News viewpoint evolution trend tracking system based on artificial intelligence

Abstract

The invention relates to the technical field of artificial intelligence, in particular to a news viewpoint evolution trend tracking system based on artificial intelligence, which comprises a collector, an extractor, an aggregator, a modeler, a detector and a presenter. A disturbance sample generation and consistency detection mechanism is combined, so that the clustering process is more robust, and the recognition precision of the viewpoint cluster semantic center is remarkably improved; according to the method, a black hole detection mechanism based on density estimation and multi-point similarity joint modeling, an anti-fact intervention mechanism and a dynamic graph density-structure joint estimation model are adopted, a node disturbance situation can be constructed, the influence of the node disturbance situation on a semantic structure can be inferred, and through fusion of semantic consistency, density offset and energy anti-fact scoring, the dynamic graph density-structure joint estimation model is obtained. High-confidence screening of abnormal nodes is realized, and the purity of a viewpoint atlas structure and the reliability of semantic evolution modeling are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to an artificial intelligence-based system for tracking the evolution of news viewpoints. Background Technology

[0002] With the proliferation of social media, mainstream media, and self-media platforms, the expression of news opinions has become increasingly diverse and time-sensitive. Existing methods for tracking opinions and modeling trends have the following shortcomings:

[0003] Insufficient semantic expression: Traditional methods rely on keywords or shallow bag-of-words models, which cannot fully characterize the deep semantic differences and emotional consistency between viewpoints;

[0004] Weak ability to process abnormal information: Existing technologies have limited ability to identify noisy viewpoints, abnormal expressions or false information, which can easily lead to distortion of viewpoint clustering.

[0005] The evolutionary trend modeling is crude: it lacks the ability to model evolution across multiple scales and time windows, and cannot effectively depict the structural evolution of viewpoint clusters across different time periods. Summary of the Invention

[0006] To address the above issues and overcome the shortcomings of existing technologies, this invention provides an AI-based news opinion evolution trend tracking system. Addressing the problem of insufficient semantic expression, this invention introduces a perturbation-robust fusion semantic clustering algorithm. By mining deep semantic similarities between opinion representations and combining perturbation sample generation and consistency detection mechanisms, the clustering process becomes more robust, significantly improving the accuracy of identifying semantic centers of opinion clusters. This achieves consistent aggregation of multi-source opinion semantics, enhancing clustering quality and opinion normalization capabilities. Addressing the problem of weak anomaly processing capabilities, this invention employs a black hole detection mechanism based on density estimation and multi-point similarity joint modeling, a counterfactual intervention mechanism, and a dynamic graph density-structure joint estimation model. This enables the construction of node perturbation scenarios and inference of their impact on semantic structure. Through the fusion of semantic consistency, density shift, and energy counterfactual scoring, high-confidence filtering of anomaly nodes is achieved. Addressing the problem of coarse evolution trend modeling, this invention constructs opinion temporal graphs and opinion evolution units, combined with a multi-scale evolutionary perceptual encoder, to finely model the temporal changes of opinions while maintaining semantic continuity.

[0007] The technical solution adopted in this invention is as follows: The news opinion evolution trend tracking system based on artificial intelligence provided by this invention includes a collector, an extractor, an aggregator, a modeler, a detector, and a presenter, specifically including the following:

[0008] The collector captures news data, including mainstream media, self-media content, and social media content;

[0009] The extractor extracts opinion information from news data;

[0010] The aggregator introduces a perturbation-robust fusion semantic clustering algorithm to perform semantic normalization and sentiment consistency clustering of opinion information, forming opinion clusters;

[0011] The modeler establishes a model of the evolutionary trend of viewpoints over time;

[0012] The detector identifies conflict detection results based on an evolutionary trend model, including conflicts, opposing relationships, and anomalous variations between viewpoints;

[0013] The presenter visualizes the evolutionary trend model and conflict detection results.

[0014] Furthermore, the aggregator specifically includes the following structure:

[0015] Feature extraction unit: Uses a pre-trained language model to vectorize opinion information to obtain opinion semantic representation vectors;

[0016] Network building unit: A heterogeneous graph structure is constructed based on the viewpoint semantic representation vector output by the feature extraction unit. Each viewpoint semantic representation is a node in the heterogeneous graph structure, and edges are established based on the semantic similarity between nodes.

[0017] Black hole detection unit: An improved anomaly detection mechanism based on density estimation and multi-point similarity joint modeling is introduced on the heterogeneous graph structure to detect and remove abnormal nodes, and obtain the heterogeneous graph structure after removal.

[0018] Core Point Extraction Module: Introduces a core point semantic clustering mechanism based on contrastive learning, performs semantic clustering of nodes in the heterogeneous graph structure after removal, and generates opinion clusters; extracts the core representative opinion node from each opinion cluster as the semantic center of the cluster;

[0019] Viewpoint Consistency Verification Unit: Performs consistency evaluation on the core representative viewpoint node of each viewpoint cluster and obtains the evaluation result;

[0020] Iterative optimization unit: The viewpoint semantic representation vector and evaluation results are iteratively optimized using an expectation-maximization strategy.

