Named entity recognition method and its device, equipment, medium, and product

The pointer vector generated by deep semantic information extracts the product attribute information of e-commerce platforms, which solves the problems of low accuracy and insufficient recall in the existing technology of product attribute information extraction, and realizes efficient naming entity recognition and massive data processing.

CN113962224BActive Publication Date: 2024-11-26GUANGZHOU HUADUO NETWORK TECH
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202111177567.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-10-09
Publication Date
2024-11-26
Estimated Expiration
2041-10-09

AI Technical Summary

Technical Problem

The existing technology has problems of low accuracy and insufficient recall in the extraction of product attribute information in the e-commerce field. Especially when facing the rapid growth of product data, keyword-based methods cannot be effectively processed, and rules-based methods only Ability to mine results that match the rule template.

Method used

Deep semantic information is used to generate first-word pointer vectors and tail-word pointer vectors. The named entities in text information are extracted through these pointer vectors, and the network model is used to extract text feature and pointer vector generation to achieve accurate naming entities. Identify.

Benefits of technology

It improves the accuracy and recall of named entity recognition, and is suitable for rapidly growing application scenarios, such as e-commerce platforms and book management, and can efficiently process massive data, reduce costs and quickly digest data processing tasks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113962224B_ABST
    Figure CN113962224B_ABST
Patent Text Reader

Abstract

The present application discloses a method for named entity recognition and its apparatus, equipment, medium, and product, the method comprising: obtaining text information of a named entity to be recognized, the text information comprising a plurality of words; extracting deep semantic information of the text information to obtain its text feature vector; generating a first word pointer vector and a last word pointer vector according to the text feature vector, each pointer vector comprising pointer elements pointing to each word in the text information in sequence, the pointer elements in the two pointer vectors corresponding to the first word and the last word of the named entity in the text information storing the index label of the entity type to which the named entity belongs; extracting the character string between the words in the text information pointed to by the pointer elements of the same index label in the two pointer vectors as the named entity. The present application can improve the efficiency of named entity recognition, and has a high recall rate and accuracy, and is particularly suitable for extracting corresponding named entities from product information as product attribute data.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to natural language processing technology, and in particular to a named entity recognition method and its corresponding apparatus, computer equipment, computer-readable storage medium, and computer program product. Background Art

[0002] With the rapid development of the e-commerce industry, the number of users and product types on e-commerce platforms has also grown rapidly. Faced with tens of millions of product data, it is particularly important to standardize and unify product information. On the one hand, for the platform, the platform can use the attribute information in the product to classify the products, or perform more scenario applications based on the attributes, such as recommendations; on the other hand, for users, users can perform limited searches based on the product attribute information, and find the products they want in a targeted manner, which improves the user experience. These scenarios are inseparable from the processing of product attribute information in product data. How to extract product attributes from messy data is a major problem in current information extraction scenarios.

[0003] In the prior art, keywords are commonly used to match product information. When a word in the attribute keyword library is hit, the product is considered to have the attribute corresponding to the keyword. There are also some technologies that mine specific text patterns in product information, sort out specific attribute rules, and use the rules to extract product attributes.

[0004] The keyword matching method ignores the semantic information of the text, and the attribute information recalled is mostly noise data, and the accuracy cannot be effectively guaranteed. The extraction method based on attribute rules has a higher accuracy rate than the keyword method, but the pattern is relatively simple, and the attribute rules need to be added and deleted by manual mining. It cannot recall new attribute information, which also has shortcomings.

[0005] The extraction of product attribute information is essentially the problem of named entity extraction in the field of natural language technology. Therefore, it is understandable that the problem of named entity extraction exists not only in the e-commerce field, but also in other fields that need to extract named entities. They also need to face the same technical problems. Summary of the invention

[0006] The primary purpose of the present application is to solve at least one of the above problems and provide a method for named entity recognition and its corresponding apparatus, computer device, computer-readable storage medium, and computer program product.

[0007] In order to meet the various objectives of this application, this application adopts the following technical solutions:

[0008] A method for named entity recognition provided for one of the purposes of this application comprises the following steps:

[0009] Acquire text information of a named entity to be identified, wherein the text information includes a plurality of words;

[0010] Extracting deep semantic information of the text information to obtain its text feature vector;

[0011] Generate a first word pointer vector and a last word pointer vector according to the text feature vector, each pointer vector includes pointer elements pointing to each single word in the text information in sequence, and the pointer elements corresponding to the first word and the last word of the named entity in the text information in the two pointer vectors store the index label of the entity type to which the named entity belongs;

[0012] The character strings between the single words in the text information pointed to by the pointer elements of the same index label in the two pointer vectors are extracted as named entities.

[0013] In a further embodiment, obtaining text information of a named entity to be identified includes the following steps:

[0014] Get the product object of the named entity to be identified;

[0015] Extracting the product title text and / or product detail information of the product object as product information;

[0016] Data cleaning is performed on the commodity information of the commodity object, and the text information after data cleaning is constructed into text information of the named entity to be identified.

[0017] In a further embodiment, a first word pointer vector and a last word pointer vector are generated according to the text feature vector, and a pre-trained pointer generation network model is called to implement the pointer generation network model, which performs the following steps:

[0018] Taking the text feature vector as the vector to be processed input, an encoder is used to perform feature operation on the vector to be processed word by word to adapt to the text information, so as to generate a first word pointer vector corresponding to the current single word;

[0019] Concatenate the text feature vector with the first word pointer vector corresponding to the current word to obtain a fused feature vector corresponding to the current word;

[0020] The fused feature vector is used as the vector to be processed input, and an encoder is used to perform feature operation on the vector to be processed word by word to adapt to the text information, so as to generate a tail word pointer vector corresponding to the current word.

[0021] In a specific embodiment, the encoder performs the following steps for each word in the text information to obtain a corresponding first word pointer vector or last word pointer vector:

[0022] Obtaining a hidden memory vector corresponding to a previous word, wherein the hidden memory vector is randomly initialized and referenced and then updated word by word;

[0023] Perform a multi-head attention mechanism operation according to the vector to be processed, and calculate the normalized dot product of the vector to be processed and the hidden memory vector corresponding to the previous word as the addressing memory vector corresponding to the current word;

[0024] The addressing memory vector corresponding to the current single word is linearly transformed and then added to the vector to be processed to obtain an intermediate vector;

[0025] The intermediate vector is linearly transformed and regularized to obtain the first word pointer vector or the last word pointer vector corresponding to the current word;

[0026] According to the first word pointer vector, the last word pointer vector and the hidden memory vector corresponding to the previous word corresponding to the current word, the hidden memory vector corresponding to the current word is updated for reference in the encoding process of the next word.

[0027] In a further embodiment, the hidden layer memory vector corresponding to the current word is updated according to the first word pointer vector, the last word pointer vector and the hidden layer memory vector corresponding to the previous word for reference in the encoding process of the next word, including the following steps:

[0028] Perform vector addition of the first word pointer vector and the last word pointer vector corresponding to the current word to obtain a sum vector;

[0029] Performing classification mapping on the summed vector to obtain classification probabilities as corresponding smoothing weights;

[0030] Using the smoothing weight as a weight parameter, the sum vector and the hidden layer memory vector corresponding to the previous word are smoothly synthesized to obtain the hidden layer memory vector corresponding to the current word;

[0031] The hidden memory vector corresponding to the current word is passed to the encoding process of the next word for reference.

[0032] In a further embodiment, extracting the character strings between the single words in the text information pointed to by the pointer elements of the same index tag in the two pointer vectors as named entities comprises the following steps:

[0033] Obtain index labels corresponding to first word pointer elements and last word pointer elements of different named entities according to the first word pointer vector and the last word pointer vector;

[0034] Query a preset mapping vocabulary to restore the entity type corresponding to the index label;

[0035] According to the boundaries defined by the single words in the text information pointed to by the first word pointer element and the last word pointer element of each named entity, the character string corresponding to each named entity is extracted, and a corresponding relationship list between the entity type and the named entity is constructed.

[0036] In an extended embodiment, the pointer generation network model is pre-trained, and the training process includes the following steps:

[0037] Acquire a sample data set, wherein the sample data set includes multiple groups of sample data, each group of sample data includes a title text of a commodity object and a first word pointer vector and a last word pointer vector corresponding to a named entity therein;

[0038] Calling each set of sample data to train the pointer generation network model, wherein the pointer generation network model predicts the corresponding first word pointer vector and last word pointer vector according to the text feature vector in the set of sample data;

[0039] According to the first word pointer vector and the last word pointer vector in the trained group of sample data, the first word pointer vector and the last word pointer vector predicted by the pointer generation network model are respectively supervised and the loss value is calculated. When the loss value is greater than a preset threshold, the next group of sample data is called to continue iterative training of the pointer generation network model.

[0040] A named entity recognition device provided to meet one of the purposes of the present application includes: a text acquisition module, a feature extraction module, a pointer generation module, and an entity extraction module, wherein the text acquisition module is used to acquire text information of a named entity to be identified, the text information including multiple words; the feature extraction module is used to extract deep semantic information of the text information and obtain its text feature vector; the pointer generation module is used to generate a first word pointer vector and a last word pointer vector according to the text feature vector, each pointer vector includes pointer elements pointing to each word in the text information in sequence, and the pointer elements corresponding to the first word and last word of the named entity in the text information in the two pointer vectors store the index label of the entity type to which the named entity belongs; the entity extraction module is used to extract the character string between the words in the text information pointed to by the pointer elements with the same index label in the two pointer vectors as the named entity.

