Identification analysis knowledge graph construction method based on adversarial neural network

Abstract

The invention provides an identification analysis knowledge graph construction method based on an adversarial neural network, and belongs to the technical field of the Internet, and the method comprises the steps: 1, obtaining various types of data related to identification analysis, and carrying out the cleaning and conversion; 2, performing entity recognition on the cleaned and converted data and text, and recognizing semantic relation extraction between entities from the text; 3, constructing an entity-relation-entity triple, and forming a preliminary knowledge graph; 4, taking the preliminary knowledge graph as a data set to train an adversarial neural network, and obtaining the discrimination capability of a discriminator and the data generation capability of a generator; and step 5, obtaining entities and relationships obtained through adversarial neural network training according to the discrimination capability and the data generation capability, and converting the entities and relationships into nodes and edges in the knowledge graph to construct a complete knowledge graph. The defects that the accuracy of entity and relation extraction is reduced, and the construction efficiency and quality of the knowledge graph are reduced are overcome.

Description

Technical Field

[0001] The present invention relates to the field of Internet technology, and in particular to a method for constructing a knowledge graph for identity resolution based on an adversarial neural network. Background Art

[0002] In the Industrial Internet or Internet of Things environment, there are a large number of devices and data resources in different systems, and there are complex upstream and downstream dependencies or functional collaboration relationships between the devices.

[0003] Traditional knowledge graph construction methods, such as rule-based or template-based methods, cannot have higher efficiency and better scalability when processing large-scale data. In key steps such as entity recognition and relationship extraction, they cannot effectively identify and process complex and ambiguous information in the text, which reduces the accuracy of entity and relationship extraction. At the same time, it reduces the efficiency and quality of knowledge graph construction.

[0004] Therefore, the present invention proposes a method for constructing a knowledge graph for identity resolution based on an adversarial neural network. Summary of the invention

[0005] The present invention provides a method for constructing a knowledge graph for identity resolution based on an adversarial neural network, which is used to solve the problems in the prior art that it is unable to have higher efficiency and better scalability when processing large-scale data, and in key steps such as entity recognition and relationship extraction, it is unable to effectively identify and process complex and fuzzy information in the text, which reduces the accuracy of entity and relationship extraction, and at the same time, reduces the construction efficiency and quality of the knowledge graph.

[0006] On the one hand, the present invention provides a method for constructing a knowledge graph for identity resolution based on an adversarial neural network, comprising: Step 1: Acquire various data related to identity resolution, and clean and convert various data related to identity resolution; Step 2: Perform entity recognition on the cleaned and converted data based on the BERT model to obtain entity recognition results, and extract semantic relationships between entities from the text based on the recurrent neural network to obtain relationship extraction results; Step 3: Based on the entity recognition results and relationship extraction results, construct the "entity-relationship-entity" triples to form a preliminary knowledge graph, and use the graph database to store the preliminary knowledge graph; Step 4: Use the preliminary knowledge graph as a data set to train the adversarial neural network, adopt an alternating training strategy and introduce a Self-Attention mechanism to obtain the discriminative ability of the discriminator and the data generation ability of the generator; Step 5: Based on the discriminative ability of the discriminator and the data generation ability of the generator, the entities and relationships trained by the adversarial neural network are obtained and converted into nodes and edges in the knowledge graph to construct a complete knowledge graph.

[0007] According to a method for constructing a knowledge graph of identity resolution based on an adversarial neural network provided by the present invention, various types of data related to identity resolution are obtained, and various types of data related to identity resolution are cleaned and converted, including: Acquire various data related to identity resolution according to the identity database, and form a data set with the various data related to identity resolution; Cleaning and converting the data set, obtaining entity names or other identifying text information other than entity names according to the cleaned and converted data set, and mapping the entity names or other identifying text information other than entity names to a unique entity ID; Determine the relationship between entities based on the unique entity ID; Obtain relationship descriptions from other identifying text information other than entity names, perform standardization processing, and obtain consistent relationship labels.

[0008] According to a method for constructing a knowledge graph for identity resolution based on an adversarial neural network provided by the present invention, the relationship between entities is determined according to the unique entity ID, including: Determine entity name features and entity attribute features according to a unique entity ID, and generate initial entity features according to the entity name features and entity attribute features; Input the initial entity features into a deep neural network based on relational representation to extract the relational features of the entities; Get the Manhattan distance between entities, determine the similarity between entities based on the Manhattan distance, and select the pairwise entities with the highest similarity as alignment entities; Obtain target relationship features between aligned entities, and determine feature loss items based on the difference between the target relationship features and the preset entity relationships; Determine the local relationship feature description component and the global relationship feature description component between the aligned entities according to the feature loss term; Obtain entity representations corresponding to the local relationship feature description component and the global relationship feature description component; Determine the predicted missing relationship of the aligned entity according to the entity representation, and determine the preset retained relationship of the aligned entity according to the predicted missing relationship and the preset relationship table; Obtaining granular feature representation information of each aligned entity, and obtaining node feature representation information between the aligned entities according to the granular feature representation information; Determine the relationship node feature representation information between the aligned entities according to a pre-built adaptive connection matrix based on the relationship type based on the node feature representation information between the aligned entities; The relationship type between the aligned nodes is determined according to the feature representation information of the relationship nodes, and the matching relationship is obtained in the preset retained relationship according to the relationship type.

[0009] According to a method for constructing a knowledge graph for identity resolution based on an adversarial neural network provided by the present invention, other identifying text information of an entity name or a non-entity name is obtained according to a cleaned and converted data set, and the entity name or other identifying text information of a non-entity name is mapped to a unique entity ID, including: Extract all information containing entity names or other identifying text information that is not an entity name from the cleaned and converted data, and determine specific descriptions and attributes of the entity names or other identifying text information that is not an entity name based on all information containing entity names or other identifying text information that is not an entity name; Determine the type and attribute value of other identifying text information of the entity name or non-entity name according to the specific description and attribute of other identifying text information of the entity name or non-entity name, and obtain a unique identifier of each other identifying text information of the entity name or non-entity name; A reverse mapping method is used to map the unique identifier of the entity name or other identifying text information of the non-entity name to the original entity name.

