Search term recommendation method, training method of recognition model, and related devices

Abstract

This application discloses a search term recommendation method, a training method for a recognition model, and related apparatus. The method involves acquiring historical viewing content and object feature data of the target account, obtained from the historical viewing content, determining multiple candidate keywords from the historical viewing content, determining search tendency parameters for each candidate keyword based on the object feature data (the search tendency parameters characterize the probability that the target account wants to search for the candidate keywords), identifying the candidate keywords corresponding to the search tendency parameters that meet the tendency conditions as target search terms, and recommending the target search terms to the target account. Thus, based on historical viewing content and object feature data, the keywords obtained for searching match the object's preferences, improving the accuracy of search term recommendation. Recommendations can be made without requiring the object to input partial keywords, better meeting the object's expectations for recommended keywords, reducing the difficulty of recommending search terms for specific search scenarios, and improving the user experience.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a search term recommendation method, a training method for a recognition model, and related apparatus. Background Technology

[0002] With the development of mobile internet, search functionality has become a primary entry point for users to obtain information. Recommendation features can be added to search functions to help users quickly complete their searches. For example, recommending search terms that users might use during the search process.

[0003] In related technologies, real-time recommendations can typically be provided when a user enters a portion of their search query. For example, after a user enters the keyword "product A," search terms like "how is product A?" might appear, thus helping the user achieve a quick search.

[0004] However, this method of supplementing search terms generally uses fixed suffixes, such as "how" or "what", which leads to lower accuracy in recommending search terms. Summary of the Invention

[0005] To address the aforementioned technical problems, this application provides a search term recommendation method, a training method for a recognition model, and related apparatus to improve the accuracy of search term recommendations.

[0006] The embodiments of this application disclose the following technical solutions:

[0007] On one hand, embodiments of this application provide a search term recommendation method, the method comprising:

[0008] Obtain the historical viewing content viewed through the target account and the object feature data corresponding to the target account;

[0009] Multiple candidate keywords were identified from the historical viewing content;

[0010] Based on the object feature data, a search tendency parameter is determined for each of the candidate keywords, wherein the search tendency parameter is used to characterize the probability that the target account wants to search for the candidate keywords;

[0011] Candidate keywords that meet the search tendency parameters are identified as target search terms;

[0012] Recommend the target search term to the target account.

[0013] On the other hand, embodiments of this application provide a method for training a recognition model, the method comprising:

[0014] Obtain a first sample pair and a second sample pair. The first sample pair includes the sample content where the sample keyword is located and the sample importance parameter of the sample keyword. The second sample pair includes the feature data of the first sample object, the sample understanding difficulty parameter of the sample keyword, and the sample relevance parameter of the sample keyword. The sample keyword is the keyword marked by the first sample object during the process of viewing the sample content, as represented by the feature data of the first sample object.

[0015] A target recognition model is obtained by performing a first-stage training based on the first sample pair and a second-stage training based on the second sample pair.

[0016] The first phase of training is as follows:

[0017] The first initial recognition model is used to extract keywords from the sample content included in the first sample pair to obtain predicted keywords and the predicted important parameters of the predicted keywords.

[0018] Based on the differences between the predicted keywords and the sample keywords included in the first sample pair, and the differences between the predicted importance parameters and the sample importance parameters included in the first sample pair, the model parameters of the first initial recognition model are adjusted to obtain the second initial recognition model.

[0019] The second phase of training is as follows:

[0020] Based on the second sample pair, a correlation analysis is performed on the feature data of the first sample object and the sample keywords using a third initial recognition model to determine the prediction comprehension difficulty parameter and the prediction correlation parameter of the sample keywords;

[0021] Based on the difference between the predicted understanding difficulty parameter of the sample keywords and the sample understanding difficulty parameter, and the difference between the predicted relevance parameter of the sample keywords and the sample relevance parameter, the model parameters of the third initial recognition model are adjusted to obtain the fourth initial recognition model.

[0022] If the first stage of training occurs before the second stage of training, then the third initial recognition model is the second initial recognition model, and the fourth initial recognition model is the target recognition model; if the second stage of training occurs before the first stage of training, then the first initial recognition model is the fourth initial recognition model, and the second initial recognition model is the target recognition model.

[0023] On the other hand, embodiments of this application provide a search term recommendation method, the method comprising:

[0024] After logging into the content application through the target account, an input control and a target search term are displayed. The input control is used to search for the input content, and the target search term is determined based on the target account's historical viewing content and the target account's object feature data.

[0025] If a selection operation is performed for the target search term, the content obtained from the search based on the target search term will be displayed.

[0026] On the other hand, embodiments of this application provide a search term recommendation device, the device comprising: an acquisition unit, a determination unit, a calculation unit, a filtering unit, and a recommendation unit;

[0027] The acquisition unit is used to acquire the historical viewing content viewed through the target account and the object feature data corresponding to the target account;

[0028] The determining unit is used to determine multiple candidate keywords from the historical viewing content;

[0029] The calculation unit is used to determine the search tendency parameters of each candidate keyword based on the object feature data. The search tendency parameters are used to characterize the probability that the target account wants to search for the candidate keywords.

[0030] The filtering unit is used to determine the candidate keywords corresponding to the search tendency parameters that meet the tendency conditions as the target search terms;

[0031] The recommendation unit is used to recommend the target search term to the target account.

[0032] On the other hand, embodiments of this application provide a training apparatus for a recognition model, the apparatus comprising: an acquisition unit and a training unit;

[0033] The acquisition unit is used to acquire a first sample pair and a second sample pair. The first sample pair includes the sample content where the sample keyword is located and the sample importance parameter of the sample keyword. The second sample pair includes the feature data of a first sample object, the sample understanding difficulty parameter of the sample keyword, and the sample relevance parameter of the sample keyword. The sample keyword is the keyword marked by the first sample object during the process of viewing the sample content, as represented by the feature data of the first sample object.

[0034] The training unit is used to perform a first-stage training based on the first sample pair and a second-stage training based on the second sample pair to obtain a target recognition model.

[0035] The first phase of training is as follows:

[0036] The first initial recognition model is used to extract keywords from the sample content included in the first sample pair to obtain predicted keywords and the predicted important parameters of the predicted keywords.

[0037] Based on the differences between the predicted keywords and the sample keywords included in the first sample pair, and the differences between the predicted importance parameters and the sample importance parameters included in the first sample pair, the model parameters of the first initial recognition model are adjusted to obtain the second initial recognition model.

[0038] The second phase of training is as follows:

[0039] Based on the second sample pair, a correlation analysis is performed on the feature data of the first sample object and the sample keywords using a third initial recognition model to determine the prediction comprehension difficulty parameter and the prediction correlation parameter of the sample keywords;

[0040] Based on the difference between the predicted understanding difficulty parameter of the sample keywords and the sample understanding difficulty parameter, and the difference between the predicted relevance parameter of the sample keywords and the sample relevance parameter, the model parameters of the third initial recognition model are adjusted to obtain the fourth initial recognition model.

[0041] If the first stage of training occurs before the second stage of training, then the third initial recognition model is the second initial recognition model, and the fourth initial recognition model is the target recognition model; if the second stage of training occurs before the first stage of training, then the first initial recognition model is the fourth initial recognition model, and the second initial recognition model is the target recognition model.

[0042] On the other hand, embodiments of this application provide a search term recommendation device, the device comprising: a first display unit and a second display unit;

[0043] The first display unit is used to display an input control and a target search term after logging into the content application through the target account. The input control is used to search for the input content, and the target search term is determined based on the target account's historical viewing content and the target account's object feature data.

[0044] The second display unit is used to display the content obtained by searching based on the target search term if a selection operation for the target search term is obtained.

[0045] On the other hand, embodiments of this application provide a computer device, the computer device including a processor and a memory:

[0046] The memory is used to store computer programs and to transfer the computer programs to the processor;

[0047] The processor is configured to execute the methods described above according to instructions in the computer program.

[0048] On the other hand, embodiments of this application provide a computer-readable storage medium for storing a computer program for performing the methods described above.

[0049] On the other hand, embodiments of this application provide a computer program product including a computer program that, when run on a computer device, causes the computer device to perform the methods described above.

[0050] As can be seen from the above technical solution, the system acquires historical content viewed by the target account and object feature data describing the characteristics of the target account. The target account is highly likely to be interested in the content it viewed, thus historical viewing content reflects the target account's interests over a period of time. Multiple candidate keywords can then be identified from this historical viewing content, and the target account is highly likely to conduct subsequent searches based on these candidate keywords. To further improve the accuracy of candidate keyword recommendations, each candidate keyword can be further filtered based on object feature data that describes the characteristics of the object. Specifically, based on the target account's object feature data, search tendency parameters for each candidate keyword are determined. These search tendency parameters characterize the probability that the target account wants to search based on the candidate keywords, or in other words, they reflect the target account's level of interest in the candidate keywords. Candidate keywords corresponding to search tendency parameters that meet the tendency criteria are identified as target search terms. Since the target account is highly likely to search based on these target search terms, these target search terms are recommended to the target account, improving search efficiency. Therefore, by using historical viewing content and object feature data, search terms that match the target's preferences are obtained, improving the accuracy of search term recommendations. Moreover, it can make recommendations without requiring users to input partial search terms, which better meets users' expectations for recommended search terms, reduces the difficulty of recommending search terms in the search function, and improves the user experience. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 A schematic diagram of a computer system for a search term recommendation method provided in an embodiment of this application;

[0053] Figure 2 This is a schematic diagram illustrating an application scenario of a search term recommendation method provided in an embodiment of this application.

[0054] Figure 3 A flowchart illustrating a search term recommendation method provided in an embodiment of this application;

[0055] Figure 4 A flowchart illustrating another search term recommendation method provided in an embodiment of this application;

[0056] Figure 5 A schematic diagram illustrating how to display an entry point for recommended content in a chat software interface, as provided in this application embodiment;

[0057] Figure 6 A schematic diagram of a recommended content page provided in an embodiment of this application;

[0058] Figure 7 This is a schematic diagram illustrating how to view target recommended content, as provided in an embodiment of this application.

[0059] Figure 8 A schematic diagram illustrating a search function provided in an embodiment of this application;

[0060] Figure 9 A schematic diagram illustrating a recommendation function provided in an embodiment of this application;

[0061] Figure 10 This is a schematic diagram corresponding to a search function provided in an embodiment of this application;

[0062] Figure 11 This is a schematic diagram of the structure of a search term recommendation device provided in an embodiment of this application;

[0063] Figure 12 This is a schematic diagram of the structure of a search term recommendation device provided in an embodiment of this application;

[0064] Figure 13 This application provides a schematic diagram of the structure of a server according to an embodiment of the present application.

[0065] Figure 14 This is a schematic diagram of the structure of a terminal device provided in an embodiment of this application;

[0066] Figure 15 This is a schematic diagram illustrating a redirection from a search page to a recommendation page, provided as an embodiment of this application.

[0067] Figure 16A schematic diagram of a search control provided in an embodiment of this application;

[0068] Figure 17 This is a schematic diagram of the structure of a training device for a recognition model provided in an embodiment of this application. Detailed Implementation

[0069] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0070] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0071] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0072] It should be understood that although the terms first, second, etc., may be used in this application to describe various information, this information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, a first parameter may also be referred to as a second parameter, and similarly, a second parameter may also be referred to as a first parameter. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0073] It should be noted that this application may display prompt interfaces, pop-ups, or output voice prompts before and during the collection of user-related data (e.g., historical viewing content, object characteristic data, etc.). These prompt interfaces, pop-ups, or voice prompts are used to inform the user that their relevant data is being collected. This ensures that the application only begins the steps related to collecting user-related data after receiving confirmation from the user regarding the prompt interface or pop-up; otherwise (i.e., without receiving confirmation from the user), the steps to collect user-related data end, meaning no user-related data is collected. In other words, all user data collected in this application is collected with the user's consent and authorization, and the collection, use, and processing of relevant user data must comply with the relevant laws, regulations, and standards of the relevant countries and regions.

[0074] In related technologies, recommendations are made based on search results by adding suffixes to keywords after they have been entered. However, this method requires users to input specific keywords before recommendations can be made, making the recommendation criteria quite stringent and resulting in a poor user experience. Furthermore, if the keywords entered are incorrect, incorrect recommendations may occur, leading to low accuracy in search term recommendations.

[0075] Based on this, embodiments of this application provide a search term recommendation method that, based on historical viewed content and object feature data, obtains search terms that match the object's preferences, thereby improving the accuracy of search term recommendations. Furthermore, recommendations can be made without requiring the object to input partial search terms, better meeting the object's expectations for recommended search terms, reducing the difficulty of search term recommendations in search functions, and improving the user experience.

[0076] To facilitate understanding of the technical solutions of this application, some technical terms involved in the embodiments of this application will be briefly introduced below.

[0077] (1) Large Language Model (LLM) is a natural language processing model based on deep learning. It has a large number of model parameters, a lot of training data, and features rich knowledge and strong text semantic capabilities.

[0078] (2) Full-parameter fine-tuning is the process of using new domain-specific data to train a general large language model and updating all the parameters of the model to make it more suitable for specific downstream tasks.

[0079] (3) Recommender System: A content distribution system. It proactively predicts user interests and preferences by analyzing historical user behavior data and pushes potentially interesting content to users. Recommender systems typically recommend content through recall and ranking. In the recall phase, the system quickly filters a small subset of content relevant to the user from a large set of candidate content; this step requires efficient processing of massive amounts of data. In the ranking phase, the model performs fine-grained ranking of the recalled content, selecting the final content to recommend to the user. This step usually uses relatively complex deep learning models, multi-objective optimization, and personalized ranking techniques to ensure high relevance and high quality of the recommendation results.

[0080] (4) Search System: An information retrieval system. Based on the user's search query, it quickly finds relevant content from a massive content library and returns it to the user. It also involves recall and ranking processes. Compared to recommendation, the user's needs are explicitly given.

[0081] (5) Deep Structured Semantic Models (DSSM) is a deep learning model with a dual-tower structure, including a user tower and an item tower. It is used to learn vector representations of object features and item features, and optimizes the model by maximizing the similarity of related samples. It is widely used in the recall process.

[0082] (6) Loss function: The objective function for training the neural network model. It is minimized by gradient descent to obtain the gradient of the model parameters during the training iteration.

[0083] (7) Vector (Representation): A method of representing objects such as text, users or items to a continuous numerical vector space. It is sometimes also called Embedding, embedding representation or representation vector.

[0084] (8) Multi-Layer Perceptron (MLP) is a type of feedforward neural network that contains multiple fully connected layers and activation functions.

[0085] To facilitate understanding of the search term recommendation method provided in the embodiments of this application, the computer system for the search term recommendation method will be described below.

[0086] See Figure 1The figure is a schematic diagram of a computer system for a search term recommendation method provided in an embodiment of this application. The computer system 100 includes multiple devices (such as a first device 110, a second device 120, a third device 130, a fourth device 140, etc.) that can communicate through a communication network. The devices can be terminal devices or servers.

