Entity recognition model hot update method and its device, equipment, medium, and product

Abstract

The present application discloses a method for hot updating of an entity recognition model and its device, equipment, medium, and product, the method comprising: inputting the recognized text into the entity recognition model to obtain multiple named entities, and outputting the named entities that accurately hit the standard vocabulary; fuzzy matching the named entities that are not accurately hit with the standard vocabulary, and adding the named entities that achieve fuzzy matching to the standard vocabulary; calculating the information entropy scores of the named entities that do not achieve fuzzy matching based on the corpus composed of the recognized texts that have been historically processed by the entity recognition model, and screening out the target named entities and their corresponding target recognized texts whose information entropy scores exceed a preset threshold; adding the target recognized text to the data set to constitute the newly added data sample for iteratively training the entity recognition model according to the license information, and adding the target named entity to the standard vocabulary. The present application can realize the hot update of the entity recognition model and improve the model's ability to completely and accurately recognize named entities.

Description

Technical Field

[0001] The present application relates to the field of e-commerce information technology, and in particular to a method for hot updating an entity recognition model and its corresponding apparatus, computer equipment, computer-readable storage medium, and computer program product. Background Art

[0002] Named entity recognition plays an important role in search, recommendation, and user profile analysis in the e-commerce field. For example, when a user searches, the search box can guide the user to search for the product words they want to search for, thereby improving search efficiency. The associated words in the word association need to be based on the products in the product library, and the named entity recognition method can identify the product words. Based on the user's search, the tags of the search keywords, such as product words and brands, can be identified to improve the effect of search ranking. At the same time, the user's search for products, click on products, add to cart, order products, etc., can use the nested entity recognition method to identify the user's preferences for these behaviors for user profile analysis.

[0003] Named entity recognition (NER) can be treated as a sequence labeling problem and solved using a trained neural network model, where entity boundaries and category labels are jointly predicted. Due to the large deficiencies in sample balance and training in sequence labeling, entity recognition tasks have many problems such as low recall rate and inaccurate boundary recognition during inference. When used in production, the accuracy is low and the usage scenarios are relatively limited.

[0004] Therefore, how to improve named entity recognition technology so that it can identify named entities from the files to be recognized as completely and accurately as possible has become a problem that needs to be continuously evolved in this field. Summary of the invention

[0005] The primary purpose of the present application is to solve at least one of the above problems and to provide an entity recognition model hot update method and its corresponding device, computer equipment, computer-readable storage medium, and computer program product.

[0006] In order to meet the various objectives of this application, this application adopts the following technical solutions:

[0007] A method for hot updating an entity recognition model provided to meet one of the purposes of the present application comprises the following steps:

[0008] The recognized text is input into the entity recognition model to obtain multiple named entities, and the named entities that accurately hit the standard words in the standard vocabulary are output as recognition results;

[0009] Performing fuzzy matching on named entities that do not accurately match standard words in the standard word library with standard words in the standard word library, and adding the named entities that achieve fuzzy matching as standard words to the standard word library;

[0010] Calculating information entropy scores of named entities that fail to achieve fuzzy matching with standard words in the standard vocabulary, and screening out target named entities and their corresponding target recognized texts whose information entropy scores exceed a preset threshold;

[0011] According to the license information, the target recognized text is added to the data set to form a new data sample for iteratively training the entity recognition model, and the target named entity is added to the standard vocabulary as a standard word.

[0012] In a further embodiment, the recognized text is input into an entity recognition model to obtain multiple named entities, and the named entities that accurately hit the standard words in the standard vocabulary are output as recognition results, including the following steps:

[0013] Encode the recognized text to obtain its embedding vector, wherein the embedding vector includes the embedding vector of each character in the recognized text, and the embedding vector of each character includes the word vector of the character and all possible word segments of the character obtained by classifying and encoding the character according to the position of the character in the word segment;

[0014] The embedded vector is input into the text feature extraction module of the entity recognition model for representation learning to obtain a text feature vector representing the deep semantic information of the embedded vector;

[0015] Inputting the text feature vector into the conditional random field module of the entity recognition model for part-of-speech tagging to extract one or more named entities in the recognized text;

[0016] The extracted named entities are precisely matched with the standard vocabulary, and the named entities that are precisely matched with the standard words in the standard vocabulary are output as recognition results.

[0017] In a further embodiment, the information entropy scores of the named entities that do not achieve fuzzy matching with the standard words in the standard vocabulary are calculated, and the target named entities and their corresponding target recognized texts whose information entropy scores exceed a preset threshold are screened out, including the following steps:

[0018] Calling a corpus consisting of recognized texts processed by the entity recognition model within a preset historical period and their corresponding named entities that have not achieved fuzzy matching;

[0019] Calculating the weighted sum of the self-information entropy and the mutual information entropy of each named entity that has not achieved fuzzy matching as the corresponding information entropy score of each named entity;

[0020] Compare whether the information entropy scores of each named entity exceed a preset threshold, and store the target named entities whose information entropy scores exceed the preset threshold and their corresponding target recognized texts as a set to be labeled.

[0021] In a further embodiment, according to the license information, the target recognized text is added to the data set to form a new data sample for iteratively training the entity recognition model, and the target named entity is added as a standard word to the standard word library, including the following steps:

[0022] Acquire the permission information of the target recognized text and its corresponding target named entity;

[0023] According to the positive and negative sample types indicated by the licensing information of the target recognized text, the target recognized text is marked as a positive or negative sample, and added to the data set to constitute a new data sample for iterative training of the entity recognition model;

[0024] According to the standard word confirmation instruction contained in the permission information of the target named entity, the corresponding target named entity is added to the standard word library as a standard word.

[0025] In an extended embodiment, after the steps of adding the target recognized text to the data set to form a data sample for iteratively training the entity recognition model according to the permission information and adding the target named entity as a standard word to the standard word library, the following steps are included:

[0026] In response to a scheduled task or a user instruction, calling the newly added data sample in the data set to perform iterative training on the entity recognition model;

[0027] After completing the iterative training of the entity recognition model, the entity recognition model is restarted to identify named entities for the recognized text.

[0028] In an extended embodiment, after the step of inputting the recognized text into the entity recognition model to obtain multiple named entities and outputting the named entities that accurately hit the standard words in the standard vocabulary as the recognition results, the following steps are included:

[0029] Constructing a search expression containing the named entities in the recognition result;

[0030] Calling a product search engine to obtain product objects matching the search expression from a product database;

[0031] The commodity objects are constructed into a commodity list, and the commodity list is pushed to a search requester that provides the recognized text.

[0032] In an extended embodiment, after the step of inputting the recognized text into the entity recognition model to obtain multiple named entities and outputting the named entities that accurately hit the standard words in the standard vocabulary as the recognition results, the following steps are included:

[0033] The commodity object carrying the recognized text is labeled using the named entities in the recognition result, so that the named entities in the recognition result constitute the image label of the commodity object.