[0010] According to a method for constructing a knowledge graph for identity resolution based on an adversarial neural network provided by the present invention, entity recognition is performed on the cleaned and converted data based on the BERT model to obtain entity recognition results, and semantic relationships between entities are extracted from the text based on a recurrent neural network to obtain relationship extraction results, including: Converting the cleaned and converted data into input text in a unified format, and converting the input text into a sequence of fixed length; Determine the corresponding pre-trained BERT model according to the text features of the input text, and input the converted input text into the pre-trained BERT model; Obtain one or more labels for each word in the input text according to the output result of the BERT model, identify entities in the input text according to the one or more labels, obtain entity recognition results, and perform classification; For the classified entities, the semantic relations between entities are extracted from the text according to the recurrent neural network to obtain the relation extraction results.

[0011] According to a method for constructing a knowledge graph for identity resolution based on an adversarial neural network provided by the present invention, a triple of "entity-relationship-entity" is constructed based on entity recognition results and relationship extraction results to form a preliminary knowledge graph, and the preliminary knowledge graph is stored in a graph database, including: Convert entity recognition results into triples, where each entity contains a unique ID, category, and text description; Convert the relation extraction result into triples, where each triple contains two attributes: antecedent and consequent. Organize the converted entities and relationships into a graph structure, and obtain the nodes corresponding to each entity and the edges corresponding to each relationship; A preliminary knowledge graph is formed according to the nodes corresponding to each entity and the edges corresponding to each relationship, and a graph database is used to store the preliminary knowledge graph.

[0012] According to a method for constructing a knowledge graph for identity resolution based on an adversarial neural network provided by the present invention, the preliminary knowledge graph is used as a data set to train an adversarial neural network, an alternating training strategy is adopted and a Self-Attention mechanism is introduced to obtain the discriminative ability of the discriminator and the data generation ability of the generator, including: Define an adversarial neural network based on a discriminator network and a generator network; When training the discriminator network, the discriminator is used to predict whether the input sample comes from the real knowledge graph. At the same time, the loss function of the discriminator output is optimized based on the cross entropy. When training the generator network, the generator generates samples corresponding to the real knowledge graph data, and the goal of the generator is set to maximize the confusion rate of the discriminator for the samples generated by the generator; An alternating training strategy is adopted and the Self-Attention mechanism is introduced. The discriminative ability of the discriminator and the data generation ability of the generator are obtained based on cross entropy optimization and confusion rate.

[0013] According to a method for constructing a knowledge graph for identity resolution based on an adversarial neural network provided by the present invention, entities and relationships obtained through adversarial neural network training are obtained according to the discriminative ability of a discriminator and the data generation ability of a generator, and are converted into nodes and edges in a knowledge graph to construct a complete knowledge graph, including: According to the discriminative ability of the discriminator and the data generation ability of the generator, the text entities and relations obtained through adversarial neural network training are obtained; Get the locations of text entities and relations, and use clustering algorithms to cluster entities and relations with similar attributes; Get the types of text entities and relations, and define the link relationships between entities and relations; The clustering results and link relationships are converted into nodes and edges in the knowledge graph, and the generated nodes and edges are stored in the knowledge graph database to build a complete knowledge graph.

[0014] According to a method for constructing a knowledge graph for identity resolution based on an adversarial neural network provided by the present invention, before using the preliminary knowledge graph as a data set to train the adversarial neural network, the method further includes: Obtain unlabeled samples in the knowledge graph; Calculate the expected loss rate of unlabeled samples being identified as positive samples and the expected loss rate of negative samples: ; in, It is expressed as the expected loss rate of unlabeled samples being identified as positive samples, Represented as a set of unlabeled triples, It is represented as the total number of unlabeled triples, i is the i-th unlabeled sample, ln is the natural logarithm, It is represented as loss factor, S is represented as decision function, t is represented as timing factor, It is expressed as the partition weight of the i-th unlabeled sample in the normal sample, Expressed as the decision weight factor of the i-th unlabeled sample in the normal sample; ; in, It is expressed as the expected loss rate of unlabeled samples being identified as negative samples, It is expressed as the number of unlabeled triple sets that are positive sample sets, It is expressed as the decision weight factor of the i-th unlabeled sample in the negative sample; Calculate the first difference between the expected loss rate of unlabeled samples identified as positive samples and the expected loss rate of unlabeled samples identified as negative samples; When the first difference is not within the preset difference range, it is judged that the sampling quantity does not meet the standard, and a reminder to resample the negative samples is issued.

[0015] Compared with the prior art, the beneficial effects of this application are as follows: By performing entity recognition and relationship extraction on the preprocessed identity resolution related data, and constructing "entity-relationship-entity" triples, a preliminary knowledge graph is formed. Based on adversarial neural network training, the trained entities and relationships are obtained and converted into nodes and edges in the knowledge graph to construct a complete knowledge graph. It can have higher efficiency and better scalability when processing large-scale data. In key steps such as entity recognition and relationship extraction, it can effectively identify and process complex and ambiguous information in the text, improve the accuracy of entity and relationship extraction, and at the same time, improve the efficiency and quality of knowledge graph construction. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. 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.

[0017] Figure 1 It is a flowchart of a method for constructing a knowledge graph for identity resolution based on an adversarial neural network provided in an embodiment of the present invention; Figure 2 It is a schematic diagram of a process for cleaning and converting various types of data related to identity resolution provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0018] 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.

[0019] Embodiment 1: The embodiment of the present invention provides a method for constructing a knowledge graph for identity resolution based on an adversarial neural network, such as Figure 1 As shown, the method mainly includes the following steps: Step 1: Acquire various data related to identity resolution, and clean and convert various data related to identity resolution; Step 2: Perform entity recognition on the cleaned and converted data based on the BERT model to obtain entity recognition results, and extract semantic relationships between entities from the text based on the recurrent neural network to obtain relationship extraction results; Step 3: Based on the entity recognition results and relationship extraction results, construct the "entity-relationship-entity" triples to form a preliminary knowledge graph, and use the graph database to store the preliminary knowledge graph; Step 4: Use the preliminary knowledge graph as a data set to train the adversarial neural network, adopt an alternating training strategy and introduce a Self-Attention mechanism to obtain the discriminative ability of the discriminator and the data generation ability of the generator; Step 5: Based on the discriminative ability of the discriminator and the data generation ability of the generator, the entities and relationships trained by the adversarial neural network are obtained and converted into nodes and edges in the knowledge graph to construct a complete knowledge graph.