[0087] The communication network uses standard communication technologies and / or protocols, typically the Internet, but can also be any network, including but not limited to Bluetooth, local area network (LAN), metropolitan area network (MAN), wide area network (WAN), mobile, private network, or any combination of virtual private network. In some embodiments, custom or dedicated data communication technologies may be used to replace or supplement the aforementioned data communication technologies.

[0088] The terminal device can be an electronic device such as a smartphone, wearable device, personal computer (PC), intelligent voice interaction device, smart home appliance, vehicle terminal, aircraft, or unmanned vending terminal. The terminal device can install and run a client application for a content application. This content application can be an application that supports content recommendation (such as search term recommendation), or other applications that support content display, content creation, and content storage; this application does not limit its scope. Furthermore, this application does not limit the form of the content application, including but not limited to applications (Apps), mini-programs, etc., installed on the terminal device, and can also be in the form of a webpage.

[0089] A server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services such as cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDNs), and basic cloud computing services such as big data. A server can serve as a backend server for the aforementioned content applications, providing backend services to the clients of these applications.

[0090] To facilitate understanding of the search term recommendation method provided in this application embodiment, the following example uses a server as the execution subject of the search term recommendation method to illustrate its application scenarios.

[0091] See Figure 2 This figure is a schematic diagram illustrating an application scenario of a search term recommendation method provided in an embodiment of this application. Figure 2In this application scenario, the following are included: Figure 1 There is a terminal device 110 and a server 120.

[0092] Terminal device 110 has a content application installed. This content application has both search and recommendation functions. The following explanation, using the target audience using terminal device 110 as an example, will illustrate both functions. Figure 2 The left side of the text describes the search function. Figure 2 The right side of the image describes the recommendation feature.

[0093] It should be noted that content applications can include multiple platforms with different functions, such as social platforms and content recommendation platforms. Each platform has its corresponding functions; for example, social platforms have communication functions, while content recommendation platforms have recommendation functions.

[0094] As one possible implementation, the search function in this application embodiment can be applied to multiple platforms in the content application, and the recommendation function can be applied to one platform in the content application. That is, the platform with the recommendation function also has the search function, and the search function is also applicable to other platforms.

[0095] As one possible implementation, the search function and recommendation function in the embodiments of this application can be applied to the same platform, that is, the platform with the recommendation function also has the search function, and the search function is no longer applicable to other platforms.

[0096] For example, a content recommendation platform has search and recommendation functions, but the search function is only used by the content recommendation platform. Figure 5 The search control 502 shown is used for searching this content across all platforms, i.e., cross-platform search functionality. For example... Figure 6 The search control 602 shown is only used for searching within this content recommendation platform.

[0097] (1) Search function.

[0098] The search function is generally provided by the aforementioned search system. The target user logs into the content application on terminal device 110 using their target account. After obtaining authorization from the target user, the content application on terminal device 110 collects the target user's object characteristic data, as well as the target user's historical viewing content during their use of the content application. Terminal device 110 then sends the historical viewing content and object characteristic data to the server.

[0099] Server 120 is used to provide corresponding services for content applications, such as recommendation services. After obtaining the target object's historical viewing content through the target account and object feature data describing the target object's characteristics, the historical viewing content reflects the target object's interests over a period of time, since the target object is highly likely to be interested in the content it viewed. Therefore, Server 120 can identify multiple candidate keywords from the historical viewing content, and the target object is highly likely to conduct subsequent searches based on these candidate keywords.

[0100] To further improve the accuracy of candidate keyword recommendations, each candidate keyword can be further filtered based on object feature data that describes the characteristics of the target object. Specifically, server 120 determines the search tendency parameters of each candidate keyword based on the object feature data of the target object. The search tendency parameters can characterize the probability that the target object wants to search based on the candidate keywords, or in other words, the search tendency parameters can reflect the target object's degree of interest in the candidate keywords. The candidate keywords corresponding to the search tendency parameters that meet the tendency conditions are determined as target search terms. Server 120 sends the target search terms to terminal device 110.

[0101] The target object is more likely to search based on the target search term, so the terminal device 110 displays the target search term to recommend the target search term to the target object so that the target object can search based on the target search term and improve the target object's search efficiency.

[0102] Therefore, by using historical viewing content and object characteristic data, the keywords obtained for searching match the object's preferences, improving the accuracy of search term recommendations. Furthermore, recommendations can be made without requiring the object to input any keywords, better meeting the object's expectations for recommended keywords, reducing the difficulty of recommending search terms for specific search scenarios, and improving the user experience.

[0103] (2) Recommendation function.

[0104] The recommendation function is generally provided by the aforementioned recommendation system. The target user logs into the content application on terminal device 110 using their target account. After obtaining authorization from the target user, the content application on terminal device 110 collects a historical search sequence consisting of the search content the target user has already searched while using the content application. Terminal device 110 then sends this historical search sequence to server 120.

[0105] Server 120 retrieves the target account's historical search sequence. The target audience is highly likely to be interested in their past search content; therefore, the historical search sequence reflects the target audience's interests over a period of time. Furthermore, the multiple historical search results within the sequence are sorted chronologically, demonstrating how the target audience's interests change over time. Based on this historical search sequence, a virtual interest vector can be generated. This virtual interest vector represents the content the target audience has recently been interested in, making it more likely that the target audience will subsequently view recommended content related to this virtual interest vector. This recommended content can be news feed content, such as articles or videos.

[0106] Server 120 retrieves recommendation content vectors corresponding to multiple pieces of content awaiting recommendation and calculates the similarity between the virtual interest vector and each recommendation content vector. Recommendation content corresponding to vectors with similarity scores meeting certain criteria is identified as supplementary recommendation content; that is, supplementary recommendation content has a high similarity to the virtual interest vector, and the target audience is more likely to view it. Therefore, supplementary recommendation content is designated as target recommendation content for the target account to improve the accuracy of the overall recommendation.

[0107] Therefore, by transforming historical search sequences into virtual interest vectors, recommended content can respond promptly to the search needs of the target audience, thereby improving the accuracy of content recommendations.

[0108] The search term recommendation method provided in this application embodiment can be executed by a server. However, in other embodiments of this application, the terminal device may also have similar functions to the server to execute the search term recommendation method provided in this application embodiment, or the terminal device and the server may jointly execute the search term recommendation method provided in this application embodiment. This embodiment does not limit this.

[0109] The following describes in detail a search term recommendation method provided in this application through method embodiments. Figure 3 This section explains the search term recommendation method used for the search function. Figure 4 This section explains the search term recommendation method used in the recommendation function.

[0110] See Figure 3 This figure is a schematic flowchart of a search term recommendation method provided in an embodiment of this application. For ease of description, the following embodiment uses a server as the execution subject of this search term recommendation method. Figure 3 As shown, the search term recommendation method includes S301-S305.

[0111] S301: Obtain the historical content viewed through the target account and the object feature data corresponding to the target account.

[0112] The target account can be the account used by the target user to log in to the content application, and the target account can represent the target user's use of the content application. The target user is the individual waiting for recommended keywords or content, including but not limited to natural persons, humanoid robots, and virtual humans (such as characters generated through augmented reality or virtual reality technology). In practical applications, the target user can view some content through a terminal device, which then sends it to the server.

[0113] Viewed history refers to one or more pieces of content viewed by a target user within a historical time period. Taking video applications as an example, if a target user logs into a video application with a target account, all videos viewed within that application using that target account can be considered viewed history. For instance, when a target account logs into a content application, all content viewed by that target account within a period prior to the login time will be considered viewed history.

[0114] Object characteristic data is data used to describe the characteristics of a target account, that is, data describing the characteristics of the target object using the target account. For example, object characteristic data can describe the target object's age, educational background, professional field, interest categories, etc., and this application does not make specific limitations in this regard.

[0115] S302: Identify multiple candidate keywords from historical content.

[0116] The content viewed by the target audience is likely to be of interest, thus the viewed history reflects the target account's interests over a past period. This allows for the identification of multiple candidate keywords from the viewed history. Candidate keywords are those keywords from the viewed history that could trigger the target audience's search intent, such as unfamiliar concepts or viewpoints. Compared to the viewed history itself, candidate keywords are more likely to be used in the search process, meaning the target account is more likely to conduct subsequent searches based on candidate keywords extracted from the viewed history.

[0117] This application does not specifically limit the method of extracting candidate keywords. For example, candidate keywords can be extracted using methods such as keyword extraction based on statistical features (term frequency–-inverse document frequency, TF-IDF), keyword extraction based on word graph models (TextRank), and keyword extraction based on topic models (Linear Discriminant Analysis, LDA). The following explanation will use the example of determining multiple candidate keywords from historically viewed content using a first identification model; see A1-A3 for details, which will not be elaborated upon here.

[0118] S303: Determine the search tendency parameters for each candidate keyword based on the object feature data.

[0119] While candidate keywords can stimulate the search desire of target audiences, different target audiences have different characteristics such as age, educational background, professional field, and interest categories, resulting in varying search desires for different keywords. In other words, the search desire of a target audience for a candidate keyword is related to their target characteristic data. Therefore, by analyzing the target account's corresponding target characteristic data, the search tendency parameters for each candidate keyword can be determined.

[0120] The search tendency parameter characterizes the probability that a target object wants to search for candidate keywords through a target account. For example, the larger the search tendency parameter, the higher the probability that the target object wants to search for that candidate keyword through a target account. This application does not specifically limit the method for determining the search tendency parameter; examples B1-B2 will be used later for illustration, and will not be repeated here.

[0121] S304: Identify the candidate keywords corresponding to the search tendency parameters that meet the tendency conditions as the target search terms.

[0122] The preference criteria are used to filter candidate keywords from multiple candidate keywords to identify those with a higher search intent for the target account. This application does not impose specific limitations on the preference criteria; those skilled in the art can set them according to actual needs. For example, multiple candidate keywords can be sorted based on search preference parameters, and the top M candidate keywords with the highest probability of the target account wanting to search for that candidate keyword can be selected as target search terms, where M is a positive integer. Alternatively, N keywords with a probability greater than a first probability threshold can be selected as target search terms, where N is a positive integer.

[0123] The target search term is a candidate keyword with a high probability of being searched by the target account from multiple candidate keywords. It can be obtained through preference conditions and search preference parameters. This application does not specifically limit the number of target search terms. There can be one or more target search terms. That is, one or more candidate keywords can be selected as target search terms from multiple candidate keywords. In order to increase the number of target search terms, the probability of hitting the search intent of the target account can be increased, thereby improving the accuracy of the recommendation.

[0124] S305: Recommend target search terms to the target account.

[0125] Among multiple candidate keywords for the target search term, the candidate keywords with the highest search probability for the target account are recommended to the target account so that the target audience can search based on the target search term, thereby improving the target audience's search efficiency.

[0126] This application does not specifically limit the method of recommending target search terms. For example, the server may directly display the target search terms to the target account. Alternatively, the server may send the target search terms to the terminal device, and the target search terms may be displayed through a content application on the terminal device. This application does not specifically limit the method of displaying keywords. For example, the target search terms may be directly displayed in the input control used for searching, such as a display box, so that the target account can directly search based on the target search terms. Alternatively, the target search terms may be displayed around the input control used for searching, so that the target account can use the target search terms to search. Furthermore, the target search terms can be converted into standardized pending search content; two methods are described below as examples.

[0127] Method 1 involves adding pre-defined collocations to the target search term, such as "what is it?" or "applications," which transforms the target search term into standardized pending search content, such as "[target search term] what is it?" or "[target search term] application."

[0128] Method two involves identifying the type of target search terms, such as product names or technical terms. Different types of keywords correspond to different supplementary terms. Based on the type of the target search term, supplementary terms are determined—vocabulary used to complement the target search term. Based on the target search term and its supplementary terms, a preliminary search content including the target search term is obtained. This preliminary search content is then recommended to the target account, thus achieving the goal of recommending the target search term.

[0129] Therefore, compared to directly recommending target search terms, the pending search content is more understandable. This makes it easier for the target account to use the pending search content, and they don't need to supplement the pending search content, further improving the search efficiency. Moreover, by pre-establishing a mapping relationship between different types of keywords and supplementary terms, suitable supplementary terms can be added for different types of keywords, thereby improving the accuracy of the pending search content and enhancing the user experience.

[0130] As can be seen from the above technical solution, the system acquires historical content viewed by the target account and object feature data describing the characteristics of the target account. The target account is highly likely to be interested in the content it viewed, thus historical viewing content reflects the target account's interests over a period of time. Multiple candidate keywords can then be identified from this historical viewing content, and the target account is highly likely to conduct subsequent searches based on these candidate keywords. To further improve the accuracy of candidate keyword recommendations, each candidate keyword can be further filtered based on object feature data that describes the characteristics of the object. Specifically, based on the target account's object feature data, search tendency parameters for each candidate keyword are determined. These search tendency parameters characterize the probability that the target account wants to search based on the candidate keywords, or in other words, they reflect the target account's level of interest in the candidate keywords. Candidate keywords corresponding to search tendency parameters that meet the tendency criteria are identified as target search terms. Since the target account is highly likely to search based on these target search terms, these target search terms are recommended to the target account, improving search efficiency. Therefore, by using historical viewing content and object feature data, search terms that match the target's preferences are obtained, improving the accuracy of search term recommendations. Moreover, it can make recommendations without requiring users to input partial search terms, which better meets users' expectations for recommended search terms, reduces the difficulty of recommending search terms in the search function, and improves the user experience.

[0131] As one possible implementation, this application provides a specific implementation of S302, namely, a specific implementation of determining multiple candidate keywords from historical viewing content, see A1-A3 for details.

[0132] A1: The first recognition model identifies the historical viewing content and obtains multiple candidate keywords to be determined in the historical viewing content.

[0133] The first recognition model is a recognition model used to identify desired content (such as candidate keywords) from input content (such as historical viewing content). Based on the first recognition model, historical viewing content is identified to obtain multiple candidate keywords. Candidate keywords are keywords that are likely to become candidate keywords.

[0134] Historical viewing content can include images, videos, music, and multimodal combinations. Therefore, a first recognition model can be determined based on the modality of the historical viewing content. For example, the first recognition model can be a large language model, which is not specifically limited in this embodiment. If the first recognition model is a large language model, the prompt information used to indicate the first recognition model is not specifically limited in this embodiment. One example is described below. Specifically, a first prompt template is pre-set, which can be as follows:

[0135] "As a content understanding assistant, please analyze the following specific content to identify potential candidate keywords that the target audience might want to learn more about or search for. These potential candidate keywords are typically:"

[0136] 1. Technical terms or field concepts;

[0137] 2. Product, technology, or platform name;

[0138] 3. Viewpoints that require background knowledge to understand.

[0139] Please provide the following for each candidate keyword:

[0140] -Important parameters (1-10 points, the higher the score, the more worthwhile the target audience is to learn more about);

[0141] - Brief explanation (to help determine if the user needs to conduct further search).

[0142] Details: [View History]

[0143] Output format: Candidate keywords to be determined | Important parameters | Explanation.

