WO2020020085A1 - Representation learning method and device - Google Patents

Abstract

A representation learning method and device, relating to the technical field of big data, and used for solving the problem of current representation learning methods being unable to portray deep semantic information in a knowledge graph. The method comprises: determining, according to the type of entity in a triplet of a knowledge graph which fuses texts, a type representation vector of the entity (S202); determining, according to the type of a relationship in the triplet, a type representation vector of the relationship (S203); determining a context representation vector of the entity according to text information of the entity (S204); determining a context representation vector of the relationship according to a weight value of the relationship (S205); constructing a scoring function of the triplet according to the type representation vector of the entity, the context representation vector of the entity, the type representation vector of the relationship and the context representation vector of the relationship (S206); constructing a target function according to the scoring function of the triplet (S207); and minimizing the target function, and learning the representation vector of the entity and the representation vector of the relationship (S208).

Description

Representation learning method and device

This application claims priority from a Chinese patent application filed with the State Intellectual Property Office on July 24, 2018, with application number 201810822334.5, and the invention name is "indicating learning methods and devices", the entire contents of which are incorporated herein by reference.

Technical field

The present application relates to the field of big data technology, and in particular, to a method and a device for representation learning.

Background technique

Knowledge Graph describes structured concepts, entities and their relationships in a structured way, expressing information on the Internet into a form closer to the human cognitive world, providing a better organization and management And the ability to understand the vast amount of information on the Internet. With the development and application of artificial intelligence technology, knowledge map has gradually become one of the key technologies, and has been widely used in intelligent search, intelligent question answering, personalized recommendation, content distribution and other fields.

Because the entities, concepts, and relationships in the knowledge graph use discrete symbolic representations, these discrete symbolic representations are difficult to apply directly to calculations or inferences. Therefore, in order to effectively use the symbolized knowledge in the knowledge graph, researchers have proposed a representation learning method for the knowledge graph. Knowledge graph representation learning method aims to represent the entities and relationships in the knowledge graph as vectors in a low-dimensional vector space, thereby transforming the calculation between entities and relationships into numerical calculations between vectors.

Current representation learning methods cannot describe the deep-seated semantic information in the knowledge graph, such as "company A-swallow-company B" and "python-swallow-rabbit". The two "swallow" have different meanings, but The current representation learning will be represented as the same vector, resulting in errors.

Summary of the Invention

The present application provides a representation learning method and device, which are used to characterize deep-level semantic information in a knowledge map and improve the accuracy of representation learning.

In order to achieve the above purpose, this application uses the following technical solutions:

According to a first aspect, a representation learning method is provided, including: determining a type representation vector of an entity according to a type of an entity in a triple of a knowledge graph of a fused text, the entity including a head entity and a tail entity; and according to a relationship in the triple The type of the relationship determines the type representation vector of the relationship; the context representation vector of the entity is determined based on the textual information of the entity; the context representation vector of the relationship is determined based on the weight value of the relationship; the vector is represented by the type of the entity, the vector of the context representation of the entity, and the relationship The type represents the vector and the context representation vector of the relationship to construct the scoring function of the triplet; the objective function is constructed based on the scoring function of the triplet; the objective function is minimized to learn the representation vector of the entity and the representation vector of the relationship. Based on this technical solution, because the type and context of the entity, the type and weight value of the relationship represent a certain level of deep semantic information, the type and weight of the entity are considered by considering the type and context of the entity, as well as the type and weight value of the relationship. Representation vectors and relational representation vectors can depict deep-level semantic information in the knowledge map and improve the accuracy of representation learning.

In a possible design, before determining the type representation vector of the entity according to the type of the entity in the triple of the knowledge graph of the fused text, the method further includes: initializing the representation vector of the head entity, the representation vector of the tail entity, and the relationship Representation vector, entity type representation matrix, relation type representation matrix, word representation matrix, and weight vector. In this way, in the subsequent representation learning process, the type representation vector of the relationship, the type representation vector of the entity, the context representation vector of the relationship, and the context representation vector of the entity can be determined.

In a possible design, determining the type representation vector of the entity according to the type of the entity in the triple of the knowledge graph of the fused text includes: determining the type identification vector of the head entity according to the type of the head entity;

Determine the type representation vector of the head entity; where f ₁ (h) represents the type representation vector of the head entity, W _etype represents the entity type representation matrix, and v _etype (h) represents the type identification vector of the head entity; according to the type of the tail entity, Determine the type identification vector of the tail entity; according to the formula

Determine the type of the tail entity to represent the vector; where f ₁ (t) represents the type of the tail entity represents the vector, and v _etype (t) represents the type identification vector of the tail entity.

In a possible design, determining the type of the relationship to represent the vector according to the type of the relationship in the triple, includes: determining the type identification vector of the relationship according to the type of the relationship; and according to the formula:

Determine the type of the relationship to represent the vector; where g ₁ (r) represents the type of the relationship represents the vector, W _rtype represents the relationship type represents the matrix, and v _rtype (r) represents the type identification vector of the relationship.

In a possible design, determining an entity's contextual representation vector according to the entity's text information includes: determining the words related to the head entity based on the text information of the head entity;

Determine the context representation vector of the head entity; where f ₂ (h) represents the context representation vector of the head entity, α, β are constants between 0 and 1, v _h represents the vector of representation of the head entity, and w _i represents Words related to the head entity, ε ₁ represents a set of all words related to the head entity, W _word represents a matrix of words, and V _vocubalary (w _i ) represents an identification vector of w _i ; according to the text information of the tail entity, determine the Tail entity related words; according to the formula:

Determine the context representation vector of the tail entity; where f ₂ (t) represents the context representation vector of the tail entity, v _t represents the vector of the tail entity, _mi represents the words related to the tail entity, and ε ₂ represents all the tail entity related. V _vocubalary (m _i ) represents the identity vector of _mi .

In a possible design, determining the contextual representation vector of the relationship according to the weight value of the relationship includes: according to the formula:

Determine the context representation vector of the relationship; where g ₂ (r) represents the context representation vector of the relationship, v _r represents the vector of the relationship, n _i represents the weight value of the relationship, and ε ₃ represents the set of weight values of the relationship.

A representation vector representing n _i .