[0034] A hot update device for an entity recognition model provided to meet one of the purposes of the present application includes: an exact matching module, a fuzzy matching module, an entity screening module, and a data update module, wherein the exact matching module is used to input the recognized text into the entity recognition model to obtain multiple named entities, and output the named entities that accurately hit the standard words in the standard vocabulary as recognition results; the fuzzy matching module is used to fuzzy match the named entities that do not accurately hit the standard words in the standard vocabulary with the standard words in the standard vocabulary, and add the named entities that achieve fuzzy matching as standard words to the standard vocabulary; the entity screening module is used to calculate the information entropy score of the named entities that do not achieve fuzzy matching of the standard words in the standard vocabulary, and screen out the target named entities and their corresponding target recognized texts whose information entropy scores exceed a preset threshold; the data update module is used to add the target recognized text to the data set to constitute the newly added data sample for iterative training of the entity recognition model according to the permission information, and add the target named entity as a standard word to the standard vocabulary.

[0035] In a further embodiment, the precise matching module includes: an encoding processing submodule, which is used to encode the recognized text to obtain its embedding vector, wherein the embedding vector includes the embedding vector of each character in the recognized text, and the embedding vector of each character includes the character word vector and all possible word segments of the character, which are classified and encoded according to the appearance position of the character in the word segment to obtain the word vector; a representation learning submodule, which is used to input the embedding vector into the text feature extraction module of the entity recognition model for representation learning to obtain a text feature vector that represents the deep semantic information of the embedding vector; an entity extraction submodule, which is used to input the text feature vector into the conditional random field module of the entity recognition model for part-of-speech tagging to extract one or more named entities in the recognized text; a vocabulary matching submodule, which is used to accurately match the extracted named entities with the standard vocabulary, and output the named entities that achieve accurate matching with the standard words in the standard vocabulary as recognition results.

[0036] In a further embodiment, the entity screening module includes: a corpus calling submodule, which is used to call a corpus consisting of recognized texts processed by the entity recognition model within a preset historical period and their corresponding named entities that have not achieved fuzzy matching; a score calculation submodule, which is used to calculate the weighted sum of the self-information entropy and mutual information entropy of each named entity that has not achieved fuzzy matching as the corresponding information entropy score of each named entity; a target screening submodule, which is used to compare whether the information entropy score of each named entity exceeds a preset threshold, and store the target named entities whose information entropy scores exceed the preset threshold and their corresponding target recognized texts as a set to be labeled.

[0037] In a further embodiment, the data update module includes: a license acquisition submodule, which is used to obtain the license information of the target recognized text and its corresponding target named entity; a classification addition submodule, which is used to mark the target recognized text as positive or negative samples according to the positive or negative sample types indicated by the license information of the target recognized text, and add them to the data set to constitute new data samples for iterative training of the entity recognition model; a confirmation addition submodule, which is used to add the corresponding target named entity as a standard word to the standard vocabulary according to the standard word confirmation instruction contained in the license information of the target named entity.

[0038] In an extended embodiment, the entity recognition model hot update device of the present application also includes: a continuous training module, which is used to respond to scheduled tasks or user instructions and call the newly added data samples in the data set to perform iterative training on the entity recognition model; a model restart submodule, which is used to restart the entity recognition model after completing the iterative training of the entity recognition model to identify named entities for the recognized text.

[0039] In an extended embodiment, the entity recognition model hot update device of the present application also includes: a search expression module, which is used to construct a search expression containing the named entities in the recognition results; a search execution module, which is used to call the product search engine to obtain product objects matching the search expression from the product database; and a product push module, which is used to construct the product objects into a product list and push the product list to the search requester who provides the recognized text.

[0040] In an extended embodiment, the entity recognition model hot update device of the present application also includes: a product labeling module, which is used to use the named entities in the recognition results to label the product objects carrying the recognized text, so that the named entities in the recognition results constitute the portrait labels of the product objects.

[0041] 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 entity recognition model hot update method described in the present application.

[0042] A computer-readable storage medium is provided to meet another purpose of the present application, which stores a computer program implemented according to the entity recognition model hot update 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.

[0043] 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.

[0044] Compared with the prior art, the advantages of this application are as follows:

[0045] First, the named entities identified from the recognized text by the entity recognition model in this application are first accurately matched with the standard vocabulary, and then fuzzy matched. On this basis, for the named entities that fail to match the tag vocabulary, the corpus obtained by processing the recognized text over a period of time by the entity recognition model is referenced, and the information entropy score of each named entity that has not achieved a match is statistically quantified. The target named entity and its corresponding target recognized text for which manual licensing information is required are determined based on the information entropy score. Finally, based on the licensing information obtained through manual annotation, the corresponding target recognized text is added to the data set corresponding to the entity recognition model, and the corresponding target named entity is added to the standard vocabulary, so that the entity recognition model can continue to iterate training based on the newly added data, and continuously improve the model's ability to completely and accurately extract named entities from the recognized text.

[0046] Secondly, this application produces new data samples in the data set and new standard words in the standard vocabulary of the entity recognition model during the production stage. Subsequently, it can continue to iterate the training based on these new data to achieve version upgrades of the model, forming a hot update mechanism that can automatically evolve over a long period of time, which can not only improve the utilization rate of the entity recognition model, but also reduce its maintenance cost.

[0047] In addition, the present application obtains keywords from the recognized text based on the improvement of accuracy. The keywords can be used in scenarios such as online search, search keyword association, data profiling, recommended search, etc., and can obtain a more semantically accurate matching effect. Therefore, especially in the field of e-commerce, due to the large amount of text data related to commodity objects and the cost of text recognition is much lower than that of images, the use of the technical solution of the present application can obtain more obvious economies of scale. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] 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:

[0049] Figure 1 A flowchart of a typical embodiment of the entity recognition model hot update method of the present application;

[0050] Figure 2 A block diagram of the principle of implementing the entity recognition model hot update method of the present application;

[0051] Figure 3 A flowchart of the working process of the entity recognition model of this application;

[0052] Figure 4 This is a flow chart of the process of determining the set of standards to be standardized according to information entropy scoring in an embodiment of the present application;

[0053] Figure 5 A flowchart of a process of expanding a data set and a standard vocabulary according to pre-annotated license information in an embodiment of the present application;

[0054] Figure 6 A flowchart of one of the extended embodiments of the entity recognition model hot update method of the present application;

[0055] Figure 7 A flowchart of another extended embodiment of the entity recognition model hot update method of the present application;

[0056] Figure 8 This is a principle block diagram of the entity recognition model hot update device of the present application;

[0057] Fig. 9 A schematic diagram of the structure of a computer device used in this application. DETAILED DESCRIPTION

[0058] 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.

[0059] 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.

[0060] 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.

[0061] 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.

[0062] 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.

[0063] 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.

[0064] 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.

[0065] 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.

[0066] 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.

[0067] 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.

[0068] 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.

[0069] A method for hot updating of an entity recognition model 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.

