An event base incremental maintenance analysis method, system, medium and program product

Abstract

This invention belongs to the field of data processing technology and discloses a method, system, medium, and program product for incremental maintenance and analysis of an event database. The method includes: filtering a set of candidate feedback information works; generating candidate event objects; generating an event database; performing event tagging and re-calling on incremental works based on the event database; obtaining the event re-call rate within the current period and performing secondary re-calling on unmatched works; setting a lifecycle for each event and marking events exceeding their lifecycle as expired; conducting a risk assessment on the aggregated works to determine the risk level; and generating an event analysis report or handling work order based on the risk level and propagation path. This invention can identify insufficient coverage of newly added works by the existing event database through event re-call rate monitoring and automatically trigger new event discovery when the re-call rate is below a threshold, improving the timeliness of event database updates.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an incremental maintenance and analysis method, system, medium, and program product for an event database. Background Technology

[0002] With the continuous development of social media platforms, enterprises need to quickly identify feedback information events from a large number of scattered social media posts in scenarios such as brand reputation management, product quality feedback, service complaint handling, and emergency public event response. They also need to assess the risk level of these events, identify the sources and key dissemination nodes, and generate event analysis reports that can be used by legal, public relations, and operations departments. Existing solutions typically monitor feedback information through keyword subscription, sentiment recognition, manual screening, and report statistics. However, these methods still have significant technical limitations in scenarios involving multiple platforms, massive amounts of data, incremental updates, and continuously evolving events.

[0003] The first type of problem is that existing solutions struggle to automatically aggregate scattered social media content into stable feedback events. Current feedback monitoring systems typically filter content based on keyword recognition, sentiment analysis, or manual rules, with manual judgment required to identify which works belong to the same event. While this approach can detect individual negative posts, it cannot reliably identify sets of events with "the same subject, the same event type, the same time and place, or the same facts." This leads to issues such as repeated output of similar events, fragmentation of the same event, and inability to merge content from the same source across platforms, resulting in a lack of consistent data foundation for subsequent risk assessment and report generation.

[0004] The second type of problem is the lack of an automatic update mechanism for incremental content in the existing event database. For existing event databases, current solutions mostly maintain them by periodically re-clustering or manually adding new content. When newly added works cannot match existing events, the system usually cannot automatically determine whether it's a matching failure, a change in event description, or the emergence of new feedback information. Therefore, the timeliness of the event database relies on manual inspection, which can easily lead to problems such as delayed discovery of new events, expired events continuing to interfere with monitoring, and the inability to reasonably renew historical events.

[0005] The third type of problem is the insufficient analysis of the event propagation process in existing solutions. Most systems can collect metrics such as volume, interaction, and platform distribution, but they struggle to further identify the initial source, dissemination nodes, outbreak nodes, authoritative intervention nodes, and key turning points, and to construct the propagation chain between accounts based on content similarity and publication time. Consequently, in scenarios involving legal tracing, public relations handling, and operational review, it is impossible to clearly answer where the event started, through which nodes it spread, when it entered the high-risk stage, and whether a work order is needed.

[0006] Therefore, how to provide an incremental maintenance and analysis method, system, medium, and program product for the event database is an urgent problem to be solved. Summary of the Invention

[0007] This invention provides an incremental maintenance and analysis method, system, medium, and program product for an event database to solve the problems mentioned above in the prior art.

[0008] According to a first aspect of the present invention, an incremental maintenance and analysis method for an event database is provided.

[0009] In one embodiment, the event database incremental maintenance analysis method includes: The standardized work data is identified and matched to select a set of candidate feedback information works; event elements are extracted from the set of candidate feedback information works to generate candidate event objects; the candidate event objects are grouped and clustered within each group based on the similarity of the candidate events to generate an event library; Based on the event database, event tags are re-applied to incremental works, and the matching score between the works and events is calculated. Based on the matching results, the works are marked as matched or unmatched. Get the event callback rate within the current period. When the event callback rate is lower than the preset threshold and the total number of candidate feedback works reaches the minimum sample size, the event elements of the unmatched works are re-extracted and clustered, the new events are added to the event library, and the unmatched works are re-called. Set a lifecycle for each event, mark events that have exceeded their lifecycle as expired, search for new works for expired events based on keywords and entity words, and if relevant new works exist, renew the lifecycle and recalculate the lifecycle. Cross-platform similarity aggregation of related works is performed to merge works from the same source; risk assessment is conducted on the aggregated works to determine the risk level; the dissemination of related works is traced to identify the initial source and key nodes, a directed dissemination link with accounts as nodes is constructed based on content similarity and release time, and an event analysis report or handling work order is generated based on the risk level and dissemination link.

[0010] In one embodiment, the standardized work data is identified and matched to filter out a set of candidate feedback information works, including: It receives content data from different social media platforms, unifies heterogeneous fields across platforms based on field mapping configurations, and retains abnormal original fields; it performs basic text cleaning, including encoding regularization, webpage marking and dirty character cleanup, full-width and half-width character unification, whitespace compression, and duplicate topic removal, and simultaneously completes text null value marking, generation of excessively long text copies, and handling of garbled characters and abnormal forwarding prefix annotations; it performs optical character recognition on video materials by extracting frame images according to screen nodes, and directly reuses existing qualified recognition content; The system retrieves the account type dictionary to complete the unified coding of various types of accounts and records the account level synchronously. It converts the original values ​​of the number of users following, the number of interactions, and the number of dissemination on the platform into cross-platform comparable standardized values. It integrates all the processed information, including field mapping, text cleaning, video recognition, account coding, and value normalization, to generate standardized work data. Standardized works are pre-verified, including whether the work can be associated with the subscription subject and whether the work contains valid text content. Works that fail the pre-verification are directly excluded and will not be included in the subsequent feedback information judgment. Emotion recognition is performed on standardized works that have passed the verification. The statements associated with the subscribers are used as the basis for judgment. The emotional tendency, negative judgment reasons and emotional confidence are output. Works that meet the negative emotional trigger are distinguished according to the preset threshold. Works with insufficient confidence or inconsistent positive and negative judgment results are marked as pending review. Standardized works are verified based on a configurable feedback information classification system. If the semantics of a work match any first- or second-level feedback category within the feedback information classification system, it is determined that the feedback category triggering condition is met. Search for risky words and their corresponding synonyms in the text of the work, and simultaneously check whether the work's interaction and dissemination indicators meet the corresponding standards of the platform. If both the word matching and indicator standards are met, it is determined that the combination of keyword and interaction dissemination threshold trigger conditions is met. Perform content exclusion checks to identify and remove advertising content, content with no subject and noisy content, and standardized works that are repeatedly collected records; For works that are not eliminated and meet any type of feedback information triggering condition, enter the candidate identifier, triggering cause, evidence text, and threshold judgment result, summarize all standardized works carrying candidate identifiers, and output a set of candidate feedback information works.

[0011] In one embodiment, extracting event elements from a set of candidate feedback information works to generate candidate event objects includes: The event judgment text is formed by reading the title, body, description, topic, optical character recognition text, subscription subject, subject entity, negative reasons, triggering evidence and feedback information classification system of the candidate feedback information works and integrating them in the order of title, body, description, topic and optical character recognition text. The event judgment text is processed by subject aliasing and identification of entity information such as time, place, person, organization, and product name in the text. It is determined whether the text has a clear event relationship where the subject has implemented or encountered a certain fact and has caused cognitive differences, changes in circumstances, or disposal results. For works that only have broad evaluations, lack factual process, or whose content is unrelated to the subscriber, it is determined that there is no clear feedback event. When a clear feedback event is determined, the event subject, event behavior, time, location, result, and evidence fragments are extracted based on the structured field template. The first-level and second-level event types are determined according to the feedback information classification system. The standardized subject, core behavior, and cognitive disagreement result are combined to generate the event name. The factual elements are integrated in chronological order to form the event description. Core content tags are extracted from the behavior, result, and cognitive disagreement point. The extracted content undergoes compliance verification, checking whether the event name includes the subject and core behavior, whether the event description can be traced back to the original evidence, and whether the event type conflicts with the triggering reason. Works that fail the verification are marked as pending review or marked as having no events, while those that pass the verification are generated as candidate event objects.

[0012] In one embodiment, grouping candidate event objects and clustering them within each group based on the similarity of the candidate events to generate an event library includes: Candidate groups are constructed using the secondary classification of feedback information, standardized subject identifiers, and event types as grouping keys. Candidate events with consistent secondary classification of feedback information and no conflicting subjects are grouped into the same comparison set, thus narrowing the scope of semantic comparison. Within each group, semantic representations are generated for event names, event descriptions, and core content tags. Event clustering similarity scores are calculated by combining subject consistency, event type consistency, keyword overlap, semantic similarity of event descriptions, and time windows. Candidate events are compared with representative events of existing event clusters within the group based on their publication time. When the similarity score is within the first preset threshold range, the event is assigned to the event cluster with the highest score. When the similarity score is within the second preset threshold range, a deep verification is initiated to check whether the time, location, factual behavior, and results are consistent. Merging can only proceed if the verification is passed. When the similarity score is less than the preset threshold or the factual elements are mutually exclusive, the candidate event is retained as an independent event. After merging, the subject, event type, core tags, and time range corresponding to the event cluster are re-summarized to form the event cluster feature information. Candidate events are categorized as follows: those with clear time, location, subjects, and facts are categorized as specific events; and those lacking time and location but consistently reflecting similar issues are categorized as summary events. The clustering results are organized into an event library for initial tag matching of incremental works. If the event library coverage is insufficient in the initial matching results, the unmatched works are grouped and clustered again, and the unmatched works are tag matched again after the event library is updated.

[0013] In one embodiment, based on an event library, incremental works are re-tagged with event labels, and a matching score between the work and the event is calculated. The work is then marked as matched or unmatched based on the matching result, including: For each incremental work, the matching scores of the event name, main entity, core keywords, semantic similarity of the event description, publication time window, and consistency of feedback information classification are combined to determine whether the work belongs to a known event in the event database; the scores of each dimension are integrated according to preset weights to calculate the event matching score between the incremental work and the known event. The event matching score is compared with a preset threshold. If the event matching score reaches the preset threshold, the match is considered successful, and the event identifier, event name, event matching score, and matching basis are written to the work. If the work matches multiple events and reaches the preset threshold, the event with the highest matching score is selected. When the difference between the highest and second highest scores meets the preset difference condition, the event with the exact same subject entity and more refined event type hierarchy is selected first. If the attribution cannot be determined, the work is marked as pending manual review. Works that fail to match any known event are marked as having no matching event.

[0014] In one embodiment, the event callback rate within the current period is obtained. When the event callback rate is lower than a preset threshold and the total number of candidate feedback works reaches the minimum sample size, the event elements of the unmatched works are re-extracted and clustered, the new events are added to the event library, and a second callback is performed on the unmatched works, including: During the periodic incremental update process, the event callback rate is calculated. The input includes the total number of works containing newly added candidate feedback information in the period, the number of works that successfully matched existing events, the number of works that did not match events, and a preset threshold. The event callback rate is calculated as the ratio of the number of successfully matched works to the total number of works containing feedback events in the period. The event callback rate is compared with a preset threshold. If the event callback rate is lower than the preset threshold, it is determined that the current event library does not cover the newly added works. If the number of newly added candidate feedback information works in a period does not reach the minimum sample size, the callback rate judgment is skipped to avoid false triggering caused by small samples. In scenarios with small sample sizes, only observation logs are retained; in scenarios with abnormal traffic on a single platform, similar works are deduplicated; in scenarios with old event expression deviation, the event renewal or extended matching process is initiated; new events to be generated need to be compared again with various events in the event library; in scenarios with concentrated noise, noisy content is removed first. The system performs new event identification processing on the collection of works with unmatched events, extracting event elements from the unmatched works; performs semantic clustering according to the feedback information; compares the clustering results with existing events in the event database to prevent duplicate event creation; and generates a corresponding event identifier and adds it to the event database after confirming that it is a completely new event. Based on the updated event library, perform a secondary event tag rollback on works that did not match successfully.

[0015] In one embodiment, a lifecycle is set for each event. Events that exceed their lifecycle are marked as expired. For expired events, new works are retrieved based on keywords and entity words. If relevant new works exist, the lifecycle is renewed and recalculated. This includes: The event lifecycle is calculated from the moment the event is generated through clustering. If the current moment exceeds the event generation time plus the lifecycle duration, the event is marked as expired and removed from the active event set in the event library. At the same time, the event's historical data and the relationship between the event and the work are retained. Perform manual or system-initiated review and update operations on expired events; During the review and update process, new works are retrieved within a specified time range based on event keywords, event descriptions, entity words in the event name, and high-frequency feature words that rank highly, and it is determined whether the retrieved new works are related to the expired event. If a new work with a related relationship is found, the event status will be adjusted to "renewed", and the event lifecycle will be recalculated from the time the renewal takes effect.

[0016] In one embodiment, cross-platform similarity aggregation of event-related works to merge works from the same source includes: Remove topic tags from works related to the event and perform a rough screening to obtain a set of candidate similar works, and exclude works whose text conditions do not meet the specifications. The candidate similar works set is screened and judged in multiple dimensions, and differentiated comparison rules are set for text and image works and video works with optical character recognition text. The structural compatibility between works is determined based on field adaptation tags, and the optical character recognition text comprehensive verification index is obtained for video works. A dual-gating mechanism is formed based on the structural compatibility conditions and the optical character recognition text verification index to comprehensively determine the similarity of works. Based on the similarity determination results, mutually exclusive similarity clusters are constructed, and at most two rounds of missing similarity relation compensation are performed. A transitive similarity attenuation rule is introduced to truncate excessive associations. After two rounds of supplementary calculations, inter-cluster conflict verification is performed, and multiple characteristic indicators are combined to determine whether to perform cluster merging; and the similar cluster structure is updated and the total amount of cluster-level propagation is recalculated based on the merging results. Based on the updated similar cluster structure, representative works of similar clusters are determined according to the scale of dissemination and publication time, and snapshots are saved.

