Entity identification method, apparatus and device, readable storage medium and program product

Abstract

The embodiment of the invention provides an entity recognition method and device, equipment, a readable storage medium and a program product. The entity recognition method comprises the following steps: calling an entity analysis model to perform text feature coding processing on an address description text associated with a to-be-queried address to obtain feature information corresponding to the address description text; calling an entity analysis model to perform entity feature analysis processing on feature information corresponding to the address to be queried and feature information corresponding to the address description text in the feature information corresponding to the address description text to obtain entity feature information of a target entity corresponding to the address to be queried; and determining an entity recognition result of the target entity according to the entity feature information, wherein the entity recognition result comprises one or two of an entity name and an entity type. By adopting the embodiment of the invention, the entity name and the entity type corresponding to the address can be determined by utilizing the entity analysis model, so that the processing efficiency of the entity identification method is effectively improved, and the accuracy of an entity identification result is also improved.

Description

Technical Field

[0001] This application relates to the field of computer technology, specifically to an entity recognition method, an entity recognition device, a computer equipment, a computer-readable storage medium, and a computer program product. Background Technology

[0002] With the development of computer technology, addresses are increasingly involved in various application scenarios, such as logistics, where it is necessary to obtain the addresses of the sender and receiver. To ensure resource security and address accuracy, entity determination processing is typically performed on the address to identify the entity name and type to which it belongs.

[0003] Typically, entity extraction tools are used to process the associated text of the address to be determined, obtaining the entity name of the entity to which the address belongs. Then, a classification tool is used to process this entity name to obtain the entity type of the entity to which the address belongs. This method requires the use of multiple different tools to determine the entity name and entity type, making the process relatively complex. It also involves data transmission between different tools, which is prone to propagation errors, resulting in low accuracy of the processing results. Summary of the Invention

[0004] This application provides entity recognition methods, apparatus, devices, readable storage media, and program products, which can use entity analysis models to determine the entity name and entity type corresponding to an address, effectively improving the processing efficiency of entity recognition methods and also helping to improve the accuracy of entity recognition results.

[0005] On one hand, embodiments of this application provide an entity recognition method, which includes:

[0006] Retrieve the address description text associated with the address to be queried;

[0007] The entity analysis model is invoked to perform text feature encoding on the address description text to obtain feature information corresponding to the address description text. The feature information corresponding to the address description text includes feature information corresponding to the address to be queried. The entity analysis model is obtained by training an initial analysis model using difference data. The difference data is determined based on sample entity information and predicted entity information of the sample address. The sample entity information includes a reference entity name and a reference entity type. The predicted entity information includes a predicted entity name and a predicted entity type obtained by processing the address description text of the sample address using the initial analysis model.

[0008] The entity analysis model is invoked to perform entity feature analysis on the feature information corresponding to the address to be queried and the feature information corresponding to the address description text, so as to obtain the entity feature information of the target entity corresponding to the address to be queried;

[0009] The entity recognition result of the target entity is determined based on the entity feature information, and the entity recognition result includes one or both of the entity name and entity type.

[0010] Accordingly, embodiments of this application provide an entity recognition device, which includes:

[0011] The acquisition unit is used to acquire the address description text associated with the address to be queried;

[0012] The processing unit is configured to invoke an entity analysis model to perform text feature encoding processing on the address description text to obtain feature information corresponding to the address description text. The feature information corresponding to the address description text includes feature information corresponding to the address to be queried. The entity analysis model is obtained by training an initial analysis model using difference data. The difference data is determined based on sample entity information and predicted entity information of the sample address. The sample entity information includes a reference entity name and a reference entity type. The predicted entity information includes a predicted entity name and a predicted entity type obtained by processing the address description text of the sample address using the initial analysis model.

[0013] The processing unit is further configured to call the entity analysis model to perform entity feature analysis processing on the feature information corresponding to the address to be queried and the feature information corresponding to the address description text, so as to obtain the entity feature information of the target entity corresponding to the address to be queried;

[0014] The determining unit is used to determine the entity recognition result of the target entity based on the entity feature information, wherein the entity recognition result includes one or both of the entity name and entity type.

[0015] In one embodiment, when the processing unit calls the entity analysis model to perform entity feature analysis processing on the feature information corresponding to the address to be queried and the feature information corresponding to the address description text, and obtains the entity feature information of the target entity corresponding to the address to be queried, it specifically performs the following steps:

[0016] The feature analysis module of the entity analysis model is invoked to perform linear transformation on the feature information corresponding to the address to be queried, so as to obtain the first transformed feature information and the second transformed feature information.

[0017] The feature analysis module is invoked to perform a linear transformation on the feature information corresponding to the address description text to obtain the third transformed feature information.

[0018] The feature analysis module is invoked to perform feature adjustment processing on the first transformation feature information, the second transformation feature information, and the third transformation feature information to obtain the adjusted feature information. Based on the feature information corresponding to the address description text and the adjusted feature information, the entity feature information of the target entity corresponding to the address to be queried is determined.

[0019] In one embodiment, the feature information corresponding to the address to be queried includes N feature vectors, where N is a positive integer; the processing unit is used to call the feature analysis module of the entity analysis model to perform linear transformation processing on the feature information corresponding to the address to be queried to obtain the first transformed feature information and the second transformed feature information, specifically to perform the following steps:

[0020] The feature analysis module is invoked to generate an address feature matrix based on the N feature vectors included in the feature information corresponding to the address to be queried, and to obtain a key weight matrix and a value weight matrix, wherein the key weight matrix and the value weight matrix include multiple weight data.

[0021] The feature analysis module is invoked to perform feature transformation processing on the address feature matrix and the key weight matrix to obtain a first weighted feature matrix, and the first transformation feature information is determined based on the first weighted feature matrix. The first transformation feature information includes N first feature vectors.

[0022] The feature analysis module is invoked to perform feature transformation processing on the address feature matrix and the value weight matrix to obtain a second weighted feature matrix, and the second transformation feature information is determined based on the second weighted feature matrix. The second transformation feature information includes N second feature vectors.

[0023] In one embodiment, when the processing unit calls the feature analysis module to perform feature adjustment processing on the first transformation feature information, the second transformation feature information, and the third transformation feature information to obtain the adjusted feature information, it specifically performs the following steps:

[0024] The feature analysis module is invoked to perform matrix transpose on the first weighted feature matrix corresponding to the first transformation feature information to obtain the first transposed feature matrix, and the third weighted feature matrix corresponding to the third transformation feature information and the first transposed feature matrix are subjected to matrix inner product processing to obtain the inner product feature matrix.

[0025] The feature analysis module is invoked to scale each feature data included in the inner product feature matrix, and the scaled inner product feature matrix is ​​normalized to obtain a normalized matrix.

[0026] The feature analysis module is invoked to perform matrix transformation processing on the normalized matrix and the second weighted feature matrix corresponding to the second transformed feature information, and the adjusted feature information is determined based on the matrix transformation result.

[0027] In one embodiment, when the processing unit determines the entity feature information of the target entity corresponding to the address to be queried based on the feature information corresponding to the address description text and the adjusted feature information, it specifically performs the following steps:

[0028] The feature analysis module is invoked to perform feature fusion processing on the feature information corresponding to the address description text and the adjusted feature information to obtain fused feature information. The fused feature information includes one or more feature layers, and the feature layer includes one or more feature vectors.

[0029] The feature analysis module is invoked to perform feature standardization processing on each target feature vector included in the target feature layer to obtain the standardized feature layer corresponding to the target feature layer. The standardized feature layer includes one or more standardized feature vectors, and the target feature layer is any one of the one or more feature layers included in the fused feature information.

[0030] After determining the standardized feature layer corresponding to each feature layer in the fused feature information, the feature analysis module is invoked to generate a standardized feature matrix based on the standardized feature layer corresponding to each feature layer in the fused feature information.

[0031] The feature analysis module is invoked to perform feature compression processing on the standardized feature matrix to obtain the entity feature information of the target entity corresponding to the address to be queried.

[0032] In one embodiment, when the processing unit calls the feature analysis module to perform feature standardization processing on each target feature vector included in the target feature layer to obtain the standardized feature layer corresponding to the target feature layer, it specifically performs the following steps:

[0033] The feature analysis module is invoked to determine the weighted average vector and vector standard deviation data based on the target feature vectors included in the target feature layer;

[0034] The feature analysis module is invoked to normalize each target feature vector based on the weighted average vector and vector standard deviation data, thereby obtaining the normalized feature vector corresponding to each target feature vector;

[0035] The feature analysis module is invoked to perform nonlinear mapping processing on the normalized feature vectors corresponding to each target feature vector to obtain a standardized feature layer.

[0036] In one embodiment, when the processing unit calls the feature analysis module to perform feature compression processing on the standardized feature matrix to obtain the entity feature information of the target entity corresponding to the address to be queried, it specifically performs the following steps:

[0037] The feature analysis module is invoked to obtain the set number of entity types and the number of text segments. The number of entity types is used to indicate the number of entity types associated with the address, and the number of text segments is used to indicate the number of text segments corresponding to the address description text.

[0038] The feature analysis module is invoked to perform feature compression processing on the standardized feature matrix based on the number of entity types and the number of text segments, resulting in a compressed feature matrix.

[0039] The feature analysis module is invoked to perform nonlinear transformation processing on the compressed feature matrix to obtain the entity feature information of the target entity corresponding to the address to be queried.

[0040] In one embodiment, when the processing unit calls the entity analysis model to perform text feature encoding processing on the address description text to obtain the feature information corresponding to the address description text, it specifically performs the following steps:

[0041] The feature encoding module of the entity analysis model is invoked to perform text segmentation processing on the address description text to obtain multiple text segments, and feature encoding processing is performed on each text segment to obtain the feature information of each text segment. The multiple text segments include the entity name of the target entity and the entity name to be determined.

[0042] The feature encoding module is invoked to perform feature association processing on the feature information of each text segmentation. The feature information corresponding to the address description text indicates that the probability of association between the query address and the entity name of the target entity is greater than the probability of association between the query address and the entity name to be determined.

[0043] In one embodiment, the processing unit is further configured to perform the following steps:

[0044] Acquire multiple sample data groups, wherein each sample data group includes a sample address, sample entity information corresponding to the sample address, and address description text of the sample address;

[0045] The feature encoding module of the initial analysis model is invoked to perform text feature encoding processing on the address description text of each sample address to obtain the feature information corresponding to the address description text of each sample address;

[0046] The feature analysis module of the initial analysis model is invoked to perform entity feature analysis processing on the feature information corresponding to the address description text of each sample address, so as to obtain the entity feature information corresponding to each sample address, and to determine the predicted entity information of each sample address based on the entity feature information corresponding to each sample address.

[0047] The feature analysis module of the initial analysis model is adjusted using the difference data to obtain the adjusted initial analysis model. The entity analysis model is then determined based on the adjusted initial analysis model. The difference data is determined based on the predicted entity information of each sample address and the sample entity information.

[0048] Accordingly, embodiments of this application provide a computer device, which includes:

[0049] A processor is a tool for implementing computer programs.

[0050] A computer-readable storage medium storing a computer program adapted to be loaded by a processor and executed by the entity recognition method described above.

[0051] Accordingly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when read and executed by a processor of a computer device, causes the computer device to perform the aforementioned entity recognition method.

[0052] Accordingly, this application also provides a computer program product, which includes a computer program stored in a computer-readable storage medium. A processor of a computer device reads the computer program from the computer-readable storage medium and executes the computer program, causing the computer device to perform the aforementioned entity recognition method.