[0041] In a further embodiment, the text acquisition module includes: an object acquisition submodule, used to acquire the product object of the named entity to be identified; a text extraction submodule, used to extract the product title text and / or product details information of the product object as product information; a text optimization submodule, used to perform data cleaning on the product information of the product object, and construct the text information after data cleaning as the text information of the named entity to be identified.

[0042] In a further embodiment, in the pointer generation module, a pre-trained pointer generation network model is called to implement it, and the pointer generation network model includes: a first word pointer network, which is used to take the text feature vector as the input of the vector to be processed, and adapt to the text information to use the encoder to perform feature operations on the vector to be processed word by word to generate a first word pointer vector corresponding to the current word; a vector splicing network, which is used to splice the text feature vector with the first word pointer vector corresponding to the current word to obtain a fused feature vector corresponding to the current word; a tail word pointer network, which is used to take the fused feature vector as the input of the vector to be processed, and adapt to the text information to use the encoder to perform feature operations on the vector to be processed word by word to generate a tail word pointer vector corresponding to the current word.

[0043] In a specific embodiment, the encoder called in the first word pointer network and the last word pointer network includes: a hidden layer acquisition unit, which is used to obtain a hidden layer memory vector corresponding to the previous word, and the hidden layer memory vector is randomly initialized and referenced and then updated word by word; an addressing operation unit, which is used to perform a multi-head attention mechanism operation according to the vector to be processed, and calculate the normalized dot product of the vector to be processed and the hidden layer memory vector corresponding to the previous word as the addressing memory vector corresponding to the current word; an intermediate processing unit, which is used to linearly transform the addressing memory vector corresponding to the current word and add it to the vector to be processed to obtain an intermediate vector; a pointer generation unit, which is used to linearly transform the intermediate vector and superimpose a regularization operation to obtain the first word pointer vector or the last word pointer vector corresponding to the current word; a hidden layer update unit, which is used to update the hidden layer memory vector corresponding to the current word according to the first word pointer vector, the last word pointer vector and the hidden layer memory vector corresponding to the previous word for reference in the encoding process of the next word.

[0044] In a further embodiment, the hidden layer update unit includes: a sum processing subunit, which is used to perform vector addition of the first word pointer vector and the last word pointer vector corresponding to the current word to obtain a sum vector; a weight mapping subunit, which is used to perform classification mapping on the sum vector to obtain a classification probability as a corresponding smoothing weight; a smoothing synthesis subunit, which is used to use the smoothing weight as a weight parameter to smoothly synthesize the sum vector and the hidden layer memory vector corresponding to the previous word to obtain the hidden layer memory vector corresponding to the current word; and a hidden layer transfer subunit, which is used to transfer the hidden layer memory vector corresponding to the current word to the encoding process of the next word for reference.

[0045] In a further embodiment, the entity extraction module includes: a character extraction submodule, which is used to obtain the index labels corresponding to the first character pointer elements and the last character pointer elements of different named entities according to the first character pointer vector and the last character pointer vector; a mapping and restoration submodule, which is used to query a preset mapping vocabulary and restore the entity type corresponding to the index label; a list construction submodule, which is used to extract the character strings corresponding to each named entity according to the boundaries defined by the single words in the text information pointed to by the first character pointer element and the last character pointer element of each named entity, and construct a list of correspondence between entity types and named entities.

[0046] In an extended embodiment, the pointer generation network model is placed in the following structure for pre-training, which structure includes: a sample acquisition submodule, used to acquire a sample data set, the sample data set includes multiple groups of sample data, each group of sample data includes a title text of a commodity object and the first word pointer vector and the last word pointer vector corresponding to the named entity therein; a sample training submodule, used to call each group of sample data to train the pointer generation network model, wherein the pointer generation network model predicts the corresponding first word pointer vector and last word pointer vector according to the text feature vector in the group of sample data; a gradient update submodule, used to supervise the first word pointer vector and the last word pointer vector predicted by the pointer generation network model respectively according to the first word pointer vector and the last word pointer vector in the trained group of sample data and calculate the loss value, and when the loss value is greater than a preset threshold, call the next group of sample data to continue iterative training of the pointer generation network model.

[0047] A computer device provided to meet one of the purposes of the present application includes a central processing unit and a memory, wherein the central processing unit is used to call and run a computer program stored in the memory to execute the steps of the named entity recognition method described in the present application.

[0048] A computer-readable storage medium is provided to meet another purpose of the present application, which stores a computer program implemented according to the named entity recognition method in the form of computer-readable instructions. When the computer program is called and executed by a computer, the steps included in the method are executed.

[0049] A computer program product provided to meet another purpose of the present application includes a computer program / instruction, which, when executed by a processor, implements the steps of the method described in any embodiment of the present application.

[0050] Compared with the prior art, the advantages of this application are as follows:

[0051] First, the existing technology uses exhaustive keyword-based methods to identify named entities, which cannot cope with the rapidly growing data fields such as e-commerce product data and shows obvious technical limitations. For rapidly added product data such as e-commerce platforms, the extraction results are less accurate. Although the rule-based naming entity method can improve the accuracy, it can only mine results that match the rule template, and the recall rate is too low. Different from these shortcomings of the prior art, the present application generates double pointers based on the deep semantic information of the text information, and uses the two pointer vectors corresponding to the head and tail to extract the named entities in the text information, which can improve the accuracy of named entity recognition. While improving the confidence of named entity extraction, it can also improve the accuracy and recall rate. It is particularly suitable for application scenarios where the text information of named entities to be identified is growing rapidly, such as e-commerce platforms, title information processing in book management and other fields.

[0052] Secondly, the first word pointer vector and the last word pointer vector generated by the present application are constructed differently from the pointer structure in the prior art. Each pointer vector includes a plurality of pointer elements whose number is the same as the word length of the text information, so that each single word of the text information has a first word pointer element and a last word pointer element corresponding to it. When a named entity is identified from the text information, the first word pointer element corresponding to the first word in the named entity will store the index label corresponding to the entity type to which the named entity belongs. Similarly, the last word pointer element corresponding to the last word in the named entity will store the index label corresponding to the entity type to which the named entity belongs. It can be seen that the text information can be identified by a single first word pointer vector and a single last word pointer vector. A unified representation of all multiple named entities, a string of a named entity in text information, a single word in the text information pointed to by the pointer element of the index label of the entity type to which the named entity belongs in the first word pointer vector, and a single word in the text information pointed to by the pointer element of the index label of the entity type to which the named entity belongs in the last word pointer vector, that is, the last word. The string from the first word to the last word is the named entity, and the mapping relationship between the named entity and its entity type is also determined. It can be seen that the two pointer vectors play a role of unified representation and can be generated by an end-to-end model with high processing efficiency, which is convenient for determining the corresponding multiple named entities from the text information with a smaller amount of computation. For batch recognition tasks, the efficiency advantage is obvious.

[0053] In addition, due to the various advantages mentioned above, such as efficient calculation and accurate recognition, the present application is particularly suitable for e-commerce platforms, and can extract various commodity attribute information from the commodity information of the massive commodity objects on the e-commerce platform, thereby further reflecting its advantages in large-scale applications, such as reducing costs and quickly digesting massive data processing tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0055] Figure 1 A flowchart of a typical embodiment of the method for named entity recognition of the present application;

[0056] Figure 2 A schematic diagram for illustrating the corresponding relationship between the first word pointer vector and the last word pointer vector and the text information regarding the index label in this application;

[0057] Figure 3 A flowchart of a process for extracting text information of a commodity object in an embodiment of the present application;

[0058] Figure 4 A schematic diagram of a process for generating a first word pointer vector and a last word pointer vector in an embodiment of the present application;

[0059] Figure 5 It is a principle block diagram of the pointer network generation model in the embodiment of the present application, in which the network structure corresponding to each single word is expanded and illustrated;

[0060] Figure 6 A schematic diagram of the network structure of an encoder for generating a first-word pointer vector in a pointer network generation model in an embodiment of the present application;

[0061] Figure 7 A schematic diagram of the network structure of an encoder for generating a tail word pointer vector in a pointer network generation model in an embodiment of the present application;

[0062] Figure 8 A schematic diagram of the common encoding business logic of the encoder in the pointer network generation model in an embodiment of the present application;

[0063] Fig. 9 Schematic diagram of the process of updating the hidden memory vector in the embodiment of the present application;

[0064] Fig.10 A flowchart of a process of obtaining a list of correspondences between entity types and named entities from text information according to a first word pointer vector and a last word pointer vector in an embodiment of the present application;

[0065] Fig.11 A flowchart of a training task for training the pointer generation network model proposed in this application;

[0066] Fig.12 This is a principle block diagram of the named entity recognition device of the present application;

[0067] Fig.13A schematic diagram of the structure of a computer device used in this application. DETAILED DESCRIPTION

[0068] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present application, and cannot be interpreted as limiting the present application.

