Model training method and device, equipment, medium and program product

Abstract

The invention discloses a model training method and device, equipment, a medium and a program product, which are used for improving the recognition effect of a model. The method can be applied to the fields of artificial intelligence, content auditing, computers and the like. Comprising the steps that a training sample and an initial recognition model are obtained, and the training sample comprises search terms, result page information corresponding to the search terms and label results corresponding to the search terms; performing feature extraction on the training sample to obtain a first feature representation and a second feature representation; determining a first weight value represented by the first feature and a second weight value represented by the second feature; performing weighted splicing on the first feature representation and the second feature representation based on the first weight value and the second weight value to obtain a to-be-predicted feature representation; performing prediction processing on the to-be-predicted feature representation to obtain a prediction result; and training the initial recognition model based on the prediction result and the label result to obtain a recognition model.

Description

Technical Field

[0001] The present application relates to the field of computers, and in particular to a model training method, device, equipment, medium and program product. Background Art

[0002] A search engine is a tool or service used to search for information on the Internet. It can find relevant content corresponding to the search terms from web pages and other resources on the Internet based on the search terms provided by the user, and present the most relevant results to the user. In order to avoid showing specific information to the user, it is usually necessary to identify whether the search term is a specific type of word.

[0003] Therefore, there is an urgent need to provide a method that can effectively identify search terms. Summary of the invention

[0004] The embodiments of the present application provide a model training method, apparatus, equipment, medium and program product for improving the recognition effect of the model.

[0005] In view of this, the present application provides, on the one hand, a model training method, comprising: obtaining a training sample and an initial recognition model, wherein the training sample includes a search term, result page information corresponding to the search term, and a label result corresponding to the search term; performing feature extraction on the training sample to obtain a first feature representation and a second feature representation, wherein the first feature representation is used to indicate a keyword feature representation of the search term and the result page information, and the second feature representation is used to indicate a feature representation of the result page information; determining a first weight value for the first feature representation and a second weight value for the second feature representation; weightedly concatenating the first feature representation and the second feature representation based on the first weight value and the second weight value to obtain a feature representation to be predicted; performing prediction processing on the feature representation to be predicted to obtain a prediction result; and training the initial recognition model based on the prediction result and the label result to obtain a recognition model.

[0006] On the other hand, the present application provides a model training device, including: an acquisition module, used to acquire a training sample and an initial recognition model, wherein the training sample includes a search term, result page information corresponding to the search term, and a label result corresponding to the search term;

[0007] A processing module is used to perform feature extraction on the training sample to obtain a first feature representation and a second feature representation, wherein the first feature representation is used to indicate a keyword feature representation of the search term and the result page information, and the second feature representation is used to indicate a feature representation of the result page information; determine a first weight value of the first feature representation and a second weight value of the second feature representation; perform weighted splicing of the first feature representation and the second feature representation based on the first weight value and the second weight value to obtain a feature representation to be predicted; perform prediction processing on the feature representation to be predicted to obtain a prediction result;

[0008] The training module is used to train the initial recognition model based on the prediction result and the label result to obtain a recognition model.

[0009] In one possible design, in another implementation of another aspect of the embodiment of the present application, the processing module is used to call the prediction network to perform prediction processing on the first feature representation to obtain a second weight value of the second feature representation, wherein the second weight value is used to indicate the attention of different types of search terms to the result page information, and the second weight value is greater than 0 and less than 1;

[0010] The first weight value is determined based on the second weight value, and the sum of the second weight value and the first weight value is 1.

[0011] In one possible design, in another implementation of another aspect of the embodiment of the present application, the processing module is used to call the fully connected network to perform fully connected processing on the first feature representation to obtain a first prediction value, wherein the fully connected network serves as the prediction network;

[0012] The first prediction value is mapped to obtain the second weight value.

[0013] In one possible design, in another implementation of another aspect of the embodiment of the present application, the processing module is used to perform keyword feature extraction on the search term in the training sample and the result page information corresponding to the search term to obtain the first feature representation;

[0014] Feature extraction is performed on the result page information corresponding to the search term in the training sample to obtain the second feature representation.

[0015] In one possible design, in another implementation of another aspect of the embodiment of the present application, the processing module is used to perform word segmentation processing on the top N result page information in the result page information corresponding to the search term to obtain a word segmentation set, where N is a positive integer;

[0016] Performing data statistics on each segmentation in the segmentation set to obtain a frequency statistical value corresponding to each segmentation in the segmentation set;

[0017] The top M words in the frequency statistics are used as the keyword set of the training sample, where M is a positive integer;

[0018] Concatenate each keyword in the keyword set with the search term to obtain a keyword sequence;

[0019] Feature extraction is performed on the keyword sequence to obtain the first feature representation.

[0020] In a possible design, in another implementation of another aspect of the embodiment of the present application, the processing module is used to segment the keyword sequence to obtain a first segmentation sequence to be processed;

[0021] Embedding the first word segmentation sequence to be processed to obtain a first word vector for each word in the first word segmentation sequence to be processed;

[0022] The first word vector is encoded to obtain the first feature representation.

[0023] In one possible design, in another implementation of another aspect of the embodiment of the present application, the processing module is used to perform word segmentation processing on the result page information corresponding to the search term to obtain a second word segmentation sequence to be processed;

[0024] Embedding the second word segmentation sequence to be processed to obtain a second word vector for each word in the second word segmentation sequence to be processed;

[0025] Using convolutional neural networks with X different convolution kernels to extract features from the second word vector, respectively, to obtain X feature representations, where X is a positive integer;

[0026] The X feature representations are concatenated to obtain a feature representation to be processed;

[0027] Feature extraction is performed on the feature representation to be processed through a fully connected layer to obtain the second feature representation.

[0028] In one possible design, in another implementation of another aspect of the embodiment of the present application, the acquisition module is used to call a data recognition model to perform recognition processing on the historical search terms to obtain the machine-labeled sample, and the data recognition model is a trained model;

[0029] Calling the data recognition model to perform recognition processing on the training samples of the historical iterative recognition model to obtain the historical annotated samples;

[0030] The machine-annotated sample and the historically-annotated sample are used as the training sample.

[0031] In one possible design, in another implementation of another aspect of the embodiment of the present application, the acquisition module is used to call the data recognition model to perform recognition processing on the historical search term to obtain a first classification result of the historical search term;

[0032] The machine-annotated sample is determined from the historical search terms according to the first classification result.

[0033] In one possible design, in another implementation of another aspect of the embodiment of the present application, the acquisition module is used to acquire a historical training sample of the historical iterative recognition model, where the historical training sample includes a historical label result;

[0034] Calling the data recognition model to perform recognition processing on each of the historical training samples to obtain a second classification result;

[0035] When the second classification result is the same as the historical label result, the historical training sample is used as the historical label sample;

[0036] When the second classification result is different from the historical label result, the historical training sample is deleted.

[0037] In one possible design, in another implementation of another aspect of an embodiment of the present application, the acquisition module is used to acquire pre-trained data, fine-tuning training data and a large language model, the pre-trained data includes a pre-trained search term and result page information corresponding to the pre-trained search term, the fine-tuning training data includes a fine-tuning search term and a label result of the fine-tuning search term, and the result page information corresponding to the pre-trained search term includes a title and an abstract;

[0038] The training module is used to train the large language model using the pre-training data to obtain a pre-training model;

[0039] The pre-training data is trained using the fine-tuning training data to obtain the data recognition model.

[0040] In one possible design, in another implementation of another aspect of the embodiment of the present application, the training module is used to encode the pre-training data to obtain a pre-training character string;

[0041] The large language model is called to train the pre-trained string according to the character prediction task to obtain the pre-trained model.

[0042] In one possible design, in another implementation of another aspect of the embodiment of the present application, the training module is used to call the pre-training data to perform classification processing on the fine-tuning training data to obtain a predicted label for the fine-tuning training data;

[0043] The pre-training data is trained using the predicted label and the label result of the fine-tuning training data to obtain the data recognition model.

[0044] In one possible design, in another implementation of another aspect of the embodiment of the present application, the acquisition module is used to acquire the search term to be processed and the result page information corresponding to the search term to be processed;

[0045] The processing module is used to perform feature extraction on the search term to be processed and the result page information corresponding to the search term to be processed based on the feature extraction network of the recognition model to obtain a third feature representation and a fourth feature representation, wherein the third feature representation is used to indicate a keyword feature representation of the search term to be processed and the result page information corresponding to the search term to be processed, and the fourth feature representation is used to indicate a feature representation of the result page information corresponding to the search term to be processed;

[0046] Determine a third weight value of the third feature representation and a fourth weight value of the fourth feature representation based on a weight classification network of the recognition model;

[0047] The feature fusion network based on the recognition model performs weighted concatenation on the third feature representation and the fourth feature representation based on the third weight value and the fourth weight value to obtain a feature representation to be processed;

[0048] The classification network based on the recognition model performs classification processing on the feature representation to be processed to obtain a classification result of the search term to be processed.

