Information processing method and system based on knowledge graph

Abstract

The invention discloses an information processing method and system based on a knowledge graph, relates to the technical field of artificial intelligence and the knowledge graph, and solves the problems that an evidence chain related to an AI decision is difficult to extract from the knowledge graph, the evaluation dimension of the evidence chain is single, a real-time conflict detection mechanism is lacked, information space-time conflicts and logical axiom violation cannot be found in time, and the reliability of the knowledge graph is improved. And the technical problem of low reliability of the knowledge graph data is solved. Comprising the steps of obtaining an AI model output result, screening a core entity through a decision centrality score, and realizing entity association matching in combination with a knowledge graph hierarchical structure and a bidirectional semantic similarity model; taking an associated entity as a starting point, extracting associated nodes and relationships through bidirectional backtracking retrieval, and screening and visualizing an optimal evidence chain by adopting a preliminary evaluation-causal contribution degree-robustness three-dimensional quantitative model; based on an OWL axiom library and a business rule library, a comprehensive conflict index is calculated through a multi-level conflict detection engine, and multi-source information conflict early warning is achieved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and knowledge graph technology, specifically relating to an information processing method and system based on knowledge graphs. Background Technology

[0002] With the deep application of artificial intelligence technology in fields such as healthcare, finance, and industry, the black-box nature of AI models leads to a lack of interpretability in the decision-making process, making it difficult to trace the basis of decisions and limiting their application in high-precision scenarios (medical diagnosis result determination, financial risk assessment). For example, in medical diagnosis, the disease determination results output by AI models cannot be linked to supporting evidence such as patient medical history and examination data, making it difficult for doctors to verify the credibility of the results.

[0003] Meanwhile, knowledge graphs, as the core carrier of structured knowledge storage and association, have been widely used in information retrieval and knowledge reasoning. However, existing technologies have the following technical problems when co-processing with AI models: Traditional entity matching relies solely on single semantic similarity calculation (cosine similarity), without combining decision-making features such as attention weights and classification confidence of AI models, leading to bias in core entity selection and insufficient reliability of the association between knowledge graphs and AI output results; existing evidence chain screenings mostly focus on semantic relevance, without considering the causal contribution of evidence to AI decision-making and its robustness, resulting in evidence chains that cannot effectively support AI decision interpretation; after newly added multi-source information (patient examination reports from different data sources, financial data from different institutions) is linked with the knowledge graph, there is a lack of a real-time conflict detection mechanism based on spatiotemporal consistency and logical axioms, which easily leads to inconsistencies in knowledge graph data and affects the accuracy of subsequent reasoning and decision-making.

[0004] While some existing technologies attempt to combine knowledge graphs with AI models, they have not formed a complete technical system that integrates precise entity association, multidimensional evidence evaluation, and real-time conflict detection. This system cannot simultaneously address the core issues of interpretability in AI decision-making and consistency of knowledge graph data. There is an urgent need for an information processing method that balances accuracy, interpretability, and reliability. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art; to this end, this invention proposes an information processing method and system based on knowledge graphs to solve the following technical problems: Existing technologies cannot extract evidence chains that are strongly relevant to AI decision-making from knowledge graphs, and the evaluation dimensions of evidence chains are singular, failing to verify the supporting value of evidence for decision-making. This results in a lack of credible explanatory basis for AI output results. Traditional entity matching does not combine the decision-making characteristics of AI models (attention weights, classification confidence) with the hierarchical structure of knowledge graphs, leading to large biases in the selection of core entities and insufficient reliability of association matching. After new information is linked to the knowledge graph, there is a lack of real-time, multi-dimensional conflict detection mechanisms, making it impossible to detect spatiotemporal conflicts and violations of logical axioms in a timely manner, resulting in low credibility of knowledge graph data.

[0006] To address the above problems, a first aspect of the present invention provides an information processing method based on knowledge graphs, comprising the following steps: S1: Obtain the output results of the AI ​​model, extract the core entities and entity relationships in the output results, wherein the core entities are entities obtained by filtering through decision centrality scores, and are determined to be core entities when the score is greater than a dynamic threshold; Based on the hierarchical structure of the knowledge graph, the entity type is used for pre-screening. The selected candidate entities in the knowledge graph are matched with the core entity using a bidirectional semantic similarity calculation model. If the matching result meets the preset confidence threshold, the corresponding candidate entity in the core entity relationship knowledge graph is determined to be successfully associated. S2: After successful entity association, starting from the successfully associated entity, perform bidirectional backtracking retrieval in the knowledge graph to extract the associated nodes and relationships that support the output results of the AI ​​model, quantitatively evaluate the generated data chain, sort and filter the evidence chain according to the final scoring function, and visualize the optimal evidence chain in a graph form. S3: When new multi-source information is added to the database and linked to the entity in the knowledge graph, the predefined OWL axiom library and business rule library are called. Through the multi-level contradiction detection engine, the comprehensive conflict index is calculated. If the comprehensive conflict index is greater than or equal to the preset conflict threshold, a contradiction warning is immediately generated and issued.

[0007] Preferably, step S1 includes the following steps: The AI ​​model outputs text decision statements, classification confidence vectors, and attention weights. It employs a domain-adaptive NER model to parse the text, with its loss function... ,in, This represents the total loss value. This represents the loss of the standard named entity recognition task. Indicates hyperparameters, This indicates the auxiliary loss based on decision-making keywords; Simultaneously, a weighted decision semantic subgraph is constructed using confidence and attention weights. The decision semantic subgraph The nodes are the extracted entities, and the edge weights are the average of the attention weights between entities; A hierarchical semantic routing table is constructed based on the ontology layer and data layer of the knowledge graph. Matching is performed by calculating the cosine similarity between entity type embedding and knowledge graph type node embedding, and joint query is performed by combining relational phrases extracted from the decision text. The data layer subgraph that simultaneously satisfies type constraints and relational predicates is used as the candidate range after filtering.

