An AI public opinion system processing method

Abstract

This invention discloses a processing method for an AI-powered public opinion system, specifically relating to the field of network management technology, addressing the problems of evidence visibility bias and difficulty in verifying authenticity caused by recommendation distribution and filtering bubbles. The method acquires multi-source public opinion records and clusters them according to semantics, time, and propagation links to form event and controversial elements; it statistically analyzes the entry point and group exposure ratio, constructs bubble profiles, and maps filtering bubble values ​​to propagation parameters; it retrieves evidence around controversial elements, constructs an evidence graph, and filters evidence subsets; it constructs counterfactual profiles based on observation bubble profiles, normalizes them, redistributes evidence weights and propagation impact weights, infers observation / counterfactual scores, takes the minimum value to obtain the bubble invariant authenticity score, and extracts bubble sensitive factors through ablation; under budget and time constraints, it solves for the minimum cost set of intervention actions and execution sequences, and records a snapshot of the recurrence.

Description

Technical Field

[0001] This invention relates to the field of network management technology, and more specifically, to a method for processing AI public opinion systems. Background Technology

[0002] Against the backdrop of rapid spread and cross-platform diffusion of online public opinion, false information, one-sided narratives, and out-of-context content are easily amplified through recommendation and distribution mechanisms, leading to problems such as "high noise in evidence, strong opposition of viewpoints, and difficulty in quickly verifying authenticity" in event assessment. To improve the automation and reliability of public opinion event authenticity assessment, existing research has begun to introduce technologies such as knowledge enhancement, question-answering evidence retrieval, and counterfactual reasoning to construct evidence chains from massive amounts of information and verify facts.

[0003] Existing literature 1 (Research on Fact Verification Methods Based on Knowledge-Enhanced Question Answering and Topic-Guided Counterfactual Reasoning, 2025) proposes automated fact verification for online misinformation, addressing issues such as insufficient semantic association mining, difficulty in filtering fine-grained noise, and the impact of dataset bias on accuracy. It also points out that traditional coarse-grained graph construction introduces noise and struggles to handle data bias. However, this type of research mainly focuses on the fact verification task itself, namely "claim-evidence," without considering the differences in filter bubbles in the context of public opinion recommendation, the evidence visibility bias caused by multi-entry exposure structures and propagation links, or providing a cost-optimization solution framework for the "intervention action set / execution sequence" for governance.

[0004] Existing public literature 2 (Research on the Network Public Opinion Interaction and Propagation Model and Governance Strategy Considering the Influence of Filter Bubbles, 2024) points out that in the context of the prevalence of recommendation algorithms on social network platforms and the formation of filter bubbles, different users are exposed to different information, and filter bubbles affect the flow and interaction patterns of various public opinion information. This literature constructs a network public opinion interaction and propagation model considering the influence of filter bubbles by quantifying the correlation between filter bubble values ​​and public opinion events, and further introduces multi-agent intervention. By increasing the probability of exiting the propagation and shifting attention to reduce the entry rate, it achieves control over the overall popularity and propagation relaxation time, such as... Figure 1 As shown. However, this type of research focuses on the evolution of dissemination and the evaluation of the effects of macro-interventions. It lacks the construction of candidate evidence sets and evidence graph reasoning for individual points of contention in an event. It cannot output a "bubble-invariant authenticity score", and it is difficult to identify the key evidence nodes or dissemination edges that are most sensitive to authenticity / risk. Furthermore, it cannot provide the "minimum set of intervention actions and execution sequence with the lowest cost" under budget and time constraints.

[0005] Therefore, there is an urgent need in this field for a method that can achieve robust assessment of the authenticity of public opinion events and output the most cost-effective and precise intervention strategy under the influence of filtering bubbles and propagation bias. Summary of the Invention

[0006] To overcome the aforementioned deficiencies in the prior art, this invention provides a processing method for an AI public opinion system. This method involves constructing an evidence graph under the conditions of filtering bubbles and propagation bias, performing counterfactual collaborative inference, obtaining bubble invariant authenticity scores and sensitivity factors, and solving for the minimum set of intervention actions and execution sequence under budget and time constraints, thereby addressing the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A method for processing public opinion using an AI system includes the following steps:

[0009] Step 1: Obtain multi-source public opinion records and cluster events to obtain event identifiers, event text sets, and disputed element sets;

[0010] Step 2: Generate topic vectors based on event text sets and count the exposure ratio of entry type and user group. Calculate bubble profile indicators and map them to obtain the filter bubble value and event relevance. At the same time, calculate entry rate, exit probability and relaxation time based on the statistics related to new reach and dissemination attenuation. Obtain candidate evidence set based on event identifier and construct evidence graph. Filter candidate evidence subsets according to the coverage threshold of disputed elements and the source type threshold.