[0070] See also Figure 1 and Figure 2 , Figure 1 This is a flow chart of the entity recognition model hot update method of the present application in its typical embodiment. Figure 2 This is a principle block diagram of the implementation principle of the technical solution of the present application. In this typical embodiment, the entity recognition model hot update method of the present application includes the following steps:

[0071] Step S1100: Input the recognized text into the entity recognition model to obtain multiple named entities, and output the named entities that accurately hit the standard words in the standard vocabulary as recognition results:

[0072] like Figure 2 As shown, the entity recognition model of the present application is used to implement named entity recognition on the recognized text. The recognized text can be obtained from a user request, or can be called from a database such as the product title of a product object in a product database or a local storage space such as a system clipboard. Depending on the different downstream tasks applied by the present application, the input text is flexibly obtained as the recognized text to perform named entity recognition, and one or more named entities are extracted from the recognized text to match with a preset standard vocabulary. The named entities that match the standard vocabulary are output as keywords in the recognition results for use as downstream tasks. In this step, when matching the named entities output by the entity recognition model with the standard vocabulary, an exact match method is adopted, that is, the character content of the named entity is required to be consistent with that of the standard word.

[0073] The entity recognition model includes a text feature extraction module and a conditional random field module. The text feature extraction module is usually implemented based on the underlying network architecture of Transformer. Basic network models suitable for processing sequence data such as Lattice LSTM and Bert can be used to represent the recognized text to obtain a corresponding text feature vector. The conditional random field module, i.e., the CRF model, performs part-of-speech tagging on the recognized text based on the text feature vector to extract the named entities therein. Since the technology used to implement named entity recognition by these models is well known to those skilled in the art, it will not be repeated here. Of course, the entity recognition model should be trained to a convergent state in advance before being put into use in this application so that it acquires the corresponding ability, which enables the model to represent the embedding vector obtained by encoding the recognized text of this application to obtain the text feature vector corresponding to the deep semantic information, and implement part-of-speech tagging on the basis of the text feature vector, and obtain each named entity according to the part-of-speech tagging result.

[0074] The standard vocabulary stores a large number of standard words, which are provided to the model output for matching, specifically matching the named entities extracted by the conditional random field module, so the standard words in the standard vocabulary are pre-defined and can be dynamically updated. The content of the standard words varies depending on the fields and scenarios served by the technical solution of the present application. For example, in the application scenarios related to the e-commerce field exemplified in the present application, the standard words can be brand words or label words. Correspondingly, when a named entity hits a standard word (or label word) in the standard vocabulary, it indicates that the named entity is a brand word (or label word). It can be seen from this that those skilled in the art can flexibly set the standard words in the standard vocabulary according to their specific task requirements.

[0075] The downstream tasks described depend on the specific application scenarios of the technical solutions of this application, such as online search, search keyword association, data profiling, recommended search, etc.

[0076] One example is that in an online search scenario, after the user submits his original search expression on the client device, the original search expression is used as the recognized text for named entity recognition to extract multiple keywords, and then the keywords are logically combined to form an optimized search expression to perform an online search.

[0077] In the second example, in the search keyword association scenario, when the user enters the search keyword, the text he has entered is used as the recognized text for named entity recognition, and then the background performs word association based on the keywords obtained by the named entity recognition, and the related associated words are clearly displayed to the user for selection.

[0078] Example three, in the data portrait scenario, taking the e-commerce field as an example, the text information such as the product title or product details of the product object can be directly obtained from the client device or from the product database as the identified text, and named entity recognition is performed to obtain the corresponding keywords, and the corresponding product objects are labeled, and these keywords are used as data portrait labels for the product objects.

[0079] Example 4: In the recommendation search scenario, still taking the e-commerce field as an example, the input of the client device can be obtained as the recognized text, and named entity recognition can be performed on it. The corresponding product objects can be searched in the product database based on the keywords obtained by recognition, and then a recommended list of product objects is constructed and pushed to the user end.

[0080] This indicates that the corresponding input text information can be obtained as the identified text according to different application scenarios, so as to start the named entity recognition process of this application. In fact, named entity recognition is a basic function in natural language processing technology, which can be widely applied to various occasions. Those skilled in the art are aware of this. Therefore, the scope of protection that should be covered by the creative spirit of this application should not be limited to the application field.

[0081] Step S1200: fuzzy matching is performed on the named entities that do not accurately match the standard words in the standard word library with the standard words in the standard word library, and the named entities that achieve fuzzy matching are added to the standard word library as standard words:

[0082] The named entities that did not achieve exact matching in the previous step, as pending named entities, can be further fuzzy matched with the standard words in the standard vocabulary in this step. Consider a situation, for example, for the named entity "Huaduo Leather Shoes", "Huaduo" is a brand word, which is pre-constructed as a standard word in the standard vocabulary. When the exact matching is performed in the previous step, since the standard word "Huaduo Leather Shoes" does not exist in the standard vocabulary, "Huaduo Leather Shoes" cannot be accurately named in the standard word in the standard vocabulary in the previous step, it is used as a pending named entity when fuzzy matching is performed in this step.

[0083] In the case of fuzzy matching, it can be performed with the help of preset regular rules, wildcards, etc., or by using the logical relationship of inclusion and inclusion. Using the above example, at this time, it is determined whether the pending named entity "Huaduo leather shoes" contains "Huaduo". If it does, the pending named entity "Huaduo leather shoes" can be added to the standard vocabulary as a new standard word to achieve the expansion of the standard vocabulary. Of course, if the pending named entity and any standard word in the standard vocabulary do not meet the rules corresponding to the fuzzy match, the fuzzy match cannot be achieved, and the pending named entity can be retained until the next step for further processing. Here, it is not difficult to understand that on the basis of using the standard vocabulary to complete the precise matching, the standard vocabulary is further used to implement fuzzy matching on the pending named entity, which can maximize the expansion of the semantic boundary of the named entity. By combining the semantic boundary to select the pending named entity from the recognition result of the entity recognition model, the expansion of the standard vocabulary can be achieved, so that the standard vocabulary can be updated and upgraded to provide the entity recognition model with the integrity and accuracy of the subsequent named entity recognition, and reduce the probability of misjudgment.

[0084] Step S1300, calculating the information entropy scores of the named entities that have not achieved fuzzy matching with the standard words in the standard vocabulary, and screening out the target named entities whose information entropy scores exceed a preset threshold and their corresponding target recognized texts:

[0085] The information entropy score is mainly composed of the mutual information entropy and self-information entropy of each named entity. The information entropy score can be calculated by referring to a data table that counts the word frequency of each named entity. The word frequency of each named entity, for example, can be counted based on the word segmentation results of the recognized text historically processed by the entity recognition model.

[0086] The self-information entropy of a named entity is used to measure the degree of freedom of the named entity. Assuming that a named entity appears N times in total, and there are n Chinese characters on its left, and each Chinese character appears N1, N2, ..., Nn times in sequence, then N = N1 + N2 + ... + Nn, so the probability of each Chinese character on the left of the word can be calculated, and the left neighbor entropy can be calculated according to the entropy formula. Similarly, its right neighbor entropy can be calculated. According to the summary and value of the left neighbor entropy and the right neighbor entropy, the self-information entropy of the named entity can be further determined. The smaller the entropy, the lower the degree of freedom, otherwise the higher the degree of freedom.

