Entity classification method, device, electronic device and storage medium

Through the dual-view classification method of entities and types, the matching of entities and types is done by using entity and type information, the problem of poor classification reliability caused by sparse relationships between entities in the knowledge graph is solved, and the accuracy and reliability of entity classification are improved.

CN113609291BActive Publication Date: 2024-11-01IFLYTEK (SUZHOU) TECH CO LTD
View PDF 9 Cites 0 Cited by

Patent Information

Application Number
CN202110860579.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-07-27
Publication Date
2024-11-01
Estimated Expiration
2041-07-27

AI Technical Summary

Technical Problem

In the prior art, entity classification depends heavily on the density of inter-entities in the knowledge graph, resulting in poor reliability of entity classification with sparse relationships.

Method used

By determining the target entity and its candidate types, and based on the matching results of entity representation and type representation, the entity type classification is realized, and the entity information and triple information are used to perform dual-view entity classification, and obtain Information-rich type representations to compensate for sparse relationship problems.

Benefits of technology

Improve the reliability and accuracy of entity classification, ensuring the availability of entity classification in the case of sparse relationships in the knowledge graph.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113609291B_ABST
    Figure CN113609291B_ABST
Patent Text Reader

Abstract

The present invention provides an entity classification method, device, electronic device and storage medium, wherein the method comprises: determining a target entity and its candidate types; determining a matching result of the target entity and the candidate types based on the entity representation of the target entity and the type representation of the candidate types, wherein the entity representation is determined based on entity information and / or entity triple information of the corresponding entity, and the type representation is determined based on type information and / or type triple information of the corresponding type; and determining the entity type of the target entity based on the matching result of the target entity and the candidate types. The method, device, electronic device and storage medium provided by the present invention implement dual-perspective entity classification based on entities and types, wherein type information and / or type triple information that is easier to obtain and maintain than entity information is applied to obtain information-rich type representations, thereby ensuring the availability of entity classification and improving the reliability and accuracy of entity classification.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of knowledge graph processing technology, and in particular to an entity classification method, device, electronic device and storage medium. Background Art

[0002] With the rapid development of big data and artificial intelligence, knowledge graph (KG) has emerged. In the knowledge graph, data is organized in the form of a graph. The nodes in the graph are mainly entities and concepts, and the edges connecting the nodes represent the attributes of the entities or the relationship between entities.

[0003] Entity types are very important in knowledge graphs. Current entity classification mostly refers to the information inside the knowledge graph, uses the relationships between entities contained in the knowledge graph to train entity representations, and classifies entities based on entity representations. However, the above solution heavily relies on the density of relationships between entities in the knowledge graph, and performs poorly for entities with sparse relationships. Summary of the invention

[0004] The present invention provides an entity classification method, device, electronic device and storage medium to solve the problem in the prior art that entity classification is heavily dependent on the density of relationships between entities in a knowledge graph, and the reliability of entity classification with diluted relationships is poor.

[0005] The present invention provides an entity classification method, comprising:

[0006] Determine the target entity and its candidate types;

[0007] Determine a matching result between the target entity and the candidate type based on an entity representation of the target entity and a type representation of the candidate type, wherein the entity representation is determined based on entity information and / or entity triple information of the corresponding entity, and the type representation is determined based on type information and / or type triple information of the corresponding type;

[0008] Based on the matching result between the target entity and the candidate types, the entity type of the target entity is determined.

[0009] According to an entity classification method provided by the present invention, the entity representation of the target entity is determined based on the following steps:

[0010] Determine the entity information representation of the target entity based on the entity name and / or entity summary text in the entity information;

[0011] and / or, determining an entity triple representation of the entity triple information based on a predicate and a tail in the entity triple information;

[0012] An entity representation of the target entity is determined based on the entity information representation or the entity triple representation, or based on a correlation between the entity information representation and the entity triple representation.

[0013] According to an entity classification method provided by the present invention, determining the entity triple representation of the entity triple information based on the predicate and the tail in the entity triple information includes:

[0014] Determining the entity triple representation based on the correlation between the predicate representation and the tail representation in the entity triple information;

[0015] The predicate representation is determined based on the overlap of the predicate and the head entity in each triple corresponding to the target entity, and the tail representation is determined based on the semantics of the tail in each triple corresponding to the target entity.

[0016] According to an entity classification method provided by the present invention, the predicate representation is determined based on the following steps:

[0017] Taking the predicates in each triple corresponding to the target entity as nodes and the overlap of the head entities of the predicates as edge weights, a predicate relationship graph is constructed;

[0018] Based on the predicate relationship graph, the predicate representation is determined.

[0019] According to an entity classification method provided by the present invention, the type representation of the candidate type is determined based on the following steps:

[0020] Determine the type information representation of the candidate type based on the type name and / or type summary text in the type information;

[0021] and / or, determining a type triple representation of the candidate type based on a type relationship and a tail type in the type triple information;

[0022] The type representation of the candidate type is determined based on the type information representation or the type triple representation, or based on a correlation between the type information representation and the type triple representation.

[0023] According to an entity classification method provided by the present invention, determining the type triple representation of the candidate type based on the type relationship and tail type in the type triple information includes:

[0024] Determine the entity triple representation based on the correlation between the type relationship representation and the tail type representation in the triple information;

[0025] The type relationship representation is determined based on the semantics of the type relationship in each triple corresponding to the candidate type, and the tail type representation is determined based on the semantics of the tail type in each triple corresponding to the candidate type.

[0026] According to an entity classification method provided by the present invention, determining the entity type of the target entity based on the matching result between the target entity and the candidate type includes:

[0027] From each candidate type, the candidate type with the highest score of the matching result with the target entity is selected as the entity type of the target entity.

[0028] The present invention also provides an entity classification device, comprising:

[0029] A determination unit, used for determining a target entity and its candidate types;

[0030] a matching unit, configured to determine a matching result between the target entity and the candidate type based on an entity representation of the target entity and a type representation of the candidate type, wherein the entity representation is determined based on entity information and / or entity triple information of the corresponding entity, and the type representation is determined based on type information and / or type triple information of the corresponding type;

[0031] A classification unit is used to determine the entity type of the target entity based on the matching result between the target entity and the candidate type.