[0020] In this embodiment, identifier resolution refers to the process of uniquely identifying and describing objects, concepts or entities, which is usually used to record and manage information in a database. These identifiers can be strings, numbers, color codes or other forms of symbols.

[0021] In this embodiment, data cleaning includes: Remove missing values: Check whether there is missing identification information in the dataset, such as entity IDs, relationship labels, etc., and perform corresponding processing, such as deleting records with missing values ​​or interpolating with appropriate values.

[0022] Remove duplicate values: For duplicate records with the same identification information, perform deduplication processing to reduce data redundancy and noise.

[0023] Outlier processing: Identify and process abnormal identification information in the data, such as identification values ​​that are incorrectly formatted, do not meet specifications, or exceed a reasonable range.

[0024] In this embodiment, data conversion includes: Text data vectorization: For identification data containing text information (such as entity names, relationship descriptions, etc.), BERT text embedding technology is used to convert it into vector representation for easy processing by neural network models.

[0025] Identification information encoding: For categorized identification information (such as entity type, relationship type, etc.), one-hot encoding is used for encoding.

[0026] In this embodiment, BERT is a pre-trained language representation model used to encode natural language text to achieve natural language understanding tasks such as text classification, named entity recognition, semantic role filling, sentiment analysis, machine translation, and speech recognition.

[0027] In this embodiment, entity recognition is an information extraction technology used to identify entities with specific meanings in text. Entities in the field of identity resolution usually include but are not limited to: companies, equipment, products, and logos. The goal of entity recognition is to find and classify these entities in the text to help understand the subject and context of the text.

[0028] Relation extraction: Relation extraction refers to the process of identifying semantic relationships between entities from text. After completing entity recognition, relationship extraction further analyzes how entities are related to each other. These relationships can be: Interaction relationship: For example, there is interaction between device A and device B; Location relationship: such as equipment or products are located in a warehouse, production line, or geographical area; Ownership relationship: For example, a device or product belongs to a company.

[0029] In this embodiment, a recurrent neural network is usually used to process sequence data, and a "memory" mechanism is used to transfer the hidden state of the previous moment to the hidden state of the current moment, thereby realizing relationship extraction of the sequence data.

[0030] In this embodiment, the "entity-relationship-entity" triple is a representation data structure used to describe the relationship between entities. The "entity-relationship-entity" triple can be written as ER-E or ERE. For example, "city-owns-university" can be written as "city E -- R -- university E".

[0031] In this embodiment, the graph database is a database based on a graph model, which can efficiently store, query and manage graph data. The graph database uses a graphical method to represent the relationship between entities, rather than using tables and columns. Graph databases usually use triangular tables to represent the relationship between entities, which is called a graph. Each node (also called a vertex) in the graph represents an entity, and the edge represents the relationship between entities. Common graph databases include Neo4j, OrientDB, and ArangoDB.

[0032] In this embodiment, the knowledge graph is a system that organizes and expresses human knowledge and relationships in a graphical way, organizing various forms of knowledge such as language, facts, and concepts into organized structures and relationships. The knowledge graph consists of multiple components, including entities, attributes, relationships, types, and ontologies. Entities are the basic units in the knowledge graph and represent things or concepts in the real world. Attributes are a set of part-of-speech words that describe entities, such as human occupations, gender, and date of birth. Relationships are connections between entities, used to describe the relationship between entities, such as the relationship between students and teachers, and the relationship between cars and drivers. Types are a set of words that classify entities, such as people, places, and time. Ontology is a set of rules that define entities, attributes, and relationships, used to describe the structure and semantics of the knowledge graph.

[0033] In this embodiment, the basic architecture of the adversarial neural network includes two main components: a generator and a discriminator. These two components compete with each other to form a dynamic "game". This mechanism is derived from the concept of "zero-sum game" in game theory, in which the gain of one party means the loss of the other party.

[0034] Generator: The task of the generator is to generate data samples from random noise, which should mimic the distribution of real data as closely as possible. It receives random noise as input and outputs a sample that is designed to fool the discriminator into thinking it is real.

[0035] Discriminator: The discriminator is a classifier whose goal is to distinguish between real data and fake data produced by the generator. It receives a data sample as input and outputs a probability value indicating the probability that the sample is real data.

[0036] In this embodiment, the Self-Attention mechanism is used to extract local dependencies in sequence data.

[0037] The beneficial effects of the above technical solution are: by performing entity recognition and relationship extraction on the pre-processed identity resolution related data, and constructing the "entity-relationship-entity" triples, a preliminary knowledge graph is formed, and training is performed based on the adversarial neural network to obtain the trained entities and relationships, and convert them into nodes and edges in the knowledge graph to construct a complete knowledge graph. It can have higher efficiency and better scalability when processing large-scale data. In key steps such as entity recognition and relationship extraction, it can effectively identify and process complex and ambiguous information in the text, improve the accuracy of entity and relationship extraction, and at the same time, improve the efficiency and quality of knowledge graph construction.

[0038] Embodiment 2: Based on Example 1, the embodiment of the present invention obtains various types of data related to identity resolution, and cleans and converts the various types of data related to the identity resolution, such as Figure 2 As shown, including: S01: acquiring various data related to identity resolution according to an identity database, and forming a data set with the various data related to identity resolution; S02: performing data cleaning and conversion on the data set, obtaining entity names or other identifying text information of non-entity names according to the cleaned and converted data set, and mapping the entity names or other identifying text information of non-entity names to a unique entity ID; S03: Determine the relationship between entities according to the unique entity ID; S04: Obtain relationship descriptions in other identifying text information other than entity names, perform standardization processing, and obtain consistent relationship labels.

[0039] In this embodiment, the identification database is a type of database used to store and manage identity information of entity objects. The identification database usually includes a primary key for uniquely identifying each entity object, and other attributes for describing detailed information of the entity object.

[0040] In this embodiment, the data set is a group of organized data samples. The data set is usually composed of a group of data, and the data set can be numbers or texts.