[0144] Therefore, the first prompt word template explicitly identifies the candidate keywords so that the first recognition model can accurately identify them. Then, historical viewing content can be filled into the first prompt word template to obtain the first prompt word information. Based on this first prompt word information, the first recognition model is guided to understand the historical viewing content, thereby obtaining multiple candidate keywords from the historical viewing content.

[0145] As one possible implementation, to improve computational efficiency, the historical viewing content, which can be quite extensive, can be divided into multiple sub-historical viewing content. Each sub-historical viewing content can then be used to populate the first prompt word template, thereby obtaining the candidate keywords included in each sub-historical viewing content. Taking an article as an example, the article D can be divided into n paragraphs, each paragraph being a sub-historical viewing content. Multiple paragraphs can then form a sequence, represented as D = {d1, d2, ..., dn}, where n is a positive integer.

[0146] A2: The important parameters corresponding to each candidate keyword are determined by the first identification model.

[0147] Key parameters describe the probability that a target account will search based on the candidate keywords, for example, a score of 1-10, with higher scores indicating greater interest from the target account. The first identification model can not only identify candidate keywords from historical content views but also score each candidate keyword to obtain key parameters, thus filtering out keywords with a lower probability of being searched by the target account.

[0148] The embodiments of this application do not specifically limit the order of A1 and A2. They can be executed sequentially or simultaneously. For example, multiple candidate keywords to be determined and important parameters corresponding to each candidate keyword can be output based on the first recognition model.

[0149] A3: Identify the candidate keywords corresponding to the important parameters that meet the important conditions as candidate keywords.

[0150] The importance criteria are used to filter candidate keywords from multiple undetermined candidate keywords to identify those with a high probability of being searched by the target account. This application does not specifically limit the importance criteria; those skilled in the art can set them according to actual needs. For example, multiple undetermined candidate keywords can be sorted based on importance parameters, and the top P undetermined candidate keywords with the highest probability of the target account wanting to search for that undetermined candidate keyword can be selected as candidate keywords, where P is a positive integer. Alternatively, Q undetermined candidate keywords with a probability greater than a second probability threshold can be selected as candidate keywords, where Q is a positive integer.

[0151] Therefore, by segmenting the historical viewed content using the first recognition model, multiple candidate keywords are obtained, along with important parameters corresponding to each candidate keyword. Based on these important parameters, multiple candidate keywords with a high probability of being searched by the target account are identified and confirmed as candidate keywords. Thus, the first recognition model effectively filters candidate keywords, removing some keywords with low search probability for the target account, thereby improving both the accuracy and efficiency of subsequent search term recommendations.

[0152] As one possible implementation, this application provides a specific implementation of S303, namely, a specific implementation of determining the search tendency parameters of each candidate keyword based on object feature data, see B1-B2 for details.

[0153] B1: By using the second recognition model to perform correlation analysis on the object feature data and multiple candidate keywords, the understanding difficulty parameter and the relevant parameter corresponding to each candidate keyword are determined.

[0154] The difficulty parameter describes how difficult it is for a target account to understand candidate keywords. The greater the difficulty for the target account to understand a candidate keyword, the greater the probability that the target account will search for that candidate keyword. In other words, the difficulty parameter reflects the probability that the target account wants to search for the corresponding candidate keyword.

[0155] The relevance parameter describes the degree of relevance between candidate keywords and the target account. The greater the relevance between the target account and the candidate keyword, the greater the probability that the target account will search for that candidate keyword. In other words, the relevance parameter reflects the probability that the target account wants to search for the corresponding candidate keyword.

[0156] The second recognition model is a recognition model. As one possible implementation, the second recognition model can be the same as the first recognition model, such as both being the same large language model. This allows the same keywords to be scored based on the same evaluation criteria, resulting in more accurate important parameters, comprehension difficulty parameters, and relevance parameters, thus improving the accuracy of candidate keyword identification.

[0157] Based on the second recognition model, correlation analysis is performed on object feature data and multiple candidate keywords to obtain the comprehension difficulty parameter and the correlation parameter corresponding to each candidate keyword.

[0158] If the second recognition model is a large language model, the embodiments of this application are not specifically limited to the prompt word information used to indicate the second recognition model. The following is an example illustrating one approach. Specifically, a second prompt word template is pre-set, which can be as follows:

[0159] "There is an object using the target account. The basic information is: [Object Feature Data]."

[0160] Please determine the difficulty and related parameters of the following candidate keywords for this target account:

[0161] [Multiple candidate keywords]

[0162] Score each candidate keyword (1-10 points):

[0163] -Understanding the difficulty parameters (1 point is the easiest, 10 points is the most difficult).

[0164] -Relevance parameters (1 point is the least relevant, 10 points is the most relevant)

[0165] B2: Based on the comprehension difficulty parameter and the relevant parameters corresponding to each candidate keyword, determine the search tendency parameter corresponding to each candidate keyword.

[0166] The search tendency parameter is used to characterize the probability that the target account wants to search for candidate keywords. Moreover, the comprehension difficulty parameter and the relevance parameter can both reflect the probability that the target account wants to search for the corresponding candidate keywords. Therefore, the search tendency parameter can be determined based on the comprehension difficulty parameter and the relevance parameter corresponding to each candidate keyword.

[0167] Therefore, the second identification model can be used to determine the comprehension difficulty parameter and relevance parameter corresponding to each candidate keyword, and then the search tendency parameter corresponding to each candidate keyword can be determined based on the comprehension difficulty parameter and relevance parameter, thereby improving the accuracy of the search tendency parameter of the candidate keyword and thus improving the accuracy of subsequent search term recommendations.

[0168] As one possible implementation, steps A1-A3 can be used in combination with steps B1-B2, or they can be used separately. This application does not make any specific limitations on this.

[0169] As one possible implementation, this application provides a specific implementation of B2, which is a specific implementation of determining the search tendency parameters corresponding to each candidate keyword based on the comprehension difficulty parameter and the relevant parameters corresponding to each candidate keyword, as detailed in C1-C2.

[0170] C1: Determine the key parameters corresponding to each candidate keyword.

[0171] The key parameter describes the probability that the target account will search based on the candidate keywords.

[0172] As one possible implementation, it can be based on the aforementioned A2, that is, by determining the important parameters corresponding to each candidate keyword through the first identification model, to obtain the important parameters of each candidate keyword. Each candidate keyword is extracted from multiple candidate keywords, so the important parameters of the candidate keywords can also be used as the important parameters of the corresponding candidate keywords.

[0173] C2: Determine the search tendency parameters for each candidate keyword based on the importance parameters, comprehension difficulty parameters, and relevance parameters corresponding to each candidate keyword.

[0174] Similarly, the search tendency parameter is used to characterize the probability that the target account wants to search for candidate keywords. Moreover, the importance parameter, the comprehension difficulty parameter, and the relevance parameter can all reflect the probability that the target account wants to search for the corresponding candidate keywords. Thus, the search tendency parameter can be determined based on the importance parameter, the comprehension difficulty parameter, and the relevance parameter corresponding to each candidate keyword.

[0175] As one possible implementation, a first weight, a second weight, and a third weight can be determined, where the first weight is used to adjust the important parameter, the second weight is used to adjust the comprehension difficulty parameter, and the third weight is used to adjust the related parameter. One possible representation is as shown in formula (1):

[0176] S(k)=α×difficulty(k)+β×relevance(k)+γ×importance(k) (1)

[0177] Where S(k) represents the search tendency parameter of the k-th candidate keyword, difficulty(k) represents the comprehension difficulty parameter of the k-th candidate keyword, relevance(k) represents the relevance parameter of the k-th candidate keyword, importance(k) represents the importance parameter of the k-th candidate keyword, α is the second weight, β is the third weight, and γ is the first weight. As one possible implementation, α can be 0.3, β can be 0.4, γ can be 0.3, etc., and this application does not make specific limitations on this.

[0178] Therefore, if candidate keywords are selected from multiple undetermined candidate keywords based on the first identification model and filtered based on important parameters, and the understanding difficulty parameter and relevance parameter of each candidate keyword are determined based on the second identification model, the important parameters, understanding difficulty parameters, and relevance parameters can all reflect the probability that the target account wants to search for the corresponding candidate keywords. Thus, based on the important parameters, understanding difficulty parameters, and relevance parameters corresponding to each candidate keyword, the search tendency parameters can be determined. This not only directly reuses existing results, which is convenient and fast, but also improves the accuracy of the search tendency parameters by enriching the dimensions of the factors determining the search tendency parameters, thereby improving the accuracy of the target search terms used for subsequent recommendations.

[0179] As mentioned above, the first recognition model and the second recognition model are the same recognition model. For ease of explanation, when the first recognition model and the second recognition model are the same recognition model, they can be referred to as the target recognition model, that is, the target recognition model is the first recognition model or the second recognition model.

[0180] In other words, the target recognition model can not only filter candidate keywords from historical viewed content based on important parameters, but also further filter target search terms from candidate keywords based on important parameters, relevance parameters, and comprehension difficulty parameters. Based on this, embodiments of this application provide a two-stage training method to train the target recognition model, as detailed in D1-D2.

[0181] D1: Obtain the first sample pair and the second sample pair.

[0182] The first sample pair includes the sample content containing the sample keywords and the sample importance parameters of the sample keywords. The second sample pair includes the feature data of the first sample object, the sample comprehension difficulty parameters of the sample keywords, and the sample relevance parameters of the sample keywords.

[0183] Among them, the sample content is the content used as sample data for training, the first sample object is the target object for viewing the sample content, the first sample object feature data is the object feature data of the first sample object, the sample keywords are the keywords marked by the first sample object during the viewing of the sample content, and the sample important parameters are the important parameters corresponding to the sample keywords.

[0184] As one possible implementation, the first sample pair can be represented as <sample content, sample importance parameters>, and the second sample pair can be represented as <first sample object feature data, sample understanding difficulty parameters, and sample relevance parameters>.

[0185] The first sample pair is used to perform the first stage of training, and the second sample pair is used to perform the second stage of training. The sample content in the first sample pair serves as the input data for the first stage of training, and the sample importance parameters in the first sample pair serve as the ground truth (or labels) for the first stage of training. The first sample object feature data in the second sample pair serves as the input data for the second stage of training, and the sample comprehension difficulty parameter and sample relevance parameter in the second sample pair serve as the ground truth for the second stage of training.

[0186] D2: Perform the first stage of training based on the first sample pair, and perform the second stage of training based on the second sample pair to obtain the target recognition model.

[0187] A target recognition model is obtained by fine-tuning all parameters of a recognition model through the first and second stages of training. For example, gradient updates are performed on the parameters of all layers of the recognition model to make the resulting target recognition model more adaptable to downstream tasks. That is, it can filter candidate keywords from historical content based on important parameters, and further filter target search terms from candidate keywords based on important parameters, relevance parameters, and comprehension difficulty parameters.

[0188] This application does not specifically limit the order of the first and second training phases; those skilled in the art can set the order according to actual needs. The two training phases are described separately below.

[0189] (1) The first phase of training, using the first sample pair as training data, is performed as follows:

[0190] E1: Extract keywords from the sample content of the first sample pair using the first initial recognition model to obtain predicted keywords and their prediction parameters.

[0191] The first initial recognition model is a target recognition model that has not yet been fully trained. The two models may have similar structures but different model parameters; this application does not impose specific limitations on this. The predicted keywords are those extracted using the first initial recognition model, and the predicted importance parameters are the important parameters of the predicted keywords. The accuracy of the predicted keywords and predicted importance parameters is related to the accuracy of the first initial recognition model.

[0192] One possible implementation is to set a third prompt word template, thereby obtaining third prompt word information based on the template, and then using this information to guide the first initial recognition model to extract keywords from the sample content. One form of the third prompt word template could be as follows:

[0193] Please extract keywords from the following content:

[0194] [Sample Content]

[0195] The output is then formatted according to the keywords in the content and their predicted key parameters, such as:

[0196] -[Predict Keyword 1] | Predict Important Parameters: [Score 1]

[0197] -[Predict Keyword 2] | Predict Important Parameter: [Score 2]

[0198] ..."

[0199] Therefore, the sample content can be filled into the third prompt word template to obtain the third prompt word information, so as to obtain each predicted keyword and the prediction important parameters corresponding to each predicted keyword through the first initial recognition model.

[0200] E2: Based on the differences between the predicted keywords and the sample keywords included in the first sample pair, as well as the differences between the predicted important parameters and the sample important parameters included in the first sample pair, adjust the model parameters of the first initial recognition model to obtain the second initial recognition model.

[0201] The accuracy of the sample keywords and the accuracy of the sample importance parameters are both relatively high. Therefore, based on the differences between the predicted keywords and the sample keywords included in the first sample pair, as well as the differences between the predicted importance parameters and the sample importance parameters included in the first sample pair, the model parameters of the first initial identification model can be adjusted. This makes the predicted keywords output by the first initial identification model increasingly closer to the sample keywords, and the differences between the predicted importance parameters output by the first initial identification model and the sample importance parameters increasingly smaller, thus obtaining a second initial identification model. For example, the model parameters of the first initial identification model can be optimized by minimizing the mean squared error (MSE) between the content output by the first initial identification model and the actual sample content.

[0202] The second initial recognition model is a recognition model trained based on the first initial recognition model.

[0203] (2) The second phase of training, using the second sample pair as training data, is performed as follows:

[0204] F1: Based on the second sample pair, the correlation analysis of the first sample object feature data and sample keywords is performed through the third initial identification model to determine the prediction comprehension difficulty parameter and the prediction correlation parameter of the sample keywords.

[0205] The third initial recognition model is a target recognition model that has not yet been fully trained. The two models may have similar structures but different model parameters; this application does not impose specific limitations on this. The predicted understanding difficulty parameter is the difficulty parameter for understanding the sample keywords, and the predicted relevance parameter is the relevance parameter for the sample keywords. The accuracy of the predicted understanding difficulty parameter and the predicted relevance parameter is related to the accuracy of the third initial recognition model.

[0206] As one possible implementation, a fourth prompt word template can be set up to obtain fourth prompt word information. This fourth prompt word information can then guide a third initial recognition model to perform correlation analysis on the feature data of the first sample object and the sample keywords. One form of the fourth prompt word template can be as follows:

[0207] Object feature data: [First sample object feature data]

[0208] Please provide personalized scores for the following keywords:

[0209] [Sample Keywords]

[0210] And output as follows:

[0211] The keyword scores for the first sample object are as follows:

[0212] -[Sample Keyword 1]|Prediction Comprehension Difficulty Parameter: [Score 1]|Prediction Relevance Parameter: [Score 1]

[0213] -[Sample Keyword 2]|Prediction Comprehension Difficulty Parameter: [Score 2]|Prediction Relevance Parameter: [Score 2]

[0214] ..."

[0215] Therefore, the feature data of the first sample object and the sample keywords can be filled into the fourth prompt word template to obtain the fourth prompt word information, so as to obtain the prediction comprehension difficulty parameter and prediction related parameter corresponding to each sample keyword through the third initial recognition model.

