WO2022057671A1

WO2022057671A1 - Neural network–based knowledge graph inconsistency reasoning method

Info

Publication number: WO2022057671A1
Application number: PCT/CN2021/116777
Authority: WO
Inventors: 陈华钧; 李娟�; 张文
Original assignee: 浙江大学
Priority date: 2020-09-16
Filing date: 2021-09-06
Publication date: 2022-03-24
Also published as: CN112100403A

Abstract

A neural network-based knowledge graph inconsistency reasoning method, comprising the following steps: performing representation learning on a triplet by using a knowledge representation learning algorithm to obtain entity representations and a relationship representation and calculating a representation score; then using the entity representations and the relationship representation as an input of a neural network; modeling an axiom by means of the neural network by using the triplet to learn parameters of the neural network used for representing the corresponding axiom, thereby obtaining an axiom model; obtaining an axiom prediction value of the triplet by using the axiom model; and implementing determination of inconsistency of the triplet and the corresponding axiom on the basis of the representation score and the axiom prediction value of the triplet. In the method, there is no need to give ontology information, an inconsistency axiom is learned by using a neural network, and whether there is inconsistency in a triplet and whether there is inconsistency on the given axiom are determined by means of a knowledge representation learning algorithm and the neural network.

Description

A Neural Network-based Inconsistency Reasoning Method for Knowledge Graphs

technical field

The invention belongs to the field of knowledge graphs and neural networks, and in particular relates to a knowledge graph inconsistency reasoning method based on neural networks.

Background technique

Knowledge graph is a knowledge system formed by structuring knowledge, and has been widely used in knowledge-driven tasks such as search engines, recommendation systems, and question answering systems. In order to store and utilize knowledge efficiently, large-scale knowledge graphs for open and vertical domains have been constructed using manual annotation, semi-automatic or automated methods. Classical knowledge graphs, such as Wikidata, Freebase, and DBpedia, use triples to store relationships between entities and entities. Each triplet corresponds to a piece of knowledge, for example (China, the capital is, Beijing) means that the capital of China is Beijing, where "China" is called the head entity, "Beijing" is called the tail entity, and "the capital is" is called for relationship.

Since the construction of knowledge graphs is often automated or semi-automated, the constructed graphs may have quality problems such as incomplete knowledge and knowledge errors. For the problem of incomplete knowledge, many knowledge graph representation learning algorithms have been proposed to predict new knowledge to complement existing knowledge graphs, mainly predicting missing entities by given entities and relationships, or predicting between two entities given two entities. possible relationship. Aiming at the problem of knowledge error, it includes two situations: not conforming to the axioms and conforming to the axioms but the knowledge content is incorrect, mainly through knowledge inconsistency reasoning to realize the detection of wrong knowledge in order to correct the errors.

The purpose of knowledge graph inconsistency reasoning is to detect wrong knowledge. It contains two types of tasks, one is to detect whether a triple is consistent, and the other is to detect whether an axiom of a triple is consistent. Perform binary classification. First of all, the existing work on inconsistent reasoning of triples in the knowledge graph requires the knowledge base ontology information, or some custom templates. Existing knowledge bases usually do not have well-defined ontologies, and manual definition by experts is not only time-intensive, but also leads to the problem of incomplete ontologies. Secondly, the current knowledge representation learning models include distance models, semantic models, and neural network models, etc., which embed knowledge graphs into low-dimensional vector spaces, and represent entities and relationships in triples as low-dimensional vectors.

The above methods obtain the score of the triplet through vector calculation to determine whether a piece of knowledge is correct, which corresponds to the triplet classification task. At the same time, these methods reflect the advantages of knowledge representation learning algorithms and neural networks in knowledge graph reasoning tasks. However, these methods can only judge the consistency of triples, and cannot judge whether the axioms corresponding to triples are consistent in fine-grained manner. Therefore, it is urgent to detect whether a certain axiom of triples is consistent to realize the detection of wrong knowledge. for error correction.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a knowledge graph inconsistency reasoning method based on neural network, which does not require given ontology information, uses neural network to learn the inconsistency axiom, and judges whether a triplet is a triple through knowledge representation learning algorithm and neural network. There is an inconsistency, whether there is an inconsistency on the given axioms.