[0021] Furthermore, in the black hole detection unit, an improved anomaly detection mechanism based on joint modeling of density estimation and multi-point similarity is used to detect and remove anomalous nodes, specifically including the following steps:

[0022] Step S1: Perturbation sample generation. Based on the viewpoint semantic representation vector output by the feature extraction unit, a vector dimension perturbation is applied to each node in the heterogeneous graph structure to generate perturbation samples corresponding to the nodes.

[0023] Step S2: Perturbation Consistency Analysis. Based on the similarity between semantic representation vectors, a semantic clustering model is constructed. Unsupervised semantic clustering is performed in the heterogeneous graph structure. Clustering is performed on the nodes and the perturbation samples corresponding to the nodes. After obtaining their respective clustering labels, the consistency is compared, and a consistency index is constructed.

[0024] Step S3: Dynamic graph density-structure joint estimation. In the heterogeneous graph structure, lightweight graph convolution is performed on each node. After fusing semantic representations, a density change model based on time windows is constructed to obtain the dynamic density shift index of nodes in the heterogeneous graph structure.

[0025] Step S4: Energy-Semantic Contrastive Reasoning Scoring. Based on neighboring nodes that are semantically similar but have significantly different activity distributions, construct counterfactual comparison samples for each node. Construct two graph structure intervention scenarios: one is to remove the original node and obtain the impact of removing it from the heterogeneous graph structure on semantic clustering; the other is to replace it with the counterfactual comparison sample corresponding to each node and obtain the response change of the heterogeneous graph. Based on the impact of semantic clustering and the response change of the heterogeneous graph, calculate the energy-semantic counterfactual score of the corresponding node.

[0026] Step S5: Multimodal anomaly score fusion, which integrates the consistency index, dynamic density shift index and energy-semantic counterfactual score into a comprehensive node anomaly score;

[0027] Step S6: Adaptive threshold generation. An adaptive threshold T is generated using Gaussian kernel density estimation combined with local anomaly factors. Nodes with scores higher than T are marked as anomalous nodes.

[0028] Step S7: Abnormal node removal. Abnormal nodes are directly removed from the heterogeneous semantic graph structure. Based on their original connection relationships in the graph, the remaining nodes are partially reconstructed to obtain the heterogeneous graph structure after removal.

[0029] Furthermore, the modeler specifically includes the following structure:

[0030] Opinion sequence graph builder: Constructs opinion clusters into a multi-stage opinion evolution graph structure, where each node represents an opinion cluster generated by the aggregator, and each edge represents the semantic evolution path between temporally adjacent clusters;

[0031] Community Evolution Compression Unit: The view cluster graph at different times is structurally aligned and community compressed. While preserving semantic continuity, stable evolving clusters are aggregated into view evolution units through a merging strategy.

[0032] Structural inductive bias injection unit: Inject structural inductive bias into each viewpoint evolution unit, including meta-information on time trend directionality, aggregation consistency, and historical stability;

[0033] Multi-scale temporal window modeling unit: Based on the Transformer architecture, an evolutionary-aware encoder that supports cross-window interaction is designed to jointly model semantic, structural and evolutionary features at different time scales, and obtain high-dimensional representation results of each viewpoint evolutionary unit in the time series.

[0034] Trend Path Prediction Unit: Path modeling is performed on the high-dimensional representation results of each viewpoint evolution unit in the time series. Using a multi-objective path search and risk perception mechanism based on semantic potential graph, the probability of viewpoint clusters appearing, trend shift direction, and semantic conflict labels of the viewpoint evolution unit in the next stage are obtained.

[0035] Furthermore, the trend path prediction unit uses a multi-objective path search and risk perception mechanism based on semantic potential graphs, with the following specific steps:

[0036] Step P1: Potential path construction. Based on the high-dimensional representation of the view evolution units in the time series, the time series structure is constructed as follows: ;

[0037] in, Indicates the first The semantic vector of each viewpoint evolution unit;

[0038] Based on the time series structure, construct a time-aware semantic potential energy graph structure. ,in A node represents a semantic vector of an idea evolution unit. The edges connect the time-proximity viewpoint evolution unit vectors;

[0039] Step P2: Path candidate generation. Based on the semantic potential graph structure, a dynamic path expansion strategy is used, starting from the current node. Perform a heuristic search on future nodes to obtain a set of potential semantic evolution paths. Each path is a set of sequence nodes, represented as follows: ;

[0040] in, Representing a path exist Timing selection;

[0041] An energy constraint rule is introduced, a threshold is set, and the cumulative semantic potential energy value is calculated for each path. When the semantic potential energy of the current node of a path is lower than the set threshold, the path expansion is terminated, and candidate paths are obtained.

[0042] Step P3: Path threat assessment embedding, using a rule-based Bayesian inference model, calculates the semantic conflict probability for each node, and weights them to obtain the semantic conflict score for the entire path;

[0043] Step P4: Multi-objective scoring of paths. Perform multi-objective comprehensive scoring on each candidate path and select the path with the highest score as the trend prediction path.

