A knowledge inconsistency detection method in knowledge graph based on entity semantic strength

Abstract

The present invention discloses a method for detecting knowledge inconsistency in a knowledge graph based on entity semantic strength, comprising: firstly performing data preprocessing on the knowledge graph and making a formal definition of knowledge inconsistency, mining potential entity association strength information of knowledge triples using three strategies of entity pair association strength, triple semantic similarity distance, and degree of combination of relationship and entity, regarding the detection of knowledge inconsistency as a classification problem, identifying the category to which the knowledge belongs, and realizing comprehensive detection of knowledge inconsistency based on the category; the present invention can effectively detect various inconsistencies of the knowledge graph, plays an important role in improving the knowledge quality of the knowledge graph, and further promotes the implementation of downstream tasks of the knowledge graph.

Description

Technical Field

[0001] The present invention relates to a method for detecting knowledge inconsistency in a knowledge graph, and in particular to a method for detecting knowledge inconsistency in a knowledge graph based on entity semantic strength. Background Art

[0002] Knowledge graph is a relatively general formal description framework of semantic knowledge. It uses nodes to represent semantic symbols and edges to represent semantic relationships between symbols. It is the most important way to represent knowledge in the era of big data. The emergence of knowledge graph has greatly promoted the process of machine understanding of language, making it possible for machines to have cognitive intelligence and "think like humans". It has been widely used in artificial intelligence fields such as intelligent information retrieval, question-answering systems, and precise recommendations.

[0003] However, although the era of big data has promoted the development of knowledge graphs, the data quality issues hidden behind the explosive growth of data scale have also laid hidden dangers for the application of knowledge graphs. On the one hand, due to the data-driven nature of knowledge graphs, the decline in data quality will inevitably lead to a threat to knowledge quality; on the other hand, errors in natural language processing and conflicting data sources cause a lot of noise in the construction of knowledge graphs. Generally speaking, inconsistency detection of large-scale knowledge graphs faces three challenges: due to the discreteness of knowledge, it is difficult to detect all inconsistent knowledge using ontology reasoning technology; although the entities and relationships in the knowledge graph are rich in semantic information, they are difficult to capture; there is a lack of golden rules for judging and detecting knowledge inconsistencies.

[0004] Due to the large-scale application of knowledge graphs in the field of artificial intelligence, the quality assurance of knowledge graphs has become a hot research field. In order to solve the difficulties of inconsistency detection in the above knowledge graphs, researchers have explored the semantic information of entities in knowledge graphs and used the semantic association features between knowledge to detect various inconsistent knowledge in knowledge graphs. Knowledge graphs can essentially be regarded as a semantic network, using nodes to represent entities in knowledge graphs and edges to represent the relationship between entities. There may be multiple paths between entity pairs, and the same path may exist between different entity pairs. This complex graph model relationship represents rich semantic information and plays an important role in inconsistency detection. Representation learning embeds knowledge into a low-dimensional dense and continuous vector space, measures its semantic similarity by the distance of different vectors, and uses representation learning methods combined with similarity measurement methods to measure the semantic similarity between entities. It is another important means of inconsistency detection.

[0005] Considering the effectiveness and sufficiency of various inconsistency detection in knowledge graphs, the present invention designs an effective and feasible inconsistency detection scheme, which regards inconsistency detection as a classification problem and can identify all knowledge inconsistencies and their types. This knowledge inconsistency detection method in knowledge graphs based on entity semantic strength not only has a good improvement in accuracy, but also has high detection efficiency and broad application prospects, and can be applied to improve the quality of knowledge graphs. Summary of the invention

[0006] In view of this, the purpose of the present invention is to provide a method for detecting knowledge inconsistency in knowledge graphs based on entity semantic strength, so as to solve the inconsistency problems occurring in knowledge graphs, improve the knowledge quality of knowledge graphs, and enable knowledge graphs to better serve downstream tasks such as question answering, recommendation, and decision-making. The present invention targets knowledge inconsistencies that may occur in knowledge graphs and can comprehensively identify all knowledge inconsistencies in an accurate and efficient manner.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] A method for detecting knowledge inconsistency in a knowledge graph based on entity semantic strength, the method comprising:

[0009] Step S1, preprocessing the knowledge graph dataset, first adding type triples as a supplement to the knowledge graph dataset according to the type information of the entities in the knowledge graph, and then generating negative samples according to the negative sample generation principle;

[0010] Step S2, abstracting the knowledge graph in the data set preprocessed in step S1 into a representation form of a graph model, extracting entity subgraphs according to the nodes that the head entity can reach, and using the resource information flow on the entity subgraphs to represent the association strength between entity pairs;

[0011] Step S3, obtaining the representation of entities and relations in the triplet in the low-dimensional vector space, and calculating the cosine similarity to represent the semantic similarity value from the head entity to the tail entity;

[0012] Step S4: Count the occurrence frequency of the tail entity corresponding to the relationship in the triple in the knowledge graph, and measure the path credibility of the triple corresponding to the relationship according to the degree of combination between the relationship and the tail entity;

[0013] Step S5: construct feature engineering based on the association strength of the triples determined in step S2, the semantic similarity value determined in step S3, and the degree of combination determined in step S4, and use multiple classifiers to train the inconsistency recognition model to complete the recognition of all possible inconsistency types.

[0014] Furthermore, the step S1 specifically includes:

[0015] Step S101: Convert the entity type information in the knowledge graph into a triple representation to ensure the completeness of the data in the knowledge graph;

[0016] Step S102: according to the triple representation of knowledge in the knowledge graph and the problems that occur during the knowledge graph construction process, the problems that occur during the knowledge graph construction process specifically include: vocabulary similarity problems that occur during natural language processing, redundancy problems that occur in entity extraction, polysemy problems that occur in knowledge fusion, and data conflicts caused by multiple data sources, the knowledge inconsistencies are divided, wherein for the target triple (h, r, t), the corresponding inconsistencies specifically include:

[0017] Relational inconsistency:

[0018]

[0019] Entity Inconsistency:

[0020]

[0021] Type inconsistency:

[0022]

[0023] Among them, G represents the knowledge contained in the knowledge graph, E and R represent the entity set and relationship set in the knowledge graph respectively, h, t∈E, r∈R, T represents the type set of entities, RI, EI, and TI are the abbreviations of relationship inconsistency, entity inconsistency, and type inconsistency respectively.

[0024] Step S103: Generate inconsistent negative samples in a 1:1 manner according to the three types of inconsistencies and features in step S102, wherein the features are specifically: features represented when similar entities and their corresponding relationships are incorrectly mapped during the knowledge graph construction process; the proportions of negative samples corresponding to entity inconsistency, relationship inconsistency, and type inconsistency are: 25%, 12.5%, and 12.5%.

[0025] Furthermore, the step S2 specifically includes:

[0026] Step S201, traverse each triple in the knowledge graph, and construct a directed graph representation of the knowledge graph with entities in the triple as nodes and relationships as edges;

[0027] Step S202: For each entity in the knowledge graph, use depth-first search to find the entity set that the entity can be connected to, and record the entity search path to construct an entity subgraph;

[0028] Step S203: For each triple in the knowledge graph, the information distribution value flowing from the head entity to the tail entity is calculated according to the resource flow in the entity subgraph where the head entity is located, which represents the association strength between the entity pairs in the triple. The information distribution value calculation expression is as follows:

[0029]

[0030] In the formula, E i-1 (·,e) indicates the path r i The set of direct predecessor nodes to reach e, E i (e′,·) represents the i The path r passed by the predecessor i The set of direct successor nodes, W e′e represents the weight of the edge e′→e, and the weight depends on the number of edges connecting e′ and e; θ is a hyperparameter set to avoid the closed loop problem and is set to 0.15.

[0031] Initially, all entities in the subgraph are set to the same semantic value. The semantic value that flows to a certain entity through multiple paths is the information assignment value of the entity pair with the entity as the node and the central node of the subgraph as the head node.

[0032] Furthermore, in step S202, the maximum value of the relationship chain in the entity subgraph is 4, and a depth-first search traversal is used to sequentially find tail nodes to which the head node can be connected within a depth range of 4 and add them to the entity subgraph.