[0011] Step 3: Execute the bubble evidence counterfactual collaborative inference algorithm. Construct a counterfactual bubble profile based on the bubble profile and normalize it. Under the observation and counterfactual bubble profile, redistribute the evidence weight and propagation influence weight of the candidate evidence subset and evidence graph respectively and infer. The redistribution is based on the bubble parameters. Output the observation authenticity score and the counterfactual authenticity score set. Take the minimum value as the bubble invariance authenticity score and set the evidence node or propagation edge to zero and ablate it. Extract the bubble sensitivity factor set according to the score decrease amount.

[0012] Step 4: Under budget and time constraints, solve the minimum set of intervention actions and execution sequence that minimizes action costs and reduces risk indicators to the threshold by using the bubble invariance authenticity score and bubble sensitivity factor set;

[0013] Step 5: Output the minimum set of intervention actions and execution sequence, and record the event identifiers, bubble parameters, candidate evidence subsets, scoring results and bubble sensitivity factor set used for reproduction.

[0014] As a further aspect of the present invention, the event clustering in step one simultaneously satisfies the semantic similarity threshold, the temporal proximity threshold, and the propagation link proximity threshold; the output set of disputed elements includes at least element type, element field, and trigger fragment, and the element field is used for subsequent evidence retrieval and consistency verification.

[0015] As a further embodiment of the present invention, in step two, the topic vector Vtopic is obtained by linearly fusing the topic structure vector T and the text center vector Vtext with a fusion coefficient α, and α is between 0.3 and 0.8 to balance structural consistency and semantic center.

[0016] As a further aspect of the present invention, in step two, the exposure ratio of entry type and the exposure ratio of user group are preferentially obtained by using the platform exposure count (impression). If the platform does not provide this information, it is obtained by log feedback or sampling estimation, and the exposure criteria of different platforms are normalized.

[0017] As a further aspect of the present invention, the candidate evidence subset selection in step two simultaneously satisfies the disputed element coverage threshold θcov and the source type threshold θsrc, wherein θcov is used to constrain the coverage of disputed elements by the evidence, and θsrc is used to constrain the credibility of the evidence source.

[0018] As a further aspect of the present invention, in step three, one or more counterfactual bubble images are constructed based on the observed bubble image. The counterfactual bubble image is obtained by adjusting one or more bubble image components related to the visibility bias, and the observed bubble image and the counterfactual bubble image are normalized and then used for weight modulation.

[0019] As a further aspect of the present invention, in step three, the redistribution of evidence weights is based on the credibility of the evidence source and the coverage of disputed elements, and modulated in conjunction with the bubble parameters related to visibility bias; the redistribution of propagation influence weights is based on the propagation intensity and cross-entry migration characteristics, and modulated in conjunction with the filter bubble values.

[0020] As a further aspect of the present invention, the solution for the minimum intervention action in step four adopts a joint optimization of budget constraints and time constraints. Under the condition of satisfying the upper limit of budget and the upper limit of time constraints, the minimum intervention action set and execution sequence that reduce the risk index R to a preset threshold Rth are output; wherein, the execution sequence satisfies that the latest effective time does not exceed the upper limit of time constraints, and the start delay of the key action does not exceed the preset start threshold.

[0021] An AI-powered public opinion system includes: a data acquisition and normalization module, which acquires and standardizes multi-source public opinion records; an event clustering module, which performs event clustering and outputs event identifiers, event text sets, and disputed element sets; a bubble parameter calculation and evidence construction module, which generates topic vectors, statistically analyzes entry types and user group exposure ratios, calculates bubble profile indicators and maps them to filter bubble values ​​and event relevance, calculates entry rate, exit probability, and relaxation time, and obtains candidate evidence sets based on event identifiers, constructs evidence graphs, and filters candidate evidence subsets; a bubble evidence counterfactual collaborative inference module, which constructs and normalizes counterfactual bubble profiles, redistributes evidence weights and dissemination impact weights under observation and counterfactual bubble profiles, and infers the results, outputting observation authenticity scores, counterfactual authenticity score sets, bubble invariant authenticity scores, and bubble sensitivity factor sets; a minimum intervention action solution module, which solves for the minimum intervention action set and execution sequence that minimizes action costs and reduces risk indicators to a threshold under budget and time constraints; and a reproduction recording module, which records event identifiers, bubble parameters, candidate evidence subsets, scoring results, and bubble sensitivity factor sets.