[0044] Step P5: Future view cluster prediction. The last node in the trend prediction path is used as the prediction starting point. A Transformer-based decoder is used to predict the next possible view evolution unit. Combining the position features of the corresponding node in the semantic potential map, the probability of the view cluster occurrence, the trend offset direction, and the semantic conflict label are output.

[0045] The beneficial effects achieved by the present invention using the above solution are as follows:

[0046] (1) To address the problem of insufficient semantic expression, this invention introduces a perturbation robust fusion semantic clustering algorithm, which mines the deep semantic similarity between viewpoint representations and combines perturbation sample generation and consistency detection mechanisms to make the clustering process more robust, significantly improves the recognition accuracy of the semantic center of viewpoint clusters, realizes the consistent aggregation of multi-source viewpoint semantics, and improves the clustering quality and viewpoint normalization ability.

[0047] (2) To address the problem of weak abnormal information processing capabilities, this invention adopts a black hole detection mechanism based on density estimation and multi-point similarity joint modeling, a counterfactual intervention mechanism and a dynamic graph density-structure joint estimation model, which can construct node perturbation scenarios and infer their impact on semantic structure. Through the fusion of semantic consistency, density shift and energy counterfactual scoring, high-confidence screening of abnormal nodes can be achieved.

[0048] (3) To address the problem of coarse modeling of evolutionary trends, this invention constructs a viewpoint time sequence diagram and viewpoint evolution unit, and combines a multi-scale evolutionary perception encoder to finely model the temporal changes of viewpoints while maintaining semantic continuity. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the structure of the AI-based news opinion evolution trend tracking system proposed in this invention.

[0050] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0051] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0052] Example 1, see Figure 1 The artificial intelligence-based news opinion evolution trend tracking system provided by this invention includes a collector, extractor, aggregator, modeler, detector, and presenter, specifically including the following:

[0053] The collector captures news data, including mainstream media, self-media content, and social media content;

[0054] The extractor extracts opinion information from news data;

[0055] The aggregator introduces a perturbation-robust fusion semantic clustering algorithm to perform semantic normalization and sentiment consistency clustering of opinion information, forming opinion clusters;

[0056] The modeler establishes a model of the evolutionary trend of viewpoints over time;

[0057] The detector identifies conflict detection results based on an evolutionary trend model, including conflicts, opposing relationships, and anomalous variations between viewpoints;

[0058] The presenter visualizes the evolutionary trend model and conflict detection results.

[0059] Example 2, based on the above examples, describes a polymerizer that specifically includes the following structure:

[0060] Feature extraction unit: Uses a pre-trained language model to vectorize opinion information to obtain opinion semantic representation vectors;

[0061] Network building unit: A heterogeneous graph structure is constructed based on the viewpoint semantic representation vector output by the feature extraction unit. Each viewpoint semantic representation is a node in the heterogeneous graph structure, and edges are established based on the semantic similarity between nodes.

[0062] Black hole detection unit: An improved anomaly detection mechanism based on density estimation and multi-point similarity joint modeling is introduced on the heterogeneous graph structure to detect and remove abnormal nodes, and obtain the heterogeneous graph structure after removal.

[0063] Core Point Extraction Module: Introduces a core point semantic clustering mechanism based on contrastive learning, performs semantic clustering of nodes in the heterogeneous graph structure after removal, and generates opinion clusters; extracts the core representative opinion node from each opinion cluster as the semantic center of the cluster;

[0064] Viewpoint Consistency Verification Unit: Performs consistency evaluation on the core representative viewpoint node of each viewpoint cluster and obtains the evaluation result;

[0065] Iterative optimization unit: The viewpoint semantic representation vector and evaluation results are iteratively optimized using an expectation-maximization strategy.

[0066] In this embodiment, the multi-scale temporal window modeling unit uses a Transformer structure as the basic network skeleton and introduces a cross-temporal window interaction mechanism to construct an evolutionary sensing encoder module. First, the viewpoint evolution unit sequence is divided for different temporal granularities, and a time window is established at each scale. By utilizing the shared attention mechanism between windows, cross-scale interactive modeling of viewpoint semantics, structural associations, and evolutionary change features is achieved. Then, through a multi-layer cross-attention module, the semantic representations at each time scale are fused to obtain a unified high-dimensional representation vector for each viewpoint evolution unit in the multi-scale semantic context.

[0067] The trend path prediction unit, based on the high-dimensional representation results output by the aforementioned multi-scale modeling unit, introduces a semantic potential graph mechanism to model future paths, constructs a time-aware semantic potential graph structure, uses viewpoint evolution units as nodes, and establishes edge weights based on semantic similarity and temporal adjacency to represent evolutionary potential. Utilizing a dynamic path expansion strategy, it performs heuristic path search to generate multiple candidate evolutionary paths, and filters paths based on the cumulative semantic potential threshold. By introducing a rule-based Bayesian inference model, it performs semantic conflict probability modeling on each node in the path, forming a path-level conflict score. Combining multiple objective indicators such as semantic coherence, trend directionality, and conflict risk, it performs a comprehensive path score, selects the path with the highest score as the trend evolution path, and predicts the probability of the next stage viewpoint cluster appearing, the trend shift direction, and the semantic conflict label.