[0033] Further, in the step S203, the calculating of the information allocation value flowing from the head entity to the tail entity according to the resource flow in the entity subgraph where the head entity is located specifically includes:

[0034] Based on the entity subgraph where the head entity is located, determine whether the tail entity is in the subgraph;

[0035] If it is not present, the missing value is processed and the entity of the triple is treated as zero value for the intensity feature;

[0036] Otherwise, the final entity information value of the triple is calculated according to the resource flow from the head entity to the tail entity, wherein the final entity information value represents the association strength of the entity pair.

[0037] Furthermore, the step S3 specifically includes:

[0038] Step S301, initialize the vector representation of entities and relationships in the triple, and obtain the representation features of entities and relationships in the low-dimensional vector space through the Embedding method as the embedding information of entities and relationships;

[0039] Step S302: Calculate the semantic distance between the head entity and the tail entity using cosine similarity according to the vector representation of the entities and the relationships, and merge the semantic distance into the feature vector of the triple as a feature of the triple.

[0040] Furthermore, the step S4 specifically includes:

[0041] Step S401: for each target triple, count the occurrence frequencies of the relationship and tail entity in the knowledge graph dataset;

[0042] Step S402: for each target knowledge triple, count the occurrence frequency of the tail entity in the knowledge graph data set at the object position;

[0043] Step S403: Calculate the degree of combination of the relationship and the tail entity according to the occurrence frequency of the tail entity and the co-occurrence frequency of the relationship and the tail entity, wherein the expression of the combination degree is:

[0044]

[0045] In the formula, O(·) represents the occurrence frequency of ·, Co(·) represents the co-occurrence frequency of e and r, e represents the head entity or tail entity in the triple, and r represents the relationship in the triple.

[0046] Furthermore, the step S5 specifically includes:

[0047] Step S501, constructing a feature vector according to the association strength of the triple determined in step S203, the semantic similarity value determined in step S302, and the degree of combination determined in step S403, to represent the semantic information feature contained in the triple;

[0048] Step S502: Based on the semantic information features and knowledge features determined in step S501, a multi-classifier is used to train the model to complete knowledge inconsistency detection.

[0049] The beneficial effects of the present invention are:

[0050] In order to address the problem of knowledge inconsistency caused by data quality issues and natural language processing errors in the process of building the knowledge graph, the present invention proposes a novel knowledge inconsistency detection method in the knowledge graph based on entity semantic strength, which can effectively detect various inconsistencies in the knowledge graph, plays an important role in improving the knowledge quality of the knowledge graph, and further promotes the implementation of downstream tasks of the knowledge graph (such as question-answering systems, intelligent recommendations, etc.). BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a schematic diagram of the overall process of a method for detecting knowledge inconsistency in a knowledge graph based on entity semantic strength provided in Example 1;

[0052] Figure 2 is a schematic diagram of the association strength of entity pairs in Example 1;

[0053] Figure 3 is a schematic diagram of the semantic similarity of triples in Example 1;

[0054] Figure 4 is a schematic diagram of the degree of integration between relationships and entities in Example 1;

[0055] Figure 5 is a flow chart of processing entity pair association strength in Example 1;

[0056] Figure 6 is a flow chart of processing triple semantic similarity in Example 1;

[0057] Figure 7 is a flow chart of processing triple relationship and entity association degree in Example 1;

[0058] Figure 8 This is a schematic diagram of identifying knowledge inconsistency in Example 1. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0060] Example 1

[0061] See also Figure 1-Figure 8 This embodiment provides a method for detecting knowledge inconsistency in a knowledge graph based on entity semantic strength. This embodiment is a method for detecting knowledge inconsistency based on entity semantic strength. The main flow chart is shown in the attached figure. Figure 1 As shown in the figure, firstly, the knowledge graph is preprocessed and the knowledge inconsistency is formally defined. The potential entity association strength information of knowledge triples is mined using three strategies: entity pair association strength, triple semantic similarity distance, and degree of combination between relationship and entity. The detection of knowledge inconsistency is regarded as a classification problem, and the category to which the knowledge belongs is identified. Based on the category, comprehensive detection of knowledge inconsistency is achieved. The specific implementation steps are as follows:

[0062] Step S1, preprocessing the knowledge graph dataset, first adding type triples as a supplement to the knowledge graph dataset according to the type information of the entities in the knowledge graph, and then generating negative samples according to the negative sample generation principle;

[0063] This embodiment explores the sources of knowledge inconsistency in the knowledge graph, and divides the knowledge inconsistency into three categories based on its manifestations: entity inconsistency, relationship inconsistency, and type inconsistency.