[0017] In one embodiment, multi-dimensional screening of the candidate similar works set includes: The works in the candidate similar works set are paired up to form candidate work pairs. For each candidate work pair, the text length ratio, term overlap ratio and longest common subsequence ratio are calculated. The corresponding preset threshold is matched according to the text length of the candidate work pair. The title and body of the candidate works are checked one by one for all applicable dimensions. The basic text dimension check is deemed to have passed only when all dimensions meet the corresponding threshold conditions. Calculate the overall similarity ranking score for candidate work pairs. The overall similarity ranking score is used only for ranking candidate work pairs and is not used as the basis for determining the similarity of candidate work pairs. If a candidate work pair contains video works with valid optical character recognition text, then additionally calculate multiple verification indicators corresponding to the optical character recognition text dimension of the candidate work pair. The optical character recognition text of candidate works is fused using a minimum value constraint mechanism. The optical character recognition text dimension verification is only passed when all basic dimension indicators meet the standards. If any basic dimension fails to meet the standards, the text cannot pass the compensation of the other high-scoring dimensions.

[0018] In one embodiment, setting differential comparison rules for graphic works and video works containing optical character recognition text includes: Identify the work type and the fill status of each text field in each pair of candidate works; If the candidate works pair is a video combination without valid optical character recognition text, then similarity comparison will only be performed if both candidate works pair have no optical character recognition text and the total length of the title and body text after removing the topic meets the standard. If the title or body field of a candidate work is empty after removing the topic tag, it will only be compared with objects whose corresponding fields are also empty, and the empty field dimension will be directly determined to pass the verification. If both the title and body of the candidate works are empty but there is valid optical character recognition text, then the similarity comparison of the candidate works will be determined only based on the optical character recognition text dimension. For candidate works that use optical character recognition (OCR) text comparison, if the title or body text dimension does not meet the standard, the work cannot be released as compensation through the OCR text dimension meeting the standard.

[0019] In one embodiment, performing a risk assessment on the aggregated work to determine the risk level includes: The system collects and aggregates eight assessment data points: qualitative level, account level, interaction level, dissemination level, platform weight, repost-like ratio, historical risk level, and data update time. When conducting event risk classification, the system performs the same risk calculation on the aggregated works associated with the event, selects the highest risk level of the valid works associated with the event as the current risk level of the event, and retains the corresponding work information and calculation basis. Qualitative, account, and quantitative levels are determined according to a preset calculation order. Qualitative levels one to four are obtained based on event semantics, content classification, risk level, and negative reasons. Account levels one to four are obtained based on account attribute information. Interaction volume and dissemination volume are matched with the historical percentile thresholds of the corresponding platform and time window to obtain interaction level and dissemination level. The highest value of interaction level and dissemination level is taken as the quantitative level. If the qualitative level is level four, high-risk scenarios are identified and judgment evidence and judgment reasons are output. The risk level is directly determined to be level four, and the comprehensive score weighted calculation is not performed. For aggregated works whose qualitative level is not level four, the comprehensive risk score is calculated based on the fixed dimension weight ratio, and the risk levels from level one to level four are mapped according to the score range of the comprehensive risk score; when an account matches multiple rules, the highest level is taken, and the interaction level and dissemination level are determined based on the platform percentile threshold updated periodically by offline tasks. Save qualitative judgment evidence, account type, interaction percentile, dissemination percentile, platform weight, weighted comprehensive score, final risk level, and calculation time. The work-level risk results are summarized to form the event-level risk results, and the risk recalculation and result tracing are supported after data updates. Monitor the interaction volume, dissemination volume, likes, reposts, and account attributes of the aggregated works; after the interaction volume, dissemination volume, likes, reposts, and account attributes are updated, recalculate the interaction level and dissemination level, update the quantitative level and calculate the comprehensive risk score; obtain the content threshold level by matching the viral risk threshold based on the repost-like ratio, and select the highest value of the weighted recalculated level and the content threshold level as the candidate risk level. Compare candidate risk levels with historical risk levels. If the candidate risk level is the highest, update the risk level and trigger the work order generation rule; otherwise, retain the historical risk level. If the event is in an active or renewed lifecycle, the risk level can only be maintained or increased, and will not be automatically downgraded with short-term data fluctuations. Risk level downgrades can only be performed when the event cycle is updated, rules are reconfigured, or manual review and confirmation is required. When the risk level is upgraded, record the update time, the level before and after the upgrade, the platform percentile threshold, the ratio of forwards to likes, and the reason for the upgrade.

[0020] In one embodiment, the dissemination of related works is traced to identify the initial source and key nodes. A directed dissemination chain with accounts as nodes is constructed based on content similarity and publication time. An event analysis report or handling work order is generated based on the risk level and dissemination chain, including: Obtain a list of works related to the event, similar clusters, publication time, platform, account attributes, interaction volume, dissemination volume, content similarity, event tags, and account weight, and trace the source of dissemination; merge reprinted works from the same source according to similar clusters, and select the earliest published works in the similar clusters as first release candidates; verify multiple dimensions of information of the first release candidates and calculate the confidence level, and classify high-confidence first release sources, candidate first release sources, and early dissemination nodes according to the confidence level; After determining the original source, identify diffusion nodes based on time windows, outbreak nodes based on sliding windows, and authoritative intervention nodes based on changes in account weight. Identify event turning points in the surrounding works corresponding to diffusion nodes, outbreak nodes, and authoritative intervention nodes, and output node type, work identifier, node time, judgment value, evidence, and confidence level. Using accounts and corresponding works as dissemination nodes, input the event-related works, similar clusters, publication time, account weight, content similarity, interaction volume, dissemination volume, and key node identification results; merge duplicate records of the same account, and retain the earliest event-related work published by the account as the time benchmark; Works related to the event are processed from earliest to latest publication time, and works with earlier publication time are selected as upstream candidates; the comprehensive similarity of title, body text and optical character recognition text is calculated; if the similarity reaches the first threshold, a directed propagation edge is generated; if the similarity reaches the second threshold, they are grouped into the same propagation cluster, and the earliest published work is selected as the cluster representative; When multiple upstream candidates exist for late-stage nodes, they are sorted according to predetermined rules, and the number of upstream and downstream edges is set to an upper limit; the first globally published works, the works with the highest dissemination data, works associated with high-weight accounts, and key nodes are selected and retained; based on the time order and the inter-cluster similarity threshold, the propagation path is supplemented between cluster representatives, and all judgment information related to nodes and edges is output; The system reads basic event information using the event identifier as the primary key, associates it with risk classification results, initial source, key nodes, propagation links, platform distribution, trend data, handling suggestions, and work order push configuration, and assembles a report data package; it also calculates statistical indicators after deduplicating the works associated with the event, sorts out the event development timeline according to the node sequence, and loads the work evidence corresponding to each node. Generate an event overview, propagation characteristics, discussion focus, event summary, and handling suggestions based on the structured data of the data packets, and bind each conclusion to a representative work identifier; read the work order push configuration, query existing handling work orders when the event meets the configuration conditions, perform new creation, update, or skip operations according to the different status of the work order, and retain the records; Output a structured event analysis report, report generation time, list of evidence works, and work order information; if the report generation or work order generation operation fails, it will retry according to the configuration, and if the number of retries exceeds the limit, it will record alarm information and the reason for failure.

[0021] According to a second aspect of the present invention, an event database incremental maintenance and analysis system is provided.

[0022] In one embodiment, the event database incremental maintenance and analysis system includes: The initial event clustering module is used to identify and match standardized work data to filter out a set of candidate feedback information works; extract event elements from the set of candidate feedback information works to generate candidate event objects; group the candidate event objects and cluster them within each group based on the similarity of the candidate events to generate an event library; The incremental tag re-call module is used to re-call event tags for incremental works based on the event library, calculate the matching score between the work and the event, and mark the work as matched or unmatched according to the matching result; The new event iteration and completion module is used to obtain the event callback rate in the current period. When the event callback rate is lower than the preset threshold and the total number of candidate feedback information works reaches the minimum sample size, the event elements of the unmatched works are re-extracted and clustered, the new events are added to the event library, and the unmatched works are re-called. The event lifecycle management module is used to set a lifecycle for each event, mark events that have exceeded their lifecycle as expired, and search for new works for expired events based on keywords and entity words. If relevant new works exist, the lifecycle is renewed and recalculated. The event in-depth analysis output module is used to aggregate similar works across platforms to merge works from the same source; to conduct risk assessment on the aggregated works to determine the risk level; to trace the spread of event-related works to identify the initial source and key nodes; to construct a directed propagation link with accounts as nodes based on content similarity and release time; and to generate event analysis reports or handling work orders based on risk level and propagation link.

[0023] According to a third aspect of the present invention, a computer-readable storage medium is provided.

[0024] In one embodiment, a computer program is stored on the computer-readable storage medium, which, when executed by a processor, implements the steps of the above method.

[0025] According to a fourth aspect of the present invention, a computer program product is provided.

[0026] In one embodiment, the computer program product stores a computer program that, when executed by a processor, implements the steps of the method.

[0027] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: (1) The present invention can identify the insufficient coverage of new works by the existing event library through event replay rate monitoring, and automatically trigger the discovery of new events when the replay rate is lower than the threshold, thereby improving the timeliness of event library updates.

[0028] (2) The present invention can set the life cycle status for feedback information events, and avoid long-term interference of expired events through the inspection, update and renewal mechanism, while retaining the ability of historical events to re-enter the monitoring range when they ferment again.

[0029] (3) The present invention can combine event semantics, account attributes, platform weight, and interactive dissemination data to achieve risk classification and dynamic upgrading, so that the priority of feedback information processing has an interpretable calculation basis.

[0030] (4) The present invention can identify the initial source, diffusion node, outbreak node, authoritative intervention node and key turning point node based on content similarity, release time and account weight, thereby generating a propagation chain that can be used for tracing and reviewing.

[0031] (5) The present invention can form a closed loop of event discovery, event update, risk classification, propagation link analysis, report generation and work order push, reduce the cost of manual screening and manual classification, and improve the efficiency of feedback information processing.

[0032] (6) Compared with the fixed-period full-scale re-clustering method, the present invention triggers a new event discovery process for unmatched works only when the event callback rate is lower than the threshold and the number of newly added candidate works reaches the minimum sample size. This can reduce the amount of invalid re-clustering calculation and improve the timeliness of new feedback information event discovery.

[0033] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0034] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0035] Figure 1 This is a flowchart illustrating an incremental maintenance and analysis method for an event database according to an exemplary embodiment; Figure 2 This is a schematic diagram illustrating the principle of an incremental maintenance and analysis system for an event database, according to an exemplary embodiment. Figure 3 This is a schematic diagram of the structure of a computer device according to an exemplary embodiment; Figure 4 This is a flowchart illustrating event extraction, event clustering, and event tag recall according to an exemplary embodiment; Figure 5 This is a flowchart illustrating event callback rate monitoring and new event discovery according to an exemplary embodiment; Figure 6 This is a flowchart illustrating event lifecycle management and renewal according to an exemplary embodiment; Figure 7 This is a schematic diagram illustrating propagation tracing, key node identification, and propagation link construction according to an exemplary embodiment. Detailed Implementation

[0036] The following description and accompanying drawings fully illustrate specific embodiments described herein to enable those skilled in the art to practice them. Some portions and features of certain embodiments may be included in or replace portions and features of other embodiments. The scope of the embodiments herein includes the entire scope of the claims and all available equivalents thereof. The various embodiments described herein are presented in a progressive manner, with each embodiment focusing on its differences from other embodiments; similar or identical parts between embodiments can be referred to interchangeably.

[0037] The modules in the apparatus or system of this application can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0038] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0039] Figure 1 An embodiment of an event database incremental maintenance analysis method of the present invention is shown.

[0040] In this optional embodiment, the event database incremental maintenance analysis method includes: S101. Identify and match standardized work data to select a set of candidate feedback information works; extract event elements from the set of candidate feedback information works to generate candidate event objects; group the candidate event objects and cluster them within each group based on the similarity of the candidate events to generate an event library. S102. Based on the event library, perform event tag re-addition on incremental works, calculate the matching score between the works and events, and mark the works as matched or unmatched according to the matching results; S103. Obtain the event callback rate within the current period. When the event callback rate is lower than the preset threshold and the total number of candidate feedback information works reaches the minimum sample size, the event elements of the unmatched works are extracted and clustered again, the new events are added to the event library, and the unmatched works are called back a second time. S104. Set a lifecycle for each event, mark events that have exceeded their lifecycle as expired, search for new works for expired events based on keywords and entity words, and if there are relevant new works, renew the lifecycle and recalculate the lifecycle. S105. Perform cross-platform similarity aggregation on related works to merge works from the same source; conduct risk assessment on the aggregated works to determine the risk level; trace the dissemination of related works to identify the initial source and key nodes, construct a directed dissemination link with accounts as nodes based on content similarity and release time, and generate an event analysis report or handling work order based on the risk level and dissemination link.