[0008] Preferably, the decision centrality score includes the following steps: Core entities are filtered through a dynamic context gating mechanism, and the computational entities are calculated. Decision centrality score:

[0009] in, Representing entities Decision centrality score, This represents the Sigmoid function. This refers to the information about entities extracted from the attention mechanism of AI models. Average attention weight, Representation and entity The relevant classification confidence, Represents the balance coefficient. Representing entities In the decision semantic subgraph Degree centrality in; Only when >Entities with dynamic thresholds are treated as core entities.

[0010] Preferably, in step S1, the bidirectional semantic similarity calculation model includes the following steps: The specific bidirectional semantic similarity calculation model is as follows:

[0011] in, Represents the core entity With knowledge graph candidate entities The overall similarity between them This represents the cosine similarity calculation function based on entity embedding vectors. This represents the Jaccard similarity function used to calculate the attribute set of an entity. This represents a function for calculating the overlap of entity relationships. , and This represents the weight coefficients that are dynamically adjusted through gradient boosting of the tree model; When the overall similarity of a candidate entity is greater than or equal to a preset confidence threshold, the core entity is determined to be successfully associated with the corresponding candidate entity in the knowledge graph.

[0012] Preferably, the weight coefficients dynamically adjusted through the gradient boosting tree model include the following steps: Extract the corresponding real-time meta-feature vector for each entity matching task; The meta-feature vector is input into an attention-enhanced gradient boosting tree model, the model comprising: Base predictor: Employs a standard gradient boosting decision tree, taking a meta-feature vector as input. This meta-feature vector includes task context features, entity pair static features, and dynamic interaction features. Through an iterative training process using gradient boosting, it outputs an initial weight vector. ; Attention-gated network: Constructs a shallow neural network structure. The input layer is a meta-feature vector, the hidden layer performs feature transformation using the ReLU activation function, and the output layer outputs an attention-gated vector. ; The final weight vector is obtained by element-wise multiplying the initial weight vector with the attention gating vector. .

[0013] Preferably, in step S2, the bidirectional backtracking retrieval in the knowledge graph includes the following steps: Define an entity-relationship association path template library at the knowledge graph ontology layer, including forward and reverse association path templates based on business scenarios; Starting from the successfully associated entity, and based on the core theme of the AI ​​model output, the system matches suitable forward and reverse path templates from the path template library. If multiple suitable templates exist, the optimal path template is selected by calculating semantic similarity. According to the optimal path template, retrieve the first-level nodes and relationships directly associated with the core entity, and record the relationship type, node attributes and path confidence of the associated path; For nodes obtained through direct association, the secondary nodes and relationships are traced back along the template path. Simultaneously, the evidence contribution of the knowledge graph nodes is calculated. ,in, This indicates the credibility of the relationship between the node and the core entity. This indicates the degree of matching between node attributes and the output of the AI ​​model. and These are the weighting coefficients for relationship credibility and matching degree, respectively; When the node evidence contribution is less than a preset contribution threshold, the backtracking of the template path is stopped. The associated nodes and relationships obtained through forward and reverse backtracking are deduplicated based on the unique identifier of the knowledge graph node and the relationship triple, and then integrated into a knowledge graph association evidence set.

[0014] Preferably, in step S2, sorting and filtering the evidence chain according to the final scoring function includes the following steps: The knowledge graph-related evidence set is used to generate multiple evidence chains according to a fixed path, and each evidence chain is stored in the form of a triple sequence. Construct a preliminary evaluation index set, including node semantic relevance, path reliability and evidence integrity, calculate the preliminary credibility score of each evidence chain, and select the Top-k candidate evidence chains; For each candidate evidence chain selected, a counterfactual scenario is constructed in the knowledge graph. By temporarily obscuring key nodes or relationships in the evidence chain, the case where the evidence does not exist is simulated. The confidence of the AI ​​model's decision result is recalculated on the counterfactual graph, and the causal contribution of the evidence chain is calculated by the change in confidence. A dynamic adversarial example generator is constructed to generate semantically preserved perturbation evidence chains for candidate evidence chains. The robustness score of the evidence chain is calculated by comparing the credibility difference between the original evidence chain and the perturbation evidence chain. A final evaluation function is constructed that integrates the initial credibility score, causal contribution score, and robustness score, and evidence chains with high causal contribution and strong robustness are preferentially selected as the optimal explanation.

[0015] Preferably, in step S3, the multi-level conflict detection engine calculates the comprehensive conflict index, including the following steps: A spatiotemporal consistency conflict detection model is constructed, and the temporal conflict degree, spatial conflict degree, and axiom violation score are calculated respectively. Design an adaptive weight allocation mechanism to dynamically adjust the weight coefficients of various conflict types based on their severity. Specifically: ,in, , and The final weights of temporal conflict, spatial conflict, and axiom violation scores in the calculation of the comprehensive conflict index are represented by Stemp, Sspatial, and Slogic, which represent the severity scores of temporal conflict, spatial conflict, and logical conflict, respectively. The comprehensive conflict index is obtained by adding the final weights corresponding to various dimensions with the time conflict degree, spatial conflict degree and axiom violation score.