[0064] Its formal definition is as follows:

[0065] Relational inconsistency:

[0066]

[0067] Entity Inconsistency:

[0068]

[0069] Type inconsistency:

[0070]

[0071] Among them, G represents the knowledge contained in the knowledge graph, E and R represent the entity set and relationship set in the knowledge graph respectively, h,t∈E,r∈R. The main workflow of knowledge preprocessing is as follows:

[0072] 1. Supplement entity type triples. Based on the entity’s pattern information, the entity’s type information is presented in the form of triples, and the type is added to the knowledge graph as an additional relationship. Representing the type information and the triples at the data layer in a unified form is more conducive to capturing semantic information of different types of inconsistencies.

[0073] 2. Processing negative samples. Since the noise in the knowledge graph mainly comes from the incorrect mapping between similar entities, the entity or relationship in the knowledge graph is randomly replaced to generate negative samples. Negative samples can help the inconsistency identification model better distinguish the performance characteristics of different inconsistencies. In order to increase the model's feature discrimination of inconsistencies and enhance the model's fault tolerance, a 1:1 method is used to generate inconsistencies. The final proportions of entity inconsistency, relationship inconsistency, and type inconsistency are 25%, 12.5%, and 12.5%, respectively.

[0074] Step S2, abstracting the knowledge graph in the data set preprocessed in step S1 into a representation form of a graph model, extracting entity subgraphs according to the nodes that the head entity can reach, and using the resource information flow on the entity subgraphs to represent the association strength between entity pairs;

[0075] Specifically, the strength of association between entity pairs is an important way to measure entity semantic information. By obtaining the representation relationship of the strength of association between entity pairs, it is possible to effectively determine whether there is a certain relationship between two entities. The possibility of a relationship between weakly related entity pairs is extremely low. According to the graph model representation of the knowledge graph, nodes represent entities and edges represent the relationship between entities. If the association path between node A and node E is relatively dense, then the association between entity A and entity E is relatively strong. See the attached diagram for a schematic diagram. Figure 2 The specific steps for mining the strength of association between entity pairs are as follows. The flowchart is attached. Figure 5 .

[0076] 1. Traverse each triple in the knowledge graph, use the tuple consisting of the head node and the tail node as the key, and the multiple paths between the head node and the tail node as the value to build a dict representation of the knowledge graph. Since there are multiple paths between nodes, use a list to store all the paths and use them as the value of the node.

[0077] 2. Traverse the dict for each entity to build an entity subgraph. Studies have shown that when the relationship chain between entities is too long, there is very little effective information flowing from the head entity to the tail entity. Therefore, the present invention stipulates that the maximum value of the relationship chain in the entity subgraph is 4. Use depth-first search traversal to find the tail nodes that the head node can connect to in the depth range of 4 and add them to the subgraph.

[0078] 3. For each triple in the knowledge graph, the subgraph where the head node is located is used as a basis to determine whether the tail entity is in the subgraph. If not, missing value processing is performed and the strength feature of the entity of the triple is treated as zero value; otherwise, the final entity information value of the triple is calculated based on the resource flow from the head entity to the tail entity.

[0079] Step S3, obtaining the representation of entities and relations in the triplet in the low-dimensional vector space, and calculating the cosine similarity to represent the semantic similarity value from the head entity to the tail entity;

[0080] Specifically, calculating the semantic distance between the head entity and the tail entity in a triple is an important basis for judging the correctness of a triple. According to the representation of entities and relations in low-dimensional vector space, the translation and addition of the head entity and the relation approximates the vector value of the tail entity, as shown in the attached diagram. Figure 3 The flowchart of processing the semantic similarity distance of triples is shown in the attached Figure 6 .

[0081] 1. Initialize the vector representation of entities and relationships in the knowledge graph, and obtain the optimal representation of entities and relationships through continuous gradient learning.

[0082] 2. Calculate the semantic similarity distance between the head entity and the tail entity. Use the vector representation of entities and relations in the knowledge triples and use cosine similarity to measure the semantic distance between the head entity and the tail entity. Merge the semantic similarity distance as a feature of the triple into the feature vector of the triple.