[0041] In this optional embodiment, the process of identifying and matching standardized work data to select a set of candidate feedback information works includes: receiving work data from different social media platforms, unifying heterogeneous fields across platforms based on field mapping configurations and retaining abnormal original fields; performing basic text cleaning, including encoding regularization, webpage marking and dirty character cleanup, full-width and half-width character unification, whitespace compression, and duplicate topic removal, while simultaneously handling text null value marking, generating excessively long text copies, and annotating garbled characters and forwarding prefixes; performing optical character recognition on video materials by extracting frame images from screen nodes, and directly reusing existing qualified recognition content; retrieving account type dictionaries to perform unified encoding on various types of accounts and simultaneously recording account levels, converting the original values ​​of platform followers, interaction volume, and dissemination volume into cross-platform comparable standardized values, integrating all processing information such as field mapping, text cleaning, video recognition, account encoding, and value normalization to generate standardized work data; and performing pre-verification on standardized work objects, including whether the work can be associated with a subscription subject and whether the work contains valid text content. Works that fail the pre-verification are directly excluded and not further processed. The process proceeds to subsequent feedback information assessment. For standardized works that pass verification, sentiment recognition is performed, using statements related to the subscriber as the basis for judgment. Sentiment tendency, negative judgment reasons, and sentiment confidence levels are output. Works meeting negative sentiment triggers are distinguished based on preset thresholds. Works with insufficient confidence or inconsistent positive / negative judgment results are marked as pending review. Standardized works are verified according to a configurable feedback information classification system. If a work's semantics match any first- or second-level feedback category within the system, it is determined to meet the feedback category triggering conditions. Risk words and corresponding synonyms are retrieved from the work's text. Simultaneously, the work's interaction and dissemination indicators are checked against the platform's corresponding standards. If both word matching and indicator compliance are met, the combination of keyword and interaction / dissemination threshold triggering conditions is satisfied. Content exclusion verification is performed, identifying and removing standardized works containing advertising content, content without a subject, or duplicate collection records. Works not removed and meeting any type of feedback information triggering conditions are entered with candidate identifiers, triggering causes, evidence text, and threshold judgment results. All standardized works carrying candidate identifiers are summarized, and a set of candidate feedback information works is output.

[0042] In this optional embodiment, extracting event elements from the candidate feedback information work set to generate candidate event objects includes: reading the title, body text, description, topic, optical character recognition text, subscriber, subject entity, negative reasons, triggering evidence, and feedback information classification system of the candidate feedback information works; integrating them according to the order of title, body text, description, topic, and optical character recognition text to form event judgment text; performing subject aliasing normalization processing on the event judgment text; identifying time, place, person, organization, and product name entity information within the text; determining whether the text contains a clear event relationship indicating that the subject has committed or encountered a certain fact, resulting in cognitive differences, changes in circumstances, or handling outcomes; and addressing cases with only broad evaluations, lacking factual details, or whose content is unrelated to the subscriber. For works that do not have a clear feedback event, if a clear feedback event is determined to exist, the event subject, event behavior, time, location, result, and evidence fragments are extracted based on the structured field template. The first-level and second-level event types are determined according to the feedback information classification system. The standardized subject, core behavior, and cognitive disagreement result are combined to generate an event name. The factual elements are integrated in chronological order to form an event description. Core content tags are extracted from the behavior, result, and cognitive disagreement point. The extracted content is subject to compliance verification to check whether the event name contains the subject and core behavior, whether the event description can be traced back to the original evidence, and whether the event type conflicts with the triggering reason. Works that fail the verification are marked as pending review or marked as no event. Works that pass the verification are generated as candidate event objects.

[0043] In this optional embodiment, candidate event objects are grouped, and clustered within each group based on the similarity of candidate events to generate an event library. This includes: constructing candidate groups using secondary classification of feedback information, standardized subject identifiers, and event types as grouping keys; classifying candidate events with consistent secondary classification of feedback information and no subject conflicts into the same comparison set, thus narrowing the semantic comparison scope; generating semantic representations for event names, event descriptions, and core content tags within each group; calculating event clustering similarity scores by combining subject consistency, event type consistency, keyword overlap, semantic similarity of event descriptions, and time windows; comparing candidate events sequentially with representative events of existing event clusters within the group based on their publication time; when the similarity score is within a first preset threshold range, the event is assigned to the event cluster with the highest score; when the similarity score is within a second preset threshold range, the event is assigned to the event cluster with the highest score. Initiate a deep verification process to check for consistency in time, location, factual actions, and results. Merging can only proceed if the verification is successful. When the similarity score is less than a preset threshold or the factual elements are mutually exclusive, the candidate events are retained as independent events. After merging, re-aggregate the subjects, event types, core tags, and time ranges corresponding to the event clusters to form event cluster feature information. Type-label the candidate events, marking those with clear time, location, subjects, and factual processes as specific events; marking those lacking time and location but continuously reflecting similar issues as summary events. Organize the clustering results into an event library for initial tag matching of incremental works. If the initial matching results show insufficient coverage of the event library, re-group clustering is performed on unmatched works, and tag matching is performed again on unmatched works after the event library is updated.

[0044] In this optional embodiment, based on the event database, incremental works are tagged with event tags and a matching score is calculated between the work and the event. The work is then marked as matched or unmatched based on the matching results, including: for each incremental work, determining whether it belongs to a known event in the event database by considering event name matching, subject entity matching, core keyword matching, event description semantic similarity, publication time window, and feedback information classification consistency; calculating the event matching score between the incremental work and known events by integrating the scores of each dimension according to preset weights; comparing the event matching score with a preset threshold; if the event matching score reaches the preset threshold, a successful match is determined, and the event identifier, event name, event matching score, and matching basis are written to the work; if the work matches multiple events and reaches the preset threshold, the event with the highest matching score is selected; when the difference between the highest and second-highest scores meets a preset difference condition, events with completely identical subject entities and more refined event type levels are prioritized; if the attribution cannot be determined, the work is marked as awaiting manual review; works that fail to match any known event are marked as having no matching event.

[0045] In this optional embodiment, the event callback rate within the current period is obtained. When the event callback rate is lower than a preset threshold and the total number of candidate feedback information works reaches the minimum sample size, the unmatched works are re-extracted and clustered, and the resulting new events are added to the event library. A second callback is performed on the unmatched works, including: calculating the event callback rate during the periodic incremental update process, inputting the total number of newly added candidate feedback information works in the period, the number of works successfully matched with existing events, the number of unmatched events, and a preset threshold; calculating the event callback rate according to the ratio of the number of successfully matched works to the total number of works containing feedback events within the period; comparing the event callback rate with the preset threshold; if the event callback rate is lower than the preset threshold, it is determined that the current event library does not adequately cover the newly added works; if new candidate feedback information is added in the period... If the number of works does not reach the minimum sample size, skip the current callback rate determination to avoid false triggering caused by small samples; in small sample scenarios, only observation logs are retained; in single-platform traffic anomaly scenarios, similar works are deduplicated; in old event expression offset scenarios, the event renewal or extended matching process is initiated; new events to be generated need to be compared again with various events in the event library; in noise concentration scenarios, noise content is removed first; new event identification processing is performed on the set of works with unmatched events, and event elements of unmatched works are extracted; semantic clustering is implemented according to the feedback information classification and grouping; the clustering results are compared again with existing events in the event library to prevent duplicate event creation; after confirming that it is a brand new event, the corresponding event identifier is generated and added to the event library; according to the updated event library, a second event tag callback is performed on the unmatched works.

[0046] In this optional embodiment, a lifecycle is set for each event. Events exceeding their lifecycle are marked as expired. For expired events, new works are retrieved based on keywords and entity words. If relevant new works exist, the event is renewed and the lifecycle is recalculated. This includes: calculating the event lifecycle from the moment the event is generated through clustering; if the current moment exceeds the event generation time plus the lifecycle duration, the event is marked as expired and removed from the active event set in the event database, while retaining historical event data and the association between the event and the works; performing manual or system-initiated review and update operations on expired events; during the review and update process, new works are retrieved within a specified time range based on event keywords, event description, entity words in the event name, and high-frequency feature words with the highest ranking, and it is determined whether the retrieved new works are related to the expired event; if related new works are found, the event status is adjusted to renewed, and the event lifecycle is recalculated from the moment the renewal takes effect.

[0047] In this optional embodiment, cross-platform similarity aggregation of event-related works to merge works from the same source includes: removing topic tags from event-related works and performing coarse screening to obtain a set of candidate similar works, and excluding works whose text conditions do not meet the specifications; performing multi-dimensional fine screening on the set of candidate similar works, and setting differentiated comparison rules for text-image works and video works with optical character recognition text; determining the structural compatibility between works based on field adaptation tags, and obtaining the comprehensive verification index of optical character recognition text for video works, forming a dual-gating mechanism based on structural compatibility conditions and optical character recognition text verification index to comprehensively determine the similarity of works; constructing mutually exclusive similar clusters based on the similarity determination results, performing at most two rounds of supplementary calculation of missed similar relationships, and introducing a transitive similarity attenuation rule to truncate excessive associations; performing inter-cluster conflict verification after the two rounds of supplementary calculation, and determining whether to perform cluster merging based on multiple feature indicators; updating the similar cluster structure and re-counting the total cluster-level propagation based on the merging results; and determining representative works of similar clusters based on the updated similar cluster structure, according to the propagation scale and release time, and saving snapshots.

[0048] In this optional embodiment, the multi-dimensional screening and determination of the candidate similar works set includes: pairing the works in the candidate similar works set into candidate works pairs; calculating the text length ratio, term overlap ratio, and longest common subsequence ratio for each candidate works pair; matching the corresponding preset threshold based on the text length of the candidate works pair; verifying all applicable dimensions of the title and body of the candidate works pair one by one; determining that the basic text dimension verification is passed only when all dimensions meet the corresponding threshold conditions; calculating the overall similarity ranking score of the candidate works pair; using the overall similarity ranking score only for ranking the candidate works pair and not as the basis for similarity determination; if the candidate works pair contains video works with valid optical character recognition text, additionally calculating multiple verification indicators corresponding to the optical character recognition text dimension of the candidate works pair; using a minimum value constraint mechanism to fuse the basic dimension indicators of the optical character recognition text of the candidate works pair; determining that the optical character recognition text dimension verification is passed only when all basic dimension indicators meet the standards; if any basic dimension fails to meet the standards, it cannot be compensated by the other high-scoring dimensions.

[0049] In this optional embodiment, the differentiated comparison rules for graphic works and video works with optical character recognition (OCR) text include: identifying the work type and text field filling status of each candidate work pair; if the candidate work pair is a video combination without valid OCR text, then similarity comparison is only performed if both sides of the candidate work pair have no OCR text and the total length of the title and body text after removing the topic meets the standard; if the title field or body text field of the candidate work pair after removing the topic tag is empty, then it is only compared with objects whose corresponding fields are also empty, and the empty field dimension verification is directly determined to be passed; if both the title and body text of the candidate work pair are empty but there is valid OCR text, then the similarity comparison judgment of the candidate work pair is only completed based on the OCR text dimension; for candidate work pairs with OCR text comparison enabled, if the title or body text dimension does not meet the standard, then compensation cannot be made for release based on the OCR text dimension meeting the standard.

[0050] In this optional embodiment, a risk assessment is performed on the aggregated works to determine the risk level, including: collecting eight assessment data points: qualitative level, account level, interaction level, dissemination level, platform weight, repost-like ratio, historical risk level, and data update time of the aggregated works; when conducting event risk grading, the same risk calculation is performed on the aggregated works associated with the event, and the highest risk level of the valid works associated with the event is selected as the current risk level of the event, retaining the corresponding work information and calculation basis; the qualitative level, account level, and quantitative level are determined according to a preset calculation order; and levels one to four are obtained based on event semantics, content classification, risk degree, and negative reasons. Qualitative risk assessment is performed by obtaining account levels one to four based on account attribute information. Interaction and dissemination volumes are matched against historical percentile thresholds for the corresponding platform and time window to determine the interaction and dissemination levels. The highest value of the interaction and dissemination levels is used as the quantitative risk level. If the qualitative level is level four, high-risk scenarios are identified, and evidence and reasons for the judgment are provided, directly determining the risk level as level four without performing a weighted comprehensive score calculation. For aggregated works with a qualitative level other than level four, a comprehensive risk score is calculated based on a fixed dimension weight ratio. The risk level is then mapped from one to four based on the score range of the comprehensive risk score. When an account matches multiple rules, the highest level is used. Interaction level and dissemination level are determined based on the platform percentile threshold updated periodically by offline tasks; qualitative judgment evidence, account type, interaction percentile interval, dissemination percentile interval, platform weight, weighted comprehensive score, final risk level, and calculation time are saved. Event-level risk results are formed by summarizing work-level risk results, supporting risk recalculation and result traceability after data updates; the interaction volume, dissemination volume, likes, reposts, and account attributes of aggregated works are monitored; after updates to interaction volume, dissemination volume, likes, reposts, and account attributes, the interaction level and dissemination level are recalculated, the quantitative level is updated, and the comprehensive risk score is calculated; the split is matched based on the repost-like ratio. The risk threshold is changed to obtain the content threshold level. The highest value of the weighted recalculated level and the content threshold level is selected as the candidate risk level. The candidate risk level is compared with the historical risk level. If the candidate risk level is the highest, the risk level is updated and the work order generation rule is triggered. Otherwise, the historical risk level is retained. If the event is in the active life cycle or the renewal life cycle, the risk level can only be maintained or increased. It will not be automatically downgraded with short-term data fluctuations. The risk level can only be downgraded when the event cycle is updated, the rule is reconfigured, or manual review and confirmation are required. When the risk level is upgraded, the update time, the level before and after the upgrade, the platform percentile threshold, the ratio of forwards to likes, and the reason for the upgrade are recorded.