[0016] A second aspect of the present invention provides an information processing system based on a knowledge graph, comprising the following modules: Core Entity Extraction and Knowledge Graph Association Matching Module: Obtain the output results of the AI ​​model, extract the core entities and entity associations from the output results, wherein the core entities are entities obtained by filtering through decision centrality scores, and are determined to be core entities when the score is greater than a dynamic threshold; Based on the hierarchical structure of the knowledge graph, the entity type is used for pre-screening. The selected candidate entities in the knowledge graph are matched with the core entity using a bidirectional semantic similarity calculation model. If the matching result meets the preset confidence threshold, the corresponding candidate entity in the core entity relationship knowledge graph is determined to be successfully associated. Knowledge Graph Evidence Chain Backtracking Retrieval and Quantitative Evaluation Display Module: After successful entity association, starting from the successfully associated entity, bidirectional backtracking retrieval is performed in the knowledge graph to extract the associated nodes and relationships that support the output results of the AI ​​model. The generated data chain is quantitatively evaluated, the evidence chain is sorted and filtered according to the final scoring function, and the optimal evidence chain is visualized in a graph form. Multi-source information conflict detection and warning generation module: When new multi-source information is added to the database and linked to the entity in the knowledge graph, the predefined OWL axiom library and business rule library are called. Through the multi-level conflict detection engine, the comprehensive conflict index is calculated. If the comprehensive conflict index is greater than or equal to the preset conflict threshold, a conflict warning is immediately generated and issued.

[0017] The beneficial effects of this invention are: This invention integrates the attention weights, classification confidence, and degree centrality of the decision semantic subgraph of an AI model, and uses the corresponding calculated scores to filter core entities in combination with dynamic thresholds (determined based on cross-validation). This achieves strong correlation between the filtered entities and AI decisions. At the same time, a bidirectional semantic similarity multi-dimensional matching model is designed. By comprehensively considering the cosine similarity of entity embedding vectors, the Jaccard similarity of attribute sets, and the overlap of relationships, and by dynamically adjusting the weight coefficients through an attention-enhanced gradient boosting tree model, the accuracy of entity association matching is improved. This invention achieves three-dimensional verification of evidence chain relevance, causality, and robustness by generating evidence chains, conducting preliminary assessments, calculating causal contribution, evaluating robustness, and finally selecting scoring functions. It also designs a graph-based visualization display, which maps evidence credibility to node size and importance to relationships to intuitively present the correlation logic between the optimal evidence chain and AI decision-making, thereby improving interpretability. This invention constructs a spatiotemporal consistency conflict detection model and axiom violation detection, covering multi-dimensional conflicts in information spatiotemporal and logical dimensions. It also proposes an adaptive weight allocation mechanism, which dynamically adjusts the weights of each conflict dimension and sets a scoring range based on the degree of business impact, ensuring that the comprehensive conflict index can reflect the severity of data information conflicts and achieving accurate early warning. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the module flow of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 As shown, this invention is an information processing method based on knowledge graphs, comprising the following steps: S1: Obtain the output results of the AI ​​model, extract the core entities and entity relationships in the output results, wherein the core entities are entities obtained by filtering through decision centrality scores, and are determined to be core entities when the score is greater than a dynamic threshold; Based on the hierarchical structure of the knowledge graph, the entity type is used for pre-screening. The selected candidate entities in the knowledge graph are matched with the core entity using a bidirectional semantic similarity calculation model. If the matching result meets the preset confidence threshold, the corresponding candidate entity in the core entity relationship knowledge graph is determined to be successfully associated. S2: After successful entity association, starting from the successfully associated entity, perform bidirectional backtracking retrieval in the knowledge graph to extract the associated nodes and relationships that support the output results of the AI ​​model, quantitatively evaluate the generated data chain, sort and filter the evidence chain according to the final scoring function, and visualize the optimal evidence chain in a graph form. S3: When new multi-source information is added to the database and linked to the entity in the knowledge graph, the predefined OWL axiom library and business rule library are called. Through the multi-level contradiction detection engine, the comprehensive conflict index is calculated. If the comprehensive conflict index is greater than or equal to the preset conflict threshold, a contradiction warning is immediately generated and issued.

[0021] In one embodiment of the present invention, step S1 includes the following steps: The AI ​​model outputs text decision statements, classification confidence vectors, and attention weights. It employs a domain-adaptive NER model to parse the text, with its loss function... ,in, This represents the total loss value. This represents the loss of the standard named entity recognition task. Indicates hyperparameters, This indicates the auxiliary loss based on decision-making keywords; Simultaneously, a weighted decision semantic subgraph is constructed using confidence and attention weights. The decision semantic subgraph The nodes are the extracted entities, and the edge weights are the average of the attention weights between entities; A hierarchical semantic routing table is constructed based on the ontology layer and data layer of the knowledge graph. Matching is performed by calculating the cosine similarity between entity type embedding and knowledge graph type node embedding, and joint query is performed by combining relational phrases extracted from the decision text. The data layer subgraph that simultaneously satisfies type constraints and relational predicates is used as the candidate range after filtering.