[0087] The mutual information entropy of a named entity is used to measure the solidification of the named entity. Assuming that the named entity is S, first calculate the probability P(S) of the word appearing in the named entity set, and then try all possible binary splits of S, that is, divide it into the left half sl and the right half sr and calculate P(sl) and P(sr). For example, there is one binary split for two-character words and two binary splits for three-character words. Next, calculate the minimum value of P(S) / (P(sl)×P(sr)) among all binary split schemes, and take the logarithm to measure the solidification. The larger the mutual information entropy, the higher the internal aggregation of the named entity, and the larger the mutual information entropy, the lower the internal aggregation of the named entity.

[0088] The self-information entropy of a named entity is fused to determine the information entropy score of the named entity, which is essentially to give a value assessment index for the named entity. The higher the information entropy score, the higher the possibility that it constitutes a central word with an independent status. Otherwise, the lower the information entropy score, the lower the possibility that it constitutes a central word. The simplest fusion method can be to directly add the mutual information entropy and the self-information entropy to determine it. In some optimized embodiments, the mutual information entropy and the self-information entropy can also be smoothed and weighted to obtain the information entropy score.

[0089] In the improved embodiment, for some specific application fields, appropriate flexibility can be made in the process of calculating the information entropy score. For example, by examining the characteristics of the title text of the commodity object in the e-commerce field where the present application is mainly applied, it can be seen that the title text is in the form of a title of an e-commerce product, including multiple semantically related and grammatically discrete words used to describe the e-commerce product, such as "Joyful gathering, disco, dance shoes, best-selling, new models, profitable and customized models", where, after being segmented, multiple named entities will be obtained in theory, such as: "Joyful gathering", "disco", "dance shoes", "best-selling", "new models", "profitable", "customized models", and by examining the characteristics of these named entities, it can be seen that they are basically stacked keywords, and there is a certain semantic correlation between words, but they are relatively discrete in grammar, and are more parallel, and the length of the entire title text will not be too long. According to these characteristics, when these title texts are used as the recognized texts and the named entities that have not achieved fuzzy matching are obtained, it is considered that the mutual information entropy will be affected by the length of the named entity. The longer the named entity, the greater the value of the mutual information entropy. Therefore, when calculating the mutual information entropy, n-party weights can be introduced for the calculation of the mutual information entropy, where n refers to the number of characters in the named entity.

[0090] After obtaining the information entropy score of each named entity by calculation, it is possible to determine whether the information entropy score of the named entity meets the requirements based on the preset threshold, and determine the named entity that meets the requirements as the target named entity, characterizing the named entity as containing a suspected standard word of higher value. Correspondingly, the recognized text containing the target named entity is also the target recognized text. Subsequently, the target recognized text and the target named entity can be provided to manual recognition.

[0091] The preset threshold can be determined based on the prior knowledge of those skilled in the art, or can be determined based on the statistical mean of the information entropy scores of all named entities. After the preset threshold is given, the information entropy score of each named entity is compared with it, and the named entity with an information entropy score higher than the preset threshold can be determined as the target named entity, otherwise it is not determined as the target hit entity.

[0092] Step S1400: Add the target recognized text to the data set to form a new data sample for iteratively training the entity recognition model according to the permission information, and add the target named entity as a standard word to the standard word library:

[0093] The target named entities and target recognized texts screened out in the previous step can be provided to relevant management users for manual annotation, so as to obtain the corresponding permission information generated by manual annotation. For the target recognized text, its sample type can be represented as permission information, and can be marked as a positive sample or a negative sample accordingly. Based on this, the target recognized text can be associated with its corresponding specific sample type and added to the data set required for iterative training of the entity recognition model to constitute a newly added data sample. For the target named entity, manual annotation can represent it as a true value or a false value, where the true value indicates that the target named entity can be regarded as a standard word, and thus, based on this permission information, the target named entity is added as a newly added standard word to the standard vocabulary required for matching the entity recognition model.

[0094] Therefore, even named entities and their corresponding recognized texts that fail to achieve precise matching with the standard vocabulary can be recycled, so that the standard vocabulary of the entity recognition model and the data set required for its training can be expanded and increased. When the training task of the entity recognition model is subsequently restarted, these data can be used to retrain the entity recognition model to achieve an upgrade of the entity recognition model. With the help of the ever-increasing data samples, the representation learning ability and the extensive recognition ability of named entities of the entity recognition model are continuously improved, reflecting the statistical recall rate advantage; with the help of the ever-growing standard vocabulary, the recognition accuracy of the recognition results output by the entity recognition model is continuously improved, reflecting the statistical precision rate advantage.

[0095] Through the disclosure of this typical embodiment, it can be understood that this application contains abundant positive effects, including but not limited to the following aspects:

[0096] First, the named entities identified from the recognized text by the entity recognition model in this application are first accurately matched with the standard vocabulary, and then fuzzy matched. On this basis, for the named entities that fail to match the tag vocabulary, the corpus obtained by processing the recognized text over a period of time by the entity recognition model is referenced, and the information entropy score of each named entity that has not achieved a match is statistically quantified. The target named entity and its corresponding target recognized text for which manual licensing information is required are determined based on the information entropy score. Finally, based on the licensing information obtained through manual annotation, the corresponding target recognized text is added to the data set corresponding to the entity recognition model, and the corresponding target named entity is added to the standard vocabulary, so that the entity recognition model can continue to iterate training based on the newly added data, and continuously improve the model's ability to completely and accurately extract named entities from the recognized text.

[0097] Secondly, this application produces new data samples in the data set and new standard words in the standard vocabulary of the entity recognition model during the production stage. Subsequently, it can continue to iterate the training based on these new data to achieve version upgrades of the model, forming a hot update mechanism that can automatically evolve over a long period of time, which can not only improve the utilization rate of the entity recognition model, but also reduce its maintenance cost.

[0098] In addition, the present application obtains keywords from the recognized text based on the improvement of accuracy. The keywords can be used in scenarios such as online search, search keyword association, data profiling, recommended search, etc., and can obtain a more semantically accurate matching effect. Therefore, especially in the field of e-commerce, due to the large amount of text data related to commodity objects and the cost of text recognition is much lower than that of images, the use of the technical solution of the present application can obtain more obvious economies of scale.

[0099] See also Figure 3 In a further embodiment, the step S1100, inputting the recognized text into the entity recognition model to obtain multiple named entities, and outputting the named entities that accurately hit the standard words in the standard vocabulary as the recognition results, includes the following steps:

[0100] Step S1110: Encode the recognized text to obtain its embedding vector, wherein the embedding vector includes the embedding vector of each character in the recognized text, and the embedding vector of each character includes the word vector of the character and all possible word segments of the character, which are obtained by classifying and encoding the character according to the position of the character in the word segment:

[0101] This step aims to encode the recognized text to achieve vectorization and obtain the corresponding embedding vector. There are various techniques for encoding text in the prior art, which can be flexibly implemented by those skilled in the art. In this application, a new encoding method is also recommended, so that the embedding vector contains the encoding information corresponding to each character of the recognized text, and the encoding information of each character contains the word vector of the character and the word vectors obtained by classifying and encoding all possible word segmentations of the character according to the position of the character in the word segmentation.