[0053] In this application, after obtaining the address description text associated with the address to be queried, an entity analysis model can be invoked to process the address description text to obtain the entity feature information of the target entity corresponding to the address to be queried, thereby determining the entity name and entity type of the target entity corresponding to the address to be queried. The entity recognition method provided in this application can determine the entity name and entity type of the entity corresponding to the address using only one entity analysis model, without needing to invoke multiple different processing tools, which can effectively improve the efficiency of determining the entity name and entity type. Furthermore, the method provided in this application does not involve data transmission between different processing tools, and there is no transmission error. Therefore, the entity recognition method provided in this application can effectively improve the accuracy of the entity recognition results. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 This is a schematic diagram of the system architecture of an entity recognition system provided in an embodiment of this application;

[0056] Figure 2 This is a flowchart illustrating an entity recognition method provided in an embodiment of this application;

[0057] Figure 3 This is a schematic diagram of a feature encoding module provided in an embodiment of this application;

[0058] Figure 4 This is a schematic diagram of an entity analysis model provided in an embodiment of this application;

[0059] Figure 5 This is a schematic diagram of an entity recognition method provided in an embodiment of this application;

[0060] Figure 6 This is a schematic flowchart of a model training method provided in an embodiment of this application;

[0061] Figure 7 This is a schematic diagram of a model training method provided in an embodiment of this application;

[0062] Figure 8 This is a structural block diagram of an entity recognition device provided in an embodiment of this application;

[0063] Figure 9 This is a structural block diagram of a computer device provided in an embodiment of this application. Detailed Implementation

[0064] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0065] It should be noted that the terms "first," "second," etc., used in the embodiments of this application are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a technical feature specified with "first" or "second" may explicitly or implicitly include at least one of those features.

[0066] Typically, when determining entity names and types based on addresses, entity extraction tools can be used to identify the entity name of the entity to which the address belongs based on the associated text; then, classification tools can be used to determine the entity type of the entity to which the address belongs based on the entity name. However, this method requires processing with multiple different tools, making the process complex and involving data transmission between different tools, which is prone to propagation errors and thus leads to low accuracy of the processing results.

[0067] Based on this, embodiments of this application provide an entity recognition method that can obtain address description text associated with a query address; call an entity analysis model to perform text feature encoding processing on the address description text to obtain feature information corresponding to the address description text, including feature information corresponding to the query address; wherein, the entity analysis model is obtained by training an initial analysis model using difference data, which is determined based on sample entity information and predicted entity information of the sample address, the sample entity information including reference entity name and reference entity type, and the predicted entity information including predicted entity name and predicted entity type obtained by processing the address description text of the sample address using the initial analysis model; call the entity analysis model to perform entity feature analysis processing on the feature information corresponding to the query address and the feature information corresponding to the address description text to obtain entity feature information of the target entity corresponding to the query address; determine the entity recognition result of the target entity based on the entity feature information, the entity recognition result including one or both of entity name and entity type. Through the method provided by embodiments of this application, the entity analysis model can be used to determine the entity name and entity type corresponding to an address, effectively improving the processing efficiency of the entity recognition method and also contributing to improving the accuracy of the entity recognition result.

[0068] The entity recognition method provided in this application can be applied to natural language processing (NLP) technology within the field of artificial intelligence. Artificial intelligence (AI) utilizes digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceiving the environment, acquiring knowledge, and using that knowledge to obtain optimal results—a set of theories, methods, technologies, and application systems. Natural language processing (NLP) studies various theories and methods that enable effective communication between humans and computers using natural language. NLP involves natural language, the language people use daily, and is closely related to linguistics research; it also involves computer science and mathematics. NLP technologies typically include text processing, semantic understanding, machine translation, robot question answering, and knowledge graphs. The entity recognition method provided in this application can be applied to information verification scenarios in the field of natural language processing. For example, if an object is of type "merchant" and a pending address is input, the entity recognition method provided in this application can be used to obtain the address description text associated with the pending address. An entity analysis model can then be called to process this address description text to obtain the entity feature information of the entity corresponding to the pending address. Based on this entity feature information, the entity name and entity type of the entity to which the pending address belongs can be determined. If the entity type of the entity to which the pending address belongs is "commercial organization," then the object information verification for that object is successful. The method provided in this application can accurately and quickly determine the entity name and entity type corresponding to an address, which helps improve the efficiency of information verification and ensures resource security.

[0069] The entity recognition method provided in this application can also be applied to the field of cloud computing. Cloud computing is a computing model that distributes computing tasks across a resource pool composed of a large number of computers, enabling various application systems to obtain computing power, storage space, and information services as needed. The cloud computing resource pool mainly includes: computing devices (virtualized machines containing operating systems), storage devices, and network devices. The entity recognition method provided in this application involves an entity analysis model, which can be trained using cloud computing devices. Specifically, the initial analysis model can be trained using cloud computing devices based on difference data to obtain the entity analysis model. This difference data can be determined based on sample entity information and predicted entity information of the sample address. The sample entity information can include the reference entity name and reference entity type of the sample address, and the predicted entity information can include the predicted entity name and predicted entity type obtained by processing the address description text of the sample address using the initial analysis model. Through the method provided in this application, cloud computing technology can be used to train the model, improving the training efficiency and prediction accuracy.

[0070] The architecture of the entity recognition system provided in the embodiments of this application will be described below with reference to the accompanying drawings.

[0071] Please see Figure 1 The figure is a schematic diagram of the system architecture of an entity recognition system provided in an embodiment of this application. The entity recognition system includes a terminal device 101, an entity recognition server 102, and a database 103. The entity recognition server 102 can interact with the terminal device 101 and the database 103. The entity recognition server 102 includes an entity analysis model. Wherein:

[0072] Terminal device 101 can run applications and interact with business objects, receiving data input from them. For example, terminal device 101 can run a resource transfer application involving operation objects. The type of operation object can be either a "seller merchant" or a "buyer individual." The business object can input the address to be queried and the object type for subsequent information verification processing. Terminal device 101 can be a handheld device (e.g., smartphone, tablet), computing device (e.g., personal computer, PC), vehicle terminal, intelligent voice interaction device, wearable device, or other intelligent device with input / output and communication functions, but is not limited to these.

[0073] The entity identification server 102 can receive the address to be queried sent by the terminal device 101 and determine the entity name and entity object to which the address belongs. The entity identification server 102 may include an entity analysis model, which it can call to process and thus determine the entity name and entity object. The entity identification server 102 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0074] Database 103 is used to store relevant data of entity identification server 102, such as the address to be queried, entity name, entity type, etc. Database 103 can be a local database within entity identification server 102, or a cloud database associated with entity identification server 102 (i.e., a database deployed in the cloud). Specifically, it can be deployed based on any of the following: private cloud, public cloud, hybrid cloud, edge cloud, etc., thus allowing the cloud database to focus on different functions. For example, a database deployed in a private cloud uses basic cloud hardware on the user's personal device, focusing more on serving a small group of users. A database deployed in a public cloud, on the other hand, is based on a third-party cloud platform, allowing data sharing; any user's data can be stored in this database, and any user can use the data in the database.

[0075] The following will elaborate on such matters. Figure 1 The working principle of the entity recognition system shown:

[0076] Terminal device 101 can run a resource transfer application. This application involves an operation object, which can be of type "seller merchant" or "buyer individual". Business objects can register operation objects through this resource transfer application. During the registration process, business objects can input the address to be queried and set the object type to terminal device 101. Terminal device 101 can generate an entity identification request (which includes the address to be queried and the set object type) and send the entity identification request to entity identification server 102.

[0077] Entity recognition server 102 can obtain address description text associated with the address to be queried, and call the entity analysis model to perform text feature encoding processing on the address description text to obtain feature information corresponding to the address description text. This feature information includes the feature information corresponding to the address to be queried. The entity analysis model in entity recognition server 102 is obtained by training an initial analysis model using difference data. The difference data is determined based on the sample entity information and predicted entity information of the sample address. The sample entity information includes the reference entity name and reference entity type, and the predicted entity information includes the predicted entity name and predicted entity type obtained by processing the address description text of the sample address using the initial analysis model. The device that trains the initial analysis model can be entity recognition server 102, or it can be a different device.

[0078] Entity recognition server 102 can call entity analysis model to perform entity feature analysis processing on the feature information corresponding to the query address and the feature information corresponding to the address description text, to obtain the entity feature information of the target entity corresponding to the query address (i.e., Figure 1 The process involves "calling the entity analysis model to process the address description text and obtaining the entity feature information corresponding to the address to be queried," and determining the entity recognition result of the target entity based on the entity feature information. The entity recognition result may include one or both of the entity name and entity type. The entity recognition server 102 can store the address to be queried and the entity recognition result of the target entity corresponding to the address to be queried in the database 103.

[0079] The entity recognition server 102 can perform corresponding operations based on the comparison results between the entity recognition results and the set object types. For example... Figure 1 As shown, assuming the entity recognition result does not match the set object type (for example, the entity recognition result indicates that the entity type of the entity to which the address to be queried belongs is "individual," while the set object type is "merchant seller"), the entity recognition server 102 can determine that the information verification of the operation object has failed and can send an information verification notification to the terminal device 101. In some cases, the entity recognition method provided in this application can also be used to determine whether the address of a business object is a "valid cluster," that is, whether the entity corresponding to the address of the business object matches the business object. For example, for a business object associated with a commercial institution, the entity type of the entity corresponding to its address should be "commercial institution." Through the entity recognition method provided in the embodiments of this application, the entity name and entity type corresponding to the address can be determined using an entity analysis model, which effectively improves the processing efficiency of the entity recognition method and also helps to improve the accuracy of the entity recognition result.

[0080] It is understood that the schematic diagrams of the entity recognition system described in the embodiments of this application are for the purpose of more clearly illustrating the entity recognition method of the embodiments of this application, and do not constitute a limitation on the entity recognition method provided in the embodiments of this application. For example, the entity recognition method provided in the embodiments of this application can be executed not only by the entity recognition server 102, but also by other devices different from the entity recognition server 102 that can communicate with the terminal device 101 and the database 103. Those skilled in the art will understand that... Figure 1 The number of terminal devices 101, entity recognition servers 102, and databases 103 shown in the examples is merely illustrative. Any number of devices can be configured according to business implementation needs. Furthermore, as system architecture evolves and new business scenarios emerge, the entity recognition method provided in this application embodiment is equally applicable to similar technical problems.

[0081] It should be noted that the collection and processing of relevant data (e.g., the address to be queried) in this application should strictly comply with the requirements of relevant laws and regulations, obtain the informed consent or separate consent of the personal information subject, and carry out subsequent data use and processing within the scope of laws and regulations and the authorization of the personal information subject.

[0082] Please see Figure 2 , Figure 2 This is a flowchart illustrating an entity recognition method provided in an embodiment of this application. This entity recognition method can be implemented by the aforementioned entity recognition server 102, or by other devices capable of implementing the entity recognition method. The following description uses the implementation of the entity recognition method by the aforementioned entity recognition server 102 as an example. The flow of the entity recognition method provided in this embodiment includes, but is not limited to:

[0083] S201. Obtain the address description text associated with the address to be queried.