[0049] On the other hand, the present application provides a search term identification method, characterized by comprising: obtaining a search term to be processed and result page information corresponding to the search term to be processed;

[0050] Performing feature extraction on the search term to be processed and the result page information corresponding to the search term to be processed to obtain a third feature representation and a fourth feature representation, wherein the third feature representation is used to indicate a keyword feature representation of the search term to be processed and the result page information corresponding to the search term to be processed, and the fourth feature representation is used to indicate a feature representation of the result page information corresponding to the search term to be processed;

[0051] Determining a third weight value represented by the third feature and a fourth weight value represented by the fourth feature;

[0052] Performing weighted concatenation of the third feature representation and the fourth feature representation based on the third weight value and the fourth weight value to obtain a feature representation to be processed;

[0053] The feature representation to be processed is classified to obtain a classification result of the search term to be processed.

[0054] In one possible design, in another implementation of another aspect of the embodiment of the present application, when the classification result indicates that the search term to be processed is a specific type of term, the search term to be processed is deleted.

[0055] Another aspect of the present application provides a computer device, including: a memory, a processor, and a bus system;

[0056] Wherein, the memory is used to store programs;

[0057] The processor is used to execute the program in the memory, and the processor is used to execute the above-mentioned methods according to the instructions in the program code;

[0058] The bus system is used to connect the memory and the processor so that the memory and the processor can communicate with each other.

[0059] Another aspect of the present application provides a computer-readable storage medium, in which instructions are stored. When the computer-readable storage medium is run on a computer, the computer is enabled to execute the above-mentioned methods.

[0060] Another aspect of the present application provides a computer program product or a computer program, the computer program product or the computer program includes computer instructions, the computer instructions are stored in a computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the methods provided by the above aspects.

[0061] It can be seen from the above technical solution that the embodiment of the present application has the following advantages: different weight values ​​are added to the search terms in the training samples and the result page information corresponding to the search terms, so that different search terms in the training samples have different attention to the result page, thereby improving the recognition effect of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 A schematic diagram of the system architecture in the embodiment of the present application;

[0063] Figure 2 A schematic diagram of an architecture for model training in an embodiment of the present application;

[0064] Figure 3 This is an illustration of the data types collected by the data enrichment module in the embodiment of the present application;

[0065] Figure 4 This is a schematic diagram of an embodiment of the model training method in the embodiment of the present application;

[0066] Figure 5 A schematic diagram showing a search result page and search results in an embodiment of the present application;

[0067] Figure 6 A schematic diagram of a training process of a data recognition model in an embodiment of the present application;

[0068] Figure 7 A schematic diagram of an embodiment of a method for identifying model data in an embodiment of the present application;

[0069] Figure 8 This is a schematic diagram of a display interface for search terms and associated search terms in an embodiment of the present application;

[0070] Fig. 9 This is a schematic diagram of an embodiment of a model training device in an embodiment of the present application;

[0071] Fig.10 This is a schematic diagram of an embodiment of a server in an embodiment of the present application;

[0072] Fig.11 This is a schematic diagram of an embodiment of a terminal in an embodiment of the present application. DETAILED DESCRIPTION

[0073] The embodiments of the present application provide a model training method, apparatus, equipment, medium and program product for improving the recognition effect of the model.

[0074] The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein, for example. In addition, the terms "including" and "corresponding to" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0075] In the embodiments of the present application, the term "module" or "unit" refers to a computer program or a part of a computer program with a predetermined function, and works together with other related parts to achieve a predetermined goal, and can be implemented in whole or in part by using software, hardware (such as processing circuits or memories) or a combination thereof. Similarly, a processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be part of an overall module or unit that includes the function of the module or unit.

[0076] A search engine is a tool or service for searching for information on the Internet. It can find relevant content corresponding to the search terms from web pages and other resources on the Internet based on the search terms provided by the user, and present the most relevant results to the user. In order to avoid showing specific information to the user, it is usually necessary to identify whether the search term is a specific type of word. The current recognition scheme is usually to identify only based on the search term; or to identify based on the search term and the search results associated with the search term after a simple splicing, which will result in poor recognition results. Therefore, it is urgent to provide a method that can effectively identify search terms.

[0077] In order to solve the above problems, the present application provides the following technical solutions: obtaining a training sample and an initial recognition model, wherein the training sample includes a search term, result page information corresponding to the search term, and a label result corresponding to the search term; performing feature extraction on the training sample to obtain a first feature representation and a second feature representation, wherein the first feature representation is used to indicate a keyword feature representation of the search term and the result page information, and the second feature representation is used to indicate a feature representation of the result page information; determining a first weight value of the first feature representation and a second weight value of the second feature representation; weighted splicing the first feature representation and the second feature representation based on the first weight value and the second weight value to obtain a feature representation to be predicted; performing prediction processing on the feature representation to be predicted to obtain a prediction result; training the initial recognition model based on the prediction result and the label result to obtain a recognition model. In this way, different weight values ​​are added to the search terms and the result page information corresponding to the search terms in the training sample, so that different search terms in the training sample have different attention to the result page, thereby improving the recognition effect of the model.

[0078] The following describes several exemplary embodiments to illustrate the technical solutions of the embodiments of the present application and the technical effects produced by the technical solutions of the present application. It should be noted that the following embodiments can refer to, draw on or combine with each other, and the same terms, similar features and similar implementation steps in different embodiments will not be described repeatedly.

[0079] The embodiments of the present application can be applied to scenarios such as search word review and online content review. The following uses search word review as an example to illustrate an exemplary architecture of this scenario. Figure 1 The architecture includes a terminal 100, a server 200, a database 300 and a network 400. The terminal 100, the server 200 and the database 300 are connected via the network 400. The database 300 is used to store historical search terms and search terms associated with the historical search terms. Figure 1The number of terminals 100, servers 200 and databases 300 in the system shown is for example only. For example, the number of terminals 100, servers 200 and databases 300 may be multiple, and the embodiments of the present application do not limit the number of terminals 100, servers 200 and databases 300.

[0080] The terminal 100 communicates with the server 200 through a network, and the database 300 can be integrated on the server 200, or placed on a cloud or other server. In the above model training and downstream applications, the terminal 100 and the server 200 can interact with each other. For example, the user inputs the corresponding search term through the terminal 100, and the terminal 100 sends the search term to the server 200. The server 200 obtains the associated search terms associated with the search term and the corresponding result page information based on the search term.

[0081] The terminal 100 is a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smart watch, an intelligent voice interaction device, a smart home appliance, a vehicle-mounted terminal, etc., but is not limited thereto. The server 200 is an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDN), big data and artificial intelligence platforms.

[0082] Database 300, in short, can be regarded as an electronic filing cabinet - a place to store electronic files, where users can add, query, update, delete, etc. the data in the files. The so-called "database" is a collection of data that is stored together in a certain way, can be shared with multiple users, has as little redundancy as possible, and is independent of the application program. Database Management System (DBMS) is a computer software system designed for managing databases, generally with basic functions such as storage, interception, security, and backup. Database management systems can be classified according to the database model it supports, such as relational, Extensible Markup Language (XML); or according to the type of computer supported, such as server clusters, mobile phones; or according to the query language used, such as Structured Query Language (SQL), XQuery; or according to the performance impulse focus, such as maximum scale, maximum operating speed; or other classification methods. Regardless of the classification method used, some DBMS can cross categories, for example, supporting multiple query languages ​​at the same time.

[0083] Based on the above description, Figure 2 The schematic diagram of the architecture shown in the figure illustrates the model training method in this application: the model training architecture includes a data enrichment module, a data recognition model training module, a training sample generation module, and a recognition model training module. Figure 3 As shown, the data enrichment module is used to collect pre-training data and fine-tuning data (i.e., manually annotated data) of the data recognition model during the training process, and is also used to collect training data of the historical iteration model. The data recognition model training module is used to train a data recognition model that can generate machine-annotated data. In this module, the data recognition model is pre-trained and fine-tuned based on a large language model. The training sample generation module is used to generate machine-annotated samples using the trained data recognition model, and at the same time denoise the training data of the historical iteration model to obtain historical annotated samples; finally, the machine annotated samples and the historical annotated samples are used as training samples for the recognition model. The recognition model training module is used to perform model training on the training samples generated by the training sample generation module to obtain the final recognition model.

[0084] It will be understood that in the specific implementation of the present application, data related to training samples, historical search terms, and historical training samples are involved. When the above embodiments of the present application are applied to specific products or technologies, user permission or consent is required, and the collection, use, and processing of relevant data must comply with relevant laws, regulations, and standards of relevant countries and regions.

[0085] Combined with the above introduction, the following introduces the model training method in this application with the server as the execution subject, please refer to Figure 4 , an embodiment of the model training method in the embodiment of the present application includes:

[0086] 401. Obtain training samples and an initial recognition model, wherein the training samples include a search term, result page information corresponding to the search term, and a label result corresponding to the search term.