[0068] Example 3, based on the above examples, describes an improved anomaly detection mechanism in the black hole detection unit that uses density estimation and multi-point similarity joint modeling to detect and remove anomalous nodes. Specifically, it includes the following steps:

[0069] Step S1: Perturbation sample generation. Based on the viewpoint semantic representation vector output by the feature extraction unit, a vector dimension perturbation is applied to each node in the heterogeneous graph structure to generate perturbation samples corresponding to the nodes.

[0070] Step S2: Perturbation Consistency Analysis. Based on the similarity between semantic representation vectors, a semantic clustering model is constructed. Unsupervised semantic clustering is performed in the heterogeneous graph structure. Clustering is performed on the nodes and the perturbation samples corresponding to the nodes. After obtaining their respective clustering labels, the consistency is compared, and a consistency index is constructed.

[0071] Step S3: Dynamic graph density-structure joint estimation. In the heterogeneous graph structure, lightweight graph convolution is performed on each node. After fusing semantic representations, a density change model based on time windows is constructed to obtain the dynamic density shift index of nodes in the heterogeneous graph structure.

[0072] Step S4: Energy-Semantic Contrastive Reasoning Scoring. Based on neighboring nodes that are semantically similar but have significantly different activity distributions, construct counterfactual comparison samples for each node. Construct two graph structure intervention scenarios: one is to remove the original node and obtain the impact of removing it from the heterogeneous graph structure on semantic clustering; the other is to replace it with the counterfactual comparison sample corresponding to each node and obtain the response change of the heterogeneous graph. Based on the impact of semantic clustering and the response change of the heterogeneous graph, calculate the energy-semantic counterfactual score of the corresponding node.

[0073] Step S5: Multimodal anomaly score fusion, which integrates the consistency index, dynamic density shift index and energy-semantic counterfactual score into a comprehensive node anomaly score;

[0074] Step S6: Adaptive threshold generation. An adaptive threshold T is generated using Gaussian kernel density estimation combined with local anomaly factors. Nodes with scores higher than T are marked as anomalous nodes.

[0075] Step S7: Abnormal node removal. Abnormal nodes are directly removed from the heterogeneous semantic graph structure. Based on their original connection relationships in the graph, the remaining nodes are partially reconstructed to obtain the heterogeneous graph structure after removal.

[0076] In this embodiment, the core code used is as follows:

[0077] import numpy as np

[0078] from sklearn.cluster import KMeans

[0079] from sklearn.neighbors import LocalOutlierFactor

[0080] from sklearn.neighbors import KernelDensity

[0081] from sklearn.metrics.pairwise import cosine_similarity

[0082] import networkx as nx

[0083] # Simulate a list of semantic vectors (one vector for each viewpoint)

[0084] semantic_vectors = np.random.rand(100, 128) # 100 nodes, 128-dimensional vector

[0085] sigma = 0.01 # Disturbance strength

[0086] # Step S1: Generation of perturbation samples

[0087] def generate_perturbations(vectors, sigma=0.01):

[0088] perturbations = vectors + np.random.normal(0, sigma,vectors.shape)

[0089] return perturbations

[0090] perturbed_vectors = generate_perturbations(semantic_vectors, sigma)

[0091] # Step S2: Perturbation Consistency Analysis

[0092] def compute_consistency_index(original, perturbed, n_clusters=10):

[0093] kmeans1 = KMeans(n_clusters=n_clusters).fit(original)

[0094] kmeans2 = KMeans(n_clusters=n_clusters).fit(perturbed)

[0095] consistency = (kmeans1.labels_ == kmeans2.labels_).astype(int)

[0096] return 1 - consistency # 1 indicates inconsistency (abnormal).

[0097] consistency_index = compute_consistency_index(semantic_vectors,perturbed_vectors)

[0098] # Step S3: Dynamic Graph Density Estimation (Simplified Version)

[0099] # Density changes are represented here using simulation plots and time window variations.

[0100] def compute_density_shift(graphs): # graphs: list of networkx Graphat different time

[0101] shifts = []

[0102] for node in graphs[0].nodes():

[0103] d0 = graphs[0].degree(node)

[0104] d1 = graphs[1].degree(node)

[0105] shift = abs(d1 - d0) / (d0 + 1e-5)

[0106] shifts.append(shift)

[0107] return np.array(shifts)

[0108] # Assume graph_t and graph_t+1 are graphs of two consecutive time windows.