In order to realize the above-mentioned purpose of the invention, the present invention provides the following technical solutions:

A neural network-based knowledge graph inconsistency reasoning method, comprising the following steps:

Taking the entity representation and relationship representation of the triplet as input, the knowledge representation learning algorithm is used to learn the entity representation and relationship representation of the triplet in the knowledge graph, and the representation score of the triplet is calculated at the same time. Representation and relational representation are used as the input of the neural network, using triples to model the axioms through the neural network to learn the parameters of the neural network used to represent the corresponding axioms to obtain the axiom model, and using the axiomatic model to obtain the axiom prediction value of the triples , based on the representation score of the triplet and the axiom prediction value to realize the judgment of the inconsistency of the triplet and the corresponding axiom.

When the knowledge representation learning algorithm is used for triple representation learning, the knowledge graph is embedded in a low-dimensional vector space, entities and relationships are represented as vectors, and the structure information of triples is preserved through the knowledge representation learning algorithm. Given a triple Group (s,r,o), you can choose any knowledge representation learning algorithm to retain structural information, such as TransE algorithm, DistMult algorithm, etc. The input is the vector representation of entities and relationships, and the output is the representation score of the current triplet. The specific calculations of the score fr are: _TransE : _fr (s,o)=||s+ro||; _DistMult : fr (s, o)=-s ^T M _r o, where M _r is a diagonal matrix, and s, r, and o are vector representations of s, r, and o, respectively. The value of the correct triple f _r (s, o) is lower than that of the error triple.

Through neural network modeling axiom models, different parameters are learned for each neural network, and each set of parameters is considered as a feature of the corresponding axiom. The knowledge representation learning algorithm and the neural network are trained at the same time, so that when learning the vector representation and the neural network parameters, it can not only encode the structural information of the triplet, but also make the triplet conform to the relevant axioms at the same time. The representation learning algorithm can be any knowledge graph representation learning algorithm, and the neural network performs binary classification to output the probability that each axiom satisfies the consistency. The specific input of the neural network during the training process comes from the elements of the current triplet, or the elements in the related triplet found through the triplet. Firstly, the elements from triples or related triples that need to be considered for the corresponding axioms are analyzed; then, according to the assumptions of the axioms, the elements of the existing triples and their related triples are regarded as neural networks through the existing triples conforming to the axiom constraints. The positive samples of each neural network module are constructed through closed assumptions. This setting allows the model to learn the axioms only by using the triples existing in the knowledge graph without ontology information.

In the above-mentioned neural network-based knowledge graph inconsistency reasoning method, when using triples to model the axioms through the neural network,

First, from the axioms defined by the OWL2 ontology language that can be used for inconsistency detection, select the axioms to be considered and analyze the constraints or conditions of the inconsistency corresponding to each axiom in the knowledge graph;

Then, the neural network model corresponding to each axiom is constructed, and the positive samples of each neural network model are constructed according to the constraints or conditions of the inconsistency corresponding to the axioms, and the negative samples are constructed based on the positive samples;

Next, for the elements of triples that need to be considered in the neural network model corresponding to each axiom, the corresponding entity representation and relationship representation are spliced and input into the neural network model, and the predicted value corresponding to each axiom is calculated and calculated based on all The prediction value corresponding to the axiom is obtained by the prediction score of the neural network model, and the representation score and prediction score calculated by the knowledge representation learning algorithm are combined to obtain the total score of the triplet;

Finally, an edge loss function is constructed according to the total score of the positive sample triplet and the corresponding total score of the negative sample triplet, and the edge loss function is used to jointly update the optimized knowledge representation learning algorithm and neural network model parameters, as well as entity representation and relation representation. After the end, the neural network model determined by the parameters is finally learned as the axiomatic model, as well as the entity representation and relation representation determined in the knowledge graph.

In the above-mentioned neural network-based knowledge graph inconsistency reasoning method, when selecting axioms for inconsistency detection, first determine whether the axioms can be used for inconsistency detection according to the conditions or constraints mentioned in the axiom definition that need to be satisfied, and then Select axioms from the axioms that can be used for inconsistency detection, and benchmark the conditions or constraints corresponding to the selected axioms to the related elements of the triples in the knowledge graph and the correlation between the elements of the triples, and then realize the positive construction of the axioms. sample.

In the above-mentioned neural network-based knowledge graph inconsistency reasoning method, when judging the inconsistency of triples, a relationship threshold is set for each relationship, and the knowledge representation is used according to the entity representation and relationship representation determined at the end of the optimization. The learning algorithm and the axiomatic model calculate the total score of the triplet. When the total score of the triplet is lower than the relation threshold, the triplet is considered to be a correct triplet, otherwise it is a wrong triplet, which means there is inconsistency sex.