[0051] In this optional embodiment, the dissemination of event-related works is traced to identify the initial source and key nodes. A directed dissemination chain with accounts as nodes is constructed based on content similarity and publication time. An event analysis report or handling work order is generated based on the risk level and dissemination chain, including: obtaining a list of event-related works, similar clusters, publication time, platform, account attributes, interaction volume, dissemination volume, content similarity, event tags, and account weight, and tracing the dissemination source; merging reprinted works from the same source according to similar clusters, selecting the earliest published works in the similar clusters as initial candidates; verifying multiple dimensions of initial candidate information and calculating confidence, and classifying high-confidence initial sources, candidate initial sources, and early dissemination nodes based on confidence; confirming... After determining the initial source, the system identifies diffusion nodes using time windows, outbreak nodes using sliding windows, and authoritative intervention nodes using changes in account weight. It then identifies event turning points in the surrounding works corresponding to diffusion, outbreak, and authoritative intervention nodes, outputting node type, work identifier, node time, judgment value, evidence, and confidence level. Using accounts and corresponding works as propagation nodes, the system inputs event-related works, similar clusters, publication time, account weight, content similarity, interaction volume, propagation volume, and key node identification results. Duplicate records from the same account are merged, retaining the earliest event-related work published by that account as the time benchmark. Event-related works are processed from earliest to latest publication time, selecting works with earlier publication times as upstream candidates. Select; calculate the comprehensive similarity of title, body text, and optical character recognition text; generate directed propagation edges when the similarity reaches the first threshold, and group works into the same propagation cluster when the similarity reaches the second threshold, selecting the earliest published work as the cluster representative; when multiple upstream candidates exist for late nodes, sort them according to predetermined rules, and set upper limits on the number of upstream and downstream edges; filter and retain globally first published works, works with the highest propagation data, works associated with high-weight accounts, and key nodes; supplement the propagation path between cluster representatives according to the time order and the inter-cluster similarity threshold, and output all judgment information related to nodes and edges; read the basic information of the event using the event identifier as the primary key, and associate it with the risk classification result, the first publication source, key nodes, propagation link, platform distribution, trend data, and handling. The system configures suggestion and work order push settings, assembles report data packages, deduplicates related works for the event, calculates statistical indicators, traces the event's development according to the time sequence of nodes, and loads evidence of works corresponding to various nodes. Based on the structured data in the data package, it generates an event overview, dissemination characteristics, discussion focus, event summary, and handling suggestions, binding representative work identifiers to each conclusion. It reads the work order push configuration; when an event meets the configuration conditions, it queries existing handling work orders, performs new creation, update, or skip operations according to different work order statuses, and retains records. It outputs a structured event analysis report, report generation time, list of evidence works, and work order-related information. If report generation or work order generation fails, it retryes according to the configuration; if the number of retries exceeds the limit, it records alarm information and the reason for failure.

[0052] Figure 2An embodiment of an event database incremental maintenance and analysis system according to the present invention is shown.

[0053] In this optional embodiment, the event database incremental maintenance and analysis system includes: The initial event clustering module 201 is used to identify and match standardized work data, and filter out a set of candidate feedback information works; extract event elements from the set of candidate feedback information works to generate candidate event objects; group the candidate event objects, and cluster them within the group according to the similarity of the candidate events to generate an event library; The incremental tag return module 202 is used to return event tags to incremental works based on the event library, calculate the matching score between the works and the events, and mark the works as matched or unmatched according to the matching results. The new event iteration completion module 203 is used to obtain the event callback rate in the current period. When the event callback rate is lower than the preset threshold and the total number of candidate feedback information works reaches the minimum sample size, the event elements of the unmatched works are re-extracted and clustered, the new events are added to the event library, and the unmatched works are re-called. The event lifecycle management module 204 is used to set a lifecycle for each event, mark events that have exceeded their lifecycle as expired, search for new works for expired events based on keywords and entity words, and if there are relevant new works, renew the lifecycle and recalculate the lifecycle. The event in-depth analysis output module 205 is used to perform cross-platform similarity aggregation of event-related works to merge works from the same source; to conduct risk assessment on the aggregated works to determine the risk level; to trace the dissemination of event-related works to identify the initial source and key nodes; to construct a directed dissemination link with accounts as nodes based on content similarity and release time; and to generate an event analysis report or handling work order based on the risk level and dissemination link.

[0054] To facilitate understanding of the above technical solutions of the present invention, the following further explains the above technical solutions of the present invention from the perspective of architecture and principle, as follows: Step 1: Data access and standardization of social media works.

[0055] The system receives content data from various social media platforms. The fields processed include content ID, platform identifier, publication time, inclusion time, title, body text, description, and topic. The text (i.e., the optical character recognition text mentioned above), author nickname, account ID, account attributes, authentication information, number of users followed on the platform, number of interactions, number of disseminations, number of plays, number of comments, number of reposts, number of likes, subscription items, keyword recognition information, etc.

[0056] During processing, the platform identifier and platform version of the work are read, and the field mapping configuration table is queried to map the title, body text, description, topic, publication time, account, and interactive dissemination fields of each platform to a unified field identifier; fields that cannot be mapped are written to an exception queue, while retaining the original field name and original value. Subsequently, the title, body text, description, topic, and... The text is processed sequentially to unify character encoding, The process includes removing tags and control characters, unifying full-width and half-width characters, folding continuous whitespace, cleaning up link parameters, and removing duplicate topics, while retaining the original text for traceability.

[0057] Null values ​​and length are handled separately for each field: Titles or body texts with a length of less than 2 characters after removing blanks are recorded as null values; when the title exceeds 200 characters or the body text exceeds 1000 characters, the original text is retained and a truncated copy is generated for calculation; garbled characters, invisible control characters, and duplicate forwarding prefixes are recorded as abnormal markers to avoid direct deletion that could lead to the loss of original evidence.

[0058] For video works, video frames are extracted at fixed time intervals or shot change points. Optical character recognition is performed on the frame images. The recognized segments with timestamps are merged in order of appearance, and consecutive duplicate segments are removed to form a new sequence. Text field; already exists The results are reused, and re-identification is only performed when the result is empty or the quality score is lower than the configured threshold.

[0059] Account attributes are uniformly coded using an account type dictionary. For example, media, industry experts, organizations, brand accounts, top or mid-tier opinion leaders, and ordinary consumer accounts are each mapped to a unified account type code, and the account level is also written into the code. Interaction volume, dissemination volume, and the number of users followed on the platform retain their original values, while logarithmic values ​​or historical percentiles are calculated according to the platform and statistical period to obtain standardized values ​​that can be compared across platforms. The unified fields, original values, standardized values, cleansing status, and anomaly markers are all written into the work view table.

[0060] The output format is a standardized artwork object, for example: { " ": " ", " ": " ", " ": "2026-06-20 10:30:00", " "Delivery issues with a certain vehicle model have led to user complaints". " ": Users reported delivery delays and questioned after-sales response", " ": ["Delivery", "After-sales"], " ": "", " ": "Car Observer", " ": " ", " 5200, " 180000 }".

[0061] In one embodiment, social media content from a car brand over the past 30 days is accessed, and titles and other data are retrieved from the media platform. The text is read from the image and text platform (title and body), and from the social platform (body and topic), and is uniformly converted into the above-mentioned standardized work objects for subsequent clustering and risk analysis.

[0062] Step 2: Feedback information triggers the screening and selection of candidate works.

[0063] Standardized work data refers to the unified work object generated in step one, which includes at least the work identifier, platform identifier, publication time, title, body text, description, and topic. Text, subscriber, entity, account attributes, sentiment, negative reasons, keyword identification results, interaction volume, dissemination volume, and corresponding standardized values.

[0064] First, verify whether the work can be linked to the subscription entity, and check the title, body text, or... The text is checked to determine if at least one valid content field exists. After passing the check, it is then judged sequentially based on negative sentiment triggers, sensitive feedback information category triggers, and keyword and interaction propagation threshold combination triggers. If any trigger condition is met, a candidate identifier and trigger reason are written, and the event element extraction process begins. If the trigger condition is not met, or the text is identified as an advertisement, lacks main noise, or involves duplicate data collection, the event will not proceed to the event analysis process.

[0065] Receive work title, body text, description, topic, The text, the subscriber, and the entity are considered. Statements related to the entity are used as evidence for judgment, and the output is a positive, neutral, or negative sentiment, a negative reason, and a confidence level. When a negative expression clearly points to the subscriber and the negative confidence level reaches the configured threshold, it is judged as negative; if the confidence level is insufficient or there is a conflict between positive and negative evidence, it is written to a pending review status.

[0066] The sensitive categories of feedback information come from a configurable feedback information classification system, including primary or secondary categories such as product quality and safety, non-compliance, high brand influence, service complaints, delivery perception disagreements, sales perception disagreements, and consumer rights. When the event semantics of a work identify one of these categories, it can be included in the candidate set even if the current interaction volume is low.

[0067] The keyword and interaction threshold combination refers to works that identify risky words or synonyms related to recall, abnormal situations, feedback requests, rights claims, information ambiguity, and information leakage, and whose interaction volume or dissemination volume reaches the absolute threshold or historical percentile threshold of the corresponding platform. Risk rules consist of negative sentiment rules, sensitive category rules, keyword and threshold combination rules, and rules excluding advertising and duplicate content; the first three types of rules allow entry based on the satisfaction of any one of them, while exclusion rules take precedence. The system writes specific identification rules, evidence text, and threshold results into candidate works as the basis for subsequent risk level interpretation.

[0068] The output is a collection of candidate feedback works, with fields including: work identifier ( Trigger type () Triggering reason () ), main entity ( Negative reasons () ) and the time of entering the candidate set ( ).

[0069] In one embodiment, if the emotion recognition result of a certain work is negative, the negative reason is "delivery delay, insufficient after-sales response", and the interaction volume exceeds a preset value, then the work is added to the candidate feedback information work set and the trigger type is marked as "negative emotion and interaction threshold".

[0070] Step 3: Extraction of candidate event elements.

[0071] like Figure 4 As shown, the input for event element extraction is the candidate feedback information output from step two. This involves reading the title, body, description, and topic of the candidate works. A classification system for text, subscribers, entities, negative reasons, triggering evidence, and feedback information is established, and each text field is categorized by title, body, description, topic, and... The sequential combination of text forms the event determination text.

[0072] The event determination text is unified by the alias of the executing entity and identified by time, place, person, organization, and product name. Then, it is determined whether there is a clear event relationship in the text that is "a clear event relationship in which the entity implements or encounters a certain fact and causes cognitive differences, changes in circumstances, or disposal results". If only a broad evaluation appears, without factual process, or is unrelated to the subscriber, it is determined to be an event that does not contain clear feedback information.

[0073] After determining that an event has clear feedback information, the event subject, event behavior, time, location, result, and evidence fragments are extracted according to a structured field template, and the primary and secondary event types are determined by the feedback information classification system. The event name is generated by combining the standardized subject, core behavior, and disputed result, the event description is generated by the factual elements in chronological order, and the core content tags are extracted from the behavior, result, and disputed points as tags of 5 to 15 characters.

[0074] After extraction, check whether the event name contains the subject and core behavior, whether the event description can be traced back to the original evidence, and whether the event type conflicts with the triggering reason. If the check fails, it will enter the pending review or output no event marker. If the check passes, a candidate event object will be generated.

[0075] The processing logic includes: first, determining whether the work contains explicit feedback information events; if so, extracting the event name ( ), Event Description ( ), core content tags ( ), event type ( ), main entity ( ) and event elements ( The core content tag is a concise tag of 5 to 15 characters, used to represent the core facts or points of contention involved in the work. If the work does not contain explicit feedback information events, then a uniform "no event tag" will be output.

[0076] After the above event element extraction and field validation, candidate event objects are output; these objects serve as input for subsequent classification grouping, semantic clustering, and event label re-recall, for example: { " ": " ", " ": "Dispute over delayed delivery of a certain vehicle model" " "Multiple users have reported delays in vehicle delivery and questioned the efficiency of after-sales communication." " ": ["Delivery delay", "Slow after-sales response"], " ": "Product Delivery Category", " ": A certain car brand" }".

[0077] In one embodiment, for a work titled "Car ordered three months ago still not delivered", core content tags such as "delivery delay", "user complaints" and "insufficient after-sales response" are extracted from the text and the event name "Dispute over delivery delay of a certain car model" is formed.

[0078] Step 4: Event clustering based on classification grouping and semantic consistency.

[0079] like Figure 4 As shown, event clustering is performed on candidate event objects. Input fields include... , , , , , and release time.

[0080] Candidate groups are formed using secondary classification of feedback information, standardized subject identifiers, and event types as grouping keys. Candidate events with different secondary classifications or clearly conflicting subjects are not included in the same comparison set, thereby narrowing the scope of semantic comparison. Within each group, semantic representations are generated for event names, event descriptions, and core content tags. The following event clustering similarity scores are calculated by combining subject consistency, event type consistency, keyword overlap, semantic similarity of event descriptions, and time windows.

