KR102273136B1 - Method and Apparatus for Completing Zero-Shot Knowledge Graph Using Multi-Hop Neighborhoods - Google Patents

Abstract

The present embodiments generate an inductive vector representation of new entities and relationships by utilizing additional shared information including words constituting the relationship names and entity descriptions shared between the entity-relationships in the knowledge graph and the entity-relationships of the new triple. By doing so, a method and apparatus for completing a zero-shot knowledge graph that can verify the validity of knowledge expressions for new entities and relationships that have not been seen in the learning process of the existing knowledge graph as knowledge without re-learning are provided.

Description

Method and Apparatus for Completing Zero-Shot Knowledge Graph Using Multi-Hop Neighborhoods

The technical field to which the present invention pertains relates to a method and apparatus for completing a zero-shot knowledge graph based on convolutional learning.

The content described in this section merely provides background information for the present embodiment and does not constitute the prior art.

A knowledge graph is one of the effective data modeling methods. The knowledge graph is being used importantly in various applications such as entity ambiguity, question and answer, and information retrieval.

The knowledge graph is composed of a set of triples <h,r,t> called facts, and the head entity h and the tail entity t are connected by a semantic relationship r. The knowledge graph integrates and interoperates with vastly disparate sources of knowledge, such as Linked Open Data.

For example, various knowledge graphs such as DBpedia, WikiData, Freebase, WordNet, and Google Knowledge Graph have been created. Since a lot of information is missing or inaccurate in these knowledge graphs, there is still a limit to being actively utilized in the real world.

State-of-the-art knowledge graphs learn only from static knowledge graphs in which a set of entities and relationships is fixed. Real-world scenarios and relationships constantly add/remove/change entities over time.

Korean Patent Publication No. 10-1914853 (2018.10.29)

Embodiments of the present invention relate to inductive expression learning supporting zero-shot triple inference, wherein the word constituting the relationship name and the entity description shared between the entity-relationships in the knowledge graph and the entity-relationships of the new triple By generating an inductive vector expression for new entities and relationships using the included additional shared information, the knowledge expression for new entities and relationships that have not been seen in the learning process of the existing knowledge graph is validated as knowledge without re-learning, and evolution Its main purpose is to generalize the entities and relationships that

Other objects not specified in the present invention may be additionally considered within the scope that can be easily inferred from the following detailed description and effects thereof.

According to an aspect of this embodiment, in a zero-shot knowledge graph completion method by a computing device, extracting specific relationship characteristic information using additional shared information shared in the knowledge graph, the multi-hop neighbor relationship in the knowledge graph There is provided a zero-shot knowledge graph completion method comprising the steps of generating comprehensive feature information on a zero-shot triple from the specific relational feature information, and calculating a validity score of the comprehensive feature information.

According to another aspect of this embodiment, in the zero-shot knowledge graph completion apparatus including one or more processors and a memory for storing one or more programs executed by the one or more processors, the processor includes additional shared information shared in the knowledge graph extracting specific relational characteristic information using , and the processor generates comprehensive characteristic information about a zero-shot triple from the specific relational characteristic information in a multi-hop neighbor relationship in the knowledge graph, and the processor generates the comprehensive characteristic information of the comprehensive characteristic information. It provides a zero-shot knowledge graph completion device, characterized in that for calculating a validity score.

As described above, according to the embodiments of the present invention, the entity-relationship in the knowledge graph and the entity-relationship of the new triple are shared between the entity description and the additional shared information including the word constituting the relationship name to create new By generating an inductive vector representation of entities and relationships, there is an effect of verifying the validity of knowledge representations about new entities and relationships that have not been seen in the learning process of existing knowledge graphs as knowledge without re-learning.

Even if it is an effect not explicitly mentioned herein, the effects described in the following specification expected by the technical features of the present invention and their potential effects are treated as if they were described in the specification of the present invention.