[0069] It will be understood by those skilled in the art that, unless expressly stated, the singular forms "one", "said", and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of the present application refers to the presence of the features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof. It should be understood that when we refer to an element as being "connected" or "coupled" to another element, it may be directly connected or coupled to the other element, or there may be an intermediate element. In addition, the "connection" or "coupling" used herein may include wireless connection or wireless coupling. The term "and / or" used herein includes all or any unit and all combinations of one or more associated listed items.

[0070] It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as those generally understood by those skilled in the art to which this application belongs. It should also be understood that terms such as those defined in general dictionaries should be understood to have meanings consistent with those in the context of the prior art, and will not be interpreted with idealized or overly formal meanings unless specifically defined as here.

[0071] It will be understood by those skilled in the art that the "client", "terminal" and "terminal device" used herein include both devices with wireless signal receivers, which are devices with only wireless signal receivers without transmission capabilities, and devices with receiving and transmitting hardware, which are devices with receiving and transmitting hardware capable of two-way communication on a two-way communication link. Such devices may include: cellular or other communication devices such as personal computers, tablet computers, which have single-line displays or multi-line displays or cellular or other communication devices without multi-line displays; PCS (Personal Communications Service, personal communication system), which can combine voice, data processing, fax and / or data communication capabilities; PDA (Personal Digital Assistant, personal digital assistant), which may include a radio frequency receiver, pager, Internet / intranet access, web browser, notepad, calendar and / or GPS (Global Positioning System, global positioning system) receiver; conventional laptop and / or palmtop computers or other devices, which have and / or include a conventional laptop and / or palmtop computer or other device with and / or including a radio frequency receiver. The "client", "terminal" and "terminal device" used herein may be portable, transportable, installed in a vehicle (air, sea and / or land), or suitable for and / or configured to run locally, and / or in a distributed form, at any other location on the earth and / or in space. The "client", "terminal" and "terminal device" used herein may also be a communication terminal, an Internet terminal, a music / video playing terminal, for example, a PDA, a MID (Mobile Internet Device) and / or a mobile phone with a music / video playing function, or a smart TV, a set-top box and other devices.

[0072] The hardware referred to by the names such as "server", "client", and "service node" in this application is essentially an electronic device with the equivalent capabilities of a personal computer. It is a hardware device with the necessary components revealed by the von Neumann principle, such as a central processing unit (including an arithmetic unit and a controller), a memory, an input device, and an output device. The computer program is stored in its memory, and the central processing unit calls the program stored in the external memory into the internal memory for execution, executes the instructions in the program, and interacts with the input and output devices to complete specific functions.

[0073] It should be pointed out that the concept of "server" referred to in this application can also be extended to the case of server clusters. According to the network deployment principle understood by those skilled in the art, the servers should be logically divided. In physical space, these servers can be independent of each other but can be called through interfaces, or integrated into a physical computer or a set of computer clusters. Those skilled in the art should understand this flexibility, and should not use it to restrict the implementation of the network deployment method of this application.

[0074] Unless expressly specified, one or more technical features of the present application can be deployed on a server for implementation and accessed by a client through a remote call to obtain an online service interface provided by the server, or can be directly deployed and run on a client for access.

[0075] The neural network models referenced or may be referenced in this application, unless expressly specified, can be deployed on a remote server and remotely called on the client, or can be deployed and directly called on a client with sufficient device capabilities. In some embodiments, when it runs on the client, its corresponding intelligence can be obtained through transfer learning to reduce the requirements for the client's hardware operating resources and avoid excessive occupation of the client's hardware operating resources.

[0076] Unless explicitly specified, the various data involved in this application can be stored remotely on a server or on a local terminal device, as long as it is suitable for being called by the technical solution of this application.

[0077] Those skilled in the art should be aware that, although the various methods of the present application are described based on the same concept and thus present commonality to each other, unless otherwise specified, these methods can be independently executed. Similarly, for each embodiment disclosed in the present application, they are all proposed based on the same inventive concept, therefore, concepts with the same expression, and concepts that are appropriately changed for convenience despite different expressions, should be understood as equivalent.

[0078] Unless the mutually exclusive relationship between the embodiments to be disclosed in this application is explicitly stated, the relevant technical features involved in each embodiment can be cross-combined to flexibly construct a new embodiment, as long as such combination does not deviate from the creative spirit of this application and can meet the needs of the prior art or solve certain deficiencies in the prior art. Those skilled in the art should be aware of this flexibility.

[0079] A named entity recognition method of the present application can be programmed as a computer program product and deployed in a client or server for execution. For example, in the e-commerce platform application scenario of the present application, it is generally deployed in a server for implementation, whereby the method can be executed by accessing an interface opened after the computer program product is running and performing human-computer interaction with the process of the computer program product through a graphical user interface.

[0080] See also Figure 1 In a typical embodiment, the named entity recognition method of the present application includes the following steps:

[0081] Step S1100: Obtain text information of a named entity to be identified, wherein the text information includes a plurality of words:

[0082] The text information is composed of computer displayable characters, usually including multiple words. The words include half-width characters or full-width characters, for example, Arabic numerals or English letters, words, etc., such as "2021", "XL", or all Chinese characters, such as "a certain brand of winter short-sleeved top red style", or any combination of these different words. In this application, whether it is a full-width character or a half-width character, it is treated as a single word.

[0083] The content expressed by the text information varies according to its application scenario. For example, in the e-commerce platform scenario of the exemplary application of this application, it can be content used to describe product information, including but not limited to the product title text and product details of the product object in the e-commerce platform. In another exemplary scenario for reviewing non-standard words, it can be the user's instant chat content.

[0084] The text information generally includes manually identifiable named entities, so that the corresponding named entities can be automatically identified by implementing the technical solution of the present application. The named entity is essentially a keyword. For example, in the e-commerce application scenario, the attribute information of the product, such as the description text used to describe the brand, style, function, applicable season, etc., can constitute the named entity.

[0085] The source of the text information does not affect the implementation of the present application. For example, it can be provided individually by users, or it can be extracted in batches from a preset database. Generally speaking, text information is the content corresponding to a specific field, depending on the content of the training data set used in the training phase of the pointer generation network model of the present application. That is, training the pointer generation network model of the present application with the content corresponding to a specific field as the training data set can enable the pointer generation network model to learn the ability to recognize named entities in the field, thereby serving the named entity requirements of the text information corresponding to the field. Those skilled in the art should be able to understand this.

[0086] Step S1200: extracting the deep semantic information of the text information to obtain its text feature vector:

[0087] In order for the present application to realize named entity recognition based on the deep semantics of text information, the present application preferably uses a neural network model that can process sequence information to extract deep semantic information from the text information, so as to obtain its corresponding text feature vector. The neural network model can be a recursive neural network connected in a chain, or it can be a recurrent neural network, both of which belong to the RNN architecture. Of course, it can also be other neural network models that can equivalently learn deep semantic representations of text information with reference to contextual information in the text information and extract its deep semantic information. In this regard, the neural network models recommended by this application include but are not limited to LSTM, BiLSTM, Transformer, Bert, AlBert, Ernie, Electra, etc., among which Bert is the best measured, so it is recommended first.

[0088] These neural network models for extracting text feature information can be collectively referred to as text feature extraction models. The text feature extraction model used in this application is a pre-trained model, that is, it has been pre-trained to a convergence state, so that it has the ability to represent the deep semantic features of text information and is suitable for extracting corresponding text feature vectors from text information. As for the training principle of the text feature extraction model, it is known to those skilled in the art, so it is omitted here.

[0089] Of course, in the present application, the text feature extraction model may also be jointly trained with the pointer generation network model used in the present application, so as to achieve fine-tuning training of the text feature extraction model in the process of training the pointer generation network model of the present application. Those skilled in the art can also understand this.

[0090] This typical embodiment takes the use of the Bert model as an example of the text feature extraction model required for extracting deep semantic information of text information in this application. After the text feature extraction model is trained to a convergent state and put into use in the technical solution of this application, the text information can be used as input to adapt to the input format requirements of the Bert model, and the text vector (Token Embedding), position vector (Position Embedding), and sentence vector (Segment Embedding) corresponding to the text information are encoded, and then fed into the Bert model. The Bert model can then perform deep semantic information extraction based on these vectors with reference to contextual information, and finally output the text feature vector corresponding to the text information.

[0091] It is not difficult to understand that the text feature vector realizes the feature representation of the deep semantic information of the text information and integrates the context information in the text information.

[0092] Step S1300: Generate a first word pointer vector and a last word pointer vector according to the text feature vector, each pointer vector includes pointer elements pointing to each word in the text information in sequence, and the pointer elements corresponding to the first word and the last word of the named entity in the text information in the two pointer vectors store the index label of the entity type to which the named entity belongs:

[0093] After obtaining the text feature vector, a preset pointer generation network model can be used to generate two corresponding pointer vectors according to the text feature vector, which are the first word pointer vector and the last word pointer vector.

[0094] The pointer generation network model preferably recommends the specific model provided in the subsequent embodiments of the present application, or it can also be implemented by those skilled in the art according to the functions and / or principles disclosed in the present application. The pointer generation network model is suitable for being transformed and implemented using a model based on a multi-head attention mechanism, such as being implemented based on the encoding structure of Transformer and Bert. Such models can process sequence information based on a multi-head attention mechanism, and realize the semantic combing according to the contextual information in the sequence. On this basis, it is transformed into two separate pointer vectors suitable for generating the two required for this application, namely the first word pointer vector and the last word pointer vector. After the pointer generation network model is trained to convergence, it can be used in this application to generate the corresponding first word pointer vector and last word pointer vector according to the text feature information.

