A method and apparatus for evaluating the credibility of generated content

Abstract

This invention provides a method and apparatus for evaluating the credibility of generated content, relating to the fields of computer science and artificial intelligence. The method includes: collecting information sources from multiple platforms; extracting features from the information sources to obtain multi-dimensional features; defining graph nodes and edges, assigning weights, and learning graph neural networks from the multi-dimensional features to obtain a directed weighted information source authority graph; performing end-to-end self-supervised learning on the information source authority graph using a graph convolutional network or graph attention network to obtain the graph neural network authority score for each node; performing conflict graph modeling and confidence propagation iterative algorithm processing on the information source authority graph to obtain a confidence propagation score; setting a timeliness factor, and weightedly fusing the graph neural network authority score, confidence propagation score, and timeliness factor to obtain a credibility score. This method achieves dynamic, accurate, and multi-platform collaborative evaluation of information source credibility.

Description

Technical Field

[0001] This invention relates to the fields of computers and artificial intelligence, and more specifically to a method and apparatus for evaluating the credibility of generated content. Background Technology

[0002] With the popularization of generative artificial intelligence, it has become the norm for users to rely on large AI models to obtain information and make consumption decisions. The frequency of a brand's appearance and recommendation order in AI-generated answers constitute a new track for corporate marketing—Generative Engine Optimization (GEO). The core goal of GEO services is to systematically improve the visibility and recommendation rate of brands in AI answers.

[0003] In current GEO practices, assessing the credibility of sources cited in AI answers is a core step. However, existing technical solutions have significant technical flaws and cannot meet the actual needs of GEO strategy formulation. Specific problems are as follows: 1. Static weight assessment has a lag: Existing solutions mostly rely on static domain authority scores such as Moz domain authority value (DA) and Ahrefs domain rating (DR), or manually maintained information source lists. However, the AI ​​big model's citation preference for information sources is driven by algorithms and has dynamic characteristics. Static scores cannot reflect the actual fluctuations in the influence of information sources in the AI ​​ecosystem in real time, and the assessment results are out of touch with the actual situation. 2. Lack of ability to identify false and conflicting information: There are a large number of noise sources on the Internet, such as false information, malicious reviews, and content from counterfeit websites targeting brands. Current technology lacks an effective mechanism to identify such sources. At the same time, it is impossible to detect contradictory statements of the same brand's core information in different sources (such as inconsistent product prices, official contact information, and parameter specifications), nor are there any technical means to use information conflicts to infer the reliability of sources. 3. Single evaluation dimension and lack of multi-platform collaboration: Most evaluation tools only monitor and analyze a single AI platform, without covering mainstream diversified AI platforms in China such as Doubao, DeepSeek, and Wenxin Yiyan; the source citation logic of different AI platforms is significantly different, and the evaluation results of a single platform do not have universal reference value and cannot support the formulation of cross-platform GEO strategies. 4. Poor interpretability and operability of evaluation results: Existing evaluation methods mostly output single scores without breaking down the scoring dimensions, making it impossible to trace the core reasons for high or low scores; at the same time, the evaluation results are out of touch with GEO's actual marketing scenarios and are difficult to directly translate into actionable source selection and content delivery strategies.

[0004] In summary, there is an urgent need in this field for a source credibility assessment method and system that can achieve dynamic evaluation, multi-platform collaboration, conflict detection capabilities, and whose evaluation results are interpretable and operable, so as to provide accurate data support for GEO strategies and solve the above-mentioned deficiencies of existing technologies. Summary of the Invention

[0005] The purpose of this invention is to provide a method and apparatus for evaluating the credibility of generated content. This method enables dynamic, accurate, and multi-platform collaborative evaluation of the credibility of information sources.

[0006] To achieve the above objectives, embodiments of the present invention provide a method for evaluating the credibility of generated content. The method includes: collecting information sources from multiple platforms; extracting features from the information sources to obtain multi-dimensional features, including basic features, content features, citation features, and conflict features; defining graph nodes and edges, assigning weights, and learning graph neural networks on the multi-dimensional features to obtain a directed weighted information source authority graph; performing end-to-end self-supervised learning on the information source authority graph using a graph convolutional network or a graph attention network to obtain the graph neural network authority score for each node; performing conflict graph modeling and confidence propagation iterative algorithm processing on the information source authority graph to obtain a confidence propagation score; setting a timeliness factor, and weightedly fusing the graph neural network authority score, the confidence propagation score, and the timeliness factor to obtain a credibility score.

[0007] Optionally, the collection of user information sources from multiple platforms includes: collecting data from multiple platform interfaces through a distributed crawler cluster; and processing the collected data through data cleaning, data verification, and data standardization to obtain user questions, generated answers, and information sources.

[0008] Optionally, the basic features include domain age, third-party domain weight, global ranking, search engine indexing volume, number of backlinks, and backlink quality; the content features include industry relevance, content update frequency, structural integrity, proportion of multimedia elements, content originality, and content compliance; the citation features include total citation frequency, first citation time, most recent citation time, citation sentiment, citation position weight, and cross-platform citation breadth; and the conflict features include baseline truth value conflict degree, number of cross-source conflicts, proportion of dimensional conflicts, and conflict information correction rate.

[0009] Optionally, the step of performing graph modeling and graph neural network learning on the multidimensional features to obtain a directed weighted source authority graph includes: constructing a node set and a directed edge set; the node set includes multiple nodes, each node corresponding to a source, and each node has an ID used to associate the basic features of the source; the directed edge set includes edge types between sources generated based on the content features, citation features, and conflict features of the source, including citation association edges, content association edges, and industry isomorphic edges; and constructing the source authority graph based on the node set and the directed edge set.