[0084] In this embodiment, the address to be queried can be input by a business object or sent to the entity recognition server by other devices. The entity recognition server can determine the entity name and entity type of the entity corresponding to the address to be queried, to ensure resource security and data accuracy. For example, if a resource allocation strategy allocates resources only once for the same entity, and one entity can correspond to multiple business objects, and business object 1 has already obtained the allocated resources corresponding to entity A, and business object 2 inputs an address, the entity corresponding to business object 2 is determined to be entity A using the entity recognition method provided in this embodiment. In this case, business object 2 cannot obtain the allocated resources corresponding to entity A again, thereby ensuring resource security. The entity recognition server can obtain address description text associated with the address to be queried, which may include the address to be queried.

[0085] For example, an entity can have multiple business objects. Such entities can be called "collective households". When registering business objects, it is usually necessary to verify whether the address of the business object meets the requirements. The method provided in this application can be used to determine the entity name and entity type corresponding to the address based on the address of the business object. If the entity type indicates that the entity corresponding to the address is a "collective household", then it can be determined that the address of the business object meets the requirements. If the entity type indicates that the entity corresponding to the address is not a "collective household", then relevant data can be further obtained to determine whether the address of the business object meets the requirements.

[0086] In one embodiment, the entity recognition server can determine the address description text associated with the address to be queried by using data stored in the server's database, or it can obtain the address description text associated with the address to be queried from the Internet through a search engine. The address description text can be a segment containing the address to be queried or N sentences (N is a positive integer).

[0087] S202. The entity analysis model is invoked to perform text feature encoding processing on the address description text to obtain the feature information corresponding to the address description text. The feature information corresponding to the address description text includes the feature information corresponding to the address to be queried. The entity analysis model is obtained by training an initial analysis model using difference data. The difference data is determined based on the sample entity information and predicted entity information of the sample address. The sample entity information includes the reference entity name and reference entity type. The predicted entity information includes the predicted entity name and predicted entity type obtained by processing the address description text of the sample address using the initial analysis model.

[0088] In this embodiment, the entity recognition server can call an entity analysis model to perform text feature encoding processing on the address description text to obtain the feature information corresponding to the address description text. This feature information may include the feature information corresponding to the address to be queried. The feature information corresponding to the address description text may include one or more feature vectors corresponding to the address description text.

[0089] The entity analysis model can be obtained by training an initial analysis model using discrepancy data. This discrepancy data can be determined based on the sample entity information and predicted entity information of the sample address. The sample entity information can include the reference entity name and reference entity type of the sample address, and the predicted entity information can include the predicted entity name and predicted entity type obtained by processing the address description text of the sample address using the entity analysis model. The process of training the initial analysis model to obtain the entity analysis model can be as shown in steps S601-S604. Through the method provided in this application embodiment, the entity analysis model can be used to process address description information, quickly determine the feature information corresponding to the address description information, which is beneficial for subsequently determining the entity recognition result corresponding to the query address based on the feature information corresponding to the address description information, thereby effectively improving the processing efficiency of entity recognition.

[0090] In one embodiment, the entity analysis model may include a feature encoding module. The implementation method for calling the entity analysis model to perform text feature encoding on the address description text to obtain the feature information corresponding to the address description text can be as follows: The feature encoding module of the entity analysis model performs text segmentation on the address description text to obtain multiple text segments, and performs feature encoding on each text segment to obtain the feature information of each text segment. The multiple text segments include the entity name of the target entity and the name of the entity to be determined. The feature encoding module in the entity analysis model can perform text segmentation on the address description text to obtain multiple text segments (the text segments may include the entity name of the target entity and the entity names of other unrelated entities to be determined), and determine the feature information of each text segment. The feature information of the text segments can be the word vectors corresponding to the text segments.

[0091] For example, if the address to be queried is "Address A" and the address description text is "School S1 is located in City B, specifically at Address A", the feature encoding module can segment the address description text to obtain the text segments "School S1", "located in", "City B", "specifically at", and "Address A". Among them, "School S1" is the entity name of the target entity, and "City B" is the entity name to be determined. The feature encoding module can determine the word vector corresponding to each text segment.

[0092] The feature encoding module performs feature association processing on the feature information of each text segment. The feature information corresponding to the address description text indicates that the probability of association between the queried address and the entity name of the target entity is greater than the probability of association between the queried address and the name of the entity to be determined. The feature encoding module performs feature association processing on the feature information of each text segment. During this process, based on contextual semantics and the relationship between entities and addresses, the importance of the target entity corresponding to the queried address is increased. Finally, the feature information corresponding to the address description text is obtained. This feature information can be a matrix composed of feature vectors, indicating the relationship between each text segment and other text segments. The feature information output by the feature encoding model, corresponding to the address description text, indicates that the probability (or degree of association) between the queried address and the entity name of the target entity is greater than the probability of association between the queried address and the name of other entities to be determined. The method provided in this application embodiment can determine feature information through the feature encoding module in the entity recognition model, so that the entity name of the target entity can be distinguished from the entity name to be determined, without using the entity extraction model to extract entities from the address description text and perform correlation analysis on the extracted entities to determine the entity name of the target entity. The method provided in this application embodiment does not involve data transmission between different models, which can effectively avoid transmission errors and improve the accuracy of entity recognition results.

[0093] In one embodiment, the result of the feature encoding module in the entity recognition model can be structurally similar to that of the Bidirectional Encoder Representations from Transformers (BERT) model. The BERT model is a pre-trained natural language processing model, primarily consisting of an input layer, an encoding layer, and a pooling layer. The input layer of the BERT model can undergo embedding transformations, layer normalization, and random deactivation operations. Embedding transformations can include word embeddings, position embeddings, and segment embeddings. During the pre-training process of the BERT model, self-attention encoding can be performed between word components (i.e., the various words in the sentence), allowing the BERT model to comprehensively consider the context during word encoding, resulting in more accurate predictions. Contextual labels between sentences can be added to the loss function, enabling the BERT model trained based on the loss function to learn the contextual relationships between sentences. Randomly masking parts of the sentence can prevent overfitting during model parameter updates, making the trained BERT model more robust. In this application, the feature encoding module in the entity analysis model can be a pre-trained BERT model.

[0094] Please see Figure 3 This figure is a schematic diagram of a feature encoding module provided in an embodiment of this application. The feature encoding module in the entity analysis model can be a BERT model. Figure 3 As shown, the feature encoding module may include two feature encoding layers, and each feature encoding layer may include multiple transformer encoder blocks. Figure 3 The term is denoted as Trm, and any transformer coding block in the feature coding layer can be connected to each transformer coding block in the previous feature coding layer, or to each input feature in the input layer. The feature coding module of the entity analysis module can perform text segmentation processing on the address description text to obtain multiple text segments, and perform feature coding processing on each text segment to obtain the feature information of each text segment. The feature information of the text segment (i.e., the input features) can be the sum of feature information determined according to the feature information of the word embedding dimension, the feature information of the position embedding dimension, and the feature information of the segment embedding dimension. The input feature information (i.e., E1, E2, ..., EN, where N is a positive integer) can be input as follows: Figure 3 The structure shown is processed. Trm can perform feature extraction on the input feature information based on a self-attention mechanism, and simultaneously perform bidirectional processing using the context of the input feature information to obtain output feature information (i.e., T1, T2, ..., TN, where N is a positive integer), which is the feature information corresponding to the address description text. Through the method provided in this application embodiment, the feature information corresponding to the address description text can be accurately determined, so that this feature information can represent the association between the feature information of the address to be processed and the feature information of the target entity, which is beneficial for subsequently determining accurate entity analysis results.

[0095] In one embodiment, the above Figure 3 This is just one possible structure similar to the BERT model for the feature encoding module in entity analysis models. In some cases, the structure of the feature encoding module can also be similar to that of a Long Short-Term Memory (LSTM) network model or a Transformer model. The LSTM model is a type of recurrent neural network suitable for processing and predicting long-term dependencies in time series data. LSTMs have hidden states, which can effectively handle error propagation. The Transformer model is a common machine learning model, mainly used for transforming and processing data; in practical applications, the Transformer model can be used for tasks such as data preprocessing, feature extraction, and data cleaning.

[0096] S203. Call the entity analysis model to perform entity feature analysis processing on the feature information corresponding to the address to be queried and the feature information corresponding to the address description text, and obtain the entity feature information of the target entity corresponding to the address to be queried.

[0097] In this embodiment, the entity recognition server can call an entity analysis model to perform entity feature analysis processing on the feature information corresponding to the query address and the feature information corresponding to the address description text, thereby obtaining the entity feature information of the target entity corresponding to the query address. The entity feature information can indicate one or both of the entity name and entity type of the target entity corresponding to the query address. Through the method provided in this embodiment, the entity analysis model can be called to process the address description text of the query address to obtain entity feature information, thereby determining the entity recognition result corresponding to the query address. It eliminates the need to call multiple different processing tools; only one entity analysis model is needed for entity recognition processing, effectively improving the processing efficiency of entity recognition and the accuracy of the entity recognition results.

[0098] In one embodiment, the entity analysis model may further include a feature analysis module. The implementation method for calling the entity analysis model to perform entity feature analysis processing on the feature information corresponding to the query address and the feature information corresponding to the address description text, to obtain the entity feature information of the target entity corresponding to the query address, can be as follows: The feature analysis module of the entity analysis model performs a linear transformation on the feature information corresponding to the query address to obtain first transformed feature information and second transformed feature information; the feature analysis module performs a linear transformation on the feature information corresponding to the address description text to obtain third transformed feature information; the feature analysis module performs feature adjustment processing on the first transformed feature information, second transformed feature information, and third transformed feature information to obtain adjusted feature information; and based on the third transformed feature information and the adjusted feature information, the entity feature information of the target entity corresponding to the query address is determined.

[0099] After determining the first, second, and third transformation feature information based on the feature information corresponding to the address description data and the feature information corresponding to the address to be queried, these triplet information (i.e., the first, second, and third transformation feature information) can be used for further feature adjustment processing to obtain the adjusted feature information. In the adjusted feature information, the correlation between the feature information corresponding to the address to be queried and the feature information corresponding to the target entity is higher, while the correlation between the feature information corresponding to the address to be queried and the feature information corresponding to the name of the entity to be determined is lower.

[0100] The entity feature information of the target entity can be determined based on the third transformed feature information and the adjusted feature information. The method provided in this application embodiment can further extract features from the feature information corresponding to the address description text, resulting in a higher correlation between the feature information corresponding to the queried address and the feature information corresponding to the target entity, thus facilitating the accurate determination of the entity recognition result.

[0101] In one embodiment, the feature information corresponding to the address to be queried includes N feature vectors, where N is a positive integer; the implementation method of calling the feature analysis module of the entity analysis model to perform linear transformation processing on the feature information corresponding to the address to be queried to obtain the first transformed feature information and the second transformed feature information can be as follows:

[0102] The feature analysis module is invoked to generate an address feature matrix based on the N feature vectors corresponding to the address to be queried, and to obtain the key weight matrix and value weight matrix, which contain multiple weight data. The entity recognition server can invoke the feature analysis module to generate an address feature matrix based on the feature information corresponding to the address to be queried, and can obtain the key weight matrix and value weight matrix, which have N columns.

[0103] The feature analysis module is invoked to perform feature transformation processing on the address feature matrix and the key weight matrix to obtain a first weighted feature matrix. Based on this first weighted feature matrix, first transformed feature information is determined, which includes N first eigenvectors. Feature transformation processing (e.g., matrix multiplication) can be performed on the address feature matrix and the key weight matrix to obtain the first weighted feature matrix, which is equivalent to weighting the variables in the address feature matrix. The first transformed feature information can then be determined based on this first weighted feature matrix, which is equivalent to performing a weighted feature transformation on the feature information corresponding to the query address using the key weight matrix to obtain the first transformed feature information.