In the above-mentioned neural network-based knowledge graph inconsistency reasoning method, when judging the inconsistency of the axioms corresponding to the triples, an axiom threshold is set for each axiom, and the entity vector representation and relationship determined when the knowledge representation learning model is generated is used. Vector representation, using each axiom model to calculate the predicted value of the triple for the axiom, when the predicted value of the triple for the axiom is lower than the corresponding axiom threshold, the triple is considered inconsistent on the axiom.

In one embodiment, for the triple (s,r,o), the selected axioms include:

The Object Property Domain, referred to as the domain axiom, defines the type of the head entity s of the relationship r should conform to the corresponding category;

Object Property Range, referred to as the range axiom, defines that the type of the tail entity o of the relation r should conform to the corresponding category;

Disjoint Object Properties, referred to as the disjoint axiom, defines that relation r and relation r ₁ are mutually exclusive, and triples (s, r, o) and triples (s, r ₁ , o) should not exist at the same time in a knowledge graph;

Irreflexive Object Property, referred to as the irreversible axiom, requires two-step judgment. First, check whether the relationship is reflexive, and then determine whether the head entity and the tail entity are equal. If the relationship r is anti-reflexive, the entity cannot pass the relationship. pointing to itself i.e. s=o;

Asymmetric Object Property, referred to as asymmetric axiom, requires two-step judgment. First, check whether the relationship is antisymmetric, and then find the triple (s ₁ ,r,o ₁ ) whose relationship is r. If judging whether s ₁ =o and o ₁ =s exist, if the relationship is antisymmetric, then the two triples (s,r,o) and (o,r,s) should not exist in a knowledge graph at the same time .

The input of the neural network module is the element of the current triplet or the triplet related to the current triplet, and the output is the probability that conforms to the corresponding axiom. The calculations of the neural network module are as follows:

Domain axiom prediction: P _dm (s,r,o)=g(W ₁ ·[r; s]+b ₁ )

Range axiom prediction: P _rg (s,r,o)=g(W ₂ ·[r; o]+b ₂ )

Disjoint axiom prediction: P _dis (s,r,o,s,r ₁ ,o)=g(W ₃ ·[r;r ₁ ]+b ₃ )

The predicted value of the irreflexive axiom: P _irre (s,r,o)=g(W ₄ ·[s;r;o]+b ₄ )

The asymmetric axiom predicts the value:

_Pasym (s,r,o,s ₁ ,r,o ₁ )=g(W ₅ ·[s;r;o;s ₁ ;r;o ₁ ]+b ₅ );

Then the total score of the triplet is:

f(s,r,o)=f _r (s,o)+α(1-P _dm )+β(1-P _rg )+∈(1-P _irre )+ζ(1-P _dis )+η (1-P _asym )

Among them, s, s ₁ , r, r ₁ , o, o ₁ respectively represent the learning representation vector, symbol of entity s, entity s ₁ , relation r, relation r ₁ , entity o, and entity o ₁ ; represent the splicing operation, g () represents the sigmoid function, W ₁ , W ₂ , W ₃ , W ₄ , W ₅ represent the weight vector, b ₁ , b ₂ , b ₃ , b ₄ , b ₅ are biases, α, β, ∈, ζ, η is a hyperparameter representing the weight, and _fr (s, o) represents the representation score of the triplet obtained by using the knowledge representation learning model.

For correct triples, the axiom probability value output by each axiomatic model is relatively higher than that of wrong triples, and the above scoring function ensures that the total score of triples of positive samples is lower than that of triples of negative samples. The constructed marginal loss function is:

Among them, F represents the set of positive samples, F' represents the set of negative samples, f(s, r, o) represents the total score of the triplet (s, r, o), and f(s', r', o') represents the The total score for the triple (s',r',o').

Based on the domain axiom, range axiom, disjoint axiom, irreflexive axiom and asymmetric axiom, when judging the inconsistency of triples, a triple (s, r, o) is given, and the training phase is used to learn entities and relationships. The vector representation and the parameters of the neural network model, calculate the final score f(s,r,o), and the values of each axiom module including P _dm , P _rg , P _irre , P _dis , and P _asym . For the consistency judgment of triples, a threshold is introduced for each relationship. When the score f(s,r,o) is lower than the threshold, it means that the triplet is a correct triplet, otherwise it is a wrong triplet, That is, there is inconsistency. In order to further distinguish whether the triples have inconsistencies in the axioms currently considered, a threshold is introduced for each axiom model according to the validation set. If P _dm , P _rg , P _irre , P _dis , and P _asym are all lower than the corresponding axiom thresholds , it indicates that the triplet has inconsistency in domain axiom, range axiom, disjoint axiom, irreflexive axiom and asymmetric axiom respectively.