[0083] Step S4: Count the occurrence frequency of the tail entity corresponding to the relationship in the triple in the knowledge graph, and measure the path credibility of the triple corresponding to the relationship according to the degree of combination of the relationship and the tail entity. For each triple, the value calculated by combining the tail entity and the relationship in the triple and the expression represents the path feasibility of the triple;

[0084] Specifically, the entity pair association strength and the semantic similarity distance of the triples can play a good role in mining the features of entity inconsistency and type inconsistency, but the features of relationship inconsistency need to rely on the combination degree of relationship and entity to complete. By counting the type distribution of the tail entity corresponding to the relationship and measuring the combination degree of relationship and entity, it is possible to effectively detect the situation where there is a relationship in the entity pair but the relationship is incorrect. The schematic diagram of the calculation of the combination degree of relationship and entity is shown in the attached figure. Figure 4 As shown, the calculation flow chart is attached Figure 7 .

[0085] 1. For each target knowledge triple, count the frequency of occurrence of the relationship and tail entity in the knowledge graph dataset.

[0086] 2. For each target knowledge triple, count the frequency of occurrence of the tail entity in the object position in the knowledge graph dataset.

[0087] 3. Calculate the degree of integration between the relationship and the tail entity based on the occurrence frequency of the tail entity and the co-occurrence frequency of the relationship and the tail entity. The calculation formula is as follows:

[0088]

[0089] Among them, O(·) represents the occurrence frequency of ·, and Co(·) represents the co-occurrence frequency of e and r.

[0090] Step S5: construct feature engineering based on the association strength of the triples determined in step S2, the semantic similarity value determined in step S3, and the degree of combination determined in step S4, and use multiple classifiers to train the inconsistency recognition model to complete the recognition of all possible inconsistency types.

[0091] Specifically, for the three different entity association strength representation methods in steps S2, S3, and S4, these three semantic strength information are combined to form the feature vector of each triple, and the detection of knowledge inconsistency is regarded as a classification problem. The feature vector of each triple is input into the classifier training model to obtain the recognition model for knowledge inconsistency. The specific steps are as follows, and the flow chart is attached. Figure 8 .

[0092] 1. Construct a feature vector based on the entity semantic information in the triple obtained in steps S2, S2, and S2, representing the semantic information features contained in the triple.

[0093] 2. Using the semantic information features of the knowledge in the knowledge graph and its knowledge features (consistency or not), use a multi-classification classifier to train the model and complete the knowledge inconsistency detection.

[0094] The knowledge inconsistency detection method proposed in this embodiment is generally described as follows:

[0095]

[0096]

[0097] The matters not described in detail in the present invention are all known technologies to those skilled in the art.