[0041] In this embodiment, an entity refers to a thing or concept with independent meaning in the real world, which can be a concrete object, an abstract concept, or a natural phenomenon, such as a table, a car, a country, or music.

[0042] In this embodiment, other identifying text information may be: binary code text information, picture and video text information, and audio text information.

[0043] In this embodiment, the relationship description means that the text information of the non-entity name usually also includes some relationship descriptions to describe the relationship between different concepts. These relationship descriptions can be simple words or phrases, or more complex sentences or paragraphs. For example, when mentioning the relationship between one person and another person in a document, simple relationship descriptions such as "are friends" or "are brothers and sisters" can be used, and more specific descriptions can be used to express the relationship, such as "she is his mother."

[0044] In this embodiment, entity ID mapping is a method of establishing an association between unstructured data such as objects, concepts, entities in the real world and specific records in a computer database. In entity ID mapping, each entity is assigned a unique ID, which is usually relative to a certain reference system. This reference system can be physical, such as a geographic location, postal code, or logical, such as a product model or customer number.

[0045] The beneficial effects of the above technical solution are: data cleaning and conversion of various data sets related to identity resolution, obtaining entity names or other identifying text information, and mapping them to unique entity IDs can improve data consistency and accuracy, and help to better understand and analyze data.

[0046] Embodiment 3: Based on Example 2, the embodiment of the present invention determines the relationship between entities according to the unique entity ID, including: Determine entity name features and entity attribute features according to the unique entity ID, and generate initial entity features according to the entity name features and entity attribute features; Input the initial entity features into a deep neural network based on relational representation to extract the relational features of the entities; Get the Manhattan distance between entities, determine the similarity between entities based on the Manhattan distance, and select the pairwise entities with the highest similarity as alignment entities; Obtain target relationship features between aligned entities, and determine feature loss items based on the difference between the target relationship features and the preset entity relationships; Determine the local relationship feature description component and the global relationship feature description component between the aligned entities according to the feature loss term; Obtain entity representations corresponding to the local relationship feature description component and the global relationship feature description component; Determine the predicted missing relationship of the aligned entity according to the entity representation, and determine the preset retained relationship of the aligned entity according to the predicted missing relationship and the preset relationship table; Obtaining granular feature representation information of each aligned entity, and obtaining node feature representation information between the aligned entities according to the granular feature representation information; Determine the relationship node feature representation information between the aligned entities according to a pre-built adaptive connection matrix based on the relationship type based on the node feature representation information between the aligned entities; The relationship type between the aligned nodes is determined according to the feature representation information of the relationship nodes, and the matching relationship is obtained in the preset retained relationship according to the relationship type.

[0047] In this embodiment, the entity name feature refers to information or features that can be obtained based on the name or title of the entity, such as: Entity Type: describes the kind of entity, such as product, service, location.

[0048] Entity name: A name or title that describes the entity itself, such as "Apple", "Starbucks Coffee", and "Beijing Sanlitun".

[0049] Entity alias: describes the alias or nickname of the entity in other contexts, such as "headquarters building" or "the world's largest bookstore chain."

[0050] Primary purpose of the entity: Describes the primary purpose or function of the entity, such as "mobile phone", "TV remote control", "bus route map".

[0051] Entity features: describe some features of the entity, such as "apple" is a red oval fruit, "TV remote control" has multiple buttons, and "bus route map" shows the bus route.

[0052] In this embodiment, entity attribute characteristics refer to a set of properties or characteristics that describe an entity, such as: Geographic location features: Features that describe the spatial location of an entity, such as longitude, latitude, and altitude.

[0053] Physical property characteristics: Characteristics that describe the physical properties of an entity, such as size, shape, color, and density.

[0054] Temporal attribute characteristics: Characteristics that describe the temporal attributes of an entity, such as production date, service life, and expiration date.

[0055] Social attribute characteristics: describe the characteristics of entities in terms of social relationships, such as brand, manufacturer, distributor, and user group.

[0056] In this embodiment, the initial entity features refer to the basic attributes of the entity when it is created, that is, the most basic features of the entity, such as a person's name or a place name.

[0057] In this embodiment, the deep neural network based on relational representation is a neural network model that can learn the relationship between entities, and can encode the semantic information between entities into the output of the neural network. The deep neural network based on relational representation usually adopts a multi-layer perceptron as the main neural network structure, and takes the relationship between entities as input, and encodes the relationship between entities into the output of the neural network, such as: bidirectional Gated Recurrent Neural Network, Graph Convolutional Network.

[0058] In this embodiment, extracting the relational features of entities refers to automatically identifying the associations between entities from unstructured data. The relational features of entities refer to the relationships between entities, such as: There is an "owns" relationship between entity A and entity B, which means that entity B belongs to entity A.

[0059] There is a "contains" relationship between entity A and entity B, which means that entity A is contained in entity B.

[0060] There is a "temporal before" relationship between entity A and entity B, which means that entity B occurs after entity A.

[0061] In this embodiment, the Manhattan distance between entities refers to the shortest path distance between two entities, that is, the actual physical distance between the entities.

[0062] In this embodiment, aligning entities refers to placing multiple objects accurately on the same plane or making their positions and orientations meet certain specific requirements.

[0063] In this embodiment, the target relationship feature between aligned entities means that when processing a group of entities, the relationship between them needs to be determined. For example, in a shopping list application, the relationship between the product and the purchaser needs to be displayed together with their names. Therefore, the relationship between the product and the purchaser needs to be determined (for example, "the purchaser bought this product") and associated with the product name.

[0064] In this embodiment, the feature loss term refers to the phenomenon that feature representation is lost or weakened due to insufficient data, noise interference, etc. when using a machine learning algorithm to train a model.

[0065] In this embodiment, the local relationship feature description component between aligned entities refers to the attributes or variables used to describe the local relationship features between entities. These features may affect aspects such as the behavior, interaction, or appearance of the entities. For example, for a robot, its movement speed may be affected by factors such as battery power, terrain complexity, and weather. Therefore, the local relationship feature description component between aligned entities may be: Position information: The position information of an entity in three-dimensional space, including coordinates, height, angle, etc.

[0066] Pose information: information about the position and orientation of an entity relative to other objects or scenes in its reference frame.