Compared with the prior art, the beneficial effects of the present invention at least include:

(1) The above-mentioned knowledge graph inconsistency reasoning method does not require well-defined ontology information, and only uses the existing triples in the knowledge graph for axiom learning, so that it can also be captured on knowledge graphs without ontology definitions or incomplete ontology definitions. Part of the axioms and inconsistency reasoning, greatly reducing labor costs. At the same time, the knowledge representation learning model and the axiomatic model can be used in any knowledge graph.

(2) Compared with the knowledge representation learning algorithm that only considers structural information, this method converts axiom learning into neural network parameter learning, and uses the knowledge representation learning algorithm and neural network joint training to allow the model to retain structural information and learn inconsistency. related axioms. Axiom learning allows the model to detect not only inconsistent triples, but also fine-grained detection of inconsistencies for several axioms under consideration. In this way, inconsistency inference can be better implemented, and it is convenient for subsequent correction of triples.

Description of drawings

In order to illustrate the embodiments of the present invention or the technical solutions in the prior art more clearly, the following briefly introduces the accompanying drawings used in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only These are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can also be obtained from these drawings without creative efforts.

Figure 1 is a flowchart of a neural network-based knowledge graph inconsistency reasoning method.

detailed description

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, and do not limit the protection scope of the present invention.

Figure 1 is a flowchart of a neural network-based knowledge graph inconsistency reasoning method. As shown in Figure 1, for a given knowledge graph containing a large number of triples (s, r, o), the knowledge graph inconsistency reasoning method includes the following steps:

Step 1, select the following five axioms from the axioms that can be used for inconsistency detection in the OWL2 object attribute axioms, and analyze the description and judgment conditions of the five axioms in OWL2 as follows:

Step 2: Obtain the training samples of the neural network corresponding to each kilometer according to the judgment conditions, that is, determine the specific input of each neural network according to the judgment conditions of the axioms. The following is a sample knowledge graph to illustrate the input of each axiom model:

s ₁ s ₁	r ₁ r ₁	o ₁ o ₁
s ₂ s ₂	r ₂ r ₂	o ₂ o ₂
s ₁ s ₁	r ₃ r ₃	o ₁ o ₁
s ₂ s ₂	r ₁ r ₁	o ₂ o ₂
s ₁ s ₁	r ₁ r ₁	s ₁ s ₁
o ₁ o ₁	r ₃ r ₃	s ₁ s ₁

The given sample graph contains 6 triples, including entities (s ₁ , s ₂ , o ₁ , o ₂ ) and relations (r ₁ , r ₂ , r ₃ ). During training, the domain axiom focuses on relations and head entities, and the module inputs are (r ₁ ,s ₁ ),(r ₂ ,s ₂ ),(r ₃ ,s ₁ ),(r ₁ ,s ₂ ),(r ₃ ,o ₁ ); the range axiom concerns relations and tail entities, the module inputs are (r ₁ ,o ₁ ),(r ₂ ,o ₂ ),(r ₃ ,o ₁ ),(r ₁ ,o ₂ ),( r ₁ ,s ₁ ),(r ₃ ,s ₁ ); the disjoint axiom focuses on the correlation of relations in triples where both head and tail entities are the same. The input of this module is (s ₁ ,r ₁ ,o ₁ , s ₁ ,r ₃ ,o ₁ ),(s ₂ ,r ₂ ,o ₂ ,s ₂ ,r ₁ ,o ₂ ); the irreflexive axiom concerns the current triple, and the module input is (s ₁ ,r ₁ ,o ₁ ),(s ₂ ,r ₂ ,o ₂ ),(s ₁ ,r ₃ ,o ₁ ),(s ₂ ,r ₁ ,o ₂ ),(s ₁ ,r ₁ ,s ₁ ),(o ₁ , r ₃ , s ₁ ); the asymmetric axiom concerns two triples with the same relationship, the input is (s ₁ ,r ₁ ,o ₁ ,s ₂ ,r ₁ ,o ₂ ),(s ₁ ,r ₁ ,o ₁ ,s ₁ ,r ₁ ,s ₁ ),(s ₂ ,r ₁ ,o ₂ ,s ₁ ,r ₁ ,s ₁ ). The knowledge representation learning algorithm focuses on the current triple, and the input is (s ₁ ,r ₁ ,o ₁ ), (s ₂ ,r ₂ ,o ₂ ), (s ₁ ,r ₃ ,o ₁ ), (s ₂ ,r ₁ ,o ₂ ), (s ₁ ,r ₁ ,s ₁ ),(o ₁ ,r ₃ ,s ₁ ). The input of each of the above modules can be regarded as a positive sample for that module.