[0102] Specifically, for the recognized text, taking Chinese as an example, each character in it is encoded separately. When encoding each character, first obtain the corresponding word segmentation set for each character according to a preset dictionary, and each word segmentation set contains all possible word segmentations of its corresponding character in the recognized text. For example, taking the character '山' in '中山西路' as an example, the possible word segmentations in the word segmentation set determined according to the dictionary may include {"中山", "中山西", "中山西路", "山西", "山西路", "山"} and so on. From this example, it can be seen that there are various position situations for the single character '山' in each word segmentation of its word segmentation set, including the first character position, the last character position, the middle character position, and the single character position. Among them, for the convenience of understanding, the first character position can be marked as B, the last character position is marked as E, the middle character position is marked as M, and the single character position is marked as S. According to these different occurrence positions, different classifications can be constructed according to the different positions, and correspondingly obtained:

[0103] B = {"山西", "山西路"}

[0104] E = {"中山"}

[0105] M = {"中山西", "中山西路"}

[0106] S = {"山"}

[0107] Thus, it can be seen that according to the different positions where each character appears in its possible word segmentations, multiple corresponding word segmentation subsets can be divided. Then, word vectors are obtained and synthesized for different word segmentation subsets. The synthesis methods include taking the mean or weighted value of each word vector to achieve word vector compression; then combining the synthesized vector representation into the word vector of the character to form the corresponding vector representation of the character, and constructing the embedding vector corresponding to the recognized text from the vector representations of each character.

[0108] In this process, multiple word segmentation subsets are obtained according to the different positions at which the characters appear in their possible word segmentations, and then a comprehensive result of the word vectors of each word segmentation is obtained for each word segmentation subset. This comprehensive result independently represents the common features of the word segmentations at the corresponding positions. Finally, these common features are superimposed on the word vectors to form the vector representation corresponding to the single word, so that the vector representation of each single word integrates the common features corresponding to each appearance position. Therefore, through the representation of these common features, the representation of the meanings corresponding to different appearance positions is highlighted, so that the embedded vector of the recognized text finally obtained integrates all aspects of semantic information. Subsequent named entity recognition can naturally obtain richer semantics based on this.

[0109] Step S1120: input the embedding vector into the text feature extraction module of the entity recognition model for representation learning to obtain a text feature vector representing the deep semantic information of the embedding vector:

[0110] The text feature extraction module is preferably implemented using Lattice LSTM, which performs representation learning on the embedded vector of the recognized text obtained based on pre-encoding, with reference to the context, to obtain a corresponding text feature vector.

[0111] Step S1130: input the text feature vector into the conditional random field module of the entity recognition model for part-of-speech tagging to extract one or more named entities in the recognized text:

[0112] The text feature vector is input into the conditional random field module (CRF) for part-of-speech tagging. Under the action of the conditional random field module, the probability matrix output by Lattice LSTM and the state transition matrix of CRF are combined for prediction to complete the part-of-speech tagging. According to the part-of-speech tagging results, multiple named entities in the recognized text can be extracted.

[0113] In this embodiment, LSTM can also be replaced by a model based on the Transformer kernel such as Bert. In addition, although these models can also independently perform the task of part-of-speech tagging, combining them with conditional random fields can significantly improve the accuracy of named entity extraction, so they are recommended.

[0114] Step S1140: accurately match the extracted named entities with the standard word library, and output the named entities that achieve accurate matching with the standard words in the standard word library as recognition results:

[0115] When performing an exact match with the standard word library, it is only necessary to determine whether there is a named entity in the standard word library that is completely consistent with each extracted name entity. When a named entity is completely consistent with a standard word in the tag word library, the named entity hits the standard word. Finally, all named entities that achieve an exact match are output as recognition results, that is, the goal of extracting standard words from the recognized text is achieved.

[0116] In this embodiment, the encoding process of the recognized text that needs to be recognized by named entity recognition is mainly improved, the recognized text is vectorized, and the embedding vector corresponding to each character in the recognized text is obtained, so that the embedding vector contains the word vector of the character and all possible word segments of the character. The word vectors are obtained by classifying and encoding according to the appearance position of the character in the word segmentation. The representation of common features of each classification is strengthened through classification, thereby guiding the representation learning process of the entity recognition model, so that the deep semantic information obtained by the model representation learning can more accurately represent the semantic correlation between each character, so that the model performs part-of-speech tagging based on more accurate deep semantic information, accurately identifies each named entity on the basis of part-of-speech tagging, and obtains the corresponding standard word through precise matching.

[0117] Secondly, during the encoding process, this application takes into account the fact that Chinese has no obvious word boundaries and different word segments have different meanings under the character sequence. Therefore, each character is classified according to its specific position in the word segment. Thus, each classification expresses the meaning of the character's position, which is more in line with the characteristics of Chinese. Therefore, the embedded vector obtained by its encoding can make it easier for the entity recognition model to recognize named entities in Chinese, making the named entity recognition task based on Chinese more efficient and accurate.

[0118] In addition, since improvements to the encoding process improve the accuracy of named entity recognition, the total amount of data samples required for the training process of the corresponding entity recognition model can be reduced, making the model easier to train to a convergence state, improving model training efficiency and saving model training costs.

[0119] See also Figure 4 In a further embodiment, the step S1300, calculating the information entropy scores of the named entities that have not achieved fuzzy matching with the standard words in the standard vocabulary, and screening out the target named entities whose information entropy scores exceed a preset threshold and their corresponding target recognized texts, comprises the following steps:

[0120] Step S1310, calling a corpus consisting of recognized texts processed by the entity recognition model in a preset historical time and their corresponding named entities that have not achieved fuzzy matching;

[0121] The entity recognition model serves various downstream tasks on a daily basis, recognizes a large amount of recognized texts, and in the process of obtaining corresponding recognition results, these historically processed recognized texts can be cached to form a corpus for statistical information entropy scoring in this application, and a data table of word frequency required for calculating information entropy scoring can also be made based on the corpus. The corpus can reflect the specific data processed by the entity recognition model in the past time period. The span of this time period can be flexibly set by a technician in this field. For example, it can be set to a cache time of one day. For this reason, when the cache is called, at least the recognized text within the past 24 hours from the current moment can be obtained. Of course, the cache can also be used to store recognized texts generated in a historical time period of any length, and the entity recognition model can also obtain the corresponding recognized text from it according to a given time span. This application is only for the needs of dynamic hot updates, and exemplarily stipulates that the recognized texts generated between 0:00 and 24:00 the previous day are used as the corpus required for the statistical information entropy scoring of named entities in this application.

[0122] Step S1320: Calculate the weighted sum of the self-information entropy and the mutual information entropy of each named entity that has not achieved fuzzy matching as the corresponding information entropy score of each named entity:

[0123] First, the self-information entropy of each named entity is counted. The self-information entropy is the sum of the left neighbor information entropy and the right neighbor information entropy of the named entity:

[0124] Entropy represents the mathematical expectation of the self-information of a random variable X with distribution P(xi), and its calculation formula is a well-known formula in information theory, which is expressed in this application as:

[0125] H(x)=∑P xi LnP xi

[0126] In this step, the self-information entropy of the named entity extracted from the recognized text is first counted. In order to calculate the self-information entropy, the left and right neighbor entropies of the named entity need to be calculated first.