[0087] In this embodiment, the search term is a term to be classified and identified. For example, the search term may be a term typed by a user in an input box of a search engine. For another example, the search term may be a term selected by a user from a plurality of candidate terms. For example, if a user enters the term "apple" in an input box of a search engine, the search term is "apple".

[0088] The tag result is used to determine the classification type corresponding to the search term. The above classification type can be set by the developer based on work experience. Optionally, the classification type can be a type that brings positive effects to users, such as words related to inspiration and encouragement; the classification type can also be a type that brings negative effects to users, such as words related to negativity and sadness. It should be understood that the specific type content of the classification type should not be understood as a limitation on this application.

[0089] The result page information is the result items presented in the search result page corresponding to the search term. Usually, the number of search results corresponding to the search term is large. When multiple search results are displayed to the user, they are usually displayed in multiple search result pages in order from high to low search relevance. The search result page is a page used to carry multiple search results. Since they are displayed in order from high to low search relevance, generally speaking, the first page carrying multiple search results, that is, the search results displayed on the first search result page, have a higher relevance to the target term. For ease of description, the multiple search results corresponding to the search term can specifically be all the search results that appear in the first search result page corresponding to the search term.

[0090] Optionally, in a specific implementation, all search results appearing in the first search result page may be presented in a structured text format, and the format may specifically be:

[0091] The title of the first search result: (content of Title 1);

[0092] Introduction of the first search result: (Content of Introduction 1);

[0093] The title of the second search result: (content of title 2);

[0094] Introduction of the second search result: (content of Introduction 2); ......

[0096] Title of the Xth search result: (content of title X);

[0097] Introduction to the Xth search result: (content of introduction X).

[0098] When displayed in the interface, the search results page can be Figure 5 As shown, the search term in the search engine is "apple", and its corresponding search result pages may include the first search result page, the second search result page and the Xth search result page (i.e. 1, 2, 3 and X in the search result page indication); wherein the first search result page includes "Title: Apple-Official Website, Introduction: Brand New Mobile Phones on the Market", "Title: Apple (Rosaceae Malus), Introduction: Apple (Maluspumila Mill.), a deciduous tree plant of the Rosaceae Malus with a tall stem", "Title: Picture of Apple, Introduction: Picture A" and "Title: Hot-selling Apple, Introduction: XX Platform Selects Good Fruits".

[0099] It should be understood that the above description of multiple search results is an exemplary description for ease of description, and the multiple search results may include more or fewer search results. The specific number of search results should not be understood as a limitation to the present application.

[0100] It should be understood that each search result displayed in the result page information corresponds to a summary information, wherein the summary information is usually presented in a concise form, aiming to provide the user with a quick overview. Optionally, in one embodiment, the summary information may specifically include a title and a summary. Among them, the title is a search result item that can be linked. If the user clicks on the title, it can jump to the page of the result text information corresponding to the search result. The link can be a hyperlink (Hyperlink), an anchor link (AnchorLink), etc. It should be understood that the specific form of the link should not be understood as a limitation of this application. The summary is text information located below the title that briefly describes the title content. It is a brief summary or generalization of the corresponding search results.

[0101] In this application, in order to reduce the labor cost of obtaining training samples in model training, the sources of the training samples include machine-annotated samples and historically annotated samples. The machine-annotated samples are obtained by identifying and processing historical search terms based on the trained data recognition model, and the historically annotated samples are obtained by identifying and processing the training samples of the historical iterative recognition model based on the trained data recognition model.

[0102] In order to introduce search domain information and improve the generalization ability of the data recognition model, the data recognition model can be fine-tuned based on the pre-trained large language model and the training data constructed based on the search information in the search domain. A large language model (LLM) is a language model composed of an artificial neural network with many parameters (usually billions of weights or more), which is trained on a large amount of unlabeled text using self-supervised learning or semi-supervised learning. Large language models can be applied to a variety of fields such as natural language processing related tasks, intelligent customer service, content creation and assisted writing, and education.

[0103] In the training process of the data recognition model, the following can be used: Figure 6 The process shown:

[0104] a. Obtain pre-training data, fine-tuning data, and an initial large language model.

[0105] In this process, the pre-training data is used to pre-train the initial large language model. The sources of the pre-training data may include the following possible implementation solutions:

[0106] 1. Construct search results data of search terms and result page information. In an exemplary solution, 500,000 search terms of a specified type are screened, and then the result pages corresponding to the search terms are obtained through the search engine; then the titles and summaries of the result pages corresponding to each search term are parsed to construct training data of search terms, title sets, and summary sets. For example, when the search term is "apple", the corresponding result page information includes "Title: Apple-official website, Introduction: New mobile phone on the market", "Title: Apple (Rosaceae Apple), Introduction: Apple (Malus pumila Mill.), Rosaceae Apple deciduous tree plant with tall stems", "Title: Apple picture, Introduction: Picture A" and "Title: Hot-selling apple, Introduction: XX platform selects good fruit", and the training data obtained by parsing and construction can be as follows: (apple) + ("Apple-official website", "Apple (Rosaceae Apple)", "Apple picture", "Hot-selling apple") + ("New mobile phone on the market", "Apple (Malus pumila Mill.), Rosaceae Apple deciduous tree plant with tall stems", "Picture A", "XX platform selects good fruit"). The training data constructed by this solution can enable the large language model to learn the text information corresponding to the specified type of search terms.

[0107] 2. Construct search sequence data of search terms and user search history. In an exemplary scheme, the search terms searched by the user over a period of time are sorted to obtain a search term sequence; then random sampling is performed from the search term sequence to construct training data for the previous search terms and the current search terms. For example, the user's search term sequence over a period of time is as follows: ("apple", "siphon drainage", "decoration points", "TV series recommendation", "XX attractions", "electric kettle recommendation", "today's weather"); then random sampling is performed to obtain "TV series recommendation"; finally, the first two search terms of "TV series recommendation" are combined with "TV series recommendation" to generate search sequence data, that is, the search sequence data can be as follows: ("siphon drainage", "decoration points", "TV series recommendation"). The training data constructed by this scheme can enable the large language model to learn knowledge such as the user's click behavior.

[0108] At the same time, in order to ensure that the trained data recognition model can be effectively applied to downstream tasks, it is also necessary to design fine-tuning data. In this embodiment, the fine-tuning data can be high-precision manually labeled data. In an exemplary solution, the source of the fine-tuning data can be as follows:

[0109] Collect manually annotated data. The manually annotated data here includes search terms and the annotated results of the search terms, as well as the search results corresponding to the search terms and the annotated results corresponding to the search results. For example, the training data composition structure can be set as follows: (Apple, normal), (Apple-official website, normal), (xx movie, vulgar).

[0110] Then, we can make reasonable data matching for the collected manually labeled data, for example, increasing the proportion of vulgar samples so that the ratio of normal samples to vulgar samples is close to 1:1, which can ensure the recall model of the data recognition model.

[0111] Finally, the sample data after reasonable data matching will be manually labeled or inspected again to obtain the final fine-tuning data.

[0112] b. Pre-train the initial large language model based on the pre-training data to obtain a pre-trained model.

[0113] In the present application, the server can pre-train the initial large language model based on the next character prediction task (i.e., Next Token Prediction). That is, the server encodes the pre-training data collected in step a through an encoder to obtain the characters (i.e., tokens) of the pre-training data, and then inputs the above characters into the initial large language model, allowing the initial large language model to perform the next character prediction task, and performs model training based on the prediction results and the actual characters to obtain the pre-trained model. In this task, the initial large language model needs to predict the probability of occurrence of the next character. In this way, the pre-trained model finally trained can learn the language model in the text, thereby understanding the relationship between words.

[0114] c. Fine-tune the pre-trained model based on the fine-tuning data to obtain the data recognition model.

[0115] In this embodiment, after obtaining the pre-trained model after the preliminary pre-training, in order to be more suitable for downstream tasks, it is necessary to use fine-tuning data to further fine-tune the parameters of the pre-trained model. This task can also be called Scalable Fine-Tuning (SFT) training of a large model, which is a technology used in natural language processing. It fine-tunes the pre-trained language model to adapt it to a specific task. The main idea of ​​SFT is to fine-tune the model for a specific task based on a large pre-trained model.

[0116] In this embodiment, the specific process of fine-tuning can be as follows: the fine-tuning data is input into the pre-trained model, and the pre-trained model obtains the predicted classification result based on the input search term, and then uses the cross entropy loss function (it should be understood that the loss function here can also use other loss functions, such as exponential loss function, hinge loss function, etc., which are not limited here) to calculate the gap between the classification result and the corresponding label result, and then adjusts its internal parameters according to the goal of minimizing this gap as much as possible, so that the model's prediction result is as close to the real label as possible. This process usually requires multiple iterations, and each iteration will improve the performance of the model.