Step 3, jointly train the knowledge representation learning algorithm and the neural network corresponding to the axiom

After the neural network corresponding to each axiom is constructed, the entity representation and relation representation of the sample are spliced and input into the neural network model, the predicted value corresponding to each axiom is calculated, and the neural network model is obtained based on the predicted value corresponding to all axioms The prediction score of the triplet is combined with the representation score of the triplet and the prediction score of the axiomatic model to obtain the total score of the triplet; finally, the edge loss function is constructed according to the total score of the positive sample triplet and the corresponding total score of the negative sample triplet , and use the edge loss function to jointly update and optimize the parameters of the knowledge representation learning algorithm and the neural network model. After the optimization, the knowledge representation learning algorithm of the determined entity vector representation and relation vector representation is the knowledge representation learning model, and the neural network model determined by the parameters is Axiom Model.

Step 4: After the knowledge representation learning model and the axiom model trained in step 3, the inconsistent reasoning of the knowledge graph can be realized. Given a triplet, calculate the final score for the current triplet, if it is lower than the threshold, the triplet is judged to be correct, if it is higher than the threshold, it is considered that the triplet is inconsistent; for each axiom module, calculate the probability of conforming to the axiom, corresponding to the The probability of is higher than the threshold value, indicating that the triplet conforms to the corresponding axiom consistency, otherwise there is inconsistency on the axiom.

The above-mentioned specific embodiments describe in detail the technical solutions and beneficial effects of the present invention. It should be understood that the above-mentioned embodiments are only the most preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, additions and equivalent substitutions made within the scope shall be included within the protection scope of the present invention.

Claims

A neural network-based knowledge graph inconsistency reasoning method, characterized in that it includes the following steps:

Taking the entity representation and relationship representation of the triplet as input, the knowledge representation learning algorithm is used to learn the entity representation and relationship representation of the triplet in the knowledge graph, and the representation score of the triplet is calculated at the same time. Representation and relational representation are used as the input of the neural network, using triples to model the axioms through the neural network to learn the parameters of the neural network used to represent the corresponding axioms to obtain the axiom model, and using the axiomatic model to obtain the axiom prediction value of the triples , based on the representation score of the triplet and the axiom prediction value to realize the judgment of the inconsistency of the triplet and the corresponding axiom.
The method for reasoning about inconsistency of knowledge graphs based on neural network according to claim 1, characterized in that, when using triples to model the axioms through a neural network,

First, from the axioms defined by the OWL2 ontology language that can be used for inconsistency detection, select the axioms to be considered and analyze the constraints or conditions of the inconsistency corresponding to each axiom in the knowledge graph;

Then, the neural network model corresponding to each axiom is constructed, and the positive samples of each neural network model are constructed according to the constraints or conditions of the inconsistency corresponding to the axioms, and the negative samples are constructed based on the positive samples;

Next, for the elements of the triplet that the neural network model corresponding to each axiom needs to consider, the corresponding entity representation and relationship representation are spliced and input into the neural network model, and the predicted value corresponding to each axiom is calculated and calculated based on all axioms. The corresponding prediction value is obtained by the prediction score of the neural network model, and the representation score and prediction score calculated by the knowledge representation learning algorithm are combined to obtain the total score of the triplet;