[0022] As a further embodiment of the present invention, the bubble evidence counterfactual collaborative inference module is configured to: construct one or more counterfactual bubble profiles based on the observed bubble profile and perform normalization processing; redistribute evidence weights and propagation influence weights to the candidate evidence subsets and evidence graphs under the observed bubble profiles and each counterfactual bubble profile, and perform evidence graph reasoning to obtain the observation authenticity score and the counterfactual authenticity score set; take the minimum value between the observation authenticity score and the counterfactual authenticity score set as the bubble invariant authenticity score; and perform zeroing ablation on the evidence nodes or propagation edges, extract the bubble sensitivity factor set based on the decrease in the bubble invariant authenticity score and output it.

[0023] The technical effects and advantages of the AI ​​public opinion system processing method of the present invention are as follows:

[0024] This invention provides a robust lower bound on truthfulness even under filtered bubbles and multi-inlet exposure bias. It achieves this by constructing an evidence heterogeneity graph, generating counterfactual bubble profiles from observed bubble profiles, and then redistributing evidence weights after normalization. s Weight w of propagation impact p Output bubble invariance authenticity score T min This addresses the problem of distorted judgment caused by the neglect of visibility bias in existing technologies.

[0025] This invention organizes evidence in a graph structure and separates and weights evidence contribution and dissemination impact. The weights are jointly modulated by calculable indicators such as credibility, element coverage, and dissemination intensity. The output simultaneously provides key evidence nodes, key links, and contribution levels, and can locate the "most sensitive factors" through ablation, making the scoring source explainable, verifiable, and locatable.

[0026] This invention abstracts the handling action into an optimization task with cost and time constraints. Based on the score and sensitivity factors, it automatically solves the action set and execution sequence with the minimum cost, so as to achieve precise intervention and priority handling of high leverage points. At the same time, it takes snapshots to solidify events, evidence graphs, weights and parameters, supports recalculation audit and effect comparison under the same conditions, and facilitates continuous iterative optimization. Attached Figure Description

[0027] Figure 1 This is a schematic diagram illustrating the technical route of existing network public opinion interaction and propagation models and governance strategies that consider the impact of filter bubbles.

[0028] Figure 2 This is a schematic diagram of the overall structure of the AI ​​public opinion system of the present invention;

[0029] Figure 3 This is a schematic diagram of the evidence diagram (heterogeneous diagram) of the present invention;

[0030] Figure 4 This is a schematic diagram of the bubble evidence counterfactual collaborative inference process of the present invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Example 1

[0033] This embodiment takes the operation of the AI ​​public opinion system of the present invention in a real emergency response scenario as an example. The AI ​​public opinion system of the present invention generates a snapshot in each scrolling window. The snapshot contains at least the event identifier, entry and group enumeration configuration, threshold and coefficient version number, candidate evidence subset, scoring result and sensitive factor set, so that consistent output can be obtained when the input public opinion records are consistent and the snapshot fields are consistent.

[0034] like Figure 2As shown, the AI ​​public opinion system of this invention consists of a data acquisition and normalization module, an event clustering module, a bubble parameter calculation and evidence construction module, a bubble evidence counterfactual collaborative inference module, a minimum intervention action solution module, and a reproduction record module. The data acquisition and normalization module obtains multi-source public opinion records from short-content platforms, social platforms, information aggregation platforms, and search aggregation portals, and maps different platform fields to a unified structure `record`. `record` at least includes a record identifier `rid`, a platform identifier `sid`, a timestamp `t`, an account identifier `uid`, an entry type `e`, a user group `g`, text content `x`, interaction counts (forwards `r`, comments `c`, likes `L`), and a dissemination association `edge`. The entry type enumeration is fixed at five categories: hot list entry, recommended entry, followed entry, search entry, and external link entry; the user group enumeration is fixed at four categories: local user group, out-of-town user group, health-conscious group, and brand fan group. The enumeration and version number are written into the reproduction snapshot field `cfg`. enum This is to constrain subsequent statistical standards.

[0035] This embodiment selects the window for a sudden public opinion event from 09:10 to 10:10 on a weekday, where online discussions focus on "a certain brand of beverage suspected of being contaminated, causing discomfort to many people." Within this window, the data acquisition and normalization module unifies the timestamps of various platforms to UTC milliseconds, and performs template noise removal, ad segment removal, URL unpacking, and reference chain parsing on the text; it performs synonym merging on entities, mapping the same brand, the same location, and the same batch number to a unified entity identifier; it performs deduplication on nearly duplicate content and retains the deduplication mapping table for backtracking. The normalization caliber is fixed by the field version number normver, and the normver is written to a snapshot to ensure that the same field definition and mapping rules are used during recalculation.