[0117] In an exemplary solution, during the model training process, the output template (i.e., prompt) of the model can be designed as follows:

[0118] {“input”: “The search terms (queries) recommended by the search can be divided into the following categories: normal, vulgar. “Vulgar” means that the query contains vulgar-related content, including vulgar descriptions, vulgar resources, vulgar websites and other information; “normal” means that there are no vulgar problems mentioned above. The known query and its result page keywords are: [ ]; the corresponding result page title and summary are: [ ]; please judge the query category based on the query text information and the result page information:”, “output”: “normal”, “vulgar”}

[0119] After the data recognition model is trained, the following scheme can be used to generate machine-annotated samples and historically annotated samples using the data recognition model:

[0120] In an exemplary solution, the data recognition model is applied to online historical search term recognition to obtain machine-labeled samples. That is, online search terms (queries) are obtained, including online random exposure queries, random click queries, exposure top queries, click top queries, and non-exposure queries, and the trained data recognition model is used to determine the classification results of these queries, and the classification results are used as the label results of the historical search terms. For example, if the training goal of the initial recognition model is to identify whether a search term has a violent label, then the search term with the recognition result of "violent" and the result page information corresponding to the search term need to be classified as the machine-labeled sample.

[0121] In another exemplary solution, the historical training samples of the historical iteration model are obtained, and then the data recognition model is used to perform recognition processing on the historical training samples to obtain the classification results after recognition; then the historical training samples are denoised according to the classification results and the label results of the historical training samples to obtain the historical labeled samples. Among them, the historical iteration model should have the same task objectives as the recognition model in this training process, such as online review of violent content, etc. For example, historical iteration model 1 is used to identify violent content, and its corresponding historical training samples are training sample set 1; historical iteration model 2 is also used to identify violent content, and its corresponding historical training samples are training sample set 2; the task of the initial recognition model in this embodiment is also to identify violent content, and the task of the trained data recognition model is also to identify violent content; then the trained data recognition model is used to perform recognition processing on each training sample in the training sample set 1 and the training sample set 2 to obtain the predicted label of each training sample; then the predicted label is compared with the original label, if the predicted label is the same as the original label, the training sample is classified as the historical labeled sample; if the predicted label is different from the original label, the training sample is deleted. This can improve the quality of training samples and thus improve the training effect of subsequent recognition models.

[0122] 402. Perform feature extraction on the training sample to obtain a first feature representation and a second feature representation, wherein the first feature representation is used to indicate a keyword feature representation of the search term and the result page information, and the second feature representation is used to indicate a feature representation of the result page information.

[0123] In this embodiment, after obtaining the training sample of the initial recognition model, the server extracts features from the training sample for subsequent model training. In order to improve feature learning of the training sample, when extracting features from the training sample, keyword features and specific text features can be learned simultaneously. The exemplary process can be as follows:

[0124] First, keyword features are extracted from the search terms and result page information of the training sample to obtain a first feature representation. The specific process may be as follows: segment the top N result page information in the result page information corresponding to the search term to obtain a segmentation set, where N is a positive integer; perform data statistics on each segmentation in the segmentation set to obtain a frequency statistic corresponding to each segmentation in the segmentation set; use the top M segmentations of the frequency statistic as the keyword set of the training sample, where M is a positive integer; concatenate each keyword in the keyword set with the search term to obtain a keyword sequence; perform feature extraction on the keyword sequence to obtain the first feature representation. At the same time, when extracting features from the keyword sequence, the method adopted may be as follows: segment the keyword sequence to obtain a first segmentation sequence to be processed; embed the first segmentation sequence to be processed to obtain the first word vector of each word in the first segmentation sequence to be processed; encode the first word vector to obtain the first feature representation.

[0125] In order to further learn the keyword features, when extracting the features of the keywords in the result page information, the text information and title information in the result page can also be counted for keywords. That is, the title and abstract are segmented respectively; then data statistics are performed separately, and finally the keywords corresponding to the title and the keywords corresponding to the abstract are obtained; finally, the two keywords are spliced ​​with the search term and embedded to obtain the final first feature representation.

[0126] For example, suppose the search term is: apple; the corresponding result page information is: "Title: Apple-official website, Introduction: Brand new mobile phone on the market", "Title: Apple (Rosaceae Apple), Introduction: Apple (Malus pumila Mill.), Rosaceae Apple deciduous tree with tall stems", "Title: Picture of apple, Introduction: Picture A and a bumper harvest of apples", and "Title: Hot-selling apples, Introduction: XX platform selects good fruits", and the training data obtained through parsing and construction can be as follows: (apple) + ("Apple-official website", "Apple (Rosaceae Apple)", "Picture of apple", "Hot-selling apple") + ("Brand new mobile phone on the market", "Apple (Malus pumila Mill.), Rosaceae Apple deciduous tree with tall stems", "Picture A a bumper harvest of apples", "XX platform selects good fruits"). When extracting keyword features from the training sample, the training sample may be preprocessed first. The preprocessing method may include unifying the text format and removing meaningless characters (such as performing half-width conversion and removing symbols). After the above processing, the text sequence of the exemplary training sample after text preprocessing may be as follows: "Apple Apple official website apple Rosaceae apple genus plant apple pictures hot sale Apple new mobile phone on the market Apple Malus pumila Mill Rosaceae apple genus deciduous tree plant stems high Apple bumper harvest XX platform selected good fruit". Then the text sequence is parsed for keywords. First, "Apple official website apple Rosaceae apple genus plant apple pictures hot sale apple" is segmented, and then all segmented words are counted for frequency, and finally the top three segmented words are selected as the keyword selection result of the title. At the same time, in order to ensure that each segmented word can be better learned and convenient for the model to use, each segmented word needs to be separated, such as using '#' to separate each segmented word. Similarly, for the abstract "Brand new mobile phone on the market, Apple Maluspumila Mill, Rosaceae, apple genus, deciduous tree, plant stem is high, apple bumper harvest, XX platform selects good fruit", it is also necessary to perform independent word segmentation; perform frequency statistics on all word segments; and finally select the top six word segments as the keyword selection results of the abstract. At the same time, in order to ensure that each word segment can be better learned and convenient for the model to use, each word segment needs to be separated, such as using '#' to separate each word segment. Finally, the keyword selection results of the title (i.e. titleterms_top), the keyword selection results of the abstract (i.e. xmlkeyterms_top), and the original query are concatenated, and then the concatenated text is embedded to obtain the word vector of the text sequence; finally, the word vector is encoded to obtain the final keyword feature representation.

[0127] Secondly, feature extraction is performed on the result page information of the training sample to obtain a second feature representation, and the specific process may be as follows: word segmentation is performed on the result page information corresponding to the search term to obtain a second word segmentation sequence to be processed; embedding representation is performed on the second word segmentation sequence to be processed to obtain a second word vector for each word in the second word segmentation sequence to be processed; feature extraction is performed on the second word vector using a convolutional neural network with X different convolution kernels to obtain X feature representations, where X is a positive integer; the X feature representations are concatenated to obtain a feature representation to be processed; and feature extraction is performed on the feature representation to be processed through a fully connected layer to obtain the second feature representation. In an exemplary scenario, assume that the training samples are as follows: (apple) + ("Apple-official website", "apple (Rosaceae apple plant)", "pictures of apples", "hot-selling apples") + ("brand new mobile phone on the market", "apple (Malus pumila Mill.), a deciduous tree plant of the genus Malus in the Rosaceae family with a tall stem", "Picture A a bumper harvest of apples", "XX platform selected good fruits"), then when extracting features from the result page information, feature extraction can be performed on "Apple Apple official website apple Rosaceae apple plant picture of apple hot-selling apple brand new mobile phone on the market apple Malus pumila Mill Rosaceae apple deciduous tree plant with a tall stem a bumper harvest of apples XX platform selected good fruits".

[0128] In the above feature extraction process, text processing is required. Therefore, in this embodiment, a model based on bidirectional encoder representation from transformers (BERT) or other text processing models can be used to process the above search terms and the result page information corresponding to the search terms.

[0129] Among them, in an exemplary scheme, when performing keyword feature extraction, feature extraction can be performed based on the pre-trained model of Macbert. First, the search terms and keywords of the result page information are spliced, and then the word segmentation is performed using BERT's own tools, and co-embedding (i.e., embedding) is performed through the vocabulary to map each word to a 768-dimensional word vector. Next, these word vectors are input into the multi-layer structure of BERT, and finally the embedding of the CLS position is obtained, which can represent the input features. Finally, the generated 768-dimensional feature vector is used as the text feature of the query and keyword.

[0130] When extracting features from the result page information, due to the long text content of the result page information, it takes a lot of time to use the BERT model for feature extraction. Therefore, a text convolutional neural network (textCNN) model can be used to extract long text features. It should be understood that the textCNN model can be trained using a manually annotated data set during training. In order to better achieve feature learning of the training samples of the initial recognition model in this embodiment, the textCNN model can set the training goal to perform two-classification of vulgar labels during the training process. Based on the textCNN model, the feature extraction process of the result page information can be as follows: first, the long text corresponding to the result page information in the training sample is segmented; then the word is converted into a vector by reading the word list and embedding; and then multiple CNNs with different convolution kernel sizes are used to extract features, and the two-dimensional features are compressed into one dimension by maximum pooling. Finally, the features of the three CNNs are spliced ​​together, extracted as 100-dimensional features through the fully connected layer, and the 100-dimensional features are used as the feature representation corresponding to the result page information. It should be understood that the dimensional description of the features involved above is only for illustration and is not limited here. In actual applications, it can be modified according to actual needs. At the same time, the number of CNNs in the textCNN model is only for illustration and is not limited here. In actual applications, it can be modified according to actual needs.