[0022] Specifically, in this embodiment, for the credit approval scenario of financial institutions, the AI ​​model is a BERT-based credit risk classification model. Its output is used to determine whether a user's "credit application is approved" or "credit application is rejected". Three types of output data are obtained from the AI ​​model's inference interface, including: 1. A text decision statement: "The applicant with user ID 20230501001 has a credit application amount of 500,000 yuan, a credit rating of A, and a comprehensive risk score of 82. Credit approval is approved."; 2. A classification confidence vector: [0.92, 0.08], where index 0 corresponds to the "approved" category and index 1 corresponds to the "rejected" category, i.e., the confidence level of "approved" is C=0.92; 3. An attention weight matrix: the dimension is [number of text tokens × number of text tokens], where the tokens are "user ID 20230501001", "500,000 yuan", and "credit rating A". The mean attention weights for "level" are A1=0.85, A2=0.62, and A3=0.78 (calculated by averaging the attention head weights of the AI ​​model); a BERT-NER model pre-trained on financial domain corpus is used to parse the above text decision statements, and the model loss function is set to... ,in, This represents a hyperparameter, which was determined to be 0.3 through 5-fold cross-validation. Calculated using negative logarithmic loss, Cross-entropy loss is employed; using the BERT-NER model, four types of entities are extracted from the text, including user entities, monetary entities, credit entities, and rating entities. Among them, the four types of entities extracted are used as decision semantic subgraphs. The nodes, whose attributes include entity type, entity value, and AI model classification confidence, are used to calculate the average attention weight among entities based on the attention weight matrix output by the AI ​​model, which is then used as a subgraph. Edge weights; statistics on the edge weights of each node. The number of connected edges (degree) in the interval is used as the degree centrality. The financial knowledge graph is divided into an ontology layer and a data layer. The ontology layer defines 5 core entity types (USER, AMOUNT, CREDIT, SCORE, LOAN), 4 relationship types (user-application-amount, user-association-credit, user-possess-score, amount-corresponding-loan), and attribute constraints (e.g., the USER entity must include user ID, age, and occupation attributes, and the CREDIT entity must include credit rating and number of overdue payments attributes). The data layer stores 1 million+ entity instances and 2 million+ relationship instances. The hierarchical semantic routing table is constructed based on the entity types of the ontology layer and the entity instances of the data layer, as shown in Table 1:

[0023] Table 1 Specifically, relational phrases “user-application-amount” and “user-association-credit” are extracted from text decision statements, and a joint query is performed using a hierarchical semantic routing table. Data layer subgraphs that simultaneously satisfy type constraints (cosine similarity ≥ 0.8) and relational predicates (matching extracted relational phrases) are used as candidate ranges.

[0024] In one embodiment of the present invention, the decision centrality score includes the following steps: Core entities are filtered through a dynamic context gating mechanism, and the computational entities are calculated. Decision centrality score:

[0025] in, Representing entities Decision centrality score, This represents the Sigmoid function. This refers to the information about entities extracted from the attention mechanism of AI models. Average attention weight, Representation and entity The relevant classification confidence, Represents the balance coefficient. Representing entities Degree centrality in the decision semantic subgraph Psnb; Only when >Entities with dynamic thresholds are treated as core entities.

[0026] Specifically, The confidence level for the AI ​​model to pass approval is 0.92. This represents the balance coefficient, which has been optimized to a value of 0.5 after testing in financial scenarios. Calculate the scores for each entity separately, including the user entity score, credit entity score, monetary entity score, and rating entity score; The dynamic threshold is determined based on the 90th percentile of 1000 historical entity matching data in the financial field, retaining only the [specific threshold]. >Entities with dynamic thresholds are ultimately filtered to identify core entities. In this embodiment, the core entities include "User ID 20230501001", "Credit Rating A", and "500,000 yuan".

[0027] In one embodiment of the present invention, step S1, the bidirectional semantic similarity calculation model, includes the following steps: The specific bidirectional semantic similarity calculation model is as follows:

[0028] in, Represents the core entity With knowledge graph candidate entities The overall similarity between them This represents the cosine similarity calculation function based on entity embedding vectors. This represents the Jaccard similarity function used to calculate the attribute set of an entity. This represents a function for calculating the overlap of entity relationships. , and This represents the weight coefficients that are dynamically adjusted through gradient boosting of the tree model; When the overall similarity of a candidate entity is greater than or equal to a preset confidence threshold, the core entity is determined to be successfully associated with the corresponding candidate entity in the knowledge graph.

[0029] Specifically, the preset credit threshold is determined based on the entity matching accuracy requirements of the financial scenario. In this embodiment, the preset credit threshold is set to 0.75.

[0030] In one embodiment of the present invention, the dynamically adjusted weight coefficients through the gradient boosting tree model include the following steps: Extract the corresponding real-time meta-feature vector for each entity matching task; The meta-feature vector is input into an attention-enhanced gradient boosting tree model, the model comprising: Base predictor: Employs a standard gradient boosting decision tree, taking a meta-feature vector as input. This meta-feature vector includes task context features, entity pair static features, and dynamic interaction features. Through an iterative training process using gradient boosting, it outputs an initial weight vector. ; Attention-gated network: Constructs a shallow neural network structure. The input layer is a meta-feature vector, the hidden layer performs feature transformation using the ReLU activation function, and the output layer outputs an attention-gated vector. ; The final weight vector is obtained by element-wise multiplying the initial weight vector with the attention gating vector. .

