamic programming to calculate LCS length;

[0330] Retrospectively construct specific sequences;

[0331] Similarity score: score = LCS_length / max(len1, len2);

[0332] c. Key information extraction:

[0333] Named Entity Recognition (NER):

[0334] Identify entities such as time, place, person, and organization;

[0335] Calculate entity matching rate: match_rate = matched_entities / total_entities; Entity importance weighting: weight = entity_type_weight * entity_frequency; Relationship extraction:

[0336] Extract subject-verb-object triples;

[0337] Calculate relationship matching;

[0338] Construct an event graph;

[0339] Calculate graph similarity;

[0340] d. Comprehensive similarity calculation:

[0341] Feature fusion:

[0342] Semantic similarity weight: 0.4;

[0343] Edit distance weight: 0.3;

[0344] Entity matching weight: 0.2;

[0345] Relationship matching weight: 0.1;

[0346] Threshold determination:

[0347] Highly similar: score ≥ 0.8;

[0348] Partial similarity: 0.6≤score<0.8;

[0349] Low similarity: score < 0.6;

[0350] (2) Difference detection algorithm:

[0351] a. Rule-based explicit difference detection;

[0352] Differences in text structure:

[0353] Paragraph structure comparison: calculate the number, length, and position changes of paragraphs;

[0354] Comparison of syntactic structures: analysis of sentence components and changes in grammatical relations;

[0355] Punctuation comparison: identifying changes in punctuation usage, placement, and frequency;

[0356] Format style comparison: detect font, typesetting, and layout changes;

[0357] Differences in content elements:

[0358] Keyword differences: additions, deletions, and modifications to core vocabulary;

[0359] Differences in numerical information: changes in numbers, dates, and amounts;

[0360] Entity information differences: changes in names of people, places, and organizations;

[0361] Differences in event descriptions: changes in event process and details;

[0362] b. Semantic-based implicit difference detection;

[0363] Semantic deviation analysis:

[0364] Word meaning level: synonym replacement, synonym changes;

[0365] Phrasal level: changes in expression and tone;

[0366] Sentence level: changes in contextual meaning and logical relationships;

[0367] Paragraph level: shift in topic focus and change in argument angle;

[0368] Differences in emotional tendencies:

[0369] Emotional polarity changes: positive / negative / neutral shifts;

[0370] Changes in emotional intensity: changes in the use of degree words;

[0371] Subjective differences: objective statements vs. subjective evaluations;

[0372] Differences in stance and attitude: support / opposition / neutral changes;

[0373] c. Multi-dimensional comparative analysis;

[0374] The time dimension comparison includes consistency of temporal relationships, changes in time span, accuracy of time points, and temporal logic coherence; the location dimension comparison includes accuracy of geographic location, changes in spatial range, consistency of geographic relationships, and completeness of scene descriptions; the character dimension comparison includes accuracy of role identity, changes in character relationships, consistency of behavior descriptions, and accuracy of speech citations; the event dimension comparison includes changes in causal relationships, completeness of process descriptions, differences in outcome impacts, and changes in background information;

[0375] d. Discrepancy assessment and handling;

[0376] Quantification of the degree of difference:

[0377] Explicit difference score: based on rule detection results;

[0378] Implicit difference score: based on semantic analysis results;

[0379] Comprehensive difference score: weighted average calculation;

[0380] Difference classification:

[0381] Significant Difference: A change that affects the nature of an event;

[0382] General Differences: Changes in detail or expression;

[0383] Negligible differences: minor changes that do not affect understanding;

[0384] Treatment strategy:

[0385] Major discrepancies: manual review is required;

[0386] General differences: automatically marked by the system;

[0387] Ignore the difference: pass directly;

[0388] 3. Interpretability Design Techniques

[0389] Design for explainability includes:

[0390] Visualization of the analysis process: Using a decision tree to display the analysis process; Evidence chain construction: Combining key information highlighting and reasoning path display; a. Basic structure of the evidence chain:

[0391] Core facts layer: basic facts of the original event;

[0392] Communication layer: information dissemination process and changes;

[0393] Influence layer: social influence and public opinion effect;

[0394] Verification layer: fact verification and supporting materials;

[0395] b. Key information extraction:

[0396] Event elements:

[0397] Time information: time of occurrence, duration, key time nodes; Location information: location of occurrence, scope of impact, geographical association;

[0398] Character information: parties involved, related parties, witnesses;

[0399] Organizational information: Involved units, regulatory authorities, and media organizations;