[0131] 403. Determine a first weight value represented by the first feature and a second weight value represented by the second feature.

[0132] After obtaining the first feature representation and the second feature representation of the training sample, the server needs to design a weight classification network to determine the different weight values ​​between the first feature representation and the second feature representation in order to adjust the degree of attention between the search term and the result page information. By judging the classification type of the search term, the search term and the result page information features are weighted so that different types of search terms pay different attention to the result page information. For example, for search terms in novels or movies, when the recognition model determines whether the search term is classified as vulgar, it will pay more attention to the result page information to judge; and for some search terms that can express vulgar information themselves, when the recognition model determines whether the search term is classified as vulgar, it will pay more attention to the information of the search term itself to judge. The specific operations can be as follows:

[0133] The prediction network is called to perform prediction processing on the first feature representation of the search term side to obtain a second weight value of the second feature representation, wherein the second weight value is used to indicate the attention of different types of search terms to the result page information, and the second weight value is greater than 0 and less than 1; the first weight value is determined based on the second weight value, and the sum of the second weight value and the first weight value is 1.

[0134] In an exemplary solution, the prediction network can be set as a fully connected layer, so that the first feature representation is predicted by the fully connected layer and the prediction result is mapped by the sigmoid function to obtain the second weight value of the second feature representation, and the value of the second weight value will be between 0 and 1. The specific representation can be as follows:

[0135] α=sigmoid(linear(F querβ ))

[0136] Wherein, the α is used to indicate the first weight value, the F query Used to indicate the first feature representation.

[0137] The first weight value can be expressed as follows: β=1-α.

[0138] 404. Perform weighted concatenation on the first feature representation and the second feature representation based on the first weight value and the second weight value to obtain a feature representation to be predicted.

[0139] In this embodiment, after the server determines the first weight value and the second weight value, the first feature representation and the second feature representation can be weighted and spliced ​​by the first weight value and the second weight value, so as to obtain the final classification input of the initial recognition model, that is, the feature representation to be predicted. In an exemplary scheme, the feature representation to be predicted after weighted splicing can be as follows:

[0140] F concat =concatenate(α·F summary , (1-α)F query )

[0141] Wherein, the α is used to indicate the first weight value, the F query It is used to indicate the first feature representation (i.e., the search term and the keyword feature of the result page corresponding to the search term). summary Used to indicate the second feature representation (ie, the text feature of the result page corresponding to the search term).

[0142] 405. Perform prediction processing on the feature representation to be predicted to obtain a prediction result.

[0143] In this embodiment, the server performs prediction processing on the feature representation to be predicted through the classification network of the initial recognition model to obtain the prediction result of the training sample.

[0144] It should be understood that the classification network of the initial recognition model can be designed as a binary classification network, and its specific structure can be designed according to actual needs, such as designing a binary classification network consisting of an input layer, a hidden layer and an output layer. In this structure, the hidden layer usually uses an activation function such as tanh, and the output layer uses a softmax function for classification.

[0145] 406. Train the initial recognition model based on the prediction result and the label result to obtain a recognition model.

[0146] The server calculates the loss value based on the label result of the training sample and the prediction result, and then trains the initial recognition model based on the loss value to obtain the recognition model. It should be understood that the training process of the initial recognition model is an iterative process. The server adjusts the various parameters of the recognition model in the iterative process so that the parameters reach the convergence condition of the model, and then outputs the final recognition model; or the server outputs the final recognition model after iterative training to the target number of times.

[0147] It should be understood that the above describes the process of recognizing model training. In actual applications, in order to improve the speed of online application, each search term and its corresponding associated terms can be recognized first, and the recognition results can be stored. Finally, when applying online, the recognition results of the corresponding search terms can be directly obtained from the stored results, and selectively displayed according to the recognition results. Recognition processing can also be performed online in real time, which is not limited here. Figure 7 The flowchart shown illustrates the recognition process of the recognition model with the server as the execution subject.

[0148] 701. Obtain a search term to be processed and result page information corresponding to the search term to be processed.

[0149] In this embodiment, the server accepts the pending search terms uploaded by the terminal device. The pending search terms obtained by the terminal device include the search terms currently input by the user and the associated search terms corresponding to the search terms. At the same time, the server can also obtain the result page information corresponding to the pending search terms. It should be understood that during the recognition process, the result page information is not displayed.

[0150] It should be understood that during the recognition process, the server can recognize the currently input search term and the associated search term separately. For example, the user enters the search term "apple" through the search engine input bar, and according to the user's historical operations, the associated search terms of "apple" may include "Apple mobile phone", "Apple price" and "apple fragrance", etc. During the recognition process, "apple", "Apple mobile phone", "Apple price" and "apple fragrance" can be recognized separately.

[0151] 702. A feature extraction network based on the recognition model performs feature extraction on the search term to be processed and the result page information corresponding to the search term to be processed to obtain a third feature representation and a fourth feature representation, wherein the third feature representation is used to indicate a keyword feature representation of the search term to be processed and the result page information corresponding to the search term to be processed, and the fourth feature representation is used to indicate a feature representation of the result page information corresponding to the search term to be processed.

[0152] It should be understood that the specific operation of this step is the same as the aforementioned step 402 and will not be repeated here.

[0153] 703. Determine a third weight value represented by the third feature and a fourth weight value represented by the fourth feature based on a weight classification network of the recognition model.

[0154] It should be understood that the specific operation of this step is the same as the aforementioned step 403 and will not be repeated here.

[0155] 704. A feature fusion network based on the recognition model performs weighted concatenation on the third feature representation and the fourth feature representation based on the third weight value and the fourth weight value to obtain a feature representation to be processed.

[0156] It should be understood that the specific operation of this step is the same as the aforementioned step 404 and will not be repeated here.

[0157] 705. Classify the feature representation to be processed based on the classification network of the recognition model to obtain a classification result of the search term to be processed.

[0158] It should be understood that the specific operation of this step is the same as the aforementioned step 405 and will not be repeated here.

[0159] In this embodiment, after obtaining the classification results of each pending search term, the server can associate the classification results of the pending search term with the pending search term and store them. Then, in actual online applications, if the search term is entered into the search engine, the classification results of the search term and the classification results of the associated search terms of the search term can be quickly found, and the associated search terms that can be displayed can be selected based on the classification results. For example, suppose the stored results are as follows: the classification result of the search term "apple" is normal, the classification result of its associated search term "Apple mobile phone" is normal, the classification result of "apple price" is normal, the classification result of "apple fragrance" is normal, and the classification result of "novel apple" is pornographic. Then, when an online user enters "apple" through the search bar of the search engine, the associated search terms displayed can be as follows: Figure 8 As shown, only "Apple", "Apple mobile phone", "Apple price" and "Apple fragrance" are displayed, and the related search term "novel apple" is no longer displayed.

[0160] The model training device in this application is described in detail below. Fig. 9 , Fig. 9 This is a schematic diagram of an embodiment of a model training device in an embodiment of the present application. The model training device 20 includes:

[0161] An acquisition module 201 is used to acquire a training sample and an initial recognition model, wherein the training sample includes a search term, result page information corresponding to the search term, and a label result corresponding to the search term;

[0162] The processing module 202 is used to perform feature extraction on the training sample to obtain a first feature representation and a second feature representation, wherein the first feature representation is used to indicate a keyword feature representation of the search term and the result page information, and the second feature representation is used to indicate a feature representation of the result page information; determine a first weight value of the first feature representation and a second weight value of the second feature representation; perform weighted splicing of the first feature representation and the second feature representation based on the first weight value and the second weight value to obtain a feature representation to be predicted; perform prediction processing on the feature representation to be predicted to obtain a prediction result;

[0163] The training module 203 is used to train the initial recognition model based on the prediction result and the label result to obtain a recognition model.

[0164] In an embodiment of the present application, a model training device is provided. Using the above device, different weight values ​​are added to the search terms in the training samples and the result page information corresponding to the search terms, so that different search terms in the training samples have different attention to the result page, thereby improving the recognition effect of the model.

[0165] Optionally, in the above Fig. 9On the basis of the corresponding embodiment, in another embodiment of the model training device 20 provided in the embodiment of the present application,

[0166] The processing module 202 is used to call the prediction network to perform prediction processing on the first feature representation to obtain a second weight value of the second feature representation, wherein the second weight value is used to indicate the attention of different types of search terms to the result page information, and the second weight value is greater than 0 and less than 1;

[0167] The first weight value is determined based on the second weight value, and the sum of the second weight value and the first weight value is 1.