[0031] Specifically, for the single-match task of core entities and knowledge graph candidate instances, a 128-dimensional real-time meta-feature vector is extracted from three dimensions: task context, entity pair static features, and dynamic interaction features. The task context features include scenario type (credit approval scenario labeled as 1 (value range 1-3, 1=credit approval, 2=post-loan monitoring, 3=credit card application)), text decision statement length (current task text length is 38 characters, normalized to 0.62: normalization formula: (actual length - minimum length) / (maximum length - minimum length)), and AI model confidence level (classification confidence 0.92, labeled as 3 (label 1 is <0.6, label 2 is 0.6-0.8, label 3 is >0.8)). The entity pair static features include entity type combination (user entity-user instance combination labeled as 1 (1=USER-USER, 2=AMOUNT-AMOUNT, 3=CREDIT-CREDIT, etc.)) and the number of core entity attributes (user ID 20). The extracted attributes for instance 230501001 are 1, and after normalization, the value is 0.2 (maximum number of attributes is 5). The completeness of candidate instance attributes (USER instance 20230501001 has 100% attribute completeness, with a value of 1.0 (completeness = actual number of attributes / required number of attributes) and entity type embedding similarity (the previously calculated cosine similarity of 0.91 is directly used as the feature value); dynamic interaction features include attribute overlap rate (the number of attribute overlaps between the core entity and the candidate instance is 1, overlap rate = 1 / 3 ≈ 0). .33), the number of matching relation types (2 types of core entity associations, 3 types of candidate instance associations, 2 matching points, value 2 / 3≈0.67), the mean attention weight (mean attention weight for core entities 0.85, normalized value 0.92), and the difference between the decision centrality score (core entity score 1.992, candidate instance association entity score 1.87, difference value 0.122, normalized value 0.15); through the above extraction logic, a 128-dimensional feature vector is finally generated; The training parameters for the base predictor (standard gradient boosting decision tree) are set using the XGBoost framework. The input is a 128-dimensional meta-feature vector, and the training data consists of meta-feature vectors from 100,000 historical entity matching tasks in the financial field and manually labeled optimal weight vectors. The core training parameters include: tree depth: 8 (determined through 5-fold cross-validation to avoid overfitting), learning rate: 0.1 (to control the contribution of each tree), number of iterations: 100 (training 100 decision trees and minimizing the loss function through gradient boosting), and loss function: mean squared error (MSE). The goal is to minimize the deviation between the model output weights and the manually labeled weights. The attention-gated network structure is designed as follows: a three-layer shallow neural network (input layer - hidden layer - output layer) is constructed, with the following specific structure: Input layer: 128 dimensions, consistent with the dimension of the meta-feature vector, activation function is Linear (no activation); Hidden layer: 64 dimensions, using ReLU activation function (activation formula: f(x) = max(0, x)), weight initialization uses He normal distribution; Output layer: 3 dimensions, corresponding to the gating values ​​of the three weight coefficients, activation function is Sigmoid (ensuring the output value is between 0 and 1, used to adjust the initial weights); Optimizer: Adam, learning rate 0.001, training loss function is cross-entropy loss; The meta-feature vector is input into the trained attention-gated network. After ReLU transformation in the hidden layer and Sigmoid activation in the output layer, the output gating vectors g1, g2, and g3 are obtained. The initial weight vector is multiplied element-wise with the attention-gating vector to obtain the final weight coefficients used for bidirectional semantic similarity calculation. The final weights are substituted into the bidirectional semantic similarity formula to calculate the comprehensive similarity.

[0032] In one embodiment of the present invention, step S2, which involves bidirectional backtracking retrieval in the knowledge graph, includes the following steps: Define an entity-relationship association path template library at the knowledge graph ontology layer, including forward and reverse association path templates based on business scenarios; Starting from the successfully associated entity, and based on the core theme of the AI ​​model output, the system matches suitable forward and reverse path templates from the path template library. If multiple suitable templates exist, the optimal path template is selected by calculating semantic similarity. According to the optimal path template, retrieve the first-level nodes and relationships directly associated with the core entity, and record the relationship type, node attributes and path confidence of the associated path; For nodes obtained through direct association, the secondary nodes and relationships are traced back along the template path. Simultaneously, the evidence contribution of the knowledge graph nodes is calculated. ,in, This indicates the credibility of the relationship between the node and the core entity. This indicates the degree of matching between node attributes and the output of the AI ​​model. and These are the weighting coefficients for relationship credibility and matching degree, respectively; When the node evidence contribution is less than a preset contribution threshold, the backtracking of the template path is stopped. The associated nodes and relationships obtained through forward and reverse backtracking are deduplicated based on the unique identifier of the knowledge graph node and the relationship triple, and then integrated into a knowledge graph association evidence set.

[0033] Specifically, the forward and reverse association path templates are shown in Table 2:

[0034] Table 2 The template library contains 12 core templates (8 forward and 4 reverse). Each template is labeled with its path length (≤3 levels), required relationship type ("application", "association", "mortgage"), and business priority (levels 1-5, with 1 being the highest). The process involves extracting the core theme from the text decision statements output by the AI ​​model: the user's credit application has been approved, with the core supporting elements being an amount of 500,000 yuan, a credit rating of A, and a risk score of 82. Based on the core theme, suitable forward and reverse templates are selected from the path template library, including: forward template matching: core themes involving "amount," "credit rating," and "repayment ability," matching T1, T2, and T3; reverse template matching: core themes involving "loan approval" and "collateral," matching T4 and T6. For the five initially matched templates, the semantic similarity between the templates and the core theme is calculated (using the TF-IDF + cosine similarity algorithm), and the two templates with the highest semantic similarity are selected as the optimal path templates. In this embodiment, the optimal template path includes forward template T2 and reverse template T6. in, and The corresponding weights are determined using the analytic hierarchy process (AHP). In this embodiment, the weights are determined accordingly. The value is 0.6. The value is 0.4; The preset contribution threshold is a minimum standard for determining the contribution of node evidence based on the scenario's business objectives (prioritizing the accuracy of approvals and reducing the risk of bad debts) and historical data statistics. In this embodiment, the value is 0.7.

[0035] In one embodiment of the present invention, step S2, which involves sorting and filtering the evidence chain according to the final scoring function, includes the following steps: The knowledge graph-related evidence set is used to generate multiple evidence chains according to a fixed path, and each evidence chain is stored in the form of a triple sequence. Construct a preliminary evaluation index set, including node semantic relevance, path reliability and evidence integrity, calculate the preliminary credibility score of each evidence chain, and select the Top-k candidate evidence chains; For each candidate evidence chain selected, a counterfactual scenario is constructed in the knowledge graph. By temporarily obscuring key nodes or relationships in the evidence chain, the case where the evidence does not exist is simulated. The confidence of the AI ​​model's decision result is recalculated on the counterfactual graph, and the causal contribution of the evidence chain is calculated by the change in confidence. A dynamic adversarial example generator is constructed to generate semantically preserved perturbation evidence chains for candidate evidence chains. The robustness score of the evidence chain is calculated by comparing the credibility difference between the original evidence chain and the perturbation evidence chain. A final evaluation function is constructed that integrates the initial credibility score, causal contribution score, and robustness score, and evidence chains with high causal contribution and strong robustness are preferentially selected as the optimal explanation.