[0104] The feature analysis module is invoked to perform feature transformation processing on the address feature matrix and the value weight matrix to obtain a second weighted feature matrix. Based on this second weighted feature matrix, second transformed feature information is determined. This second transformed feature information includes N second eigenvectors. Performing feature transformation processing (e.g., matrix multiplication) on the address feature matrix and the value weight matrix to obtain the second weighted feature matrix is ​​equivalent to weighting each variable in the address feature matrix. The second transformed feature information can then be determined based on this second weighted feature matrix, which is equivalent to performing a weighted feature transformation on the feature information corresponding to the query address using the value weight matrix to obtain the second transformed feature information. For example: If the feature information corresponding to the address to be queried includes N feature vectors, each of which is an M-dimensional vector, then the address feature matrix is ​​an N-row, M-column matrix, and the key weight matrix and value weight matrix are both M-row, I-column matrices (I can be the same as N or different from N). The address feature matrix and the key weight matrix can be multiplied together to obtain the first weighted feature matrix, thereby determining the first transformation feature information; the address feature matrix and the value weight matrix can be multiplied together to obtain the second weighted feature matrix, thereby determining the second transformation feature information.

[0105] It should be noted that the method for determining the third transformation feature information is similar to that for determining the first transformation feature information: the feature analysis module can be called to perform feature transformation processing on the address feature matrix and the query weight matrix to obtain the third weighted feature matrix, and the third transformation feature information can be determined based on the third weighted feature matrix. The weight data included in the query weight matrix, key weight matrix, and value weight matrix can be the same or different. Through the method provided in this application embodiment, the first transformation feature information (i.e., key feature information), the second transformation feature information (i.e., value feature information), and the third transformation feature information (i.e., query feature information) can be determined, which is beneficial for subsequent multi-head attention mechanism processing based on these three transformation feature information (also known as ternary feature information), thereby facilitating the obtaining of accurate entity recognition results.

[0106] In one embodiment, the implementation method of calling the feature analysis module to perform feature adjustment processing on the first transformation feature information, the second transformation feature information, and the third transformation feature information to obtain the adjusted feature information can be as follows:

[0107] The feature analysis module is invoked to transpose the first weighted feature matrix corresponding to the first transformation feature information, obtaining the first transposed feature matrix. Then, the inner product of the third weighted feature matrix corresponding to the third transformation feature information and the first transposed feature matrix is ​​performed, obtaining the inner product feature matrix. The first weighted feature matrix can be represented by K, and the first transposed matrix can be represented by K. T If the third weighted eigenma matrix can be represented by Q, then the inner product eigenma matrix can be represented as Q*K. T .

[0108] The feature analysis module is invoked to scale the feature data included in the inner product feature matrix, and then the scaled inner product feature matrix is ​​normalized to obtain a normalized matrix. The scaling process for the feature data in the inner product feature matrix can be achieved by scaling the inner product feature matrix with the text encoding dimension parameter (the text encoding dimension parameter can be represented by d). k The text encoding dimension parameter can be determined based on the encoding dimension of the text segmentation. For example, if the encoding dimension of the text segmentation is 128 dimensions, then d... k It can be 128. The text encoding dimension parameter can also be determined based on the number of heads in the multi-head attention mechanism. The process of normalizing the scaled inner product feature matrix can be achieved using the normalization exponential function (Softmax function). The normalized matrix can be represented as Softmax. The Softmax function can be used to "compress" or "map" a feature vector to another real number vector, such that the value of each element is between (0,1) and the sum of all elements is 1. The Softmax function is commonly used in multi-class classification problems, especially in machine learning and deep learning. Its function is similar to transforming the model's output into a probability distribution, allowing the output value to be directly interpreted as the probability of different classes.

[0109] The feature analysis module is called to perform matrix transformation on the normalized matrix and the second weighted feature matrix corresponding to the second transformed feature information, and the adjusted feature information is determined based on the matrix transformation result. The second weighted feature matrix can be represented by V. The adjusted feature information can be determined using the normalized matrix and the second weighted feature matrix. The formula for calculating the adjusted feature information is as follows (1):

[0110]

[0111] In equation (1) above, Attention(Q,K,V) represents the adjusted feature information. In some cases, the process of determining the adjusted feature information can also be described as follows: perform a matrix linear transformation on the feature information corresponding to the address to be queried to obtain a key matrix Key (i.e., the first weighted feature matrix) and a value matrix Value (i.e., the second weighted feature matrix); perform a matrix linear transformation on the feature information corresponding to the address description text to obtain a query matrix Query (i.e., the third weighted feature matrix); and construct a multi-head attention output (i.e., the adjusted feature information) based on the key matrix, the value matrix, and the query matrix. Through the method provided in the embodiments of this application, the multi-head attention mechanism can be used to linearly transform the matrix into a multi-head dimension based on the feature information output by the feature encoding module, and calculate its self-attention value, so that the correlation between the feature information of the address to be queried and the feature information of the target entity in the adjusted feature information is further increased, and the correlation between the feature information of the address to be queried and the feature information of other irrelevant entities is further reduced, which is beneficial to determine the accurate entity recognition result based on the adjusted feature information.

[0112] In one embodiment, the method for determining the entity feature information of the target entity corresponding to the address to be queried based on the feature information corresponding to the address description text and the adjusted feature information can be as follows:

[0113] The feature analysis module is invoked to perform feature fusion processing on the feature information corresponding to the address description text and the adjusted feature information to obtain fused feature information. The fused feature information includes one or more feature layers, and each feature layer includes one or more feature vectors. The feature fusion of the feature information corresponding to the address description text and the adjusted feature information can be implemented by performing matrix summation on the feature matrix corresponding to the feature information corresponding to the address description text and the feature matrix corresponding to the adjusted feature information to obtain a fused feature matrix. The fused feature information can be determined based on the fused feature matrix. The fused feature information can include one or more feature layers, and each feature layer can include one or more feature vectors. For example, if the address description text of the address to be queried includes 10 sentences, each sentence includes 5 words, and after the address description text is input into the entity analysis model for processing, the word vector of each word is a 128-dimensional vector, then the fused feature information is determined to be a 5*10*128 matrix using the method provided in this application. This fused feature information includes 5 feature layers, and each feature layer includes 10 128-dimensional feature vectors.

[0114] The feature analysis module is invoked to perform feature standardization processing on each target feature vector included in the target feature layer, resulting in a standardized feature layer corresponding to the target feature layer. The standardized feature layer includes one or more standardized feature vectors, and the target feature layer is any one of the one or more feature layers included in the fused feature information. For any feature layer, feature standardization processing can be performed on the feature vectors included in that feature layer to obtain a standardized feature layer. The feature vectors included in the standardized feature layer are all standardized feature vectors.

[0115] After determining the standardized feature layers corresponding to each feature layer in the fused feature information, the feature analysis module is invoked to generate a standardized feature matrix based on the standardized feature layers corresponding to each feature layer in the fused feature information. The standardized feature matrix can be determined based on each standardized feature layer. In some cases, the process of determining the standardized feature matrix can also be described as follows: The multi-head attention output (i.e., the adjusted feature information) is summed with the feature information corresponding to the address description text, and the summed feature result is standardized using layer normalization to obtain the standardized feature matrix.

[0116] The feature analysis module is invoked to perform feature compression processing on the standardized feature matrix, obtaining the entity feature information of the target entity corresponding to the address to be queried; the standardized feature matrix can be compressed to obtain entity feature information that meets application requirements. The method provided in this application embodiment can standardize the output feature information of the multi-head attention mechanism to further improve the prediction accuracy of the model; feature compression of the standardized feature matrix can determine entity feature information, which is beneficial for subsequently determining the entity name and entity type of the target entity based on the entity feature information.

[0117] In one embodiment, the implementation of calling the feature analysis module to perform feature standardization processing on each target feature vector included in the target feature layer to obtain the standardized feature layer corresponding to the target feature layer can be as follows:

[0118] The feature analysis module is invoked to determine the weighted average vector and vector standard deviation data based on the target feature vectors included in the target feature layer. The weighted average vector is used to indicate the average feature vector of the target feature layer, and the vector standard deviation data is used to indicate the differences between the target feature vectors included in the target feature layer. The formula for determining the weighted average vector can be shown in the following formula (2):

[0119]

[0120] In equation (2) above, u lLet H represent the l-th feature layer in the fused feature information, and let H represent the number of feature vectors contained in the l-th feature layer. This represents the i-th feature vector in the l-th feature layer.

[0121] The formula for calculating the standard deviation of vector data can be shown in equation (3) below:

[0122]

[0123] In the above formula (3), o l This represents the vector standard deviation data of the l-th feature layer, and the remaining identifiers are the same as those expressed by the identifiers in equation (2) above.

[0124] The feature analysis module is called to normalize each target feature vector based on the weighted average vector and vector standard deviation data, so as to obtain the normalized feature vector corresponding to each target feature vector. The normalization process for any target feature vector can be implemented as shown in the following equation (4):

[0125]

[0126] In the above formula (4), g represents the normalized eigenvector of the i-th eigenvector in the l-th feature layer. l b represents the gain parameter, and b represents the bias parameter. The gain parameter and the bias parameter are obtained by model training. The other identifiers are the same as those expressed by the identifiers in the above equation (3).

[0127] The feature analysis module is called to perform nonlinear mapping processing on the normalized feature vectors corresponding to each target feature vector to obtain the standardized feature layer. The implementation process of nonlinear mapping processing on the normalized feature vectors can be shown in the following equation (5):

[0128]

[0129] In equation (5) above, h represents the standardized eigenvector. ReLU represents the normalized eigenvector of the i-th eigenvector in the l-th feature layer, and ReLU represents the activation function (ReLU function). The Linear Rectification Function (ReLU function) is a commonly used activation function that can increase the non-linear relationship between the layers of a neural network.

[0130] The normalized feature vectors in the feature layer are processed as shown in Equation (5) to obtain the standardized feature layer. The method for determining the standardized feature layer described in Equations (2) to (5) is called Layer Normalization (Layer-Norm). The method provided in this application embodiment can normalize the feature layer, which can further improve the accuracy of entity feature information and improve the training efficiency of the model during the model training process.

[0131] In one embodiment, the implementation method of calling the feature analysis module to perform feature compression processing on the standardized feature matrix to obtain the entity feature information of the target entity corresponding to the address to be queried can be:

[0132] The feature analysis module is invoked to obtain the set number of entity types and the number of text segments. The number of entity types indicates the number of entity types associated with the address, and the number of text segments indicates the number of text segments corresponding to the address description text. The set number of entity types can be determined based on the number of existing entity types. For example, if there are 10 existing entity types, the number of entity types can be set to 10. The number of text segments is the number of text segments into which the address description text is divided. For example, if the address description text includes 10 sentences, and each sentence is divided into 5 text segments, then the number of text segments can be set to 5.

[0133] The feature analysis module is called to perform feature compression processing on the standardized feature matrix based on the number of entity types and the number of text segments, resulting in a compressed feature matrix. The process of compressing the standardized feature matrix based on the number of entity types and the number of text segments can be achieved by using a linear transformation to compress the standardized feature matrix into a 2n*n_saq*2 compressed feature matrix, where n represents the number of entity types and n_saq represents the number of text segments.