[0168] In an embodiment of the present application, a model training device is provided. Using the above device, the keyword features of the search term and the result page are predicted through a prediction network to obtain a weight value represented by the feature of the result page, and the sum of the weight values ​​represented by the keyword feature and the feature of the result page is set to 1, so that the weight correlation between the keyword and the result page information can be easily determined, thereby improving the recognition effect of the model.

[0169] Optionally, in the above Fig. 9 On the basis of the corresponding embodiment, in another embodiment of the model training device 20 provided in the embodiment of the present application, the processing module 202 is used to call the first feature representation of the fully connected network to perform fully connected processing to obtain a first prediction value, wherein the fully connected network serves as the prediction network;

[0170] The first prediction value is mapped to obtain the second weight value.

[0171] In an embodiment of the present application, a model training device is provided. The above device is used to predict the first feature representation using a fully connected layer as a prediction network, thereby improving the feasibility of the scheme. At the same time, the predicted value is mapped to the mapping function so that the prediction result can be normalized to obtain a weight value ranging from 0 to 1, thereby improving the convergence speed of the model training and improving the power generation and stability of the model.

[0172] Optionally, in the above Fig. 9 On the basis of the corresponding embodiment, in another embodiment of the model training device 20 provided in the embodiment of the present application,

[0173] The processing module 202 is used to extract keyword features from the search words in the training sample and the result page information corresponding to the search words to obtain the first feature representation;

[0174] Feature extraction is performed on the result page information corresponding to the search term in the training sample to obtain the second feature representation.

[0175] In an embodiment of the present application, a model training device is provided. Using the above device, different types of feature extraction are performed on the search terms and result page information in the training samples, so that the feature representation between the search terms and result page information can be learned more comprehensively, thereby improving the recognition effect of the model.

[0176] Optionally, in the above Fig. 9 On the basis of the corresponding embodiment, in another embodiment of the model training device 20 provided in the embodiment of the present application,

[0177] The processing module 202 is used to perform word segmentation on the top N result page information in the result page information corresponding to the search term to obtain a word segmentation set, where N is a positive integer;

[0178] Performing data statistics on each segmentation in the segmentation set to obtain a frequency statistical value corresponding to each segmentation in the segmentation set;

[0179] The top M words in the frequency statistics are used as the keyword set of the training sample, where M is a positive integer;

[0180] Concatenate each keyword in the keyword set with the search term to obtain a keyword sequence;

[0181] Feature extraction is performed on the keyword sequence to obtain the first feature representation.

[0182] In an embodiment of the present application, a model training device is provided. Using the above device, the result page information with a high ranking is selected for word segmentation, and the word segmentation with a high statistical number ranking in the word segmentation is selected as the keyword, so that a more comprehensive feature representation of the keyword can be learned as much as possible while reducing the amount of data processing, thereby improving the recognition effect of the model.

[0183] Optionally, in the above Fig. 9 On the basis of the corresponding embodiment, in another embodiment of the model training device 20 provided in the embodiment of the present application, the processing module 202 is used to segment the keyword sequence to obtain a first segmentation sequence to be processed;

[0184] Embedding the first word segmentation sequence to be processed to obtain a first word vector for each word in the first word segmentation sequence to be processed;

[0185] The first word vector is encoded to obtain the first feature representation.

[0186] In the embodiment of the present application, a model training device is provided. The above device is used to perform feature learning on the word segment to be processed using an embedded representation, so that high-dimensional information can be mapped to a low-dimensional continuous space, thereby reducing computational complexity and storage requirements, and at the same time, the potential structure and relationship in the text data can be captured, so as to learn more semantic relationships, thereby improving the recognition effect of the model.

[0187] Optionally, in the above Fig. 9 On the basis of the corresponding embodiment, in another embodiment of the model training device 20 provided in the embodiment of the present application, the processing module 202 is used to perform word segmentation processing on the result page information corresponding to the search term to obtain a second word segmentation sequence to be processed;

[0188] Embedding the second word segmentation sequence to be processed to obtain a second word vector for each word in the second word segmentation sequence to be processed;

[0189] Using convolutional neural networks with X different convolution kernels to extract features from the second word vector, respectively, to obtain X feature representations, where X is a positive integer;

[0190] The X feature representations are concatenated to obtain a feature representation to be processed;

[0191] Feature extraction is performed on the feature representation to be processed through a fully connected layer to obtain the second feature representation.

[0192] In an embodiment of the present application, a model training device is provided. The above device is used to extract features of text information using convolutional neural networks of various specifications, and finally multiple features are spliced ​​to obtain the final feature representation, so that feature representations of different granularities can be captured, thereby improving the expressiveness of the model. When splicing feature representations of different granularities, the use of multi-semantic information can be realized, the gradient message problem of model training can be solved, the training efficiency can be improved, and overfitting can be reduced.

[0193] Optionally, in the above Fig. 9 On the basis of the corresponding embodiment, in another embodiment of the model training device 20 provided in the embodiment of the present application,

[0194] The acquisition module 201 is used to call the data recognition model to perform recognition processing on the historical search terms to obtain the machine-labeled samples, and the data recognition model is a trained model;

[0195] Calling the data recognition model to perform recognition processing on the training samples of the historical iterative recognition model to obtain the historical annotated samples;

[0196] The machine-annotated sample and the historically-annotated sample are used as the training sample.

[0197] In an embodiment of the present application, a model training device is provided. By using the above device, training samples are generated through a trained data recognition model, and historical training samples are also processed for recognition to generate the training samples, which can effectively reduce the labor cost of obtaining training samples, increase the number of training samples, and thus improve the recognition effect of the model.

[0198] Optionally, in the above Fig. 9 On the basis of the corresponding embodiment, in another embodiment of the model training device 20 provided in the embodiment of the present application,

[0199] The acquisition module 201 is used to call the data recognition model to perform recognition processing on the historical search word to obtain a first classification result of the historical search word;

[0200] The machine-annotated sample is determined from the historical search terms according to the first classification result.

[0201] In an embodiment of the present application, a model training device is provided. By using the above device, a large number of historical search terms can be identified and processed using a trained data recognition model, and corresponding classification results can be obtained. The historical search terms whose classification results meet the requirements of the training samples and the result page information corresponding to the historical search terms are used as the training samples. Since the number of historical search terms is extremely large, this can effectively reduce the labor cost of obtaining training samples, while increasing the number of training samples, thereby improving the recognition effect of the model.

[0202] Optionally, in the above Fig. 9 On the basis of the corresponding embodiment, in another embodiment of the model training device 20 provided in the embodiment of the present application,

[0203] The acquisition module 201 is used to acquire the historical training samples of the historical iterative recognition model, where the historical training samples include historical label results;

[0204] Calling the data recognition model to perform recognition processing on each of the historical training samples to obtain a second classification result;

[0205] When the second classification result is the same as the historical label result, the historical training sample is used as the historical label sample;

[0206] When the second classification result is different from the historical label result, the historical training sample is deleted.

[0207] In an embodiment of the present application, a model training device is provided. By using the above device, the trained data recognition model is used to perform recognition processing on historical training samples, which can effectively reduce noise samples and thus improve the recognition results of the model.

[0208] Optionally, in the above Fig. 9 On the basis of the corresponding embodiment, in another embodiment of the model training device 20 provided in the embodiment of the present application,

[0209] The acquisition module 201 is used to acquire pre-training data, fine-tuning training data and a large language model, wherein the pre-training data includes pre-training search terms and result page information corresponding to the pre-training search terms, the fine-tuning training data includes fine-tuning search terms and label results of the fine-tuning search terms, and the result page information corresponding to the pre-training search terms includes a title and an abstract;

[0210] The training module 203 is used to train the large language model using the pre-training data to obtain a pre-training model;

[0211] The pre-training data is trained using the fine-tuning training data to obtain the data recognition model.

[0212] In the embodiment of the present application, a model training device is provided. The above device is used to provide a small amount of manually annotated models to pre-train the large language model, thereby obtaining a general text prediction model; and then the text prediction model is fine-tuned using fine-tuning training data, so that the text prediction model generates the data recognition model, so that the knowledge of the large model can be transferred to the data recognition model, thereby improving the recognition effect of the data recognition model.

[0213] Optionally, in the above Fig. 9 On the basis of the corresponding embodiment, in another embodiment of the model training device 20 provided in the embodiment of the present application,

[0214] The training module 203 is used to encode the pre-training data to obtain a pre-training character string;

[0215] The large language model is called to train the pre-trained string according to the character prediction task to obtain the pre-trained model.

[0216] In an embodiment of the present application, a model training device is provided. By using the above device and pre-training data to pre-train the large language model, the generalization ability of the large model can be utilized to improve the applicability of the data recognition model.

[0217] Optionally, in the above Fig. 9 On the basis of the corresponding embodiment, in another embodiment of the model training device 20 provided in the embodiment of the present application, the training module 203 is used to call the pre-training data to perform classification processing on the fine-tuning training data to obtain a predicted label of the fine-tuning training data;

[0218] The pre-training data is trained using the predicted label and the label result of the fine-tuning training data to obtain the data recognition model.