[0036] The event clustering module performs event clustering on the normalized records within a scrolling window, obtaining the event identifier `eventid`, the event text set `X` (`eventid`), and the disputed element set `Q` (`eventid`). Clustering employs a joint constraint of semantic similarity, temporal proximity, and propagation link proximity: semantic similarity is measured by the cosine similarity of the text encoding vectors, temporal proximity is measured by the record time difference, and propagation link proximity is measured by the shortest path hop count of reference or forwarding edges; when a candidate record simultaneously satisfies a similarity threshold θ... sim With time threshold θ t Or satisfy the propagation hop count threshold θ p When merged into the same event cluster, where θ sim Take 0.82, θ t Take 30 minutes, θ pTake 2 hops. The window forms the main event eventid E2026A001, the event text set X(E2026A001) converges to 18420 records, and the event cluster statistics are written to the snapshot field clusterstat.

[0037] To ensure that subsequent evidence screening and inference are aligned with verifiable points of contention, the event clustering module extracts a set of disputed elements Q (E2026A001) from the event text set. Disputed elements are represented as structured entries, each containing at least an element type (qtype), element fields (qfields), and a triggering claim fragment (qspan). The qtype covers six categories: timeline, responsible party, scope of impact, outcome, credibility of evidence source, and causal chain. Each qfield contains at least two key-value pairs for retrieval; for example, a timeline element includes the occurrence time and notification time, and a responsible party element includes the entity and attribution of responsibility. Field completeness is defined as "the number of entries in qfields with at least two keys divided by the total number of disputed element entries." In this window, the field completeness is 0.86, and this is written to the snapshot field q. stat .

[0038] In this embodiment, the bubble parameter calculation and evidence construction module receives the eventid, the event text set, and the disputed element set, and outputs the topic vector, entry point and group exposure ratio, bubble profile vector, filtered bubble value and event relevance, propagation dynamic parameters, candidate evidence set, evidence graph, and candidate evidence subset. First, a fused topic vector V is generated to simultaneously characterize the topic structure and semantic center. topic The topic structure vector T is constructed from the distribution of topic terms and the set of fields pointed to by controversial elements, and the text center vector V... text Obtained from the center of the event text vector, the two are fused according to the fusion coefficient α, which is defined as follows:

[0039]

[0040] in, The fusion coefficient ranges from 0 to 1. In this embodiment, α = 0.6 is used to ensure both structural constraints and semantic center are considered; T and V text All vectors are 768-dimensional and have the same dimensions; L2 normalization is performed before computation to avoid dimensional drift; V topic Write snapshot field v topic This is for subsequent recalculation and auditing.

[0041] For entry point and group exposure ratio statistics, the platform exposure count (impression) is used first. When the platform does not provide an exposure count, the interaction-weighted exposure proxy (expo) is used as the exposure proxy value, defined as expo = 0.5·r + 0.3·c + 0.2·L. In this formula, r, c, and L are the forward count, comment count, and like count, respectively, all of which are non-negative integers. Outliers are truncated in the data acquisition and normalization module, and the truncation ratio is recorded. After normalizing the exposure or exposure proxy of each platform according to the same scale, the exposure ratio P of each entry point type is calculated separately. entry Exposure percentage of user groups P group This window P entry For the trending topics: 0.31, recommendations: 0.27, followers: 0.19, searches: 0.15, backlinks: 0.08, P group The distribution is 0.34 for local users, 0.28 for users from other regions, 0.21 for health-related users, and 0.17 for brand followers. This distribution is written into the snapshot field p. entry With p group .

[0042] After obtaining the exposure structure, the bubble parameter calculation and evidence construction module calculates the bubble profile vector B and maps it to obtain the filtered bubble value k and the event relevance m. The bubble profile vector B consists of the homogenization index H, the diversity lower limit D, the cross-position exposure ratio C, and the entry concentration G, denoted as B=(H,D,C,G). H represents the degree of information homogenization, taking the concentration of the event text topic distribution; D represents the diversity lower limit, taking the lower bound of coverage for different source types and different position claims; C represents the cross-position exposure ratio, taking the visibility ratio of anti-position evidence in the recommendation and hot list entry points; G represents the entry concentration, taking the reverse quantization of the Pentry's normalized entropy. In this embodiment, the window calculation yields B=(0.73,0.78,0.26,0.64), and the filtered bubble value k is obtained from B through a mapping function and clipped to (0.5,2.0]. In this embodiment, linear mapping followed by clipping is used: k=clip(k0+w H ·Hw D ·Dw C ·C+w G ·G,0.5,2.0), where k0 is the bias term with a value of 1.0, w H =0.6, w D =0.4, w C =0.3, w G =0.5, clip is the clipping function; the event relevance m is obtained by weighting topic matching and popularity intensity, and in this window m=0.72. Map version number map ver Weighting coefficients and clipping boundaries are written to the snapshot field k. map With k clip To avoid discrepancies that could lead to non-reproducibility.