[0010] Optionally, the reference association edge is: when information source A is the information reference source of information source B, a directed edge A→B is established; the content association edge is: when the content relevance of the two information sources is greater than or equal to a preset threshold and there is no information conflict, a bidirectional directed edge is established; the industry isomorphic edge is: when the two information sources belong to the same industry field and the target brand content attributes are consistent, a bidirectional directed edge is established.

[0011] Optionally, the step of using a graph convolutional network or a graph attention network to perform end-to-end self-supervised learning on the source authority graph to obtain the graph neural network authority score of each node includes: obtaining the node weights and edge weights of the nodes in the source authority graph, and using the node weights and multi-dimensional features as the input feature matrix; weighting and aggregating the multi-dimensional features and edge weights of the first-order or second-order neighbors of the nodes in the input feature matrix through a hidden layer to obtain the embedding vector of each node; and mapping the embedding vector to the 0-1 interval through a fully connected layer to obtain the graph neural network authority score of each node.

[0012] Optionally, the step of performing conflict graph modeling and confidence propagation iterative algorithm processing on the source authority graph to obtain a confidence propagation score includes: constructing a brand entity knowledge base, including brand product parameters, official pricing, and contact information; constructing a conflict graph based on the brand entity knowledge base, and establishing an undirected conflict edge between the two sources when the inconsistency level of two sources regarding the same brand and the same dimension is ≥ a preset threshold; setting an initial confidence level and message passing function based on the graph neural network authority score to determine the confidence level function, iterating the confidence level function according to the convergence threshold to obtain a stable confidence level value for each node; and applying a truncation function (clamp) to the stable confidence level value of each node to obtain a confidence propagation score for each node.

[0013] Optionally, the method for obtaining the timeliness factor includes: setting a configurable time window, extracting the cumulative citation frequency and citation frequency growth trend of the information source within the window, performing extreme value normalization on the cumulative citation frequency and citation frequency growth rate respectively to obtain two standardized values, and performing a weighted summation of the two standardized values ​​to obtain the timeliness factor.

[0014] Optionally, the weighted fusion of the graph neural network authority score, confidence propagation score, and timeliness factor to obtain the credibility score includes: S = 100 × clamp(α × graph neural network authority score + β × confidence propagation score + γ × timeliness factor, 0, 1), α + β + γ = 1; where S is the credibility score, clamp is the truncation function in the 0-1 interval, α, β, and γ are the weights of the graph neural network authority score, confidence propagation score, and timeliness factor, respectively, and left and right are the lower and upper bound thresholds of the clamp function, respectively.

[0015] On the other hand, this application also proposes an apparatus for evaluating the credibility of generated content. The apparatus includes: a collection module for collecting information sources from multiple platforms; a first processing module for extracting features from the information sources to obtain multi-dimensional features, including basic features, content features, citation features, and conflict features; a second processing module for defining graph nodes and edges, assigning weights, and learning graph neural networks on the multi-dimensional features to obtain a directed weighted information source authority graph; and a third processing module for performing end-to-end self-supervised learning on the information source authority graph using a graph convolutional network or a graph attention network to obtain the graph neural network authority score for each node. The fourth processing module is used to perform conflict graph modeling and confidence propagation iterative algorithm processing on the source authority map to obtain the confidence propagation score; The fifth processing module is used to set a timeliness factor and to perform a weighted fusion of the graph neural network authority score, confidence propagation score, and timeliness factor to obtain a credibility score.

[0016] On the other hand, this application also proposes a machine-readable storage medium storing instructions for causing a machine to perform the method for evaluating the credibility of generated content as described in any one of the claims of this application.

[0017] Through the aforementioned technical solutions, this application achieves real-time synchronization between evaluation results and the AI ​​platform's source citation behavior by collecting data from multiple platforms, dynamically updating source features, iteratively learning authoritative graphs, and introducing timeliness factors, overcoming the lag of traditional static weight evaluation. This application introduces conflict detection and confidence propagation algorithms into the GEO source evaluation field. By constructing a brand entity knowledge base as the benchmark truth, representing source contradictions with a conflict graph, and adjusting confidence through inhibitory message passing, it can accurately identify false and conflicting noise sources. This application extracts multiple specific features from four dimensions: basic qualifications, content quality, AI ecosystem influence, and information consistency. It uses graph neural networks to fuse node features with graph topology features, and weighted fusion to achieve a comprehensive evaluation of the three core indicators: authority, consistency, and timeliness, avoiding bias from single-dimensional evaluation. Through multi-dimensional scoring decomposition, source detail analysis, and conflict relationship tracing, it achieves dynamic, accurate, and multi-platform collaborative evaluation of source credibility, while ensuring the interpretability and operability of the evaluation results, providing a complete and reliable decision-making basis for the formulation and implementation of brand GEO strategies.

[0018] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of a method for evaluating the credibility of generated content according to the present invention; Figure 2 This is a flowchart of the multidimensional feature extraction of information sources according to the present invention; Figure 3 This is a schematic diagram illustrating the construction of the source authority map of the present invention; Figure 4 This is a flowchart of the conflict detection and belief propagation algorithm of the present invention; Figure 5 This is a schematic diagram of the evaluation system module architecture of the present invention. Detailed Implementation

[0020] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0021] It should be noted that the acquisition, transmission, storage, use, and processing of data in the technical solution of this application all comply with relevant laws and regulations. In the embodiments of this application, certain existing industry solutions such as software, components, and models may be mentioned. These should be considered exemplary, intended only to illustrate the feasibility of implementing the technical solution of this application, and do not imply that the applicant has already used or necessarily used such solutions.