[0216] F2: Based on the difference between the prediction difficulty parameter of the sample keywords and the sample understanding difficulty parameter, as well as the difference between the prediction relevance parameter of the sample keywords and the sample relevance parameter, adjust the model parameters of the third initial recognition model to obtain the fourth initial recognition model.

[0217] The accuracy of the sample comprehension difficulty parameter and the sample relevance parameter are both relatively high. Therefore, based on the differences between the predicted comprehension difficulty parameter and the actual sample comprehension difficulty parameter, as well as the differences between the predicted relevance parameter and the actual sample relevance parameter, the model parameters of the third initial recognition model can be adjusted. This will reduce the differences between the predicted comprehension difficulty parameter and the actual sample comprehension difficulty parameter output by the third initial recognition model, and also reduce the differences between the predicted relevance parameter and the actual sample relevance parameter, thus obtaining the fourth initial recognition model. For example, the model parameters of the third initial recognition model can be optimized by minimizing the mean squared error (MSE) between the output of the third initial recognition model and the actual sample content.

[0218] The fourth initial recognition model is a recognition model trained based on the third initial recognition model.

[0219] It should be noted that if the first stage of training is performed first, followed by the second stage of training—that is, after obtaining the second initial recognition model based on the first stage of training, the second initial recognition model is then trained again based on the second stage of training to obtain the target recognition model—then the third initial recognition model is the second initial recognition model, and the fourth initial recognition model is the target recognition model.

[0220] Similarly, if the second stage of training is performed first, followed by the first stage of training—that is, after obtaining the fourth initial recognition model based on the second stage of training, the fourth initial recognition model is trained again based on the first stage of training—then the target recognition model is obtained. In other words, the first initial recognition model is the fourth initial recognition model, and the second initial recognition model is the target recognition model.

[0221] Therefore, by fine-tuning all parameters in two stages, the target recognition model can more accurately extract candidate keywords and determine the important parameters, comprehension difficulty parameters, and related parameters corresponding to each candidate keyword. This enables the model to extract target search terms that are more of interest to the target account from multiple candidate keywords, thereby providing more accurate search suggestion services to the target account and improving the user experience.

[0222] The embodiments of this application do not specifically limit the method of obtaining the first sample pair and the second sample pair. The following description uses one method as an example. Please refer to G1-G7 for details.

[0223] G1: Obtain the feature data of the first sample object, multiple sample keywords, and the number of times the sample keywords are labeled by multiple sample objects.

[0224] The first sample object feature data refers to the object feature data of the first sample object, used to describe the characteristics of the first sample object. The sample keywords are the keywords marked by the first sample object during the process of viewing the sample content. Different sample objects may mark different keywords when viewing the same sample content, thus allowing us to obtain the number of times the sample keywords are marked by multiple sample objects. The number of markings refers to the number of times the user (i.e., the sample object) marks the sample keywords in the content application.

[0225] This application does not specifically limit the form of sample keywords, such as content obtained by identifying objects through content applications. This application does not specifically limit the identification method, such as underlining, bolding, bordering, highlighting, or italicizing parts of the text (e.g., a word, a sentence, or a paragraph). If the marking method involves underlining a sentence in the content, the marked text can be called an underlined sentence, etc.

[0226] Taking the collection of training data by highlighting text in an article as an example, when users read an article, they can highlight certain parts of the text. This behavior reflects the user's special attention to that part of the text, and thus the following data can be obtained:

[0227] (1) Underlined content: The sample object can underline any content in the article. The following example illustrates how to underline a sentence.

[0228] (2) Sample object feature data: Relevant information of the sample object that performed the line drawing operation.

[0229] (3) Underline frequency: The number of times different sample objects underline the same sentence.

[0230] To obtain sample keywords, keywords can be extracted from the underlined sentence. This can be done using word segmentation or named entity recognition tools, such as the NLP toolkit (HanLP), which consists of a series of models and algorithms, to identify noun phrases, proper nouns, etc., from the underlined sentence as sample keywords. This method is convenient, fast, simple, and reliable.

[0231] G2: Determine the sample importance parameters of the target sample keywords based on the total number of times the target sample keywords are tagged.

[0232] The target sample keyword is one of the sample keywords. The total number of times the target sample keyword is marked by multiple sample objects can be obtained based on the number of times the target sample keyword is marked by multiple sample objects, and then the sample importance parameter of the target sample keyword can be determined.

[0233] For example, the more times a sentence containing a target keyword is underlined, the higher the importance of that sentence. Therefore, based on the total number of times the target keyword is marked, the importance parameter of that target keyword can be determined.

[0234] G3: Determine the sample-related parameters of the target sample keyword based on the total number of times the target sample keyword is tagged by the sample objects included in the target sample object set.

[0235] The target sample object set includes the first sample object. The similarity between the object feature data of the sample objects included in the target sample object set and the feature data of the first sample object is greater than the similarity threshold. In other words, the sample objects included in the target sample object set are all target objects similar to the target sample object.

[0236] The total number of times a target sample keyword is tagged by the same type of sample object can reflect the degree of relevance between the keyword and that type of sample object. The more times the keyword is tagged, the more relevant it is. Therefore, the sample relevance parameters of the target sample keyword can be determined based on the total number of times the target sample keyword is tagged by the sample objects included in the target sample object set.

[0237] G4: Determine the sample comprehension difficulty parameter of the target sample keywords based on the frequency of occurrence of the target sample keywords and the frequency of occurrence of the tagged keywords.

[0238] For example, the lower the frequency of the target sample keyword in all the underlined sentences, the more professional the keyword representation in the MuBao sample, and the higher the difficulty of understanding it. Therefore, the sample comprehension difficulty parameter of the target sample keyword can be determined based on the frequency of occurrence of the target sample keyword and the frequency of occurrence of the tagged keywords.

[0239] G5: Take each sample keyword as the target sample keyword, and obtain the sample importance parameter, sample relevance parameter, and sample comprehension difficulty parameter corresponding to each sample keyword.

[0240] Each sample keyword is used as the target sample keyword. G2-G4 are then executed to obtain the sample importance parameters, sample relevance parameters, and sample comprehension difficulty parameters corresponding to each sample keyword.

[0241] G6: Determine the first sample pair based on the sample content containing the sample keywords and the sample importance parameters of the sample keywords.

[0242] The first sample pair includes the sample content containing the sample keywords and the important parameters of the sample keywords.

[0243] G7: Determine the second sample pair based on the first sample object feature data, the sample understanding difficulty parameter of the sample keywords, and the sample relevance parameter of the sample keywords.

[0244] The second sample pair includes the first sample object feature data, the sample understanding difficulty parameter of the sample keywords, and the sample relevance parameter of the sample keywords.

[0245] Therefore, by observing the marking actions performed by each target object on each keyword, the degree of attention each target object pays to each keyword can be determined. This allows us to determine the importance parameters, comprehension difficulty parameters, and related parameters of each keyword, thereby obtaining the first and second sample pairs for training the target recognition model. This method is convenient, fast, and highly accurate.

[0246] The search term recommendation method provided in this application embodiment will be described below from the perspective of user operation, as shown in L1-L2. For ease of description, the following embodiments will be described using a terminal device as the executing entity of the search term recommendation method.

[0247] L1: After logging into the content application with the target account, display the input controls and target search terms.

[0248] The target account is the account used by the target object to log in to the content application; it can represent the target object. The content application is an application that provides search and recommendation functions for content.

[0249] Input controls are used to search for input content, which refers to the content the object uses for searching, thus providing a search service based on the input control. For example... Figure 8 The input control 802 shown is used for searching based on input content.

[0250] The target search terms are determined based on the target account's historical viewing content and the target account's target characteristic data, as detailed in S301-S305. For example... Figure 8 As shown, seven target search terms are displayed through area 801.

[0251] After logging into the content application with the target account, users can use the application and enjoy its search services. One possible implementation is to directly display the input controls and target search terms after logging in. Alternatively, some actions can be performed after logging in before displaying the input controls and target search terms; this application does not specify a particular method.

[0252] L2: If a selection operation is performed on the target search term, the content obtained from the search based on the target search term will be displayed.

[0253] The select operation is used to select the desired target search term, such as selecting one target search term from multiple target search terms, or selecting only one target search term to display.

[0254] If a selection operation is performed for a target search term, it means that the target object wants to be searched using the target search term. In response to the selection operation for the target search term, a search is performed based on the target search term, thereby displaying the content obtained from the search based on the target search term.

[0255] In response to the conditions or states on which the operation is performed depends, one or more operations may be performed in real time or with a set delay when the conditions or states on which they depend are met. Unless otherwise specified, there is no restriction on the order in which the multiple operations are performed.

[0256] Therefore, by using historical viewing content and object characteristic data, the search terms used for searching are aligned with the object's preferences, improving the accuracy of search term recommendations. Displaying the target search terms the object might use simultaneously with the input controls allows for search term recommendations even without the object needing to input part of the search term. This not only better meets the object's expectations for recommended search terms but also reduces the difficulty of recommending search terms within the search function, improving the user experience.

[0257] As one possible implementation, if a selection action is obtained for a target search term, indicating that the target account wants to search based on that target search term, then the target search term and the content obtained from the search based on the target search term can be displayed in the input control.

[0258] This application does not specifically limit the order in which the target search term and the search results are displayed. Several methods are described below as examples.

[0259] Method one allows you to display the search results based on the target search term after the input control shows the target search term. See also... Figure 15 After selecting the target search term "computing cost," the target search term is automatically entered into input control 802, thus displaying the target search term in the input control and then redirecting the user to the recommendation page. This recommendation page displays content obtained based on the target search term. Furthermore, displaying the target search term in the input control on this recommendation page helps users clearly understand which search term the current content is based on, improving search efficiency while clearly identifying the search term and enhancing the user experience.

[0260] Method two allows you to display the search results based on the target search term while simultaneously displaying the input control. This is equivalent to... Figure 15 In the search bar, after selecting the target search term "computing cost", you are directly redirected to the recommendation page. Figure 15 Jump to (A) in the middle Figure 15 (C) diagram. This reduces the page refresh rate, thus saving power for the node terminal device, etc.

[0261] One possible implementation is to display input controls and pending search terms after the user logs into the content application with the target account. These pending search terms are supplemented based on the target search term. For example, pre-defined collocations such as "what is it?" or "applications" can be added to the target search term, thus converting it into standardized pending search terms such as "[target search term] what is it?" or "[target search term] application?". Another approach is to identify the type of the target search term, such as product name or technical terminology. Different types of keywords correspond to different supplementary terms, thus determining the supplementary terms—the vocabulary used to supplement the target search term—based on its type. Based on the target search term and its supplementary terms, pending search terms including the target search term are obtained. These pending search terms are then recommended to the target account, thereby achieving the goal of recommending the target search term.

[0262] Therefore, compared to directly recommending target search terms, the pending search content is more understandable. This makes it easier for the target account to use the pending search content, and they don't need to supplement the pending search content, further improving the search efficiency. Moreover, by pre-establishing a mapping relationship between different types of keywords and supplementary terms, suitable supplementary terms can be added for different types of keywords, thereby improving the accuracy of the pending search content and enhancing the user experience.

[0263] As one possible implementation, after logging into the content application through the target account, a search control and recommended content are displayed. The recommended content is content recommended to the target account. In response to a trigger operation on the search control, an input control and the target search term are displayed. This application embodiment does not specifically limit the trigger operation; it can be a click operation, a double-click operation, a swipe operation, etc. This application embodiment does not specifically limit the number of recommended content items; it can be one or multiple. This application embodiment also does not specifically limit the type of recommended content, such as videos, articles, etc.

[0264] See Figure 6 After logging into the content application with the target account, the user is taken to a recommendation page, which displays a search control 602 and multiple recommended items located in area 601. When the user (i.e., the target object) performs a trigger operation on the search control 602, they can be redirected to... Figure 8 The search page shown displays input control 802 and target search term 801.

[0265] Therefore, this content application not only adds a search function to the recommendation feature, achieving integrated search and recommendation, but also, to avoid interfering with the display of recommended content on the corresponding recommendation page, displays the search control only in a small area. Only when this search control is triggered will the user be redirected to the corresponding search page, thus ensuring a good user experience.

[0266] As one possible implementation, to further reduce the space occupied by the search control on the recommendation page, after logging into the content application with the target account, only the recommendation page including the recommended content can be displayed. This is equivalent to hiding the search control on the recommendation page. When the user performs a drag-and-drop operation on the recommendation page, the search control and recommended content are displayed in response to the drag-and-drop operation. This application embodiment does not specifically limit the drag-and-drop operation; it can be a drop-down operation, a left-swipe operation, a right-swipe operation, etc.

[0267] See Figure 16 This figure is a schematic diagram of a search control provided in an embodiment of this application. Figure 16 In this example, we'll use a pull-down operation as a drag-and-drop operation. After logging into the content application with the target account, only a recommendation page containing recommended content can be displayed. This recommendation page doesn't show a search control, effectively hiding it. When the user performs a pull-down operation on the recommendation page, in response to a drag-and-drop operation, the search control 1601 and recommended content will be displayed. When the user clicks the search control 1601, they will be redirected to the search page.

[0268] Therefore, by hiding the search control in the recommendation page, the search control will only be displayed when there is a search need, i.e., after responding to a drag operation on the recommendation page. This allows for displaying as much recommended content as possible within the limited recommendation page, further reducing the space occupied by the search control on the recommendation page.

[0269] As one possible implementation, when a target user selects one recommended item from multiple recommendations, or performs a selection operation on only one displayed recommended item, detailed information about that recommended item is displayed in response to the selection operation. This detailed information includes one or more search keywords, each used to help the target user perform a quick search. In response to a trigger action based on a search keyword, the user is redirected to a search page for that search keyword. The search page includes input controls and search results for the search keyword.

[0270] This detailed information includes one or more search keywords. Each search keyword is determined based on the target account's historical viewing content and the target account's target characteristics data, as detailed in S301-S304. The difference lies in that candidate keywords corresponding to search tendency parameters that meet the tendency criteria are determined as search keywords, which will not be elaborated further here. Search keywords are terms that the target audience may be interested in for further searching.

[0271] As one possible implementation, the search keywords can be displayed in a highlighted form to help the target audience quickly identify which parts of the detailed content are search keywords. This application does not specifically limit the highlighting form; it can be displayed using methods such as highlighting, underlining, or special icons.

[0272] Therefore, by displaying search keywords that the target account might be interested in for further searching within the detailed content of the recommended content, the target account can directly trigger an action based on the search keywords when needed, thereby jumping to the search page for those keywords. This eliminates the need for the target to exit the currently viewed recommended content and then search through the search function. Instead, it directly identifies the search keywords that the target might be interested in in advance and helps them jump directly to the search page, thus improving search efficiency.

[0273] After introducing the search term recommendation methods for the search function, the following section combines... Figure 4 This section explains the content recommendation methods for the recommendation function.