[0043] The propagation dynamic parameter is determined by the new reach rate r. new Attenuation rate r decay With the rate of decline r fall Statistical analysis yielded the entry rate η, exit probability μ, and relaxation time δ. new The method used to characterize the intensity of "new exposure" within a window is as follows: the window is divided according to the time slice length Δt, and the exposure amount Imp of adjacent time slices is calculated. t The increment is calculated by accumulating only positive increments as the contribution of new exposure, and then normalized using the total exposure within the window to obtain r between 0 and 1. new r decay The statistical method used to characterize the intensity of "discussion decay" within the window is as follows: the discussion volume Vol is calculated based on the same Δt. t The decrease in adjacent time slices is calculated, and only the decrease is accumulated as the decay contribution. Then, it is normalized using the discussed value within the window as a benchmark to obtain r between 0 and 1. decay r fall The statistical definition used to characterize the strength of the "relative peak decline" is: the peak discussion volume (Vol) obtained within the window. peak For each time slice, the amount of decline below the peak value is calculated and accumulated, and then normalized using the peak value as a reference to obtain r between 0 and 1. fall In the formula, Imp t For the exposure of time slice t, Vol t For the amount of discussion on time slice t, Vol peak This represents the peak number of discussions within the window.

[0044] Based on this, this embodiment uses a fixed linear combination rule to obtain the entry rate η and the exit probability μ: η is derived from r new The relevance m to the event is obtained by weighting with fixed weights κ1 and κ2, and then cropping to 0 to 1; μ is obtained by r decay (1−m) is obtained by weighting it with fixed weights λ1 and λ2, and then cropping it to 0 to 1, where κ1=0.7, κ2=0.3, λ1=0.75, and λ2=0.25. The relaxation time δ is used to characterize the time scale required for propagation to fall back to steady state. Its calculation is based on the exit probability μ and the time slice length Δt, and a lower bound protection constant 1e−6 is set for (1−μ) to ensure numerical stability. In this embodiment, Δt is taken as 0.5 hours. All coefficients and input / output values ​​are written to the reproduction snapshot field rate. coef with rate pack This is to ensure that the calculation is computable and recalculated.

[0045] The acquisition of the candidate evidence set employs a deterministic retrieval strategy: a search key is generated using the qfields of the disputed element set. The search key must include at least the fields for brand entity, location entity, batch number, handling action, and reporting agency, and is prefixed with V.topic Perform semantic expansion retrieval to form a candidate evidence set S cand Each piece of evidence must include at least the evidence identifier sid, the source type stype, and the publication timestamp t. s Evidence content text Evidence links and link hashes hash This window retrieved 268 candidate pieces of evidence, and the hash digest of the candidate evidence list was written to the snapshot field evicanddigest.

[0046] like Figure 3 As shown, the evidence graph is a heterogeneous graph. Node types include text nodes, evidence nodes, entity nodes, and topic nodes; edge types include text-text reference or forwarding edges, text-evidence similarity edges, evidence-entity inclusion edges, topic-text assignment edges, and evidence-topic association edges. For any text node and evidence node, their similarity is calculated using their vector representations. The vector dimension is 768, and L2 normalization is performed before calculation. A similarity threshold of 0.72 is used to control noisy edges. Each edge records its type, direction, timestamp, and propagation strength. Propagation strength is obtained by weighting forwarding depth, cross-platform migration counts, and coverage exposure. Statistical information such as the number of graph nodes, the number of edges, and the proportion of edge types are written into the "graph" field of the reproduction snapshot. stat .

[0047] Candidate evidence subset S sub Used to compress the size of evidence and ensure coverage and diversity before inference, and to screen for elements that simultaneously meet the coverage threshold θ of disputed elements. cov With source type threshold θ src , where θ cov Take 0.75, θ src The value is set to 0.60. Coverage(s) is calculated based on the matching hits and field consistency of the six categories of disputed elements, requiring coverage of at least five of the six categories of disputed elements, with at least one piece of evidence in each category reaching the matching threshold; the source type threshold θ src To ensure diverse and credible sources, the sources must cover at least five categories: authoritative media, government agencies, mainstream media, certified self-media, and fact-checking databases. After meeting two thresholds, redundancy is removed by time proximity and content repetition, resulting in a subset of 96 candidate evidence entries for this window. The cover(s), source type, credibility level, and hash of each piece of evidence in the candidate evidence subset are written into the evisublist field, providing definitive input for subsequent inferences.