[0400] Event process:

[0401] Cause: triggering events, background conditions;

[0402] Pass: key behaviors, important nodes;

[0403] Results: direct impact, indirect impact;

[0404] Development: subsequent progress and derivative events;

[0405] Related information:

[0406] Related events: historical cases, similar events;

[0407] Background information: policies, regulations, and industry standards;

[0408] Expert opinions: authoritative interpretation and professional analysis;

[0409] Public reaction: public opinion and mass opinions;

[0410] c. Evidence Grading and Assessment:

[0411] Level of evidence:

[0412] Level 1 evidence: official documents, direct evidence (weight: 0.5); Level 2 evidence: authoritative media reports, expert analysis (weight: 0.3); Level 3 evidence: indirect evidence, public feedback (weight: 0.2); Evidence quality assessment:

[0413] Authenticity: reliability of source and authenticity of content;

[0414] Timeliness: freshness of information and timely updates;

[0415] Relevance: degree of relevance to the core facts;

[0416] Completeness: the degree of completeness of the information;

[0417] d. Example of building a chain of evidence:

[0418] Take a product quality incident as an example:

[0419] 1. Core facts layer:

[0420] Level 1 evidence:

[0421] Quality inspection report from the market supervision department;

[0422] Official company statement;

[0423] On-site inspection records;

[0424] Evidence Link:

[0425] Time: The problem was discovered on X / X / 2023;

[0426] Location: Specific production base and sales area;

[0427] Involved entities: manufacturers and distributors;

[0428] Problem description: Specific quality problem and scope of impact;

[0429] 2. Communication layer:

[0430] Secondary evidence:

[0431] investigative reporting in mainstream media;

[0432] Analytical articles by industry experts;

[0433] Survey reports from consumer associations;

[0434] Information dissemination path:

[0435] First release on social media → Media follow-up reports → Official response;

[0436] Propagation timeline and nodes;

[0437] the evolution of important ideas;

[0438] 3. Impact layer:

[0439] Level 3 evidence includes consumer complaints and feedback, social media discussions and comments, and changes in market sales data;

[0440] Impact assessment includes brand reputation impact, market share changes, and consumer confidence index;

[0441] 4. Validation layer:

[0442] Cross-validation includes comparison of multi-source information, timeline consistency checks, and mutual verification of key details;

[0443] Credibility assessment includes completeness of evidence: 90%, source reliability: 85%, logical consistency: 95%;

[0444] Result scoring mechanism: including credibility scoring and evidence sufficiency assessment;

[0445] Audit feedback records: support problem annotation and feedback.

[0446] The foregoing description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the form disclosed herein and should not be construed as excluding other embodiments. Rather, the present invention can be used in various other combinations, modifications, and environments and can be modified within the scope of the concept described herein through the above teachings or techniques or knowledge in the relevant field. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention are intended to be protected by the appended claims.

Claims

1. The intelligent recognition system of public opinion means based on the artificial intelligence large model is characterized by: It includes a data layer and a service layer. The data layer includes a public opinion data storage module, a knowledge base storage module, and a result cache module. The public opinion data storage module uses a distributed database to store original public opinion data, the knowledge base storage module uses a graph database to store the public opinion tactics knowledge system, and the result cache module uses a high-speed cache to store analysis results. The service layer includes a data collection and preprocessing module, an event element extraction module, a fact verification module, a public opinion tactics identification module, and a human-machine collaborative verification module. The data collection and preprocessing module obtains public opinion event data, cleans and standardizes it, and after standardizing the data, the event element extraction module identifies and extracts the core element information of the public opinion event. The fact verification module then reviews and confirms the true development context of the event. The public opinion tactics identification module then performs intelligent identification of public opinion means. Finally, the human-machine collaborative verification module confirms and corrects the identification results. The public opinion tactics identification module is also used to build a public opinion tactics knowledge base. Building a public opinion tactics knowledge base specifically includes: S201. Conduct formal definition of public opinion warfare tactics; The fueling type is formally defined as follows: Basic characteristics: exaggeration, embellishment, or over-interpretation during the dissemination of public opinion events; Core elements: secondary dissemination, emotional rendering, and influence expansion; Judgment dimensions: temporal characteristics, content characteristics, and dissemination characteristics; Quantitative indicator system: Text deviation: similarity threshold with the original facts; Sentiment intensity: emotional word ratio and polarity value threshold; Dissemination impact: topic growth rate and dissemination range threshold; Keyword features: frequency threshold of inflammatory words; Comprehensive score: multi-dimensional weighted calculation method; The types of frame-up are formally defined as follows: Basic characteristics: the act of fabricating or distorting facts to attribute blame to a specific person; Core elements: false information, causal distortion, and targeting; Judgment dimensions: fact deviation, target orientation, and causal relationship; Quantitative indicator system: Factual consistency: the matching threshold with the verified facts; Target-directedness: threshold for frequency of mention of a specific object; Causal credibility: threshold for semantic relationship analysis; Comprehensive score: multi-dimensional weighted calculation method; S202. Build a knowledge base of public opinion tactics; Construct a basic knowledge system, including a classification system for public opinion tactics, formal definition standards, quantitative indicator calculation methods, and evidence chain construction rules; Case database construction, including typical case annotation specifications, case feature extraction methods, case classification indexing mechanism, and similar case matching rules; Dynamic update mechanism, including extraction of new tactics features, dynamic adjustment of indicator system, adaptive optimization of thresholds, and continuous iteration of rules.