[0095] The first word pointer vector includes a plurality of pointer elements corresponding to the total number of words in the text information, and each pointer element points to the words one by one according to the order of the words in the text information. Thus, each pointer element points to a single word in the text information, and the order of the pointer element in the first word pointer vector is the same as the order of the word it points to in the text information. The pointer element in the first word pointer vector stores the index label of the entity type to which the named entity corresponding to the word it points belongs in the form of a numerical value. Specifically, when the word pointed to by the pointer element belongs to the first word of a named entity, the pointer element is used to store the index label of the entity type to which the named entity belongs; if the word in the text information pointed to by a pointer element does not belong to the first word of any possible named entity, the pointer element can store a predetermined identifier such as a null value or a "0" value to distinguish it.

[0096] Similarly, the tail word pointer vector includes a plurality of pointer elements corresponding to the total number of words in the text information, and each pointer element points to these words one by one according to the order of the words in the text information. Thus, each pointer element points to a single word in the text information, and the order of the pointer element in the tail word pointer vector is the same as the order of the word it points to in the text information. The pointer element in the tail word pointer vector stores the index label of the entity type to which the named entity corresponding to the word it points belongs in the form of a numerical value. Specifically, when the word pointed to by the pointer element belongs to the last word of a named entity, the pointer element is used to store the index label of the entity type to which the named entity belongs; if the word in the text information pointed to by a pointer element does not belong to the last word of any possible named entity, the pointer element can store a predetermined identifier such as a null value or a "0" value to distinguish it.

[0097] It can be seen that the structures adopted by the first word pointer vector and the last word pointer vector are the same, but the single words indicated by each pointer element when pointing to the same named entity are different. The pointer element in the first word pointer vector is used to indicate the position of the first word of the named entity in the text information, and stores the index label corresponding to the entity type to which the corresponding named entity belongs; the pointer element in the last word pointer vector is used to indicate the position of the last word of the named entity in the text information, and stores the index label corresponding to the entity type to which the corresponding named entity belongs. The first word pointer vector and the last word pointer vector are both unified pointer vectors, and both realize the centralized indication of multiple named entities contained in the text information.

[0098] The index tags are numerical tags mapped to the entity types that are compiled in advance to correspond to the entity types to which the named entities belong. Taking the various attribute information in the product information on the e-commerce platform as an example, the mapping vocabulary is shown below:

[0099] Entity Type Mapping Values Non-Entity 0 brand 1 Version 2 shape 3 Fabric 4 Use season 5 Function 6

[0100] This table is for example purposes only. As can be seen from the table, when a word in the text information does not belong to any named entity, it is a non-entity, represented by a value of "0"; when a word in the text information points to a named entity corresponding to an entity type such as "brand", "version", "style", "fabric", "season of use", or "function", it is represented by index labels such as "1", "2", "3", "4", "5", or "6".

[0101] Based on this, it is not difficult to understand that the head pointer vector and the tail pointer vector have determined the entity type corresponding to a named entity by storing the index label of the entity type to which a named entity belongs in the pointer element, that is, the mapping value in the above example table.

[0102] Adapting the above example, the construction of the first word pointer vector and the last word pointer vector is further illustrated, for example: Figure 2 As shown, there is a text information corresponding to the product title: "A certain company's 2021 summer anti-ultraviolet ice silk loose and comfortable windbreaker". According to this text information, the following two pointer vectors will be obtained:

[0103] First word pointer vector: [1,0,0,0,0,0,5,0,6,0,0,0,4,0,2,0,0,0,3,0]

[0104] Tail pointer vector: [0,1,0,0,0,0,0,5,0,0,0,6,0,4,0,2,0,0,0,3]

[0105] Among them, the order of each non-zero numerical value in the first word pointer vector in the vector corresponds to the single word in the text information that is sorted accordingly, and the single word is the first word of a named entity, that is, the first word. The numerical value itself is used as an index label and mapped to the entity type in the previous example table. Similarly, the order of each non-zero numerical value in the last word pointer vector in the vector corresponds to the single word in the text information that is sorted accordingly, and the single word is the last word of a named entity, that is, the last word. The numerical value itself is used as an index label and mapped to the entity type in the previous example table. Among them, according to the previous exemplary table, the "0" value is a single word corresponding to a non-entity type.

[0106] It can be seen that as long as the pointer generation network model is controlled to generate the first word pointer vector and the last word pointer vector according to the above rules, named entity extraction can be performed accordingly.

[0107] Step S1400: extract the character strings between the single words in the text information pointed to by the pointer elements of the same index tag in the two pointer vectors as named entities:

[0108] Continue to use Figure 2 From the example, it is not difficult to understand that for the entity type "brand", its mapping label is "1", and the first and last characters of the corresponding named entity appear in the first and second characters of the order of the text information, namely "a certain company". Similarly, for the entity type "function", its mapping label is "6", and the first and last characters of the corresponding named entity appear in the ninth and twelfth characters of the order of the text information, namely "ultraviolet protection". It can be seen that, as long as the two characters are taken as the first and last characters according to the different characters in the text information pointed to by the same index label in the first character pointer vector and the last character pointer vector, and the entire character string is extracted including the middle part, the character string is the named entity corresponding to the same index label. Based on this, multiple named entities in the text information can be extracted at one time.

[0109] Through the principles disclosed in this embodiment, it can be understood that the technical solution of the present application has a wide range of advantages over the prior art, including but not limited to:

[0110] First, the existing technology uses exhaustive keyword-based methods to identify named entities, which cannot cope with the rapidly growing data fields such as e-commerce product data and shows obvious technical limitations. For rapidly added product data such as e-commerce platforms, the extraction results are less accurate. Although the rule-based naming entity method can improve the accuracy, it can only mine results that match the rule template, and the recall rate is too low. Different from these shortcomings of the prior art, the present application generates double pointers based on the deep semantic information of the text information, and uses the two pointer vectors corresponding to the head and tail to extract the named entities in the text information, which can improve the accuracy of named entity recognition. While improving the confidence of named entity extraction, it can also improve the accuracy and recall rate. It is particularly suitable for application scenarios where the text information of named entities to be identified is growing rapidly, such as e-commerce platforms, title information processing in book management and other fields.

[0111] Secondly, the first word pointer vector and the last word pointer vector generated by the present application are constructed differently from the pointer structure in the prior art. Each pointer vector includes a plurality of pointer elements whose number is the same as the word length of the text information, so that each single word of the text information has a first word pointer element and a last word pointer element corresponding to it. When a named entity is identified from the text information, the first word pointer element corresponding to the first word in the named entity will store the index label corresponding to the entity type to which the named entity belongs. Similarly, the last word pointer element corresponding to the last word in the named entity will store the index label corresponding to the entity type to which the named entity belongs. It can be seen that the text information can be identified by a single first word pointer vector and a single last word pointer vector. A unified representation of all multiple named entities, a string of a named entity in text information, a single word in the text information pointed to by the pointer element of the index label of the entity type to which the named entity belongs in the first word pointer vector, and a single word in the text information pointed to by the pointer element of the index label of the entity type to which the named entity belongs in the last word pointer vector, that is, the last word. The string from the first word to the last word is the named entity, and the mapping relationship between the named entity and its entity type is also determined. It can be seen that the two pointer vectors play a role of unified representation and can be generated by an end-to-end model with high processing efficiency, which is convenient for determining the corresponding multiple named entities from the text information with a smaller amount of computation. For batch recognition tasks, the efficiency advantage is obvious.

[0112] In addition, due to the various advantages mentioned above, such as efficient calculation and accurate recognition, the present application is particularly suitable for e-commerce platforms, and can extract various commodity attribute information from the commodity information of the massive commodity objects on the e-commerce platform, thereby further reflecting its advantages in large-scale applications, such as reducing costs and quickly digesting massive data processing tasks.

[0113] See also Figure 3 In a further embodiment, for the present application, in combination with a more specific application scenario, the step S1100, obtaining text information of the named entity to be identified, includes the following steps:

[0114] Step S1110: Obtain the commodity object of the named entity to be identified:

[0115] The application scenario of this embodiment is the e-commerce field that the applicant is engaged in. In the e-commerce field, there are a large number of commodity objects. Whether it is the entire e-commerce platform or each store in the e-commerce platform, there are a large number of commodity objects in the corresponding commodity data. Therefore, once it is necessary to perform named entity recognition on these commodity objects, the commodity database can be called to obtain the commodity objects of the named entities to be recognized, as the target commodity objects for subsequent named entity extraction.

[0116] Individual extraction is also allowed. For example, when a user of a merchant instance publishes a newly launched product, he enters various product-related information, including product title text, product details text, etc., to form text information. After the user publishes the product, the backend server treats it as a product object for storage. In this process, the product object can also be automatically determined as the target product object that needs to be recognized by named entities.

[0117] Step S1120: extract the product title text and / or product details information of the product object as product information:

[0118] For a commodity object, in order to identify the named entity therein, it is necessary to obtain the text type information therein. As mentioned above, the text type information of the commodity object usually includes the title text and / or the commodity detail text (commodity detail information). In a simplified embodiment, only the title text or the commodity detail text can be selected. In a more thoughtful embodiment, both can be selected, depending on actual needs. These text type information constitutes the commodity information required for named entity recognition in this application.