[0134] The feature analysis module performs a nonlinear transformation on the compressed feature matrix to obtain the entity feature information of the target entity corresponding to the query address. This nonlinear transformation can be achieved by inputting the compressed feature matrix into an activation function (Softmax function) for processing to obtain entity feature information. This entity feature information indicates the degree of association between the query address and each entity, and the probability of each entity type (this probability is a binary classification probability, used to indicate whether the target entity is of that entity type; for example, if the target entity is entity type A, the probability is 1; if the target entity is not entity type A, the probability is 0). Since the target entity corresponding to the address can be any of multiple entity types, in some cases, the process of determining entity feature information can also be described as: inputting the standardized feature matrix into a linear transformation layer (i.e., the following...). Figure 4 The linear transformation module in the module outputs multi-type binary classification prediction results (i.e. entity feature information) based on the linear transformation layer. These multi-type binary classification prediction results indicate the binary classification prediction probability of the target entity in each of the multiple entity types.

[0135] Please see Figure 4 This figure is a schematic diagram of an entity analysis model provided in an embodiment of this application. The entity analysis model can be configured in the above-mentioned... Figure 1 In the entity recognition server 102 shown. For example... Figure 4 The entity analysis model shown may include a feature encoding module and a feature analysis module. The feature analysis module may include a multi-head attention module, a summation and normalization module, and a linear transformation module. During entity recognition processing, the entity recognition server can obtain address description text associated with the address to be queried (this address description text may include the address to be queried), and input this address description text into the entity analysis model for processing. The feature encoding module in the entity analysis model can perform text segmentation and feature encoding processing on the address description text to obtain feature information from multiple text segmentation words. For example... Figure 4 The text segmentation feature information shown includes S1, S2, S3, T1, T2, ..., Tn; the [CLS] symbol represents classification, instructing the BERT model to perform a classification task. The output feature information corresponding to [CLS] can then be used as the semantic representation of the text. The [SEP] symbol is used to segment two sentences to distinguish between them. Feature association processing can be performed on the text segmentation feature information (i.e., the feature information of each text segmentation word, [CLS], and [SEP] can be input into the pre-trained BERT module for processing) to obtain the feature information corresponding to the address description text (i.e.,...). Figure 4 The information in the table includes Ec, ES1, ES2, ES3, ET1, ET2, ..., ETn, Ep, where ES1, ES2, and ES3 are the feature information corresponding to the address to be queried.

[0136] ES1, ES2, and ES3 can be processed separately using key weight matrices and value weight matrices to obtain a first weighted feature matrix (Key) and a second weighted feature matrix (Value). The Key, Value, and the feature information corresponding to the address description text (i.e., Ec to Ep) are then input into the multi-head attention mechanism module for processing to obtain the adjusted feature information (i.e.,...). Figure 4 The adjusted feature information can be determined by the method described in equation (1) above.

[0137] The adjusted feature information (ETc to ETp) and the feature information corresponding to the address description text (Ec to Ep) can be input into the summation and standardization module for processing to obtain the standardized feature matrix. The implementation method of the standardized feature matrix can be determined as described in equations (2) to (5) above.

[0138] The standardized feature matrix can be input into the linear transformation module for processing to obtain the entity feature information of the target entity corresponding to the address to be queried. For example... Figure 4 As shown, the determined entity feature information is a feature matrix, and the features in this feature matrix can be used to indicate the probability of the target entity in each entity type, as well as the entity name of the target entity. The method provided in this application embodiment can determine accurate entity feature information, which is beneficial for subsequently determining accurate entity recognition results based on this entity feature information. Furthermore, the entity recognition model provided in this application embodiment can be directly used for entity recognition processing, effectively improving the efficiency of the entity recognition process, reducing the complexity of the entity recognition process, and saving resources.

[0139] It should be noted that, Figure 4 The entity analysis model shown is only one feasible model structure. In practical applications, the model structure of the entity analysis model can be adaptively adjusted according to different application requirements. For example, the model structure of the entity analysis model can also be similar to that of LSTM models, BERT models, Transformer models, etc.

[0140] S204. Determine the entity recognition result of the target entity based on the entity feature information. The entity recognition result includes one or both of the entity name and entity type.

[0141] In this embodiment, the entity recognition server can determine the entity recognition result of the target entity based on entity feature information. This entity recognition result may include one or both of the entity name and entity type. For example, if the address to be queried is "Address X", the entity recognition method provided in this application can determine that the entity name of the target entity corresponding to the address is "Company D" and the entity type is "Commercial Organization". Alternatively, if the type of the target entity is a new type that has never appeared before, the entity recognition result may not include the entity type. The method provided in this embodiment allows for the determination of the entity name and entity type of the target entity corresponding to the address to be queried using a single model, effectively improving the efficiency of entity recognition processing. Furthermore, the method provided in this application does not involve data transmission between different models or processing tools, thus eliminating transmission errors and effectively improving the accuracy of the entity recognition result.

[0142] In one embodiment, the implementation of determining the entity recognition result of a target entity based on entity feature information can be as follows: entity name analysis processing is performed on the entity feature information to obtain the entity name of the target entity; entity type analysis processing is performed on specific feature bits in the entity feature information to obtain the entity type of the target entity. The entity feature information may include feature information of the entity name of the target entity corresponding to the address to be queried, and the feature information of the address to be queried and the feature information of the entity name of the target entity have a high degree of correlation, so the entity name of the target entity can be determined based on the entity feature information. Specific feature bits in the entity feature information can be analyzed to determine the entity type of the target entity.

[0143] For example: entity feature information as described above Figure 4 As shown, the first and second rows of entity feature information indicate the probability that the target entity is of the first entity type. Specifically, if the first and second rows of entity feature information satisfy the feature requirements of the start and end feature bits for the first entity type, then the entity type of the target entity can be determined to be the first entity type. Assuming the start feature bit is the first position in the first row, the end feature bit is the first position in the second row, and the feature requirement is that the values ​​of the start and end feature bits in the entity feature information are both 1, then... Figure 4 The entity feature information shown does not meet the feature requirements, and the entity type of the target entity is the first entity type. The method provided in this application embodiment can quickly and accurately determine the entity name and entity type of the target entity based on entity feature information, effectively improving the processing efficiency and accuracy of entity recognition.

[0144] Please see Figure 5 This figure is a schematic diagram of an entity recognition method provided in an embodiment of this application. Figure 5 The execution device for the entity recognition method shown can be the one described above. Figure 1 The entity recognition server 102 in the middle. Figure 5The entity recognition method shown may include the following steps: S501, determining the address to be queried and obtaining the address description text associated with the address to be queried; the entity recognition server receives the address to be queried sent by other devices and obtains the address description text associated with the address to be queried, which may include the address to be queried. S502, calling an entity analysis model to perform entity analysis processing on the address description text to obtain the entity feature information of the target entity corresponding to the address to be queried; the entity recognition server may call an entity analysis model to perform entity analysis processing on the address description text to obtain the entity feature information of the target entity, which may include relevant feature information of the entity type of the target entity and relevant feature information of the entity name of the target entity. Specifically, the execution of step S502 can be the entity recognition server executing steps S5021-S5022.

[0145] S5021. The feature encoding module is invoked to perform text feature encoding processing on the address description text to obtain the feature information corresponding to the address description text. The entity analysis model may include a feature encoding module, and the entity recognition server may invoke this feature encoding module to perform text feature encoding processing on the address description text to obtain the feature information corresponding to the address description text. The feature information corresponding to the address description text may include the feature information corresponding to the address to be queried. S5022. The feature analysis module is invoked to perform entity feature analysis processing on the feature information corresponding to the address to be queried and the feature information corresponding to the address description text to obtain the entity feature information of the target entity corresponding to the address to be queried. The entity analysis model may also include a feature analysis module, and the entity recognition server may invoke this feature analysis module to perform entity feature analysis processing on the feature information corresponding to the address to be queried and the feature information corresponding to the address description text to obtain the entity feature information of the target entity corresponding to the address to be queried. S503. The entity recognition result of the target entity is determined based on the entity feature information. The entity recognition server may determine the entity recognition result of the target entity based on the entity feature information. The entity recognition result may include one or both of the entity name and entity type. The method provided in this application embodiment can utilize an entity analysis model to achieve entity recognition of the address to be queried, effectively improving the processing efficiency of entity recognition.

[0146] The entity recognition method provided in this application can realize the entity recognition process of an address through an entity analysis model, which effectively improves the processing efficiency of entity recognition; it can avoid data transmission between different processing tools or models, thereby avoiding transmission errors and improving the prediction accuracy of the model and the accuracy of the entity recognition results; it can use the entity analysis model to process the address description text once to obtain the entity name and entity type, which effectively reduces the operational complexity of entity recognition and helps to save resources.

[0147] Please see Figure 6 , Figure 6 This is a flowchart illustrating a model training method provided in an embodiment of this application. The entity analysis model in the above embodiment can be... Figure 6 The model was trained using the method shown above. This model training method can be derived from the above... Figure 1 The entity recognition server 102 in this embodiment can be used to implement the model training method, but it can also be implemented by other devices capable of implementing this model training method. Taking the entity recognition server 102 as an example, the process of the model training method provided in this embodiment includes, but is not limited to:

[0148] S601. Obtain multiple sample data groups, wherein the sample data group includes a sample address, sample entity information corresponding to the sample address, and address description text of the sample address.

[0149] In this embodiment, the entity recognition server can acquire multiple sample data sets. Any sample data set may include a sample address, sample entity information corresponding to the sample address, and address description text for the sample address. The sample entity information corresponding to the sample address may include a reference entity name and reference entity type, and this information is used to implement a supervised model training process. The address description text for the sample address may be obtained by the entity recognition server from a database or the internet based on the sample address. The address description text for the sample address may be a text segment associated with the sample address, and this segment may include multiple sentences. The address description text for the sample address may include the sample address itself.

[0150] S602. Call the feature encoding module of the initial analysis model to perform text feature encoding processing on the address description text of each sample address to obtain the feature information corresponding to the address description text of each sample address.

[0151] In this embodiment of the application, the model structure of the initial analysis model can be as described above. Figure 4 As shown. The entity recognition server can call the feature encoding module of the initial analysis model to perform text feature encoding processing on the address description text of each sample address in multiple sample data groups, thereby obtaining the feature information corresponding to the address description text of each sample address. The feature information corresponding to the address description text of the sample address may include the feature information corresponding to the sample address. The implementation process of the feature encoding module performing text feature encoding processing on the address description text of the sample address can be as shown in step S202 above.

[0152] S603. The feature analysis module of the initial analysis model is invoked to perform entity feature analysis processing on the feature information corresponding to the address description text of each sample address, so as to obtain the entity feature information corresponding to each sample address, and to determine the predicted entity information of each sample address based on the entity feature information corresponding to each sample address.

[0153] In this embodiment, the initial analysis model may further include a feature analysis module. The entity recognition server can call the feature analysis module to perform entity feature analysis processing on the feature information corresponding to the address description text of each sample address, thereby obtaining the entity feature information corresponding to each sample address. Specifically, the feature information corresponding to the address description text of the sample address and the feature information corresponding to the sample address can be input into the feature analysis module for entity feature analysis processing to obtain the entity feature information corresponding to the sample address. The implementation process of the feature analysis module performing entity feature analysis processing on the feature information corresponding to the address description text of the sample address can be as shown in step S203 above. At this time, the key weight matrix, value weight matrix, gain parameter, and bias parameter involved in the implementation process can be initially set. The entity recognition server can determine the predicted entity information of each sample address based on the entity feature information corresponding to each sample address. The predicted entity information may include the predicted entity name and predicted entity type of the sample address predicted by the initial analysis model. The implementation method of determining the predicted entity information of the sample address based on the entity feature information corresponding to the sample address can be as shown in step S204 above.