[0109] graph_t = nx.erdos_renyi_graph(100, 0.05)

[0110] graph_t1 = nx.erdos_renyi_graph(100, 0.07)

[0111] density_shift = compute_density_shift([graph_t, graph_t1])

[0112] # Step S4: Counterfactual Scoring (Simplified Simulation)

[0113] def counterfactual_score(vectors):

[0114] sim = cosine_similarity(vectors)

[0115] np.fill_diagonal(sim, 0)

[0116] max_sim = np.max(sim, axis=1)

[0117] return 1 - max_sim # Lower similarity → more "abnormal"

[0118] cf_score = counterfactual_score(semantic_vectors)

[0119] # Step S5: Multimodal Fusion Scoring

[0120] def fuse_scores(consistency, density, cf, weights=(0.4, 0.3, 0.3)):

[0121] score = weights[0]*consistency + weights[1]*density + weights[2]*cf

[0122] return score

[0123] fused_score = fuse_scores(consistency_index, density_shift, cf_score)

[0124] # Step S6: Adaptive Threshold Determination

[0125] def adaptive_threshold(score):

[0126] kde = KernelDensity(kernel='gaussian', bandwidth=0.1).fit(score.reshape(-1, 1))

[0127] log_dens = kde.score_samples(score.reshape(-1, 1))

[0128] lof = LocalOutlierFactor(n_neighbors=20).fit_predict(score.reshape(-1, 1))

[0129] combined = log_dens + lof

[0130] threshold = np.percentile(combined, 95) # top 5% considered as anomaly

[0131] return score > threshold

[0132] is_anomaly = adaptive_threshold(fused_score)

[0133] # Step S7: Remove anomalous nodes

[0134] def remove_anomalies(graph, anomaly_flags):

[0135] cleaned_graph = graph.copy()

[0136] for i, flag in enumerate(anomaly_flags):

[0137] if flag:

[0138] cleaned_graph.remove_node(i)

[0139] return cleaned_graph

[0140] final_graph = remove_anomalies(graph_t, is_anomaly)

[0141] print("Number of anomaly nodes removed:", np.sum(is_anomaly)).

[0142] Example 4, based on the above examples, describes a modeler that specifically includes the following structure:

[0143] Opinion sequence graph builder: Constructs opinion clusters into a multi-stage opinion evolution graph structure, where each node represents an opinion cluster generated by the aggregator, and each edge represents the semantic evolution path between temporally adjacent clusters;

[0144] Community Evolution Compression Unit: The view cluster graph at different times is structurally aligned and community compressed. While preserving semantic continuity, stable evolving clusters are aggregated into view evolution units through a merging strategy.

[0145] Structural inductive bias injection unit: Inject structural inductive bias into each viewpoint evolution unit, including meta-information on time trend directionality, aggregation consistency, and historical stability;

[0146] Multi-scale temporal window modeling unit: Based on the Transformer architecture, an evolutionary-aware encoder that supports cross-window interaction is designed to jointly model semantic, structural and evolutionary features at different time scales, and obtain high-dimensional representation results of each viewpoint evolutionary unit in the time series.

[0147] Trend Path Prediction Unit: Path modeling is performed on the high-dimensional representation results of each viewpoint evolution unit in the time series. Using a multi-objective path search and risk perception mechanism based on semantic potential graph, the probability of viewpoint clusters appearing, trend shift direction, and semantic conflict labels of the viewpoint evolution unit in the next stage are obtained.

[0148] In this embodiment, the social event of "adjustment of national subsidy policy for new energy vehicles" is taken as an example;

[0149] Between June and September 2023, thousands of news articles and comments from mainstream media, self-media platforms, and social networks were collected, and the aggregator normalized them into multiple semantically consistent clusters of viewpoints:

[0150] The government is gradually reducing subsidies for new energy vehicles;

[0151] Reduced subsidies will dampen consumer enthusiasm for purchasing;

[0152] Automakers are shifting towards technology-driven alternatives to subsidy-based models;

[0153] The modeler constructs these opinion clusters into a multi-stage opinion evolution graph at different time stages (such as weekly), where each node in the graph represents an opinion cluster and each edge represents a semantic evolution path between adjacent time periods.

[0154] Through semantic similarity and topic continuity analysis, the system established an evolutionary path, such as a three-stage path from "subsidy reduction" → "market reaction" → "industry strategy adjustment";

[0155] Based on the constructed multi-stage graph structure, the system performs structural alignment on graphs from adjacent time points and discovers the following patterns:

[0156] One type of viewpoint (P2, P4, P6) differs only in language style and tone across multiple stages, but the essential expression remains the same;

[0157] Another set of viewpoints (P3, P5) form a consistent discussion flow on the topic of "manufacturer adaptability" in terms of semantics;

[0158] Therefore, the system uses semantic aggregation and topic stability rules to group these viewpoints into two stable "viewpoint evolution units":

[0159] Consumer subsidy perception path;

[0160] Manufacturer's response path;

[0161] Structural inductive bias is injected into each viewpoint evolution unit to assist subsequent modelers in focusing on its evolutionary characteristics:

[0162] The injected metadata includes:

[0163] Directional trend over time: negative, with semantic representation declining continuously, indicating low sentiment;

[0164] Consistency: High; most viewpoints express similar concerns.