[0119] Step S1130: clean the commodity information of the commodity object, and construct the text information after data cleaning into the text information of the named entity to be identified:

[0120] After obtaining the product information of the product object, various technical means commonly used in the field can be used to clean the product information, such as removing spaces, removing punctuation marks, etc., so as to remove relevant redundant information, which helps to improve the efficiency of named entity recognition.

[0121] This embodiment applies the technical solution of the present application to an e-commerce platform, and can provide named entity extraction services for commodity objects in the e-commerce platform. Relevant named entities are extracted as commodity attribute data, which can provide practical help for data entry of commodity objects, and can also unify the organizational dimensions of commodity attribute information on the e-commerce platform. In particular, for cross-border e-commerce based on independent sites, each merchant independently maintains its own site and commodity database. The e-commerce platform applies the technical solution of the present application to provide them with organizational management services for commodity attribute information, which can unify information management standards while reducing the organization and maintenance costs of commodity information for each merchant.

[0122] See also Figure 4 In a further embodiment, the step S1300 generates a first word pointer vector and a last word pointer vector according to the text feature vector, and calls a pre-trained pointer generation network model to implement the pointer generation network model, which performs the following steps:

[0123] Step S2100: Taking the text feature vector as the vector to be processed as input, an encoder is used to perform feature operation on the vector to be processed word by word to adapt to the text information, so as to generate a first word pointer vector corresponding to the current word:

[0124] In this embodiment, the Figure 5 , Figure 6 as well as Figure 7 The pointer generation network model shown is used to realize the generation of the first word pointer vector and the last word pointer vector. Figure 5 The network architecture is composed of a schematic framework after the pointer generation model of this application is expanded word by word in sequence and a text feature extraction model. Figure 6 and Figure 7 Schematic diagrams of the internal structures of an encoder for generating a first word pointer vector and an encoder for generating a last word pointer vector in the pointer generation network model respectively.

[0125] like Figure 5 As shown, after the text information is input into the text feature extraction model (taking the Bert model as an example), the corresponding text feature vector is extracted, and the text feature vector is quoted word by word in the process of encoding the text information word by word in the pointer generation network model of the present application.

[0126] When the pointer generation network model encodes the first word pointer vector corresponding to each single word in the corresponding text information, it takes the text feature vector as input, regards it as a vector to be processed, and then calls the vector to be processed as follows Figure 6 The encoder shown performs feature operations, specifically operations based on a multi-head attention mechanism, so as to fully refer to context information. To this end, in the encoding business logic corresponding to each single word, the first word pointer vector can be updated until the last single word is encoded, and the final first word pointer vector output can be obtained.

[0127] Step S2200: concatenate the text feature vector with the first word pointer vector corresponding to the current word to obtain a fused feature vector corresponding to the current word:

[0128] On the other hand, for each word in the text information, a tail word pointer vector needs to be generated for it. For this purpose, the encoding business logic needs to be executed for each word in the text information. In order to generate the tail word pointer vector with reference to the first word pointer vector, Figure 7As shown, in each encoding business logic of the tail word pointer vector, the first word pointer vector corresponding to the current single word is fused with the text feature vector extracted by the text feature extraction model to obtain a corresponding fused feature vector. The specific fusion method can be to concatenate the text feature vector with the corresponding first word pointer vector, and then perform linear transformation through a linear layer to obtain a corresponding fused feature vector. The formula is expressed as:

[0129]

[0130] in, represents the text feature vector obtained when corresponding to the i-th word, Y i,s Represents the first word pointer vector generated corresponding to the i-th word, Lin i That is the fused feature vector corresponding to the i-th word.

[0131] Step S2300: Taking the fused feature vector as the vector to be processed, an encoder is used to perform feature operation on the vector to be processed word by word to adapt to the text information, so as to generate a tail word pointer vector corresponding to the current word:

[0132] For the tail word pointer vector, its input is the fused feature vector corresponding to each word, that is, the comprehensive semantic vector that fuses the text feature vector and the first word pointer vector corresponding to the corresponding word. Then, it can be understood that when the pointer generation network model encodes the tail word pointer vector corresponding to each word in the corresponding text information, it takes the fused feature vector as input and regards it as a vector to be processed, and then is as follows Figure 7 As shown, feature operation is performed on the vector to be processed, specifically, operation is performed based on a multi-head attention mechanism so as to fully refer to context information. To this end, in the encoding business logic corresponding to each single word, the update of the tail word pointer vector can be realized, until the last single word is encoded, the final tail word pointer vector output can be obtained.

[0133] like Figure 5 As shown, the pointer generation network model of the present application is configured as a bidirectional structure. Therefore, for the generation of the first word pointer vector and the last word pointer vector, when encoding each single word, both use a hidden layer memory vector to memorize the intermediate feature information. The hidden layer memory vector will be referenced by the encoding business logic of the first word pointer vector and the last word pointer vector corresponding to each word, and is also updated word by word with reference to the two pointer vectors. Subsequent embodiments will further explain this, but those skilled in the art should be able to implement it by themselves according to the principles disclosed herein.

[0134] In this embodiment, by applying a bidirectional pointer generation network model, two pointer vectors are encoded word by word for the text information, and the generation of the two pointer vectors refers to each other's context information, making the entire model's understanding of semantics more accurate, thereby ensuring that accurate first-word pointer vectors and last-word pointer vectors can be obtained, laying a fundamental foundation for the extraction of named entities.

[0135] In a specific embodiment, the encoder used in the pointer generation network model of the present application is based on Figure 6 and Figure 7 It is not difficult to see that the structure shown is realized. Figure 6 and Figure 7 Has exactly the same structure, except that in the encoding business logic of the first word pointer vector, Figure 6 The encoder shown takes the text feature vector output by the text feature extraction model as direct input and encodes it as a vector to be processed. In the encoding business logic of the tail word pointer vector, Figure 7 The encoder shown takes the fused feature vector obtained by concatenating the first word pointer vector corresponding to the same single word and the text feature vector as input, and encodes it as a vector to be processed. That is to say, Figure 7 The encoder shown is relative to Figure 6 The encoder shown in the figure only adds a concatenation layer and a linear layer to the front end of the tail word pointer vector encoding. Figure 6 and Figure 7 The same structure in the above examples will be used to further explain the common parts of the two. Figure 6 and Figure 7 The encoder with the same structure executes the following for each word in the text information: Figure 8 The following steps are shown to obtain the corresponding first word pointer vector or last word pointer vector:

[0136] Step S3100, obtaining a hidden memory vector corresponding to the previous word, wherein the hidden memory vector is randomly initialized and referenced and then updated word by word:

[0137] When the encoder starts the encoding business logic for a single word in the text information, it first references the hidden memory vector generated by the previous single word. For the encoding business logic of the first word in the text information, there is a slight exception because there is no hidden memory vector corresponding to the previous single word. The hidden memory vector can be randomly initialized and referenced first. When the encoding is completed to update the first word pointer vector or the last word pointer vector, the hidden memory vector is updated accordingly. For other single words, the hidden memory vector corresponding to the previous single word is referenced in turn. The hidden memory vector is used to memorize the implicit feature information of the pointer corresponding to the current single word, and is constructed using a multi-head attention mechanism, which can be understood by those skilled in the art.

[0138] For the two encoders corresponding to the same word, each will reference the hidden memory vector corresponding to the previous word in order to update the corresponding first word pointer vector and last word pointer vector respectively. Then, the updated first word pointer vector and last word pointer vector will be used to jointly update the hidden memory vector corresponding to the current word and pass it to the two encoders corresponding to the next word.

[0139] It can be seen that the first word pointer vector, the last word pointer vector and the hidden memory vector are updated word by word according to each single word in the text information, and transmitted sequentially to form a serialized processing capability.

[0140] Step S3200: Perform a multi-head attention mechanism operation according to the vector to be processed, and calculate the normalized dot product of the vector to be processed and the hidden memory vector corresponding to the previous word as the addressing memory vector corresponding to the current word:

[0141] The encoder applies a multi-head attention mechanism. Therefore, the attention layer Attention is used to multiply the vector to be processed with the hidden memory vector corresponding to the previous word referenced, and then normalize it to obtain a vector composed of the normalized dot product as the addressing memory vector of the vector to be processed with respect to the hidden memory vector corresponding to the previous word.

[0142] Step S3300, linearly transform the addressing memory vector corresponding to the current single word and add it to the vector to be processed to obtain an intermediate vector:

[0143] On the basis of obtaining the addressing memory vector, a linear layer Linear is used to perform linear transformation on it, and it is activated through an activation layer. On this basis, the addressing memory vector is added to the vector to be processed to obtain an intermediate vector.

[0144] Step S3400: linearly transform the intermediate vector and perform regularization operation to obtain the first word pointer vector or the last word pointer vector corresponding to the current word:

[0145] The intermediate vector is further linearly transformed through a linear layer, and then regularized LN processing is applied. At this point, the first word pointer vector or the last word pointer vector corresponding to the current word can be obtained.

[0146] Step S3500: Update the hidden layer memory vector corresponding to the current word according to the first word pointer vector, the last word pointer vector and the hidden layer memory vector corresponding to the previous word for reference in the encoding process of the next word:

[0147] According to the first word pointer vector and the last word pointer vector generated by the above process and the hidden layer memory vector corresponding to the previous word, the hidden layer vector corresponding to the current word can be updated and output to the encoding process of the next word for reference. The specific operation process can be referred to the next embodiment.