[0081] Candidate events are compared sequentially with representative events from existing event clusters within the group, based on their publication time: events with a score greater than or equal to 0.75 (i.e., the first preset threshold range) are merged into the highest-scoring event cluster; events with scores between 0.60 and 0.75 (i.e., the second preset threshold range) undergo further verification to ensure consistency in time, location, factual behavior, and outcome, and are only merged if they match; events with scores less than 0.60 (i.e., less than the threshold) or mutually exclusive factual elements are retained as independent events. After merging, the subject, event type, core tags, and time range of the event clusters are re-summarized as event descriptions for subsequent work matching.

[0082] Candidate events with a clear time, place, subject, and factual sequence are marked as specific events; candidate events lacking a clear time and place but consistently reflecting similar issues are marked as summary events. Specific events are subject to stricter time windows and factual consistency conditions, while summary events allow for longer time windows, but still require consistency in subject and event type to avoid merging works that are only similar in theme but different in fact.

[0083] Clustering results are not directly processed as a one-time analysis report, but are instead written to or updated in the event database. ),in, Indicates an event, This represents a structured collection of data used for the continuous storage and maintenance of event objects. The event library is then used for the first re-call of incremental works; only when the first re-call rate indicates insufficient coverage of the event library will the system perform this clustering process for unmatched works, and perform a second re-call after the event library is updated.

[0084] The output format is a structured event library. ,For example: { " ": " ", " ": "Dispute over delayed delivery of a certain vehicle model" " ": "Product Delivery Category", " ": "Specific event", " ": User complaints arising from delivery delays of a certain vehicle model", " ": [" ", " ", " "], " ": "2026-06-20 02:00:00" }".

[0085] In one embodiment, candidate events such as "delivery delay", "repeated changes in vehicle pick-up time" and "no explanation given by after-sales service" are grouped into the same event; while "vehicle system lag" is grouped into another product experience event.

[0086] When performing event clustering, a clustering similarity score can be calculated for any two candidate events. This is used to determine whether an event belongs to the same event cluster. The calculation method is as follows: =0.30×Consistency score of main entity + 0.20×Consistency score of event type + 0.20×Score of overlap of core keywords + 0.20×Semantic similarity score of event description + 0.10×Score of time window.

[0087] Specifically, the event clustering similarity scoring items and their value rules are shown in Table 1, which illustrates the value selection method of each component scoring item and its role in event clustering: Table 1 Event Clustering Similarity Scoring Items and Value Rules when When the threshold is greater than or equal to 0.75 (i.e., the first preset threshold range), two candidate events can be grouped into the same event cluster; when... When the value is between 0.60 and 0.75 (i.e., the second preset threshold range), the system enters a secondary verification to check for conflicts in the facts, time, location, and subjects; when If the value is less than 0.60 (i.e., less than the preset threshold), the data will not be merged. If two candidate events have the same subject entity but are mutually exclusive in terms of facts, such as one being "delivery delay" and the other being "accidental triggering of assisted driving", they will not be merged even if there is overlap in the keywords.

[0088] Step 5: Recall the event label.

[0089] like Figure 4 As shown, based on Perform event tag callbacks on all or incremental works. Input fields include work title, body text, topic, text, The event name, event description, core keywords, and event category are listed in the text.

[0090] The processing logic includes: determining whether each work belongs to [the relevant category / category]. A known event in the work is identified; the criteria for judgment include event name recognition, subject entity matching, core keyword matching, semantic similarity of event description, publication time window, and consistency of feedback information classification. If the match is successful, the work is written to... , , , If no known event is matched, it is marked as "no matching event".

[0091] The output format is a record of event attribution for the work, for example: { " ": " ", " ": " ", " ": "Dispute over delayed delivery of a certain vehicle model" " 0.86, " The main points are consistent, the core keywords include delivery delays and after-sales communication, and the facts are consistent. }".

[0092] In one embodiment, although the newly added work on a social media platform does not directly mention the event name, the text mentions "the delivery cycle of the same model of car has been delayed again and again". Based on the similarity of the main entity, keywords of delivery delay and event description, it is reclassified as "dispute over delivery delay of a certain model".

[0093] When the event tag is retried, the event matching score can be calculated. It is used to determine whether a work belongs to a known event. The calculation method is as follows: =0.25×Main entity matching score +0.20×Feedback information classification matching score +0.20×Core keyword matching score +0.25×Event description semantic similarity score +0.10×Publication time window score.

[0094] Specifically, the event tag callback matching scoring items and value rules are shown in Table 2, which explains the calculation basis and matching function of each component scoring item: Table 2 Event Tag Recall Matching Scoring Items and Value Rules when When the value is greater than or equal to 0.75, the work is pushed back to this event; when multiple events meet the threshold, the event is selected. For the highest-scoring event, if the difference between the highest score and the second-highest score (i.e., the second-highest score) is less than 0.05, then the event with the exact same subject entity and a more specific event type will be prioritized; if it still cannot be determined, it will be marked as pending manual review. The first callback is used for existing events. The initial determination of ownership of a work, followed by a second review, is used for... After the new incident discovery process was updated, the previous The works were rematched.

[0095] The callback output fields are defined as shown in Table 3, which describes the name, status, and business meaning of the callback fields: Table 3. Event Attribution Fields for Works Step Six: Event Recall Rate Monitoring and New Event Discovery Triggering.

[0096] like Figure 5 As shown, the event callback rate is calculated during the daily incremental update process. The input fields include the total number of newly added candidate feedback works, the number of works that successfully matched existing events, the number of works that did not match events, and the configured threshold.

[0097] The processing logic includes: calculating the event callback rate, using the following formula: Event callback rate = Number of works that successfully match existing events / Total number of works containing feedback information on the day.

[0098] When the event callback rate is lower than the configured threshold, the system determines the current... Insufficient coverage of newly added works may lead to drift in the expression of new or existing events, thus triggering the new event discovery process. The default threshold can be set to 70%, but this threshold can be configured according to business scope, platform, or industry. If the number of newly added candidate feedback works on a given day is lower than the minimum sample size, the callback rate judgment will be skipped to avoid false triggering due to a small sample size.

[0099] Specifically, the event recall rate monitoring parameters are shown in Table 4, which illustrate the new event discovery recall rate threshold, minimum sample size, update frequency, and monitoring scope: Table 4 Event Recall Rate Monitoring Parameters The output format is a callback rate monitoring record, for example: { " ": "2026-06-21", " 120, " 72, " 48, " 0.60, " 0.70, " ": }".

[0100] In one embodiment, 120 new candidate feedback entries are added on a certain day. Only 72 of them match existing events, resulting in a callback rate of 60%, which is below the 70% threshold. The system automatically triggers the new event discovery process, re-extracting and clustering the 48 unmatched entries. To avoid false triggers caused by sudden traffic spikes, small sample sizes, or expression drift, the following anomaly handling rules can be further set, as shown in Table 5, to illustrate the handling methods for scenarios such as small sample sizes, sudden traffic spikes, expression drift of old events, repeated new events, and noise concentration: Table 5 Abnormal Recall Rate Scenarios and Handling Rules Step 7: New event discovery and event database update.

[0101] like Figure 5 As shown, new event discovery is performed on a collection of works with no matching events. Input fields include the title, body, and topic of the unmatched works. Text, main entity, publication time, account attributes, and interaction data.

[0102] The processing logic includes: re-extracting event elements from unmatched works; performing semantic clustering after classifying and grouping according to feedback information; and comparing the clustering results with existing data. A second comparison is performed to avoid duplicate creation; if it is confirmed to be a new event, a new one is generated. And added to ; Updated Then perform a second round of testing on the unmatched works.

[0103] The output format is a structured event library update record, for example: { " ": "2026-06-21", " ": [ { " ": " ", " "Controversy over accidental triggering of driver assistance system in a certain vehicle model" " ": "Product Safety Category", " ": [" ", " "] } ], " 41 }".

[0104] In one embodiment, if the unmatched works contain a cluster of content such as "false triggering of assisted driving" and "abnormal emergency braking," these are clustered into a new event, "Controversy over false triggering of assisted driving of a certain vehicle model," and written into the database. They also relabeled the relevant works with the event tag.

[0105] Step 8: Event lifecycle management and renewal.

[0106] like Figure 6 As shown, set the lifecycle state for each event. Input fields include... , , , , , , and .

[0107] The processing logic includes: calculating the event's lifecycle from the time the cluster was generated, with a default lifecycle of 14 days; when the current time exceeds... When adding a lifecycle event, the system marks the event as expired and removes it from the default active event set in the event library, but retains the event's historical data and related works. For expired events, users or the system can trigger an update check; during the check, the system considers event keywords, event description, entity words in the event name, and high-frequency feature words. Within a specified time window, newly added works are retrieved, and it is determined whether these works are related to expired events. If new related works exist, the event status is changed to "renewed," and the lifecycle is recalculated from the renewal date. The event lifecycle management parameters are shown in Table 6, which illustrates the time parameters used for event creation, update checks, renewal, and cooldown control. Table 6 Event Lifecycle Management Parameters The output format is an event state change record, for example: { " ": " ", " ": " ", " ": " ", " ": "2026-07-05 09:00:00", " "In the past 14 days, 6 new works related to disputes over delivery delays were found." }".

[0108] In one embodiment, the "delivery delay dispute of a certain model" expired after 14 days. When the system checked for updates, it found that multiple related complaints had appeared in the past 14 days, and the main entity and high-frequency feature words were consistent. The system then renewed the event and re-included it in the active event set of the event database.

[0109] Step 9: Aggregate similar works across platforms.

[0110] Before event analysis and dissemination chain construction, aggregate works with the same source or high similarity. Input fields include title, body text, OCR text, topic, platform, publication time, account information, and dissemination interaction data.

[0111] During the initial screening, topic tags are first removed from the title and body text, and then the search engine performs similar content searches on the title and body text separately; in one embodiment, the following is adopted: The search method retains the top 5000 results with a similarity score greater than 75 for each field. The union of the title and body text results is then used to obtain a candidate set of similar works. Results with a body text exceeding 1000 characters, a title exceeding 200 characters, or both a title and body text less than 2 characters and no valid similar works are considered similar. Text submissions marked as having abnormal length during the rough screening stage will not proceed to the subsequent fine screening.

[0112] For each candidate work pair, the text length ratio, term overlap ratio, and longest common subsequence ratio are calculated separately. The text length ratio is calculated by dividing the shorter text length by the longer text length, with a threshold of 0.5. The term overlap ratio is calculated by dividing the number of intersecting words by the number of union words, with a threshold of 0.8 for texts shorter than 15 characters, 0.6 for texts between 15 and 50 characters, and 0.4 for texts longer than 50 characters. The threshold for the longest common subsequence ratio in the title is 0.8, and in the body text it is 0.7. All six applicable dimensions for both the title and body text must pass simultaneously. The total similarity score is only used for candidate ranking and does not replace the threshold determination.

[0113] Video works exist When dealing with text, The text now includes three refined screening dimensions: length ratio, term overlap ratio, and longest common subsequence ratio. Furthermore, it requires that the title, body text, and... All applicable dimensions passed simultaneously. The video work did not. When text is present, only with the same type of text that does not exist. Works with a combined title and body text length exceeding 20 characters after removing the topic are compared. If a title or body text field is empty after removing the topic, that dimension is only compared with works where the corresponding field is also empty, and that dimension is considered to have passed; if both the title and body text are empty but... Non-empty video works can still be based solely on Dimensions are used for similarity determination.

[0114] To adapt to mixed feedback information data of text, images, and short videos, the title, body text, and... Instead of comparing concatenated texts according to the same text rules, a mask is generated for each work's fields. Separate tags from empty fields. For titles or body text, both works' corresponding fields are valid. Then, content comparison is initiated. If both fields are empty, only structural similarity is determined, but no similarity score is contributed. If only one field is empty, the comparison is complete. And directly determine that the field is incompatible. Works with both the title and body text empty will only be considered valid if both parties have valid entries. Text comparison is entered; none The video still needs to meet the threshold for the combined effective length of the title and body text. This processing makes the field missing pattern a constraint for similarity determination, preventing empty fields from being treated as identical text and causing disordered aggregation.

[0115] For those with For video works containing text, retain deduplicated subtitle segments with timestamps, filter out segments with low confidence and those shorter than the configured value, and calculate separately. Length ratio Term overlap ratio The proportion of the longest common subsequence Effective fragment coverage Consistency rate of segment order .Will The joint check value is denoted as Its calculation formula can be expressed as: ;in, The number of valid characters matched is relatively long. The proportion of valid characters, To maintain the proportion of time-sequential matching segments to the total number of matching segments, a minimum constraint is used instead of a weighted average, ensuring that a low score in any base dimension cannot be compensated for by other high-scoring dimensions.

[0116] The structural compatibility parameters obtained from the field application mask and Joint check value As a two-parameter gating: the work only applies to... All applicable dimensions for the title and body text have reached their respective thresholds and are enabled. hour achieve Similarity is only determined when a threshold is met; Scores cannot replace the failure of a single non-empty title or body text. Therefore, videos from the same source press release can pass due to their wide subtitle coverage and consistent segment order, while different videos sharing only a few high-frequency subtitles will be excluded due to insufficient coverage or order consistency. Simultaneously, videos lacking titles or body text will not be mistakenly merged with text and image works that have corresponding non-empty fields, thus balancing recall in short videos without body text and mismatch suppression in fragmented subtitle scenarios.