Finally, an edge loss function is constructed according to the total score of the positive sample triplet and the corresponding total score of the negative sample triplet, and the edge loss function is used to jointly update the optimized knowledge representation learning algorithm and neural network model parameters, as well as entity representation and relation representation. After the end, the neural network model determined by the parameters is finally learned as the axiomatic model, as well as the entity representation and relation representation determined in the knowledge graph.
The method for reasoning about inconsistency of knowledge graphs based on neural network according to claim 2, characterized in that, when selecting axioms for inconsistency detection, the axioms are first judged according to the conditions or constraints mentioned in the axiom definitions that need to be satisfied. Whether it can be used for inconsistency detection, and then select the axioms from the axioms that can be used for inconsistency detection, and mark the conditions or constraints corresponding to the selected axioms to the related elements of the triplet in the knowledge graph, and the relationship between the triplet elements. Correlation, and then realize the positive sample of the construction axiom.
The neural network-based knowledge graph inconsistency reasoning method according to claim 2, wherein when judging the inconsistency of triples, a relationship threshold is set for each relationship, and the entity representation determined at the end of the optimization is based on and relational representation, using the knowledge representation learning algorithm and axiomatic model to calculate the total score of the triplet, when the total score of the triplet is lower than the relational threshold, the triplet is considered a correct triplet, otherwise it is an error The triplet of , there is inconsistency.
The method for reasoning about inconsistency of knowledge graphs based on neural network according to claim 2, wherein when judging the inconsistency of axioms corresponding to triples, an axiom threshold is set for each axiom, and each axiom model is used Calculate the predicted value of the triplet for the axiom. When the predicted value of the triplet for the axiom is lower than the corresponding axiom threshold, it is considered that the triplet is inconsistent on the axiom.
The method for reasoning about inconsistency of knowledge graph based on neural network according to claim 2, wherein, for triples (s, r, o), the selected axioms include:

The object attribute domain, referred to as the domain axiom, defines that the type of the head entity s of the relationship r should conform to the corresponding category;

The range of object attributes, referred to as the range axiom, defines that the type of the tail entity o of the relation r should conform to the corresponding category;

The disjoint object attribute, referred to as the disjoint axiom, defines that relation r and relation r 1 are mutually exclusive, and triples (s, r, o) and triples (s, r 1 , o) should not exist in a knowledge graph at the same time ;

The irreversible object property, referred to as the irreversible axiom, requires two-step judgment. First, check whether the relationship is anti-reflexive, and then determine whether the head entity and the tail entity are equal. If the relationship r is anti-reflexive, the entity cannot point to itself through the relationship, that is, s= o;

The asymmetric object property, referred to as the asymmetric axiom, requires two-step judgment. First, check whether the relationship is antisymmetric, and then find the triple (s 1 , r, o 1 ) with the relationship r, and determine whether there is s 1 = o and o 1 = s, if the relation is antisymmetric, the two triples (s,r,o) and (o,r,s) should not exist in a knowledge graph at the same time.
The method for reasoning about inconsistency of knowledge graphs based on neural network according to claim 6, wherein the input of the neural network module is the element of the current triplet or the triplet related to the current triplet, and the output is conforming to the corresponding axiom The probability of , the calculation of the neural network module is as follows:

Domain axiom prediction: P dm (s,r,o)=g(W 1 ·[r; s]+b 1 )

Range axiom prediction: P rg (s,r,o)=g(W 2 ·[r; o]+b 2 )

Disjoint axiom prediction: P dis (s,r,o,s,r 1 ,o)=g(W 3 ·[r;r 1 ]+b 3 )

The predicted value of the irreflexive axiom: P irre (s,r,o)=g(W 4 ·[s;r;o]+b 4 )

The asymmetric axiom predicts the value:

Pasym (s,r,o,s 1 ,r,o 1 )=g(W 5 ·[s;r;o;s 1 ;r;o 1 ]+b 5 );

Then the total score of the triplet is:

f(s,r,o)=f r (s,o)+α(1-P dm )+β(1-P rg )+ε(1-P irre )+ζ(1-P dis )+η (1-P asym )

Among them, s, s 1 , r, r 1 , o, o 1 respectively represent the learning representation vector, symbol of entity s, entity s 1 , relation r, relation r 1 , entity o, and entity o 1 ; represent the splicing operation, g () represents the sigmoid function, W 1 , W 2 , W 3 , W 4 , W 5 represent the weight vector, b 1 , b 2 , b 3 , b 4 , b 5 are biases, α, β, ε, ζ, η is a hyperparameter representing the weight, and fr (s, o) represents the representation score of the triplet obtained by using the knowledge representation learning model.
The neural network-based knowledge graph inconsistency reasoning method according to claim 1, wherein the constructed edge loss function is:

Among them, F represents the set of positive samples, F' represents the set of negative samples, f(s, r, o) represents the total score of the triplet (s, r, o), and f(s', r', o') represents the The total score for the triple (s',r',o').
The neural network-based inconsistency reasoning method for knowledge graphs according to claim 1, wherein the knowledge representation learning algorithm comprises a TransE algorithm or a DistMult algorithm.