[0274] See Figure 4 This figure is a flowchart illustrating another content recommendation method provided in an embodiment of this application. For ease of description, the following embodiment uses a server as the execution subject of this content recommendation method. Figure 4As shown, the recommended methods for this content include S401-S405.

[0275] S401: Obtain multiple recommended content vectors and the target account's historical search sequence.

[0276] The target account can be the account used by the target object to log in to the content application, and the target account can represent the target object's use of the content application.

[0277] The recommended content vector is used to describe the recommended content recommended to the target object, such as advertisements, articles, videos, etc., and this application does not specifically limit it. The embodiments of this application also do not specifically limit the method of obtaining the recommended content vector based on the recommended content; for example, it can be obtained through one-hot encoding, TF-IDF, and Word2Vec (a natural language processing technique). H3-H4 will be used as examples later, and will not be elaborated upon here.

[0278] In recommendation functions, the interests of a target audience are typically represented by their historical interactions (such as clicks and favorites). However, the intent expressed by users in search functions is often more explicit and timely. Therefore, transforming the target audience's search behavior into an interest representation usable in recommendation functions can be called virtual interest. This virtual interest has the same representation as the content of the target audience's actual interactions and can be directly used in the recall and ranking processes of recommendation functions.

[0279] The historical search sequence includes multiple search results from the target account, ordered chronologically (e.g., older searches appear first, followed by searches closer to the current time). These historical searches reflect the target account's interests, revealing changes in those interests. For example, if the first eight searches are related to technology and the last two to entertainment, it indicates the target account's strong interest in technology and a recent increase in interest in entertainment.

[0280] The embodiments of this application do not specifically limit the number of historical search contents included in the historical search sequence, and those skilled in the art can set it according to actual needs.

[0281] S402: Based on historical search sequences, predict and obtain the virtual interest vector of the target account.

[0282] As mentioned above, a target account is highly likely to be interested in its past search history. Therefore, the historical search sequence can reflect the target account's interests over a period of time. Furthermore, the multiple historical search results within the sequence are ordered chronologically, demonstrating how the target account's interests change over time. Thus, predictions can be made based on the historical search sequence to obtain a virtual interest vector. This virtual interest vector is used to represent the content the target account has recently been interested in.

[0283] The embodiments in this application do not specifically limit the prediction method. For example, it may use an Autoregressive Integrated Moving Average (ARIMA) model, a Long Short-Term Memory (LSTM) model, etc. These will be explained based on I1-I2 later, and will not be repeated here.

[0284] Therefore, the generated virtual interest vector can represent the overall interests reflected in the target account's recent search behavior and can be stored in a database. Whenever the target account makes a new search, the historical search sequence is updated to update the target account's new virtual interest vector, thereby dynamically following changes in the target object's interests and improving the accuracy of recommended content.

[0285] S403: Calculate the similarity between the virtual interest vector and each recommended content vector.

[0286] Similarity is used to compare the similarity between two things. For example, it can be calculated by the distance between the features of the things. If the distance is small, the similarity is large; if the distance is large, the similarity is small.

[0287] The embodiments of this application do not specifically limit the method of calculating similarity, such as using Minkowski distance, Manhattan distance, Euclidean distance, etc.

[0288] S404: The recommended content corresponding to the recommended content vector that meets the similarity condition is identified as supplementary recommended content.

[0289] The similarity condition is used to filter content from multiple recommended content vectors that the target account is more interested in. This application does not specifically limit the similarity condition; those skilled in the art can set it according to actual needs. For example, multiple recommended content vectors can be sorted based on similarity, and the top U recommended content vectors with high interest to the target account can be filtered out as supplementary recommended content, where U is a positive integer. Alternatively, V recommended content vectors with similarity greater than a similarity threshold can be filtered out as supplementary recommended content, where V is a positive integer.

[0290] The supplementary recommendations are content that the target account may be interested in, determined based on historical search sequences. Moreover, since the historical search sequences generally include content actively entered by the target account, they can more accurately meet the target account's explicit and immediate information needs, thereby reflecting the target account's interests. Consequently, the virtual interest vector determined based on the historical search sequences can also reflect the target account's interests. Supplementary recommendations with a high similarity to the virtual interest vector are more likely to be content that the target account is interested in.

[0291] The embodiments of this application do not specifically limit the number of supplementary recommendations. There may be one or more. Those skilled in the art can set them according to actual needs, such as by setting them through similar conditions.

[0292] S405: Based on the supplementary recommendations, determine the target recommendations for the target account.

[0293] The target recommended content is content recommended to the target account. As one possible implementation, the recommended content displayed together with the search control, as mentioned above, can be the target recommended content, or it can be recommended content not determined by S401-S405; this application does not specifically limit it in this regard.

[0294] This application does not specifically limit the method of determining the target recommended content based on supplementary recommended content. For example, the supplementary recommended content can be directly used as the target recommended content. Alternatively, supplementary recommended content can be added to the existing recommended content for recommendation. Yet another example is that the supplementary recommended content and the recall results can be finely sorted together based on the recall results of the existing recommended system to obtain the final recommended content.

[0295] As one possible implementation, taking the K highest similarity scores as an example (where K is a positive integer), the supplementary recommended content corresponds to the K most similar content vectors among multiple recommended content options. Multiple recall results are obtained for the target account; these recall results are those whose accuracy as target recommended content exceeds a preset threshold. Here, recall (match) refers to triggering as many correct results as possible from the full information set and returning these results to the ranking; the results obtained from the recall are the recall outcomes.

[0296] This process uses K supplementary recommendations as a complement to the multiple recall results. From these multiple recall results and the K supplementary recommendations, some content is arbitrarily deleted, resulting in multiple unranked contents. For example, after mixing the multiple recall results and the K supplementary recommendations, the merged result is shuffled to ensure content diversity and truncated to an appropriate size, effectively deleting arbitrary parts of the content. Finally, the unranked contents are sorted to obtain the target recommendations, and fine-tuning yields more accurate target recommendations.

[0297] Therefore, supplementing the original recall results with additional recommended content, breaking it down to ensure content diversity, and truncating it to an appropriate scale can guarantee the speed of subsequent fine ranking and improve the efficiency and accuracy of recommendations.

[0298] As described in the above technical solution, the process involves obtaining recommendation content vectors corresponding to multiple pending recommendations, as well as a historical search sequence composed of the target account's past search content. The target account is highly likely to be interested in its past search content; therefore, the historical search sequence reflects the target account's interests over a period of time. Furthermore, the historical search content within the sequence is ordered chronologically, reflecting changes in the target account's interests over time. Based on this historical search sequence, a virtual interest vector can be obtained. This virtual interest vector reflects the content the target account has recently been interested in, making it more likely that the target account will subsequently view recommendations related to the virtual interest vector. The similarity between the virtual interest vector and each recommendation content vector is calculated. Recommendations corresponding to vectors with similarity scores that meet the similarity criteria are identified as supplementary recommendations. This means that if the supplementary recommendations have a high similarity to the virtual interest vector, the target account is more likely to view them. These supplementary recommendations are then designated as target recommendations for the target account, improving the accuracy of the recommendations. Thus, transforming the historical search sequence into a virtual interest vector allows the recommended content to respond promptly to the user's search needs, improving the accuracy of content recommendations.

[0299] To further improve the accuracy of the similarity between virtual interest vectors and recommended content vectors, they can be placed in the same feature space. This avoids a lack of awareness of their co-occurrence relationship and improves the accuracy of similarity calculation. This application does not specifically limit the method of placing virtual interest vectors and recommended content vectors in the same feature space; two methods are described below as examples.

[0300] Method 1: Extract features using the same recognition model.

[0301] In other words, the third recognition model is used to predict the virtual interest vector (see H1-H2 for details); the third recognition model is also used to extract the recommended content vector (see H3-H4 for details). Thus, both the recommended content vector and the virtual interest vector are extracted using the same recognition model, allowing them to be extracted into the same feature space, which will be explained in detail below.

[0302] H1: The target account's interest is obtained by analyzing historical search sequences through a third identification model.

[0303] The third recognition model is a recognition model, and this application does not specifically limit it. The historical search sequence can be a sequence composed of images, videos, music, etc., or it can be a combination of multimodal content. Therefore, the third recognition model can be determined based on the modality of the historical search sequence. For example, the third recognition model can be a large language model, and this application does not specifically limit it. If the third recognition model is an open-source large language model, it is used to extract the intermediate layer output later. The third recognition model is used to understand the target account's interests reflected in the historical search sequence and convert it into a semantic vector for subsequent similarity calculation. As a possible implementation, if the third recognition model is a large language model, a fifth prompt word template can be set, thereby obtaining the fifth prompt word information based on the fifth prompt word template, and then guiding the third recognition model to analyze the historical search sequence based on the fifth prompt word information. One form of the fifth prompt word template can be as follows:

[0304] Below are the N most recent search records of a target account. Please analyze the interests of the target account reflected in these search records:

[0305] Historical search sequences: [Historical Search Sequences]

[0306] Please analyze the overall interests of the target account as reflected in these historical search results.

[0307] The most recent N historical search results of the target account can be denoted as H. These results are sorted chronologically from earliest to latest to obtain the historical search sequence, which can be denoted as [H1, H2, ..., HN]. This application does not specifically limit the number of historical search results included in the historical search sequence; for example, it can include 5 results.

[0308] Therefore, the fifth prompt word information can be input into the third recognition model, and the third recognition model can be guided to output the overall interest tendency through the fifth prompt word information.

[0309] H2: Extract the vector output from the intermediate layer of the third recognition model to obtain the virtual interest vector of the target account.

[0310] Since the final output of the third recognition model is generally the final result, used for descriptive text in dialogue with the target account, rather than a vector, the vectors output from the intermediate layers of the third recognition model can be extracted to obtain the virtual interest vector of the target account. This vector, output from the intermediate layer, describes the target account's interests. For example, the output of the penultimate layer of the third recognition model can be determined as the virtual interest vector of the target account.

[0311] H3: Retrieve multiple recommended content

[0312] Recommended content refers to content that can be recommended to a target audience, such as videos and text.

[0313] H4: The third recognition model is used to extract features from multiple recommended content items to obtain multiple recommended content vectors.

[0314] The third recognition model extracts features from multiple recommended content items, resulting in a recommended content vector for each item. Each recommended content vector describes the corresponding recommended content.

[0315] Therefore, both the recommended content vector and the virtual interest vector are obtained through feature extraction using the same model, namely the third recognition model. The logic of converting data into vectors is the same for both, which makes it easier for the virtual interest vector and the recommended content vector to be in the same feature space. This avoids the lack of awareness of the co-occurrence relationship between the two and improves the accuracy of similarity calculation.

[0316] Method 2: Achieve vector space alignment through the alignment module.

[0317] In other words, the fourth recognition model is used to predict the virtual interest vector to be determined, and the alignment module is used to transform the virtual interest vector to be determined, so that the transformed virtual interest vector and the recommended content vector are in the same feature space. This will be explained in detail below, see I1-I3.

[0318] I1: By analyzing historical search sequences through the fourth identification model, the interest of the target account can be obtained.

[0319] The third recognition model is just one type of recognition model, and this application does not specifically limit it. For example, an open-source recognition model can be used to extract the intermediate layer output later. Or, the same model as the third recognition model can be used as the fourth recognition model, etc.

[0320] If the fourth identification model and the third identification model are the same model, the target account's interests can be obtained based on the aforementioned H1 method. However, it is no longer limited to extracting features from the recommended content based on the fourth identification model. Those skilled in the art can set it according to actual needs.

[0321] I2: Extract the vector output from the intermediate layer of the fourth recognition model to obtain the undetermined virtual interest vector of the target account.

[0322] Since the final output of the fourth recognition model is generally the final result, used as descriptive text for dialogue with the target account, rather than a vector, the vectors output from the intermediate layers of the fourth recognition model can be extracted to obtain the undetermined virtual interest vector of the target account. This vector output from the intermediate layer describes the target account's interests. For example, the output of the penultimate layer of the fourth recognition model can be determined as the undetermined virtual interest vector of the target account.

[0323] Although the undetermined virtual interest vector is a vector used to describe the interests of the target account, it may reside in a different feature space than the recommended content vector. Therefore, this application proposes to transform the undetermined virtual interest vector into the same feature space as the recommended content vector using an alignment module, thereby obtaining the virtual interest vector. See I3 for details.

[0324] I3: The virtual interest vector is obtained by transforming the virtual interest vector to be determined according to the alignment module.

[0325] The alignment module is used to transform the vector into the feature space of the recommended content vector. In this way, the alignment module aligns the virtual interest vector to be placed in the same feature space as the recommended content vector.

[0326] The embodiments of this application do not specifically limit the structure of the alignment module, and the alignment module can be a multilayer perceptron, etc. Taking a multilayer perceptron as an example, the embodiments of this application do not specifically limit the number of layers of the multilayer perceptron, such as two layers, three layers, etc. The following description uses three layers as an example, see formulas (2)-(4).

[0327] h1 = ReLU(W1×Eh+b1) (2)

[0328] h2 = ReLU(W2×h1+b2) (3)

[0329] Vh = W3×h2+b3 (4)

[0330] Where h1 is the output of the first layer of the multilayer perceptron, h2 is the output of the second layer of the multilayer perceptron, Vh is the output of the third layer of the multilayer perceptron, or the output of the multilayer perceptron, W1, W2 and W3 are three weight matrices, b1, b2 and b3 are three bias vectors, ReLU is the activation function, Eh is the undetermined virtual interest vector, and Vh is the virtual interest vector.

[0331] Therefore, a virtual interest vector describing the interests of a target account can be obtained through a general recognition model. Then, the virtual interest vector is transformed into the same feature space as the content recommendation vector through an alignment module. Since the alignment module and the recognition model are relatively easy to obtain, it is convenient and fast while avoiding the lack of awareness of the co-occurrence relationship between the virtual interest vector and the content recommendation vector, thus improving the accuracy of similarity calculation.

[0332] The embodiments of this application do not specifically limit the training process of the alignment module. The following description uses one training method as an example. Please refer to J1-J6 for details.

[0333] J1: Obtain the third and fourth sample pairs.

[0334] The third sample pair includes a sample search sequence and first sample recommended content. The sample search sequence includes multiple sample search contents, ordered chronologically (e.g., from earliest to latest). The sample search contents are those searched by the second sample object, who is aware of the first sample recommended content. The first sample recommended content is the recommended content viewed based on the most recent sample search content in the sample search sequence. For example, the sample search sequence includes five sample search contents searched on January 1st, 2nd, 3rd, 4th, and 5th, and the first sample recommended content is the recommended content viewed by the second sample object from among the multiple recommendations obtained from the sample search content on January 5th.

[0335] The fourth sample pair includes the sample search sequence and the second sample recommended content, which are all recommendations other than those in the first sample. In other words, the third sample pair can be considered a positive sample, and the fourth sample pair can be considered a negative sample.

[0336] One possible approach is to obtain multiple third sample pairs, then split and arbitrarily combine them to obtain fourth sample pairs, thereby reducing the difficulty of constructing fourth sample pairs. For example, each batch can include multiple third sample pairs, where each sample search sequence corresponds to the first sample recommended content clicked by a second sample object, i.e., a positive sample, while the first sample recommended content included in other third sample pairs within the batch is the negative sample of that sample search sequence.