[0148] See also Fig. 9 In a further embodiment, the step S3500, updating the hidden memory vector corresponding to the current word according to the first word pointer vector, the last word pointer vector and the hidden memory vector corresponding to the previous word for reference in the encoding process of the next word, comprises the following steps:

[0149] Step S3510: perform vector addition of the first word pointer vector and the last word pointer vector corresponding to the current word to obtain a sum vector:

[0150] For the two encoders corresponding to the current word, the first word pointer vector and the last word pointer vector respectively inputted thereto are vector-added to obtain a corresponding sum vector.

[0151] Step S3520: Classify and map the summed vector to obtain classification probabilities as corresponding smoothing weights:

[0152] For the sum vector, a binary classifier may be applied to classify the sum vector after matching it with a corresponding weight, and a Sigmoid function may be used for classification, so that the classification probability corresponding to the sum vector may be obtained as a smoothing weight.

[0153] Step S3530: Using the smoothing weight as a weight parameter, the sum vector and the hidden layer memory vector corresponding to the previous word are smoothly synthesized to obtain the hidden layer memory vector corresponding to the current word:

[0154] At this point, the smoothing weight can be used as a weight parameter to smoothly synthesize the sum vector and the hidden memory vector corresponding to the previous word to determine the extent to which the two hidden memory vectors are fused, and finally obtain the hidden memory vector corresponding to the current word.

[0155] Step S3540: pass the hidden memory vector corresponding to the current word to the encoding process of the next word for reference:

[0156] After the hidden memory vector corresponding to the current word is updated, it can be passed to the encoding process corresponding to the next word for reference. It can be understood that the updating principle of the hidden memory vector is to update and transmit word by word according to the word sequence of the text information, so that the entire pointer generation network model fully refers to the context information in the entire text information.

[0157] In order to make the description of the above two embodiments more formal, the encoding business logic of each individual word is further described in conjunction with mathematical formula as follows:

[0158] First, assuming that the product title text is X, define the output of the text feature extraction model (taking the Bert model as an example) as:

[0159] V bert =Bert(X)

[0160] Define the hidden memory vector corresponding to the current word stored in the pointer memory layer as Y i,c , representing the memory elements of the first word pointer vector and the last word pointer vector at the i-th moment (i.e., corresponding to the i-th word in the text information), which are used to store the implicit feature information shared by the first word pointer vector and the last word pointer vector.

[0161] Next, generate the first word pointer vector:

[0162] 1.1, such as Figure 6 The encoder structure is shown in Figure 1. The hidden memory vector Y corresponding to the previous word is input. i-1,c and text feature vector First, calculate the normalized dot product between the two Attention i ,get About Y i-1,c The addressing memory vector is:

[0163]

[0164] 1.2. Perform linear transformation on the addressing memory vector and activate the output, plus the input The value of , to obtain the intermediate vector after vector addition:

[0165]

[0166] 1.3. Finally, the intermediate vector is subjected to another layer of linear transformation and regularization processing to obtain the first word pointer vector corresponding to the current word:

[0167] Y i,s =LN(Linear(Add i ))

[0168] At this point, it is completed Figure 6 The encoder shown implements the encoding process corresponding to the single words in the text information.

[0169] Then, generate the tail word pointer vector:

[0170] 2.1, such as Figure 7 As shown in the editor structure, enter the hidden memory vector Y corresponding to the previous wordi-1,c , text feature vector The first word pointer vector Y corresponding to the current word i,s , first and Y i,s Perform vector concatenation and linear transformation to obtain the fused feature vector:

[0171]

[0172] The following calculation process Figure 6 The encoders shown are identical.

[0173] 2.2. Calculate the fusion feature vector Lin i The hidden memory vector Y corresponding to the previous word i-1,c Normalized dot product between the two Attention i , get the fusion feature about Y i-1,c The addressing memory vector is:

[0174]

[0175] 2.3. Perform linear transformation on the addressing memory vector and activate the output, and then add the fusion feature vector Lin i , get the intermediate vector after adding the vectors:

[0176]

[0177] 2.4. Finally, the intermediate vector is subjected to another layer of linear transformation and regularization processing to obtain the first word pointer vector corresponding to the current word:

[0178]

[0179] At this point, it is completed Figure 7 The encoder shown implements the encoding process corresponding to the single words in the text information.

[0180] Finally, after the two encoding business logics corresponding to the current word obtain the first word pointer vector and the last word pointer vector corresponding to the current word respectively, the hidden layer memory vector corresponding to the current word can be updated. Specifically, the following formula is applied:

[0181] Y i,c =αY′ i,c +(1-α)Y i-1,c

[0182] α=Sigmoid(W i Y i,c +b)

[0183] Y′ i,c =Sum(Yi,s , Y i,e )

[0184] Among them, α is a smoothing weight, which is used as a weight parameter and represents the hidden memory vector Y at the current moment. i,c In the example, the ratio of the first word pointer vector and the last word pointer vector information is retained, and 1-α represents the hidden memory vector Y at the current moment (current single word) i,c The proportion of memory information from the previous moment (previous word) retained.

[0185] It can be seen from the above formula that after the first word pointer vector and the last word pointer vector corresponding to the current word are summed to obtain the sum vector, the sum vector is matched with the corresponding weight and bias calculation, and then classified and mapped by the Sigmoid function to obtain the smoothing weight, and then the sum vector corresponding to the current word and the hidden layer memory vector corresponding to the previous word are smoothly synthesized using the smoothing weight as a weight parameter, so as to obtain the hidden layer memory vector corresponding to the current word, which can be passed to the encoding business logic reference corresponding to the next word.

[0186] It can be seen from the above embodiments that in the present application, the pointer generation network model applies multi-head attention mechanism (Attention), full connection (Linear), layer normalization (Layer Normalization) and other technologies in the encoder for generating the first word pointer vector and the last word pointer vector, and the specific process structure stacked is used to calculate and convert the pointer vector for the text feature vector output by the pre-trained model. Among them, the memory structure of the pointer vector is also realized by the hidden layer memory vector, which is used to fuse all the pointer feature information of the current moment (current single word) and the previous moment (previous single word) to enhance the entity extraction ability at the current moment. Therefore, the pointer network generation model of the present application can realize the effective conversion of the text feature vector of the text information into the first word pointer vector and the last word pointer vector. In addition, the pointer memory layer of the present application can transmit the feature information of the index tag extracted in the current text information, increase the information dimension that the subsequent named entity extraction can pay attention to, thereby enhancing the extraction ability of the model for named entities.

[0187] For example, there will be implicit connection information between product attribute information. For example, when the "brand" in the product-related named entity is "XXX Shoes" or "XXX Clothing", in the subsequent text of the same product information, there will be a high probability of named entities related to clothing and footwear, such as "fabric", "version", "style", etc., while there will be very few named entities related to electronic products, such as "resolution", "graphics card model", etc. Therefore, when judging whether the current word should be recognized as an entity boundary or which entity type corresponds to a named entity, adding the identified named entity feature information, on the one hand, increases the feature information dimension of entity recognition, and on the other hand, enhances the model's ability to judge the named entity boundaries and entity types at the current moment. That is, by increasing the memory information, the model's ability to extract and generalize the correct named entities is enhanced.

[0188] Based on the advantages described above, it can be understood that the pointer generation network model of the present application is particularly effective when applied to the e-commerce field to extract product attribute information from product information.

[0189] See also Fig.10 In a further embodiment, the step S1400, extracting the character strings between the single words in the text information pointed to by the pointer elements of the same index tag in the two pointer vectors as named entities, comprises the following steps:

[0190] Step S1410: Obtain index tags corresponding to the first word pointer element and the last word pointer element of different named entities according to the first word pointer vector and the last word pointer vector:

[0191] For example, let the text information be X, the pointer generation network model be M, and the model will output the first word pointer vector Y s , tail pointer vector Y e .

[0192] Apply the following formula to obtain the index labels corresponding to the first word pointer element and the last word pointer element of different named entities:

[0193] Index s =Argmax(Y s )

[0194] Index e =Argmax(Y e )

[0195] Still Figure 2 For example: Taking the product title text X = "A certain company's 2021 summer anti-ultraviolet ice silk loose and comfortable windbreaker" as an example, the index label structures of the first word pointer vector and the last word pointer vector obtained are:

[0196] Indexs =[10000050600040200030]

[0197] Index e =[01000005000604020003]

[0198] It is not difficult to understand that the two pointer vectors store the index labels of multiple named entities respectively.

[0199] Step S1420: query the preset mapping vocabulary to restore the entity type corresponding to the index tag:

[0200] Please review the typical embodiments of the present application and query the mapping vocabulary as illustrated in the typical embodiments of the present application to obtain the entity type corresponding to each index tag.

[0201] Step S1430: extract the character string corresponding to each named entity according to the boundaries defined by the single words in the text information pointed to by the first character pointer element and the last character pointer element of each named entity, and construct a list of corresponding relationships between entity types and named entities:

[0202] Get the boundary information of different named entities in the text information based on the first word pointer vector and the last word pointer vector:

[0203] Entity k ={(m,n)|m,n∈1,2,...,l∪m<n}

[0204] Among them, k∈1, 2, ..., K represents the entity type corresponding to the named entity, K represents the total number of entity types corresponding to the named entity, l represents the length of the text information X, and (m, n) is the boundary information that defines the boundary of the named entity in the text information.