[0154] It should be noted that when calling the feature analysis module to perform entity feature analysis on the feature information corresponding to the address description text of each sample address, the feature information needs to be standardized using the Layer-Norm method. This can ensure the stability of the feature information, enhance the generalization ability of the model, avoid the gradient vanishing problem during training, and improve the training speed of the model.

[0155] S604. The feature analysis module of the initial analysis model is adjusted using the difference data to obtain the adjusted initial analysis model, and the entity analysis model is determined based on the adjusted initial analysis model. The difference data is determined based on the predicted entity information of each sample address and the sample entity information.

[0156] In this embodiment, after the entity recognition server determines the predicted entity information of each sample address, it can determine the difference data based on the predicted entity information of each sample address and its corresponding sample entity information. This difference data can reflect the model prediction accuracy of the initial analysis model to a certain extent. The calculation formula for determining the difference data can be shown in the following formula (6):

[0157] Loss=∑-[ *log(y ′ )+(1- )*log (1- ′ (6)

[0158] In equation (6) above, y represents the sample entity information of the sample address. ′ The predicted entity information represents the sample address, and Loss represents the difference data. For example, for a sample address, Loss1 can be obtained by performing the calculation as shown in Equation (6) above based on the reference entity type in its sample entity information and the predicted entity type in its predicted entity information. Loss2 can be obtained by performing the calculation as shown in Equation (6) above based on the reference entity name in its sample entity information and the predicted entity name in its predicted entity information. The difference data Loss is then determined based on Loss1 and Loss2.

[0159] Since the feature encoding module in the initial analysis model can be a pre-trained BERT module, which has high prediction accuracy and involves a large number of parameters, the entity recognition server can use difference data to adjust the parameters of the feature analysis module of the initial analysis model in order to improve the training efficiency of the model and save computing resources. The parameters involved in the feature analysis module can include weight data in the key weight matrix, weight data in the value weight matrix, weight data in the query weight matrix, gain parameters involved in the standardization process, bias parameters, etc.

[0160] The entity recognition server can adjust the model parameters of the feature analysis module in the initial analysis model to obtain the adjusted initial analysis model. The entity recognition server can adjust the model parameters of the adjusted initial analysis model multiple times. When the training conditions are met (e.g., the loss function shown in equation (6) converges, or the number of model parameter adjustments reaches a preset number), the adjusted initial analysis model can be determined as the entity analysis model. It should be noted that the process of adjusting the model parameters using difference data can be a parameter gradient update method or other applicable model parameter adjustment methods.

[0161] Please see Figure 7 This figure is a schematic diagram of a model training method provided in an embodiment of this application. Figure 7The initial analysis model shown may include a feature encoding module and a feature analysis module. The feature analysis module may include a multi-head attention module, a summation and standardization module, and a linear transformation module. During model training, the address description text of the sample address can be input into the initial analysis model for processing. Specifically, the feature encoding module of the initial analysis model can be called to perform text feature encoding on the address description text of the sample address to obtain the feature information corresponding to the address description text. The feature information corresponding to the address description text of the sample address can be input into the multi-head attention module for processing to obtain the adjusted feature information corresponding to the address description text of the sample feature. This feature information is then input into the summation and standardization module for processing, and the processing result is input into the linear transformation module to obtain the entity feature information corresponding to the sample address. The entity recognition server can determine the predicted entity information of the sample address based on the feature information of the sample address, and determine the difference data based on the predicted entity information and the sample entity information of the sample address. The entity recognition server can use the difference data to adjust the model parameters of the feature analysis module in the initial analysis model to obtain the adjusted initial analysis model. Through the method provided in this embodiment, model parameters of some modules in the model can be adjusted, effectively improving the model training efficiency while improving the prediction accuracy of the model.

[0162] The model training method provided in this application allows for parameter adjustment of some modules in the model, improving the overall prediction accuracy while effectively ensuring training efficiency. The addition of a standardization module ensures the stability of feature information, enhances the model's generalization ability, avoids gradient vanishing during training, and improves training speed. It enables the training of an entity analysis model capable of predicting entity names and types, facilitating improved processing efficiency and accuracy of entity analysis results.

[0163] Please see Figure 8 , Figure 8 This is a structural block diagram of an entity recognition device provided in an embodiment of this application. The entity recognition device can be disposed in the computer device provided in this embodiment of the application; the computer device can be as described above. Figure 1 The entity recognition server 102 in the entity recognition system shown. Figure 8 The entity recognition device shown can be a computer program running on a computer device, which can be used to execute... Figure 2 or Figure 6 Some or all of the steps in the method embodiments shown. Please refer to [link / reference]. Figure 8 The entity recognition device may include the following units:

[0164] The acquisition unit 801 is used to acquire the address description text associated with the address to be queried;

[0165] Processing unit 802 is used to call an entity analysis model to perform text feature encoding processing on the address description text to obtain feature information corresponding to the address description text. The feature information corresponding to the address description text includes feature information corresponding to the address to be queried. The entity analysis model is obtained by training an initial analysis model using difference data. The difference data is determined based on sample entity information and predicted entity information of the sample address. The sample entity information includes a reference entity name and a reference entity type. The predicted entity information includes a predicted entity name and a predicted entity type obtained by processing the address description text of the sample address using the initial analysis model.

[0166] The processing unit 802 is further configured to call the entity analysis model to perform entity feature analysis processing on the feature information corresponding to the address to be queried and the feature information corresponding to the address description text, so as to obtain the entity feature information of the target entity corresponding to the address to be queried;

[0167] The determining unit 803 is used to determine the entity recognition result of the target entity based on the entity feature information, wherein the entity recognition result includes one or both of the entity name and entity type.

[0168] In one embodiment, the processing unit 802, when invoking the entity analysis model to perform entity feature analysis processing on the feature information corresponding to the address to be queried and the feature information corresponding to the address description text, to obtain the entity feature information of the target entity corresponding to the address to be queried, specifically performs the following steps:

[0169] The feature analysis module of the entity analysis model is invoked to perform linear transformation on the feature information corresponding to the address to be queried, so as to obtain the first transformed feature information and the second transformed feature information.

[0170] The feature analysis module is invoked to perform a linear transformation on the feature information corresponding to the address description text to obtain the third transformed feature information.

[0171] The feature analysis module is invoked to perform feature adjustment processing on the first transformation feature information, the second transformation feature information, and the third transformation feature information to obtain the adjusted feature information. Based on the feature information corresponding to the address description text and the adjusted feature information, the entity feature information of the target entity corresponding to the address to be queried is determined.

[0172] In one embodiment, the feature information corresponding to the address to be queried includes N feature vectors, where N is a positive integer; the processing unit 802 is used to call the feature analysis module of the entity analysis model to perform linear transformation processing on the feature information corresponding to the address to be queried to obtain the first transformed feature information and the second transformed feature information, specifically to perform the following steps:

[0173] The feature analysis module is invoked to generate an address feature matrix based on the N feature vectors included in the feature information corresponding to the address to be queried, and to obtain a key weight matrix and a value weight matrix, wherein the key weight matrix and the value weight matrix include multiple weight data.

[0174] The feature analysis module is invoked to perform feature transformation processing on the address feature matrix and the key weight matrix to obtain a first weighted feature matrix, and the first transformation feature information is determined based on the first weighted feature matrix. The first transformation feature information includes N first feature vectors.

[0175] The feature analysis module is invoked to perform feature transformation processing on the address feature matrix and the value weight matrix to obtain a second weighted feature matrix, and the second transformation feature information is determined based on the second weighted feature matrix. The second transformation feature information includes N second feature vectors.

[0176] In one embodiment, the processing unit 802 is used to call the feature analysis module to perform feature adjustment processing on the first transformation feature information, the second transformation feature information, and the third transformation feature information to obtain the adjusted feature information. Specifically, it is used to perform the following steps:

[0177] The feature analysis module is invoked to perform matrix transpose on the first weighted feature matrix corresponding to the first transformation feature information to obtain the first transposed feature matrix, and the third weighted feature matrix corresponding to the third transformation feature information and the first transposed feature matrix are subjected to matrix inner product processing to obtain the inner product feature matrix.

[0178] The feature analysis module is invoked to scale each feature data included in the inner product feature matrix, and the scaled inner product feature matrix is ​​normalized to obtain a normalized matrix.

[0179] The feature analysis module is invoked to perform matrix transformation processing on the normalized matrix and the second weighted feature matrix corresponding to the second transformed feature information, and the adjusted feature information is determined based on the matrix transformation result.

[0180] In one embodiment, the processing unit 802, when determining the entity feature information of the target entity corresponding to the address to be queried based on the feature information corresponding to the address description text and the adjusted feature information, specifically performs the following steps:

[0181] The feature analysis module is invoked to perform feature fusion processing on the feature information corresponding to the address description text and the adjusted feature information to obtain fused feature information. The fused feature information includes one or more feature layers, and the feature layer includes one or more feature vectors.

[0182] The feature analysis module is invoked to perform feature standardization processing on each target feature vector included in the target feature layer to obtain the standardized feature layer corresponding to the target feature layer. The standardized feature layer includes one or more standardized feature vectors, and the target feature layer is any one of the one or more feature layers included in the fused feature information.

[0183] After determining the standardized feature layer corresponding to each feature layer in the fused feature information, the feature analysis module is invoked to generate a standardized feature matrix based on the standardized feature layer corresponding to each feature layer in the fused feature information.

[0184] The feature analysis module is invoked to perform feature compression processing on the standardized feature matrix to obtain the entity feature information of the target entity corresponding to the address to be queried.

[0185] In one embodiment, the processing unit 802, when invoking the feature analysis module to perform feature standardization processing on each target feature vector included in the target feature layer to obtain the standardized feature layer corresponding to the target feature layer, specifically performs the following steps:

[0186] The feature analysis module is invoked to determine the weighted average vector and vector standard deviation data based on the target feature vectors included in the target feature layer;

[0187] The feature analysis module is invoked to normalize each target feature vector based on the weighted average vector and vector standard deviation data, thereby obtaining the normalized feature vector corresponding to each target feature vector;

[0188] The feature analysis module is invoked to perform nonlinear mapping processing on the normalized feature vectors corresponding to each target feature vector to obtain a standardized feature layer.

[0189] In one embodiment, when the processing unit 802 calls the feature analysis module to perform feature compression processing on the standardized feature matrix to obtain the entity feature information of the target entity corresponding to the address to be queried, it specifically performs the following steps:

[0190] The feature analysis module is invoked to obtain the set number of entity types and the number of text segments. The number of entity types is used to indicate the number of entity types associated with the address, and the number of text segments is used to indicate the number of text segments corresponding to the address description text.

[0191] The feature analysis module is invoked to perform feature compression processing on the standardized feature matrix based on the number of entity types and the number of text segments, resulting in a compressed feature matrix.

[0192] The feature analysis module is invoked to perform nonlinear transformation processing on the compressed feature matrix to obtain the entity feature information of the target entity corresponding to the address to be queried.