[0337] J2: The sample search sequence is analyzed by the fourth recognition model to obtain the sample interest. The vector output by the intermediate layer of the fourth recognition model is extracted to obtain the sample undetermined interest vector.

[0338] The sample search sequence is analyzed using the fourth recognition model to obtain the sample interest of the second sample object. The vector output from the intermediate layer of the fourth recognition model is then extracted to obtain the sample's undetermined interest vector. For example, the output of the penultimate layer of the fourth recognition model is determined as the sample's undetermined interest vector for the second sample object.

[0339] J3: Transform the undetermined interest vector of the sample according to the initial alignment model to obtain the sample interest vector.

[0340] The initial alignment model is an alignment model that has not yet been fully trained. The two can be models with the same structure but different model parameters. By training the initial alignment model, the alignment model can be obtained.

[0341] The sample interest vector is a vector obtained by transforming the sample undetermined interest vector through the initial alignment model, and is used to describe the interest of the second sample object based on the sample search sequence.

[0342] J4: Extract features from the recommended content of the first sample to obtain the recommended content vector of the first sample.

[0343] The first sample recommendation vector is used to describe the first sample recommendation content.

[0344] J5: Extract features from the recommended content of the second sample to obtain the recommended content vector of the second sample.

[0345] The second sample recommendation content vector is used to describe the second sample recommendation content.

[0346] J6: Based on the training objective that the distance between the sample interest vector and the first sample recommended content vector is getting closer and closer, and the distance between the sample interest vector and the second sample recommended content vector is getting farther and farther, the model parameters of the initial alignment model are adjusted to obtain the alignment model.

[0347] The first sample's recommended content belongs to positive samples, and the second sample's recommended content belongs to negative samples. Therefore, we can use the distance between the sample interest vector and the first sample's recommended content vector getting closer and closer, and the distance between the sample interest vector and the second sample's recommended content vector getting farther and farther as the training objective. We can then adjust the model parameters of the initial alignment model to obtain the alignment model, so that the transformed sample interest vector becomes more and more similar to positive samples and less and less similar to negative samples.

[0348] The embodiments of this application do not specifically limit the form of the training objective. The training objective is described below using a loss function as an example, see formula (5).

[0349] L align = -log[exp(sim(Ve, Spos) / τ) / Σexp(sim(Ve, Si) / τ)] (5)

[0350] Where sim(·,·) is the cosine similarity function, τ is the temperature parameter used to control the difficulty of contrastive learning, Ve is the sample interest vector, Spos is the first sample recommendation content vector in the same third sample pair as Ve, and Si is the second sample recommendation content vector in the same fourth sample pair as Ve, such as traversing the second sample recommendation content vectors in all third sample pairs within a batch except for Spos.

[0351] Therefore, training the alignment module through contrastive learning can shorten the distance between the sample interest vector and the corresponding first sample recommended content vector, thereby helping to bring the aligned semantic vector closer to the recommendation recall vector space. At the same time, by contrastive learning of positive and negative samples, co-occurrence relationships can be captured, which improves the accuracy of subsequent similarity calculations.

[0352] To minimize the need to modify the recommendation system, an alignment module can be added internally. The following example illustrates how a recommendation system implements recommendation functionality using a deep semantic matching model.

[0353] Deep semantic matching models are a type of deep learning model with a dual-tower structure, including a user tower and an item tower. They are used to learn vector representations of object features and item features, learning feature representations by maximizing the similarity of related samples. This type of model is widely used in the recall process. Here, "item" refers to the recommended content in this embodiment. The user tower includes the aforementioned alignment module.

[0354] Based on this, the aforementioned J1-J6 can be further subdivided into K1-K7, which will be explained in detail below.

[0355] K1: The second sample object feature data is obtained by extracting features from the second sample object data through the user tower in the deep semantic matching model.

[0356] User towers are used to learn vector identifiers of object features. For example, feature extraction can be performed on the second sample object data of the second sample object through user towers to obtain the feature data of the second sample object.

[0357] The second sample object is the sample object of the collected sample search sequence, and the dual-tower model also includes an item tower.

[0358] K2: Obtain the third and fourth sample pairs.

[0359] K3: Analyze the sample search sequence through the fourth recognition model to obtain the sample interest, extract the vector output by the intermediate layer of the fourth recognition model, and obtain the sample undetermined interest vector.

[0360] K4: Transform the sample interest vectors based on the initial alignment model to obtain the sample interest vectors.

[0361] K2-K4 can be found in J1-J3 mentioned above.

[0362] K5: Extract features from the first sample's recommended content using the item pyramid to obtain the first sample's recommended content vector.

[0363] K6: Extract features from the recommended content of the second sample using the item pyramid to obtain the recommended content vector of the second sample.

[0364] Item pyramids are used to learn vector representations of item features. For example, features can be extracted from the first sample of recommended content using an item pyramid to obtain the first sample's recommended content vector. Similarly, features can be extracted from the second sample of recommended content using an item pyramid to obtain the second sample's recommended content vector.

[0365] K7: Based on the first and second sub-training objectives, adjust the model parameters of the initial alignment model to obtain the alignment model.

[0366] The training objective includes a first sub-training objective and a second sub-training objective, which will be explained below.

[0367] The first sub-training objective is to make the distance between the feature data of the second sample object and the recommended content vector of the first sample increasingly closer, and to make the distance between the feature data of the second sample object and the recommended content vector of the second sample increasingly farther. This measures the degree of matching between the sample object feature data and the sample recommended content vector.

[0368] The embodiments of this application do not specifically limit the form of the first sub-training objective. The first sub-training objective is described below using a loss function as an example, see formula (6).

[0369] L dssm= -log[exp(sim(U, Spos) / τ) / Σexp(sim(U, Si) / τ)] (6)

[0370] Where sim(·,·) is the cosine similarity function, τ is the temperature parameter used to control the difficulty of contrastive learning, U is the feature data of the second sample object, which is the sample object of the search sequence of the collected samples, Spos is the first sample recommended content vector of the first sample recommended content clicked by the second sample object, and Si is the second sample recommended content vector of the second sample recommended content not clicked by the second sample object, such as traversing the second sample recommended content vectors of all third sample pairs in the batch except for Spos.

[0371] The second sub-training objective is to make the distance between the sample interest vector and the first sample recommended content vector increasingly closer, and the distance between the sample interest vector and the second sample recommended content vector increasingly farther, as shown in the aforementioned formula (5).

[0372] Therefore, by introducing a deep semantic matching model, we can not only filter recommended content that users may be interested in from the semantic matching level, but also train the alignment module in the process of downstream recommendation tasks. This allows the alignment module to be applied to downstream recommendation tasks as well. The trained alignment module can not only achieve better feature space alignment, but also reuse the results, thereby improving computational accuracy while reducing computational complexity.

[0373] Furthermore, in order to ensure that the aligned feature space has good structural properties, this application provides a specific implementation of K7, namely, adjusting the model parameters of the initial alignment model based on the first sub-training objective and the second sub-training objective to obtain a specific implementation of the alignment model.

[0374] Specifically, based on the first sub-training objective, the second sub-training objective, and the third sub-training objective, the model parameters of the initial alignment model are adjusted to obtain the alignment model.

[0375] In other words, the aforementioned training objective includes not only the first and second sub-training objectives, but also a third sub-training objective. The third sub-training objective is that the sum of the regularization loss for the sample interest vector and the regularization loss for the first sample recommendation content vector becomes smaller and smaller. This application does not specifically limit the form of the third sub-training objective. The third sub-training objective is described below using a loss function as an example, see formula (7).

[0376] L reg = ||Ve||2 + ||Spos||2 (7)

[0377] Where Ve is the sample interest vector, and Spos is the first sample recommendation content vector that is in the same third sample pair as Ve.

[0378] As one possible implementation, if the training objective includes a first sub-training objective, a second sub-training objective, and a third sub-training objective, it can be expressed as Equation (8).

[0379] L total = L dssm + α×L align + β×L reg (8)

[0380] Among them, L total For the training objective, L dssm For the first sub-training objective, refer to the aforementioned formula (6), L align For the second sub-training objective, please refer to the aforementioned formula (5), L reg For the third sub-training objective, refer to the aforementioned formula (7). α and β are weight coefficients used to balance the importance of different loss terms. In practice, the optimal value is determined by the effect on the validation set.

[0381] Therefore, by introducing regularization, such as constraining the L2 norm of vectors, we can not only prevent the alignment model from learning excessively large numerical representations, but also make the vector distributions of the two feature spaces more similar. The smaller the difference in the norm distributions of the two vector spaces, the more similar their geometric structures, which helps to achieve better semantic alignment. This ensures that the aligned feature space has good structural properties.

[0382] As one possible approach to improve recommendation speed, the recommendation process can be divided into two parts: an offline phase and an online service.

[0383] For the offline phase, after obtaining the virtual interest vectors of each target object and the vectors of each recommended content, multiple virtual interest vectors and multiple recommended content vectors can be stored in the database. In the database, a mapping relationship is established between the virtual interest vectors and each target object, such as establishing a mapping relationship between the object identifier of the target account and the virtual interest vector of the target account.

[0384] For online services, if a pre-search request is received from an object identifier, the virtual interest vector corresponding to the object identifier is determined according to the mapping relationship, and steps S403-S405 are executed to recommend relevant target content to the target object corresponding to that object identifier. The pre-search request refers to an action where the target object corresponding to the identifier has a search request, such as opening a search control or opening an application that provides search services.

[0385] For example, after obtaining the pre-search request, the object identifiers included in the pre-search request can be obtained. Based on the mapping relationship, the virtual interest vector corresponding to the object identifier can be determined. Then, the Approximate Nearest Neighbor (ANN) method can be used to find the top K recommended content vectors most similar to the virtual interest vector. The recommended content corresponding to the K recommended content vectors is determined as supplementary recommended content. Multiple supplementary recommended content are used to supplement the original multiple recall results. After mixing the multiple recall results and the K supplementary recommended content, the merged result is shuffled to ensure content diversity and truncated to an appropriate size, which is equivalent to deleting arbitrary parts of the content. Finally, the multiple unranked contents are sorted to obtain the target recommended content, and fine-tuning yields more accurate target recommended content.

[0386] Therefore, by pre-extracting virtual interest vectors for each target object during the offline phase and establishing a mapping relationship between target objects and virtual interest vectors, when a pre-search request for a certain target object is received, the virtual interest vector of that target object can be directly obtained based on the mapping relationship, and target recommended content can be determined for the target object based on this virtual interest vector. This reduces the computational load, ensures the efficiency of recall, and fully utilizes the interest information reflected in the user's search behavior, achieving an effective conversion from search intent to recommendation recall.

[0387] To facilitate a further understanding of the technical solutions provided in the embodiments of this application, the following description uses the server as the execution subject of the search term recommendation method provided in the embodiments of this application, and takes the recommendation of articles and other content in chat software as an example to provide an overall exemplary introduction to the search term recommendation method.

[0388] With the development of mobile internet, search and recommendation functions have become the two main entry points for users to obtain information. Search functions can accurately meet users' explicit, immediate information needs, while recommendation functions can proactively recommend products based on users' potential interests. However, the systems corresponding to these two functions—search systems and recommendation systems—usually operate independently, leading to a fragmented user experience and failing to fully leverage their respective advantages. Therefore, the industry has begun exploring the deep integration of search and recommendation systems to provide users with a more intelligent and seamless service experience.

[0389] However, existing technical solutions suffer from the following main problems: 1) Shallow fusion level: Existing solutions mainly focus on simple data integration or module merging in the system architecture, lacking a deep understanding and transformation of user intent and interests. For example, using data with more dimensions in the hope of training a better model. 2) Primarily one-way fusion: Most solutions may consider using search data to enhance recommendation effects, but ignore the feedback effect of recommendation scenarios on search, failing to form a virtuous cycle. 3) Lack of personalized consideration: Existing fusion solutions often adopt a uniform processing logic, failing to fully consider the different knowledge backgrounds and interest preferences of different users, resulting in insufficiently accurate fusion effects.

[0390] Based on this, the embodiments of this application provide corresponding solutions, namely: 1) Introducing a recognition model to understand the semantics of users' historical search sequences, and mapping the semantic representation extracted by the recognition model to the item representation space of the recommendation system through feature space alignment, thereby achieving accurate conversion from search intent to recommendation interest and making the integration more in-depth. 2) Designing a two-stage keyword extraction scheme based on the recognition model: first, identifying important concepts and professional terms from the article, and then combining object feature data for personalized scoring. 3) Proposing a method for fully parameter-supervised fine-tuning of the recognition model based on users' underlining behavior, utilizing real user spontaneous behavior data to improve the accuracy of keyword mining by the recognition model. Furthermore, through a mechanism of guiding search through recommended content, the recognition model intelligently extracts keywords and generates personalized search suggestions, thus achieving bidirectional integration of search and recommendation.

[0391] These solutions break down the boundaries between search and recommendation scenarios, providing users with a more intelligent and seamless content service experience. A detailed explanation follows.

[0392] See Figure 5 This figure is a schematic diagram illustrating how to display recommended content entry points in a chat software interface, according to an embodiment of this application. Figure 5 In this context, the recommended content belongs to an information feed product, and the entry point for the recommended content is located on page 501 of the chat software's main page. Performing actions such as clicking or double-clicking on this entry point redirects the user to the recommended content page.

[0393] See Figure 6 This figure is a schematic diagram of a recommended content page provided in an embodiment of this application. Figure 6 In this context, the content recommendation function can not only perform recommendation functions but also search functions, such as... Figure 6 Section 601 is the content recommendation area corresponding to the recommendation function, used to display the target recommended content. Users can access the corresponding target recommended content for detailed viewing through trigger actions, such as... Figure 7 As shown. Figure 6602 is the search control corresponding to the search function. After executing a trigger operation on the search control, a search can be performed, such as... Figure 8 As shown.

[0394] The following is combined Figure 7 and Figure 8 The search and recommendation functions will be explained separately.

[0395] See Figure 8 The figure is a schematic diagram of a search function provided in an embodiment of this application.

[0396] In the first application scenario (i.e., the search scenario), user search behavior within the search function is leveraged to enhance recommendation effectiveness. When a user enters search terms in the search control and browses related content, the system analyzes the user's search intent in real time and transforms it into an interest vector usable by the recommendation system. For example, when a user searches for "low-carb diet," the system not only returns directly related search results but also transforms this search intent into a representation of the user's interest in areas such as healthy eating and nutrition, thereby pushing relevant high-quality articles to the user in subsequent content recommendations. Furthermore, even before a user enters keywords, the system determines target recommendation terms based on the user's historical browsing history, as shown in Figure 801, thus helping the user quickly complete the search.

[0397] See Figure 7 This figure is a schematic diagram of viewing target recommended content provided in an embodiment of this application.