[0067] Motion information: information about the entity’s position change over time, including velocity, acceleration, and displacement.

[0068] In this embodiment, the global relationship feature description component between aligned entities refers to describing the global relationship between two or more entities, such as: Common ancestor: Two entities share the same parent concept or classification.

[0069] Similarity: Two entities have similar characteristics or attributes; they may belong to the same category or be in a similar environment.

[0070] Dependency: One entity depends on the existence or behavior of another entity.

[0071] Interaction: There is an interaction between two entities, and they may affect each other's behavior or state.

[0072] Coexistence: Two entities can coexist in the same environment and they may not have direct dependencies.

[0073] In this embodiment, entity representation refers to a way of describing an entity, linking the entity with its relevant features and attributes. The entity representation can be a text description, a graphical representation or other, such as: Text description: Use natural language or other character languages ​​to describe the attributes and characteristics of an entity, such as "a human is an animal that walks upright."

[0074] Graphical representation: Use icons, lines, shapes, etc. to represent the attributes and characteristics of entities, such as using a rectangle to represent a room, a circle to represent a table, and a triangle to represent a chair.

[0075] In this embodiment, the predicted missing relationship between entities refers to the situation where, given an entity, there may be a lack of certain features or attributes between other entities and the entity. For example, given a person entity, it can be predicted that there is a lack of certain attributes between other entities (such as places, organizations) and the person entity, such as time, place, and organizational structure.

[0076] In this embodiment, the preset reservation relationship between entities refers to a logically necessary relationship between entities that describes the relationship that must exist between the entities. For example, in a reservation system, the relationship between a person entity ("guest") and an order entity ("order") can be defined as "order is related to guest", which means that each order must be associated with a specific guest, and the relationship between guests and orders is a one-to-many relationship, which can be achieved by associating one guest with multiple orders, but one order cannot be associated with multiple guests.

[0077] In this embodiment, the granularity characteristics of an entity refer to information describing the most basic attributes and states of the entity, such as: Human entities: name, gender, date of birth.

[0078] Vehicle entity: make, model, color, year.

[0079] Order entity: order number, customer name, product list, total price.

[0080] Address entity: address, city, state, country.

[0081] In this embodiment, the adaptive connection matrix based on relationship type means that a relationship type needs to be defined for each relationship type first, which can be achieved by associating each relationship with a known parent class or keyword. For example, the "friend" relationship can be associated with the "relationship" type in the "social network", and then a connection matrix representing the relationship between entities is constructed. This matrix is ​​a two-dimensional array, in which each row and each column represents a relationship type. Each element in the matrix is ​​associated with a corresponding relationship type and indicates whether the type appears in any relationship of the current entity. If an entity does not have a specific relationship type associated with a certain relationship, a value of "none" or "undetermined" is placed in the corresponding position.

[0082] In this embodiment, aligning node feature representation information between entities refers to using nodes to represent entities when drawing an entity relationship diagram or chart. A node is an abstract representation of an entity and can be used to describe the relationship between entities.

[0083] In this embodiment, the relationship node feature representation information between the aligned entities refers to the visual elements representing the relationship between the entities in the graphical user interface, such as: Directed edge: Draw a directed edge connecting two entities on the screen. The edge can use arrows or other symbols to represent the relationship between the entities.

[0084] Circle Node: Draw a circle node on the screen to represent an endpoint of an entity. When the user clicks the node, they can view the entities related to it and the relationship between them.

[0085] In this embodiment, an adaptive connection matrix based on relationship type is used to map the aligned entities to the node feature representation information between them, and the relationship node feature representation information between the aligned entities is determined. For each entity, we need to find the adaptive connection matrix cell associated with it and extract the feature vector and other information related to the entity from the cell.

[0086] The beneficial effects of the above technical solution are: determining the relationship type between the aligned nodes based on the relationship node feature representation information, obtaining the matching relationship in the preset retained relationship based on the relationship type, and using predefined relationship types to help quickly locate the associated entities. At the same time, it can avoid the trouble of manually arranging all entities and ensure that the relationship between entities is always displayed correctly.

[0087] Embodiment 4: Based on Example 3, the embodiment of the present invention obtains other identifying text information of the entity name or non-entity name according to the cleaned and converted data set, and maps the entity name or other identifying text information of the non-entity name to a unique entity ID, including: Extract all information containing entity names or other identifying text information that is not an entity name from the cleaned and converted data, and determine specific descriptions and attributes of the entity names or other identifying text information that is not an entity name based on all information containing entity names or other identifying text information that is not an entity name; Determine the type and attribute value of other identifying text information of the entity name or non-entity name according to the specific description and attribute of other identifying text information of the entity name or non-entity name, and obtain a unique identifier of each other identifying text information of the entity name or non-entity name; A reverse mapping method is used to map the unique identifier of the entity name or other identifying text information of the non-entity name to the original entity name.

[0088] In this embodiment, the information containing entity names refers to extracting named entities from the text. These entities can be names of people, places, names of organizations, proper nouns, and their values ​​are usually unique and related to the context. For example, in the sentence "Dr. Johnson and his research team are studying climate change", "Dr. Johnson" is an entity and its value is Dr. Johnson.

[0089] In this embodiment, the attributes of an entity refer to the characteristics or properties of the entity. For example, in a named entity recognition system, the attributes of an entity may include: name, position, and professional field.

[0090] In this embodiment, the entity type refers to the category to which the entity belongs. The entity type can be used to classify and organize entities. The entity type is usually divided according to certain rules. For example, it can be classified according to factors such as the function, purpose, and time range of the entity. Common entity types include: person name entity, place name entity, and organizational entity.

[0091] In this embodiment, the entity attribute value refers to the characteristics or description of the entity. The entity attribute value can be data in the form of specific numbers, text, dates, etc., or it can be the performance of the entity on certain indicators. For example, if the entity is a telephone number, its entity attribute value is the actual value of the number; if the entity is a person's name, his entity attribute value is his full name.

[0092] The beneficial effects of the above technical solution are: determining the specific description and attributes of the entity based on the information containing the entity name, and associating it with the entity type and attribute value to obtain a unique identifier for each entity. The unique identifier can reduce errors caused by similar entity names. At the same time, it can ensure the uniqueness and consistency of the entity, which helps to improve the quality and reliability of the data.