[0048] Bubble evidence counterfactual collaborative inference module, such as Figure 4 The algorithm for counterfactual inference based on bubble evidence is shown. This module constructs three counterfactual bubble profiles B based on the observed bubble profile B. cf(1) B cf(2) B cf(3)The counterfactual construction modifies only the components related to visibility bias while maintaining dimensional consistency. Then, min-max normalization is performed on the observed profile and each counterfactual profile to obtain a set of normalized bubble parameters for weight modulation. Based on the normalization results, evidence weights w are redistributed to the candidate evidence subset and the evidence map. s Weight w of propagation impact p Weight of evidence w s The propagation impact weight w is jointly determined by the credibility of the evidence source, the coverage of disputed elements, and the visibility bias modulation term. p The propagation intensity, cross-entry migration, and filter bubble value k are jointly modulated; after normalization of the two types of weights, the evidence graph reasoning is driven, and the observation authenticity score T is output. obs With three counterfactual truth ratings T cf(1) T cf(2) T cf(3) .

[0049] Bubble Invariance Authenticity Score T min The minimum value among the four scores is taken, defined as follows:

[0050]

[0051] This window received =0.64, { }={0.59,0.71,0.66}, therefore =0.59; score package pack At least the eventid, four score values, hash of the evidence subset used in the reasoning, and version number of the weight function should be written, and the reproduction record module should write them to the reproduction snapshot field score. pack To support recalculation.

[0052] To obtain interpretable governance leverage, the bubble evidence counterfactual collaborative inference module performs zeroing ablation on evidence nodes and propagation edges to obtain the bubble invariance authenticity score after ablation, and measures the sensitivity by the score decrease ΔT.

[0053] Sensitivity contribution of evidence nodes Definition:

[0054]

[0055] in, This indicates that the evidence node s i The bubble inference score is recalculated after setting it to zero to ensure its authenticity.

[0056] Propagation edge sensitivity contribution Definition

[0057]

[0058] in, This indicates that the edge p will be propagated. j The authenticity score of the bubble obtained after recalculation is set to zero is unchanged.

[0059] Sensitive threshold θ T Take 0.03, when ΔT is not less than θ T The corresponding object will be included in the bubble sensitivity factor set F. sens This window identifies 9 sensitive evidence nodes and 7 sensitive propagation edges, which are output as a structured list and written to the senslist field of the reproduction snapshot for subsequent minimum intervention solution.

[0060] To verify the effects of counterfactual construction and bubble parameter modulation on robustness and interpretability, this embodiment performs an internal comparative evaluation on the same window data, and the results are shown in Table 1.

[0061] Table 1. Robustness and interpretability comparison results (Window E2026A001)

[0062] Observational rating only 0.64 none none none none Counterfactual but does not change with bubble parameter redistribution 0.64 0.54 / 0.69 / 0.60 0.15 0.54 6 / 4 This embodiment (with bubble parameter reallocation) 0.64 0.59 / 0.71 / 0.66 0.12 0.59 9 / 7

[0063] As shown in Table 1, the observation scores T for the three settings under the same window E2026A001 are... obs All scores were 0.64, but after introducing counterfactual factors, the counterfactual score set without bubble parameter redistribution was 0.54 / 0.69 / 0.60, lowering the minimum value and causing T to... min The score was only 0.54 with a fluctuation range of 0.15, and only 6 sensitive pieces of evidence and 4 sensitive propagation edges were identified. However, in this embodiment, after redistributing the evidence weight and propagation influence weight according to the bubble parameters, the counterfactual score set became 0.59 / 0.71 / 0.66, and the minimum value increased, causing T to... min The value was increased to 0.59 and the fluctuation range decreased to 0.12. Nine sensitive pieces of evidence and seven sensitive propagation edges could be extracted, indicating that the lower bound of authenticity is more robust under the filter bubble perturbation, and the key evidence and key links are more fully located, thus resulting in stronger robustness and interpretability.

[0064] The minimum intervention action solution module uses bubble invariance realism score T. min With the bubble-sensitive factor set F sensAs input, under budget and time constraints, find the minimum set of intervention actions that minimizes total cost and reduces the risk index below a threshold, and output the execution sequence. The risk index R ranges from 0 to 100 and is obtained by summing four sub-items with fixed weights: propagation speed, negative proportion, authenticity uncertainty, and manipulation risk. Each sub-item is first normalized to 0–1, and if it exceeds the limit, it is truncated to 0–1 according to a pruning rule, and then mapped to 0–100. The propagation speed term is based on the discussion volume Vol within the window. t The growth rate fitting results were normalized to a preset lower bound of 0 and an upper bound of 0.8; the negative proportion term is the ratio of negative text exposure to total exposure; the authenticity uncertainty term is determined by T. min Characterization of complementary quantities, T min The lower the value, the higher the uncertainty. The manipulation risk item is the proportion of the propagation intensity of the "rapid migration edge across the entry point" and "abnormal collaborative edge" in the sensitive propagation edge to the total propagation intensity of all sensitive propagation edges. The weights are fixed as follows: propagation speed item 0.25, negative proportion item 0.30, authenticity uncertainty item 0.30, manipulation risk item 0.15. The weights, growth rate upper and lower bounds, and all sub-item inputs and outputs are written into the reproduction snapshot to solidify the caliber. This window sets the initial risk R0 to 76 and the target threshold R th The budget limit is 45, the maximum cost unit is 100, the time limit is 2 hours, and key actions must be initiated within 40 minutes.

