Knowledge graph error detection method based on relation symmetry
By defining relationally symmetric entities and triples, generating complementary views and calculating similarity scores, the accuracy problem of error detection in knowledge graphs is solved, and efficient automated cleaning and fine-grained detection are achieved, which is suitable for large-scale knowledge graphs.
Patent Information
- Application Number
- CN202510951730.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-10
- Publication Date
- 2025-09-12
AI Technical Summary
Existing technologies have the problem of insufficient detection accuracy when detecting erroneous triples in knowledge graphs, especially in the case of diverse data sources and omissions in manual annotation. Traditional graph comparative learning methods may change the error distribution or introduce noise.
Define relation-symmetric entities and triplets, generate relation-symmetric triplet views and head-tail entity connection triplet views, capture global and local structures through graph attention networks, calculate similarity scores to distinguish normal and erroneous triplets, and design a contrastive loss function based on the information maximization criterion to optimize model parameters.
It improves the recognition accuracy of erroneous triples in knowledge graphs and is suitable for large-scale automated cleaning without relying on labeled data or manual rules. It is suitable for scenarios where text information is missing or the labeling cost is high, reduces noise introduction, and has a low parameter count.
Smart Images
Figure CN120633801A_ABST
Abstract
Description
Technical Field
[0001] This field relates to the fields of artificial intelligence and data mining technology, and specifically to a knowledge graph error detection method based on relational symmetry. Background Art
[0002] A knowledge graph is a knowledge base that represents entities and their relationships using a graph structure. It is widely used in fields such as natural language processing, recommendation systems, and question-answering systems. However, during the construction of a knowledge graph, due to the diversity of data sources or omissions in manual annotation, incorrect triples (entity-relationship-entity) are inevitably introduced. These triples are semantically inconsistent or inconsistent with real-world facts. The accumulation of these errors significantly degrades the quality of the knowledge graph, thereby affecting the performance of downstream tasks. Therefore, how to efficiently and accurately detect incorrect triples in knowledge graphs has become an urgent research topic.
[0003] In the existing technology, graph contrastive learning methods are widely used for the analysis and processing of graph data. Their core lies in generating different views through data augmentation techniques and learning consistent representations. However, traditional graph contrastive learning frameworks mainly rely on graph-level enhancement methods, such as node dropping, edge perturbation, subgraph sampling, or matrix diffusion. These methods have certain applicability on general graph data, but have significant limitations in the task of error detection in knowledge graphs. Since knowledge graphs are composed of multiple groups of triplets, the essence of the error detection task is to identify abnormal triplets. However, graph-level enhancement techniques may change the distribution of errors and even introduce new noise, resulting in a decrease in the accuracy of detection results. Therefore, it is necessary to propose a new method that fully utilizes the structural characteristics of knowledge graphs to improve the ability to identify erroneous triplets.
[0004] Application Contents
[0005] The purpose of this invention is to provide a knowledge graph error detection method based on relational symmetry. The specific technical solution is as follows:
[0006] A knowledge graph error detection method based on relational symmetry, comprising: S1, defining relational symmetric entities and relational symmetric triples, which are used to describe the semantic similarity or consistency of entities and triples in a knowledge graph under a specific relational pattern; S2, using triples in S1 as nodes to generate relational symmetric triple views and head-tail entity connection triple views, which are used to describe triple characteristics from different dimensions to complement each other; S3, calculating the relational symmetric triple views in S2. Figure 3 Tuple representation X i and the head and tail entity connection triplet view Figure 3 Tuple representation Z i The similarity score between them is used to distinguish normal triplets from erroneous triplets.
[0007] Based on S3, a contrastive loss function based on the information maximization criterion is designed to optimize model parameters by minimizing the contrastive loss to enhance the distinguishing ability of triple representation. The contrastive loss function is specifically:
[0008]
[0009] Among them, τ is the temperature parameter, which controls the scaling of similarity.
[0010] The relational symmetric entities in S1 include 1-hop relational symmetric entities and multi-hop relational symmetric entities, where 1-hop relational symmetric entities are entity pairs connected to the same pivot entity through the same relation, and multi-hop relational symmetric entities are entity pairs connected to the same pivot entity through the same multi-hop relation chain.
[0011] The relation-symmetric triples in S1 are extended based on relation-symmetric entities, including one-hop relation-symmetric triples and multi-hop relation-symmetric triples. One-hop relation-symmetric triples are triples that share head and tail entities and have the same relationship, and multi-hop relation-symmetric triples are triples whose head and tail entities are both relation-symmetric entities and have the same relationship, which are used to reflect the symmetric structure between triplets.