[0022] Figure 1 This is a flowchart illustrating a method for evaluating the credibility of generated content according to the present invention, such as... Figure 1 As shown, this embodiment of the invention provides a method for evaluating the credibility of generated content, the method comprising: Step S101 involves collecting information sources from multiple platforms and extracting features from the information sources to obtain multi-dimensional features, which include basic features, content features, citation features, and conflict features. Step S102 involves defining graph nodes and edges, assigning weights, and learning a graph neural network to obtain a directed weighted source authority graph for the multidimensional features. Step S103 involves using a graph convolutional network or a graph attention network to perform end-to-end self-supervised learning on the source authority graph to obtain the graph neural network authority score for each node. Step S104 involves performing conflict graph modeling and confidence propagation iterative algorithm processing on the source authority graph to obtain the confidence propagation score; Step S105 involves setting a timeliness factor and weighting and fusing the graph neural network authority score, confidence propagation score, and timeliness factor to obtain a credibility score.

[0023] Through the aforementioned technical solutions, this application achieves real-time synchronization between evaluation results and the AI ​​platform's source citation behavior by collecting data from multiple platforms, dynamically updating source features, iteratively learning authoritative graphs, and introducing timeliness factors, overcoming the lag of traditional static weight evaluation. This application introduces conflict detection and confidence propagation algorithms into the GEO source evaluation field. By constructing a brand entity knowledge base as the benchmark truth, representing source contradictions with a conflict graph, and adjusting confidence through inhibitory message passing, it can accurately identify false and conflicting noise sources. This application extracts multiple specific features from four dimensions: basic qualifications, content quality, AI ecosystem influence, and information consistency. It uses graph neural networks to fuse node features with graph topology features, and weighted fusion to achieve a comprehensive evaluation of the three core indicators: authority, consistency, and timeliness, avoiding bias from single-dimensional evaluation. Through multi-dimensional scoring decomposition, source detail analysis, and conflict relationship tracing, it achieves dynamic, accurate, and multi-platform collaborative evaluation of source credibility, while ensuring the interpretability and operability of the evaluation results, providing a complete and reliable decision-making basis for the formulation and implementation of brand GEO strategies.

[0024] The aforementioned platforms include, but are not limited to, Doubao, DeepSeek, Wenxin Yiyan, Kimi, and Tongyi Qianwen, and support dynamic platform expansion.

[0025] The aforementioned information sources, including user questions and generated answers, are the data source for the entire GEO information source credibility assessment. Their functions include: 1) Locating the assessment target: User questions define the brand / industry assessment scope, ensuring that all collected information sources revolve around the target brand and industry, avoiding interference from irrelevant sources, and aligning with GEO's goal of serving the brand; 2) Extracting core information sources: AI-generated answers explicitly cite source information such as URLs and media names, as well as the cited fragments, positions, and order of the sources, serving as the sole data source for subsequent extraction of citation features; 3) Supporting feature verification: By using the cited fragments in the answers, the original source text can be matched in reverse to verify the source content features and conflict features, providing the basic verification basis for feature extraction.

[0026] Specifically, for platforms with open API interfaces, data collection is achieved through structured API requests to quickly obtain standardized structured data; for platforms without open APIs, a distributed crawler cluster with a master-slave architecture is built, adapted to anti-crawling mechanisms such as dynamic IP proxies and request frequency control, and data is crawled and structured by page parsing technologies such as XPath and CSS selectors; the data collected by both modes is uniformly aggregated into the system data buffer pool to achieve normalized management.

[0027] This method achieves comprehensive, timely, and structured data collection from mainstream domestic AI platforms through a dual-mode parallel data collection using a distributed crawler cluster and a standardized API interface. It addresses the problems of existing technologies, such as single data collection dimensions, incomplete platform coverage, and non-standard information extraction.

[0028] The aforementioned multidimensional features include extracting and quantifying multidimensional features of information sources from four dimensions, using standardized feature extraction algorithms, industry knowledge base matching, and multidimensional quantification models, with each independent information source as the unit. It uses the unique URL of the referencing information source as an identifier, combines the information source name to complete information source entityization and deduplication, merges the citation records of the same information source on different platforms and in different answers, constructs an independent information source entity database, and assigns a unique information source ID to each information source; deploys toolsets and algorithm models such as domain name information query tools, NLP natural language processing models, brand entity knowledge bases, and standardized quantification models; and formulates unified extraction rules, quantification standards, and data formats for the four dimensions of features.

[0029] The above-mentioned process of defining graph nodes and edges, assigning weights, and learning graph neural networks to obtain a directed weighted source authority graph is based on the multidimensional features of the source. Through graph modeling, defining nodes and edges, assigning initial weights, and learning graph neural networks, a directed weighted source authority graph is constructed.

[0030] The aforementioned Graph Neural Network (GNN) is a type of neural network for graph-structured data; Graph Convolutional Network (GCN) is a method that uses convolutional algorithms to aggregate information on a graph; Graph Attention Network (GAT) adds an attention mechanism to GCN, allowing nodes to focus on more important neighbors; End-to-End Self-Supervised Learning (ESL) generates supervisory signals from the data itself without manual labeling, training the entire model in one go; Belief Propagation (BP) is a classic algorithm that iteratively propagates messages on a graph to infer probabilistic relationships between nodes. The aforementioned timeliness factor refers to the citation trend and latest influence based on the impact of time.

[0031] This method enables accurate identification of false or contradictory information sources and dynamic adjustment of the confidence level of information source nodes, solving the problem that existing technologies cannot use information conflicts to infer the reliability of information sources.

[0032] In some embodiments, the collection of user information sources from multiple platforms includes: collecting data from multiple platform interfaces through a distributed crawler cluster; and processing the collected data through data cleaning, data verification, and data standardization to obtain user questions, generated answers, and information sources.