Optionally, the scoring function for the triples is:

among them,

Represents compound operations, f ₁ (h) represents the type of the head entity represents a vector, g ₁ (r) represents the type of a relation represents a vector, f ₁ (t) represents the type of a tail entity represents a vector, and f ₂ (h) represents a head entity The context represents the vector of, g ₂ (r) represents the context of the relationship, and f ₂ (t) represents the context of the tail entity.

Optionally, the objective function is: L = ∑ _{(h, r, t) ∈ Δ} (∑ _{(h ′, r, t ′) ∈ Δ ′} max (0, S (h, r, t) + MS (h ′, R, t ′))). Among them, (h, r, t) represents a positive example triplet, Δ represents a positive example triplet set, (h ′, r, t ′) represents a negative example triplet, and h ′ represents a head entity of a negative example, t ′ represents the tail entity of the negative example, Δ ′ represents the triple set of negative examples, and M is a constant.

In a possible design, before determining the type representation vector of the entity according to the type of the entity in the triple of the knowledge map of the fused text, the method further includes: obtaining an initial knowledge map; constructing a fusion based on a framework of the initial knowledge map The framework of textual knowledge graphs; the framework of textual knowledge graphs defines at least the following: extended attributes of entities, extended attributes of relations, and extended relations between entities; extended attributes of entities include textual information of entities; according to the initial knowledge graph Obtain external data from the entity or relationship information in the external data; from the external data, determine the extended attribute value of the entity and the extended attribute value of the relationship to build a knowledge map of the fused text. Based on this technical solution, the server constructs a knowledge graph of the fused text by extending the framework of the initial knowledge graph and supplementing the extended attribute values of the related entities and the extended attribute values of the relationship. In this way, compared with the original knowledge map, the knowledge map of the fused text is more complete in content.

In a second aspect, a representation learning device is provided, including: a type representation module for determining a type representation vector of an entity according to a type of an entity in a triple of a knowledge graph of a fused text, and the entity includes a head entity and a tail entity; The type of the relationship in the triple, determines the type of the relationship to represent the vector. The context representation module is used to determine the context representation vector of the entity according to the text information of the entity; and determine the context representation vector of the relationship according to the weight value of the relationship. A processing module for constructing a scoring function of a triple based on the type representation vector of the entity, the context representation vector of the entity, the type representation vector of the relationship, and the context representation vector of the relationship; constructing an objective function based on the scoring function of the triple Minimize the objective function and learn the representation vector of the entity and the representation vector of the relationship.

In a possible design, the processing module is further used to initialize the representation vector of the head entity, the representation vector of the tail entity, the relationship vector, the entity type representation matrix, the relationship type representation matrix, the word representation matrix, and the weight vector. .

In a possible design, the type representation module is used to determine the type identification vector of the head entity according to the type of the head entity; according to the formula

Determine the type representation vector of the head entity; where f ₁ (h) represents the type representation vector of the head entity, W _etype represents the entity type representation matrix, and v _etype (h) represents the type identification vector of the head entity; according to the type of the tail entity, Determine the type identification vector of the tail entity; according to the formula

Determine the type of the tail entity to represent the vector; where f ₁ (t) represents the type of the tail entity represents the vector, and v _etype (t) represents the type identification vector of the tail entity.

In a possible design, the type representation module is used to determine the type identification vector of the relationship according to the type of the relationship; according to the formula:

Determine the type of the relationship to represent the vector; where g ₁ (r) represents the type of the relationship represents the vector, W _rtype represents the relationship type represents the matrix, and v _rtype (r) represents the type identification vector of the relationship.

In a possible design, the context representation module is used to determine the words related to the head entity based on the text information of the head entity; according to the formula:

Determine the context representation vector of the head entity; where f ₂ (h) represents the context representation vector of the head entity, α, β are constants between 0 and 1, v _h represents the vector of representation of the head entity, and w _i represents Words related to the head entity, ε ₁ represents a set of all words related to the head entity, W _word represents a matrix of words, and V _vocubalary (w _i ) represents an identification vector of w _i ; according to the text information of the tail entity, determine the Tail entity related words; according to the formula:

Determine the context representation vector of the tail entity; where f ₂ (t) represents the context representation vector of the tail entity, v _t represents the vector of the tail entity, _mi represents the words related to the tail entity, and ε ₂ represents all the tail entity related. V _vocubalary (m _i ) represents the identity vector of _mi .

In one possible design, the context representation module is used according to the formula:

Determine the context representation vector of the relationship; where g ₂ (r) represents the context representation vector of the relationship, v _r represents the vector of the relationship, n _i represents the weight value of the relationship, and ε ₃ represents the set of weight values of the relationship.

A representation vector representing n _i .

Optionally, the scoring function for the triples is:

among them,

Represents compound operations, f ₁ (h) represents the type of the head entity represents a vector, g ₁ (r) represents the type of a relation represents a vector, f ₁ (t) represents the type of a tail entity represents a vector, and f ₂ (h) represents a head entity The context represents the vector of, g ₂ (r) represents the context of the relationship, and f ₂ (t) represents the context of the tail entity.

Optionally, the objective function is: L = ∑ _{(h, r, t) ∈ Δ} (∑ _{(h ′, r, t ′) ∈ Δ ′} max (0, S (h, r, t) + MS (h ′, R, t ′))). Among them, (h, r, t) represents a positive example triplet, Δ represents a positive example triplet set, (h ′, r, t ′) represents a negative example triplet, and h ′ represents a head entity of a negative example, t ′ represents the tail entity of the negative example, Δ ′ represents the set of triples of the negative example, and M is a constant.

In a possible design, the representation learning device further includes: a framework extension module, a data acquisition module, and an extension mapping module. Framework extension module, used to obtain the initial knowledge map; based on the framework of the initial knowledge map, construct a framework of the text knowledge map; the framework of the text knowledge map defines at least the following: extended attributes of entities, extended attributes of relationships, and entities The extended relationship between them; the extended attributes of the entity include the textual information of the entity. A data acquisition module is configured to acquire external data according to entity information or relationship information in the initial knowledge map. The extended mapping module is used to determine the extended attribute value of the entity and the extended attribute value of the relationship from the external data to construct a knowledge map of the fused text.

According to a third aspect, a server is provided, including: a processor, a memory, a bus, and a communication interface; the memory is configured to store a computer to execute instructions, the processor is connected to the memory through the bus, and when the server runs When the processor executes the computer execution instructions stored in the memory, so that the server executes the representation learning method according to any one of the first aspects.