[0012] When generating relation-symmetric triple views in S2, the triple representation is initialized through the embedding model, and the graph attention network is used to aggregate the symmetric neighbor information to capture the global symmetric structure.
[0013] When generating the head-tail entity connection triplet view in S2, the local neighbor information is aggregated through the graph attention network to capture the local connection pattern of the triplet.
[0014] Computing relational symmetric triples in S3 Figure 3 Tuple representation X i and the head and tail entity connection triplet view Figure 3 Tuple representation Z i When the similarity score between the two is calculated, the similarity C(h,r,t)=
[0015] sim(X i ,Z i ) is used as the confidence score of the triple, where sim(.,.) can use functions such as cosine similarity; the confidence score is used to distinguish normal triples from erroneous triples: a high score indicates strong consistency and is a normal triple; a low score indicates weak consistency and is an erroneous triple.
[0016] The beneficial effect of this application is that it proposes the definition of "relational symmetric entity" and "relational symmetric triple", captures the implicit symmetric structure in the knowledge graph through one-hop and multi-hop symmetric patterns, and can more accurately identify errors caused by inconsistent symmetric relationships (such as redundant or contradictory triples). Symmetry is used to generate complementary views (relational symmetric view and entity connection view), and the modeling of global semantic consistency is enhanced through contrastive learning. Based on the multi-view contrastive learning framework, the correctness is directly distinguished by the triple confidence score, without relying on labeled data or manual rules, and is suitable for the automated cleaning of large-scale knowledge graphs. In response to the problem that graph-level enhancement technology in traditional graph contrastive learning may destroy the error distribution, views are generated through symmetric structures to reduce noise introduction, which is more suitable for fine-grained error detection of knowledge graphs. No text description of entities / relationships or other external semantic resources are required, and semantic consistency can be modeled only through relational symmetry. It is suitable for scenarios where text information is missing or the annotation cost is high (such as industrial knowledge graphs). Based on a lightweight graph attention network (GAT) and a contrastive learning framework. Only the triple symmetric structure needs to be processed, and there is no text encoding module. The number of parameters is lower, and the model parameters are concentrated on symmetric relationship modeling, without large pre-trained language models.
[0017] Figures in the specification
[0018] Figure 1 This is a schematic diagram of the application process;
[0019] Figure 2 A schematic diagram of defining relational symmetric entities and relational symmetric triples in this application;
[0020] Figure 3 A schematic diagram of generating a relational symmetric triple view and a head-tail entity connection triple view in this application. Specific embodiments
[0021] To make the objectives, technical solutions, and advantages of the present invention more clearly understood, the present invention will be further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings. It should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present invention. In addition, in the following description, descriptions of well-known structures and technologies are omitted to avoid unnecessary confusion of the concepts of the present invention.
[0022] like Figure 1-3 As shown:
[0023] A knowledge graph error detection method based on relational symmetry, comprising: S1, defining relational symmetric entities and relational symmetric triples, which are used to describe the semantic similarity or consistency of entities and triples in a knowledge graph under a specific relational pattern; S2, using triples in S1 as nodes to generate relational symmetric triple views and head-tail entity connection triple views, which are used to describe triple characteristics from different dimensions to complement each other; S3, calculating the relational symmetric triple views in S2. Figure 3 Tuple representation X i and the head and tail entity connection triplet view Figure 3 Tuple representation Z i Similarity scores between entities are used to distinguish between valid and erroneous triples. The definitions of "relationally symmetric entities" and "relationally symmetric triples" are proposed. Using one-hop and multi-hop symmetric patterns to capture the implicit symmetric structure in knowledge graphs, this approach can more accurately identify errors caused by inconsistent symmetric relations (such as redundant or contradictory triples). Symmetry is leveraged to generate complementary views (relationally symmetric and entity-connected), enhancing the modeling of global semantic consistency through contrastive learning. Based on a multi-view contrastive learning framework, correctness is directly distinguished from error based on triple confidence scores, without relying on annotated data or manual rules. This approach is suitable for automated cleaning of large-scale knowledge graphs. To address the problem that graph-level augmentation techniques in traditional graph contrastive learning can disrupt the distribution of errors, views are generated based on symmetric structures, reducing the introduction of noise and making them more suitable for fine-grained error detection in knowledge graphs. Semantic consistency can be modeled solely through relational symmetry, eliminating the need for textual descriptions of entities / relationships or other external semantic resources. This approach is suitable for scenarios where textual information is scarce or annotation costs are high, such as industrial knowledge graphs. It is based on a lightweight graph attention network (GAT) and a contrastive learning framework. It only processes triplet symmetric structures and does not require a text encoding module. It has a lower parameter count, with model parameters focused on modeling symmetric relationships and no large pre-trained language model is required.