[0033] Specifically, the collected data undergoes integrity verification (removing data with missing core elements), validity verification (verifying the accessibility of the source URL and the validity of the answer content), deduplication (removing duplicates based on AI platform identifier + original user question + timestamp and source URL), format standardization (unifying timestamp, URL, and encoding formats), and compliance filtering (filtering out inappropriate information) to ensure the quality of the collected data. The cleaned and verified collected data is stored in relational and non-relational databases, categorized by AI platform, collection time, and target brand industry. Indexes are created for core fields, and each data entry is assigned a unique ID and associated with traceability information to establish a data traceability mechanism. Recently accessed high-frequency data is cached in memory to improve the efficiency of subsequent data retrieval.

[0034] The method also includes building a real-time monitoring dashboard for data collection status to enable automatic early warning of data collection anomalies; the distributed crawler cluster supports elastic scaling to enable dynamic allocation of data collection resources; and the system reserves an interface for AI platform adaptation to support dynamic expansion of platform coverage.

[0035] This method solves the problems of existing technologies, such as single feature extraction dimensions, inconsistent quantification standards, and inability to capture the core value of information sources.

[0036] In some embodiments, such as Figure 2 As shown, the basic features include domain age, third-party domain weight, global ranking, search engine indexing volume, number of backlinks, and backlink quality; the content features include industry relevance, content update frequency, structural integrity, proportion of multimedia elements, content originality, and content compliance; the citation features include total citation frequency, first citation time, most recent citation time, citation sentiment, citation position weight, and cross-platform citation breadth; and the conflict features include baseline truth value conflict degree, number of cross-source conflicts, proportion of dimensional conflicts, and conflict information correction rate.

[0037] The sources and methods for obtaining the basic features, content features, citation features, and conflict features include: using Moz / Ahrefs, Alexa / SimilarWeb, Baidu and Google search engines, the Ministry of Industry and Information Technology's filing database, and external link query tools (such as Majestic) to call platform APIs / web crawlers to capture raw data, and calculating it using standardized formulas; and using the original text of the source webpage, NLP natural language processing models, compliance detection models, and webpage structure parsing tools to obtain webpage elements (update time, image / video ratio, title / body structure), and NLP models to calculate semantic similarity (TF-IDF + cosine similarity), text repetition rate, etc.

[0038] This method extracts and quantifies source features from four dimensions, enabling quality verification and correction of feature data.

[0039] In some embodiments, the step of performing graph modeling and graph neural network learning on the multidimensional features to obtain a directed weighted source authority graph includes: constructing a node set and a directed edge set; the node set includes multiple nodes, each node corresponding to a source, and each node has an ID used to associate the basic features of the source; the directed edge set includes edge types between sources generated based on the content features, citation features, and conflict features of the source, including citation association edges, content association edges, and industry isomorphic edges; and constructing the source authority graph based on the node set and the directed edge set.

[0040] The above graph modeling includes defining a set of nodes V, where each node uniquely corresponds to an independent information source, based on "independent domain name, official account, and exclusive content carrier"; attaching multi-dimensional features as basic attributes to each node; and assigning a unique ID to each node, associating it with basic information such as the information source name and URL to achieve rapid source tracing.

[0041] In some embodiments, the reference association edge is: when information source A is the information reference source of information source B, a directed edge A→B is established; the content association edge is: when the content relevance of the two information sources is greater than or equal to a preset threshold and there is no information conflict, a bidirectional directed edge is established; the industry isomorphic edge is: when the two information sources belong to the same industry field and the target brand content attributes are consistent, a bidirectional directed edge is established. Specifically, a set of directed edges E is defined to represent the association relationship between information sources. The generation of content association edges is based on the actual reference logic of the information sources, the content relevance, and the information synergy, and conflict filtering rules are executed.

[0042] In some embodiments, such as Figure 3As shown, the step of using a graph convolutional network or a graph attention network to perform end-to-end self-supervised learning on the source authority graph to obtain the graph neural network authority score of each node includes: obtaining the node weights and edge weights of the nodes in the source authority graph, and using the node weights and multi-dimensional features as the input feature matrix; weighting and aggregating the multi-dimensional features and edge weights of the first-order or second-order neighbors of the nodes in the input feature matrix through a hidden layer to obtain the embedding vector of each node; and mapping the embedding vector to the 0-1 interval through a fully connected layer to obtain the graph neural network authority score of each node.

[0043] The node weight W(v) mentioned above takes a value of 0-1 and is obtained by weighted summation of the comprehensive scores of basic features, content features, and citation features. The formula is W(v)=αF{base}(v)+βF{content}(v)+γF{cite}(v), where α, β, and γ are the adjustable weights of basic features, content features, and citation features, respectively, satisfying α+β+γ=1, and can be dynamically adjusted according to industry attributes and brand type. The basic features mentioned above are calculated using equal-weighted summation to obtain the comprehensive score F{base}(v); the content features are calculated using weighted summation to obtain the comprehensive score F{content}(v), with the weights adjustable dynamically according to industry requirements; the citation features are calculated using weighted summation to obtain the comprehensive score F{cite}(v), with the weights adjustable according to GEO requirements; the conflict features are not included in the comprehensive score calculation but serve as the core input for subsequent conflict detection.

[0044] The edge weights ω(e) range from 0 to 1, representing the strength of the relationship between nodes. The reference-related edges are determined by the proportion of reference frequency, the weight of the reference position, and the sentiment of the reference. The content-related edges are taken from the standardized value of content relevance. The industry isomorphic edges are obtained by weighting the industry matching degree and the content attribute matching degree.

[0045] This method employs Graph Convolutional Networks (GCNs) or Graph Attention Networks (GATs) for end-to-end self-supervised learning of the graph. Node weights and multidimensional features are used as the input feature matrix. Hidden layers weight and aggregate the multidimensional features of a node's first- or second-order neighbors and edge weights to generate a low-dimensional, dense embedding vector for each node. The embedding vector is then mapped to the 0-1 interval through a fully connected layer to obtain the node's graph neural network authority score, GNNScore(v). The source authority graph is dynamic; as new sources are added, existing source features are updated, and relationships between sources change, the system automatically initializes / updates nodes, constructs / deletes edges, adjusts weights, and re-executes graph neural network learning to update the node embedding vector and the graph neural network authority score, GNNScore(v).