[0219] In an embodiment of the present application, a model training device is provided. By using the above device and fine-tuning the pre-trained model using a small amount of data, the knowledge of the large model can be effectively transferred to the data recognition model, thereby improving the recognition effect of the data recognition model.

[0220] Optionally, in the above Fig. 9 On the basis of the corresponding embodiment, in another embodiment of the model training device 20 provided in the embodiment of the present application, the acquisition module 201 is used to obtain the search term to be processed and the result page information corresponding to the search term to be processed;

[0221] The processing module 202 is used to perform feature extraction on the search term to be processed and the result page information corresponding to the search term to be processed based on the feature extraction network of the recognition model to obtain a third feature representation and a fourth feature representation, wherein the third feature representation is used to indicate a keyword feature representation of the search term to be processed and the result page information corresponding to the search term to be processed, and the second feature representation is used to indicate a feature representation of the result page information corresponding to the search term to be processed;

[0222] Determine a third weight value of the third feature representation and a fourth weight value of the fourth feature representation based on a weight classification network of the recognition model;

[0223] The feature fusion network based on the recognition model performs weighted concatenation on the third feature representation and the fourth feature representation based on the third weight value and the fourth weight value to obtain a feature representation to be processed;

[0224] The classification network based on the recognition model performs classification processing on the feature representation to be processed to obtain a classification result of the search term to be processed.

[0225] In an embodiment of the present application, a model training device is provided. Using the above device, after the recognition model is trained, the recognition model can be applied to content review to ensure the user's online experience; at the same time, different weight values ​​are added to the search terms and the result page information corresponding to the search terms, so that different search terms have different attention to the result page, thereby improving the recognition effect.

[0226] The model training device provided in this application can be used for the server, see Fig.10 , Fig.10: is a schematic diagram of a server structure provided in an embodiment of the present application. The server 300 may have relatively large differences due to different configurations or performances, and may include one or more central processing units (CPU) 322 (for example, one or more processors) and memory 332, and one or more storage media 330 (for example, one or more mass storage devices) storing application programs 342 or data 344. Among them, the memory 332 and the storage medium 330 can be short-term storage or permanent storage. The program stored in the storage medium 330 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations on the server. Furthermore, the central processing unit 322 can be configured to communicate with the storage medium 330 to execute a series of instruction operations in the storage medium 330 on the server 300.

[0227] The server 300 may also include one or more power supplies 326, one or more wired or wireless network interfaces 350, one or more input and output interfaces 358, and / or one or more operating systems 341, such as Windows Server 2000. TM , Mac OS X TM , Unix TM ,Linux TM , FreeBSD TM etc.

[0228] The steps performed by the server in the above embodiment can be based on the Fig.10 The server structure shown.

[0229] The model training device provided in this application can be used in terminal devices. Fig.11 For the sake of convenience, only the parts related to the embodiments of the present application are shown. For specific technical details not disclosed, please refer to the method part of the embodiments of the present application. In the embodiments of the present application, the terminal device is a smart phone as an example for explanation:

[0230] Fig.11 The block diagram shows a partial structure of a smart phone related to the terminal device provided in the embodiment of the present application. Fig.11 The smart phone includes: a radio frequency (RF) circuit 410, a memory 420, an input unit 430, a display unit 440, a sensor 450, an audio circuit 460, a wireless fidelity (WiFi) module 470, a processor 480, and a power supply 490. Those skilled in the art will appreciate that Fig.11The structure of the smartphone shown in the figure does not constitute a limitation of the smartphone, and may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently.

[0231] Combine the following Fig.11 A detailed introduction to the various components of a smartphone:

[0232] The RF circuit 410 can be used for receiving and sending signals during the process of sending and receiving information or making calls. In particular, after receiving the downlink information of the base station, it is sent to the processor 480 for processing; in addition, the designed uplink data is sent to the base station. Generally, the RF circuit 410 includes but is not limited to an antenna, at least one amplifier, a transceiver, a coupler, a low noise amplifier (LNA), a duplexer, etc. In addition, the RF circuit 410 can also communicate with the network and other devices through wireless communication. The above-mentioned wireless communication can use any communication standard or protocol, including but not limited to the global system of mobile communication (GSM), general packet radio service (GPRS), code division multiple access (CDMA), wideband code division multiple access (WCDMA), long term evolution (LTE), email, short messaging service (SMS), etc.

[0233] The memory 420 can be used to store software programs and modules. The processor 480 executes various functional applications and data processing of the smartphone by running the software programs and modules stored in the memory 420. The memory 420 may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function (such as a sound playback function, an image playback function, etc.), etc.; the data storage area may store data created according to the use of the smartphone (such as audio data, a phone book, etc.), etc. In addition, the memory 420 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0234] The input unit 430 can be used to receive input digital or character information, and to generate key signal input related to the user settings and function control of the smart phone. Specifically, the input unit 430 may include a touch panel 431 and other input devices 432. The touch panel 431, also known as a touch screen, can collect the user's touch operation on or near it (such as the user's operation on the touch panel 431 or near the touch panel 431 using any suitable object or accessory such as a finger, stylus, etc.), and drive the corresponding connection device according to a pre-set program. Optionally, the touch panel 431 may include two parts: a touch detection device and a touch controller. Among them, the touch detection device detects the user's touch orientation, detects the signal brought by the touch operation, and transmits the signal to the touch controller; the touch controller receives the touch information from the touch detection device, converts it into contact coordinates, and then sends it to the processor 480, and can receive and execute commands sent by the processor 480. In addition, the touch panel 431 can be implemented in various types such as resistive, capacitive, infrared, and surface acoustic waves. In addition to the touch panel 431, the input unit 430 may also include other input devices 432. Specifically, other input devices 432 may include but are not limited to one or more of a physical keyboard, function keys (such as volume control keys, switch keys, etc.), a trackball, a mouse, a joystick, etc.

[0235] The display unit 440 can be used to display information input by the user or information provided to the user and various menus of the smartphone. The display unit 440 may include a display panel 441. Optionally, the display panel 441 may be configured in the form of a liquid crystal display (LCD), an organic light-emitting diode (OLED), or the like. Further, the touch panel 431 may cover the display panel 441. When the touch panel 431 detects a touch operation on or near it, it is transmitted to the processor 480 to determine the type of touch event. Subsequently, the processor 480 provides a corresponding visual output on the display panel 441 according to the type of touch event. Although in Fig.11 In the embodiment, the touch panel 431 and the display panel 441 are used as two independent components to implement the input and output functions of the smartphone, but in some embodiments, the touch panel 431 and the display panel 441 can be integrated to implement the input and output functions of the smartphone.

[0236] The smartphone may also include at least one sensor 450, such as a light sensor, a motion sensor, and other sensors. Specifically, the light sensor may include an ambient light sensor and a proximity sensor, wherein the ambient light sensor may adjust the brightness of the display panel 441 according to the brightness of the ambient light, and the proximity sensor may turn off the display panel 441 and / or the backlight when the smartphone is moved to the ear. As a type of motion sensor, the accelerometer sensor can detect the magnitude of acceleration in all directions (generally three axes), and can detect the magnitude and direction of gravity when stationary. It can be used for applications that identify the posture of the smartphone (such as horizontal and vertical screen switching, related games, magnetometer posture calibration), vibration recognition related functions (such as pedometer, tapping), etc.; as for other sensors that can be configured in smartphones, such as gyroscopes, barometers, hygrometers, thermometers, infrared sensors, etc., they will not be repeated here.

[0237] The audio circuit 460, the speaker 461, and the microphone 462 can provide an audio interface between the user and the smartphone. The audio circuit 460 can transmit the received audio data to the speaker 461 after converting the received audio data into an electrical signal, which is converted into a sound signal for output; on the other hand, the microphone 462 converts the collected sound signal into an electrical signal, which is received by the audio circuit 460 and converted into audio data, and then the audio data is output to the processor 480 for processing, and then sent to another smartphone through the RF circuit 410, or the audio data is output to the memory 420 for further processing.

[0238] WiFi is a short-range wireless transmission technology. Smartphones can help users send and receive emails, browse web pages, and access streaming media through WiFi module 470, which provides users with wireless broadband Internet access. Fig.11 A WiFi module 470 is shown, but it is understandable that it is not an essential component of the smartphone and can be omitted as needed without changing the essence of the invention.

[0239] Processor 480 is the control center of the smartphone, which uses various interfaces and lines to connect various parts of the entire smartphone, and executes various functions of the smartphone and processes data by running or executing software programs and / or modules stored in memory 420, and calling data stored in memory 420, thereby monitoring the smartphone as a whole. Optionally, processor 480 may include one or more processing units; optionally, processor 480 may integrate an application processor and a modem processor, wherein the application processor mainly processes the operating system, user interface, and application programs, and the modem processor mainly processes wireless communications. It is understandable that the above-mentioned modem processor may not be integrated into processor 480.