[0024] Based on S3, a contrastive loss function based on the information maximization criterion is designed to optimize model parameters by minimizing the contrastive loss to enhance the distinguishing ability of triple representation. The contrastive loss function is specifically:
[0025]
[0026] Among them, τ is the temperature parameter, which controls the scaling of similarity.
[0027] The relational symmetric entities in S1 include 1-hop relational symmetric entities and multi-hop relational symmetric entities, where 1-hop relational symmetric entities are entity pairs connected to the same pivot entity through the same relation, and multi-hop relational symmetric entities are entity pairs connected to the same pivot entity through the same multi-hop relation chain.
[0028] The relation-symmetric triples in S1 are extended based on relation-symmetric entities, including one-hop relation-symmetric triples and multi-hop relation-symmetric triples. One-hop relation-symmetric triples are triples that share head and tail entities and have the same relationship, and multi-hop relation-symmetric triples are triples whose head and tail entities are both relation-symmetric entities and have the same relationship, which are used to reflect the symmetric structure between triplets.
[0029] When generating relation-symmetric triple views in S2, the triple representation is initialized through the embedding model, and the graph attention network is used to aggregate the symmetric neighbor information to capture the global symmetric structure.
[0030] When generating the head-tail entity connection triplet view in S2, the local neighbor information is aggregated through the graph attention network to capture the local connection pattern of the triplet.
[0031] Computing relational symmetric triples in S3 Figure 3 Tuple representation X i and the head and tail entity connection triplet view Figure 3 Tuple representation Z i When the similarity score between the two is calculated, the similarity C(h,r,t)=
[0032] sim(X i ,Z i ) is used as the confidence score of the triple, where sim(.,.) can use functions such as cosine similarity; the confidence score is used to distinguish normal triples from erroneous triples: a high score indicates strong consistency and is a normal triple; a low score indicates weak consistency and is an erroneous triple.
Claims
1. A knowledge graph error detection method based on relational symmetry, characterized in that: include: S1. Define relational symmetric entities and relational symmetric triples to describe the semantic similarity or consistency of entities and triples in a knowledge graph under a specific relational model; S2. Using the triples in S1 as nodes, generate a relation-symmetric triple view and a head-tail entity connection triple view, for describing triple characteristics from different dimensions; S3, calculating the relational symmetric triple view triple representation X in S2 i The triplet view triplet representation Z is connected with the head and tail entities i The similarity score between them is used to distinguish normal triplets from erroneous triplets.
2. The knowledge graph error detection method based on relational symmetry according to claim 1, characterized in that: Based on S3, a contrast loss function based on the information maximization criterion is designed to optimize the model parameters by minimizing the contrast loss to enhance the distinguishing ability of triple representation; the contrast loss function is specifically: Among them, τ is the temperature parameter, which controls the scaling of similarity.
3. The knowledge graph error detection method based on relational symmetry according to claim 1, characterized in that: The relational symmetric entities in S1 include 1-hop relational symmetric entities and multi-hop relational symmetric entities, wherein the 1-hop relational symmetric entity is an entity pair connected to the same pivot entity through the same relationship, and the multi-hop relational symmetric entity is an entity pair connected to the same pivot entity through the same multi-hop relationship chain.
4. The knowledge graph error detection method based on relational symmetry according to claim 2, characterized in that: The relation-symmetric triples described in S1 are extended based on the relation-symmetric entities, including 1-hop relation-symmetric triples and multi-hop relation-symmetric triples, wherein the 1-hop relation-symmetric triples are triples that share head and tail entities and have the same relationship, and the multi-hop relation-symmetric triples are triples whose head and tail entities are both relation-symmetric entities and have the same relationship, which are used to reflect the symmetrical structure between triplets.
5. The knowledge graph error detection method based on relational symmetry according to claim 1, characterized in that: When generating the relational symmetric triple view in S2, the triple representation is initialized by the embedding model, and the graph attention network is used to aggregate the symmetric neighbor information to capture the global symmetric structure.
6. The knowledge graph error detection method based on relational symmetry according to claim 1, characterized in that: When generating the head-tail entity connection triplet view in S2, local neighbor information is aggregated through the graph attention network to capture the local connection pattern of the triplet.
7. The knowledge graph error detection method based on relational symmetry according to claim 1, characterized in that: The calculation relation of the symmetric triple view triple representation X in S3 i And the head and tail entity connection triples view triples represent Z i When the similarity score between i ,Z i ) as the confidence score of the triple, where sim(.,.) can use functions such as cosine similarity; The confidence score is used to distinguish normal triplets from incorrect triplets: a high score indicates strong consistency and is a normal triplet; a low score indicates weak consistency and is an incorrect triplet.