According to a fourth aspect, a computer-readable storage medium is provided. The computer-readable storage medium stores instructions that, when run on a computer, enable the computer to perform the representation learning according to any one of the first aspects. method.

According to a fifth aspect, a computer program product containing instructions is provided, which, when run on a computer, enables the computer to execute the representation learning method according to any one of the first aspects.

According to a sixth aspect, a chip system is provided. The chip system includes a processor for supporting a server to implement the functions involved in the first aspect. In a possible design, the chip system further includes a memory, and the memory is configured to store program instructions and data necessary for the server. The chip system can be composed of chips, and can also include chips and other discrete devices.

The technical effects brought by any one of the design methods in the second aspect to the sixth aspect may refer to the technical effects brought by the different design methods in the first aspect, and are not repeated here.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a schematic diagram of a communication system according to an embodiment of the present application; FIG.

2 is a schematic structural diagram of a server according to an embodiment of the present application;

3 is a flowchart of a method for constructing a knowledge map according to an embodiment of the present application;

4 is a schematic diagram of an initial knowledge map provided by an embodiment of the present application;

FIG. 5 is a schematic diagram of a fused text knowledge map according to an embodiment of the present application; FIG.

6 is a flowchart illustrating a learning method according to an embodiment of the present application;

FIG. 7 is a schematic structural diagram of a learning device according to an embodiment of the present application.

detailed description

Before introducing the method provided in the embodiments of the present application, the terms involved in the embodiments of the present application are briefly introduced.

The knowledge map is a symbolic expression of the objective world. The knowledge graph itself is a networked knowledge base where entities with attributes are connected through relationships. From the perspective of the graph, the knowledge graph is essentially a network, where nodes represent entities (or concepts) in the objective world, and edges represent various relationships or attributes of entities.

Among them, entity refers to specific things that are distinguishable and exist independently. For example, "apple", "banana", etc. can all be entities.

Concept refers to the conceptual representation of objective things that people form in the process of cognizing the world, such as people, animals, and plants. In other words, a concept can be understood as a collection of entities with the same characteristics.

Relations are used to describe objectively existing associations between entities and concepts. Exemplarily, the relationship between the entities may be an include relationship, a subordinate relationship, or the like. For example, a mobile phone contains a camera, that is, an inclusion relationship exists between the mobile phone and the camera.

A property is a characterization of an abstract aspect of an object. It is worth noting that an entity (or concept) generally has many properties and relationships. These properties and relationships can be referred to as the attributes of the entity (or concept). For example, if the entity is Beijing, the attributes of Beijing include population, area, and so on.

The attribute value is the value of the specified attribute of the object. For example, China's area is 9.6 million square kilometers, and 9.6 million square kilometers is the value of the area attribute.

A triple is a universal representation of a knowledge graph. The basic forms of triples include (first entity-relation-tail entity) and (concept-attribute-attribute value). For example, China-Capital-Beijing is an example of a triad (head entity-relation-tail entity), where China is the head entity, Beijing is the tail entity, and the capital is the relationship between China and Beijing. Beijing-population-2069.3 million constitutes an example of (concept-attribute-attribute value) triples, in which population is an attribute and 20.693,000 is the attribute value. It should be noted that in the embodiments of the present application, unless otherwise specified, the triples refer to the basic form of (first entity-relationship-tail entity).

The schema of the knowledge graph is a specification for modeling concepts, an abstract model describing the objective world, and a clear definition of concepts and their relationships in a formal way. Understandably, the schema defines the data model in the knowledge graph. Specifically, the schema defines the types of entities and the types of relationships.

FIG. 1 shows a communication system to which the technical solution provided in this application is applicable. The communication system includes a server 10 and a terminal device 20. The server 10 and the terminal device 20 communicate through a wireless network or a wired network.

The terminal device 20 may be a mobile phone, a tablet computer, a notebook computer, an ultra-mobile personal computer (UMPC), a netbook, a personal digital assistant (PDA), or the like. The terminal device 20 may be installed with a client having functions of intelligent search, intelligent question answering, and the like.

The server 10 is configured to provide services such as intelligent search, intelligent question answering, and the like for the terminal device. The server 10 includes a frame extension unit, a data acquisition unit, an extension mapping unit, a feature calculation unit, and a storage unit.

Wherein, the frame expansion unit is configured to construct a frame of a knowledge graph fused with text according to a frame of an initial knowledge graph.

The data obtaining unit is configured to obtain external data from the Internet according to entity information or relationship information in the initial knowledge map.

The extended mapping unit is configured to generate an extended attribute value of an entity and an extended attribute value of a relationship from external data; add the extended attribute value of the entity and the extended attribute value of the relationship to a knowledge map of the fused text, Construct a knowledge map of the fused text.

The feature calculation unit is configured to determine a representation vector of an entity and a representation vector of a relationship in a knowledge map of the fused text.

The storage unit is configured to store related data of the constructed knowledge map of the fused text.

FIG. 2 is a schematic diagram of a hardware structure of a server according to an embodiment of the present application. The server includes at least one processor 101, a communication line 102, a memory 103, and at least one communication interface 104.

The processor 101 may be a general-purpose central processing unit (CPU), a microprocessor, an application-specific integrated circuit (ASIC), or one or more processors for controlling the execution of the program of the solution of the present application. integrated circuit.

The communication line 102 may include a path for transmitting information between the aforementioned components.

The communication interface 104 uses any device such as a transceiver to communicate with other devices or communication networks, such as Ethernet, radio access network (RAN), wireless local area networks (WLAN), etc. .

The memory 103 may be a read-only memory (ROM) or other types of static storage devices that can store static information and instructions, a random access memory (random access memory, RAM), or other types that can store information and instructions The dynamic storage device can also be electrically erasable programmable read-only memory (EEPROM-ready-only memory (EEPROM)), compact disc (read-only memory (CD-ROM)) or other optical disk storage, optical disk storage (Including compact discs, laser discs, optical discs, digital versatile discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or can be used to carry or store desired program code in the form of instructions or data structures and can be used by a computer Any other media accessed, but not limited to this. The memory 103 may exist independently, and is connected to the processor through the communication line 102. The memory 103 may also be integrated with the processor 101.

The memory 103 is configured to store computer-executable instructions for executing the solution of the present application. The processor 101 is configured to execute computer-executable instructions stored in the memory 103, so as to implement the technical solution provided by the following embodiments of the present application.