[0032] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any of the above-mentioned entity classification methods when executing the computer program.

[0033] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of any of the above-mentioned entity classification methods.

[0034] The entity classification method, device, electronic device and storage medium provided by the present invention realize dual-perspective entity classification based on entity and type by matching the entity representation of the target entity with the type representation of the candidate type, wherein type information and / or type triple information that is easier to obtain and maintain than entity information is used to obtain information-rich type representation, thereby compensating for the problem of less reference information due to sparse relationships of the target entity in the knowledge graph, ensuring the availability of entity classification, and improving the reliability and accuracy of entity classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly described below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0036] Figure 1 It is one of the flow charts of the entity classification method provided by the present invention;

[0037] Figure 2 It is a flowchart of a method for determining an entity representation of a target entity provided by the present invention;

[0038] Figure 3 is a flow chart of the predicate representation determination method provided by the present invention;

[0039] Figure 4 is a flow chart of a method for determining a type representation of a candidate type provided by the present invention;

[0040] Figure 5 This is the second flow chart of the entity classification method provided by the present invention;

[0041] Figure 6 It is a structural schematic diagram of the entity classification device provided by the present invention;

[0042] Figure 7 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0044] In the knowledge graph, data is organized in the form of a graph. The nodes in the graph are mainly entities and concepts, and the edges between nodes represent the attributes of the entities or the relationships between entities. For example, in the knowledge graph in the field of product recommendation, products can be regarded as nodes, and the similarities or associations between products can be regarded as edges between nodes.

[0045] Traditional entity classification methods for knowledge graphs mostly classify entities into a small set of coarse-grained types, such as only classifying entities into people, organizations, or places. However, too coarse classification will lose a lot of information, so more and more research tends to be more fine-grained entity classification.

[0046] Fine-grained entity classification aims to give entities more fine-grained type information, which can be represented in the form of paths. For example, Leonardo da Vinci can be associated with a more specific type path "person / artist / painter". Obviously, fine-grained types (such as painters, artists) can provide more specific semantic information than coarse-grained types (such as people) in data mining, and are therefore more meaningful. It can be seen that the finer the granularity of the entity type, the more useful it is in knowledge graph-based tasks, such as knowledge graph completion, entity linking, relationship extraction, and question-answering systems.

[0047] Currently, fine-grained entity classification tends to infer the type of missing entities based on the internal information of the knowledge graph. For example, entities are first embedded into a low-dimensional vector space based on the relationships, continuous attribute values, descriptions, etc. contained in the knowledge graph, and then the entities are divided into different semantic types based on the embedding results. However, although the above methods and their variations are widely used in many knowledge graph-related applications, they are powerless for entities that have only sparse relationships with other entities.

[0048] Moreover, most traditional representation learning methods represent entities and relationships in a closed knowledge graph, which makes such methods inapplicable to newly added entities. However, in the real knowledge graph construction process, the emergence of new entities is inevitable. If the representation of new entities cannot be obtained through traditional representation learning methods, it will make downstream tasks for such entities difficult to carry out.

[0049] In view of the above problems, an embodiment of the present invention provides an entity classification method to achieve fine-grained entity classification. Figure 1 is one of the flow charts of the entity classification method provided by the present invention, such as Figure 1 As shown, the method includes:

[0050] Step 110, determining the target entity and its candidate types.

[0051] Here, the target entity is the entity that needs to be classified, and the candidate type of the target entity is the candidate item of the entity type obtained by entity classification of the target entity. There can be multiple candidate types, and the candidate type can be a pre-set fixed optional type, or it can be various fine-grained types under the coarse-grained type obtained by coarse-grained classification of the target entity, or it can be a fixed optional type determined according to the technical field to which the target entity belongs. The embodiment of the present invention does not make specific limitations on this.

[0052] For example, the target entity is literary work A, and its candidate types can be various fine-grained types under the novel type, such as "martial arts novels", "human novels", "science fiction novels", "urban novels", etc.

[0053] Step 120, based on the entity representation of the target entity and the type representation of the candidate type, determine the matching result of the target entity and the candidate type, the entity representation is determined based on the entity information and / or entity triple information of the corresponding entity, and the type representation is determined based on the type information and / or type triple information of the corresponding type.

[0054] Specifically, the entity information of the target entity is used to reflect the information of the entity itself, such as the name of the entity, the introductory text of the entity, etc. The entity triple information of the target entity is used to reflect the information reflected by the triple with the target entity as the head entity. Here, the triple with the target entity as the head entity can be expressed in the form of "target entity, predicate, tail", where the predicate can be the relationship between the target entity and the tail entity, such as "literary work A, author, writer C", and the predicate can also be the attribute of the target entity, and the corresponding tail is the attribute value, such as "literary work A, creation year, XX years". The entity triple information can include the information reflected by the predicate and tail in the triple with the target entity as the head entity.

[0055] The entity representation of the target entity can be determined based on the entity information and / or entity triple information of the target entity, wherein the entity representation determined based on the entity information can cover the information of the target entity itself, and the entity representation determined based on the entity triple information can absorb the information related to the target entity implied in the entity triple, thereby enriching the content contained in the entity representation. The determination of the entity representation can rely solely on the entity information or the entity triple information, or can be a combination of the two, and the embodiments of the present invention do not specifically limit this.

[0056] The type information of the candidate type is used to reflect the information of the type itself, such as the name of the type, the introductory text of the type, etc. The type triple information of the candidate type is used to reflect the information reflected by the triple with the candidate type as the head entity. Here, the triple with the candidate type as the head entity can be regarded as a type entity triple, which is specifically expressed in the form of "candidate type, relationship, other types", that is, the type triple is used to show the relationship between various types. The relationship here can be used to express the similarity or correlation between two types, for example, it can be determined by the degree of overlap of entities belonging to the two types. The type triple information can include the information reflected by the relationship and the tail type in the triple with the candidate type as the head entity.