[0127] The left neighbor entropy calculation formula is:

[0128] EL=∑P(wi|str)LnP(wi|str)

[0129] Among them, Wi is the set of left-neighboring strings, str is the combination of strings, and P(.) is the conditional probability of wi under the given condition str.

[0130] The right neighbor entropy calculation formula is:

[0131] ER=∑P(wi|str)LnP(wi|str)

[0132] Among them, wi is the set of right-neighboring strings, str is the combination of strings, and P(.) is the conditional probability of wi under the given condition str.

[0133] According to the left neighbor entropy and right neighbor entropy formulas, the larger the wi set and the larger the value of p, the larger the entropy. The larger the entropy, the greater the amount of information represented.

[0134] On the basis of obtaining the left neighbor entropy and the right neighbor entropy, the following self-information entropy evaluation function is applied:

[0135] L=ERlnEL+ELlnER+|EL-ER|ln(ER+EL)-(ER+EL)ln|ER-EL|

[0136] It can be seen that when the information entropy of both sides is relatively large, the possibility that the named entity can be independently used as a central word is greater, and the evaluation function after optimizing the richness of the left and right neighbor words is better. This application has also achieved good results in actual tests.

[0137] Secondly, the mutual information entropy of each named entity is calculated:

[0138] In general, the mutual information of two discrete random variables X and Y can be defined as:

[0139]

[0140] Among them, p(x, y) is the joint probability distribution function of X and Y, and p(x) and p(y) are the marginal probability distribution functions of X and Y respectively.

[0141] The mutual information between two strings evaluates their significance together. The greater the mutual information, the stronger the mutual dependence, rather than being dispensable. Therefore, the following function is used to calculate the mutual information entropy:

[0142]

[0143] Among them, P(str) is the word frequency corresponding to the named entity.

[0144] Furthermore, since the mutual information between objects is affected by the length of the named entity (the longer the named entity, the larger the mutual information value), in the e-commerce scenario, the use of text is likely to be a stack of keywords, and the length of the text is basically not too long, so the weight of n parties is added to MI, and the correction is as follows:

[0145]

[0146] Then, for each named entity, its self-information entropy and its mutual information entropy are linearly fused with different weights to obtain the information entropy score of the corresponding named entity:

[0147] In order to obtain the sum of the self-information entropy and the mutual information entropy of each named entity and to facilitate the reconciliation, different weight parameters are assigned to the self-information entropy and the mutual information entropy respectively to calculate the information entropy score. Those skilled in the art can flexibly determine the different weights according to actual conditions. The corresponding formula is as follows:

[0148] score=α·L(W)+β·AMI(W)

[0149] Among them, L(W) is the self-information entropy, AMI(W) is the mutual information entropy, and α and β are their corresponding weight parameters.

[0150] According to this formula, the comprehensive information entropy corresponding to each named entity can be obtained as its corresponding information entropy score, so the named entity can be selected as the target named entity based on this.

[0151] Step S1330: compare whether the information entropy scores of each named entity exceed a preset threshold, and store the target named entity whose information entropy score exceeds the preset threshold and its corresponding target recognized text as a to-be-annotated set:

[0152] In order to achieve the optimization of named entities, a preset threshold can be introduced to determine whether the information entropy score of each named entity meets the requirements. When the information entropy score corresponding to a named entity is higher than the preset threshold, the named entity is determined as a target named entity with higher information value, and may be marked as a new standard word after subsequent manual inspection. Correspondingly, the recognized text where the target named entity is located is determined as the target recognized text, and may be a new data sample of a certain sample type in the data set required for entity recognition model training after subsequent manual inspection.

[0153] In order to facilitate manual annotation, the mapping relationship data of the target named entity and the target recognized text in which it is located can be constructed as a to-be-annotated set for ready call.

[0154] In the process of optimizing the named entities that fail to achieve fuzzy matching with the standard vocabulary to identify the target named entities, this embodiment applies the idea of ​​weighted fusion of self-information entropy and mutual information entropy in information theory, taking into account both the stability of the aggregation degree within the named entity itself and the degree of freedom of association between the named entity and other words, so that the new words discovered in this way are more semantically accurate, and the selection of pending named entities and pending recognized texts that need to be manually labeled is achieved.

[0155] See also Figure 5In a further embodiment, the step S1400, adding the target recognized text to the data set to form a new data sample for iteratively training the entity recognition model according to the permission information, and adding the target named entity as a standard word to the standard word library, includes the following steps:

[0156] Step S1410: Obtain the permission information of the target recognized text and its corresponding target named entity:

[0157] The target named entity and target recognized text obtained from step S1300 and its various variant embodiments are provided to manual labeling in advance, and the corresponding labeling information can be regarded as the permission information corresponding to the target named entity and the target recognized text. According to the corresponding permission information, the corresponding target named entity and target recognized entity can be processed.

[0158] Step S1420: According to the positive and negative sample types indicated by the permission information of the target recognized text, the target recognized text is marked as a positive or negative sample, and added to the data set to constitute a new data sample for iterative training of the entity recognition model:

[0159] When manually annotating the target recognized text, the main thing is to annotate its sample type in order to provide corresponding supervision labels for the training of the entity recognition model. Here, there are mainly two types of sample types, namely positive sample types and negative sample types. When manually annotating, the target recognized text can be determined as a positive sample or a negative sample based on actual conditions, such as whether the target recognized text contains a target named entity that can be used as a new standard word, thereby indicating the corresponding licensing information of the target recognized text.

[0160] According to this principle, for the target recognized text, the computer identifies the sample type contained in its licensing information, and accordingly associates the target recognized text with its specific sample type, and adds it to the data set as a new data sample in the data set. When the user subsequently trains the entity recognition model, its specific sample type constitutes the supervisory label of the entity recognition model.

[0161] Step S1430: according to the standard word confirmation instruction contained in the permission information of the target named entity, add the corresponding target named entity as a standard word to the standard word library:

[0162] When manually labeling the target named entity, the main thing is to mark whether it is suitable for constituting a new standard word in the standard vocabulary. For this purpose, its corresponding permission information can be expressed as a logical value, such as True and False. Accordingly, for the target named entity, the computer recognizes the logical value of its permission information and regards the logical value marked as True as a standard word confirmation instruction, thereby responding to the instruction and adding the corresponding target named entity as a new standard word to the standard vocabulary, thereby expanding the capacity of the standard vocabulary.

[0163] In this embodiment, manual annotation is allowed to further refine the target named entities and target recognized texts determined based on information entropy scores, and according to the annotated permission information, the target named entities and target recognized texts are respectively added to the standard vocabulary and data set corresponding to the entity recognition model, so that the entity recognition model can amplify the standard words required to realize its functions and the data samples required to realize its recognition capabilities, thereby building the basis for the hot update mechanism of the entity recognition model.