[0193] In one embodiment, the processing unit 802, when invoking an entity analysis model to perform text feature encoding processing on the address description text to obtain the feature information corresponding to the address description text, specifically performs the following steps:

[0194] The feature encoding module of the entity analysis model is invoked to perform text segmentation processing on the address description text to obtain multiple text segments, and feature encoding processing is performed on each text segment to obtain the feature information of each text segment. The multiple text segments include the entity name of the target entity and the entity name to be determined.

[0195] The feature encoding module is invoked to perform feature association processing on the feature information of each text segmentation. The feature information corresponding to the address description text indicates that the probability of association between the query address and the entity name of the target entity is greater than the probability of association between the query address and the entity name to be determined.

[0196] In one embodiment, the processing unit 802 is further configured to perform the following steps:

[0197] Acquire multiple sample data groups, wherein each sample data group includes a sample address, sample entity information corresponding to the sample address, and address description text of the sample address;

[0198] The feature encoding module of the initial analysis model is invoked to perform text feature encoding processing on the address description text of each sample address to obtain the feature information corresponding to the address description text of each sample address;

[0199] The feature analysis module of the initial analysis model is invoked to perform entity feature analysis processing on the feature information corresponding to the address description text of each sample address, so as to obtain the entity feature information corresponding to each sample address, and to determine the predicted entity information of each sample address based on the entity feature information corresponding to each sample address.

[0200] The feature analysis module of the initial analysis model is adjusted using the difference data to obtain the adjusted initial analysis model. The entity analysis model is then determined based on the adjusted initial analysis model. The difference data is determined based on the predicted entity information of each sample address and the sample entity information.

[0201] Figure 8 The entities identified in the device can be individually or entirely merged into one or more other units, or some of the units can be further divided into multiple functionally smaller units. This achieves the same operation without affecting the technical effects of the embodiments of this application. The above units are based on logical functions. In practical applications, the function of one unit can be implemented by multiple units, or the function of multiple units can be implemented by one unit. In other embodiments of this application, the blockchain-based data processing device may also include other units. In practical applications, these functions can also be implemented with the assistance of other units, and can be implemented collaboratively by multiple units.

[0202] In one embodiment, the ability to perform such operations can be achieved by running on a general-purpose computing device, such as a computer, which includes processing elements and storage elements such as a central processing unit (CPU), random access memory (RAM), and read-only memory (ROM). Figure 2 or Figure 6 Computer programs for the steps involved in some or all of the methods shown, to construct, for example... Figure 8 The entity recognition device shown herein, and the entity recognition method for implementing the embodiments of this application, are described. A computer program may be recorded on, for example, a computer-readable storage medium, loaded onto the aforementioned computing device via the computer-readable storage medium, and executed therein.

[0203] The entity recognition device provided in this application embodiment can realize the entity recognition process of an address through an entity analysis model, which effectively improves the processing efficiency of entity recognition processing; it can avoid data transmission between different processing tools or models, thereby avoiding transmission errors and improving the prediction accuracy of the model and the accuracy of the entity recognition results; it can use the entity analysis model to process the address description text once to obtain the entity name and entity type, which effectively reduces the operational complexity of entity recognition processing and helps to save resources.

[0204] Based on the above methods and apparatus embodiments, this application provides a computer device. Please refer to... Figure 9 , Figure 9 This is a structural block diagram of a computer device provided in an embodiment of this application. Figure 9The computer device shown can be the one described above. Figure 1 The entity recognition server 102 in the middle. Figure 9 The computer device shown includes at least a processor 901, an input interface 902, an output interface 903, and a computer-readable storage medium 904. The processor 901, input interface 902, output interface 903, and computer-readable storage medium 904 can be connected via a bus or other means.

[0205] Computer-readable storage medium 904 can be stored in the memory of a computer device. Computer-readable storage medium 904 is used to store computer programs, including computer instructions. Processor 901 is used to execute the computer program stored in computer-readable storage medium 904. Processor 901 (or CPU (Central Processing Unit)) is the computing and control core of the computer device, suitable for implementing computer programs, specifically for loading and executing computer programs to achieve corresponding methods or functions.

[0206] This application also provides a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both built-in storage media in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space for storing the operating system of the computer device. Furthermore, the storage space also stores computer programs suitable for loading and execution by a processor. It should be noted that the computer-readable storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device; optionally, it can also be at least one computer-readable storage medium located remotely from the aforementioned processor.

[0207] In a specific implementation, the processor 901 can load and execute the computer program stored in the computer-readable storage medium 904 to achieve the aforementioned related... Figure 2 The corresponding steps in the entity recognition method are shown. Specifically, the computer program in the computer-readable storage medium 904 is loaded and executed by the processor 901 as follows:

[0208] Retrieve the address description text associated with the address to be queried;

[0209] The entity analysis model is invoked to perform text feature encoding on the address description text to obtain feature information corresponding to the address description text. The feature information corresponding to the address description text includes feature information corresponding to the address to be queried. The entity analysis model is obtained by training an initial analysis model using difference data. The difference data is determined based on sample entity information and predicted entity information of the sample address. The sample entity information includes a reference entity name and a reference entity type. The predicted entity information includes a predicted entity name and a predicted entity type obtained by processing the address description text of the sample address using the initial analysis model.

[0210] The entity analysis model is invoked to perform entity feature analysis on the feature information corresponding to the address to be queried and the feature information corresponding to the address description text, so as to obtain the entity feature information of the target entity corresponding to the address to be queried;

[0211] The entity recognition result of the target entity is determined based on the entity feature information, and the entity recognition result includes one or both of the entity name and entity type.

[0212] In one embodiment, when the computer program in the computer-readable storage medium 904 is loaded and executed by the processor 901 to call the entity analysis model to perform entity feature analysis processing on the feature information corresponding to the query address and the feature information corresponding to the address description text, and obtain the entity feature information of the target entity corresponding to the query address, the program specifically performs the following steps:

[0213] The feature analysis module of the entity analysis model is invoked to perform linear transformation on the feature information corresponding to the address to be queried, so as to obtain the first transformed feature information and the second transformed feature information.

[0214] The feature analysis module is invoked to perform a linear transformation on the feature information corresponding to the address description text to obtain the third transformed feature information.

[0215] The feature analysis module is invoked to perform feature adjustment processing on the first transformation feature information, the second transformation feature information, and the third transformation feature information to obtain the adjusted feature information. Based on the feature information corresponding to the address description text and the adjusted feature information, the entity feature information of the target entity corresponding to the address to be queried is determined.

[0216] In one embodiment, the feature information corresponding to the address to be queried includes N feature vectors, where N is a positive integer; when the computer program in the computer-readable storage medium 904 is loaded and executed by the processor 901, and calls the feature analysis module of the entity analysis model to perform linear transformation processing on the feature information corresponding to the address to be queried to obtain the first transformed feature information and the second transformed feature information, it is specifically used to perform the following steps:

[0217] The feature analysis module is invoked to generate an address feature matrix based on the N feature vectors included in the feature information corresponding to the address to be queried, and to obtain a key weight matrix and a value weight matrix, wherein the key weight matrix and the value weight matrix include multiple weight data.

[0218] The feature analysis module is invoked to perform feature transformation processing on the address feature matrix and the key weight matrix to obtain a first weighted feature matrix, and the first transformation feature information is determined based on the first weighted feature matrix. The first transformation feature information includes N first feature vectors.

[0219] The feature analysis module is invoked to perform feature transformation processing on the address feature matrix and the value weight matrix to obtain a second weighted feature matrix, and the second transformation feature information is determined based on the second weighted feature matrix. The second transformation feature information includes N second feature vectors.

[0220] In one embodiment, when the computer program in the computer-readable storage medium 904 is loaded and executed by the processor 901 to call the feature analysis module to perform feature adjustment processing on the first transformation feature information, the second transformation feature information, and the third transformation feature information to obtain the adjusted feature information, it is specifically used to perform the following steps:

[0221] The feature analysis module is invoked to perform matrix transpose on the first weighted feature matrix corresponding to the first transformation feature information to obtain the first transposed feature matrix, and the third weighted feature matrix corresponding to the third transformation feature information and the first transposed feature matrix are subjected to matrix inner product processing to obtain the inner product feature matrix.

[0222] The feature analysis module is invoked to scale each feature data included in the inner product feature matrix, and the scaled inner product feature matrix is ​​normalized to obtain a normalized matrix.

[0223] The feature analysis module is invoked to perform matrix transformation processing on the normalized matrix and the second weighted feature matrix corresponding to the second transformed feature information, and the adjusted feature information is determined based on the matrix transformation result.

[0224] In one embodiment, when a computer program in a computer-readable storage medium 904 is loaded and executed by a processor 901 to determine the entity feature information of the target entity corresponding to the address to be queried based on the feature information corresponding to the address description text and the adjusted feature information, it is specifically used to perform the following steps:

[0225] The feature analysis module is invoked to perform feature fusion processing on the feature information corresponding to the address description text and the adjusted feature information to obtain fused feature information. The fused feature information includes one or more feature layers, and the feature layer includes one or more feature vectors.

[0226] The feature analysis module is invoked to perform feature standardization processing on each target feature vector included in the target feature layer to obtain the standardized feature layer corresponding to the target feature layer. The standardized feature layer includes one or more standardized feature vectors, and the target feature layer is any one of the one or more feature layers included in the fused feature information.

[0227] After determining the standardized feature layer corresponding to each feature layer in the fused feature information, the feature analysis module is invoked to generate a standardized feature matrix based on the standardized feature layer corresponding to each feature layer in the fused feature information.

[0228] The feature analysis module is invoked to perform feature compression processing on the standardized feature matrix to obtain the entity feature information of the target entity corresponding to the address to be queried.

[0229] In one embodiment, when the computer program in the computer-readable storage medium 904 is loaded and executed by the processor 901 to call the feature analysis module to perform feature standardization processing on each target feature vector included in the target feature layer to obtain the standardized feature layer corresponding to the target feature layer, it is specifically used to perform the following steps:

[0230] The feature analysis module is invoked to determine the weighted average vector and vector standard deviation data based on the target feature vectors included in the target feature layer;

[0231] The feature analysis module is invoked to normalize each target feature vector based on the weighted average vector and vector standard deviation data, thereby obtaining the normalized feature vector corresponding to each target feature vector;

[0232] The feature analysis module is invoked to perform nonlinear mapping processing on the normalized feature vectors corresponding to each target feature vector to obtain a standardized feature layer.

[0233] In one embodiment, when the computer program in the computer-readable storage medium 904 is loaded and executed by the processor 901 to call the feature analysis module to perform feature compression processing on the standardized feature matrix to obtain the entity feature information of the target entity corresponding to the address to be queried, it is specifically used to perform the following steps:

[0234] The feature analysis module is invoked to obtain the set number of entity types and the number of text segments. The number of entity types is used to indicate the number of entity types associated with the address, and the number of text segments is used to indicate the number of text segments corresponding to the address description text.

[0235] The feature analysis module is invoked to perform feature compression processing on the standardized feature matrix based on the number of entity types and the number of text segments, resulting in a compressed feature matrix.

[0236] The feature analysis module is invoked to perform nonlinear transformation processing on the compressed feature matrix to obtain the entity feature information of the target entity corresponding to the address to be queried.