[0165] Historical stability: Medium, with high volatility in the early stages and increased concentration in the recent period;

[0166] The injected metadata includes:

[0167] Directional trend over time: positive, indicating steady growth;

[0168] Aggregate consistency: Medium;

[0169] Historical stability: High;

[0170] The modeler employs an improved Transformer encoder, setting multi-scale windows (such as daily, weekly, and stage windows) to perform joint modeling for each viewpoint evolution unit as follows:

[0171] Semantic modeling: The input token is formed by fusing word vectors (BERT) with the representation vectors of cluster centers;

[0172] Structural modeling: Introducing the local adjacency structure of the view graph as an attention bias;

[0173] Evolutionary modeling: Cross-window attention linkage modeling of trend changes in EVU1 and EVU2;

[0174] The final system obtains a high-dimensional representation vector of each evolutionary unit in the time series, including its historical semantic trajectory, structural trend and evolutionary potential characteristics;

[0175] The system takes a high-dimensional representation vector as input and constructs a "semantic potential graph," in which:

[0176] Nodes are representation vectors;

[0177] Edges connect temporally adjacent nodes and assign energy weights to the evolution direction;

[0178] The system performs path search from the latest moment to future nodes:

[0179] A multi-objective heuristic strategy (combining trend energy, semantic consistency, and conflict risk) is used to generate potential paths;

[0180] Each path is assigned:

[0181] Probability of the occurrence of opinion clusters;

[0182] Trend deviation direction;

[0183] Semantic conflict tags;

[0184] Output:

[0185] Potential themes and risks for the next phase;

[0186] Predicting trend direction;

[0187] The system recommends key paths for manual review, and supports source tracing and emergency assessment.

[0188] Example 5, based on the above examples, describes a trend path prediction unit that uses a multi-objective path search and risk perception mechanism based on semantic potential graphs. The specific steps are as follows:

[0189] Step P1: Potential path construction. Based on the high-dimensional representation of the view evolution units in the time series, the time series structure is constructed as follows:

[0190] ;

[0191] in, Indicates the first The semantic vector of each viewpoint evolution unit;

[0192] Based on the time series structure, construct a time-aware semantic potential energy graph structure. ,in A node represents a semantic vector of an idea evolution unit. The edges connect the time-proximity viewpoint evolution unit vectors;

[0193] Step P2: Path candidate generation. Based on the semantic potential graph structure, a dynamic path expansion strategy is used, starting from the current node. Perform a heuristic search on future nodes to obtain a set of potential semantic evolution paths. Each path is a set of sequence nodes, represented as follows:

[0194] ;

[0195] in, Representing a path exist Timing selection;

[0196] An energy constraint rule is introduced, a threshold is set, and the cumulative semantic potential energy value is calculated for each path. When the semantic potential energy of the current node of a path is lower than the set threshold, the path expansion is terminated, and candidate paths are obtained.

[0197] Step P3: Path threat assessment embedding, using a rule-based Bayesian inference model, calculates the semantic conflict probability for each node, and weights them to obtain the semantic conflict score for the entire path;

[0198] Step P4: Multi-objective scoring of paths. Perform multi-objective comprehensive scoring on each candidate path and select the path with the highest score as the trend prediction path.

[0199] Step P5: Future view cluster prediction. The last node in the trend prediction path is used as the prediction starting point. A Transformer-based decoder is used to predict the next possible view evolution unit. Combining the position features of the corresponding node in the semantic potential map, the probability of the view cluster occurrence, the trend offset direction, and the semantic conflict label are output.

[0200] In this embodiment, the core code used is as follows:

[0201] import torch

[0202] import torch.nn as nn

[0203] import networkx as nx

[0204] import numpy as np

[0205] # Input assumption: High-dimensional representation vector H_t of the time series of viewpoint evolution units

[0206] # H_t = [E_1, E_2, ..., E_t], where each E_i is a d-dimensional vector.

[0207] H_t = torch.randn((10, 128)) # 10 time points, each with 128 dimensions

[0208] # -------------------------

[0209] # Step P1: Construct the semantic potential graph G_s

[0210] # -------------------------

[0211] G_s = nx.DiGraph()

[0212] for i in range(len(H_t)):

[0213] G_s.add_node(i, vector=H_t[i])

[0214] for i in range(len(H_t) - 1):

[0215] dist = torch.norm(H_t[i+1] - H_t[i]).item()

[0216] G_s.add_edge(i, i+1, weight=dist)

[0217] # -------------------------

[0218] # Step P2: Dynamic path expansion + potential energy threshold control

[0219] # -------------------------

[0220] def expand_paths(G, current, max_len=4, energy_threshold=1.5):

[0221] paths = []

[0222] def dfs(path, acc_energy):

[0223] if len(path) > max_len or acc_energy > energy_threshold:

[0224] return

[0225] paths.append(list(path))

[0226] for succ in G.successors(path[-1]):

[0227] energy = G.edges[path[-1], succ]['weight']

[0228] if energy < energy_threshold:

[0229] dfs(path + [succ], acc_energy + energy)

[0230] dfs([current], 0)

[0231] return paths

[0232] candidate_paths = expand_paths(G_s, current=len(H_t)-1)

[0233] # -------------------------

[0234] # Step P3: Path Threat Assessment Embedding (Bayesian Conflict Probability)