[0398] In the second application scenario (i.e., the recommendation scenario), the system analyzes the articles a user is reading to predict what content the user might want to learn more about, and presents this as search suggestions. For example, when a user is reading an article introducing artificial intelligence technology, the system analyzes the article content in real time, identifies technical terms or important concepts mentioned in the text (such as "computing power cost"), and transforms these terms into personalized search suggestions based on the user's knowledge background and interests. Figure 7 Control 701 in the image illustrates search suggestions. When a user performs a trigger action on this control, corresponding search suggestions can be displayed. Alternatively, search suggestions can be displayed directly. Users can use these search suggestions to directly jump to the search function for further information; for example, selecting a search suggestion will directly lead to... Figure 8 .

[0399] In addition, articles can include a built-in highlighting feature to help users mark content of interest, such as... Figure 7 As shown in 702, there are sentences underlined by the user (not necessarily the user currently reading the article).

[0400] The search and recommendation functions are explained below.

[0401] (1) Recommendation function.

[0402] In recommender systems, user interests are typically represented by their historical interactions (such as clicks and favorites). However, user intent expressed in search scenarios is often more explicit and timely. The core idea of ​​this module is to transform user search behavior into an interest representation usable by the recommender system, called "virtual interest." This virtual interest has the same representation as the content the user actually interacts with and can be directly used in the recall and ranking processes of the recommender system. This will be explained in detail below.

[0403] See Figure 9 This figure is a schematic diagram of a recommendation function provided in an embodiment of this application.

[0404] S1: Obtain the historical search sequence of the target object.

[0405] The historical search sequence includes multiple search results for the target object, sorted by time.

[0406] S2: Analyze the historical search sequence using the fourth identification model to obtain the interest of the target object.

[0407] S3: Extract the vector output from the intermediate layer of the fourth recognition model to obtain the undetermined virtual interest vector of the target object.

[0408] The vector output by the intermediate layer is used to describe the interests of the target account.

[0409] S4: The virtual interest vector is obtained by transforming the virtual interest vector in the user tower of the deep semantic matching model according to the alignment module.

[0410] The alignment module is used to transform the vectors into the feature space of the recommended content vectors. The training process of the alignment module can be found in the aforementioned K1-K7, etc.

[0411] S5: Extract features from multiple recommended items using the item pyramid in the deep semantic matching model to obtain multiple recommended item vectors.

[0412] S6: Store multiple recommended content vectors and virtual interest vectors of target objects in the database.

[0413] Whenever a new search is performed on the target object, the historical search sequence is updated and a virtual interest vector is regenerated, thereby dynamically following the changes in the target object's interests.

[0414] S7: If a pre-search request for the object identifier of the target object is obtained, the virtual interest vector corresponding to the object identifier of the target object is determined according to the mapping relationship.

[0415] S8: Calculate the similarity between the virtual interest vector and each recommended content vector.

[0416] S9: The recommended content corresponding to the recommended content vector that meets the similarity condition is identified as supplementary recommended content.

[0417] S10: Determine the target recommended content based on the supplementary recommendations.

[0418] After obtaining the pre-search request, the object identifiers included in the pre-search request can be obtained. Based on the mapping relationship, the virtual interest vector corresponding to the object identifier can be determined. Then, the Approximate Nearest Neighbor (ANN) method can be used to find the top K recommended content vectors most similar to the virtual interest vector. The recommended content corresponding to the K recommended content vectors is determined as supplementary recommended content. Multiple supplementary recommended content are used to supplement the original multiple recall results. After the multiple recall results and the K supplementary recommended content are mixed, the merged result is shuffled to ensure content diversity and truncated to an appropriate size, which is equivalent to deleting arbitrary parts of the content. Finally, the multiple unranked contents are sorted to obtain the target recommended content. Fine-tuning yields more accurate target recommended content.

[0419] (2) Search function.

[0420] This module aims to extract keywords that users might want to explore further from the articles they read in recommended scenarios and transform them into personalized search suggestions. This approach helps users shift from passively receiving recommended content to actively exploring knowledge areas of interest. This process consists of two stages: keyword extraction and personalized filtering, which will be explained below.

[0421] See Figure 10 The figure is a schematic diagram of a search function provided in an embodiment of this application.

[0422] S11: Retrieve the target account's historical viewing content.

[0423] S12: Extract the top N candidate keywords based on the ranking of important parameters in the article using a target recognition model.

[0424] For example, a target recognition model is used to identify historical viewed content, resulting in multiple candidate keywords. The target recognition model then determines the important parameters corresponding to each candidate keyword. Candidate keywords whose important parameters meet the important criteria are then identified as candidate keywords. This process allows for the identification of multiple candidate keywords from historical viewed content; see A1-A3 for details.

[0425] S13: Obtain the candidate word set.

[0426] The candidate word set includes multiple candidate keywords.

[0427] S14: Obtain the object characteristic data of the target object.

[0428] S15: Perform personalized filtering and comprehensive scoring through the target recognition model.

[0429] For example, by using a target recognition model to perform correlation analysis on object feature data and multiple candidate keywords, the understanding difficulty parameter and the relevant parameter corresponding to each candidate keyword are determined. Based on the importance parameter, understanding difficulty parameter and relevant parameter corresponding to each candidate keyword, the search tendency parameter corresponding to each candidate keyword is determined, as shown in formula (1). The candidate keywords corresponding to the search tendency parameters that meet the tendency conditions are determined as the target search terms.

[0430] S16: Recommend target search terms to the target account.

[0431] For example, identify the type of the target search term; determine the complementary terms of the target search term based on the type of the target search term; obtain the pending search content including the target search term based on the target search term and the complementary terms of the target search term; and recommend the pending search content to the target account.

[0432] S17: Obtain the first sample pair and the second sample.

[0433] S18: Perform the first phase of training based on the first sample pair, and perform the second phase of training based on the second sample pair.

[0434] See D1-D2 for details.

[0435] S19: Obtain the target recognition model.

[0436] Therefore, after the integrated search and recommendation system with recommendation and search functions provided in this application was launched to serve users of a certain chat software, comparative experiments verified that the average daily search usage per user increased by 1.2%, and the average daily reading volume per user in the recommendation scenario increased by 0.4%. The increased user consumption of the search function and recommended articles indicates a significant improvement in the user experience in both scenarios.

[0437] In response to the search term recommendation method described above, this application also provides a corresponding search term recommendation device so that the above search term recommendation method can be applied and implemented in practice.

[0438] See Figure 11The figure is a schematic diagram of the structure of a search term recommendation device provided in an embodiment of this application. Figure 11 As shown, the search term recommendation device 1100 includes: an acquisition unit 1101, a determination unit 1102, a calculation unit 1103, a filtering unit 1104, and a recommendation unit 1105;

[0439] The acquisition unit 1101 is used to acquire the historical viewing content viewed through the target account and the object feature data corresponding to the target account;

[0440] The determining unit 1102 is used to determine multiple candidate keywords from the historical viewing content;

[0441] The calculation unit 1103 is used to determine the search tendency parameters of each candidate keyword based on the object feature data. The search tendency parameters are used to characterize the probability that the target account wants to search for the candidate keywords.

[0442] The filtering unit 1104 is used to determine the candidate keywords corresponding to the search tendency parameters that meet the tendency conditions as the target search terms;

[0443] The recommendation unit 1105 is used to recommend the target search term to the target account.

[0444] As can be seen from the above technical solution, the system acquires historical content viewed by the target account and object feature data describing the characteristics of the target account. The target account is highly likely to be interested in the content it viewed, thus historical viewing content reflects the target account's interests over a period of time. Multiple candidate keywords can then be identified from this historical viewing content, and the target account is highly likely to conduct subsequent searches based on these candidate keywords. To further improve the accuracy of candidate keyword recommendations, each candidate keyword can be further filtered based on object feature data that describes the characteristics of the object. Specifically, based on the target account's object feature data, search tendency parameters for each candidate keyword are determined. These search tendency parameters characterize the probability that the target account wants to search based on the candidate keywords, or in other words, they reflect the target account's level of interest in the candidate keywords. Candidate keywords corresponding to search tendency parameters that meet the tendency criteria are identified as target search terms. Since the target account is highly likely to search based on these target search terms, these target search terms are recommended to the target account, improving search efficiency. Therefore, by using historical viewing content and object feature data, search terms that match the target's preferences are obtained, improving the accuracy of search term recommendations. Moreover, it can make recommendations without requiring users to input partial search terms, which better meets users' expectations for recommended search terms, reduces the difficulty of recommending search terms in the search function, and improves the user experience.

[0445] As one possible implementation, the computing unit 1103 is specifically used for:

[0446] The second identification model is used to perform correlation analysis on the object feature data and the multiple candidate keywords to determine the comprehension difficulty parameter and the correlation parameter corresponding to each candidate keyword. The comprehension difficulty parameter is used to describe the difficulty for the target account to understand the candidate keywords, and the correlation parameter is used to describe the degree of correlation between the candidate keywords and the target account.

[0447] Based on the comprehension difficulty parameter and the relevant parameters corresponding to each candidate keyword, the search tendency parameter corresponding to each candidate keyword is determined.

[0448] As one possible implementation, the determining unit 1102 is further configured to determine important parameters corresponding to each of the candidate keywords, the important parameters being used to describe the probability that the target account will search based on the candidate keywords;

[0449] The calculation unit 1103 is specifically used to determine the search tendency parameters corresponding to each candidate keyword based on the importance parameters, the comprehension difficulty parameters, and the relevance parameters corresponding to each candidate keyword.

[0450] As one possible implementation, the determining unit 1102 is specifically used for:

[0451] The historical viewing content is identified using a first recognition model to obtain multiple candidate keywords included in the historical viewing content;

[0452] The first identification model determines the important parameters corresponding to each of the candidate keywords to be determined. The important parameters are used to describe the probability that the target account will search based on the candidate keywords to be determined.

[0453] The candidate keywords corresponding to the important parameters that meet the important conditions are determined as the candidate keywords.

[0454] As one possible implementation, the recommendation unit 1105 is specifically used for:

[0455] Identify the type of the target search term;

[0456] Based on the type of the target search term, determine the supplementary terms for the target search term;

[0457] Based on the target search term and the supplementary terms of the target search term, the pending search content including the target search term is obtained;

[0458] Recommend the pending search content to the target account.

[0459] See Figure 17 This figure is a schematic diagram of the structure of a training device for a recognition model provided in an embodiment of this application. Figure 17 As shown, the training device 1700 for the recognition model includes an acquisition unit 1701 and a training unit 1702.

[0460] The acquisition unit 1701 is used to acquire a first sample pair and a second sample pair. The first sample pair includes the sample content where the sample keyword is located and the sample importance parameter of the sample keyword. The second sample pair includes the feature data of the first sample object, the sample understanding difficulty parameter of the sample keyword, and the sample relevance parameter of the sample keyword. The sample keyword is the keyword marked by the first sample object during the process of viewing the sample content, as represented by the feature data of the first sample object.

[0461] The training unit 1702 is used to perform a first stage of training based on the first sample pair and a second stage of training based on the second sample pair to obtain a target recognition model, wherein the target recognition model is either the first recognition model or the second recognition model.

[0462] The first phase of training is as follows:

[0463] The first initial recognition model is used to extract keywords from the sample content included in the first sample pair to obtain predicted keywords and the predicted important parameters of the predicted keywords.

[0464] Based on the differences between the predicted keywords and the sample keywords included in the first sample pair, and the differences between the predicted importance parameters and the sample importance parameters included in the first sample pair, the model parameters of the first initial recognition model are adjusted to obtain the second initial recognition model.

[0465] The second phase of training is as follows:

[0466] Based on the second sample pair, a correlation analysis is performed on the feature data of the first sample object and the sample keywords using a third initial recognition model to determine the prediction comprehension difficulty parameter and the prediction correlation parameter of the sample keywords;

[0467] Based on the difference between the predicted understanding difficulty parameter of the sample keywords and the sample understanding difficulty parameter, and the difference between the predicted relevance parameter of the sample keywords and the sample relevance parameter, the model parameters of the third initial recognition model are adjusted to obtain the fourth initial recognition model.

[0468] If the first stage of training occurs before the second stage of training, then the third initial recognition model is the second initial recognition model, and the fourth initial recognition model is the target recognition model; if the second stage of training occurs before the first stage of training, then the first initial recognition model is the fourth initial recognition model, and the second initial recognition model is the target recognition model.

[0469] As one possible implementation, the training device 1700 for the recognition model further includes a construction unit for:

[0470] Obtain the feature data of the first sample object, multiple sample keywords, and the number of times the sample keywords are marked by multiple sample objects;

[0471] The sample importance parameter of the target sample keyword is determined based on the total number of times the target sample keyword is tagged. The target sample keyword is one of the sample keywords. The total number of times the target sample keyword is tagged is determined based on the number of times the target sample keyword is tagged by multiple sample objects.

[0472] Based on the total number of times the target sample keyword is marked by the sample objects included in the target sample object set, the sample-related parameters of the target sample keyword are determined. The target sample object set includes the first sample object, and the similarity between the object feature data of the sample objects included in the target sample object set and the feature data of the first sample object is greater than a similarity threshold.

[0473] The sample comprehension difficulty parameter of the target sample keywords is determined based on the frequency of occurrence of the target sample keywords and the frequency of occurrence of the tagged keywords.

[0474] Each of the sample keywords is used as the target sample keyword to obtain the sample importance parameter, the sample relevance parameter, and the sample comprehension difficulty parameter corresponding to each of the sample keywords.

[0475] The first sample pair is determined based on the sample content containing the sample keywords and the sample importance parameters of the sample keywords;

[0476] The second sample pair is determined based on the first sample object feature data, the sample comprehension difficulty parameter of the sample keywords, and the sample relevance parameter of the sample keywords.

[0477] See Figure 12 The figure is a schematic diagram of the structure of a search term recommendation device provided in an embodiment of this application. Figure 12 As shown, the search term recommendation device 1200 includes: a first display unit 1201 and a second display unit 1202;

[0478] The first display unit 1201 is used to display an input control and a target search term after logging into the content application through the target account. The input control is used to search for the input content, and the target search term is determined based on the target account's historical viewing content and the target account's object feature data.

[0479] The second display unit 1202 is used to display the content obtained by searching based on the target search term if a selection operation for the target search term is obtained.

[0480] As described in the above technical solution, the process involves acquiring historical content viewed by the target account and object feature data describing the target account's characteristics. The target account is highly likely to be interested in the content it viewed, thus historical viewing content reflects the target account's interests over a period of time. Multiple candidate keywords can then be identified from this historical viewing content, and the target account is highly likely to conduct subsequent searches based on these candidate keywords. To further improve the accuracy of candidate keyword recommendations, each candidate keyword can be further filtered based on object feature data that describes the object's characteristics. Specifically, based on the target account's object feature data, search tendency parameters for each candidate keyword are determined. These search tendency parameters characterize the probability that the target account wants to search based on the candidate keywords, or in other words, they reflect the target account's level of interest in the candidate keywords. Candidate keywords corresponding to search tendency parameters that meet the tendency criteria are identified as target search terms. Since the target account is highly likely to search based on these target search terms, these target search terms are recommended to the target account, improving search efficiency. Therefore, by combining historical viewing content and object feature data, search terms that match the target's preferences are obtained, improving the accuracy of search term recommendations. Moreover, it can make recommendations without requiring users to input partial search terms, which better meets users' expectations for recommended search terms, reduces the difficulty of recommending search terms in the search function, and improves the user experience.