[0164] See also Figure 6 In an extended embodiment, after the step S1400, adding the target recognized text to the data set to form a data sample for iteratively training the entity recognition model according to the permission information, and adding the target named entity as a standard word to the standard word library, the following steps are included:

[0165] Step S1500: In response to a scheduled task or a user instruction, calling the newly added data sample in the data set to perform iterative training on the entity recognition model:

[0166] In this embodiment, for the entity recognition model, a scheduled task can be set to start the task of continuing to train the entity recognition model. Of course, it can also be started manually by triggering a user command. When the entity recognition model starts to implement training, it can iteratively call the newly added data samples in its data set to implement training until it is trained to a convergence state. Based on this, it is not difficult to understand that the target named entities generated by the recognized texts processed daily by the entity recognition model will be matched with the output of the conditional random field module as the newly added standard words in the standard vocabulary during the training process, so that the reference standard words in the training process are more comprehensive, which also helps to promote the rapid convergence of the model.

[0167] Step S1600: After completing the iterative training of the entity recognition model, restart the entity recognition model to identify named entities for the recognized text:

[0168] When the entity recognition model completes its iterative training, the entity recognition model can be republished to achieve version upgrade, and the entity recognition model can be restarted to identify named entities for the recognized text. This cycle continues to form a hot update mechanism for the entity recognition model.

[0169] This embodiment improves the closed loop of the hot update mechanism of the entity recognition model, so that the background can realize continuous cyclic upgrading of the entity recognition model. On the one hand, the entity recognition model is used daily to serve the entity recognition needs of the recognized text, and new standard words and new data samples are generated in the process of providing services, so as to expand the corresponding standard vocabulary and data set; on the other hand, the corresponding new standard words in the standard vocabulary are used regularly or irregularly for model training and precise matching in the production stage, and the new data samples in the data set are used for upgrading training of the entity recognition model. After the entity recognition model training is completed, it is republished and restarted to provide named entity recognition services, and the cycle continues, forming a self-update mechanism.

[0170] See also Figure 7 In an extended embodiment, after the step S1100, in which the recognized text is input into the entity recognition model to obtain multiple named entities, and the named entities that accurately hit the standard words in the standard vocabulary are output as recognition results, the following steps are included:

[0171] Step S1101: construct a search expression containing the named entities in the recognition result:

[0172] In this embodiment, the needs of online search application scenarios in the e-commerce field are met. The recognized text is the original search expression input by the user. The user completes the input of the recognized text on his terminal interface. After confirming the submission, it is obtained by the server. The server encodes the recognized text according to the technical solution of the present application and the relevant steps of the aforementioned embodiments to obtain its corresponding embedding vector, and then uses the entity recognition model to predict the named entity according to the embedding vector, and further obtains the corresponding recognition result through precise matching. The recognition result contains multiple named entities that are precisely matched with the corresponding standard words, which can be used as keywords required for the search.

[0173] On this basis, the various keywords can be combined according to preset rules, for example, the various keywords can be set to logical AND operations to form a new search expression.

[0174] Step S1102: calling a product search engine to obtain product objects matching the search expression from a product database;

[0175] According to the search expression corrected in the previous step, the product search engine is called to perform a search on the product database to obtain multiple product objects matching the search expression.

[0176] Step S1103: construct the commodity object into a commodity list, and push the commodity list to the search requester who provides the recognized text:

[0177] Furthermore, corresponding summary information of each matched product object, such as its product image, product title, product price, etc., can be obtained, the summary information can be encapsulated in a product list, and then the product list can be pushed to the client device of the user as the requester for display.

[0178] This embodiment further applies the technical solution of the present application to the online search application scenario in the field of e-commerce, demonstrating the technical advantages of the present application. It is not difficult to understand that due to the beneficial effects previously introduced in the present application, the keywords obtained through naming and recognition are more accurate. Therefore, the new search expression determined according to the preset logic and rules, and the product list obtained by searching in the product database, can better match the user's intention expressed in the recognized text, and realize accurate search of products.

[0179] In an extended embodiment, after step S1101, inputting the recognized text into the entity recognition model to obtain multiple named entities, and outputting the named entities that accurately hit the standard words in the standard vocabulary as the recognition results, the following steps are included: using the named entities in the recognition results to mark the commodity objects carrying the recognized text, so that the named entities in the recognition results constitute the portrait labels of the commodity objects:

[0180] In this embodiment, the need for data profiling of commodity objects in a commodity database in the e-commerce field is met. The commodity title and / or commodity details text of the commodity object is used as the material for determining the data profiling label, as the identified text. Then, through the relevant steps of the aforementioned embodiments, the identified text is encoded to obtain its corresponding embedding vector. An entity recognition model is used to predict and accurately match the recognition result based on the embedding vector. The recognition result contains multiple named entities that are consistent with the standard words in the standard vocabulary, constituting keywords. These keywords can be used as the labels required for data profiling of the commodity object. The commodity object can be directly labeled with these labels to complete the data profiling of the commodity object.

[0181] This embodiment further applies the technical solutions of other embodiments of the present application to the online search application scenario in the e-commerce field, demonstrating the technical advantages of the present application. It is not difficult to understand that due to the beneficial effects previously introduced in the present application, the keywords obtained by naming and identification are more accurate. Therefore, using these keywords as data portrait tags for commodity objects can quickly, efficiently and accurately implement data portrait labeling for commodity objects.

[0182] See also Figure 8 , an entity recognition model hot update device provided to meet one of the purposes of the present application is a functional embodiment of the entity recognition model hot update method of the present application, the device includes: an accurate matching module 1100, a fuzzy matching module 1200, an entity screening module 1300, and a data update module 1400, wherein the accurate matching module 1100 is used to input the recognized text into the entity recognition model to obtain multiple named entities, and output the named entities that accurately hit the standard words in the standard vocabulary as the recognition results; the fuzzy matching module 1200 is used to output the named entities that do not accurately hit the standard words in the standard vocabulary The named entities are fuzzy matched with standard words in the standard vocabulary, and the named entities that achieve fuzzy matching are added to the standard vocabulary as standard words; the entity screening module 1300 is used to calculate the information entropy score of the named entities that fail to achieve fuzzy matching with the standard words in the standard vocabulary, and screen out the target named entities and their corresponding target recognized texts whose information entropy scores exceed a preset threshold; the data update module 1400 is used to add the target recognized text to the data set to form a new data sample for iterative training of the entity recognition model according to the permission information, and add the target named entity as a standard word to the standard vocabulary.

[0183] In a further embodiment, the precise matching module 1100 includes: an encoding processing submodule, which is used to encode the recognized text to obtain its embedding vector, wherein the embedding vector includes the embedding vector of each character in the recognized text, and the embedding vector of each character includes the character vector and all possible word segments of the character, which are classified and encoded according to the appearance position of the character in the word segment to obtain the word vector; a representation learning submodule, which is used to input the embedding vector into the text feature extraction module of the entity recognition model for representation learning to obtain a text feature vector that represents the deep semantic information of the embedding vector; an entity extraction submodule, which is used to input the text feature vector into the conditional random field module of the entity recognition model for part-of-speech tagging to extract one or more named entities in the recognized text; a vocabulary matching submodule, which is used to accurately match the extracted named entities with the standard vocabulary, and output the named entities that achieve accurate matching with the standard words in the standard vocabulary as recognition results.