[0117] The process refines the output of work pairs, determining their similarity or dissimilarity, scores across various dimensions, and overall ranking. Mutually exclusive similarity clusters are constructed starting with similar relationships involving a large number of related members, with at most two rounds of recalculation for any missed relationships; similarity relationships are not propagated indefinitely. The representative work for each similar cluster is the one with the highest dissemination volume at the time of cluster construction; if dissemination volumes are equal, the work with the earlier publication date is selected. Once the representative work is confirmed, it is saved as a snapshot. After refinement, the following data is obtained: similar work relationships, similar cluster identifiers, a list of member works, representative work identifiers, and cluster-level dissemination and interaction data, which can be used for event analysis and dissemination chain construction.

[0118] The two rounds of missing relationship calculation are performed on a limited candidate set: the first round only constructs mutually exclusive initial clusters among works with direct high-confidence similarity edges; the second round only recalculates boundary work pairs that share at least one directly adjacent work, reach the candidate threshold with the cluster representative work, or are in the same subject and time window, and does not continue to expand the next round based on the indirect relationships added in the second round. Each similarity edge record... , , and ,in Distinguish between the initial direct calculation, the second round of supplementary calculation, and the indirect derivation.

[0119] For indirect relationships, the system is configured to transmit similar attenuation coefficients. ,and For a network consisting of h directly similar edges... The hop path, whose transit score is defined as Among them, the first on the path The direct similarity score of 6 directly similar edges is denoted as ; Indicates the number of hops; the similarity threshold is denoted as The maximum configured pass-through level is denoted as When the transmission score is below the similarity threshold, i.e. Or, the number of hops exceeds the maximum transmission level, i.e. The system immediately disconnects the association. By multiplying the minimum direct score of the path by the hierarchical attenuation, a weakly similar edge can be prevented from being covered by other strongly similar edges, and irrelevant works in multi-layer forwarding links can be pulled into the same similar cluster due to continuous transmission.

[0120] After two rounds of recalculation, inter-cluster conflict verification is performed. If there are at least a configured number of bidirectional direct similar edges between two independent clusters, the system first calculates the cross-cluster direct edge density, representative work similarity, consistency of feedback information subjects, consistency of event types, and overlap of time windows. Clusters are merged only when the subjects and event types do not conflict, and the cross-cluster direct edge density and representative work similarity both reach the threshold; otherwise, they remain independent and are recorded. and After merging or retaining, the total propagation amount at the cluster level is recalculated, but merging is not triggered solely by the total propagation amount, thereby preventing high-propagation-noise clusters from swallowing small-scale real-event clusters.

[0121] After completing the above-mentioned restricted supplementary calculation, attenuation truncation, and conflict verification, representative works are determined based on the dissemination volume and release time and saved as snapshots. The sources of direct edges, supplementary edges, and indirect edges, the scores after attenuation, the conflict verification results, and the final member list are output for subsequent event analysis and dissemination chain construction.

[0122] The output format is similar cluster records, for example: { " ": " ", " ": " ", " ": [" ", " ", " "], " ": [" ", " ", " "], " 230000 }".

[0123] In one embodiment, the same complaint video is forwarded to various social media platforms by different accounts, through the text, title, and... Similarity scores are used to group works into the same similarity cluster, with the work from the social media platform with the highest dissemination being used as the representative work for the event report.

[0124] Step 10: Risk Classification and Dynamic Upgrading.

[0125] When classifying the risk of a work, the system reads the qualitative feedback information level, account attribute level, interaction level, dissemination level, social platform weight, repost-like ratio, historical level, and update time. When classifying the risk of an event, the system performs the same calculation on the works associated with the event, and uses the highest risk level among the currently valid associated works as the current level of the event, while retaining the corresponding works and calculation basis.

[0126] The specific calculation order is as follows: qualitative levels from 1 to 4 are obtained based on the event semantics, feedback information classification, risk level, and negative reasons; account levels from 1 to 4 are obtained based on the account attribute dictionary; interaction volume and dissemination volume are mapped to the historical percentile levels of the corresponding social platforms and time windows, and the higher value is taken as the quantitative level; the comprehensive score is obtained by substituting it into the weighted formula, and then mapped to risk levels from 1 to 4 according to the score range.

[0127] Each calculation simultaneously saves qualitative judgment evidence, account type, the quantile range of interaction and dissemination, platform weight, weighted score, final grade, and calculation time, so that work-level results can be summarized into event-level risks and can be recalculated and interpreted when data is updated in subsequent periods.

[0128] The processing logic includes: qualitative dimensions determined by event semantics and feedback information classification; account dimensions determined by posting account attributes, account level, authentication information, etc.; and quantitative dimensions determined by the percentile thresholds of dissemination and interaction volume on the corresponding social media platforms. A comprehensive risk score is then calculated. =Platform weight × (Qualitative rating score × 50%) (Interaction level, dissemination level) × 30% + Account level score × 20%)

[0129] If the qualitative feedback information level is four, meaning a high-impact attention scenario is identified, and the model outputs corresponding evidence and reasons for the judgment, it is directly determined as a level four high-impact event, skipping the weighted calculation. The risk grading input items are defined in Table 7, which explains the generation method and value range of qualitative level, account level, interaction level, dissemination level, and social media platform weight: Table 7. Input Items for Feedback Information Risk Classification The aforementioned qualitative, account, interaction, and dissemination dimensions all use levels 1 to 4, with level 4 being the highest level for each dimension. Account levels are generated according to the following rules: media accounts and industry expert accounts are level 4; organization accounts, brand accounts, top opinion leaders, and public figure accounts are level 3; mid-tier or lower-tier opinion leaders and individual merchants are level 2; ordinary consumer accounts and other accounts are level 1. If an account meets multiple conditions simultaneously, the highest level is used. Interaction and dissemination levels are determined based on the platform's historical percentile thresholds. For example, exceeding the 99th percentile is level 4, exceeding the 95th percentile is level 3, exceeding the 80th percentile is level 2, and the rest are level 1; specific percentile values ​​are updated daily or weekly by offline tasks. The risk level mapping rules are shown in Table 8, which illustrates the correspondence between the comprehensive risk score range and the feedback information levels from level 1 to level 4. Table 8. Risk Score and Level Mapping for Feedback Information For works or events that have already been tagged, the update time of interaction volume, dissemination volume, likes, reposts, or account attributes is used as the recalculation trigger point. After triggering, the latest cumulative value is read, and the interaction level and dissemination level are regenerated according to the current platform percentile threshold. The higher value of the two is taken to replace the original quantitative level, and the comprehensive score is recalculated with the original qualitative level, account level, and social platform weight.

[0130] Calculate the repost-like ratio; when this ratio reaches the viral risk threshold corresponding to the current historical level, generate a content threshold level. The new candidate level is the higher of the reweighted level and the content threshold level, and is compared with the historical level; if the candidate level is higher than the historical level, update the risk level and trigger the work order rule again; if the candidate level is equal to or lower than the historical level, retain the historical level.

[0131] The principle of "only upgrading, never downgrading" means that within the active or extended lifecycle of the same event, automatic recalculation can only maintain or increase the already confirmed risk level, and will not automatically downgrade due to a single decrease in interaction, platform data replenishment, or a short-term drop in popularity. A downgrade can only be implemented after the event enters a new lifecycle, rules are reconfigured, or manual review and confirmation. Each upgrade records the update time, the level before and after the upgrade, the percentile threshold, the ratio of reposts to likes, and the specific triggering reason.

[0132] The output format is a risk classification record, for example: { " ": " ", " 3, " 4, " 3, " 1.2, " 3.24, " Level 3 " "The transmission threshold has risen to reach level three." }".

[0133] In one embodiment, an event is initially a Level 2 attention alert. Subsequently, it is forwarded by a media account and the interaction volume exceeds the platform's percentile threshold. After recalculation, the system upgrades it to a Level 3 sensitivity alert and records the reason for the upgrade.

[0134] Step 11: Source tracing and key node identification.

[0135] like Figure 7 As shown, this method traces the dissemination and identifies key nodes of related works that have formed an event. Inputs include a list of related works, similar clusters, publication time, social media platforms, account attributes, interaction volume, dissemination volume, content similarity, event tags, and account weight. Works reposted from the same source are merged by similar clusters to avoid duplicate publication of the same content and repeated use in initial candidate and peak statistics.

[0136] When tracing the source, the earliest works in each similar cluster are sorted in ascending order of publication time, and the top 5 are selected as the first publication candidates; the originality of the content, whether there are reprint or citation marks, account verification, and source credibility are verified in sequence. The consistency between the region and the event, as well as the citation relationship in subsequent works, are considered, and a confidence score of 0 to 100 is calculated according to the formula for the confidence score of the first source. A score of 80 is output as a high-confidence first source, a score of 60 to 80 is output as a candidate first source, and a score below 60 is only used as an early dissemination node.

[0137] Once the initial source is determined, the number of works is counted in hourly or daily windows to identify global and local peaks, forming diffusion nodes; a sliding window of interaction or dissemination volume over the past 24 hours is used to identify sudden increases in works, forming outbreak nodes; an account's weight reaching or rising to level 2 or above for the first time is used to identify authoritative intervention nodes; and then, turning points such as official responses from the main body, information clarifications, platform intervention, and changes in feedback trends are identified from the works before and after these nodes. Outputs include node type, work identifier, node time, judgment value, evidence, and confidence level.

[0138] The processing logic includes: sorting related works by publication time in ascending order, selecting the earliest published works as initial candidates; verifying the initial candidates based on content originality, source credibility, account attributes, IP location, content similarity, and subsequent dissemination impact, and outputting the initial candidates and their confidence levels. Key nodes include diffusion nodes, outbreak nodes, authoritative intervention nodes, and event turning point nodes. Diffusion nodes can be determined based on a global or local peak in the number of works within a time window; outbreak nodes can be determined based on a sudden increase in the dissemination or interaction volume of a single work relative to the sliding window mean and standard deviation; authoritative intervention nodes can be determined based on the first increase in account weight level; event turning point nodes can be identified by combining works near the aforementioned nodes, such as official responses from the main body, brand responses, information clarifications, platform intervention, changes in feedback information tendencies, and third-party intervention.

[0139] The output format is key node records, for example: { " ": " ", " ": " ", " 82, " ": [ { " ": " ", " ": " ", " ": "2026-06-21 18:20:00", " "Interaction volume exceeded the average of the past 24 hours plus three standard deviations." }, { " ": " ", " ": " ", " "High-influence media accounts intervene for the first time." } ] }".

[0140] In one embodiment, an event is identified as first posted by an ordinary user, then reposted by a media account to form an authoritative intervention node, and finally triggered by an abnormal increase in the interaction volume of a media video to form an outbreak node.

[0141] Confidence of the first source It can be calculated as follows: =0.35 × score for leading publication time + 0.25 × score for originality of content + 0.20 × score for credibility of source + 0.10 × score for consistency of IP or region + 0.10 × score for subsequent dissemination and citation.

[0142] Among them, the publication time priority score indicates the publication time of the candidate work relative to other similar works; the content originality score excludes obviously reprinted, copied, or meaningless compilation content; the source credibility score is generated by account level and authentication information; the IP or regional consistency score determines whether the work is consistent with the location of the event or the location of the subject; and the subsequent dissemination and citation score determines whether subsequent works reuse or cite the content of the candidate work. A value greater than or equal to 80 can be used as a high-confidence initial source; a value between 60 and 80 can be used as a candidate initial source; and a value less than 60 can only be used as an early propagation node.

[0143] Step 12: Constructing the propagation chain.

[0144] like Figure 7 As shown, the system uses accounts and their representative works as propagation nodes. Input includes event-related works, similar clusters, publication time, account weight, content similarity between works, interaction volume, propagation volume, and key node results. Duplicate records with the same nickname or the same account are merged into one node, and the earliest event-related work published by that account is retained as the time reference.

[0145] Each work is processed from earliest to latest publication time, with only works published earlier considered as upstream candidates. The title, body text, and... are calculated for both the upstream candidates and the current work. The similarity is considered. When the similarity is greater than or equal to 0.60, a directed propagation edge is generated from the earlier account to the later account. When the similarity is greater than or equal to 0.85, it is regarded as a copy, reprint or press release relationship, and is first classified into the same propagation cluster with the earliest published work as the cluster representative.

[0146] When a late-stage node has multiple upstream candidates, they are first sorted by similarity from high to low. If the similarity is the same, they are sorted by publication time interval from shortest to longest, retaining a maximum of 3 upstream edges; downstream edges are also retained at a maximum of 3. Works with the highest global first release, the highest dissemination or interaction volume, accounts with a weight greater than or equal to 2 and with a dissemination relationship, and key nodes are retained. Paths are supplemented between representatives of dissemination clusters according to chronological order and an inter-cluster similarity threshold of 0.50. Output includes node identifier, work identifier, node role, start and end nodes of the edge, similarity, chronological order, and judgment criteria.

[0147] The processing logic includes: comparing the content similarity of works in chronological order of publication; when the similarity between a later work and an earlier work exceeds a first threshold, generating a propagation edge from the earlier account to the later account; when the similarity exceeds a second threshold, they can be grouped into the same propagation cluster, retaining only the earliest publishing account as the cluster representative; the system filters key nodes, retaining those with higher weight, propagation relationships, global sources, or propagation interactions. The nodes are configured such that the number of upstream and downstream edges for each node is capped to avoid an overly dense propagation graph. For the central node, inter-cluster propagation relationships can be supplemented based on publication time and semantic similarity.

[0148] The output format is a propagation path diagram, for example: { " ": " ", " ": [ {" ": " ", " ": " ", " ": " "}, {" ": " ", " ": " ", " ": " "}, {" ": " ", " ": " ", " ": " "} ], " ": [ {" ": " ", " ": " ", " ": 0.72, " ": }, {" ": " ", " ": " ", " ": 0.58, " ": } ] }".