[0057] The type representation of the candidate type can be determined based on the type information and / or type triple information of the candidate type, wherein the type representation determined based on the type information can cover the information of the candidate type itself, and the type representation determined based on the type triple information can absorb the information related to the candidate type implicit in the type triple, thereby enriching the content contained in the type representation. The determination of the type representation can rely only on the type information or the type triple information, or can combine the two, which is not specifically limited in the embodiments of the present invention.

[0058] After obtaining the entity representation of the target entity and the type representation of the candidate type, the target entity and the candidate type can be matched based on the entity representation and the type representation, thereby obtaining a matching result of the two. The matching result here can be a matching score of the target entity and the candidate type. The higher the matching score, the higher the probability that the candidate type is the entity type of the target entity. The matching result can also be whether there is a match, which is not specifically limited in the embodiment of the present invention.

[0059] By matching entity representation and type representation, entity classification no longer relies solely on entity-related information obtained based on the knowledge graph, but instead adds type-related information, so that entity classification can be performed from both entity and type perspectives, thereby improving the reliability and accuracy of entity classification. Moreover, from the type perspective, compared with entities, the number of candidate types is smaller, the probability of adding new types is lower, the acquisition and maintenance of type information and / or type triple information is less difficult, and it is easier to obtain information-rich type representation. Therefore, entity classification combined with type representation can make up for the problem of less reference information due to sparse relationships of target entities in the knowledge graph, thereby improving the reliability of entity classification.

[0060] Step 130 , determining the entity type of the target entity based on the matching result between the target entity and the candidate types.

[0061] Specifically, the entity type of the target entity can be obtained by matching the target entity and one or more candidate types corresponding to the target entity. For example, based on the matching score of a candidate type, it is possible to determine whether the candidate type is an entity type based on the size relationship between the matching score and a preset threshold; based on the matching scores of multiple candidate types, the candidate type with the highest matching score can be selected as the entity type, or the highest matching score can be compared with a preset threshold. If it is greater than the threshold, the candidate type corresponding to the matching score is determined to be an entity type.

[0062] The method provided by the embodiment of the present invention realizes dual-perspective entity classification based on entity and type by matching the entity representation of the target entity with the type representation of the candidate type, wherein type information and / or type triple information that is easier to obtain and maintain than entity information is applied to obtain information-rich type representation, thereby compensating for the problem of less reference information due to sparse relationships of the target entity in the knowledge graph, ensuring the availability of entity classification, and improving the reliability and accuracy of entity classification.

[0063] Based on the above embodiments, Figure 2 is a flow chart of a method for determining an entity representation of a target entity provided by the present invention, such as Figure 2 As shown in Figure 1, the entity representation of the target entity is determined based on the following steps:

[0064] Step 210: Determine the entity information representation of the target entity based on the entity name and / or entity summary text in the entity information.

[0065] Specifically, entity information covers the information of the entity itself, among which the entity name contained therein is the direct identifier of the entity, and the entity summary text can concisely and accurately record the relevant information of the entity. For example, if the target entity is a literary work, the entity name of the target entity is the name of the work, and the entity summary text is the text used to introduce the literary work, such as "Literary work A is a novel written by writer C in the XX century. The novel takes the love tragedy of the protagonists Xiao Ming and Xiao Li as the main line, and shows the various aspects of ancient society."

[0066] The entity information representation of the target entity can be determined based on the entity name and / or entity summary text of the target entity. For example, the semantic encoding of the entity name can be used as the entity information, or the semantic encoding of the entity summary text can be used as the entity information, or the semantic encoding of the entity name and the entity summary text can be fused as the entity information. The specific fusion method can be splicing, weighted addition, etc., which is not specifically limited in the embodiments of the present invention.

[0067] And / or, step 220, determining the entity triple representation of the entity triple information based on the predicate and the tail in the entity triple information.

[0068] Specifically, the entity triple information includes the predicate and tail of the triple with the target entity as the head entity. The predicate and tail in the entity triple information can be one-to-one corresponding, or one predicate can correspond to multiple tails. The embodiment of the present invention does not specifically limit the number of predicates in the entity triple information.

[0069] The entity triple representation of the target entity can be determined based on the predicate and tail in the entity triple information. For example, the semantics of the corresponding predicate and tail can be concatenated according to the correspondence between the predicate and tail in the entity triple information to obtain the entity triple representation. The semantics of the predicate and tail can also be selectively fused based on the correlation between the predicate and tail to obtain the entity triple representation. The embodiments of the present invention do not specifically limit this.

[0070] Step 230 , determining the entity representation of the target entity based on the entity information representation or the entity triple representation, or based on the correlation between the entity information representation and the entity triple representation.

[0071] Specifically, when determining the entity representation, only the entity information representation obtained in step 210 can be applied to determine the entity representation of the target entity, for example, directly using the entity information representation as the entity representation, or performing further feature extraction on the entity information representation to obtain the entity representation; in addition, only the entity triple representation obtained in step 220 can be applied to determine the entity representation of the target entity, for example, directly using the entity triple representation as the entity representation, or performing further feature extraction on the entity triple representation to obtain the entity representation.

[0072] In addition, the entity information representation obtained in step 210 and the entity triple representation obtained in step 220 can be combined to determine the entity representation. In this process, the hidden connection and interaction between the entity information and the entity triple information are taken into account. For example, when people read the entity information of the target entity, especially the entity summary text, they can clearly perceive that the content displayed by the entity information may be closer to the content represented by certain predicates and tails in the entity triple information, and the correlation with the content represented by other predicates and tails is not high. Therefore, in the embodiment of the present invention, the correlation between the entity information representation and the entity triple representation can be applied. When the entity information representation and the entity triple representation are fused, the part with higher correlation between the two is highlighted, and the part with lower correlation between the two is weakened, so that the fused entity representation can be closer to the focus of human thinking in entity classification, and the entity representation can improve the representation ability of the target entity and its related information.