[0235] # -------------------------

[0236] def bayesian_conflict_score(E):

[0237] # Pseudo-Bayesian model illustration: High semantic gradient = High conflict probability

[0238] grad = torch.norm(E[1:] - E[:-1], dim=1)

[0239] prob = torch.sigmoid(grad * 2.0) # Convert to collision probability

[0240] return prob.mean().item()

[0241] path_scores = []

[0242] for path in candidate_paths:

[0243] vectors = torch.stack([G_s.nodes[i]['vector'] for i in path])

[0244] score = bayesian_conflict_score(vectors)

[0245] path_scores.append(score)

[0246] # -------------------------

[0247] # Step P4: Multi-objective scoring (combining indicators such as conflict and potential stability)

[0248] # -------------------------

[0249] def multi_objective_score(path, conflict_score):

[0250] # Example: Multi-objective comprehensive score (can be expanded with more indicators)

[0251] path_len = len(path)

[0252] smoothness = np.mean([G_s.edges[path[i], path[i+1]]['weight'] fori in range(len(path)-1)])

[0253] return 0.5 * (1 - conflict_score) + 0.5 * (1 / (1 + smoothness))

[0254] final_scores = [multi_objective_score(p, s) for p, s in zip(candidate_paths, path_scores)]

[0255] best_path_idx = np.argmax(final_scores)

[0256] best_path = candidate_paths[best_path_idx]

[0257] # -------------------------

[0258] # Step P5: Predict future view clusters using the Transformer decoder

[0259] # -------------------------

[0260] class FuturePredictor(nn.Module):

[0261] def __init__(self, d_model=128, num_layers=2):

[0262] super().__init__()

[0263] self.decoder_layer = nn.TransformerDecoderLayer(d_model=d_model, nhead=4)

[0264] self.decoder = nn.TransformerDecoder(self.decoder_layer, num_layers=num_layers)

[0265] self.linear = nn.Linear(d_model, d_model)

[0266] def forward(self, past_sequence):

[0267] tgt = torch.zeros(1, 1, past_sequence.size(-1)) # Initialize the decoding start point

[0268] memory = past_sequence.unsqueeze(1) # seq_len x batch x dim

[0269] output = self.decoder(tgt, memory)

[0270] return self.linear(output.squeeze(0))

[0271] # Instantiate the predictor

[0272] predictor = FuturePredictor()

[0273] best_path_vectors = torch.stack([G_s.nodes[i]['vector'] for i inbest_path])

[0274] pred_E = predictor(best_path_vectors) # Predict the next view vector

[0275] # Example: Output trend prediction labels

[0276] print("

Prediction Result

[0277] print("- Next viewpoint evolution unit semantic vector:", pred_E.detach().numpy())

[0278] print("- Trend offset direction: positive" if pred_E.mean().item() > 0 else "negative")

[0279] print("- Conflict tag: High risk" if bayesian_conflict_score(torch.cat((best_path_vectors, pred_E.unsqueeze(0)))) > 0.7 else "Low risk").

[0280] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0281] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

[0282] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A news opinion evolution trend tracking system based on artificial intelligence, characterized by: It includes collectors, extractors, aggregators, modelers, detectors, and renderers, specifically including the following: The data collector captures news data; The extractor extracts opinion information from news data; The aggregator introduces a perturbation-robust fusion semantic clustering algorithm to perform semantic normalization and sentiment consistency clustering of opinion information, forming opinion clusters; The modeler establishes a model of the evolutionary trend of viewpoints over time; The detector identifies conflict detection results based on an evolutionary trend model; The presenter visualizes the evolutionary trend model and conflict detection results.

2. The news opinion evolution trend tracking system based on artificial intelligence according to claim 1, characterized in that: The aggregator specifically includes the following structure: Feature extraction unit: Uses a pre-trained language model to vectorize opinion information to obtain opinion semantic representation vectors; Network building unit: A heterogeneous graph structure is constructed based on the viewpoint semantic representation vector output by the feature extraction unit. Each viewpoint semantic representation is a node in the heterogeneous graph structure, and edges are established based on the semantic similarity between nodes. Black hole detection unit: An improved anomaly detection mechanism based on density estimation and multi-point similarity joint modeling is introduced on the heterogeneous graph structure to detect and remove abnormal nodes, and obtain the heterogeneous graph structure after removal. Core Point Extraction Module: Introduces a core point semantic clustering mechanism based on contrastive learning, performs semantic clustering of nodes in the heterogeneous graph structure after removal, and generates opinion clusters; extracts the core representative opinion node from each opinion cluster as the semantic center of the cluster; Viewpoint Consistency Verification Unit: Performs consistency evaluation on the core representative viewpoint node of each viewpoint cluster and obtains the evaluation result; Iterative optimization unit: The viewpoint semantic representation vector and evaluation results are iteratively optimized using an expectation-maximization strategy.