Optionally, the computer-executable instructions in the embodiments of the present application may also be referred to as application program codes, which are not specifically limited in the embodiments of the present application.

In a specific implementation, as an embodiment, the processor 101 may include one or more CPUs, such as CPU0 and CPU1 in FIG. 2.

In a specific implementation, as an embodiment, the server may include multiple processors, such as the processor 101 and the processor 107 in FIG. 2. Each of these processors may be a single-CPU processor or a multi-CPU processor. A processor herein may refer to one or more devices, circuits, and / or processing cores for processing data (such as computer program instructions).

As shown in FIG. 3, a method for constructing a knowledge map provided by an embodiment of the present application includes the following steps:

S101. The server obtains an initial knowledge map.

The initial knowledge map is obtained by the server from the Internet; or, the initial knowledge map is manually entered into the server.

S102. Based on the framework of the initial knowledge map, the server constructs a framework of the text knowledge map.

Compared to the framework of the initial knowledge graph, the framework of the text-fused knowledge graph further includes at least: extended attributes of entities, extended attributes of relationships, and extended relations between entities.

The extended attributes of entities are the attributes of undefined entities in the framework of the initial knowledge graph. For example, for the Beijing entity, there are only two attributes defined in the framework of the initial knowledge map: area and population, while the framework of the fused knowledge map also defines another attribute: latitude and longitude. In this way, latitude and longitude are the extended attributes of Beijing. Optionally, the extended attributes of the entity may be determined according to the following manner: the server extracts the entity name information from the existing text information or external text information using a text mining algorithm according to the entity name information in the initial knowledge map Related high-frequency words, combined with part-of-speech filtering techniques, form extended attributes of entities. The text mining algorithm includes a topic model, core word extraction, and named entity recognition.

It should be noted that, in the embodiment of the present application, the extended attributes of the entity include at least: text information of the entity. That is, the knowledge graph of the fused text is a knowledge graph having at least the attribute of the text information of the entity.

The extended attributes of relationships are the attributes of undefined relationships in the framework of the initial knowledge graph. Exemplarily, the extended attributes of the relationship include: the type of the head entity, the type of the tail entity, and the like. Optionally, the extended attributes of the relationship are manually defined by an expert.

The extended relationship between entities is the relationship between undefined entities in the framework of the initial knowledge graph. Exemplarily, the extended relationship between entities includes a distance relationship, a proximity relationship, and the like. Optionally, the extended relationship may be obtained from a relational database, which includes relationships between various entities. It should be noted that in order to ensure the rationality of the extended relationship between entities, after determining the extended relationship between the entities, the server uses the relationship between the entities specified in the knowledge map of the current fused text as training data, and uses weak supervision Algorithm and reinforcement learning algorithm to verify the rationality of the extended relationship, thereby removing the unreasonable extended relationship.

S103. The server obtains external data according to entity information or relationship information in the knowledge map of the fused text.

The information of the entity includes a name of the entity. The information of the relationship includes the name of the relationship. The external data includes information of extended attributes of entities or relationships. The external data may be structured data, semi-structured data, or unstructured data. If the external data is unstructured data, the external data may be text information or multimedia information, and the multimedia information includes videos, pictures, and web pages.

In an optional implementation manner, the server obtains external data from the Internet by using a crawler or other technology according to entity information or relationship information in the fused text knowledge map.

Exemplarily, the server directly extracts external data from an encyclopedia website (such as Baidu Encyclopedia, Wikipedia) or a vertical website (such as an electronic product website, a book website, a movie website, or a music website). Because encyclopedia websites and vertical websites include a lot of entity attribute information, for example, the book website includes information such as the author, publisher, and publishing time of the book, the server can generate a certain wrapper (or template) by generating certain rules. Use a wrapper to extract external data that contains attribute information. It should be noted that the method of generating a wrapper can be divided into: a manual method (that is, writing a wrapper manually), a supervised method, a semi-supervised method, and an unsupervised method.

S104. The server determines an extended attribute value of the entity and an extended attribute value of the relationship from external data to construct a knowledge map of the fused text.

Optionally, if the external data is structured data or semi-structured data, the server extracts the extended attribute value of the entity or the extended attribute value of the relationship from the external data by using a manually defined or automatically generated matching mode.

Optionally, if the external data is unstructured data, for example, the external data is text information, the server uses data mining methods to mine the relationship mode between the attribute and the attribute value from the text information, so as to realize the attribute name and attribute. Positioning of the value in the text. It can be understood that in a real language environment, there are some keywords (such as attribute names) used to limit and define the meaning of the attribute value near many attribute values, so these keywords can be used to locate the attribute value.

After that, the server adds the extended attribute value of the entity and the extended attribute value of the relationship to the knowledge graph of the fused text to complete the creation of the knowledge graph of the fused text.

For example, FIG. 4 shows a schematic diagram of an initial knowledge map, and FIG. 5 shows a schematic diagram of a knowledge map that fuses text. As shown in Figure 4, the framework of the initial knowledge graph defines two entity types: products and components. The products are: Huawei P10 and Huawei P8, the parts are: camera, lens 1, lens 2 and lens. Lens 1, lens 2 and lens have two attributes: sensor and pixel. The knowledge map of the fused text shown in FIG. 5 is obtained based on the expansion of the initial knowledge map shown in FIG. 4. As shown in Figure 5, there is an extended relationship between Huawei P10 and Huawei P8: an orderly co-occurrence relationship. Huawei P10 has extended attributes: theme and frequency. Containment relationships have extended attributes: htype, hr frequency, ttype, and rt frequency. Among them, htype indicates the type of the head entity, hr frequency indicates the frequency of the head entity and the relationship, ttype indicates the type of the tail entity, and rt frequency indicates the frequency of the relationship and the tail entity. It should be noted that although not shown in FIG. 5, any entity in FIG. 5 also has an extended attribute of text information.

The method for constructing a knowledge graph provided in the embodiments of the present application constructs a knowledge graph of the fused text by extending the framework of the initial knowledge graph and supplementing the extended attribute values of the related entities and the extended attribute values of the relationship. In this way, compared with the original knowledge map, the knowledge map of the fused text is more complete in content.