1 is a block diagram illustrating an apparatus for completing a zero-shot knowledge graph according to an embodiment of the present invention.
2 is a flowchart illustrating a zero-shot knowledge graph completion method according to another embodiment of the present invention.
3 is a diagram illustrating each operation of a zero-shot knowledge graph completion method according to another embodiment of the present invention.
4 is a diagram illustrating a knowledge graph processed by a zero-shot knowledge graph completion method according to another embodiment of the present invention.
5 is a diagram illustrating an operation of extracting specific relational characteristic information by a zero-shot knowledge graph completion method according to another embodiment of the present invention.
6 is a diagram illustrating an operation of performing word embedding in a zero-shot knowledge graph completion method according to another embodiment of the present invention.
7 is a diagram illustrating an operation of performing an attention-based convolution in a zero-shot knowledge graph completion method according to another embodiment of the present invention.
8 is a diagram illustrating an operation of performing type matching in a zero-shot knowledge graph completion method according to another embodiment of the present invention.
9 is a diagram illustrating an operation in which a zero-shot knowledge graph completion method according to another embodiment of the present invention generates comprehensive feature information from specific relationship feature information in a multi-hop neighbor relationship.
10 is a diagram illustrating an operation of generating comprehensive feature information by applying self-attention in a zero-shot knowledge graph completion method according to another embodiment of the present invention.

Hereinafter, in the description of the present invention, if it is determined that the subject matter of the present invention may be unnecessarily obscured as it is obvious to those skilled in the art with respect to related known functions, the detailed description thereof will be omitted, and some embodiments of the present invention will be described. It will be described in detail with reference to exemplary drawings.

1 is a block diagram illustrating an apparatus for completing a zero-shot knowledge graph according to an embodiment of the present invention.

The zero-shot knowledge graph completion device 110 includes at least one processor 120 , a computer-readable storage medium 130 , and a communication bus 170 .

The processor 120 may control to operate as the zero-shot knowledge graph completion device 110 . For example, the processor 120 may execute one or more programs stored in the computer-readable storage medium 130 . The one or more programs may include one or more computer-executable instructions, which, when executed by the processor 120 , cause the zero-shot knowledge graph completion apparatus 110 to perform operations according to the exemplary embodiment. can be configured to

Computer-readable storage medium 130 is configured to store computer-executable instructions or program code, program data, and/or other suitable form of information. The program 140 stored in the computer-readable storage medium 130 includes a set of instructions executable by the processor 120 . In one embodiment, the computer-readable storage medium 130 includes memory (volatile memory, such as random access memory, non-volatile memory, or a suitable combination thereof), one or more magnetic disk storage devices, optical disk storage devices, It may be flash memory devices, other types of storage media that can be accessed by the knowledge graph completion apparatus 110 and store desired information, or a suitable combination thereof.

The communication bus 170 interconnects various other components of the zero-shot knowledge graph completion apparatus 110 including the processor 120 and the computer-readable storage medium 140 .

The zero-shot knowledge graph completion device 110 may also include one or more input/output interfaces 150 and one or more communication interfaces 160 that provide interfaces for one or more input/output devices. The input/output interface 150 and the communication interface 160 are connected to the communication bus 170 . The input/output device may be connected to other components of the zero-shot knowledge graph completion device 110 through the input/output interface 150 .

The zero-shot knowledge graph completion device 110 relates to inductive expression learning supporting zero-shot triple inference, and the entity description and relationship name shared between the entity-relationship in the knowledge graph and the entity-relationship of the new triple By generating an inductive vector expression for new entities and relationships by using additional shared information including constituting words, the knowledge expression for new entities and relationships that have not been seen in the learning process of the existing knowledge graph can be used as knowledge without re-learning. Plausibility), generalize evolving entities and relationships, and extend zero-shot knowledge graphs.

A knowledge graph is a data model that structures real-world facts related to entities and organically connects them for machine understanding. Each fact in the knowledge graph is generally expressed as a relation <h,r,t>, which is an edge with directionality, and a head entity h and a tail entity t, which are nodes, are semantic which is an edge type. They are connected to each other through a Semantic Relation.

In knowledge graph completion that performs link prediction and triple classification, Knowledge Graph Embedding (KGE) transforms structural information at the entity-relationship level into task-based latent features expressed as low-dimensional vectors.