[0073] For example, when people read the entity summary text of the target entity person "Xiaohong", they find that the entity summary text focuses on introducing Xiaohong's paintings. Then, in the entity triple information of "Xiaohong", the triples with the predicate "works" or "paintings" are obviously more relevant and more important than the triples with the predicates "birth", "family", etc. Therefore, when generating entity representation, the content related to the work can be highlighted, and the content that is less related to the work or irrelevant can be weakened.

[0074] Furthermore, the correlation between the entity information representation and the entity triple representation can be realized by performing attention interaction on the entity information representation and the entity triple representation. Accordingly, the entity representation determined based on the correlation can be the result of performing attention interaction on the entity information representation and the entity triple representation. For example, the entity representation can be embodied in the following form:

[0075]

[0076] In the formula, E e is the entity representation, att(·,·) is the attention interaction function, [E en ; E ed ] is the entity information representation, where E en is the encoding representation of the entity name, E ed is the encoding representation of the entity summary text; It is represented by entity triples.

[0077] The method provided by the embodiment of the present invention uses entity information representation and / or entity triple representation to realize entity representation of the target entity. When combining entity information representation and entity triple representation for entity representation, the correlation between the two is applied, which helps to improve the representation ability of entity representation for the target entity and its related information, thereby improving the accuracy and reliability of entity classification.

[0078] Based on any of the above embodiments, step 220 includes:

[0079] Determine the entity triple representation based on the correlation between the predicate representation and the tail representation in the entity triple information;

[0080] The predicate representation is determined based on the overlap of the predicate and the head entity in each triple corresponding to the target entity, and the tail representation is determined based on the semantics of the tail in each triple corresponding to the target entity.

[0081] Specifically, considering that the predicates in the triple of the target entity may be diverse, and there may be similarities or associations between the diverse predicates, integrating the associations between the predicates can reflect the selection bias of the predicates of the target entity in the entity triple in a more integrated way in the predicate representation, thereby reflecting the tendency of the target entity to express meaning. Here, the predicate representation can be determined based on the predicates in each triple corresponding to the target entity, and the correlation between the predicates. The correlation between the predicates can be specifically expressed as the overlap of the head entity in the triple of the predicate. For example, through the existing knowledge graph, the i-th predicate p can be obtained. i and the jth predicate p j The set of head entities in each triplet is denoted as ES(p i ) and ES(pj ), can be measured by measuring ES(p i ) and ES(p j ) to determine the predicate p i and p j The relevance of the word further affects the determination of the predicate representation.

[0082] The tail representation can be determined based on the semantics of the tail in the triple in which the target entity is located. The semantics of the tail here can be the semantics of the tail noun itself, or it can be combined with information related to the tail noun, such as the semantics obtained from the description text of the tail noun. The embodiment of the present invention does not specifically limit this. For example, when the tail is an attribute value, the semantic information of the attribute value can be directly used as the tail representation. For another example, when the tail is a tail entity, the semantic information of the entity summary text of the tail entity can be used as the tail representation.

[0083] There are also hidden connections and interactions between the predicates and tails in different entity triples. For example, for the same predicate "work", assuming that the type of the head entity is a painter, the corresponding tail is usually a painting. If the type of the head entity is a musician, the corresponding tail is usually a musical work. Obviously, paintings and musical works are two different types of works. On the contrary, under the premise of understanding the predicate and tail of the entity triple, the information provided by the predicate and tail can help distinguish the type of the head entity. It can be seen that the hidden interactions and associations between the predicate and tail can help enhance the semantic representation of the entity triple information and even enhance the semantic representation of the target entity.

[0084] Furthermore, the correlation between the predicate representation and the tail representation can be realized by performing attention interaction on the predicate representation and the tail representation. Accordingly, the entity triple representation determined based on the correlation can be the result of performing attention interaction on the predicate representation and the tail representation. For example, the entity triple representation can be embodied in the following form:

[0085]

[0086] In the formula, That is, the entity triple is represented by Representation by Predicate Attention Interaction Representation of interaction with tail attention Splicing obtained.

[0087] Among them, the representation of predicate attention interaction That is, the set of attention representations of each predicate in the entity triple information, to is the attention representation of the 1st to nth predicates, n is the total number of predicates in the entity triple information, and the attention representation of the i-th predicate is It can be:

[0088]

[0089] Where att(·, ·) is the attention interaction function, is the predicate representation of the i-th predicate, E et Represents the tail.

[0090] Representation of tail attention interactions That is, the set of attention representations of each tail in the entity triple information, to is the attention representation from the 1st tail to the nth tail, where the attention representation of the i-th tail is It can be:

[0091]

[0092] Where att(·, ·) is the attention interaction function, is the tail representation of the i-th tail, E ep For the predicate representation.

[0093] The method provided by the embodiment of the present invention determines the entity triple representation based on the correlation between the predicate representation and the tail representation, which helps to improve the reliability of the entity triple representation, thereby improving the reliability of entity classification.

[0094] Based on any of the above embodiments, Figure 3 is a flow chart of the predicate representation determination method provided by the present invention, such as Figure 3 As shown, the predicate representation is determined based on the following steps:

[0095] Step 310 , construct a predicate relationship graph using the predicates in each triple corresponding to the target entity as nodes and the overlap of the head entities of the predicates as edge weights.

[0096] Step 320: determine the predicate representation based on the predicate relationship graph.

[0097] Specifically, considering that there may be similarities or associations between various predicates in the triple of the target entity, in order to apply the associations between predicates to improve the reliability of the entity representation of the target entity, the predicate representation containing the associations between predicates can be obtained by constructing a predicate relationship graph and further extracting features from the predicate relationship graph.