[0046] In some embodiments, the step of performing conflict graph modeling and confidence propagation iterative algorithm processing on the source authority graph to obtain a confidence propagation score includes: constructing a brand entity knowledge base, including brand product parameters, official pricing, and contact information; constructing a conflict graph based on the brand entity knowledge base, and establishing an undirected conflict edge between the two sources when the degree of inconsistency between two sources regarding the same brand and the same dimension is ≥ a preset threshold; setting an initial confidence level and a message passing function based on the graph neural network authority score to determine the confidence level function, iterating the confidence level function according to a convergence threshold to obtain a stable confidence level value for each node; and applying a truncated function (clamp) linear mapping to the stable confidence level value of each node to obtain the confidence propagation score for each node.

[0047] like Figure 4 As shown, the system receives the raw data and configuration parameters required for computation, serving as the input foundation for the entire process. Through node traversal, conflict determination, and edge weighting, the input data is transformed into a conflict graph structure for subsequent confidence propagation calculations. Initial confidence values ​​are set for each node in the conflict graph, completing the preparation work for iterative computation. A loop operation is executed, involving message passing, updating confidence, and calculating the confidence difference, progressively updating the confidence relationships between nodes. The system determines whether the iteration meets the termination condition: the maximum confidence difference is less than the threshold ε (max(Δb)≤ε), or the number of iterations reaches the maximum number Tmax (t≥Tmax). If the convergence condition is not met, the system returns to the confidence propagation iterative computation step to continue execution; if it is met, the system proceeds to the next stage. Based on the converged confidence results, the final BPScore score is calculated and generated.

[0048] Output the results and proceed to score fusion: Output the BPScore calculation result and pass it as input to the subsequent score fusion stage to complete the entire calculation process.

[0049] Specifically, for nodes in the node set V, conflict feature values ​​are supplemented and compared with the baseline true values; only when two information sources have clear descriptions of the same brand and the same core dimension and the degree of inconsistency is ≥ a preset threshold, an undirected conflict edge E_c is established, and a weight ω(e_c) (value 0-1, positively correlated with the degree of information contradiction) is assigned to the undirected conflict edge according to the following formula: ω(e_c) = \frac{Conf_{i,j} - Conf_{min}}{Conf_{max} - Conf_{min}} Wherein, Conf_{i,j} is the actual inconsistency quantification value between the two information sources, and Conf_{min} and Conf_{max} are the extreme values ​​of the inconsistency quantification value in this evaluation; the conflict diagram is a dynamic structure that is adjusted in real time as the information source content and brand entity knowledge base are updated.

[0050] The initial confidence level b_0(v) takes a value of 0-1 and is determined by the authority score GNNScore(v) of the graph neural network, i.e., b_0(v) = GNNScore(v); the message passing function m_{i→j}(t) is the confidence message passed from node i to j in the t-th iteration, taking a value of 0-1, and the formula is m_{i→j}(t) = 1-ω(e_{ij})×b_{t-1}(i), which represents the degree of inhibitory influence.

[0051] The confidence function b_t(v) is determined based on the initial confidence and the message passing function: the confidence update value of node v in the t-th iteration, with the formula b_t(v)=b_0(v)×\prod_{i∈N(v)}\frac{1}{|N(v)|}×m_{i→v}(t), where N(v) is the set of conflicting neighbors of node v; A convergence threshold ε (default 0.001) is set. Iteration stops when the maximum change in confidence between two consecutive iterations of all nodes is ≤ ε, or when convergence is not achieved after reaching the maximum number of iterations (default 100). The stable confidence values ​​of the nodes after the algorithm terminates are linearly mapped to the 0-1 interval (0 for values ​​less than 0, 1 for values ​​greater than 1) using the truncation function clamp, resulting in the confidence propagation score BPScore(v) for each node. A higher confidence propagation score indicates stronger consistency of the information source and higher confidence. This method achieves a linked evaluation of source authority and information consistency through strong suppression of false sources, enhancement of confidence in consistent sources, and refined handling of local conflicts.

[0052] In some embodiments, the method for obtaining the timeliness factor includes: setting a configurable time window, extracting the cumulative citation frequency and citation frequency growth trend of the information source within the window, performing extreme value normalization on the cumulative citation frequency and citation frequency growth rate respectively to obtain two standardized values, and performing a weighted summation of the two standardized values ​​to obtain the timeliness factor. The cumulative citation frequency and citation frequency growth rate are determined based on the citation characteristics.

[0053] For example, based on citation characteristics, the TemporalFactor(v) is calculated, with a value ranging from 0 to 1. A configurable time window is set (default 30 days). The cumulative citation frequency C_t and the citation frequency growth trend T_r (citation frequency growth rate) of the information source within the window are extracted. Extreme value normalization is performed on the cumulative citation frequency and the citation frequency growth rate to obtain standardized values ​​F_c and F_r. The weighted sum of the above standardized values ​​F_c and F_r yields the TemporalFactor(v) = λ × F_c + (1-λ) × F_r, where λ is the trend weight coefficient (0-1, default 0.5), which can be adjusted according to GEO requirements. For information sources with no citation records within the time window, TemporalFactor(v) = 0; for newly added information sources with no growth trend, F_r = 0. This method quantifies the recent citation trend and latest influence of information sources in the AI ​​platform.