[0093] Embodiment 5: Based on Example 4, the embodiment of the present invention performs entity recognition on the cleaned and converted data based on the BERT model to obtain entity recognition results, and extracts semantic relationships between entities from the text according to a recurrent neural network to obtain relationship extraction results, including: Converting the cleaned and converted data into input text in a unified format, and converting the input text into a sequence of fixed length; Determine the corresponding pre-trained BERT model according to the text features of the input text, and input the converted input text into the pre-trained BERT model; Obtain one or more labels for each word in the input text according to the output result of the BERT model, identify entities in the input text according to the one or more labels, obtain entity recognition results, and perform classification; For the classified entities, the semantic relations between entities are extracted from the text according to the recurrent neural network to obtain the relation extraction results.

[0094] In this embodiment, the input text in a unified format refers to a group of text data with the same structure and attributes, and these data are all about the same subject or field, such as product description, service description or organizational charter.

[0095] In this embodiment, the fixed-length sequence refers to a sequence with a fixed length and a fixed number of elements, the position of each element in the sequence is fixed, and the intervals between the elements are also the same.

[0096] In this embodiment, text features refer to some key information extracted from the text, which can be used to describe the content and semantics of the text.

[0097] In this embodiment, the BERT model is a pre-trained language representation model that can perform self-supervised pre-training through a large amount of unlabeled data and generate general language understanding capabilities.

[0098] The beneficial effects of the above technical solution are: the preprocessed data is converted into a sequence of fixed length and input into the corresponding pre-trained BERT model for entity recognition. The semantic relationship between entities is determined based on the recurrent neural network through the identified entities, which can more efficiently extract useful information, quickly determine and understand the semantic relationship between each entity, and at the same time, can convert unstructured information into structured data, which helps to build a high-quality knowledge graph and enhance the searchability and scalability of the knowledge graph.

[0099] Embodiment 6: Based on Example 5, the embodiment of the present invention constructs a triple of "entity-relationship-entity" based on the entity recognition results and the relationship extraction results to form a preliminary knowledge graph, and uses a graph database to store the preliminary knowledge graph, including: Convert entity recognition results into triples, where each entity contains a unique ID, category, and text description; Convert the relation extraction result into triples, where each triple contains two attributes: antecedent and consequent. Organize the converted entities and relationships into a graph structure, and obtain the nodes corresponding to each entity and the edges corresponding to each relationship; A preliminary knowledge graph is formed according to the nodes corresponding to each entity and the edges corresponding to each relationship, and a graph database is used to store the preliminary knowledge graph.

[0100] In this embodiment, the entity recognition result is converted into a triple form. For example, if the entity recognition result is "person", it can be represented as (1, "human", "John Doe").

[0101] In this embodiment, the relationship extraction result is converted into a triple form. For example, for the relationship "purchase-commodity", it can be expressed as (("person", "purchase"), ("person", "commodity")).

[0102] In this embodiment, the node corresponding to the entity may be a simple representation of the entity or relationship.

[0103] In this embodiment, an edge corresponding to each relationship may connect two nodes and represent the relationship between them.

[0104] The beneficial effect of the above technical solution is: converting the entity recognition results and relationship extraction results into triples, and obtaining the corresponding nodes and edges to form a preliminary knowledge graph, which can quickly locate the relevant information of a specific entity and save the time of searching in the text.

[0105] Embodiment 7: Based on Example 6, this embodiment of the present invention uses the preliminary knowledge graph as a data set to train an adversarial neural network, adopts an alternating training strategy and introduces a Self-Attention mechanism to obtain the discriminant ability of the discriminator and the data generation ability of the generator, including: Define an adversarial neural network based on a discriminator network and a generator network; When training the discriminator network, the discriminator is used to predict whether the input sample comes from the real knowledge graph. At the same time, the loss function of the discriminator output is optimized based on the cross entropy. When training the generator network, the generator generates samples corresponding to the real knowledge graph data, and the goal of the generator is set to maximize the confusion rate of the discriminator for the samples generated by the generator; An alternating training strategy is adopted and the Self-Attention mechanism is introduced. The discriminative ability of the discriminator and the data generation ability of the generator are obtained based on cross entropy optimization and confusion rate.

[0106] In this embodiment, the input of the discriminator network is a randomly generated sample, and the output is the probability of whether the sample comes from the real knowledge graph.

[0107] In this embodiment, the input of the generator network is samples in the real knowledge graph, and the output is the probability of different samples of the same sample seen in the real knowledge graph.

[0108] In this embodiment, cross entropy is a commonly used machine learning loss function, which is used to measure the gap between the output predicted by the model during training and the actual label. It is usually used in classification problems, where the goal of the model is to minimize the cross entropy between the predicted output and the actual label.

[0109] In this embodiment, optimization refers to minimizing the loss function using an optimization algorithm such as gradient descent. During the optimization process, an early stopping strategy may be adopted. Early stopping may prevent the model from overfitting the training data, which may lead to poor performance in testing and actual applications.

[0110] In this embodiment, the confusion rate is a useful data for evaluating the performance of a classifier, and the actual label is compared with the predicted label to determine the accuracy of the classifier.

[0111] The beneficial effects of the above technical solution are: defining an adversarial neural network based on a discriminator network and a generator network, and training the discriminator network and the generator network, obtaining the discriminant ability of the discriminator and the data generation ability of the generator according to cross entropy optimization and confusion rate, which can improve the robustness and generalization ability of the model, so that the subsequently constructed knowledge graph can still maintain high accuracy and stability when facing noisy data or abnormal situations.

[0112] Embodiment 8: Based on Example 7, the embodiment of the present invention obtains entities and relationships obtained through adversarial neural network training according to the discriminant ability of the discriminator and the data generation ability of the generator, and converts them into nodes and edges in the knowledge graph to construct a complete knowledge graph, including: According to the discriminative ability of the discriminator and the data generation ability of the generator, the text entities and relations obtained through adversarial neural network training are obtained; Get the location of text entities and relations and perform clustering; Get the types of text entities and relations, and define the link relationships between entities and relations; The clustering results and link relationships are converted into nodes and edges in the knowledge graph, and the generated nodes and edges are stored in the knowledge graph database to build a complete knowledge graph.