[0149] In one embodiment, based on the content similarity greater than 0.6 and the order of publication, a propagation path is generated from the initial user to the short video account and then to the media account, and the media account is marked as an authoritative intervention node.

[0150] Table 9 shows the parameters for constructing the propagation link, which are used to illustrate the propagation edge threshold, inter-cluster relationship threshold, upper limit of the number of nodes and edges, and node retention conditions. Table 9 Parameters for Propagation Link Construction When generating propagation edges, earlier works must be published before later works, and their content similarity must be greater than or equal to the first similarity threshold. If two works are similar to the second similarity threshold, they are considered to be copies, reposts, or porters within the same propagation cluster. Instead of directly expanding with too many duplicate edges, the earliest published work or the work with the highest propagation volume is retained as the cluster representative. Similarity relationships are not propagated infinitely by default. and resemblance, and Similarity does not necessarily lead to and Similarity is supplemented only between cluster representatives according to time order and inter-cluster similarity to avoid chain-like erroneous merging.

[0151] Step 13: Event analysis report generation and work order push.

[0152] Using the event identifier as the primary key, the system reads the event name, description, category, lifecycle, and associated works from the event database, and associates them with risk classification results, initial source, key nodes, propagation links, platform distribution, volume trend, interaction trend, feedback information focus, processing suggestions, and work order push configuration to generate a report data package.

[0153] The system deduplicates related works by time, platform, and similar clusters, calculating total volume, number of accounts, dissemination volume, interaction volume, platform distribution, and trend peaks. It generates an event timeline based on key moments, incorporating evidence of works corresponding to initial release, spread, outbreak, authoritative intervention, and turning points into the report. The large language model generates an event overview, dissemination characteristics, discussion focus, event summary, and handling suggestions solely based on the above structured data. Each conclusion is associated with a representative work identifier to avoid generating conclusions detached from the input data.

[0154] After the report is assembled, the rules engine reads the overall switch for work order push, feedback information type, risk level, interaction threshold, propagation threshold, and the AND or OR relationships between conditions. When an event meets the configuration, it queries existing work orders by event identifier: if no work order exists, a new work order is created; if the work order is pending or being processed, the level, handler, and update time are updated; if the work order has been rejected or completed, it is skipped according to the configuration and automatically created with the reason recorded.

[0155] Output a structured event analysis report, report generation time, list of evidence works, as well as work order actions, work order levels, identification rules, handlers, and execution status. If report or work order generation fails, it will retry according to the configuration; after exceeding the maximum number of retries, an alarm and the reason for failure will be recorded.

[0156] The processing logic includes: assembling the core overview of the event, core dissemination data, event development timeline, dissemination characteristics, discussion focus, event summary, and handling suggestions according to the report template; automatically creating or updating a work order if the event level, event type, interaction volume, or playback volume meets the work order push rules; updating the work order level and handler if there is already a work order being processed; and skipping or prompting manual handling according to the configuration if the work order has been rejected or completed.

[0157] Output formats include structured reports and work order records, for example: { " ": " ", " ": [ "Key Overview of the Incident" "Core Communication Data" "The Development Context of Feedback Information" "Characteristics of event propagation" "Feedback and Discussion Focus", Incident Summary and Handling Recommendations ], " ": " ", " Level 3 }".

[0158] In one embodiment, an event report is generated for Level 3 sensitive early warning events, and a processing work order is automatically created based on the configuration of Level 3 and Level 4 events with more than 1,000 interactions or more than 10,000 plays.

[0159] It should be noted that event clustering can be achieved through semantic clustering using a large language model, vector similarity clustering, rule-based clustering, or multi-model fusion. The callback rate threshold, lifecycle days, check update time window, social platform weight, risk classification weight, and propagation link similarity threshold can all be configured. Event renewal can also be confirmed manually by the user, automatically, or through human-machine collaboration. The report generation template can be replaced or expanded according to different scenarios such as brand, product, activity, person, and event.

[0160] Specifically, 120 new social media posts from a certain car brand between June 20, 2026 and June 21, 2026 were accessed, including titles, text, or... There were 120 candidate works in the text that mentioned the brand and had a negative sentiment. Event elements were extracted from the candidate works to obtain several candidate events, including "delivery delay disputes", "accidental triggering of assisted driving disputes", and "in-vehicle infotainment system lag feedback".

[0161] Compare candidate events with existing ones Event tagging was performed. The existing event database already contained the term "delivery delay dispute for a certain vehicle model." 72 entries were successfully tagged back to this event or other existing events. The remaining 48 entries were tagged as... Therefore, the daily event callback rate was... If the new event discovery rate is below 70% of the threshold and the number of candidate works is greater than 20, the new event discovery process is triggered.

[0162] Event element extraction and semantic clustering were re-performed on the 48 unmatched works, and it was found that 31 of them involved core content such as "false triggering of assisted driving", "abnormal emergency braking", and "false alarm in high-speed scene". The clustering score reached 0.82, exceeding the clustering threshold of 0.75, therefore a new event, "Dispute over False Triggering of Assisted Driving System for a Certain Vehicle Model," was generated. This new event was then written to... A second round of testing was conducted on 48 unmatched works, of which 31 were matched with new events. Updated to , Updated to .

[0163] In subsequent lifecycle management, the existing "delivery delay dispute for a certain vehicle model" entered an expired status after 14 days. During the check and update process, based on entity words in the event name, event description, and high-frequency characteristic words such as "delivery delay," "vehicle delivery delay," and "after-sales communication," a search was conducted for newly added works in the past 14 days, revealing 6 relevant works. All are greater than 0.75, therefore the event status is changed from... Updated to And recalculate the 14-day life cycle.

[0164] Regarding the newly discovered "controversy over erroneous triggering of driver assistance system in a certain vehicle model," further risk grading will be implemented. The event's qualitative level is 3, the account level is 4, the interaction level is 3, the dissemination level is 2, and the platform weight is 1.2. =1.2×(3×50%+) (3,2)×30%+4×20%)=3.84, which is mapped to a Level 4 high-impact event. Record the classification criteria and automatically create feedback information handling work orders based on the work order push configuration.

[0165] Perform propagation chain analysis to identify the earliest published and The system identifies 82 works from a certain social media platform as the original source, identifies media video works with interaction volumes exceeding the sliding window mean plus three standard deviations as breakout nodes, and identifies works where a media account first intervenes as authoritative intervention nodes. Based on the rule of content similarity greater than 0.60 and sequential publication time, propagation edges are generated, and reposted works with similarity greater than 0.85 are formed into propagation clusters. The output is an event propagation chain diagram containing the source account, breakout nodes, authoritative nodes, and major propagation edges. Based on the above results, the system generates a core event overview, core propagation data, feedback information development, propagation characteristics, content focus points, event summary, and handling suggestions.

[0166] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 3As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores static and dynamic information data. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in the above method embodiments.

[0167] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the computer device to which the present invention is applied. A specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0168] In addition, the present invention also provides a computer program product, on which a computer program is stored, wherein when the computer program is executed by a processor, the steps in the embodiments of the method are implemented.

[0169] In addition, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0170] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0171] This invention is not limited to the structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this invention is limited only by the appended claims.

Claims

1. A method for incremental maintenance and analysis of an event database, characterized in that, include: The standardized work data is identified and matched to select a set of candidate feedback information works; Extract event elements from the collection of candidate feedback information works to generate candidate event objects; Candidate event objects are grouped, and clustered within each group based on the similarity of candidate events to generate an event library; Based on the event database, event tags are re-applied to incremental works, and the matching score between the works and events is calculated. Based on the matching results, the works are marked as matched or unmatched. Get the event callback rate within the current period. When the event callback rate is lower than the preset threshold and the total number of candidate feedback works reaches the minimum sample size, the event elements of the unmatched works are re-extracted and clustered, the new events are added to the event library, and the unmatched works are re-called. Set a lifecycle for each event, mark events that have exceeded their lifecycle as expired, search for new works for expired events based on keywords and entity words, and if relevant new works exist, renew the lifecycle and recalculate the lifecycle. Cross-platform similarity aggregation of related works is performed to merge works from the same source; a risk assessment is conducted on the aggregated works to determine the risk level; The dissemination of related works is traced to identify the initial source and key nodes. A directed dissemination chain with accounts as nodes is constructed based on content similarity and release time. Event analysis reports or handling work orders are generated based on risk level and dissemination chain.

2. The event database incremental maintenance and analysis method according to claim 1, characterized in that, The process of identifying and matching standardized work data to filter out a set of candidate feedback information works includes: It receives content data from different social media platforms, unifies heterogeneous fields across platforms based on field mapping configurations, and retains abnormal original fields; it performs basic text cleaning, including encoding regularization, webpage marking and dirty character cleanup, full-width and half-width character unification, whitespace compression, and duplicate topic removal, and simultaneously completes text null value marking, generation of excessively long text copies, and handling of garbled characters and abnormal forwarding prefix annotations; it performs optical character recognition on video materials by extracting frame images according to screen nodes, and directly reuses existing qualified recognition content; The system retrieves the account type dictionary to complete the unified coding of various types of accounts and records the account level synchronously. It converts the original values ​​of the number of users following, the number of interactions, and the number of dissemination on the platform into cross-platform comparable standardized values. It integrates all the processed information, including field mapping, text cleaning, video recognition, account coding, and value normalization, to generate standardized work data. Standardized works are pre-verified, including whether the work can be associated with the subscription subject and whether the work contains valid text content. Works that fail the pre-verification are directly excluded and will not be included in the subsequent feedback information judgment. Emotion recognition is performed on standardized works that have passed the verification. The statements associated with the subscribers are used as the basis for judgment. The emotional tendency, negative judgment reasons and emotional confidence are output. Works that meet the negative emotional trigger are distinguished according to the preset threshold. Works with insufficient confidence or inconsistent positive and negative judgment results are marked as pending review. Standardized works are verified based on a configurable feedback information classification system. If the semantics of a work match any first- or second-level feedback category within the feedback information classification system, it is determined that the feedback category triggering condition is met. Search for risky words and their corresponding synonyms in the text of the work, and simultaneously check whether the work's interaction and dissemination indicators meet the corresponding standards of the platform. If both the word matching and indicator standards are met, it is determined that the combination of keyword and interaction dissemination threshold trigger conditions is met. Perform content exclusion checks to identify and remove advertising content, content with no subject and noisy content, and standardized works that are repeatedly collected records; For works that are not eliminated and meet any type of feedback information triggering condition, enter the candidate identifier, triggering cause, evidence text, and threshold judgment result, summarize all standardized works carrying candidate identifiers, and output a set of candidate feedback information works.

3. The event database incremental maintenance and analysis method according to claim 1, characterized in that, The step of extracting event elements from the candidate feedback information work set to generate candidate event objects includes: The event judgment text is formed by reading the title, body, description, topic, optical character recognition text, subscription subject, subject entity, negative reasons, triggering evidence and feedback information classification system of the candidate feedback information works and integrating them in the order of title, body, description, topic and optical character recognition text. The event judgment text is processed by subject aliasing and identification of entity information such as time, place, person, organization, and product name in the text. It is determined whether the text has a clear event relationship where the subject has implemented or encountered a certain fact and a cognitive difference, change of situation, or disposal result has occurred. For works that only have broad evaluations, lack factual process, or whose content is unrelated to the subscriber, it is determined that there is no clear feedback event. When a clear feedback event is determined, the event subject, event behavior, time, location, result, and evidence fragments are extracted based on the structured field template. The first-level and second-level event types are determined according to the feedback information classification system. The standardized subject, core behavior, and cognitive disagreement result are combined to generate the event name. The factual elements are integrated in chronological order to form the event description. Core content tags are extracted from the behavior, result, and cognitive disagreement point. The extracted content undergoes compliance verification, checking whether the event name includes the subject and core behavior, whether the event description can be traced back to the original evidence, and whether the event type conflicts with the triggering reason. Works that fail the verification are marked as pending review or marked as having no events, while those that pass the verification are generated as candidate event objects.

4. The event database incremental maintenance and analysis method according to claim 1, characterized in that, The step of grouping candidate event objects and clustering them within each group based on the similarity of the candidate events to generate an event library includes: Candidate groups are constructed using the secondary classification of feedback information, standardized subject identifiers, and event types as grouping keys. Candidate events with consistent secondary classification of feedback information and no conflicting subjects are grouped into the same comparison set, thus narrowing the scope of semantic comparison. Within each group, semantic representations are generated for event names, event descriptions, and core content tags. Event clustering similarity scores are calculated by combining subject consistency, event type consistency, keyword overlap, semantic similarity of event descriptions, and time windows. Candidate events are compared with representative events of existing event clusters within the group based on their publication time. When the similarity score is within the first preset threshold range, the event is assigned to the event cluster with the highest score. When the similarity score is within the second preset threshold range, a deep verification is initiated to check whether the time, location, factual behavior, and results are consistent. Merging can only proceed if the verification is passed. When the similarity score is less than the preset threshold or the factual elements are mutually exclusive, the candidate event is retained as an independent event. After merging, the subject, event type, core tags, and time range corresponding to the event cluster are re-summarized to form the event cluster feature information. Candidate events are categorized as follows: those with clear time, location, subjects, and facts are categorized as specific events; and those lacking time and location but consistently reflecting similar issues are categorized as summary events. The clustering results are organized into an event library for initial tag matching of incremental works. If the event library coverage is insufficient in the initial matching results, the unmatched works are grouped and clustered again, and the unmatched works are tag matched again after the event library is updated.