[0098] Here, the construction of the predicate relationship graph is realized by taking each predicate in the entity triple information of the target entity as a node and taking the overlap of the head entities of each two predicates as the weight of the edge between the corresponding two nodes. Among them, the overlap of the head entities in the triples where the predicates are located can be expressed as the intersection ratio of the head entity sets in the triples where the two predicates are located. For example, the predicate p in the predicate relationship graph can be expressed by the following formula i to the predicate p j The edge weights are:

[0099]

[0100] In the formula, That is, the predicate p i to the predicate p j The weight of the edge, that is, for the predicate p i For the predicate p j The degree of association, len(·) indicates the number, and the predicate p i and the predicate p j The head entity sets in the triples are ES(p i ) and ES(p j ), len(ES(p i )∩ES(p j )) is the number of head entities in both head entity sets, len(ES(p i )) is ES(p i ) in the header entity.

[0101] After obtaining the predicate relationship graph, the predicate relationship graph can be feature extracted to obtain the predicate representation. The predicate representation here includes the information of each node in the predicate relationship graph and the connection relationship between the nodes in the predicate relationship graph, so it can fully and completely display the characteristics of the predicate part contained in the entity triple information.

[0102] Furthermore, feature extraction of the predicate relationship graph can be achieved through a graph convolutional network (GCN) or various variations of the graph convolutional network. For example, the predicate representation can be obtained by performing multiple convolutions on the predicate relationship graph, where a single convolution can be expressed as follows:

[0103]

[0104]

[0105] in, The kth convolution obtains the predicate p i Neighbour(p i) is the predicate p i In the predicate relationship diagram, the predicate p i A connected set of predicates, The k-1th convolution obtains the predicate p j When k = 0, For the predicate p i The encoding representation of , for example, can be obtained by one-hot encoding Another example is the semantic representation obtained through bert-as-service encoding.

[0106] After multiple convolutions, the predicate representation can be obtained

[0107] Based on any of the above embodiments, Figure 4 is a flow chart of a method for determining a type representation of a candidate type provided by the present invention, such as Figure 4 As shown, the type representation of the candidate type is determined based on the following steps:

[0108] Step 410: Determine the type information representation of the candidate type based on the type name and / or type summary text in the type information.

[0109] Specifically, the type information covers the information of the type itself, among which the type name is a direct identifier of the type, and the type summary text can concisely and accurately record the relevant information of the type. For example, if the candidate type is martial arts novels, the type name of the candidate type is martial arts novels, and the type summary text is the text used to introduce the type of martial arts novels, for example, "martial arts novels often feature knights and righteous men as protagonists, describing their unique skills and courageous acts."

[0110] The type information representation of the candidate type can be determined based on the type name and / or type summary text of the candidate type. For example, the semantic encoding of the type name can be used as the type information, or the semantic encoding of the type summary text can be used as the type information, or the semantic encoding of the type name and the type summary text can be fused as the type information. The specific fusion method can be splicing, weighted addition, etc., which is not specifically limited in the embodiments of the present invention.

[0111] And / or, step 420, determining the type triple representation of the candidate type based on the type relationship and tail type in the type triple information.

[0112] Specifically, the type triple information includes the type relationship and tail type of the triple with the candidate type as the head entity. The type relationship and tail type in the type triple information can be one-to-one corresponding, or one type relationship can correspond to multiple tail types. The embodiment of the present invention does not specifically limit the number of type relationships in the type triple information.

[0113] The type triple representation of the candidate type can be determined based on the type relationship and tail type in the type triple information. For example, according to the correspondence between the type relationship and the tail type in the type triple information, the corresponding type relationship and the semantics of the tail type can be spliced ​​to obtain the type triple representation. Alternatively, the semantics of the type relationship and the tail type can be selectively fused in combination with the correlation between the type relationship and the tail type to obtain the type triple representation. The embodiments of the present invention do not specifically limit this.

[0114] Step 430 : Determine the type representation of the candidate type based on the type information representation or the type triple representation, or based on the correlation between the type information representation and the type triple representation.

[0115] Specifically, when determining the type representation, only the type information representation obtained in step 410 can be applied to determine the type representation of the candidate type, for example, directly using the type information representation as the type representation, or performing further feature extraction on the type information representation to obtain the type representation; in addition, only the type triple representation obtained in step 420 can be applied to determine the type representation of the candidate type, for example, directly using the type triple representation as the type representation, or performing further feature extraction on the type triple representation to obtain the type representation.

[0116] In addition, the type information representation obtained in step 410 and the type triple representation obtained in step 420 can be combined to determine the type representation. In this process, the hidden connection and interaction between the type information and the type triple information are taken into account. For example, when people read the type information of a candidate type, especially the type summary text, they can clearly perceive that the content displayed by the type information may be closer to certain type relationships in the type triple information and the content represented by the tail type, while the correlation with the content represented by other type relationships and the tail type is not high. Therefore, in an embodiment of the present invention, the correlation between the type information representation and the type triple representation can be applied. When the type information representation and the type triple representation are fused, the part with higher correlation between the two is highlighted, and the part with lower correlation between the two is weakened, so that the fused type representation can be closer to the focus of human thinking on the understanding and cognition of candidate types during entity classification, thereby improving the representation ability of the type representation for candidate types and related information.

[0117] Furthermore, the correlation between the type information representation and the type triple representation can be realized by performing attention interaction on the type information representation and the type triple representation, and the type representation determined based on the correlation can be the result of performing attention interaction on the type information representation and the type triple representation. For example, the type representation can be embodied in the following form:

[0118]

[0119] In the formula, E t is the type representation, att(·,·) is the attention interaction function, [E tn ; E td ] is the type information representation, where E tn is the encoding representation of the type name, E td is the encoded representation of the type summary text; It is represented by a type triple.

[0120] The method provided by the embodiment of the present invention uses type information representation and / or type triple representation to realize type representation of candidate types. When combining type information representation and type triple representation for type representation, the correlation between the two is applied, which helps to improve the characterization capability of type representation for candidate types and related information, thereby improving the accuracy and reliability of entity classification.