[0065] The minimum intervention action solution module maintains an action library, which includes at least the following action types: authoritative evidence delivery, clarification Q&A package delivery, abnormal collaborative link handling, unified statement announcement, diverse delivery, and manual review. Each action 'a' includes at least the action cost (cost(a), target entry type set E(a), target user group set G(a), intensity parameter (intensity(a), expected effective delay lag(a), and expected duration dur(a), and the action parameter version number 'act' is fixed. ver Write a snapshot of the reproduction. The objective is to minimize the risk indicator R under the constraints of budget and time limits. 下降 To the threshold R th Furthermore, the total cost is minimized; the solution employs a constrained search combined with a marginal contribution pruning strategy to achieve the joint constraints of minimizing the set and minimizing the cost.

[0066] In this embodiment, under window E2026A001, the minimum intervention action set A* output by the minimum intervention action solution module contains 5 items: A1 is to prioritize and distribute evidence from government agencies and authoritative media in the sensitive evidence node set across entry points; A2 is to de-weight the cross-entry rapid migration edges in the sensitive propagation edge set; A3 is to deliver clarification Q&A packages with verifiable evidence links to the elements with the largest coverage gaps in the disputed elements; A4 is to perform diversified delivery to the recommended entry points to reduce homogenization and improve cross-position visibility; A5 is to trigger a manual review work order to perform secondary verification of high-risk claims and write back to the evidence database. The corresponding execution sequence II* is A2, A1, A3, A4, A5, where A2 is executed first to suppress abnormal propagation links, then A1 improves the visibility of highly credible evidence, then A3 fills in the evidence coverage of disputed elements, A4 is used to reduce filter bubble bias and shorten relaxation time, and finally A5 performs closed-loop verification of the remaining high-risk claims. The total cost of this solution is... Furthermore, it meets the budget cap of 100, and the latest effective time max(lag(a)+dur(a)) does not exceed 2 hours, satisfying the timeliness constraint. After the action takes effect, the system recalculates T for the updated evidence weight and propagation weight in the same window. min With risk Rwithin-window, we obtain T min The risk decreases from R0=76 ​​to R'=41 when the value increases from 0.59 to 0.71, satisfying R'≤R th =45.

[0067] The reproduction recording module generates a reproduction snapshot each time the window refreshes. Each snapshot records at least the event identifier, bubble parameters, candidate evidence subset, scoring results, and bubble sensitivity factor set, and also records the necessary version number and random seed to ensure consistency in the recalculation. When the input public opinion records are consistent and the reproduction snapshot fields are consistent, the AI ​​public opinion system of this invention can recalculate event clustering, bubble parameters, evidence subset, scoring set, sensitivity factors, and action solution output.

[0068] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0069] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for processing public opinion in an AI system, characterized in that, Includes the following steps: Step 1: Obtain multi-source public opinion records and cluster events to obtain event identifiers, event text sets, and disputed element sets; Step 2: Generate topic vectors based on event text sets and count the exposure ratio of entry type and user group. Calculate bubble profile indicators and map them to obtain the filter bubble value and event relevance. At the same time, calculate entry rate, exit probability and relaxation time based on the statistics related to new reach and dissemination attenuation. Obtain candidate evidence set based on event identifier and construct evidence graph. Filter candidate evidence subsets according to the coverage threshold of disputed elements and the source type threshold. Step 3: Execute the bubble evidence counterfactual collaborative inference algorithm. Construct a counterfactual bubble profile based on the bubble profile and normalize it. Under the observation and counterfactual bubble profile, redistribute the evidence weight and propagation influence weight of the candidate evidence subset and evidence graph respectively and infer. The redistribution is based on the bubble parameters. Output the observation authenticity score and the counterfactual authenticity score set. Take the minimum value as the bubble invariance authenticity score and set the evidence node or propagation edge to zero and ablate it. Extract the bubble sensitivity factor set according to the score decrease amount. Step 4: Under budget and time constraints, solve the minimum set of intervention actions and execution sequence that minimizes action costs and reduces risk indicators to the threshold by using the bubble invariance authenticity score and bubble sensitivity factor set; Step 5: Output the minimum set of intervention actions and execution sequence, and record the event identifiers, bubble parameters, candidate evidence subsets, scoring results and bubble sensitivity factor set used for reproduction.