Link prediction is in fact given two elements in <h,r,t> (e.g. <?,r,t>, <h,?,t>, <h,r,?>) for ? A validity score is measured by evaluating candidate facts for entity or relationship candidates (eg, h, r, t). Determine the ? that produces the highest plausibility score.

Triple classification checks whether facts not seen in the training data set are valid or not based on a specific relational threshold learned in advance. Knowledge graph completion performs learning while maximizing the overall validity score of the facts present in the knowledge graph. Based on the validity score of a given fact, the knowledge graph completion task can refine the incomplete and noisy knowledge graph.

In order to solve the problem that the existing knowledge graph technology can learn only for the static knowledge graph in which the entity and the relationship set are fixed, the zero-shot knowledge graph completion apparatus according to the present embodiment is provided between the existing knowledge graph and the new (unseen) triple. From the shared side information, we generalize a triple containing new entities and relationships, and inductively generate an embedding vector representation.

Considering that the relationship in the triple includes additional information constraining the types of the head entity and the tail entity, the zero-shot knowledge graph completion device utilizes the entity type constraint and relationship name, which are additional information of the relationship, to describe and type the entity. , extract the characteristic information of the triple from the name. In order to generate an embedding vector expression with a sophisticated level of judgment, in the process of collecting feature information of a neighboring triple, it is collected while minimizing unnecessary information according to the importance of a neighboring triple. That is, the attention mechanism is applied to the neighboring triples of the graph, and the neighbor information is propagated based on the type constraint of the relationship to minimize unnecessary information collection.

2 is a flowchart illustrating a zero-shot knowledge graph completion method according to another embodiment of the present invention, FIG. 3 is a diagram illustrating each operation of the zero-shot knowledge graph completion method, and FIG. 4 is a zero-shot knowledge graph completion method It is a figure which exemplifies the knowledge graph to be processed. The zero-shot knowledge graph completion method may be performed by a computing device or a zero-shot knowledge graph completion apparatus.

In step S210, the processor obtains a knowledge graph. The knowledge graph may be defined to satisfy a Resource Description Framework (RDF) standard. 3A illustrates an acquired knowledge graph.

In step S220, the processor extracts specific relational characteristic information by using the additional shared information shared in the knowledge graph. In step S220, the processor extracts specific relationship characteristic information using (i) relationship information including relationship name and type constraints and (ii) additional shared information of entity information including entity description and type. 3B illustrates specific relational characteristic information extracted from the knowledge graph.

In step S230, the processor generates comprehensive feature information about the zero-shot triple from specific relationship feature information in a multi-hop neighbor relationship in the knowledge graph. 3( c ) exemplifies comprehensive feature information generated based on the multi-hop neighbor relationship.

In step S240, the processor calculates a validity score of the comprehensive feature information. 3D illustrates a score calculated through a model based on a fully connected layer.

Referring to the knowledge graph of FIG. 4 , a head entity (ex. Harvard_University) and a tail entity (ex. Massachusetts) have various additional information such as description and a plurality of types. The relationship name (ex. State) and type constraint (ex. Thing, PopulatedPlace) actually provides information for obtaining the associated description for <Harvard_University, State, Massachusetts>. The zero-shot knowledge graph completion method extracts specific relational characteristic information from additional shared information shared between facts of the knowledge graph. In the zero-shot scenario, it is assumed that the target fact <h, r, t> includes an entity-relationship not belonging to the existing knowledge graph and at least one entity belonging to the existing knowledge graph. In fact, vectors are generated inductively from feature information.

Hereinafter, an operation of generating specific relational characteristic information will be described with reference to FIGS. 5 to 8 .

5 is a diagram illustrating an operation of extracting specific relational characteristic information by a zero-shot knowledge graph completion method.

Since most knowledge graph completion models have unique vector representations for learning, it is not easy to create new embeddings for new entity-relationships.

The processor extracts specific relation characteristic information from the head entity-relation vector and the tail entity-relation vector for each triple through word embedding, attention-based convolution, and type matching.

6 is a flowchart illustrating an operation in which the zero-shot knowledge graph completion method performs word embedding.