[0205] On this basis, through the boundary information corresponding to each named entity, the character string in the text information X is intercepted accordingly, so as to obtain the data pair consisting of "entity type-named entity", and obtain the mapping relationship data between the entity type and the named entity, thus completing the named entity extraction.

[0206] For example, suppose Figure 2The sorting index value of the text information shown starts with 1, and the sorting index value of the original product title is from 1 to 20. After restoring the entity type and extracting the boundary information, the data list obtained is: {"Brand": [1, 2], "Use Season": [7, 8], "Function": [9, 12], "Fabric": [13, 14], "Type": [15, 16], "Style": [19, 20]}. Therefore, the corresponding relationship list of entity types and named entities obtained after entity extraction is ("Brand", "Company"), ("Applicable Season", "Summer"), ("Function", "Ultraviolet Protection"), ("Fabric", "Ice Silk"), ("Type", "Loose"), ("Style", "Windbreaker").

[0207] This embodiment further illustrates the specific method of extracting each named entity from the text information using the first word pointer vector and the last word pointer vector obtained by this application. According to this embodiment, the following understanding can be obtained more vividly: In the field of e-commerce, the traditional keyword method can only be listed in an exhaustive manner, and can only match the information in the vocabulary. For the rapidly growing commodity data, this method has obvious technical limitations. For the newly added commodity data, the extraction result accuracy is low, because the keyword does not necessarily refer to the commodity attribute, it may just be the description information that happens to appear with other words, but the keyword method cannot understand this semantic information. The method based on the rule template can improve the accuracy of attribute extraction, but the recall rate is too low, and only the results that match the design pattern can be mined. If you want to define a template that can match all commodity attributes, it is undoubtedly a costly and inefficient method. The method based on neural network, pre-training model, and first and last double pointers used in this application allows the neural network to learn and understand the semantic information of the entire commodity title text, automatically learn the boundary segmentation ability of commodity attribute entities and the discrimination ability of entity types, while improving the confidence of commodity attribute extraction, it can also improve the accuracy and recall rate.

[0208] See also Fig.11 In an extended embodiment, the pointer generation network model is pre-trained, and is trained as a named entity extraction service for product information in the e-commerce field, so as to use the named entities as product attribute data. Accordingly, the training process includes the following steps:

[0209] Step S4100: Obtain a sample data set, wherein the sample data set includes multiple groups of sample data, each group of sample data includes a title text of a commodity object and a first word pointer vector and a last word pointer vector corresponding to a named entity therein:

[0210] The sample data set used when the pointer network model of the present application is trained, wherein the sample data can be obtained from the network, and the sample data set can include a large number of groups of sample data, each group of sample data correspondingly includes a title text of a commodity object, and according to Figure 2 The example of the mapping vocabulary in the typical embodiment of the present application is used to pre-compile the first word pointer vector and the last word pointer vector corresponding to each title text, so as to provide the boundary information of multiple named entities in the title text through the two pointer vectors. The data set obtained in this step is represented as:

[0211] D: {(X j , Y j,s , Y j,e )|j∈1,...,n}

[0212] Where X j Represents the jth product title text, represented by l j The words are:

[0213]

[0214] Y j,s Indicates the entity type label corresponding to each word in the jth sample, represented by l j tags, expressed as:

[0215]

[0216] Y j,e Indicates the entity type label corresponding to each word in the j-th text, represented by l j tags, expressed as:

[0217]

[0218] Step S4200: Call each set of sample data to train the pointer generation network model, wherein the pointer generation network model predicts the corresponding first word pointer vector and last word pointer vector according to the text feature vector in the set of sample data:

[0219] A group of sample data is called once to train the pointer generation network model. According to the structure and principle of the pointer generation network model disclosed in the previous text of this application, it can be understood that the pointer generation network model will predict two pointer vectors corresponding to the title text, including the first word pointer vector and the last word pointer vector, based on the text feature vector extracted by the text feature extraction model in the group of sample data.

[0220] The extraction process of the first word pointer vector and the last word pointer vector by the pointer generation network model and the transformation formula thereof can be referred to as described above and will not be repeated here.

[0221] Step S4300: supervise the first word pointer vector and the last word pointer vector predicted by the pointer generation network model according to the first word pointer vector and the last word pointer vector in the trained set of sample data respectively and calculate the loss value; when the loss value is greater than a preset threshold, call the next set of sample data to continue iterative training of the pointer generation network model:

[0222] After the two pointer vectors corresponding to the product title text in a set of sample data are predicted, the two pointer vectors prepared in the set of sample data can be used to calculate the prediction loss one by one with the two predicted pointer vectors, and then the losses of the two are superimposed to obtain the loss value corresponding to the entire model. The weight parameters of the pointer network generation model are corrected by backpropagation based on the loss value, thereby realizing gradient update.

[0223] If the loss value of this training does not reach the expected preset threshold, it means that the pointer generation network model has not reached the convergence state. Based on this, another set of sample data can be further called to continue iterative training of the model. And so on, continuous iteration will eventually make the loss value close to 0 or reach the expected preset threshold. This can be regarded as model convergence, and the model can be put into use.

[0224] This embodiment provides an example of training the pointer generation network model implemented by the present application to serve the named entity extraction of product information in the e-commerce field, so that those skilled in the art can apply the pointer generation network model of the present application to similar fields by analogy. It can be understood that as long as the relevant sample data and mapping vocabulary of the corresponding field are applied, the pointer generation network model can be suitable for serving the named entity extraction needs of the corresponding field.

[0225] See also Fig.12A named entity recognition device provided to meet one of the purposes of the present application is a functional embodiment of the named entity recognition method of the present application, the device includes: a text acquisition module 1100, a feature extraction module 1200, a pointer generation module 1300, and an entity extraction module 1400, wherein the text acquisition module 1100 is used to obtain text information of a named entity to be identified, the text information including multiple words; the feature extraction module 1200 is used to extract deep semantic information of the text information and obtain its text feature vector; the pointer generation module 1300 is used to generate a first word pointer vector and a last word pointer vector according to the text feature vector, each pointer vector includes pointer elements pointing to each word in the text information in sequence, and the pointer elements corresponding to the first word and the last word of the named entity in the text information in the two pointer vectors store the index label of the entity type to which the named entity belongs; the entity extraction module 1400 is used to extract the character string between the words in the text information pointed to by the pointer elements of the same index label in the two pointer vectors as the named entity.

[0226] In a further embodiment, the text acquisition module 1100 includes: an object acquisition submodule, used to acquire the product object of the named entity to be identified; a text extraction submodule, used to extract the product title text and / or product detail information of the product object as product information; and a text optimization submodule, used to perform data cleaning on the product information of the product object, and construct the text information after data cleaning as the text information of the named entity to be identified.

[0227] In a further embodiment, the pointer generation module 1300 is implemented by calling a pre-trained pointer generation network model, which includes: a first-word pointer network, which is used to take the text feature vector as the input vector to be processed, and adapt to the text information to use the encoder to perform feature operations on the vector to be processed word by word to generate a first-word pointer vector corresponding to the current word; a vector concatenation network, which is used to concatenate the text feature vector with the first-word pointer vector corresponding to the current word to obtain a fused feature vector corresponding to the current word; and a last-word pointer network, which is used to take the fused feature vector as the input vector to be processed, and adapt to the text information to use the encoder to perform feature operations on the vector to be processed word by word to generate a last-word pointer vector corresponding to the current word.

[0228] In a specific embodiment, the encoder called in the first word pointer network and the last word pointer network includes: a hidden layer acquisition unit, which is used to obtain a hidden layer memory vector corresponding to the previous word, and the hidden layer memory vector is randomly initialized and referenced and then updated word by word; an addressing operation unit, which is used to perform a multi-head attention mechanism operation according to the vector to be processed, and calculate the normalized dot product of the vector to be processed and the hidden layer memory vector corresponding to the previous word as the addressing memory vector corresponding to the current word; an intermediate processing unit, which is used to linearly transform the addressing memory vector corresponding to the current word and add it to the vector to be processed to obtain an intermediate vector; a pointer generation unit, which is used to linearly transform the intermediate vector and superimpose a regularization operation to obtain the first word pointer vector or the last word pointer vector corresponding to the current word; a hidden layer update unit, which is used to update the hidden layer memory vector corresponding to the current word according to the first word pointer vector, the last word pointer vector and the hidden layer memory vector corresponding to the previous word for reference in the encoding process of the next word.

[0229] In a further embodiment, the hidden layer update unit includes: a sum processing subunit, which is used to perform vector addition of the first word pointer vector and the last word pointer vector corresponding to the current word to obtain a sum vector; a weight mapping subunit, which is used to perform classification mapping on the sum vector to obtain a classification probability as a corresponding smoothing weight; a smoothing synthesis subunit, which is used to use the smoothing weight as a weight parameter to smoothly synthesize the sum vector and the hidden layer memory vector corresponding to the previous word to obtain the hidden layer memory vector corresponding to the current word; and a hidden layer transfer subunit, which is used to transfer the hidden layer memory vector corresponding to the current word to the encoding process of the next word for reference.