3. The news opinion evolution trend tracking system based on artificial intelligence according to claim 2, characterized in that: In the black hole detection unit, an improved anomaly detection mechanism based on joint modeling of density estimation and multi-point similarity is used to detect and remove anomalous nodes. Specifically, this includes the following steps: Step S1: Perturbation sample generation. Based on the viewpoint semantic representation vector output by the feature extraction unit, a vector dimension perturbation is applied to each node in the heterogeneous graph structure to generate perturbation samples corresponding to the nodes. Step S2: Perturbation Consistency Analysis. Based on the similarity between semantic representation vectors, a semantic clustering model is constructed. Unsupervised semantic clustering is performed in the heterogeneous graph structure. Clustering is performed on the nodes and the perturbation samples corresponding to the nodes. After obtaining their respective clustering labels, the consistency is compared, and a consistency index is constructed. Step S3: Dynamic graph density-structure joint estimation. In the heterogeneous graph structure, lightweight graph convolution is performed on each node. After fusing semantic representations, a density change model based on time windows is constructed to obtain the dynamic density shift index of nodes in the heterogeneous graph structure. Step S4: Energy-Semantic Contrastive Reasoning Scoring. Based on neighboring nodes that are semantically similar but have significantly different activity distributions, construct counterfactual comparison samples for each node. Construct two graph structure intervention scenarios: one is to remove the original node and obtain the impact of removing it from the heterogeneous graph structure on semantic clustering; the other is to replace it with the counterfactual comparison sample corresponding to each node and obtain the response change of the heterogeneous graph. Based on the impact of semantic clustering and the response change of the heterogeneous graph, calculate the energy-semantic counterfactual score of the corresponding node. Step S5: Multimodal anomaly score fusion, which integrates the consistency index, dynamic density shift index and energy-semantic counterfactual score into a comprehensive node anomaly score; Step S6: Adaptive threshold generation. An adaptive threshold T is generated using Gaussian kernel density estimation combined with local anomaly factors. Nodes with scores higher than T are marked as anomalous nodes. Step S7: Abnormal node removal. Abnormal nodes are directly removed from the heterogeneous semantic graph structure. Based on their original connection relationships in the graph, the remaining nodes are partially reconstructed to obtain the heterogeneous graph structure after removal.

4. The news opinion evolution trend tracking system based on artificial intelligence according to claim 1, characterized in that: The modeler specifically includes the following structure: Opinion sequence graph builder: Constructs opinion clusters into a multi-stage opinion evolution graph structure, where each node represents an opinion cluster generated by the aggregator, and each edge represents the semantic evolution path between temporally adjacent clusters; Community Evolution Compression Unit: The view cluster graph at different times is structurally aligned and community compressed. While preserving semantic continuity, stable evolving clusters are aggregated into view evolution units through a merging strategy. Structural inductive bias injection unit: Inject structural inductive bias into each viewpoint evolution unit, including meta-information on time trend directionality, aggregation consistency, and historical stability; Multi-scale temporal window modeling unit: Based on the Transformer architecture, an evolutionary-aware encoder that supports cross-window interaction is designed to jointly model semantic, structural and evolutionary features at different time scales, and obtain high-dimensional representation results of each viewpoint evolutionary unit in the time series. Trend Path Prediction Unit: Path modeling is performed on the high-dimensional representation results of each viewpoint evolution unit in the time series. Using a multi-objective path search and risk perception mechanism based on semantic potential graph, the probability of viewpoint clusters appearing, trend shift direction, and semantic conflict labels of the viewpoint evolution unit in the next stage are obtained.

5. The news opinion evolution trend tracking system based on artificial intelligence according to claim 4, characterized in that: The trend path prediction unit uses a multi-objective path search and risk perception mechanism based on semantic potential graphs. The specific steps are as follows: Step P1: Potential path construction. Based on the high-dimensional representation of the view evolution units in the time series, the time series structure is constructed as follows: ； in, Indicates the first The semantic vector of each viewpoint evolution unit; Based on the time series structure, construct a time-aware semantic potential energy graph structure. ,in A node represents a semantic vector of an idea evolution unit. The edges connect the time-proximity viewpoint evolution unit vectors; Step P2: Path candidate generation. Based on the semantic potential graph structure, a dynamic path expansion strategy is used, starting from the current node. Perform a heuristic search on future nodes to obtain a set of potential semantic evolution paths. Each path is a set of sequence nodes, represented as follows: ； in, Representing a path exist Timing selection; An energy constraint rule is introduced, a threshold is set, and the cumulative semantic potential energy value is calculated for each path. When the semantic potential energy of the current node of a path is lower than the set threshold, the path expansion is terminated, and candidate paths are obtained. Step P3: Path threat assessment embedding, using a rule-based Bayesian inference model, calculates the semantic conflict probability for each node, and weights them to obtain the semantic conflict score for the entire path; Step P4: Multi-objective scoring of paths. Perform multi-objective comprehensive scoring on each candidate path and select the path with the highest score as the trend prediction path. Step P5: Future view cluster prediction. The last node in the trend prediction path is used as the prediction starting point. A Transformer-based decoder is used to predict the next possible view evolution unit. Combining the position features of the corresponding node in the semantic potential map, the probability of the view cluster occurrence, the trend offset direction, and the semantic conflict label are output.

Images