The zero-shot knowledge graph completion method converts words into symbolic representations through word embeddings. Since most relationship names do not exist in the lexicon, we encode the relationship names at the word level shared by the same lexical set.

In step S610, the processor divides the relationship name into a set of words shared with the entity description. In fact, a relationship and a correlation between two entities can be obtained from the name and description. Relational modeling at the word level processes relationships that do not belong to the existing knowledge graph from correlation in order to predict the validity of facts including relationships that do not belong to the existing knowledge graph.

In step S620, the processor sequentially performs stemming for each word focusing on the abbreviation of the word instead of the part-of-speech analysis (lemmaitzation).

In step S630, the processor encodes the word into a word level shared by the same vocabulary set. The entity-related information includes an entity description and a type, and a word set included in the entity description and a word set included in the plurality of types represent symbolic expressions. The relationship information includes a relationship name and type constraint, and the type constraint includes a domain and a range. Domain and Range are defined by the resource description framework.

의 d 차원 벡터 표현으로 인코딩된다. |w|는 어휘 집합의 크기이다.Each word has an embedding matrix

is encoded as a d-dimensional vector representation of |w| is the size of the vocabulary set.

7 is a flowchart illustrating an operation in which the zero-shot knowledge graph completion method performs attention-based convolution.

The zero-shot knowledge graph completion method generates specific relational characteristic information through an attention-based convolution model.

A convolution model includes a layer that extracts feature vectors by applying a convolution filter. The layer may include a convolution layer for extracting features and a max pooling layer for performing subsampling by selecting representative features. Convolutional models may share convolutional filters, layers may include parameters, and parameters of a layer include a set of learnable filters. The parameter may include a weight (ω) and/or a bias (b) between nodes.

In step S710, the processor transforms the word vector of the entity description into the entity description matrix through the embedding lookup, the word vector of the relationship name into the relationship name matrix, and the word vector of the type constraint into the type constraint matrix. The embedding lookup table is used by looking up data in the form of a list according to the index it receives.

In step S720, the processor applies an attention weight to the entity description matrix based on the relationship name matrix and the type constraint matrix. The processor combines the relationship name matrix and the type constraint matrix and passes it to the attention-based convolution model. The attention matrix is expressed as Equation (1).

A is the attention matrix, and W is the weight matrix. D' _h is the convolution result.

The extracted specific relational feature vector is expressed as Equation (2). The processor combines the specific relational feature vector for the head entity-relationship and the specific relational feature vector for the tail entity-relationship.

8 is a flowchart illustrating an operation in which the zero-shot knowledge graph completion method performs type matching.

The zero-shot knowledge graph completion method determines the validity of the fact expression for the head entity-relation vector and the tail entity-relation vector through type matching. Types are used to narrow the scope of validity.

In step S810, the processor converts the types of the domain and the head entity into a hot vector among the type constraints through the mask and calculates the compared head entity-relationship validity (Validity), and the range and Convert the type of the tail entity to a hot vector and compute the compared tail entity-relationship validity.

In step S810, the processor applies the head entity-relationship validity and the tail entity-relationship validity to the specific relationship characteristic information. Validity can be expressed as a value of 0 or 1. Specific relational characteristic information to which validity is applied is expressed as in Equation (3).

Hereinafter, an operation of generating comprehensive feature information will be described with reference to FIGS. 9 and 10 .

9 is a diagram illustrating an operation in which the zero-shot knowledge graph completion method generates comprehensive feature information from specific relationship feature information in a multi-hop neighbor relationship, and FIG. 10 is a zero-shot knowledge graph completion method synthesized by applying self-attention. It is a diagram illustrating an operation of generating characteristic information.

Due to the arbitrary structure of the knowledge graph, the existing knowledge graph completion model uses a local neighbor structure such as a one-hop neighbor or a relationship path. The zero-shot knowledge graph completion method can add global structural information to entity-relationship embeddings.

The processor generates comprehensive feature information by accumulating specific relationship feature information through a propagation path according to a multi-hop neighbor relationship. When searching for a multi-hop neighbor, a depth search is performed according to the maximum depth K. A validity score for the comprehensive feature information may be calculated.