After constructing the knowledge graph of the fused text, it is necessary to perform learning on the constructed knowledge graph of the fused text in order to effectively utilize the knowledge in the knowledge graph of the fused text. FIG. 6 shows a flowchart of a learning method according to an embodiment of the present application. The method includes the following steps:

S201. The server initializes the representation vector of the head entity, the representation vector of the tail entity, the representation vector of the relationship, the entity type representation matrix, the relationship type representation matrix, the word representation matrix, and the weight value in the triple of the knowledge graph of the fused text. Represents a vector.

Specifically, the server uses methods such as uniform distribution initialization and Bernoulli distribution initialization to initialize the representation vector of the head entity, the representation vector of the tail entity, and the relationship in the triple of the knowledge graph of the fused text. A representation vector, the entity type representation matrix, the relation type representation matrix, the word representation matrix, and a representation vector of the weight value.

It should be noted that the dimensions of the representation vector of the head entity, the dimensions of the representation vector of the tail entity, the dimensions of the relationship representation vector, and the weight value of the representation vector are all preset. Fixed. Furthermore, the dimensions of the representation vector of the head entity, the dimensions of the representation vector of the tail entity, the dimensions of the relationship representation vector, and the weight value of the representation vector are equal.

In the embodiment of the present application, the entity type representation matrix is used to map the type identification vector of the entity to the type representation vector of the entity. The type identification vector of the entity is used to directly characterize the type to which the entity belongs. The type representation vector of the entity is used to indirectly characterize the type to which the entity belongs.

It should be noted that the number of rows of the entity type representation matrix is equal to the number of dimensions of the entity type representation vector. The number of columns of the entity type representation matrix is equal to the number of dimensions of the type identification vector of the entity. Wherein, the dimension of the type representation vector of the entity is equal to the dimension of the representation vector of the entity. The dimension of the type identification vector of the entity is equal to the total number of types of the entity. Each dimension of the entity type identification vector corresponds to a type of the entity, and each dimension of the entity type identification vector has a value of 0 or 1. If the dimension of the entity type identification vector is 1, it indicates that the entity belongs to the type corresponding to the dimension. If one dimension of the entity's type identification vector is 0, it indicates that the entity does not belong to the type corresponding to the dimension. Exemplarily, the framework of the knowledge graph defines the types of entities including type 1, type 2, type 3, and type 4. If the type to which entity A belongs is type 1 and type 4, the type identification vector of entity A is (1,0 , 0,1).

In the embodiment of the present application, the relationship type representation matrix is used to map the type identification vector of the relationship to the type representation vector of the relationship. The type identification vector of the relationship is used to directly characterize the type to which the relationship belongs. The type representation vector of the relationship is used to indirectly characterize the type to which the relationship belongs.

It should be noted that the number of rows of the relationship type representation matrix is equal to the number of dimensions of the relationship type representation vector. The relationship type indicates that the number of columns of the matrix is equal to the dimension of the type identification vector of the relationship. The dimension of the type-representation vector of the relation is equal to the dimension of the relation-representation vector. The dimension of the type identification vector of the relationship is equal to the total number of types of the relationship. Each dimension of the type identification vector of the relationship corresponds to a type of the entity, and each dimension of the type identification vector of the entity has a value of 0 or 1. If the value of the dimension identification vector of the relationship is 1, the relationship belongs to the corresponding type of the dimension; if the value of the dimension identification vector of the relationship is 0, the relationship does not belong to the corresponding type of the dimension.

In the embodiment of the present application, the word representation matrix is used to map an identification vector of a word to a representation vector of a word. The identification vector of the word is used to directly characterize the position of the word in the vocabulary. The word type representation vector is used to indirectly characterize the position of the word in the vocabulary. The vocabulary contains all entity-related words in the knowledge graph.

It should be noted that the number of rows of the word representation matrix is equal to the number of dimensions of the word representation vector. The number of columns of the word representation matrix is equal to the dimension of the identification vector of the word. The dimension of the representation vector of the word is equal to the dimension of the representation vector of the entity. The dimension of the identification vector of the word is equal to the total number of words in the vocabulary. Each dimension of the word's identification vector corresponds to a position in the vocabulary. The value of the word's identification vector is 0 or 1. If a dimension of the word's identification vector is 0, it means that the word is not in the position of the vocabulary corresponding to that dimension; if a dimension of the word's identification vector is 1, it means that the word's corresponding vocabulary in that dimension position.

In the embodiment of the present application, the representation vector of the weight value is used to represent the weight value of the relationship. The weight value of the relationship is used to explain the degree of correlation between the two entities to which the relationship is connected.

S202. The server determines a type representation vector of the entity according to the type of the entity in the triple of the knowledge map of the fused text.

The entities include a head entity and a tail entity.

In the embodiment of the present application, the type of the entity is defined by a framework of a knowledge graph of fused text. And, the type of the entity is not unique. In other words, an entity can correspond to multiple types. Exemplarily, it is assumed that the entity is a mobile phone, and the mobile phone may be an electronic product or a communication tool. Here, the electronic product or communication tool is the type to which the mobile phone belongs.

Specifically, the server first determines the type identification vector of the head entity based on the type of the head entity; then, the server then determines the vector based on the formula

A type representation vector of the head entity is determined. Among them, f ₁ (h) represents a type representation vector of the head entity, W _etype represents the entity type representation matrix, v _etype (h) represents a type identification vector of the head entity, and ‖ ‖ represents a second norm.

Specifically, the server first determines the type identification vector of the tail entity according to the type of the tail entity; then, the server then determines the vector based on the formula

A type representation vector of the tail entity is determined. Wherein, f ₁ (t) represents a type representation vector of the tail entity, and v _etype (t) represents a type identification vector of the tail entity.

S203. The server determines a type representation vector of the relationship according to the type of the relationship in the triple.

Wherein, the type of the relationship is defined by a framework of a knowledge graph of fused text. The type of the relationship includes an include relationship, a subordinate relationship, a side-by-side relationship, and the like.

Specifically, the server may first determine the type identification vector of the relationship according to the type of the relationship. The server then according to the formula:

Determine the type representation vector of the relationship; where g ₁ (r) represents the type representation vector of the relationship, W _rtype represents the relationship type representation matrix, and v _rtype (r) represents the type identification vector of the relationship.

S204. The server determines a context representation vector of the entity according to the text information of the entity.

The context representation vector of the entity is used to characterize the context features of the entity. The text information is stored in advance by the server.