[0098] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A method for detecting knowledge inconsistency in a knowledge graph based on entity semantic strength, characterized in that: The method comprises: Step S1, preprocessing the knowledge graph dataset, first adding type triples as a supplement to the knowledge graph dataset according to the type information of the entities in the knowledge graph, and then generating negative samples according to the negative sample generation principle; Step S2, abstracting the knowledge graph in the data set preprocessed in step S1 into a representation form of a graph model, extracting entity subgraphs according to the nodes that the head entity can reach, and using the resource information flow on the entity subgraphs to represent the association strength between entity pairs; Step S3, obtaining the representation of entities and relations in the triplet in the low-dimensional vector space, and calculating the cosine similarity to represent the semantic similarity value from the head entity to the tail entity; Step S4: Count the occurrence frequency of the tail entity corresponding to the relationship in the triple in the knowledge graph, and measure the path credibility of the triple corresponding to the relationship according to the degree of combination between the relationship and the tail entity; Step S5, constructing feature engineering according to the association strength of the triples determined in step S2, the semantic similarity value determined in step S3, and the degree of combination determined in step S4, and training the inconsistency recognition model using multiple classifiers to complete the recognition of all possible inconsistency types; The step S1 specifically includes: Step S101: Convert the entity type information in the knowledge graph into a triple representation to ensure the completeness of the data in the knowledge graph; Step S102: according to the triple representation of knowledge in the knowledge graph and the problems that occur during the knowledge graph construction process, the problems that occur during the knowledge graph construction process specifically include: vocabulary similarity problems that occur during natural language processing, redundancy problems that occur in entity extraction, polysemy problems that occur in knowledge fusion, and data conflicts caused by multiple data sources, the knowledge inconsistencies are divided, wherein for the target triple (h, r, t), the corresponding inconsistencies specifically include: Relational inconsistency: Entity Inconsistency: Type inconsistency: Among them, G represents the knowledge contained in the knowledge graph, E and R represent the entity set and relationship set in the knowledge graph respectively, h,t∈E,r∈R; Step S103: Generate inconsistent negative samples in a 1:1 manner according to the three types of inconsistencies and features in step S102, wherein the features are specifically: features represented when similar entities and their corresponding relationships are incorrectly mapped during the knowledge graph construction process; the proportions of negative samples corresponding to entity inconsistency, relationship inconsistency, and type inconsistency are: 25%, 12.5%, and 12.5%; The step S2 specifically includes: Step S201, traverse each triple in the knowledge graph, and construct a directed graph representation of the knowledge graph with entities in the triple as nodes and relationships as edges; Step S202: For each entity in the knowledge graph, use depth-first search to find the entity set that the entity can be connected to, and record the entity search path to construct an entity subgraph; Step S203: For each triple in the knowledge graph, the information distribution value flowing from the head entity to the tail entity is calculated according to the resource flow in the entity subgraph where the head entity is located, which represents the association strength between the entity pairs in the triple. The information distribution value calculation expression is as follows: In the formula, E i-1 (·,e) indicates the path r i The set of direct predecessor nodes to reach e, E i (e′,·) represents the i The path r passed by the predecessor i The set of direct successor nodes of ′ e represents the weight of the edge e′→e, which depends on the number of edges connecting e′ and e; θ is a hyperparameter set to avoid the closed loop problem and is set to 0.15; The step S4 specifically includes: Step S401: for each target triple, count the occurrence frequencies of the relationship and tail entity in the knowledge graph dataset; Step S402: for each target knowledge triple, count the occurrence frequency of the tail entity in the knowledge graph data set at the object position; Step S403: Calculate the degree of combination of the relationship and the tail entity according to the occurrence frequency of the tail entity and the co-occurrence frequency of the relationship and the tail entity, wherein the expression of the combination degree is: In the formula, O(·) represents the frequency of occurrence of ·, Co(·) represents the co-occurrence frequency of e and r, e represents the head entity or tail entity in the triple, and r represents the relationship in the triple; The step S5 specifically includes: Step S501, constructing a feature vector according to the association strength of the triple determined in step S203, the semantic similarity value determined in step S302, and the degree of combination determined in step S403, to represent the semantic information feature contained in the triple; Step S502: Based on the semantic information features and knowledge features determined in step S501, a multi-classifier is used to train the model to complete knowledge inconsistency detection.

2. According to claim 1, a method for detecting knowledge inconsistency in a knowledge graph based on entity semantic strength is characterized in that: In step S202, the maximum value of the relationship chain in the entity subgraph is 4, and a depth-first search traversal is used to sequentially find tail nodes to which the head node can be connected within a depth range of 4 and add them to the entity subgraph.

3. According to claim 2, a method for detecting knowledge inconsistency in a knowledge graph based on entity semantic strength is characterized in that: In step S203, the calculating of the information allocation value flowing from the head entity to the tail entity according to the resource flow in the entity subgraph where the head entity is located specifically includes: Based on the entity subgraph where the head entity is located, determine whether the tail entity is in the subgraph; If it is not present, the missing value is processed and the entity of the triple is treated as zero value for the intensity feature; Otherwise, the final entity information value of the triple is calculated based on the resource flow from the head entity to the tail entity.

4. According to claim 3, a method for detecting knowledge inconsistency in a knowledge graph based on entity semantic strength is characterized in that: The step S3 specifically includes: Step S301, initialize the vector representation of entities and relationships in the triple, and obtain the representation features of entities and relationships in the low-dimensional vector space through the Embedding method as the embedding information of entities and relationships; Step S302: Calculate the semantic distance between the head entity and the tail entity using cosine similarity according to the vector representation of the entities and the relationships, and merge the semantic distance into the feature vector of the triple as a feature of the triple.

Images