[0230] In a further embodiment, the entity extraction module 1400 includes: a character extraction submodule, which is used to obtain the index labels corresponding to the first character pointer elements and the last character pointer elements of different named entities according to the first character pointer vector and the last character pointer vector; a mapping restoration submodule, which is used to query a preset mapping vocabulary and restore the entity type corresponding to the index label; a list construction submodule, which is used to extract the character strings corresponding to each named entity according to the boundaries defined by the single words in the text information pointed to by the first character pointer element and the last character pointer element of each named entity, and construct a list of correspondence between entity types and named entities.

[0231] In an extended embodiment, the pointer generation network model is placed in the following structure for pre-training, which structure includes: a sample acquisition submodule, used to acquire a sample data set, the sample data set includes multiple groups of sample data, each group of sample data includes a title text of a commodity object and the first word pointer vector and the last word pointer vector corresponding to the named entity therein; a sample training submodule, used to call each group of sample data to train the pointer generation network model, wherein the pointer generation network model predicts the corresponding first word pointer vector and last word pointer vector according to the text feature vector in the group of sample data; a gradient update submodule, used to supervise the first word pointer vector and the last word pointer vector predicted by the pointer generation network model respectively according to the first word pointer vector and the last word pointer vector in the trained group of sample data and calculate the loss value, and when the loss value is greater than a preset threshold, call the next group of sample data to continue iterative training of the pointer generation network model.

[0232] In order to solve the above technical problems, the present application also provides a computer device. Fig.13 As shown, a schematic diagram of the internal structure of a computer device. The computer device includes a processor, a computer-readable storage medium, a memory, and a network interface connected via a system bus. Among them, the computer-readable storage medium of the computer device stores an operating system, a database, and computer-readable instructions. The database may store a control information sequence. When the computer-readable instructions are executed by the processor, the processor can implement a named entity recognition method. The processor of the computer device is used to provide computing and control capabilities to support the operation of the entire computer device. The memory of the computer device may store computer-readable instructions. When the computer-readable instructions are executed by the processor, the processor can execute the named entity recognition method of the present application. The network interface of the computer device is used to connect and communicate with a terminal. Those skilled in the art can understand that Fig.13 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0233] In this embodiment, the processor is used to execute Fig.12 The memory stores the program code and various data required to execute the above modules or submodules. The network interface is used to transmit data between user terminals or servers. The memory in this embodiment stores the program code and data required to execute all modules / submodules in the named entity recognition device of this application, and the server can call the program code and data of the server to execute the functions of all submodules.

[0234] The present application also provides a storage medium storing computer-readable instructions. When the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the named entity recognition method of any embodiment of the present application.

[0235] The present application also provides a computer program product, including a computer program / instruction, which implements the steps of the method described in any embodiment of the present application when executed by one or more processors.

[0236] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments of the present application can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, the aforementioned storage medium can be a computer-readable storage medium such as a disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0237] To summarize, after the present application performs deep semantic extraction based on text information to obtain a text feature vector, two pointer vectors, namely a first-word pointer vector and a last-word pointer vector, are uniformly generated on the basis of the text feature vector, which are used to indicate the first and last words of the named entities in the text information respectively. These two pointer vectors can be used to extract multiple named entities in the text information in batches, thereby improving the efficiency of named entity recognition, and having high recall and accuracy rates. This is particularly suitable for application scenarios that require massive data processing, such as extracting relevant product attribute information from product information on an e-commerce platform. That is, the product information of massive product objects on an e-commerce platform can be batch processed using the method of this application, thereby extracting the named entities in each product object for constructing its product attribute information. It can be seen that the technical solution of this application has a very broad application prospect.

[0238] Those skilled in the art will appreciate that the various operations, methods, steps, measures, and schemes in the processes discussed in this application may be alternated, altered, combined, or deleted. Further, other steps, measures, and schemes in the various operations, methods, and processes discussed in this application may also be alternated, altered, rearranged, decomposed, combined, or deleted. Further, the steps, measures, and schemes in the prior art that are similar to those disclosed in this application may also be alternated, altered, rearranged, decomposed, combined, or deleted.

[0239] The above description is only a partial implementation method of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A method for named entity recognition, characterized in that: include: Acquire text information of a named entity to be identified, wherein the text information includes a plurality of words; Extracting deep semantic information of the text information to obtain its text feature vector; A first word pointer vector and a last word pointer vector are generated according to the text feature vector, each pointer vector includes pointer elements pointing to each word in the text information in sequence, and the pointer elements corresponding to the first word and the last word of the named entity in the text information in the two pointer vectors store the index label of the entity type to which the named entity belongs, wherein the pointer vector is generated by calling a pre-trained pointer generation network model, and the pointer generation network model performs the following steps: taking the text feature vector as the vector to be processed input, adapting to the text information, using an encoder to perform feature operation on the vector to be processed word by word to generate a first word pointer vector corresponding to the current word; splicing the text feature vector with the first word pointer vector corresponding to the current word to obtain a fused feature vector corresponding to the current word; taking the fused feature vector as the vector to be processed input, adapting to the text information, using an encoder to perform feature operation on the vector to be processed word by word to generate a last word pointer vector corresponding to the current word; The character strings between the single words in the text information pointed to by the pointer elements of the same index label in the two pointer vectors are extracted as named entities.

2. The method for named entity recognition according to claim 1, characterized in that: Obtaining text information of the named entity to be identified includes the following steps: Get the product object of the named entity to be identified; Extracting the product title text and / or product detail information of the product object as product information; Data cleaning is performed on the commodity information of the commodity object, and the text information after data cleaning is constructed into text information of the named entity to be identified.

3. The method for named entity recognition according to claim 1, characterized in that: The encoder performs the following steps for each word in the text information to obtain a corresponding first word pointer vector or last word pointer vector: Obtaining a hidden memory vector corresponding to the previous word, wherein the hidden memory vector is randomly initialized and referenced and then updated word by word; Perform a multi-head attention mechanism operation according to the vector to be processed, and calculate the normalized dot product of the vector to be processed and the hidden memory vector corresponding to the previous word as the addressing memory vector corresponding to the current word; The addressing memory vector corresponding to the current single word is linearly transformed and then added to the vector to be processed to obtain an intermediate vector; The intermediate vector is linearly transformed and superimposed with a regularization operation to obtain the first word pointer vector or the last word pointer vector corresponding to the current word; The hidden layer memory vector corresponding to the current word is updated according to the first word pointer vector, the last word pointer vector and the hidden layer memory vector corresponding to the previous word for reference in the encoding process of the next word.

4. The method for named entity recognition according to claim 3, characterized in that: The hidden layer memory vector corresponding to the current word is updated according to the first word pointer vector, the last word pointer vector and the hidden layer memory vector corresponding to the previous word for reference in the encoding process of the next word, including the following steps: Perform vector addition of the first word pointer vector and the last word pointer vector corresponding to the current single word to obtain a sum vector; Performing classification mapping on the summed vector to obtain classification probabilities as corresponding smoothing weights; Using the smoothing weight as a weight parameter, the sum vector and the hidden layer memory vector corresponding to the previous word are smoothly synthesized to obtain the hidden layer memory vector corresponding to the current word; The hidden memory vector corresponding to the current word is passed to the encoding process of the next word for reference.

5. The method for named entity recognition according to any one of claims 1 to 4, characterized in that: Extracting character strings between single words in the text information pointed to by the pointer elements of the same index label in the two pointer vectors as named entities includes the following steps: Obtain index labels corresponding to first word pointer elements and last word pointer elements of different named entities according to the first word pointer vector and the last word pointer vector; Query a preset mapping vocabulary to restore the entity type corresponding to the index label; According to the boundaries defined by the single words in the text information pointed to by the first word pointer element and the last word pointer element of each named entity, the character string corresponding to each named entity is extracted, and a corresponding relationship list between the entity type and the named entity is constructed.

6. The method for named entity recognition according to any one of claims 3 to 4, characterized in that: The pointer generation network model is pre-trained, and its training process includes the following steps: Acquire a sample data set, wherein the sample data set includes multiple groups of sample data, each group of sample data includes a title text of a commodity object and a first word pointer vector and a last word pointer vector corresponding to a named entity therein; Calling each set of sample data to train the pointer generation network model, wherein the pointer generation network model predicts the corresponding first word pointer vector and last word pointer vector according to the text feature vector in the set of sample data; According to the first word pointer vector and the last word pointer vector in the trained group of sample data, the first word pointer vector and the last word pointer vector predicted by the pointer generation network model are respectively supervised and the loss value is calculated. When the loss value is greater than a preset threshold, the next group of sample data is called to continue iterative training of the pointer generation network model.

7. A computer device comprising a central processing unit and a memory, characterized in that: The central processing unit is used to call and run the computer program stored in the memory to execute the steps of the method according to any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that: It stores a computer program implemented according to the method described in any one of claims 1 to 6 in the form of computer-readable instructions, and when the computer program is called and executed by a computer, the steps included in the corresponding method are executed.

9. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method described in any one of claims 1 to 6 are implemented.

Citation Information

Patent Citations

  • Entity linking method and device, equipment and storage medium

    CN111523326A

  • Named entity extraction method and device, electronic equipment and storage medium

    CN112069821A