[0054] In some embodiments, the weighted fusion of the graph neural network authority score, confidence propagation score, and timeliness factor to obtain the credibility score includes: S = 100 × clamp(α × graph neural network authority score + β × confidence propagation score + γ × timeliness factor, 0, 1), α + β + γ = 1; where S is the credibility score, clamp is the truncation function in the 0-1 interval, α, β, and γ are the weights of the graph neural network authority score, confidence propagation score, and timeliness factor, respectively, and left and right are the lower and upper bound thresholds of the clamp function, respectively.

[0055] For example, a linear weighted fusion algorithm can be used to fuse the graph neural network authority score, confidence propagation score, and timeliness factor into a dynamic credibility score S_v on a 0-100 scale: S_v = 100 × clamp(α × GNNScore(v) + β × BPScore(v) + γ × TemporalFactor(v), 0, 1), where clamp is the truncation function to ensure that the fused intermediate value is in the range of 0-1, and α, β, and γ are dimension weight parameters that satisfy α + β + γ = 1. These can be dynamically adjusted according to industry attributes, brand development stage, and GEO evaluation goals. The weight adaptation can be based on the following: mature brands α = 0.4, β = 0.4, γ = 0.2; emerging brands α = 0.2, β = 0.3, γ = 0.5; and highly information-sensitive industries α = 0.3, β = 0.5, γ = 0.2. Automatic optimization via machine learning is also supported.

[0056] Specifically, information sources can be divided into 5 credibility levels based on a 0-100 score scale: 90-100 points is Level 1 (high credibility, core targeting), 80-89.9 points is Level 2 (relatively high credibility, key targeting), 60-79.9 points is Level 3 (medium credibility, general targeting), 40-59.9 points is Level 4 (low credibility, restricted targeting), and 0-39.9 points is Level 5 (low credibility, prohibited targeting). The levels can be slightly adjusted according to industry. This method outputs full-dimensional information of a single information source in a structured data format, including source ID, name, URL, three core indicators, dynamic credibility score, credibility level, weight parameters, and scoring timestamp. It sorts all information sources from high to low scores and supports filtering by credibility level, industry type, citation frequency, and other dimensions to generate a list of high-credibility information sources and a warning list of low-credibility information sources.

[0057] Its scoring results are dynamic and real-time, adaptable and flexible, interpretable and traceable, operable and implementable.

[0058] Furthermore, this application proposes a multi-dimensional interactive visualization dashboard to transform quantitative assessment data into intuitive, interpretable, and actionable guidance for GEO strategies. This addresses the problem of existing assessment results being abstract and disconnected from actual marketing scenarios, achieving a closed loop from data assessment to strategy implementation. The specific execution process includes: normalizing, classifying, aggregating, and annotating outliers in the data to generate standardized datasets for visualization and reporting; building an interactive visualization dashboard that supports PC / mobile adaptation, data drill-down, and real-time updates, with all charts supporting export, screenshots, and linked data queries; and presenting basic assessment information and sources on a large screen. The system provides quick access to a list of core / warning information sources, including grade distribution, core scoring metrics, and GEO core value metrics. It displays a ranking of information source scores, supporting single-source dimension breakdown, graph-based display, and conflict details presentation, addressing the issue of score interpretability. It presents the time-varying trends of information source scores, citation frequency, and timeliness factors, reflecting the dynamics of AI platform information source citation preferences. A heatmap displays the conflict network, statistically analyzes high-conflict dimensions, and generates a red alert list of low-credibility information sources, enabling real-time warnings of false information sources. It compares the information source coverage, scoring preferences, and high-credibility information source lists of various AI platforms, providing a basis for cross-platform GEO strategies.

[0059] According to a specific implementation method, such as Figure 5As shown, this application completes the collection and cleaning of raw information through data acquisition and preprocessing, outputting standardized source data to provide basic data in a unified format for subsequent analysis. Feature extraction performs feature mining on the standardized source data to generate multi-dimensional features for graph construction, conflict detection, and confidence propagation. The source authority graph construction is based on the multi-dimensional feature dataset, providing support for subsequent score fusion and conflict detection. Conflict detection and confidence propagation simultaneously receive the multi-dimensional features and the source authority graph, generating a confidence propagation score by detecting contradictions and conflicts between information and the confidence of propagation nodes. Dynamic score fusion integrates the confidence propagation score output from the graph construction with the conflict detection and confidence propagation score to obtain a credibility score, completing the quantitative assessment and result presentation of information credibility.

[0060] On the other hand, this application also proposes an apparatus for evaluating the credibility of generated content. This apparatus includes: a collection module for collecting information sources from multiple platforms, extracting features from the information sources to obtain multi-dimensional features, including basic features, content features, citation features, and conflict features. It serves as the system data input entry point, configuring multi-platform collection adaptation interfaces to achieve dual-mode data collection via distributed crawler clusters and standardized API interfaces; supporting flexible configuration and triggering of collection tasks, completing the cleaning, verification, and standardization of collected data; implementing data storage according to the principle of database sharding and table partitioning, establishing data indexing and traceability mechanisms, providing collection status monitoring and anomaly warnings, and supporting elastic scaling of collection resources and dynamic platform adaptation. Furthermore, it achieves information source entityization and deduplication, constructing and maintaining an independent information source entity library; deploying the toolset and algorithm models required for feature extraction, automatically extracting and quantifying information source features from four dimensions, completing the quality verification and correction of feature data; constructing and dynamically updating the information source feature library, achieving full-link association with the original collected data, and supporting rapid querying and indexing of feature data.

[0061] The first processing module is used to define graph nodes and edges, assign weights, and learn a directed weighted source authority graph by performing graph neural network learning on the multidimensional features. This module automatically constructs and dynamically updates the directed weighted source authority graph, completing the definition of nodes and edges and the assignment of initial weights.