The processor applies self-attention to specific relational characteristic information in order to collect specific relational characteristic information of a neighbor with high importance.

를 획득한다. N _fhrti 는 사실 f _hrti 의 다양한 크기를 갖는 자식 사실 세트이다. f _hrtj 는 사실 f _hrti 의 하나의 자식 사실이다. N_hrti는 특정 관계 사실 벡터 집합에 관한 매트릭스 표현이다. For a fact f _hrti of depth K, the collection function AGGREGATE _k is a relational fact vector of fact f _{hrti .}

to acquire N _fhrti is in fact a set of child facts of varying sizes of f _{hrti .} f _hrtj is actually a child fact of f _{hrti .} N _hrti is a matrix representation of a particular set of relational fact vectors.

Attention Attention mechanism to calculate the weighted using the vector N _fhrti fact that a particular relationship of the child with the parent N _fhrti facts Facts vector f _hrti. The relationship between each child fact and the parent fact is expressed as Equation (4).

의 합을 1로 만드는 소프트맥스 함수를 통해 N _fhrti 의 어텐션 점수로부터 획득된다. 어텐션 점수는 수학식 5와 같이 표현된다.The attention weight a _ij is the child fact f _hrtj coupled to the child facts _{N fhrti} while taking the parent fact f _{hrti into account.} indicates the importance of

It is obtained from the attention score of N _fhrti through the softmax function that makes the sum of . The attention score is expressed as in Equation 5.

W _a and w _a are parameters of the acquisition function that share the same depth K.

는 어텐션 가중치와 결합된다.specific relation fact vector

is combined with the attention weight.

c _fhrti is an associative vector representing the contextual information of the parent fact f _{hrti .} As shown in Equation 7, the context information c _fhrti is combined with the parent fact vector f _hrti.

는 f _hrti 의 문맥 기반 특정 관계 사실 벡터이다. 종합 특징 벡터가 생성될 때까지 깊이 탐색을 수행하고, 경파 경로를 따라 특정 관계 사실 벡터 f _hrti 를 문맥 기반 특정 관계 사실 벡터

로 교체한다.

is the context-specific relational fact vector of f _{hrti .} A depth search is performed until a synthetic feature vector is generated, and the specific relational fact vector f _hrti along the hard wave path is drawn into the context specific relational fact vector.

replace with

The processor calculates a plausibility score. The processor evaluates whether a certain relationship threshold for zero-shot triples is satisfied through a score model based on a fully connected layer.

The knowledge graph completion method scores the given <h, r, t> facts to check the validity of the facts. Evaluate the fact <h, r, t> as a scoring function for each of the candidate entities h and t.

The zero-shot knowledge graph completion apparatus may be implemented in a logic circuit by hardware, firmware, software, or a combination thereof, or may be implemented using a general-purpose or special-purpose computer. The device may be implemented using a hardwired device, a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), or the like. In addition, the device may be implemented as a system on chip (SoC) including one or more processors and controllers.

The zero-shot knowledge graph completion apparatus may be mounted in the form of software, hardware, or a combination thereof on a computing device or server provided with hardware elements. A computing device or server is all or part of a communication device such as a communication modem for performing communication with various devices or a wired/wireless communication network, a memory for storing data for executing a program, and a microprocessor for executing operations and commands by executing the program It can mean a variety of devices, including

In Figures 2, 6 to 8, it is described as executing each process sequentially, but this is only an exemplary description, and if you are skilled in the art, in the range that does not depart from the essential characteristics of the embodiment of the present invention, Figure 2, It will be possible to apply various modifications and variations by changing the order described in FIGS. 6 to 8 , or executing one or more processes in parallel or adding other processes.

The operations according to the present embodiments may be implemented in the form of program instructions that can be performed through various computer means and recorded in a computer-readable medium. Computer-readable medium represents any medium that participates in providing instructions to a processor for execution. Computer-readable media may include program instructions, data files, data structures, or a combination thereof. For example, there may be a magnetic medium, an optical recording medium, a memory, and the like. A computer program may be distributed over a networked computer system to store and execute computer readable code in a distributed manner. Functional programs, codes, and code segments for implementing the present embodiment may be easily inferred by programmers in the technical field to which the present embodiment pertains.