[0036] Specifically, the knowledge graph-related evidence set is used to generate multiple evidence chains according to a fixed path, denoted as... ,in, The core entity indicating a successful association. Indicates associated nodes, This represents the relationship between nodes; where the fixed path is core entity → associated node 1 → associated node 2 → … → support node. Specifically, the semantic relevance of nodes is assessed by using a pre-trained language model to obtain the embedding vectors of the node description text and the core topic text, respectively. The semantic relevance of a single node is calculated using cosine similarity, and the average of the relevance of all nodes in the evidence chain is taken as the score for this dimension. Evaluation is based on the prior confidence of relations in the knowledge graph and the path structure. This evaluation is performed by combining the product of the reliability of all relations in the path with a penalty term for the path length to obtain the path reliability. Specifically: ,in, Representing relations confidence level Indicates the path length. The length penalty coefficient is a configurable parameter greater than 0. Its physical meaning is the rate of credibility decay caused by a unit increase in path length. It is determined through a grid search on the validation set. Its optimal value aims to maximize the average final credibility of the initially selected Top-K evidence chains in subsequent refined evaluations. In this embodiment, The value is 0.5; Evidence completeness is used to evaluate the degree to which the evidence chain covers the key attributes of the core entity. It is scored by the ratio of the number of core theme-related attributes involved in the evidence chain to the preset total number of core attributes that should theoretically be present to arrive at the conclusion; Finally, the preliminary credibility score is the weighted sum of the scores of the above three indicators, where the weight coefficients corresponding to the scores of the three indicators are obtained by regression analysis training on historical labeled data. Among them, calculating the chain of evidence The causal contribution is as follows: ,in, This indicates the chain of evidence. Causal contribution This represents the confidence level of the original decision. It represents the confidence level of the original decision in a counterfactual scenario; the causal contribution directly quantifies the chain of evidence. Specifically, a dynamically updated adversarial example generator is constructed. For each candidate evidence chain EC, a semantically preserved perturbation evidence chain is generated by replacing synonyms, inserting irrelevant nodes, or slightly distorting the semantics of the relationships. The robustness score of the evidence chain is calculated by reassessing its ability to support the original conclusion on the perturbed evidence chain. ,in, The credibility evaluation function represents the chain of evidence; a high robustness score indicates that the conclusion of the chain of evidence does not depend on fragile semantic matching. Construct the final credibility fusion function Specifically:

[0037] in, This indicates the initial credibility score. , and This represents the configurable fusion weights, and > , In this embodiment, The value is 0.2. The value is 0.5. The value is 0.3; The evidence chain with high causal contribution and strong robustness is given priority as the optimal explanation. When there is a conflict in causal contribution (multiple evidence chains have high causal contribution but point to different conclusions), the evidence tracing mechanism is activated to trace back to the original data source for consistency verification.

[0038] In one embodiment of the present invention, step S3, in which the multi-level conflict detection engine calculates the comprehensive conflict index, includes the following steps: A spatiotemporal consistency conflict detection model is constructed, and the temporal conflict degree, spatial conflict degree, and axiom violation score are calculated respectively. Design an adaptive weight allocation mechanism to dynamically adjust the weight coefficients of various conflict types based on their severity. Specifically: ,in, , and This indicates the final weight of temporal conflict degree, spatial conflict degree, and axiom violation score in the calculation of the comprehensive conflict index. , and The scores represent the severity of time conflicts, spatial conflicts, and logical conflicts. The comprehensive conflict index is obtained by adding the final weights corresponding to various dimensions with the time conflict degree, spatial conflict degree and axiom violation score.

[0039] Specifically, a spatiotemporal consistency conflict detection model is constructed, and multi-dimensional conflict values ​​are calculated, including the calculation of temporal conflict degree. ,in, The timestamp representing newly added information. The timestamps represent the existing data in the knowledge graph. This represents the maximum time difference preset for the scenario; for credit approval scenarios, it is set to 6 months; calculating spatial conflict degree. ,in, This represents the straight-line distance between the location associated with the new information and the location associated with the existing data. This represents the maximum spatial distance preset for the scenario (500km for credit approval scenarios; exceeding this is considered a strong spatial difference); axiom violation fractions. Based on a predefined OWL axiom base and business rule base, it is determined whether new information violates logical constraints, according to the following rules: If a core axiom / rule is violated (e.g., a conflict between no overdue records in credit reporting and newly added overdue records, or an assessment agency's qualification suddenly changing from Grade A to no qualification), then... =1; If a minor axiom / rule is violated (e.g., monthly income fluctuation ≤30%, exceeding this is considered a minor conflict), then =0.5; if there is no violation, then =0; Among them, examples from the OWL axiom library and business rule library: the overdue number attribute of a CREDIT instance cannot change from 0 to ≥1, the valuation fluctuation range of an ASSET instance is ≤20%, and the monthly income fluctuation range of a USER instance is ≤30% (without special proof of occupational change); Regarding the axiom violation score calculation result for one type of new information, specifically: new data shows "1 overdue payment", old data shows "no overdue payments", although there is an official source from the central bank, it violates the core axiom "the overdue number cannot change from 0 to ≥1" (further verification of the authenticity of the overdue payment is required) → =1; Among them, an adaptive weight allocation mechanism is designed to score the severity of the three types of conflicts from 1 to 5 levels (level 5 being the most severe) based on the degree of risk impact in the financial credit scenario: Time Conflict Severity Rating: 2-month time difference, moderate impact on loan approval (income verification timeliness is acceptable) → =2; Spatial conflict severity score: Maximum spatial distance 25km, no cross-city differences, minor impact → =1; Logical Conflict Severity Score: Credit delinquency conflicts are a core risk (affecting user credit rating), while income fluctuations are a secondary risk → Take the highest value between the two types of logical conflicts → =5; The preset conflict threshold is set to 0.6 in this embodiment, based on the risk tolerance (medium risk) of the financial credit scenario.