5. The event database incremental maintenance and analysis method according to claim 1, characterized in that, The process of tagging incremental works based on the event database, calculating the matching score between the work and the event, and marking the work as matched or unmatched according to the matching result includes: For each incremental work, the matching scores of the event name, main entity, core keywords, semantic similarity of the event description, publication time window, and consistency of feedback information classification are combined to determine whether the work belongs to a known event in the event database; the scores of each dimension are integrated according to preset weights to calculate the event matching score between the incremental work and the known event. The event matching score is compared with a preset threshold. If the event matching score reaches the preset threshold, the match is considered successful, and the event identifier, event name, event matching score, and matching basis are written to the work. If the work matches multiple events and reaches the preset threshold, the event with the highest matching score is selected. When the difference between the highest and second highest scores meets the preset difference condition, the event with the exact same subject entity and more refined event type hierarchy is selected first. If the attribution cannot be determined, the work is marked as pending manual review. Works that fail to match any known event are marked as having no matching event.

6. The event database incremental maintenance and analysis method according to claim 1, characterized in that, The process of obtaining the event callback rate within the current period, and when the event callback rate is lower than a preset threshold and the total number of candidate feedback works reaches the minimum sample size, involves re-extracting and clustering event elements for unmatched works, adding the new events to the event library, and performing a second callback on the unmatched works, including: During the periodic incremental update process, the event callback rate is calculated. The input includes the total number of works containing newly added candidate feedback information in the period, the number of works that successfully matched existing events, the number of works that did not match events, and a preset threshold. The event callback rate is calculated as the ratio of the number of successfully matched works to the total number of works containing feedback events in the period. The event callback rate is compared with a preset threshold. If the event callback rate is lower than the preset threshold, it is determined that the current event library does not cover the newly added works. If the number of newly added candidate feedback information works in a period does not reach the minimum sample size, the callback rate judgment is skipped to avoid false triggering caused by small samples. In scenarios with small sample sizes, only observation logs are retained; in scenarios with abnormal traffic on a single platform, similar works are deduplicated; in scenarios with old event expression deviation, the event renewal or extended matching process is initiated; new events to be generated need to be compared again with various events in the event library; in scenarios with concentrated noise, noisy content is removed first. The system performs new event identification processing on the collection of works with unmatched events, extracting event elements from the unmatched works; performs semantic clustering according to the feedback information; compares the clustering results with existing events in the event database to prevent duplicate event creation; and generates a corresponding event identifier and adds it to the event database after confirming that it is a completely new event. Based on the updated event library, perform a secondary event tag rollback on works that did not match successfully.

7. The event database incremental maintenance and analysis method according to claim 1, characterized in that, The process of setting a lifecycle for each event, marking events that have exceeded their lifecycle as expired, and retrieving new works for expired events based on keywords and entity words, including renewing the lifecycle and recalculating it if relevant new works exist, involves: The event lifecycle is calculated from the moment the event is generated through clustering. If the current moment exceeds the event generation time plus the lifecycle duration, the event is marked as expired and removed from the active event set in the event library. At the same time, the event's historical data and the relationship between the event and the work are retained. Perform manual or system-initiated review and update operations on expired events; During the review and update process, new works are retrieved within a specified time range based on event keywords, event descriptions, entity words in the event name, and high-frequency feature words that rank highly, and it is determined whether the retrieved new works are related to the expired event. If a new work with a related relationship is found, the event status will be adjusted to "renewed", and the event lifecycle will be recalculated from the time the renewal takes effect.

8. The event database incremental maintenance and analysis method according to claim 1, characterized in that, The method of cross-platform similarity aggregation of related works to merge works of the same origin includes: Remove topic tags from works related to the event and perform a rough screening to obtain a set of candidate similar works, and exclude works whose text conditions do not meet the specifications. The candidate similar works set is screened and judged in multiple dimensions, and differentiated comparison rules are set for text and image works and video works with optical character recognition text. The structural compatibility between works is determined based on field adaptation tags, and the optical character recognition text comprehensive verification index is obtained for video works. A dual-gating mechanism is formed based on the structural compatibility conditions and the optical character recognition text verification index to comprehensively determine the similarity of works. Based on the similarity determination results, mutually exclusive similar clusters are constructed, and at most two rounds of missing similarity relation compensation are performed. A transitive similarity attenuation rule is introduced to truncate excessive associations. After two rounds of supplementary calculations, inter-cluster conflict verification is performed, and multiple characteristic indicators are combined to determine whether to perform cluster merging; and the similar cluster structure is updated and the total amount of cluster-level propagation is recalculated based on the merging results. Based on the updated similar cluster structure, representative works of similar clusters are identified and snapshots are saved according to the scale of dissemination and publication time.

9. The event database incremental maintenance and analysis method according to claim 8, characterized in that, The multi-dimensional screening and determination of the candidate similar works set includes: The works in the candidate similar works set are paired up to form candidate work pairs. For each candidate work pair, the text length ratio, term overlap ratio and longest common subsequence ratio are calculated. The corresponding preset threshold is matched according to the text length of the candidate work pair. The title and body of the candidate works are checked one by one for all applicable dimensions. The basic text dimension check is deemed to have passed only when all dimensions meet the corresponding threshold conditions. Calculate the overall similarity ranking score for candidate work pairs. The overall similarity ranking score is used only for ranking candidate work pairs and is not used as the basis for determining the similarity of candidate work pairs. If a candidate work pair contains video works with valid optical character recognition text, then additionally calculate multiple verification indicators corresponding to the optical character recognition text dimension of the candidate work pair. The optical character recognition text of candidate works is fused using a minimum value constraint mechanism. The optical character recognition text dimension verification is only passed when all basic dimension indicators meet the standards. If any basic dimension fails to meet the standards, the text cannot pass the compensation of the other high-scoring dimensions.

10. The event database incremental maintenance and analysis method according to claim 8, characterized in that, The differentiated comparison rules for graphic works and video works with optical character recognition text include: Identify the work type and the fill status of each text field in each pair of candidate works; If the candidate works pair is a video combination without valid optical character recognition text, then similarity comparison will only be performed if both candidate works pair have no optical character recognition text and the total length of the title and body text after removing the topic meets the standard. If the title or body field of a candidate work is empty after removing the topic tag, it will only be compared with objects whose corresponding fields are also empty, and the empty field dimension will be directly determined to pass the verification. If both the title and body of the candidate works are empty but there is valid optical character recognition text, then the similarity comparison of the candidate works will be determined only based on the optical character recognition text dimension. For candidate works that use optical character recognition (OCR) text comparison, if the title or body text dimension does not meet the standard, the work cannot be released as compensation through the OCR text dimension meeting the standard.

11. The event database incremental maintenance and analysis method according to claim 1, characterized in that, The risk assessment of the aggregated works to determine the risk level includes: The system collects and aggregates eight assessment data points: qualitative level, account level, interaction level, dissemination level, platform weight, repost-like ratio, historical risk level, and data update time. When conducting event risk classification, the system performs the same risk calculation on the aggregated works associated with the event, selects the highest risk level of the valid works associated with the event as the current risk level of the event, and retains the corresponding work information and calculation basis. Qualitative, account, and quantitative levels are determined according to a preset calculation order. Qualitative levels one to four are obtained based on event semantics, content classification, risk level, and negative reasons. Account levels one to four are obtained based on account attribute information. Interaction volume and dissemination volume are matched with the historical percentile thresholds of the corresponding platform and time window to obtain interaction level and dissemination level. The highest value of interaction level and dissemination level is taken as the quantitative level. If the qualitative level is level four, high-risk scenarios are identified and judgment evidence and judgment reasons are output. The risk level is directly determined to be level four, and the comprehensive score weighted calculation is not performed. For aggregated works whose qualitative level is not level four, the comprehensive risk score is calculated based on the fixed dimension weight ratio, and the risk levels from level one to level four are mapped according to the score range of the comprehensive risk score; when an account matches multiple rules, the highest level is taken, and the interaction level and dissemination level are determined based on the platform percentile threshold updated periodically by offline tasks. Save qualitative judgment evidence, account type, interaction percentile, dissemination percentile, platform weight, weighted comprehensive score, final risk level, and calculation time. The work-level risk results are summarized to form the event-level risk results, and the risk recalculation and result tracing are supported after data updates. Monitor the interaction volume, dissemination volume, likes, reposts, and account attributes of the aggregated works; after the interaction volume, dissemination volume, likes, reposts, and account attributes are updated, recalculate the interaction level and dissemination level, update the quantitative level and calculate the comprehensive risk score; obtain the content threshold level by matching the viral risk threshold based on the repost-like ratio, and select the highest value of the weighted recalculated level and the content threshold level as the candidate risk level. Compare candidate risk levels with historical risk levels. If the candidate risk level is the highest, update the risk level and trigger the work order generation rule; otherwise, retain the historical risk level. If the event is in an active or renewed lifecycle, the risk level can only be maintained or increased, and will not be automatically downgraded with short-term data fluctuations. Risk level downgrades can only be performed when the event cycle is updated, rules are reconfigured, or manual review and confirmation is required. When the risk level is upgraded, record the update time, the level before and after the upgrade, the platform percentile threshold, the ratio of forwards to likes, and the reason for the upgrade.

12. The event database incremental maintenance and analysis method according to claim 1, characterized in that, The process of tracing the dissemination of related works to identify the initial source and key nodes, constructing a directed dissemination chain with accounts as nodes based on content similarity and publication time, and generating an event analysis report or handling work order based on risk level and dissemination chain includes: Obtain a list of works related to the event, similar clusters, publication time, platform, account attributes, interaction volume, dissemination volume, content similarity, event tags, and account weight, and trace the source of dissemination; merge reprinted works from the same source according to similar clusters, and select the earliest published works in the similar clusters as first release candidates; verify multiple dimensions of information of the first release candidates and calculate the confidence level, and classify high-confidence first release sources, candidate first release sources, and early dissemination nodes according to the confidence level; After determining the original source, identify diffusion nodes based on time windows, outbreak nodes based on sliding windows, and authoritative intervention nodes based on changes in account weight. Identify event turning points in the surrounding works corresponding to diffusion nodes, outbreak nodes, and authoritative intervention nodes, and output node type, work identifier, node time, judgment value, evidence, and confidence level. Using accounts and corresponding works as dissemination nodes, input the event-related works, similar clusters, publication time, account weight, content similarity, interaction volume, dissemination volume, and key node identification results; merge duplicate records of the same account, and retain the earliest event-related work published by the account as the time benchmark; Works related to the event are processed from earliest to latest publication time, and works with earlier publication time are selected as upstream candidates; the comprehensive similarity of title, body text and optical character recognition text is calculated; if the similarity reaches the first threshold, a directed propagation edge is generated; if the similarity reaches the second threshold, they are grouped into the same propagation cluster, and the earliest published work is selected as the cluster representative; When multiple upstream candidates exist for late-stage nodes, they are sorted according to predetermined rules, and the number of upstream and downstream edges is set to an upper limit; the first globally published works, the works with the highest dissemination data, works associated with high-weight accounts, and key nodes are selected and retained; based on the time order and the inter-cluster similarity threshold, the propagation path is supplemented between cluster representatives, and all judgment information related to nodes and edges is output; The system reads basic event information using the event identifier as the primary key, associates it with risk classification results, initial source, key nodes, propagation links, platform distribution, trend data, handling suggestions, and work order push configuration, and assembles a report data package; it also calculates statistical indicators after deduplicating the works associated with the event, sorts out the event development timeline according to the node sequence, and loads the work evidence corresponding to each node. Generate an event overview, propagation characteristics, discussion focus, event summary, and handling suggestions based on the structured data of the data packets, and bind each conclusion to a representative work identifier; read the work order push configuration, query existing handling work orders when the event meets the configuration conditions, perform new creation, update, or skip operations according to the different status of the work order, and retain the records; Output a structured event analysis report, report generation time, list of evidence works, and work order information; if the report generation or work order generation operation fails, it will retry according to the configuration, and if the number of retries exceeds the limit, it will record alarm information and the reason for failure.

13. An incremental maintenance and analysis system for an event database, characterized in that, include: The initial event clustering module is used to identify and match standardized work data and filter out a set of candidate feedback information works; Extract event elements from the collection of candidate feedback information works to generate candidate event objects; Candidate event objects are grouped, and clustered within each group based on the similarity of candidate events to generate an event library; The incremental tag re-call module is used to re-call event tags for incremental works based on the event library, calculate the matching score between the work and the event, and mark the work as matched or unmatched according to the matching result; The new event iteration and completion module is used to obtain the event callback rate in the current period. When the event callback rate is lower than the preset threshold and the total number of candidate feedback information works reaches the minimum sample size, the event elements of the unmatched works are re-extracted and clustered, the new events are added to the event library, and the unmatched works are re-called. The event lifecycle management module is used to set a lifecycle for each event, mark events that have exceeded their lifecycle as expired, and search for new works for expired events based on keywords and entity words. If relevant new works exist, the lifecycle is renewed and recalculated. The event in-depth analysis output module is used to aggregate similar works related to an event across platforms to merge works from the same source; and to conduct a risk assessment on the aggregated works to determine the risk level. The dissemination of related works is traced to identify the initial source and key nodes. A directed dissemination chain with accounts as nodes is constructed based on content similarity and release time. Event analysis reports or handling work orders are generated based on risk level and dissemination chain.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 12.

15. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the method according to any one of claims 1 to 12.

Images