[0121] Based on any of the above embodiments, step 420 includes:

[0122] Determine the entity triple representation based on the correlation between the type relationship representation in the triple information and the tail type representation;

[0123] The type relationship representation is determined based on the semantics of the type relationship in each triple corresponding to the candidate type, and the tail type representation is determined based on the semantics of the tail type in each triple corresponding to the candidate type.

[0124] Specifically, compared to the various predicates in the triple where the target entity is located, the types of type relations of the candidate type are relatively small, and the type relations that may exist in each type are relatively similar. Therefore, considering the relationship between various type relations of the candidate type cannot obtain richer information as considering the relationship between various predicates of the target entity. Instead, it may lead to the representation results of various type relations being too similar. Therefore, in the embodiment of the present invention, for various type relations in the triple where the candidate type is located, only the semantics of various type relations are obtained to construct the type relation representation.

[0125] The tail type representation can be determined based on the semantics of the tail type in the triple in which the candidate type is located. Here, the semantics of the tail type can be the semantics of the tail type noun itself, or it can be combined with information related to the tail type noun, such as the semantics obtained from the description text of the tail type noun. The embodiment of the present invention does not specifically limit this. For example, the semantic information of the tail type can be directly used as the tail type representation, or the semantic information of the type summary text of the tail type can be used as the tail type representation.

[0126] There are also hidden connections and interactions between type relations and tail types in different type triples. The correlation between type relations and tail types can be helpful for enhancing the semantic representation of type triple information and even the semantic representation of candidate types. Furthermore, the correlation between type relation representation and tail type can be achieved by performing attention interaction on type relation representation and tail type representation. Accordingly, the type triple representation determined based on the correlation can be the result of performing attention interaction on type relation representation and tail type representation. For example, the type triple representation can be embodied in the following form:

[0127]

[0128] In the formula, That is, the type triple is expressed as, Representation of Attention Interactions by Type Relations Representation of interaction with tail-type attention Splicing obtained.

[0129] Among them, the representation of type relationship attention interaction That is, the set of attention representations of each type relationship in the type triple information, to is the attention representation of the 1st type relationship to the nth type relationship, m is the total number of type relationships in the entity triple information, and the attention representation of the i-th type relationship is It can be:

[0130]

[0131] Where att(·, ·) is the attention interaction function, is the type relation representation of the i-th type relation, E tt The tail type is represented.

[0132] Representation of tail-type attention interactions That is, the set of the meanings of each tail type in the type triple information, to is the attention representation of the 1st tail type to the mth tail type, where the attention representation of the i-th tail type It can be:

[0133]

[0134] Where att(·, ·) is the attention interaction function, is the tail type representation of the i-th tail type, E tpRepresents a type relationship.

[0135] The method provided by the embodiment of the present invention determines the type triple representation based on the correlation between the type relationship representation and the tail type representation, which helps to improve the reliability of the type triple representation, thereby improving the reliability of entity classification.

[0136] Based on any of the above embodiments, step 130 includes:

[0137] From each candidate type, the candidate type with the highest score in the matching result with the target entity is selected as the entity type of the target entity.

[0138] Specifically, when a target entity corresponds to multiple candidate types, the entity representation of the target entity can be matched with the type representation of each candidate type to obtain the matching results of the target entity and each candidate type. Here, the matching results of the target entity and each candidate type include a score to measure the degree of matching. The higher the score, the higher the probability that the target entity belongs to the candidate type.

[0139] Therefore, after obtaining the matching results between the target entity and each candidate type, the candidate type corresponding to the matching result with the highest score can be selected as the entity type of the target entity.

[0140] Based on any of the above embodiments, Figure 5 This is the second flow chart of the entity classification method provided by the present invention. Figure 5 As shown, the entity classification method can obtain the entity representation of the target entity and the type representation of the candidate type from two perspectives, and then combine the two for matching.

[0141] Among them, for the target entity, you can first obtain the entity name, entity description text, and the predicate set and tail set in the triple of the target entity as the head entity. For the entity name and entity description text, you can obtain the semantic representation of the entity name and the semantic representation of the entity description text respectively, and splice the two as the entity information representation. Figure 5 In the figure, the symbol with a plus sign in a circle represents a splicing operation.

[0142] For the predicate set and tail set in the entity triple, each predicate in the predicate set can be one-hot encoded, thereby obtaining the predicate vector of each predicate with a length equal to the number of predicates and the subscript corresponding to the predicate sequence number set to 1 and all other subscripts set to 0. On this basis, the predicate can be used as a node, and the overlap of the head entities of each two predicates can be used as the weight of the edge between the two nodes to construct a predicate relationship graph, and the predicate representation of each predicate can be obtained by feature extraction of the predicate relationship graph. For each tail in the tail set, the semantics of each tail can be obtained to construct the tail representation.

[0143] On this basis, attention interaction calculation (Attention) can be performed on the predicate representation of the predicate set and the tail representation of the tail set, and the predicate attention representation and tail attention representation obtained by the attention interaction can be concatenated to obtain the entity triple representation.

[0144] Next, attention interaction is performed on the entity information representation and the entity triple representation to obtain the entity representation.

[0145] For any candidate type, you can first obtain the type name, type description text, and type relationship set and tail type set in the triple of the candidate type as the head entity. For the type name and type description text, you can obtain the semantic representation of the type name and the semantic representation of the type description text respectively, and splice the two as the type information representation.

[0146] For the type relation set and the tail type set in the triple of the type relation of the candidate type, the semantic representation of each type relation in the type relation set can be used as the type relation representation of each type relation. For each tail type in the tail type set, the semantics of each tail type can be obtained to construct the tail type representation.

[0147] On this basis, attention interaction calculation (Attention) can be performed on the type relationship representation of the type relationship set and the tail type representation of the tail type set, and the type relationship attention representation and the tail type attention representation obtained by the attention interaction can be spliced ​​to obtain the type triple representation.

[0148] Next, attention interaction is performed on the type information representation and the type triple representation to obtain the type representation.