The present embodiments are for explaining the technical idea of the present embodiment, and the scope of the technical idea of the present embodiment is not limited by these embodiments. The protection scope of this embodiment should be interpreted by the following claims, and all technical ideas within the equivalent range should be interpreted as being included in the scope of the present embodiment.

Claims (14)

In the zero-shot knowledge graph completion method by a computing device,
extracting specific relational characteristic information using additional shared information shared in the knowledge graph;
generating comprehensive feature information on a zero-shot triple from the specific relationship feature information in a multi-hop neighbor relationship in the knowledge graph; and
Comprising the step of calculating the validity score of the comprehensive feature information,
The calculating of the validity score comprises evaluating whether or not a specific relation threshold for the zero-shot triple is satisfied through a score model based on a fully connected layer. According to claim 1,
The step of extracting the specific relationship characteristic information,
Zero-shot knowledge graph, characterized in that the specific relationship characteristic information is extracted using (i) relationship information including relationship name and type constraints and (ii) additional shared information of entity information including entity description and type How to complete. 3. The method of claim 2,
The step of extracting the specific relationship characteristic information,
A zero-shot knowledge graph completion method, characterized in that the specific relation feature information is extracted from the head entity-relation vector and the tail entity-relation vector for each triple through word embedding, attention-based convolution, and type matching. 4. The method of claim 3,
The word embedding is
dividing the relationship name into a set of words shared with the entity description, sequentially morphological analysis centering on an abbreviation instead of a part-of-speech for each word, and encoding the word at the word level shared by the same vocabulary set, A zero-shot knowledge graph completion method, characterized in that the word is converted into a symbolic expression. 4. The method of claim 3,
The attention-based convolution is
transform the word vector of the entity description into an entity description matrix through an embedding lookup, transform the word vector of the relationship name into a relationship name matrix, transform the word vector of the type constraint into a type constraint matrix;
and applying an attention weight to the entity description matrix based on the relationship name matrix and the type constraint matrix. 4. The method of claim 3,
The type matching is
Through the mask, the type of the domain and the head entity among the type constraints is converted into a hot vector and the comparison head entity-relationship validity is calculated, and the type of the range and the tail entity among the type constraints is hot Transform to a vector and compute the compared tail entity-relationship validity,
and applying the head entity-relationship validity and the tail entity-relationship validity to the specific relationship characteristic information. According to claim 1,
The step of generating the comprehensive feature information comprises:
The zero-shot knowledge graph completion method, characterized in that the comprehensive characteristic information is generated by accumulating the specific relational characteristic information through a propagation path according to the multi-hop neighbor relation. According to claim 1,
The step of generating the comprehensive feature information comprises:
A method of completing a zero-shot knowledge graph, characterized in that self-attention is applied to the specific relational characteristic information in order to collect specific relational characteristic information of a neighbor with high importance. delete A zero-shot knowledge graph completion apparatus comprising one or more processors and a memory for storing one or more programs executed by the one or more processors,
The processor extracts specific relational characteristic information using additional shared information shared in the knowledge graph,
The processor generates comprehensive feature information about a zero-shot triple from the specific relationship feature information in a multi-hop neighbor relationship in the knowledge graph,
The processor calculates a validity score of the comprehensive feature information, and calculating the validity score evaluates whether a specific relation threshold for the zero-shot triple is satisfied through a score model based on a fully connected layer Zero-shot knowledge graph completion device. 11. The method of claim 10,
The processor is
Zero-shot knowledge graph, characterized in that the specific relationship characteristic information is extracted using (i) relationship information including relationship name and type constraints and (ii) additional shared information of entity information including entity description and type complete device. delete 11. The method of claim 10,
The processor is
A zero-shot knowledge graph completion device, characterized in that the specific relation feature information is extracted from the head entity-relation vector and the tail entity-relation vector for each triple through word embedding, attention-based convolution, and type matching. 11. The method of claim 10,
The processor is
and generating the comprehensive feature information by accumulating the specific relationship feature information through a propagation path according to the multi-hop neighbor relationship.

Images