2. The intelligent recognition system for public opinion means based on artificial intelligence large model according to claim 1 is characterized in that: The system also includes an interface layer, which includes a standard service interface, a real-time push interface and an asynchronous task interface. The standard service interface is used to provide standardized service calls, the real-time push interface is used to support real-time data push, and the asynchronous task interface is used to process asynchronous tasks.

3. The intelligent recognition system for public opinion means based on artificial intelligence large model according to claim 1 is characterized in that: The event element extraction module identifies and extracts the core element information of public opinion events, including: The event element extraction module is used to determine the basic theme and scope of the public opinion event to be analyzed, and to extract the key elements of the public opinion event, including time, place, people, and the course of events.

4. The intelligent recognition system for public opinion means based on artificial intelligence large model according to claim 1 is characterized in that: Intelligent identification of public opinion tactics through the public opinion tactics identification module specifically includes: S301, scene classification mechanism; Scene classification includes scene feature quantification, which includes the following indicators: Reach indicators: Platform coverage: the number of platforms involved or the total number of monitored platforms; Geographical distribution: the number of regions involved or the total number of monitored regions; User engagement: number of interactive users or total number of exposed users; Feature significance index: Keyword matching degree: the number of hit keywords or the total number of keywords; Topic relevance: cosine similarity with the core topic; Sentiment polarity: standard deviation of sentiment values; Impact indicators: Propagation speed: the growth rate per unit time; KOL participation: the weighted sum of the influence of participating KOLs; Interaction intensity: the time-decayed weighted sum of comments and reposts; Scene judgment rules: Common scenarios: Platform coverage < 0.3; Feature significance > 0.7; Influence index < 0.4; Processing strategy: routine monitoring, response within 24 hours; Complex scenarios: 0.3 ≤ platform coverage < 0.6; 0.4 ≤ Feature significance ≤ 0.7; 0.4 ≤ influence index < 0.7; Handling strategy: Focus on it and respond within 12 hours; Emergency scenarios: Platform coverage ≥ 0.6; Feature significance < 0.4; Influence index ≥ 0.7; Treatment strategy: Immediate treatment, response within 2 hours; Dynamic adjustment mechanism: Real-time update: indicator values ​​are updated every 5 minutes; Threshold optimization: adjust the threshold once a week based on historical data; Manual intervention: supports manual adjustment of grading results; Emergency response: Rapid escalation mechanism under set circumstances; S302, identification specification definition; Identification specifications include: Input specifications, including basic information about the event, verified facts, and public opinion to be analyzed; Analysis requirements, including feature matching analysis, fact comparison and verification, impact assessment, and comprehensive research and judgment; Output specifications, including identification results, evidence description, and disposal recommendations; The identification results include the type and confidence level of the public opinion tactics; the evidence description includes characteristics and factual deviations; and the handling recommendations include impact assessment and response plans; Feature matching analysis includes: Public opinion tactics feature matching: Matching with the promotion type, including calculating text deviation, evaluating sentiment intensity, analyzing communication impact, and counting keyword features; Matching the frame-up type, including assessing factual consistency, calculating target orientation, analyzing causal credibility, and calculating a comprehensive score; Formal definition matching: Basic feature comparison, including calculating text semantic similarity, calculating core element coverage, and determining dimension conformity; Quantitative indicator matching, including indicator value calculation, threshold judgment, weight allocation, and comprehensive score calculation; Fact comparison verification includes: Basic fact verification, which includes time dimension, space dimension, and subject dimension; The time dimension includes: Accuracy of event time, completeness of time series, temporal relationship of key nodes, and rationality of time span; The spatial dimensions include: Accuracy of location information, rationality of spatial scope, verification of geographical relevance, and authenticity of scene description; The subject dimensions include: Verify identity of individuals, verify organizational structure, sort out relationship networks, and confirm role positioning; Content consistency verification, including factual element comparison and logical relationship verification; The comparison of factual elements includes the matching degree of core information, accuracy of detailed description, reliability of data citation, and