[0240] The smart phone also includes a power source 490 (such as a battery) for supplying power to various components. Optionally, the power source can be logically connected to the processor 480 through a power management system, thereby managing functions such as charging, discharging, and power consumption management through the power management system.

[0241] Although not shown, the smartphone may also include a camera, a Bluetooth module, etc., which will not be described in detail here.

[0242] The steps performed by the terminal device in the above embodiment can be based on the Fig.11 The terminal device structure shown.

[0243] A computer-readable storage medium is also provided in an embodiment of the present application. The computer-readable storage medium stores a computer program, which, when executed on a computer, enables the computer to execute the methods described in the aforementioned embodiments.

[0244] The present application also provides a computer program product including a program, which, when executed on a computer, enables the computer to execute the methods described in the aforementioned embodiments.

[0245] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0246] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0247] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0248] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0249] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0250] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A model training method, characterized in that: include: Acquire a training sample and an initial recognition model, wherein the training sample includes a search term, result page information corresponding to the search term, and a label result corresponding to the search term; Performing feature extraction on the training sample to obtain a first feature representation and a second feature representation, wherein the first feature representation is used to indicate a keyword feature representation of the search term and the result page information, and the second feature representation is used to indicate a feature representation of the result page information; Determining a first weight value for the first feature representation and a second weight value for the second feature representation; Performing weighted concatenation of the first feature representation and the second feature representation based on the first weight value and the second weight value to obtain a feature representation to be predicted; Performing prediction processing on the feature representation to be predicted to obtain a prediction result; The initial recognition model is trained based on the prediction result and the label result to obtain a recognition model.

2. The method according to claim 1, characterized in that Determining a first weight value of the first feature representation and a second weight value of the second feature representation comprises: Calling a prediction network to perform prediction processing on the first feature representation to obtain a second weight value of the second feature representation, wherein the second weight value is used to indicate the attention of different types of search terms to the result page information, and the second weight value is greater than 0 and less than 1; The first weight value is determined based on the second weight value, and the sum of the second weight value and the first weight value is 1.

3. The method according to claim 2, characterized in that Calling the prediction network to perform prediction processing on the first feature representation to obtain a second weight value of the second feature representation includes: Calling the first feature representation of a fully connected network to perform full connection processing to obtain a first prediction value, wherein the fully connected network serves as the prediction network; The first prediction value is mapped to obtain the second weight value.

4. The method according to any one of claims 1 to 3, characterized in that Extracting features from the training sample to obtain a first feature representation and a second feature representation includes: Performing keyword feature extraction on the search terms in the training sample and the result page information corresponding to the search terms to obtain the first feature representation; Feature extraction is performed on the result page information corresponding to the search term in the training sample to obtain the second feature representation.

5. The method according to claim 4, characterized in that Performing keyword feature extraction on the search words in the training sample and the result page information corresponding to the search words to obtain the first feature representation includes: Performing word segmentation processing on the top N result page information in the result page information corresponding to the search term to obtain a word segmentation set, where N is a positive integer; Performing data statistics on each segmentation in the segmentation set to obtain a frequency statistical value corresponding to each segmentation in the segmentation set; The top M word segments of the frequency statistics are used as the keyword set of the training sample, where M is a positive integer; Concatenate each keyword in the keyword set with the search term to obtain a keyword sequence; Feature extraction is performed on the keyword sequence to obtain the first feature representation.

6. The method according to claim 5, characterized in that Extracting features from the keyword sequence to obtain the first feature representation includes: Segmenting the keyword sequence to obtain a first segmented word sequence to be processed; Embedding the first word segmentation sequence to be processed to obtain a first word vector for each word in the first word segmentation sequence to be processed; The first word vector is encoded to obtain the first feature representation.

7. The method according to claim 4, characterized in that Performing feature extraction on the result page information corresponding to the search term in the training sample to obtain the second feature representation includes: Performing word segmentation processing on the result page information corresponding to the search term to obtain a second word segmentation sequence to be processed; Embedding the second to-be-processed word segmentation sequence to obtain a second word vector for each word in the second to-be-processed word segmentation sequence; Using convolutional neural networks with X different convolution kernels to perform feature extraction on the second word vector to obtain X feature representations, where X is a positive integer; Concatenate the X feature representations to obtain a feature representation to be processed; Feature extraction is performed on the feature representation to be processed through a fully connected layer to obtain the second feature representation.

8. The method according to any one of claims 1 to 3 and 5 to 7, characterized in that: Obtaining training samples includes: Calling a data recognition model to perform recognition processing on historical search terms to obtain the machine-labeled samples, wherein the data recognition model is a trained model; Calling the data recognition model to perform recognition processing on the training samples of the historical iteration recognition model to obtain the historical annotated samples; The machine-annotated samples and the historically-annotated samples are used as the training samples.

9. The method according to claim 8, characterized in that Calling the data recognition model to recognize and process the historical search terms to obtain the machine-annotated samples includes: Calling the data recognition model to perform recognition processing on the historical search words to obtain a first classification result of the historical search words; The machine-annotated samples are determined from the historical search terms according to the first classification result.

10. The method according to claim 8, characterized in that The calling of the data recognition model to perform recognition processing on the training samples of the historical iteration recognition model to obtain the historical annotation samples includes: Acquire historical training samples of the historical iterative recognition model, wherein the historical training samples include historical label results; Calling each of the historical training samples in the data recognition model to perform recognition processing to obtain a second classification result; When the second classification result is the same as the historical label result, using the historical training sample as the historical label sample; When the second classification result is different from the historical label result, the historical training sample is deleted.

11. The method according to claim 8, characterized in that The method further comprises: Acquire pre-training data, fine-tuning training data and a large language model, wherein the pre-training data includes pre-training search terms and result page information corresponding to the pre-training search terms, the fine-tuning training data includes fine-tuning search terms and label results of the fine-tuning search terms, and the result page information corresponding to the pre-training search terms includes a title and an abstract; Using the pre-training data to train the large language model to obtain a pre-training model; The pre-training data is trained using the fine-tuning training data to obtain the data recognition model.

12. The method according to claim 5, characterized in that The using the pre-training data to train the large language model to obtain a pre-training model includes: Encoding the pre-training data to obtain a pre-training character string; The large language model is called to train the pre-trained character string according to the character prediction task to obtain the pre-trained model.

13. The method according to claim 5, characterized in that Using the fine-tuning training data to train the pre-training data to obtain the data recognition model includes: Calling the pre-training data to perform classification processing on the fine-tuning training data to obtain a predicted label for the fine-tuning training data; The pre-training data is trained using the predicted labels and label results of the fine-tuning training data to obtain the data recognition model.

14. The method according to any one of claims 1 to 3, 5 to 7, 9 to 13, characterized in that The method further comprises: Obtaining a search term to be processed and result page information corresponding to the search term to be processed; The feature extraction network based on the recognition model performs feature extraction on the search term to be processed and the result page information corresponding to the search term to be processed to obtain a third feature representation and a fourth feature representation, wherein the third feature representation is used to indicate a keyword feature representation of the search term to be processed and the result page information corresponding to the search term to be processed, and the fourth feature representation is used to indicate a feature representation of the result page information corresponding to the search term to be processed; Determine a third weight value of the third feature representation and a fourth weight value of the fourth feature representation based on a weight classification network of the recognition model; The feature fusion network based on the recognition model performs weighted concatenation on the third feature representation and the fourth feature representation based on the third weight value and the fourth weight value to obtain a feature representation to be processed; The classification network based on the recognition model performs classification processing on the feature representation to be processed to obtain a classification result of the search term to be processed.

15. A model training device, characterized in that: include: An acquisition module, used to acquire training samples and an initial recognition model, wherein the training samples include search terms, result page information corresponding to the search terms, and label results corresponding to the search terms; A processing module, configured to perform feature extraction on the training sample to obtain a first feature representation and a second feature representation, wherein the first feature representation is used to indicate a keyword feature representation of the search term and the result page information, and the second feature representation is used to indicate a feature representation of the result page information; determine a first weight value of the first feature representation and a second weight value of the second feature representation; perform weighted splicing of the first feature representation and the second feature representation based on the first weight value and the second weight value to obtain a feature representation to be predicted; perform prediction processing on the feature representation to be predicted to obtain a prediction result; A training module is used to train the initial recognition model based on the prediction result and the label result to obtain a recognition model.

16. A computer device, characterized in that: include: Memory, processor, and bus system; Wherein, the memory is used to store programs; The processor is used to execute the program in the memory, and the processor is used to perform the method according to any one of claims 1 to 14 according to the instructions in the program code; The bus system is used to connect the memory and the processor so that the memory and the processor can communicate with each other.

17. A computer-readable storage medium comprising instructions, which, when executed on a computer, causes the computer to perform the method according to any one of claims 1 to 14.

18. A computer program product, comprising a computer program, characterized in that The computer program is used by a processor to execute the method according to any one of claims 1 to 14.

Images