[0237] In one embodiment, when the computer program in the computer-readable storage medium 904 is loaded and executed by the processor 901 to call the entity analysis model to perform text feature encoding processing on the address description text to obtain the feature information corresponding to the address description text, it is specifically used to perform the following steps:

[0238] The feature encoding module of the entity analysis model is invoked to perform text segmentation processing on the address description text to obtain multiple text segments, and feature encoding processing is performed on each text segment to obtain the feature information of each text segment. The multiple text segments include the entity name of the target entity and the entity name to be determined.

[0239] The feature encoding module is invoked to perform feature association processing on the feature information of each text segmentation. The feature information corresponding to the address description text indicates that the probability of association between the query address and the entity name of the target entity is greater than the probability of association between the query address and the entity name to be determined.

[0240] In one embodiment, when the computer program in the computer-readable storage medium 904 is loaded and executed by the processor 901, it is further configured to perform the following steps:

[0241] Acquire multiple sample data groups, wherein each sample data group includes a sample address, sample entity information corresponding to the sample address, and address description text of the sample address;

[0242] The feature encoding module of the initial analysis model is invoked to perform text feature encoding processing on the address description text of each sample address to obtain the feature information corresponding to the address description text of each sample address;

[0243] The feature analysis module of the initial analysis model is invoked to perform entity feature analysis processing on the feature information corresponding to the address description text of each sample address, so as to obtain the entity feature information corresponding to each sample address, and to determine the predicted entity information of each sample address based on the entity feature information corresponding to each sample address.

[0244] The feature analysis module of the initial analysis model is adjusted using the difference data to obtain the adjusted initial analysis model. The entity analysis model is then determined based on the adjusted initial analysis model. The difference data is determined based on the predicted entity information of each sample address and the sample entity information.

[0245] The computer device provided in this application embodiment can realize the entity recognition process of an address through an entity analysis model, which effectively improves the processing efficiency of entity recognition processing; it can avoid data transmission between different processing tools or models, thereby avoiding transmission errors and improving the prediction accuracy of the model and the accuracy of entity recognition results; it can use the entity analysis model to process the address description text once to obtain the entity name and entity type, which effectively reduces the operational complexity of entity recognition processing and helps to save resources.

[0246] This application also provides a computer-readable storage medium storing computer instructions. When executed on a computer device, the instructions cause the computer device to perform the steps in the various method embodiments of this application to implement the entity recognition method provided in this application. Specific implementation details can be found in the foregoing description and will not be repeated here.

[0247] This application also provides a computer program product, which includes a computer program or computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer program or computer instructions from the computer-readable storage medium and executes the computer program or computer instructions, causing the computer device to perform the steps in the various method embodiments of this application to implement the entity recognition method provided in this application. Specific implementation details can be found in the foregoing description and will not be repeated here.

[0248] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application. In the embodiments of this application, the terms "module" or "unit" refer to a computer program or part of a computer program with a predetermined function, which works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0249] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in or transmitted through a computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0250] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An entity recognition method, characterized in that, The method includes: Retrieve the address description text associated with the address to be queried; The entity analysis model is invoked to perform text feature encoding on the address description text to obtain feature information corresponding to the address description text. The feature information corresponding to the address description text includes feature information corresponding to the address to be queried. The entity analysis model is obtained by training an initial analysis model using difference data. The difference data is determined based on sample entity information and predicted entity information of the sample address. The sample entity information includes a reference entity name and a reference entity type. The predicted entity information includes a predicted entity name and a predicted entity type obtained by processing the address description text of the sample address using the initial analysis model. The entity analysis model is invoked to perform entity feature analysis on the feature information corresponding to the address to be queried and the feature information corresponding to the address description text, so as to obtain the entity feature information of the target entity corresponding to the address to be queried; The entity recognition result of the target entity is determined based on the entity feature information, and the entity recognition result includes one or both of the entity name and entity type.

2. The method as described in claim 1, characterized in that, The step of calling the entity analysis model to perform entity feature analysis processing on the feature information corresponding to the address to be queried and the feature information corresponding to the address description text, to obtain the entity feature information of the target entity corresponding to the address to be queried, includes: The feature analysis module of the entity analysis model is invoked to perform linear transformation on the feature information corresponding to the address to be queried, so as to obtain the first transformed feature information and the second transformed feature information. The feature analysis module is invoked to perform a linear transformation on the feature information corresponding to the address description text to obtain the third transformed feature information. The feature analysis module is invoked to perform feature adjustment processing on the first transformation feature information, the second transformation feature information, and the third transformation feature information to obtain the adjusted feature information. Based on the feature information corresponding to the address description text and the adjusted feature information, the entity feature information of the target entity corresponding to the address to be queried is determined.

3. The method as described in claim 1 or 2, characterized in that, The feature information corresponding to the address to be queried includes N feature vectors, where N is a positive integer; The feature analysis module of the entity analysis model performs a linear transformation on the feature information corresponding to the address to be queried to obtain first transformed feature information and second transformed feature information, including: The feature analysis module is invoked to generate an address feature matrix based on the N feature vectors included in the feature information corresponding to the address to be queried, and to obtain a key weight matrix and a value weight matrix, wherein the key weight matrix and the value weight matrix include multiple weight data. The feature analysis module is invoked to perform feature transformation processing on the address feature matrix and the key weight matrix to obtain a first weighted feature matrix, and the first transformation feature information is determined based on the first weighted feature matrix. The first transformation feature information includes N first feature vectors. The feature analysis module is invoked to perform feature transformation processing on the address feature matrix and the value weight matrix to obtain a second weighted feature matrix, and the second transformation feature information is determined based on the second weighted feature matrix. The second transformation feature information includes N second feature vectors.

4. The method as described in claim 2, characterized in that, The feature analysis module is invoked to perform feature adjustment processing on the first transformation feature information, the second transformation feature information, and the third transformation feature information to obtain adjusted feature information, including: The feature analysis module is invoked to perform matrix transpose on the first weighted feature matrix corresponding to the first transformation feature information to obtain the first transposed feature matrix, and the third weighted feature matrix corresponding to the third transformation feature information and the first transposed feature matrix are subjected to matrix inner product processing to obtain the inner product feature matrix. The feature analysis module is invoked to scale each feature data included in the inner product feature matrix, and the scaled inner product feature matrix is ​​normalized to obtain a normalized matrix. The feature analysis module is invoked to perform matrix transformation processing on the normalized matrix and the second weighted feature matrix corresponding to the second transformed feature information, and the adjusted feature information is determined based on the matrix transformation result.

5. The method as described in claim 2, characterized in that, The step of determining the entity feature information of the target entity corresponding to the address to be queried based on the feature information corresponding to the address description text and the adjusted feature information includes: The feature analysis module is invoked to perform feature fusion processing on the feature information corresponding to the address description text and the adjusted feature information to obtain fused feature information. The fused feature information includes one or more feature layers, and the feature layer includes one or more feature vectors. The feature analysis module is invoked to perform feature standardization processing on each target feature vector included in the target feature layer to obtain the standardized feature layer corresponding to the target feature layer. The standardized feature layer includes one or more standardized feature vectors, and the target feature layer is any one of the one or more feature layers included in the fused feature information. After determining the standardized feature layer corresponding to each feature layer in the fused feature information, the feature analysis module is invoked to generate a standardized feature matrix based on the standardized feature layer corresponding to each feature layer in the fused feature information. The feature analysis module is invoked to perform feature compression processing on the standardized feature matrix to obtain the entity feature information of the target entity corresponding to the address to be queried.

6. The method as described in claim 5, characterized in that, The step of calling the feature analysis module to perform feature standardization processing on each target feature vector included in the target feature layer to obtain the standardized feature layer corresponding to the target feature layer includes: The feature analysis module is invoked to determine the weighted average vector and vector standard deviation data based on the target feature vectors included in the target feature layer; The feature analysis module is invoked to normalize each target feature vector based on the weighted average vector and vector standard deviation data, thereby obtaining the normalized feature vector corresponding to each target feature vector; The feature analysis module is invoked to perform nonlinear mapping processing on the normalized feature vectors corresponding to each target feature vector to obtain a standardized feature layer.

7. The method as described in claim 5, characterized in that, The step of calling the feature analysis module to perform feature compression processing on the standardized feature matrix to obtain the entity feature information of the target entity corresponding to the address to be queried includes: The feature analysis module is invoked to obtain the set number of entity types and the number of text segments. The number of entity types is used to indicate the number of entity types associated with the address, and the number of text segments is used to indicate the number of text segments corresponding to the address description text. The feature analysis module is invoked to perform feature compression processing on the standardized feature matrix based on the number of entity types and the number of text segments, resulting in a compressed feature matrix. The feature analysis module is invoked to perform nonlinear transformation processing on the compressed feature matrix to obtain the entity feature information of the target entity corresponding to the address to be queried.

8. The method as described in claim 1 or 2, characterized in that, The entity analysis model is invoked to perform text feature encoding on the address description text to obtain the feature information corresponding to the address description text, including: The feature encoding module of the entity analysis model is invoked to perform text segmentation processing on the address description text to obtain multiple text segments, and feature encoding processing is performed on each text segment to obtain the feature information of each text segment. The multiple text segments include the entity name of the target entity and the entity name to be determined. The feature encoding module is invoked to perform feature association processing on the feature information of each text segmentation. The feature information corresponding to the address description text indicates that the probability of association between the query address and the entity name of the target entity is greater than the probability of association between the query address and the entity name to be determined.

9. The method as described in claim 1 or 2, characterized in that, The method further includes: Acquire multiple sample data groups, wherein each sample data group includes a sample address, sample entity information corresponding to the sample address, and address description text of the sample address; The feature encoding module of the initial analysis model is invoked to perform text feature encoding processing on the address description text of each sample address to obtain the feature information corresponding to the address description text of each sample address; The feature analysis module of the initial analysis model is invoked to perform entity feature analysis processing on the feature information corresponding to the address description text of each sample address, so as to obtain the entity feature information corresponding to each sample address, and to determine the predicted entity information of each sample address based on the entity feature information corresponding to each sample address. The feature analysis module of the initial analysis model is adjusted using the difference data to obtain the adjusted initial analysis model. The entity analysis model is then determined based on the adjusted initial analysis model. The difference data is determined based on the predicted entity information of each sample address and the sample entity information.

10. An entity recognition device, characterized in that, include: The acquisition unit is used to acquire the address description text associated with the address to be queried; The processing unit is configured to invoke an entity analysis model to perform text feature encoding processing on the address description text to obtain feature information corresponding to the address description text. The feature information corresponding to the address description text includes feature information corresponding to the address to be queried. The entity analysis model is obtained by training an initial analysis model using difference data. The difference data is determined based on sample entity information and predicted entity information of the sample address. The sample entity information includes a reference entity name and a reference entity type. The predicted entity information includes a predicted entity name and a predicted entity type obtained by processing the address description text of the sample address using the initial analysis model. The processing unit is further configured to call the entity analysis model to perform entity feature analysis processing on the feature information corresponding to the address to be queried and the feature information corresponding to the address description text, so as to obtain the entity feature information of the target entity corresponding to the address to be queried; The determining unit is used to determine the entity recognition result of the target entity based on the entity feature information, wherein the entity recognition result includes one or both of the entity name and entity type.

11. A computer device, characterized in that, The computer device includes: A processor is a tool for implementing computer programs. A computer-readable storage medium storing a computer program adapted to be loaded by the processor and to implement the entity recognition method as described in any one of claims 1-9.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program adapted to be loaded by a processor and to implement the entity recognition method as described in any one of claims 1-9.

13. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the entity recognition method as described in any one of claims 1-9.

Images