[0184] In a further embodiment, the entity screening module 1300 includes: a corpus calling submodule, which is used to call a corpus consisting of recognized texts processed by the entity recognition model in a preset historical period and their corresponding named entities that have not achieved fuzzy matching; a score calculation submodule, which is used to calculate the weighted sum of the self-information entropy and mutual information entropy of each named entity that has not achieved fuzzy matching as the corresponding information entropy score of each named entity; a target screening submodule, which is used to compare whether the information entropy score of each named entity exceeds a preset threshold, and store the target named entities whose information entropy scores exceed the preset threshold and their corresponding target recognized texts as a set to be labeled.

[0185] In a further embodiment, the data update module 1400 includes: a license acquisition submodule, which is used to obtain the license information of the target recognized text and its corresponding target named entity; a classification addition submodule, which is used to mark the target recognized text as positive or negative samples according to the positive or negative sample types indicated by the license information of the target recognized text, and add them to the data set to constitute new data samples for iterative training of the entity recognition model; a confirmation addition submodule, which is used to add the corresponding target named entity as a standard word to the standard vocabulary according to the standard word confirmation instruction contained in the license information of the target named entity.

[0186] In an extended embodiment, the entity recognition model hot update device of the present application also includes: a continuous training module, which is used to respond to scheduled tasks or user instructions and call the newly added data samples in the data set to perform iterative training on the entity recognition model; a model restart submodule, which is used to restart the entity recognition model after completing the iterative training of the entity recognition model to identify named entities for the recognized text.

[0187] In an extended embodiment, the entity recognition model hot update device of the present application also includes: a search expression module, which is used to construct a search expression containing the named entities in the recognition results; a search execution module, which is used to call the product search engine to obtain product objects matching the search expression from the product database; and a product push module, which is used to construct the product objects into a product list and push the product list to the search requester who provides the recognized text.

[0188] In an extended embodiment, the entity recognition model hot update device of the present application also includes: a product labeling module, which is used to use the named entities in the recognition results to label the product objects carrying the recognized text, so that the named entities in the recognition results constitute the portrait labels of the product objects.

[0189] In order to solve the above technical problems, the present application also provides a computer device. Fig. 9As 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 method for hot updating an entity recognition model. 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 method for hot updating an entity recognition model 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 will understand that Fig. 9 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.

[0190] In this embodiment, the processor is used to execute Figure 8 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 entity recognition model hot update device of this application, and the server can call the program code and data of the server to execute the functions of all submodules.

[0191] 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 entity recognition model hot update method of any embodiment of the present application.

[0192] 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.

[0193] 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).

[0194] In summary, this application establishes a hot update mechanism for the entity recognition model, so that the entity recognition model can continuously improve its named entity recognition capabilities through production, training, and reproduction. It is particularly suitable for processing Chinese information. The recognition results obtained thereby can improve the execution effects of various downstream tasks.

[0195] 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.

[0196] 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 hot updating of an entity recognition model, characterized in that: The steps include: The recognized text is input into the entity recognition model to obtain multiple named entities, and the named entities that accurately hit the standard words in the standard vocabulary are output as recognition results; Performing fuzzy matching on named entities that do not accurately match standard words in the standard word library with standard words in the standard word library, and adding the named entities that achieve fuzzy matching as standard words to the standard word library; Calculating information entropy scores of named entities that fail to achieve fuzzy matching with standard words in the standard vocabulary, and screening out target named entities and their corresponding target recognized texts whose information entropy scores exceed a preset threshold; According to the license information, the target recognized text is added to the data set to form a new data sample for iteratively training the entity recognition model, and the target named entity is added as a standard word to the standard word library, including: Acquire the permission information of the target recognized text and its corresponding target named entity; According to the positive and negative sample types indicated by the licensing information of the target recognized text, the target recognized text is labeled as positive and negative samples and added to the data set to constitute new data samples for iterative training of the entity recognition model; according to the standard word confirmation instructions contained in the licensing information of the target named entity, the corresponding target named entity is added as a standard word to the standard vocabulary.

2. The entity recognition model hot update method according to claim 1, characterized in that: The identified text is input into the entity recognition model to obtain multiple named entities, and the named entities that accurately hit the standard words in the standard vocabulary are output as recognition results, including the following steps: Encode the recognized text to obtain its embedding vector, wherein the embedding vector includes the embedding vector of each character in the recognized text, and the embedding vector of each character includes the word vector of the character and all possible word segments of the character obtained by classifying and encoding the character according to the position of the character in the word segment; The embedded vector is input into the text feature extraction module of the entity recognition model for representation learning to obtain a text feature vector representing the deep semantic information of the embedded vector; Inputting the text feature vector into the conditional random field module of the entity recognition model for part-of-speech tagging to extract one or more named entities in the recognized text; The extracted named entities are precisely matched with the standard vocabulary, and the named entities that are precisely matched with the standard words in the standard vocabulary are output as recognition results.

3. The entity recognition model hot update method according to claim 1, characterized in that: Calculating the information entropy scores of the named entities that do not achieve fuzzy matching with the standard words in the standard vocabulary, and screening out the target named entities whose information entropy scores exceed a preset threshold and their corresponding target recognized texts, including the following steps: Calling a corpus consisting of recognized texts processed by the entity recognition model within a preset historical period and their corresponding named entities that have not achieved fuzzy matching; Calculating the weighted sum of the self-information entropy and the mutual information entropy of each named entity that has not achieved fuzzy matching as the corresponding information entropy score of each named entity; Compare whether the information entropy scores of each named entity exceed a preset threshold, and store the target named entities whose information entropy scores exceed the preset threshold and their corresponding target recognized texts as a set to be labeled.

4. The entity recognition model hot update method according to any one of claims 1 to 3, characterized in that: After adding the target recognized text to the data set to form a data sample for iteratively training the entity recognition model according to the license information and adding the target named entity as a standard word to the standard word library, the following steps are included: In response to a scheduled task or a user instruction, calling the newly added data sample in the data set to perform iterative training on the entity recognition model; After completing the iterative training of the entity recognition model, the entity recognition model is restarted to identify named entities for the recognized text.

5. The entity recognition model hot update method according to any one of claims 1 to 3, characterized in that: After inputting the recognized text into the entity recognition model to obtain multiple named entities, and outputting the named entities that accurately hit the standard words in the standard vocabulary as the recognition results, the following steps are included: Constructing a search expression containing the named entities in the recognition result; Calling a product search engine to obtain product objects matching the search expression from a product database; The commodity objects are constructed into a commodity list, and the commodity list is pushed to a search requester that provides the recognized text.

6. The entity recognition model hot update method according to any one of claims 1 to 3, characterized in that: After inputting the recognized text into the entity recognition model to obtain multiple named entities, and outputting the named entities that accurately hit the standard words in the standard vocabulary as the recognition results, the following steps are included: The commodity object carrying the recognized text is labeled using the named entities in the recognition result, so that the named entities in the recognition result constitute the image label of the commodity object.

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.

Images