[0149] Afterwards, the entity representation and the type representation can be matched and scored, and the matching score can be output as the matching result, thereby determining the type of the target entity. Figure 5In the figure, the symbol with a multiplication sign in the circle represents the function used for matching score, such as sigmoid function or other types of activation functions. Taking sigmoid function as an example, the matching score can be expressed as follows:

[0150] y′=sigmod(Score(E e , E t ))

[0151] In the formula, y′ is the final output matching score, that is, the matching result. Score(E e , E t ) represents entity-based representation E e and type representation E t The scoring function for the operation.

[0152] The method provided by the embodiment of the present invention mines the hidden interactions between various parts through a multi-attention mechanism, thereby obtaining higher-quality entity representation and type representation. The introduction of summary text of entities and types and triple information not only enhances the semantic representation ability of entities and types, but also can be applied to newly added entities and entities with sparse relationships, effectively ensuring the reliability of the entity classification method.

[0153] For newly added entities, only the entity information including the entity name and entity summary text can be used to construct the entity representation, without the need for global retraining. For entities with sparse relationships, rich semantic information can also be obtained through non-relational triples and entity summary text to determine the entity representation, ensuring the robustness of the entity classification method.

[0154] Based on any of the above embodiments, Figure 5 The entity classification method shown can be implemented through an end-to-end model, where the model has an initial representation layer, a hidden representation layer, and a score function layer.

[0155] Among them, the initial representation layer is used to initialize the representation of the input entity information, entity triple information, type information and type triple information, so as to obtain entity information representation, type information representation, predicate representation and tail representation of the target entity, type relationship representation and tail entity representation of the candidate type.

[0156] The hidden representation layer is used to perceive the correlation between the predicate representation and the tail representation of the target entity through attention interaction, thereby obtaining the entity triple representation; perceive the correlation between the type relationship representation and the tail entity representation of the candidate type, thereby obtaining the type triple representation. And on this basis, the hidden representation layer is also used to perceive the correlation between the entity information representation and the entity triple representation of the target entity through attention interaction, thereby obtaining the entity representation, and perceive the correlation between the type information representation and the type triple representation of the candidate type, thereby obtaining the type representation.

[0157] The scoring function layer is used to perform matching scores based on the entity representation and type representation, thereby obtaining the matching scores of the candidate types for the target entity.

[0158] In addition, the training of the above model can be implemented with the entity information, entity triple information, type information of the candidate type and type triple information of the sample entity as samples, and whether the sample entity belongs to the candidate type as a label. For example, if the sample entity belongs to the candidate type, the target value of the matching score is 1, otherwise the target value of the matching score is 0. Specifically, during training, the standard cross entropy function can be applied as the loss function to make the model's prediction value as close to the actual label as possible.

[0159] Based on any of the above embodiments, Figure 6 is a schematic diagram of the structure of the entity classification device provided by the present invention, such as Figure 6 As shown, the device comprises:

[0160] A determination unit 610, configured to determine a target entity and its candidate types;

[0161] A matching unit 620, configured to determine a matching result between the target entity and the candidate type based on an entity representation of the target entity and a type representation of the candidate type, wherein the entity representation is determined based on entity information and / or entity triple information of the corresponding entity, and the type representation is determined based on type information and / or type triple information of the corresponding type;

[0162] The classification unit 630 is configured to determine the entity type of the target entity based on the matching result between the target entity and the candidate types.

[0163] The device provided by the embodiment of the present invention realizes dual-perspective entity classification based on entity and type by matching the entity representation of the target entity with the type representation of the candidate type, wherein type information and / or type triple information that is easier to obtain and maintain than entity information is applied to obtain information-rich type representation, thereby compensating for the problem of less reference information due to sparse relationships of the target entity in the knowledge graph, ensuring the availability of entity classification, and improving the reliability and accuracy of entity classification.

[0164] Based on any of the above embodiments, the device further includes an entity representation unit, including:

[0165] An entity information representation subunit, configured to determine the entity information representation of the target entity based on the entity name and / or entity summary text in the entity information;

[0166] and / or, an entity triple representation subunit, for determining an entity triple representation of the entity triple information based on a predicate and a tail in the entity triple information;

[0167] The entity representation subunit is used to determine the entity representation of the target entity based on the entity information representation or the entity triple representation, or based on the correlation between the entity information representation and the entity triple representation.

[0168] Based on any of the above embodiments, the entity triple representation subunit is used for:

[0169] Determining the entity triple representation based on the correlation between the predicate representation and the tail representation in the entity triple information;

[0170] The predicate representation is determined based on the overlap of the predicate and the head entity in each triple corresponding to the target entity, and the tail representation is determined based on the semantics of the tail in each triple corresponding to the target entity.

[0171] Based on any of the above embodiments, the entity triple representation subunit is used for:

[0172] Taking the predicates in each triple corresponding to the target entity as nodes and the overlap of the head entities of the predicates as edge weights, a predicate relationship graph is constructed;

[0173] Based on the predicate relationship graph, the predicate representation is determined.

[0174] Based on any of the above embodiments, the device further includes a type indicating unit, configured to:

[0175] A type information representation subunit, configured to determine the type information representation of the candidate type based on the type name and / or type summary text in the type information;

[0176] and / or, a type triple representation subunit, for determining a type triple representation of the candidate type based on a type relationship and a tail type in the type triple information;

[0177] The type representation subunit is used to determine the type representation of the candidate type based on the type information representation or the type triple representation, or based on the correlation between the type information representation and the type triple representation.

[0178] Based on any of the above embodiments, the type triple represents a subunit for:

[0179] Determine the entity triple representation based on the correlation between the type relationship representation and the tail type representation in the triple information;

[0180] The type relationship representation is determined based on the semantics of the type relationship in each triple corresponding to the candidate type, and the tail type representation is determined based on the semantics of the tail type in each triple corresponding to the candidate type.

[0181] Based on any of the above embodiments, the classification unit 630 is used for:

[0182] From each candidate type, the candidate type with the highest score of the matching result with the target entity is selected as the entity type of the target entity.