[0040] Please see Figure 2 As shown, this invention is an information processing system based on knowledge graphs, comprising the following modules: Core Entity Extraction and Knowledge Graph Association Matching Module: Obtain the output results of the AI ​​model, extract the core entities and entity associations from the output results, wherein the core entities are entities obtained by filtering through decision centrality scores, and are determined to be core entities when the score is greater than a dynamic threshold; Based on the hierarchical structure of the knowledge graph, the entity type is used for pre-screening. The selected candidate entities in the knowledge graph are matched with the core entity using a bidirectional semantic similarity calculation model. If the matching result meets the preset confidence threshold, the corresponding candidate entity in the core entity relationship knowledge graph is determined to be successfully associated. Knowledge Graph Evidence Chain Backtracking Retrieval and Quantitative Evaluation Display Module: After successful entity association, starting from the successfully associated entity, bidirectional backtracking retrieval is performed in the knowledge graph to extract the associated nodes and relationships that support the output results of the AI ​​model. The generated data chain is quantitatively evaluated, the evidence chain is sorted and filtered according to the final scoring function, and the optimal evidence chain is visualized in a graph form. Multi-source information conflict detection and warning generation module: When new multi-source information is added to the database and linked to the entity in the knowledge graph, the predefined OWL axiom library and business rule library are called. Through the multi-level conflict detection engine, the comprehensive conflict index is calculated. If the comprehensive conflict index is greater than or equal to the preset conflict threshold, a conflict warning is immediately generated and issued.

[0041] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A knowledge graph-based information processing method, characterized in that, Includes the following steps: S1: Obtain the output results of the AI ​​model, extract the core entities and entity relationships in the output results, wherein the core entities are entities obtained by filtering through decision centrality scores, and are determined to be core entities when the score is greater than a dynamic threshold; Based on the hierarchical structure of the knowledge graph, the entity type is used for pre-screening. The selected candidate entities in the knowledge graph are matched with the core entity using a bidirectional semantic similarity calculation model. If the matching result meets the preset confidence threshold, the corresponding candidate entity in the core entity relationship knowledge graph is determined to be successfully associated. S2: After successful entity association, starting from the successfully associated entity, perform bidirectional backtracking retrieval in the knowledge graph to extract the associated nodes and relationships that support the output results of the AI ​​model, quantitatively evaluate the generated data chain, sort and filter the evidence chain according to the final scoring function, and visualize the optimal evidence chain in a graph form. S3: When new multi-source information is added to the database and linked to the entity in the knowledge graph, the predefined OWL axiom library and business rule library are called. Through the multi-level contradiction detection engine, the comprehensive conflict index is calculated. If the comprehensive conflict index is greater than or equal to the preset conflict threshold, a contradiction warning is immediately generated and issued.

2. The information processing method based on knowledge graphs according to claim 1, characterized in that, Step S1 includes the following steps: The AI ​​model output includes text decision statements, classification confidence vectors, and attention weights. It employs a domain-adaptive NER model to parse the text, and its loss function... ,in, This represents the total loss value. This represents the loss of the standard named entity recognition task. Indicates hyperparameters, This indicates the auxiliary loss based on decision-making keywords; Simultaneously, a weighted decision semantic subgraph is constructed using confidence and attention weights. The decision semantic subgraph The nodes are the extracted entities, and the edge weights are the average of the attention weights between entities; A hierarchical semantic routing table is constructed based on the ontology layer and data layer of the knowledge graph. Matching is performed by calculating the cosine similarity between entity type embedding and knowledge graph type node embedding, and joint query is performed by combining relational phrases extracted from the decision text. The data layer subgraph that simultaneously satisfies type constraints and relational predicates is used as the candidate range after filtering.

3. The information processing method based on knowledge graphs according to claim 2, characterized in that, The decision centrality score includes the following steps: Core entities are filtered through a dynamic context gating mechanism, and the computational entities are calculated. Decision centrality score: in, Representing entities Decision centrality score, This represents the Sigmoid function. This refers to the information about entities extracted from the attention mechanism of AI models. Average attention weight, Representation and entity The relevant classification confidence, Represents the balance coefficient. Representing entities In the decision semantic subgraph Degree centrality in; Only when >Entities with dynamic thresholds are treated as core entities.

4. The information processing method based on knowledge graphs according to claim 1, characterized in that, In step S1, the bidirectional semantic similarity calculation model includes the following steps: The specific bidirectional semantic similarity calculation model is as follows: in, Represents the core entity With knowledge graph candidate entities The overall similarity between them This represents the cosine similarity calculation function based on entity embedding vectors. This represents the Jaccard similarity function used to calculate the attribute set of an entity. This represents a function for calculating the overlap of entity relationships. , and This represents the weight coefficients that are dynamically adjusted through gradient boosting of the tree model; When the overall similarity of a candidate entity is greater than or equal to a preset confidence threshold, the core entity is determined to be successfully associated with the corresponding candidate entity in the knowledge graph.