2. The processing method of an AI public opinion system according to claim 1, characterized in that... In step one, the event clustering simultaneously satisfies the semantic similarity threshold, the temporal proximity threshold, and the propagation link proximity threshold; the output set of disputed elements includes at least the element type, the element field, and the trigger fragment, and the element field is used for subsequent evidence retrieval and consistency verification.

3. The processing method of an AI public opinion system according to claim 1, characterized in that... In step two, the topic vector V topic The topic structure vector T and the text center vector V text The fusion coefficient α is used to obtain the linear fusion, and α is taken between 0.3 and 0.8 to balance structural consistency and semantic center.

4. The processing method of an AI public opinion system according to claim 1, characterized in that, In step two, the exposure ratio of entry type and the exposure ratio of user group are preferentially obtained by using the platform's exposure count (impression). If the platform does not provide this information, it is obtained by log feedback or sampling estimation, and the exposure criteria of different platforms are normalized.

5. The processing method of an AI public opinion system according to claim 1, characterized in that, In step two, the selection of candidate evidence subsets simultaneously satisfies the disputed element coverage threshold θ. cov With source type threshold θ src , where θ cov Used to constrain the coverage of evidence to disputed elements, θ src Used to constrain the credibility of evidence sources.

6. The processing method of an AI public opinion system according to claim 1, characterized in that, In step three, the counterfactual bubble profile is constructed based on the observed bubble profile. The counterfactual bubble profile is obtained by adjusting one or more bubble profile components related to the visibility bias. The observed bubble profile and the counterfactual bubble profile are then normalized and used for weight modulation.

7. The processing method of an AI public opinion system according to claim 1, characterized in that, In step three, the redistribution of evidence weights is based on the credibility of the evidence source and the coverage of disputed elements, and modulated in conjunction with the bubble parameters related to visibility bias; the redistribution of propagation influence weights is based on propagation intensity and cross-entry migration characteristics, and modulated in conjunction with the filter bubble values.

8. The processing method of an AI public opinion system according to claim 1, characterized in that, In step four, the minimum intervention action is determined using a joint optimization approach that combines budget and time constraints. Under the condition of satisfying both the budget and time limits, the output is the action that reduces the risk indicator R to a preset threshold R. th The minimum set of intervention actions and the execution sequence; wherein the execution sequence satisfies that the latest effective time does not exceed the time limit and the start delay of the key action does not exceed the preset start threshold.

9. An AI-powered public opinion system, characterized in that, The system is used to perform the processing method according to any one of claims 1-8, and the system includes a data acquisition and normalization module for acquiring multi-source public opinion records and standardizing them; The event clustering module performs event clustering and outputs event identifiers, event text sets, and disputed element sets. The bubble parameter calculation and evidence construction module generates topic vectors, statistically analyzes the exposure ratio of entry types and user groups, calculates bubble profile indicators and maps them to filter bubble values ​​and event relevance, calculates entry rate, exit probability and relaxation time, and obtains candidate evidence set based on event identifiers, constructs evidence graph and filters out candidate evidence subsets. The bubble evidence counterfactual collaborative inference module constructs and normalizes counterfactual bubble profiles, redistributes evidence weights and propagation impact weights under observation and counterfactual bubble profiles respectively, and infers the results, outputting observation authenticity scores, a set of counterfactual authenticity scores, bubble invariant authenticity scores, and a set of bubble sensitivity factors; the minimum intervention action solution module solves for the minimum intervention action set and execution sequence that minimizes action costs and reduces risk indicators to the threshold under budget and time constraints; the reproduction recording module records event identifiers, bubble parameters, candidate evidence subsets, scoring results, and a set of bubble sensitivity factors.

10. An AI public opinion system according to claim 9, characterized in that, The bubble evidence counterfactual collaborative inference module is configured to: construct one or more counterfactual bubble profiles based on the observed bubble profile and perform normalization processing; redistribute evidence weights and propagation influence weights to candidate evidence subsets and evidence graphs under the observed bubble profiles and each counterfactual bubble profile, and perform evidence graph reasoning to obtain the set of observed authenticity scores and counterfactual authenticity scores; take the minimum value between the observed authenticity score and the set of counterfactual authenticity scores as the bubble invariant authenticity score; and perform zeroing ablation on evidence nodes or propagation edges, extract the bubble sensitivity factor set based on the decrease in the bubble invariant authenticity score, and output it.

Images