Specifically, the server determines words related to the head entity according to the text information of the head entity; then, the server according to the formula:

Determine the context representation vector of the head entity; where f ₂ (h) represents the context representation vector of the head entity, α, β are constants between 0 and 1, and v _h represents the head entity's Represents a vector, w _i represents a word related to the head entity, ε ₁ represents a set of all words related to the head entity, W _word represents a matrix of words, and V _vocubalary (w _i ) represents an identification vector of w _i .

Specifically, the server determines words related to the tail entity according to the text information of the tail entity; then, the server according to the formula:

Determining a context representation vector of the tail entity; wherein f ₂ (t) represents a context representation vector of the tail entity, v _t represents a representation vector of the tail entity, and _mi represents a word related to the tail entity, ε ₂ represents a set of all words related to the tail entity, and V _vocubalary (m _i ) represents an identification vector of _mi .

It should be noted that the server may determine the words related to the entity in the following implementation manner: the server selects a text sequence within a certain distance from the entity name from the text information, and uses word segmentation technology to divide the text sequence into individual Words, these individual words are words related to the entity.

Optionally, the above word segmentation technology may be a word segmentation technology based on string matching, a word segmentation technology based on understanding, and a word segmentation technology based on statistics. For specific implementations of these word segmentation technologies, reference may be made to the prior art, which is not described in the embodiments of the present application.

S205. The server determines a context representation vector of the relationship according to a weight value of the relationship.

Wherein, the context representation vector of the relationship is used to represent the weight feature of the relationship.

Specifically, the server uses the formula:

Determine the contextual representation vector of the relationship. Among them, g ₂ (r) represents a context representation vector of the relationship, v _r represents a representation vector of the relationship, n _i represents a weight value of the relationship, and ε ₃ represents a set of ownership weight values of the relationship.

A representation vector representing n _i .

S206. The server constructs a scoring function for the triplet according to a type representation vector of the entity, a context representation vector of the entity, a type representation vector of the relationship, and a context representation vector of the relationship.

The scoring function of the triples is:

among them,

Represents a compound operation. It should be noted that the compound operation includes vector addition or bitwise multiplication.

Bitwise multiplication refers to multiplying the value of each dimension of the first vector by the value of the corresponding dimension of the second vector to generate the value of the corresponding dimension of the third vector. Exemplary,

S207. The server constructs an objective function according to the scoring function of the triple.

The objective function is:

Among them, (h, r, t) represents a positive example triplet, Δ represents a positive example triplet set, (h ′, r, t ′) represents a negative example triplet, h ′ represents a negative example head entity, and t ′ Represents a negative example tail entity, Δ ′ represents a negative example triple set, and M is a constant.

It should be noted that the positive triples are triples that exist in the knowledge graph of the fused text, and the negative triples are triples that do not exist in the knowledge graph of the fused text. The negative triples are obtained by randomly replacing the head or tail entities with the positive triples. The negative triples are not included in the negative triples.

In addition, for the method for generating S (h ′, r, t ′), reference may be made to the above-mentioned method for generating S (h, r, t), which is not described in the embodiment of the present application.

S208. The server minimizes the objective function and learns a representation vector of the entity and a representation vector of the relationship.

In an optional implementation manner, the server uses a gradient descent algorithm to iteratively update the representation vector of the head entity, the representation vector of the tail entity, the representation vector of the relationship, the entity type representation matrix, and the relationship. The type representation matrix, the word representation matrix, the representation vector of weight values, the representation vector of negative example head entities, and the representation vector of negative example tail entities, so that the objective function can solve the minimum value, and then the head entity can be determined. A representation vector of, a representation vector of the tail entity, and a representation vector of the relationship.

It should be noted that, for the gradient descent algorithm, reference may be made to the prior art, which is not described in the embodiment of the present application.

In the representation learning method provided in the embodiments of the present application, since the type and context of the entity, the type and weight value of the relationship represent certain deep-level semantic information, therefore, by considering the type and context of the entity and the type and weight value of the relationship, The identified vector of the entity and the vector of the relation can describe the deep-level semantic information in the knowledge map, and improve the accuracy of the representation learning.

The above mainly introduces the solution provided by the embodiment of the present application from the perspective of the server. It can be understood that, in order to implement the above functions, the server includes a hardware structure and / or a software module corresponding to each function. Those skilled in the art should easily realize that, with reference to the units and algorithm steps of each example described in the embodiments disclosed herein, this application can be implemented in the form of hardware or a combination of hardware and computer software. Whether a certain function is performed by hardware or computer software-driven hardware depends on the specific application of the technical solution and design constraints. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

In the embodiment of the present application, the server may be divided according to the foregoing method example. For example, each module or unit may be divided corresponding to each function, or two or more functions may be integrated into one processing module. The above integrated modules may be implemented in the form of hardware, or in the form of software modules or units. The division of modules or units in the embodiments of the present application is schematic, and is only a logical function division. In actual implementation, there may be another division manner.

For example, in a case where each functional module is divided corresponding to each function, FIG. 7 illustrates a possible structural diagram of a learning device according to the foregoing embodiment. As shown in FIG. 7, the representation learning device includes a type representation module 701, a context representation module 702, a processing module 703, a framework extension module 704, a data acquisition module 705, and an extension mapping module 706. The type indicating module 701 is configured to support the server to perform steps S202 and S203 in FIG. 6. The context representation module 702 is configured to support the server to perform steps S204 and S205 in FIG. 6. The processing module 703 is configured to support the server to execute steps S201, S206, S207, and S208 in FIG. The frame expansion module 704 is configured to support the server to perform steps S101 and S102 in FIG. 3. The data acquisition module 705 is configured to support the server to execute step S103 in FIG. 3. The extended mapping module 706 is configured to support the server to execute step S104 in FIG. 3.

In the embodiment of the present application, the presentation learning device is presented in the form of dividing each functional module corresponding to each function, or the presentation learning device is presented in the form of dividing each functional module in an integrated manner. The “module” here may include application-specific integrated circuits (ASICs), circuits, processors and memories executing one or more software or firmware programs, integrated logic circuits, or other devices that can provide the above functions . In a simple embodiment, those skilled in the art can think that the representation learning apparatus may be implemented by using the server shown in FIG. 2. For example, the type representation module 701, context representation module 702, processing module 703, framework extension module 704, and extension mapping module 706 in FIG. 7 may be implemented by the processor 101 in FIG. 2, and the data acquisition module 705 in FIG. 7 may It is implemented by the communication interface 104 in FIG. 2. The embodiment of the present application does not make any limitation on this.