[0062] The second processing module is used to perform end-to-end self-supervised learning on the source authority graph using a graph convolutional network or a graph attention network to obtain the graph neural network authority score for each node. This module deploys a graph convolutional network (GCN) / graph attention network (GAT) model to achieve end-to-end self-supervised learning of the graph, generating low-dimensional embedding vectors for nodes; and calculates and outputs the graph neural network authority score GNNScore(v) for each node, providing data support for subsequent conflict detection and score fusion.

[0063] The third processing module is used to perform conflict graph modeling and confidence propagation iterative algorithm processing on the source authority graph to obtain the confidence propagation score. This module constructs and synchronously updates the brand entity knowledge base as the benchmark truth value for information comparison; reuses the node set of the authority graph to automatically construct a dynamic undirected weighted conflict graph, and completes the definition and weight assignment of conflict edges; executes the confidence propagation iterative algorithm to realize the dynamic adjustment and convergence determination of the confidence of the source nodes; calculates and outputs the confidence propagation score BPScore(v) of each node, and generates a source conflict relationship report.

[0064] The fourth processing module is used to set a timeliness factor and perform weighted fusion of the graph neural network authority score, confidence propagation score, and timeliness factor to obtain a credibility score. This module automatically calculates the timeliness factor TemporalFactor(v) based on the source citation feature; configures adjustable weight parameters, calculates a dynamic credibility score of 0-100 points through a linear weighted fusion algorithm, and completes the hierarchical classification of the scoring results; outputs the single source full-dimensional evaluation results in a structured data format, sorts and classifies all sources, and generates a high-credibility source list and a low-credibility source warning list.

[0065] The method for evaluating the credibility of generated content according to the present invention includes: collecting information sources from multiple platforms; extracting features from the information sources to obtain multi-dimensional features, including basic features, content features, citation features, and conflict features; defining graph nodes and edges, assigning weights, and learning graph neural networks to obtain a directed weighted information source authority graph; using graph convolutional networks or graph attention networks to perform end-to-end self-supervised learning on the information source authority graph to obtain the graph neural network authority score of each node; performing conflict graph modeling and confidence propagation iterative algorithm processing on the information source authority graph to obtain a confidence propagation score; setting a timeliness factor, and weighting and fusing the graph neural network authority score, confidence propagation score, and timeliness factor to obtain a credibility score.

[0066] Through the aforementioned technical solutions, this application achieves real-time synchronization between evaluation results and the AI ​​platform's source citation behavior by collecting data from multiple platforms, dynamically updating source features, iteratively learning authoritative graphs, and introducing timeliness factors, overcoming the lag of traditional static weight evaluation. This application introduces conflict detection and confidence propagation algorithms into the GEO source evaluation field. By constructing a brand entity knowledge base as the benchmark truth, representing source contradictions with a conflict graph, and adjusting confidence through inhibitory message passing, it can accurately identify false and conflicting noise sources. This application extracts multiple specific features from four dimensions: basic qualifications, content quality, AI ecosystem influence, and information consistency. It uses graph neural networks to fuse node features with graph topology features, and weighted fusion to achieve a comprehensive evaluation of the three core indicators: authority, consistency, and timeliness, avoiding bias from single-dimensional evaluation. Through multi-dimensional scoring decomposition, source detail analysis, and conflict relationship tracing, it achieves dynamic, accurate, and multi-platform collaborative evaluation of source credibility, while ensuring the interpretability and operability of the evaluation results, providing a complete and reliable decision-making basis for the formulation and implementation of brand GEO strategies.

[0067] The device for assessing the credibility of the generated content includes a processor and a memory. The aforementioned acquisition module, first processing module, second processing module, third processing module, and fourth processing module are all stored in the memory as program units. The processor executes the aforementioned program units stored in the memory to achieve the corresponding functions.

[0068] The processor contains a kernel, which retrieves the corresponding program units from memory. One or more kernels can be configured, and by adjusting kernel parameters, dynamic, accurate, and multi-platform collaborative evaluation of information source reliability can be achieved.

[0069] The memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0070] This invention provides a storage medium storing a program that, when executed by a processor, implements the method for evaluating the credibility of generated content.

[0071] This invention provides a processor for running a program, wherein the program executes the method for evaluating the credibility of generated content during runtime.

[0072] This invention provides a device including a processor, a memory, and a program stored in the memory and executable on the processor. When the processor executes the program, it performs the following steps: acquiring information sources from multiple platforms; extracting features from the information sources to obtain multi-dimensional features, including basic features, content features, citation features, and conflict features; defining graph nodes and edges, assigning weights, and learning a graph neural network to obtain a directed weighted information source authority graph; performing end-to-end self-supervised learning on the information source authority graph using a graph convolutional network or a graph attention network to obtain the graph neural network authority score for each node; performing conflict graph modeling and a confidence propagation iterative algorithm on the information source authority graph to obtain a confidence propagation score; setting a timeliness factor, and weightedly fusing the graph neural network authority score, the confidence propagation score, and the timeliness factor to obtain a credibility score. The device in this paper can be a server, PC, PAD, mobile phone, etc.

[0073] This application also provides a computer program product, which, when executed on a data processing device, is suitable for executing an initialization program with the following method steps: collecting information sources from multiple platforms; extracting features from the information sources to obtain multi-dimensional features, the multi-dimensional features including basic features, content features, citation features, and conflict features; defining graph nodes and edges, assigning weights, and learning graph neural networks on the multi-dimensional features to obtain a directed weighted information source authority graph; using a graph convolutional network or graph attention network to perform end-to-end self-supervised learning on the information source authority graph to obtain the graph neural network authority score of each node; performing conflict graph modeling and confidence propagation iterative algorithm processing on the information source authority graph to obtain a confidence propagation score; setting a timeliness factor, and weightedly fusing the graph neural network authority score, confidence propagation score, and timeliness factor to obtain a credibility score.