[0183] Figure 7 An example of a physical structure diagram of an electronic device is shown in FIG. Figure 7 As shown, the electronic device may include: a processor 710, a communication interface 720, a memory 730 and a communication bus 740, wherein the processor 710, the communication interface 720 and the memory 730 communicate with each other through the communication bus 740. The processor 710 may call the logic instructions in the memory 730 to execute the entity classification method, which includes: determining a target entity and its candidate types; determining a matching result between the target entity and the candidate type based on the entity representation of the target entity and the type representation of the candidate type, wherein the entity representation is determined based on the entity information and / or entity triple information of the corresponding entity, and the type representation is determined based on the type information and / or type triple information of the corresponding type; determining the entity type of the target entity based on the matching result between the target entity and the candidate type.

[0184] In addition, the logic instructions in the above-mentioned memory 730 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0185] On the other hand, the present invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the entity classification method provided by the above-mentioned methods, and the method includes: determining a target entity and its candidate types; based on the entity representation of the target entity and the type representation of the candidate type, determining a matching result between the target entity and the candidate type, the entity representation is determined based on entity information and / or entity triple information of the corresponding entity, and the type representation is determined based on type information and / or type triple information of the corresponding type; based on the matching result between the target entity and the candidate type, determining the entity type of the target entity.

[0186] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the above-mentioned entity classification methods, the methods comprising: determining a target entity and its candidate types; determining a matching result between the target entity and the candidate type based on an entity representation of the target entity and a type representation of the candidate type, the entity representation being determined based on entity information and / or entity triple information of a corresponding entity, and the type representation being determined based on type information and / or type triple information of a corresponding type; determining the entity type of the target entity based on the matching result between the target entity and the candidate type.

[0187] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0188] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0189] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An entity classification method, characterized in that: include: Determine the target entity and its candidate types; Determining a matching result between the target entity and the candidate type based on the entity representation of the target entity and the type representation of the candidate type, wherein the entity representation is determined based on the entity triple representation of the corresponding entity, or based on the correlation between the entity information representation and the entity triple representation of the corresponding entity, and the type representation is determined based on the type information and / or type triple information of the corresponding type; Determining the entity type of the target entity based on the matching result between the target entity and the candidate types; The entity triple representation is determined based on the correlation between the predicate representation and the tail representation in the entity triple information of the entity, the predicate representation is determined based on the coincidence of the predicate in each triple corresponding to the target entity and its head entity, and the tail representation is determined based on the semantics of the tail in each triple corresponding to the target entity; The type representation of the candidate type is determined based on the following steps: Determine the type information representation of the candidate type based on the type name and / or type summary text in the type information; and / or, determining a type triple representation of the candidate type based on a type relationship and a tail type in the type triple information; Determine the type representation of the candidate type based on the type information representation or the type triple representation, or based on a correlation between the type information representation and the type triple representation; The predicate representation is determined based on the following steps: A predicate relationship graph is constructed by taking the predicates in each triple corresponding to the target entity as nodes and the overlap of the head entities of the predicates as edge weights, wherein the overlap is the intersection ratio of the head entity sets in the triples where the two predicates are located; Based on the predicate relationship graph, the predicate representation is determined.

2. The entity classification method according to claim 1, characterized in that: The entity information representation of the target entity is determined based on the following steps: Based on the entity name and / or entity summary text in the entity information, an entity information representation of the target entity is determined.

3. The entity classification method according to claim 1, characterized in that: The determining, based on the type relationship and the tail type in the type triple information, the type triple representation of the candidate type comprises: Determine the entity triple representation based on the correlation between the type relationship representation and the tail type representation in the triple information; The type relationship representation is determined based on the semantics of the type relationship in each triple corresponding to the candidate type, and the tail type representation is determined based on the semantics of the tail type in each triple corresponding to the candidate type.

4. The entity classification method according to any one of claims 1 to 3, characterized in that: The determining the entity type of the target entity based on the matching result between the target entity and the candidate type includes: From each candidate type, the candidate type with the highest score of the matching result with the target entity is selected as the entity type of the target entity.

5. An entity classification device, characterized in that: include: A determination unit, used for determining a target entity and its candidate types; a matching unit, configured to determine a matching result between the target entity and the candidate type based on an entity representation of the target entity and a type representation of the candidate type, wherein the entity representation is determined based on an entity triple representation of a corresponding entity, or based on a correlation between an entity information representation and an entity triple representation of a corresponding entity, and the type representation is determined based on type information and / or type triple information of the corresponding type; A classification unit, configured to determine an entity type of the target entity based on a matching result between the target entity and the candidate types; The entity triple representation is determined based on the correlation between the predicate representation and the tail representation in the entity triple information of the entity, the predicate representation is determined based on the coincidence of the predicate in each triple corresponding to the target entity and its head entity, and the tail representation is determined based on the semantics of the tail in each triple corresponding to the target entity; The type representation of the candidate type is determined based on the following steps: Determine the type information representation of the candidate type based on the type name and / or type summary text in the type information; and / or, determining a type triple representation of the candidate type based on a type relationship and a tail type in the type triple information; Determine the type representation of the candidate type based on the type information representation or the type triple representation, or based on a correlation between the type information representation and the type triple representation; The predicate representation is determined based on the following steps: A predicate relationship graph is constructed by taking the predicates in each triple corresponding to the target entity as nodes and the overlap of the head entities of the predicates as edge weights, wherein the overlap is the intersection ratio of the head entity sets in the triples where the two predicates are located; Based on the predicate relationship graph, the predicate representation is determined.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the entity classification method according to any one of claims 1 to 4 are implemented.

7. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the entity classification method according to any one of claims 1 to 4 are implemented.

Citation Information

Patent Citations

  • Entity classification method, system and device

    CN111324749A

  • Information analysis method and device based on knowledge graph, equipment and storage medium

    CN111368096A

  • Entity classification method and device, computer equipment and storage medium

    CN111444344A

  • Entity classification method and related device

    CN111460826A

  • Entity relationship classification model construction method and device and storage medium

    CN112948505A