[0113] In this embodiment, the positions of text entities and relations refer to the positions of the entities in the text and the starting points of the relations associated therewith.

[0114] In this embodiment, the types of entities and relationships refer to the specific meanings they represent. Entities can be of various types, such as names of people, places, and organization names, while relationships can represent logical relationships between entities, such as time relationships, spatial relationships, and causal relationships. Depending on the application requirements, the types of entities and relationships may also vary. For example, in a system that represents news reports, entity types may include names of people, places, and organization names, while relationship types may include event relationships and time relationships.

[0115] In this embodiment, the link relationship between entities refers to the association between two or more entities, and this association can be realized by direct connection or indirect connection between entities. For example, there is a link relationship between a person's name and another person's name, because their names are both used to identify a person. Similarly, there may be a link relationship between a city and a country, because they are both different ways of identifying a region.

[0116] In this embodiment, converting the clustering results into nodes in the knowledge graph means that each location information can be converted into a place node in the knowledge graph, creating a place node associated with the location in the graph, and marking it as a specific place or geographic area. For example, assuming that we have a knowledge graph that represents business activities in a city, the location of each business activity can be converted into a place node in the knowledge graph and marked as a specific city in the graph.

[0117] The beneficial effects of the above technical solution are: by obtaining the location and type of entity text and relationships, determining link relationships and clustering, and constructing a complete knowledge graph, the efficiency of knowledge graph construction can be improved, and at the same time, the intelligence level of identity resolution can be improved.

[0118] Embodiment 9: Based on Example 8, before using the preliminary knowledge graph as a data set to train the adversarial neural network, the embodiment of the present invention further includes: Obtain unlabeled samples in the knowledge graph; Calculate the expected loss rate of unlabeled samples being identified as positive samples and the expected loss rate of negative samples: ; in, It is expressed as the expected loss rate of unlabeled samples being identified as positive samples, Represented as a set of unlabeled triples, It is represented as the total number of unlabeled triples, i is the i-th unlabeled sample, ln is the natural logarithm, It is represented as loss factor, S is represented as decision function, t is represented as timing factor, It is expressed as the partition weight of the i-th unlabeled sample in the normal sample, Expressed as the decision weight factor of the i-th unlabeled sample in the normal sample; ; in, It is expressed as the expected loss rate of unlabeled samples being identified as negative samples, It is expressed as the number of unlabeled triple sets that are positive sample sets, It is expressed as the decision weight factor of the i-th unlabeled sample in the negative sample; Calculate the first difference between the expected loss rate of unlabeled samples identified as positive samples and the expected loss rate of unlabeled samples identified as negative samples; When the first difference is not within the preset difference range, it is judged that the sampling quantity does not meet the standard, and a reminder to resample the negative samples is issued.

[0119] In this embodiment, in the knowledge graph, unlabeled samples refer to instances that do not have clearly labeled category information. These unlabeled samples may not be labeled due to multiple reasons, such as: omissions in the data collection process, data labeling errors, and uncertainties in the data generation process.

[0120] In this embodiment, the expected loss rate of positive samples refers to the probability of the model correctly predicting an instance when it is a positive sample minus the probability of misclassifying the instance as a negative sample, which describes the balance between the benefits of correctly identifying positive samples and the losses caused by misclassifying them as negative samples.

[0121] In this embodiment, the expected loss rate of negative samples refers to the probability of the model correctly predicting a certain instance when it is a negative sample minus the probability of misclassifying the instance as a positive sample, which describes the balance between the benefits of correctly classifying a negative sample as a negative sample and the losses caused by misclassifying it as a positive sample.

[0122] In this embodiment, the probability difference refers to the difference in probability between two events.

[0123] The beneficial effect of the above technical solution is that obtaining the probability difference between the expected loss rate of unlabeled samples identified as positive samples and the expected loss rate of unlabeled samples identified as negative samples can better balance the number of positive and negative samples and avoid the situation where certain categories are ignored.

[0124] 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, 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.

[0125] 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. A method for constructing a knowledge graph for identity resolution based on an adversarial neural network, characterized in that: include: Step 1: Acquire various data related to identity resolution, and clean and convert various data related to identity resolution; Step 2: Perform entity recognition on the cleaned and converted data based on the BERT model to obtain entity recognition results, and extract semantic relationships between entities from the text based on the recurrent neural network to obtain relationship extraction results; Step 3: Based on the entity recognition results and relationship extraction results, construct "entity-relationship-entity" triples to form a preliminary knowledge graph, and use the graph database to store the preliminary knowledge graph; Step 4: Use the preliminary knowledge graph as a data set to train the adversarial neural network, adopt an alternating training strategy and introduce a Self-Attention mechanism to obtain the discriminative ability of the discriminator and the data generation ability of the generator; Step 5: Based on the discriminative ability of the discriminator and the data generation ability of the generator, the entities and relationships trained by the adversarial neural network are obtained and converted into nodes and edges in the knowledge graph to construct a complete knowledge graph.

2. The method for constructing a knowledge graph for identity resolution based on an adversarial neural network according to claim 1, characterized in that: Obtain various types of data related to identity resolution, and clean and convert various types of data related to identity resolution, including: Acquire various data related to identity resolution according to the identity database, and form a data set with the various data related to identity resolution; Cleaning and converting the data set, obtaining entity names or other identifying text information other than entity names according to the cleaned and converted data set, and mapping the entity names or other identifying text information other than entity names to a unique entity ID; Determine the relationship between entities based on the unique entity ID; Obtain relationship descriptions from other identifying text information other than entity names, perform standardization processing, and obtain consistent relationship labels.