[0074] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0075] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0076] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0077] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0078] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0079] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0080] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0081] It should also be noted that 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. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0082] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for evaluating the credibility of generated content, characterized in that, The method includes: Information sources from multiple platforms are collected, and feature extraction is performed on the information sources to obtain multidimensional features, including basic features, content features, citation features, and conflict features. The multidimensional features are subjected to graph node and edge definition, weight assignment, and graph neural network learning to obtain a directed weighted source authority graph; The graph neural network authority score of each node is obtained by end-to-end self-supervised learning of the source authority graph using graph convolutional networks or graph attention networks. The confidence propagation score is obtained by performing conflict graph modeling and confidence propagation iterative algorithm processing on the aforementioned source authority graph; A timeliness factor is set, and the credibility score is obtained by weighting and fusing the authority score, confidence propagation score, and timeliness factor of the graph neural network.

2. The method according to claim 1, characterized in that, The user information sources collected from multiple platforms include: Data was collected from multiple platform interfaces using a distributed crawler cluster. After data cleaning, data verification, and data standardization, the collected data is processed to obtain user questions, generated answers, and information sources.

3. The method according to claim 1, characterized in that, The basic characteristics include domain age, third-party domain authority, global ranking, search engine indexing volume, number of backlinks, and quality of backlinks; The content characteristics include industry relevance, content update frequency, structural integrity, proportion of multimedia elements, content originality, and content compliance. The citation features include total citation frequency, first citation time, most recent citation time, citation sentiment, citation position weight, and cross-platform citation breadth; The conflict characteristics include the baseline truth value conflict degree, the number of cross-source conflicts, the proportion of dimensional conflicts, and the conflict information correction rate.

4. The method according to claim 1, characterized in that, The process of performing graph modeling and graph neural network learning on the multidimensional features to obtain a directed weighted source authority graph includes: Construct the set of nodes and the set of directed edges; The node set includes multiple nodes, each corresponding to a source. Each node has an ID to associate with the basic features of the source. The directed edge set includes edge types between sources generated based on the content features, reference features, and conflict features of the source, including reference association edges, content association edges, and industry isomorphic edges. Construct a source authority graph based on the set of nodes and the set of directed edges.

5. The method according to claim 4, characterized in that, The reference association edge is: when information source A is the information reference source of information source B, a directed edge A→B is established; The content association edge is defined as follows: when the content relevance between two information sources is greater than or equal to a preset threshold and there is no information conflict, a bidirectional directed edge is established; The industry isomorphic edge is defined as follows: two information sources belong to the same industry field and have the same target brand content attributes, and a bidirectional directed edge is established.

6. The method according to claim 1, characterized in that, The method of using graph convolutional networks or graph attention networks to perform end-to-end self-supervised learning on the source authority graph to obtain the graph neural network authority score of each node includes: Obtain the node weights and edge weights of the nodes in the source authority graph, and use the node weights and multidimensional features as the input feature matrix; The embedding vector of each node is obtained by weighted aggregation of the multidimensional features and edge weights of the first-order or second-order neighbors of the nodes in the input feature matrix through the hidden layer. The embedded vectors are mapped to the 0-1 interval through a fully connected layer to obtain the graph neural network authority score of each node.

7. The method according to claim 1, characterized in that, The process of obtaining a confidence propagation score by performing conflict graph modeling and confidence propagation iterative algorithm processing on the source authority graph includes: Build a brand entity knowledge base, including brand product parameters, official pricing, and contact information; A conflict graph is constructed based on the brand entity knowledge base. When the degree of inconsistency between two sources regarding the same brand and the same dimension is greater than or equal to a preset threshold, an undirected conflict edge is established between the two sources. An initial confidence level and message passing function are set based on the authority score of the graph neural network to determine the confidence level function. The confidence level function is then iterated based on the convergence threshold to obtain the stable confidence level value of the node. The confidence propagation score of each node is obtained by applying a truncation function (clamp) to the node stability confidence value.

8. The method according to claim 1, characterized in that, The method for obtaining the timeliness factor includes: setting a configurable time window, extracting the cumulative citation frequency and citation frequency growth trend of the information source within the window, performing extreme value normalization on the cumulative citation frequency and citation frequency growth rate respectively to obtain two standardized values, and performing a weighted summation of the two standardized values ​​to obtain the timeliness factor.

9. The method according to claim 1, characterized in that, The credibility score is obtained by weighting and fusing the authority score, confidence propagation score, and timeliness factor of the graph neural network, including: S = 100 × clamp (α × graph neural network authority score + β × confidence propagation score + γ × timeliness factor, 0, 1), α + β + γ = 1; Where S is the credibility score, clamp is the truncation function for the 0-1 interval, α, β, and γ are the weights of the graph neural network authority score, confidence propagation score, and timeliness factor, respectively, and left and right are the lower and upper bound thresholds of the clamp function, respectively.

10. An apparatus for evaluating the credibility of generated content, characterized in that, The device includes: The acquisition module is used to acquire information sources from multiple platforms, and to extract features from the information sources to obtain multi-dimensional features, including basic features, content features, citation features, and conflict features. The first processing module is used to define graph nodes and edges, assign weights, and learn a directed weighted source authority graph by performing graph neural network learning on the multidimensional features. The second processing module is used to perform end-to-end self-supervised learning on the source authority graph using a graph convolutional network or a graph attention network to obtain the graph neural network authority score of each node; The third processing module is used to perform conflict graph modeling and confidence propagation iterative algorithm processing on the source authority graph to obtain the confidence propagation score. The fourth processing module is used to set a timeliness factor and perform weighted fusion of the graph neural network authority score, confidence propagation score, and timeliness factor to obtain a credibility score.

11. A machine-readable storage medium storing instructions thereon, characterized in that, This instruction is used to cause the machine to perform the method for evaluating the credibility of generated content as described in any one of claims 1-9 of this application.

Images