source credibility assessment; The verification of logical relationships includes the rationality of causal relationships, the integrity of the reasoning process, the logical rigor of the argument, and the reliability evaluation of the conclusions; Impact assessment includes: communication dimension assessment, public opinion impact assessment, and social impact assessment; The evaluation of communication dimensions includes: platform influence and user engagement; Platform influence includes mainstream media coverage, social media activity, self-media communication power, and comprehensive influence score calculation; User engagement includes interaction level assessment, user portrait analysis, engagement depth analysis, and KOL influence assessment; Public opinion impact assessment includes: sentiment tendency analysis, issue development assessment; Sentiment tendency analysis includes sentiment polarity distribution, sentiment intensity changes, group sentiment evolution, and polarization degree assessment; Topic development assessment includes topic popularity trends, topic diffusion scope, related topic linkage, and development trend prediction; Social impact assessment includes: cognitive impact, behavioral impact; Cognitive influence includes the degree of public cognitive bias, the trend of group attitude change, the degree of value conflict, and the influence of social consensus; Behavioral impacts include online behavior changes, offline action conversions, group behavior trends, and social governance impacts; Comprehensive assessment includes: multi-dimensional scoring and risk warning; Multi-dimensional scoring includes calculation of feature matching score, fact verification score, impact assessment score, comprehensive score calculation and grading; Risk early warning includes risk level determination, development trend prediction, intervention suggestion generation, and effect evaluation mechanism; S303, result processing mechanism; The result processing mechanism includes confidence assessment and processing strategy; The confidence assessment includes feature matching degree assessment, evidence integrity assessment, impact assessment, and comprehensive confidence calculation; The feature matching degree evaluation includes semantic feature matching degree, emotional feature matching degree, and factual feature matching degree; The semantic feature matching degree includes cosine similarity with the standard pattern, key concept coverage, and semantic coherence score; The emotional feature matching degree includes emotional polarity consistency and emotional intensity similarity; The fact feature matching degree includes entity alignment rate and event element matching degree; The evidence integrity assessment includes the integrity of the evidence chain and the quality of evidence; The integrity of the evidence chain includes key node coverage, logical link integrity, and temporal coherence; The quality of evidence includes source reliability score, freshness of evidence, and degree of mutual confirmation of evidence; The impact assessment includes communication influence and timeliness impact; The communication influence includes platform weight score, user influence, and communication scope; The timeliness impact includes real-time popularity and development trends; The comprehensive confidence calculation includes a weighted average calculation, which is calculated as follows: Confidence = 0.4*Feature Matching Degree + 0.35*Evidence Completeness + 0.25*Influence Degree, where Confidence represents confidence, and the confidence levels are as follows: Confidence greater than or equal to 0.8 is high confidence, Confidence between 0.6 and 0.8 is medium confidence, and Confidence less than 0.6 is low confidence; The processing strategy specifically includes: directly accepting high confidence results, manually reviewing medium confidence results, and reanalyzing low confidence results. S304, optimization feedback; The optimization feedback includes result accumulation and continuous optimization; The accumulation of results includes the storage of typical cases, updating of feature database, and summary of handling experience; The continuous optimization includes adjustment of grading standards, improvement of identification specifications, and optimization of processing strategies.

5. The intelligent recognition system for public opinion means based on artificial intelligence large model according to claim 1 is characterized in that: Confirming and correcting the recognition results through the human-machine collaborative verification module specifically includes: The review process includes a three-level review mechanism: preliminary review, re-review, and final review. The average review time for a single piece of content does not exceed 2 minutes. It supports real-time updating of prompt word templates, regular weekly updates to the knowledge base, and an emergency update mechanism. The preliminary review focuses on basic feature matching, the re-review emphasizes the integrity of the evidence chain, and the final review ensures the reliability of the judgment results; Feedback mechanism, which includes real-time feedback of audit results, continuous optimization and updating of models, dynamic expansion of knowledge base, and continuous improvement of system performance.

Images