3. The method for constructing a knowledge graph for identity resolution based on an adversarial neural network according to claim 2, characterized in that: Determining the relationship between entities according to the unique entity ID includes: Determine entity name features and entity attribute features according to a unique entity ID, and generate initial entity features according to the entity name features and entity attribute features; Input the initial entity features into a deep neural network based on relational representation to extract the relational features of the entities; Get the Manhattan distance between entities, determine the similarity between entities based on the Manhattan distance, and select the pairwise entities with the highest similarity as alignment entities; Obtain target relationship features between aligned entities, and determine feature loss items based on the difference between the target relationship features and the preset entity relationships; Determine the local relationship feature description component and the global relationship feature description component between the aligned entities according to the feature loss term; Obtain entity representations corresponding to the local relationship feature description component and the global relationship feature description component; Determine the predicted missing relationship of the aligned entity according to the entity representation, and determine the preset retained relationship of the aligned entity according to the predicted missing relationship and the preset relationship table; Obtaining granular feature representation information of each aligned entity, and obtaining node feature representation information between the aligned entities according to the granular feature representation information; Determine the relationship node feature representation information between the aligned entities according to a pre-built adaptive connection matrix based on the relationship type based on the node feature representation information between the aligned entities; The relationship type between the aligned nodes is determined according to the feature representation information of the relationship nodes, and the matching relationship is obtained in the preset retained relationship according to the relationship type.

4. The method for constructing a knowledge graph for identity resolution based on an adversarial neural network according to claim 2, characterized in that: Obtain entity names or other identifying text information of non-entity names based on the cleaned and transformed data set, and map entity names or other identifying text information of non-entity names to unique entity IDs, including: Extract all information containing entity names or other identifying text information that is not an entity name from the cleaned and converted data, and determine specific descriptions and attributes of the entity names or other identifying text information that is not an entity name based on all information containing entity names or other identifying text information that is not an entity name; Determine the type and attribute value of other identifying text information of the entity name or non-entity name according to the specific description and attribute of other identifying text information of the entity name or non-entity name, and obtain a unique identifier of each other identifying text information of the entity name or non-entity name; A reverse mapping method is used to map the unique identifier of the entity name or other identifying text information of the non-entity name to the original entity name.

5. The method for constructing a knowledge graph for identity resolution based on an adversarial neural network according to claim 1, characterized in that: The cleaned and converted data is used to perform entity recognition on the text based on the BERT model to obtain entity recognition results, and the semantic relationship between entities is extracted from the text based on the recurrent neural network to obtain the relationship extraction results, including: Converting the cleaned and converted data into input text in a unified format, and converting the input text into a sequence of fixed length; Determine the corresponding pre-trained BERT model according to the text features of the input text, and input the converted input text into the pre-trained BERT model; Obtain one or more labels for each word in the input text according to the output result of the BERT model, identify entities in the input text according to the one or more labels, obtain entity recognition results, and perform classification; For the classified entities, the semantic relations between entities are extracted from the input text according to the recurrent neural network to obtain the relation extraction results.

6. The method for constructing a knowledge graph for identity resolution based on an adversarial neural network according to claim 1, characterized in that: Based on the entity recognition results and relationship extraction results, the "entity-relationship-entity" triples are constructed to form a preliminary knowledge graph, and the preliminary knowledge graph is stored in a graph database, including: Convert entity recognition results into triples, where each entity contains a unique ID, category, and text description; Convert the relation extraction result into triples, where each triple contains two attributes: antecedent and consequent. Organize the converted entities and relationships into a graph structure, and obtain the nodes corresponding to each entity and the edges corresponding to each relationship; A preliminary knowledge graph is formed according to the nodes corresponding to each entity and the edges corresponding to each relationship, and a graph database is used to store the preliminary knowledge graph.

7. The method for constructing a knowledge graph for identity resolution based on an adversarial neural network according to claim 1, characterized in that: The preliminary knowledge graph is used as a data set to train the adversarial neural network, using an alternating training strategy and introducing the Self-Attention mechanism to obtain the discriminant ability of the discriminator and the data generation ability of the generator, including: Define an adversarial neural network based on a discriminator network and a generator network; When training the discriminator network, the discriminator is used to predict whether the input sample comes from the real knowledge graph. At the same time, the loss function of the discriminator output is optimized based on the cross entropy. When training the generator network, the generator generates samples corresponding to the real knowledge graph data, and the goal of the generator is set to maximize the confusion rate of the discriminator for the samples generated by the generator; An alternating training strategy is adopted and the Self-Attention mechanism is introduced. The discriminative ability of the discriminator and the data generation ability of the generator are obtained based on cross entropy optimization and confusion rate.

8. The method for constructing a knowledge graph for identity resolution based on an adversarial neural network according to claim 1, characterized in that: According to the discriminant ability of the discriminator and the data generation ability of the generator, the entities and relationships obtained through adversarial neural network training are obtained and converted into nodes and edges in the knowledge graph to construct a complete knowledge graph, including: According to the discriminative ability of the discriminator and the data generation ability of the generator, the text entities and relations obtained through adversarial neural network training are obtained; Get the location of text entities and relations and perform clustering; Get the types of text entities and relations, and define the link relationships between entities and relations; The clustering results and link relationships are converted into nodes and edges in the knowledge graph, and the generated nodes and edges are stored in the knowledge graph database to build a complete knowledge graph.

9. The method for constructing a knowledge graph for identity resolution based on an adversarial neural network according to claim 1, characterized in that: Before using the preliminary knowledge graph as a data set to train the adversarial neural network, it also includes: Obtain unlabeled samples in the knowledge graph; Calculate the expected loss rate of unlabeled samples being identified as positive samples and the expected loss rate of negative samples: ； in, It is expressed as the expected loss rate of unlabeled samples being identified as positive samples, Represented as a set of unlabeled triples, It is represented as the total number of unlabeled triples, i is the i-th unlabeled sample, ln is the natural logarithm, It is represented as loss factor, S is represented as decision function, t is represented as timing factor, It is expressed as the partition weight of the i-th unlabeled sample in the normal sample, Expressed as the decision weight factor of the i-th unlabeled sample in the normal sample; ； in, It is expressed as the expected loss rate of unlabeled samples being identified as negative samples, It is expressed as the number of unlabeled triple sets that are positive sample sets, It is expressed as the decision weight factor of the i-th unlabeled sample in the negative sample; Calculate the first difference between the expected loss rate of unlabeled samples identified as positive samples and the expected loss rate of unlabeled samples identified as negative samples; When the first difference is not within the preset difference range, it is judged that the sampling quantity does not meet the standard, and a reminder to resample the negative samples is issued.

Images