Optionally, the embodiment of the present application further provides a computer-readable storage medium, where the computer-readable storage medium stores instructions; when the computer-readable storage medium runs on the server shown in FIG. 2, The server executes the representation learning method shown in FIG. 3 or FIG. 6.

Optionally, the embodiment of the present application further provides a computer program product containing instructions, which when run on a computer enables the computer to execute the representation learning method shown in FIG. 3 or FIG. 6.

Optionally, an embodiment of the present application provides a chip system including a processor, which is configured to support a server to implement the representation learning method shown in FIG. 3 or FIG. 6. In a possible design, the chip system further includes a memory. This memory is used to store the necessary program instructions and data of the receiver. Of course, the memory may not be in the chip system. The chip system may be composed of a chip, and may also include a chip and other discrete devices, which are not specifically limited in the embodiments of the present application.

In the above embodiments, it may be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using a software program, it may be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the processes or functions according to the embodiments of the present application are wholly or partially generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions may be from a website site, a computer, a server, or a data center. Transmission by wire (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.) to another website site, computer, server, or data center. The computer-readable storage medium may be any available medium that can be accessed by a computer or a data storage device including one or more servers, data centers, and the like that can be integrated with the medium. The usable medium may be a magnetic medium (for example, a floppy disk, a hard disk, a magnetic tape), an optical medium (for example, a DVD), or a semiconductor medium (for example, a solid state disk (SSD)).

Although the present application is described in conjunction with the embodiments herein, in the process of implementing the claimed application, those skilled in the art can understand and understand by looking at the drawings, the disclosure, and the appended claims. Other variations of the disclosed embodiments are implemented. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude the case of a plurality. A single processor or other unit may fulfill the functions of several items recited in the claims. Certain measures are recited in mutually different dependent claims, but this does not mean that these measures cannot be combined to produce good results.

Although the present application has been described with reference to specific features and embodiments thereof, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of the application. Accordingly, the specification and drawings are merely exemplary illustrations of the application as defined by the appended claims, and are deemed to have covered any and all modifications, changes, combinations, or equivalents that fall within the scope of the application. Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. In this way, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application also intends to include these changes and variations.

Claims (20)

1. A representation learning method, characterized in that the method comprises:

   Determining a type representation vector of the entity according to the type of the entity in the triple of the knowledge graph of the fused text, the entity including a head entity and a tail entity;

   Determine the type representation vector of the relationship according to the type of the relationship in the triple;

   Determining a context representation vector of the entity according to the text information of the entity;

   Determining a contextual representation vector of the relationship according to a weight value of the relationship;

   Constructing a scoring function for the triplet according to the type representation vector of the entity, the context representation vector of the entity, the type representation vector of the relationship, and the context representation vector of the relationship;

   Constructing an objective function according to the scoring function of the triplet;

   Minimize the objective function and learn a representation vector of the entity and a representation vector of the relationship.
2. The representation learning method according to claim 1, wherein before determining the type representation vector of the entity in the triple of the knowledge graph of the fused text knowledge map, the method further comprises:

   Initialize a representation vector of the head entity, a representation vector of the tail entity, a representation vector of the relationship, an entity type representation matrix, a relationship type representation matrix, a word representation matrix, and a weight value representation vector.
3. The representation learning method according to claim 2, wherein determining the type representation vector of the entity according to the type of the entity in the triple of the knowledge graph of the fused text comprises:

   Determining a type identification vector of the head entity according to the type of the head entity;

   According to formula

   

   Determine the type representation vector of the head entity; where f ₁ (h) represents the type representation vector of the head entity, W _etype represents the entity type representation matrix, and v _etype (h) represents the type identifier of the head entity vector;

   Determining a type identification vector of the tail entity according to the type of the tail entity;

   According to formula

   

   The type representation vector of the tail entity is determined; wherein f ₁ (t) represents the type representation vector of the tail entity, and v _etype (t) represents the type identification vector of the tail entity.
4. The representation learning method according to claim 2, wherein determining the type representation vector of the relationship according to the type of the relationship in the triplet comprises:

   Determining a type identification vector of the relationship according to the type of the relationship;

   According to the formula:

   

   Determine the type representation vector of the relationship; where g ₁ (r) represents the type representation vector of the relationship, W _rtype represents the relationship type representation matrix, and v _rtype (r) represents the type identification vector of the relationship.
5. The representation learning method according to claim 2, wherein determining the context representation vector of the entity based on the text information of the entity comprises:

   Determining words related to the head entity according to the text information of the head entity;

   According to the formula:

   

   Determine the context representation vector of the head entity; where f ₂ (h) represents the context representation vector of the head entity, α, β are constants between 0 and 1, and v _h represents the head entity's Represents a vector, w _i represents a word related to the head entity, ε ₁ represents a set of all words related to the head entity, W _word represents a matrix of words, and V _vocubalary (w _i ) represents an identification vector of w _i ;

   Determining words related to the tail entity according to the text information of the tail entity;

   According to the formula:

   

   Determining a context representation vector of the tail entity; wherein f ₂ (t) represents a context representation vector of the tail entity, v _t represents a representation vector of the tail entity, and _mi represents a word related to the tail entity, ε ₂ represents a set of all words related to the tail entity, and V _vocubalary (m _i ) represents an identification vector of _mi .
6. The representation learning method according to claim 2, wherein the determining a context representation vector of the relationship according to a weight value of the relationship comprises:

   According to the formula:

   

   Determine the context representation vector of the relationship; where g ₂ (r) represents the context representation vector of the relationship, v _r represents the representation vector of the relationship, n _i represents the weight value of the relationship, and ε ₃ represents the A collection of relationship weights,

   

   A representation vector representing n _i .
7. The representation learning method according to any one of claims 1 to 6, wherein the scoring function of the triples is:

   

   Among them, ο represents a compound operation, f ₁ (h) represents a type of a head entity represents a vector, g ₁ (r) represents a type of a relation represents a vector, f ₁ (t) represents a type of a tail entity represents a vector, and f ₂ (h) The context representation vector represents the head entity, g ₂ (r) represents the context representation vector of the relationship, and f ₂ (t) represents the context representation vector of the tail entity.
8. The representation learning method according to claim 7, wherein the objective function is:

   

   Among them, (h, r, t) represents a positive example triplet, Δ represents a positive example triplet set, (h ′, r, t ′) represents a negative example triplet, and h ′ represents a head entity of a negative example, t ′ represents the tail entity of the negative example, Δ ′ represents the set of triples of the negative example, and M is a constant.
9. The representation learning method according to any one of claims 1 to 8, characterized in that before determining the type representation vector of the entity according to the type of the entity in the triple of the knowledge graph of the fused text, the The method also includes:

   Obtain the initial knowledge map;

   Based on the framework of the initial knowledge map, construct a framework of the knowledge map of the fused text; the framework of the knowledge map of the fused text defines at least the following: the extended attributes of the entity, the extended attributes of the relationship, and the extended relationships between the entities; The extended attributes of the entity include text information of the entity;

   Obtaining external data according to entity information or relationship information in the initial knowledge map;

   From the external data, determine the extended attribute value of the entity and the extended attribute value of the relationship to construct a knowledge map of the fused text.
10. A representation learning device, comprising:

    A type representation module, configured to determine the type representation vector of the entity according to the type of the entity in the triple of the knowledge graph of the fused text, the entity including a head entity and a tail entity; and according to the type of the relationship in the triple To determine a type representation vector of the relationship;

    A context representation module, configured to determine a context representation vector of the entity according to text information of the entity; and determine a context representation vector of the relationship according to a weight value of the relationship;

    A processing module, configured to construct a scoring function of the triple according to the type representation vector of the entity, the context representation vector of the entity, the type representation vector of the relationship, and the context representation vector of the relationship; The scoring function of the triplet is described to construct an objective function; the objective function is minimized, and a representation vector of the entity and a representation vector of the relationship are learned.
11. The representation learning device according to claim 10, wherein:

    The processing module is further configured to initialize a representation vector of the head entity, a representation vector of the tail entity, a representation vector of the relationship, an entity type representation matrix, a relationship type representation matrix, a word representation matrix, and a weight value representation. vector.
12. The representation learning device according to claim 11, wherein:

    The type indicating module is configured to determine a type identification vector of the head entity according to the type of the head entity;

    According to formula

    

    Determine the type representation vector of the head entity; where f ₁ (h) represents the type representation vector of the head entity, W _etype represents the entity type representation matrix, and v _etype (h) represents the type identifier of the head entity vector;

    Determining a type identification vector of the tail entity according to the type of the tail entity;

    According to formula

    

    The type representation vector of the tail entity is determined; wherein f ₁ (t) represents the type representation vector of the tail entity, and v _etype (t) represents the type identification vector of the tail entity.
13. The representation learning device according to claim 11, wherein:

    The type indicating module is configured to determine a type identification vector of the relationship according to the type of the relationship;

    According to the formula:

    

    Determine the type representation vector of the relationship; where g ₁ (r) represents the type representation vector of the relationship, W _rtype represents the relationship type representation matrix, and v _rtype (r) represents the type identification vector of the relationship.
14. The representation learning device according to claim 11, wherein:

    The context representation module, configured to determine words related to the head entity according to the text information of the head entity;

    According to the formula:

    

    Determine the context representation vector of the head entity; where f ₂ (h) represents the context representation vector of the head entity, α, β are constants between 0 and 1, and v _h represents the head entity's Represents a vector, w _i represents a word related to the head entity, ε ₁ represents a set of all words related to the head entity, W _word represents a matrix of words, and V _vocubalary (w _i ) represents an identification vector of w _i ;

    Determining words related to the tail entity according to the text information of the tail entity;

    According to the formula:

    

    Determining a context representation vector of the tail entity; wherein f ₂ (t) represents a context representation vector of the tail entity, v _t represents a representation vector of the tail entity, and _mi represents a word related to the tail entity, ε ₂ represents a set of all words related to the tail entity, and V _vocubalary (m _i ) represents an identification vector of _mi .
15. The representation learning device according to claim 11, wherein:

    The context representation module is used to formulate according to the formula:

    

    Determine the context representation vector of the relationship; where g ₂ (r) represents the context representation vector of the relationship, v _r represents the representation vector of the relationship, n _i represents the weight value of the relationship, and ε ₃ represents the A collection of relationship weights,

    

    A representation vector representing n _i .
16. The representation learning device according to any one of claims 10 to 15, wherein the scoring function of the triples is:

    

    Among them, ο represents a compound operation, f ₁ (h) represents a type of a head entity represents a vector, g ₁ (r) represents a type of a relation represents a vector, f ₁ (t) represents a type of a tail entity represents a vector, and f ₂ (h) The context representation vector represents the head entity, g ₂ (r) represents the context representation vector of the relationship, and f ₂ (t) represents the context representation vector of the tail entity.
17. The representation learning device according to claim 16, wherein the objective function is:

    

    Among them, (h, r, t) represents a positive example triplet, Δ represents a positive example triplet set, (h ′, r, t ′) represents a negative example triplet, and h ′ represents a head entity of a negative example, t ′ represents the tail entity of the negative example, Δ ′ represents the set of triples of the negative example, and M is a constant.
18. The representation learning device according to any one of claims 10 to 17, wherein the representation learning device further comprises: a frame extension module, a data acquisition module, and an extension mapping module;

    A framework extension module is used to obtain an initial knowledge map; based on the framework of the initial knowledge map, construct a framework of the knowledge map of the fused text; the framework of the knowledge map of the fused text defines at least the following: the extended attributes of the entity, the Extended attributes and extended relationships between entities; the extended attributes of the entities include textual information of the entities;

    The data acquisition module is configured to acquire external data according to information of entities or relationships in the initial knowledge map;

    The extended mapping module is configured to determine an extended attribute value of the entity and an extended attribute value of the relationship from external data to construct a knowledge map of the fused text.
19. A server is characterized in that it includes: a processor, a memory, a bus, and a communication interface; the memory is used to store a computer to execute instructions, the processor is connected to the memory through the bus, and when the server is running The processor executes the computer execution instructions stored in the memory, so that the server executes the representation learning method according to any one of claims 1 to 9.
20. A computer-readable storage medium, characterized in that instructions are stored in the computer-readable storage medium, and when the computer-readable storage medium is run on a computer, the computer causes the computer to execute any one of claims 1 to 9 The representation learning method described above.

Images