5. The information processing method based on knowledge graphs according to claim 4, characterized in that, The weight coefficients dynamically adjusted through the gradient boosting tree model include the following steps: Extract the corresponding real-time meta-feature vector for each entity matching task; The meta-feature vector is input into an attention-enhanced gradient boosting tree model, the model comprising: Base predictor: Employs a standard gradient boosting decision tree, taking a meta-feature vector as input. This meta-feature vector includes task context features, entity pair static features, and dynamic interaction features. Through an iterative training process using gradient boosting, it outputs an initial weight vector. ; Attention-gated network: Constructs a shallow neural network structure. The input layer is a meta-feature vector, the hidden layer performs feature transformation using the ReLU activation function, and the output layer outputs an attention-gated vector. ; The final weight vector is obtained by element-wise multiplying the initial weight vector with the attention gating vector. .

6. The information processing method based on knowledge graphs according to claim 1, characterized in that, In step S2, bidirectional backtracking retrieval is performed in the knowledge graph, including the following steps: Define an entity-relationship association path template library at the knowledge graph ontology layer, including forward and reverse association path templates based on business scenarios; Starting from the successfully associated entity, and based on the core theme of the AI ​​model output, the system matches suitable forward and reverse path templates from the path template library. If multiple suitable templates exist, the optimal path template is selected by calculating semantic similarity. According to the optimal path template, retrieve the first-level nodes and relationships directly associated with the core entity, and record the relationship type, node attributes and path confidence of the associated path; For nodes obtained through direct association, the secondary nodes and relationships are traced back along the template path. Simultaneously, the evidence contribution of the knowledge graph nodes is calculated. ,in, This indicates the credibility of the relationship between the node and the core entity. This indicates the degree of matching between node attributes and the output of the AI ​​model. and These are the weighting coefficients for relationship credibility and matching degree, respectively; When the node evidence contribution is less than a preset contribution threshold, the backtracking of the template path is stopped. The associated nodes and relationships obtained through forward and reverse backtracking are deduplicated based on the unique identifier of the knowledge graph node and the relationship triple, and then integrated into a knowledge graph association evidence set.

7. The information processing method based on knowledge graphs according to claim 1, characterized in that, In step S2, the evidence chain is sorted and filtered according to the final scoring function, including the following steps: The knowledge graph-related evidence set is used to generate multiple evidence chains according to a fixed path, and each evidence chain is stored in the form of a triple sequence. Construct a preliminary evaluation index set, including node semantic relevance, path reliability and evidence integrity, calculate the preliminary credibility score of each evidence chain, and select the Top-k candidate evidence chains; For each candidate evidence chain selected, a counterfactual scenario is constructed in the knowledge graph. By temporarily obscuring key nodes or relationships in the evidence chain, the case where the evidence does not exist is simulated. The confidence of the AI ​​model's decision result is recalculated on the counterfactual graph, and the causal contribution of the evidence chain is calculated by the change in confidence. A dynamic adversarial example generator is constructed to generate semantically preserved perturbation evidence chains for candidate evidence chains. The robustness score of the evidence chain is calculated by comparing the credibility difference between the original evidence chain and the perturbation evidence chain. A final evaluation function is constructed that integrates the initial credibility score, causal contribution score, and robustness score, and evidence chains with high causal contribution and strong robustness are preferentially selected as the optimal explanation.

8. The information processing method based on knowledge graphs according to claim 1, characterized in that, In step S3, the multi-level conflict detection engine calculates the comprehensive conflict index, including the following steps: A spatiotemporal consistency conflict detection model is constructed, and the temporal conflict degree, spatial conflict degree, and axiom violation score are calculated respectively. Design an adaptive weight allocation mechanism to dynamically adjust the weight coefficients of various conflict types based on their severity. Specifically: ,in, , and This indicates the final weight of temporal conflict degree, spatial conflict degree, and axiom violation score in the calculation of the comprehensive conflict index. , and The scores represent the severity of time conflicts, spatial conflicts, and logical conflicts. The comprehensive conflict index is obtained by adding the final weights corresponding to various dimensions with the time conflict degree, spatial conflict degree and axiom violation score.

9. A knowledge graph-based information processing system, used to implement the knowledge graph-based information processing method according to any one of claims 1-8, characterized in that, Includes the following modules: Core Entity Extraction and Knowledge Graph Association Matching Module: Obtain the output results of the AI ​​model, extract the core entities and entity associations from the output results, wherein the core entities are entities obtained by filtering through decision centrality scores, and are determined to be core entities when the score is greater than a dynamic threshold; Based on the hierarchical structure of the knowledge graph, the entity type is used for pre-screening. The selected candidate entities in the knowledge graph are matched with the core entity using a bidirectional semantic similarity calculation model. If the matching result meets the preset confidence threshold, the corresponding candidate entity in the core entity relationship knowledge graph is determined to be successfully associated. Knowledge Graph Evidence Chain Backtracking Retrieval and Quantitative Evaluation Display Module: After successful entity association, starting from the successfully associated entity, bidirectional backtracking retrieval is performed in the knowledge graph to extract the associated nodes and relationships that support the output results of the AI ​​model. The generated data chain is quantitatively evaluated, the evidence chain is sorted and filtered according to the final scoring function, and the optimal evidence chain is visualized in a graph form. Multi-source information conflict detection and warning generation module: When new multi-source information is added to the database and linked to the entity in the knowledge graph, the predefined OWL axiom library and business rule library are called. Through the multi-level conflict detection engine, the comprehensive conflict index is calculated. If the comprehensive conflict index is greater than or equal to the preset conflict threshold, a conflict warning is immediately generated and issued.

Images