[0481] As one possible implementation, the second display unit 1202 is specifically used for:

[0482] If a selection operation is obtained for the target search term, the target search term is displayed in the input control;

[0483] Displays the results obtained from a search based on the target search term.

[0484] As one possible implementation, the first display unit 1201 is specifically used for:

[0485] After logging into the content application through the target account, the input control and pending search terms are displayed. The pending search terms are obtained by supplementing the target search terms.

[0486] As one possible implementation, the first display unit 1201 is specifically used for:

[0487] After logging into the content application through the target account, a search control and recommended content are displayed, wherein the recommended content is content recommended to the target account;

[0488] In response to a trigger operation on the search control, the input control and the target search term are displayed.

[0489] As one possible implementation, the first display unit 1201 is specifically used for:

[0490] After logging into the content application through the target account, a recommendation page including the recommended content is displayed;

[0491] In response to a drag-and-drop operation on the recommended page, the search control and the recommended content are displayed.

[0492] As one possible implementation, the search term recommendation device 1200 also includes a third display unit for:

[0493] In response to a selection operation on the recommended content, the detailed content of the recommended content is displayed, including search keywords, which are determined based on the target account's historical viewing content and the target account's object characteristic data;

[0494] In response to a triggering action for the search keyword, the user is redirected to a search page for the search keyword, the search page including input controls and search results for the search keyword.

[0495] As one possible implementation, if the recommended content includes the target recommended content, the search term recommendation device 1200 further includes a content recommendation sub-device, which includes an acquisition unit, a prediction unit, a calculation unit, a determination unit, and a recommendation unit.

[0496] The acquisition unit is used to acquire multiple recommended content vectors and the historical search sequence of the target account. The historical search sequence includes multiple historical search contents that the target account has searched, sorted by time.

[0497] The prediction unit is used to make predictions based on the historical search sequence to obtain the virtual interest vector of the target account.

[0498] The computing unit is used to calculate the similarity between the virtual interest vector and each of the recommended content vectors;

[0499] The determining unit is used to determine the recommended content corresponding to the recommended content vector that satisfies the similarity condition as supplementary recommended content;

[0500] The recommendation unit is used to determine the target recommendation content for the target account based on the supplementary recommendation content.

[0501] This application also provides a computer device, which can be a server or a terminal device. The computer device provided in this application will be described below from a hardware implementation perspective. Figure 13 The diagram shown is a structural schematic of the server. Figure 14 The diagram shown is a structural schematic of the terminal device.

[0502] See Figure 13 This figure is a schematic diagram of a server structure provided in an embodiment of this application. The server 1400 can vary considerably due to different configurations or performance. It may include one or more processors 1422, such as a central processing unit (CPU), memory 1432, and one or more application programs 1442 or data storage media 1430 (e.g., one or more mass storage devices). The memory 1432 and storage media 1430 can be temporary or persistent storage. The program stored in the storage media 1430 may include one or more modules (not shown in the figure), each module may include a series of instruction operations on the server. Furthermore, the processor 1422 may be configured to communicate with the storage media 1430 and execute the series of instruction operations in the storage media 1430 on the server 1400.

[0503] Server 1400 may also include one or more power supplies 1426, one or more wired or wireless network interfaces 1450, one or more input / output interfaces 1458, and / or one or more operating systems 1441, such as Windows Server. TM Mac OS X TM Unix TM Linux TM FreeBSD TM etc.

[0504] The steps performed by the server in the above embodiments can be based on this Figure 13 The server structure shown.

[0505] The processor 1422 is used to execute the method steps of any specific implementation of the search term recommendation method in the embodiments of this application, or to execute the method steps of any specific implementation of the recognition model training method in the embodiments of this application.

[0506] See Figure 14 This figure is a schematic diagram of the structure of a terminal device provided in an embodiment of this application. The description will be based on a smartphone as an example. Figure 14The diagram shown is a partial structural block diagram of the smartphone, which includes: a radio frequency (RF) circuit 1510, a memory 1520, an input unit 1530, a display unit 1540, a sensor 1550, an audio circuit 1560, a Wi-Fi module 1570, a processor 1580, and a power supply 1590, among other components. Those skilled in the art will understand that... Figure 14 The smartphone structure shown does not constitute a limitation on smartphones and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0507] The following is combined Figure 14 A detailed introduction to the various components of a smartphone:

[0508] The RF circuit 1510 can be used to receive and transmit signals during information transmission or calls. In particular, it receives downlink information from the base station and processes it with the processor 1580; in addition, it transmits uplink data to the base station.

[0509] The memory 1520 can be used to store software programs and modules, and the processor 1580 runs the software programs and modules stored in the memory 1520 to realize various functions and data processing of the smartphone.

[0510] Input unit 1530 can be used to receive input numeric or character information and generate key signal inputs related to user settings and function control of the smartphone. Specifically, input unit 1530 may include touch panel 1531 and other input devices 1532. Touch panel 1531, also known as a touch screen, can collect touch operations on or near the user and drive corresponding connected devices according to a pre-set program. In addition to touch panel 1531, input unit 1530 may also include other input devices 1532. Specifically, other input devices 1532 may include, but are not limited to, one or more of the following: physical keyboard, function keys (such as volume control buttons, power buttons, etc.), trackball, mouse, joystick, etc.

[0511] The display unit 1540 can be used to display information input by the user or information provided to the user, as well as various menus of the smartphone. The display unit 1540 may include a display panel 1541, which may optionally be configured as a liquid crystal display (LCD), an organic light-emitting diode (OLED), or the like.

[0512] Smartphones may also include at least one sensor 1550, such as a light sensor, a motion sensor, and other sensors. Other sensors that smartphones may also be equipped with, such as gyroscopes, barometers, hygrometers, thermometers, and infrared sensors, will not be detailed here.

[0513] Audio circuit 1560, speaker 1561, and microphone 1562 provide an audio interface between the user and the smartphone. Audio circuit 1560 converts received audio data into electrical signals and transmits them to speaker 1561, where speaker 1561 converts them into sound signals for output. On the other hand, microphone 1562 converts collected sound signals into electrical signals, which are received by audio circuit 1560, converted into audio data, and then processed by processor 1580 before being transmitted via RF circuit 1510 to, for example, another smartphone, or the audio data can be output to memory 1520 for further processing.

[0514] The processor 1580 is the control center of the smartphone, connecting various parts of the smartphone through various interfaces and lines. It performs various functions and processes data by running or executing software programs and / or modules stored in the memory 1520, and by calling data stored in the memory 1520. Optionally, the processor 1580 may include one or more processing units.

[0515] The smartphone also includes a power supply 1590 (such as a battery) that supplies power to various components. Preferably, the power supply can be logically connected to the processor 1580 through a power management system, thereby enabling functions such as charging, discharging, and power consumption management through the power management system.

[0516] Although not shown, smartphones may also include a camera, Bluetooth module, etc., which will not be described in detail here.

[0517] In this embodiment of the application, the memory 1520 included in the smartphone can store computer programs and transmit the computer programs to the processor.

[0518] The processor 1580 included in the smartphone can execute method steps of any specific implementation of the search term recommendation method provided in the above embodiments according to the instructions in the computer program, or execute method steps of any specific implementation of the recognition model training method in the embodiments of this application.

[0519] This application also provides a computer-readable storage medium for storing a computer program that executes the method steps of any specific implementation of the search term recommendation method provided in the above embodiments, or executes the method steps of any specific implementation of the recognition model training method in the embodiments of this application.

[0520] On the other hand, embodiments of this application provide a computer program product including a computer program, which, when run on a computer device, causes the computer device to perform method steps of any specific implementation of the search term recommendation method provided in the various optional implementations of the above aspects, or to perform method steps of any specific implementation of the training method of the recognition model in the embodiments of this application.

[0521] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium can be at least one of the following media: read-only memory (ROM), RAM, magnetic disk or optical disk, and other media that can store computer programs.

[0522] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “corresponding,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0523] In this application embodiment, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0524] It should be noted that the various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for the device and system embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method embodiments. The device and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the solution in this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0525] The above description is merely one specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Based on the implementation methods provided in the above aspects, this application can also be further combined to provide more implementation methods. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A search term recommendation method, characterized in that, The method includes: Obtain the historical viewing content viewed through the target account and the object feature data corresponding to the target account; Multiple candidate keywords were identified from the historical viewing content; Based on the object feature data, a search tendency parameter is determined for each of the candidate keywords, wherein the search tendency parameter is used to characterize the probability that the target account wants to search for the candidate keywords; Candidate keywords that meet the search tendency parameters are identified as target search terms; Recommend the target search term to the target account.

2. The method according to claim 1, characterized in that, The step of determining the search tendency parameters of each candidate keyword based on the object feature data includes: The second identification model is used to perform correlation analysis on the object feature data and the multiple candidate keywords to determine the comprehension difficulty parameter and the correlation parameter corresponding to each candidate keyword. The comprehension difficulty parameter is used to describe the difficulty for the target account to understand the candidate keywords, and the correlation parameter is used to describe the degree of correlation between the candidate keywords and the target account. Based on the comprehension difficulty parameter and the relevant parameters corresponding to each candidate keyword, the search tendency parameter corresponding to each candidate keyword is determined.

3. The method according to claim 2, characterized in that, The method further includes: Determine the important parameters corresponding to each of the candidate keywords, wherein the important parameters are used to describe the probability that the target account will search based on the candidate keywords; The step of determining the search tendency parameters corresponding to each candidate keyword based on the comprehension difficulty parameter and the relevance parameter corresponding to each candidate keyword includes: Based on the importance parameters, comprehension difficulty parameters, and relevance parameters corresponding to each candidate keyword, the search tendency parameters corresponding to each candidate keyword are determined.

4. The method according to claim 1, characterized in that, The step of identifying multiple candidate keywords from the historical viewing content includes: The historical viewing content is identified using a first recognition model to obtain multiple candidate keywords included in the historical viewing content; The first identification model determines the important parameters corresponding to each of the candidate keywords to be determined. The important parameters are used to describe the probability that the target account will search based on the candidate keywords to be determined. The candidate keywords corresponding to the important parameters that meet the important conditions are determined as the candidate keywords.

5. The method according to any one of claims 1-4, characterized in that, The recommendation of the target search term to the target account includes: Identify the type of the target search term; Based on the type of the target search term, determine the supplementary terms for the target search term; Based on the target search term and the supplementary terms of the target search term, the pending search content including the target search term is obtained; Recommend the pending search content to the target account.

6. A method for training a recognition model, characterized in that, The method includes: Obtain a first sample pair and a second sample pair. The first sample pair includes the sample content where the sample keyword is located and the sample importance parameter of the sample keyword. The second sample pair includes the feature data of the first sample object, the sample understanding difficulty parameter of the sample keyword, and the sample relevance parameter of the sample keyword. The sample keyword is the keyword marked by the first sample object during the process of viewing the sample content, as represented by the feature data of the first sample object. A target recognition model is obtained by performing a first-stage training based on the first sample pair and a second-stage training based on the second sample pair. The first phase of training is as follows: The first initial recognition model is used to extract keywords from the sample content included in the first sample pair to obtain predicted keywords and the predicted important parameters of the predicted keywords. Based on the differences between the predicted keywords and the sample keywords included in the first sample pair, and the differences between the predicted importance parameters and the sample importance parameters included in the first sample pair, the model parameters of the first initial recognition model are adjusted to obtain the second initial recognition model. The second phase of training is as follows: Based on the second sample pair, a correlation analysis is performed on the feature data of the first sample object and the sample keywords using a third initial recognition model to determine the prediction comprehension difficulty parameter and the prediction correlation parameter of the sample keywords; Based on the difference between the predicted understanding difficulty parameter of the sample keywords and the sample understanding difficulty parameter, and the difference between the predicted relevance parameter of the sample keywords and the sample relevance parameter, the model parameters of the third initial recognition model are adjusted to obtain the fourth initial recognition model. If the first stage of training occurs before the second stage of training, then the third initial recognition model is the second initial recognition model, and the fourth initial recognition model is the target recognition model; if the second stage of training occurs before the first stage of training, then the first initial recognition model is the fourth initial recognition model, and the second initial recognition model is the target recognition model.

7. The method according to claim 6, characterized in that, The method further includes: Obtain the feature data of the first sample object, multiple sample keywords, and the number of times the sample keywords are marked by multiple sample objects; The sample importance parameter of the target sample keyword is determined based on the total number of times the target sample keyword is tagged. The target sample keyword is one of the sample keywords. The total number of times the target sample keyword is tagged is determined based on the number of times the target sample keyword is tagged by multiple sample objects. Based on the total number of times the target sample keyword is marked by the sample objects included in the target sample object set, the sample-related parameters of the target sample keyword are determined. The target sample object set includes the first sample object, and the similarity between the object feature data of the sample objects included in the target sample object set and the feature data of the first sample object is greater than a similarity threshold. The sample comprehension difficulty parameter of the target sample keywords is determined based on the frequency of occurrence of the target sample keywords and the frequency of occurrence of the tagged keywords. Each of the sample keywords is used as the target sample keyword to obtain the sample importance parameter, the sample relevance parameter, and the sample comprehension difficulty parameter corresponding to each of the sample keywords. The first sample pair is determined based on the sample content containing the sample keywords and the sample importance parameters of the sample keywords; The second sample pair is determined based on the first sample object feature data, the sample comprehension difficulty parameter of the sample keywords, and the sample relevance parameter of the sample keywords.

8. A search term recommendation method, characterized in that, The method includes: After logging into the content application through the target account, an input control and a target search term are displayed. The input control is used to search for the input content, and the target search term is determined based on the target account's historical viewing content and the target account's object feature data. If a selection operation is performed for the target search term, the content obtained from the search based on the target search term will be displayed.

9. The method according to claim 8, characterized in that, If a selection operation is performed on the target search term, the content obtained from the search based on the target search term is displayed, including: If a selection operation is obtained for the target search term, the target search term is displayed in the input control; Displays the results obtained from a search based on the target search term.

10. The method according to claim 8, characterized in that, The process of displaying input controls and target search terms after logging into the content application through the target account includes: After logging into the content application through the target account, the input control and pending search terms are displayed. The pending search terms are obtained by supplementing the target search terms.

11. The method according to claim 8, characterized in that, The process of displaying input controls and target search terms after logging into the content application through the target account includes: After logging into the content application through the target account, a search control and recommended content are displayed, wherein the recommended content is content recommended to the target account; In response to a trigger operation on the search control, the input control and the target search term are displayed.

12. The method according to claim 11, characterized in that, After logging into the content application through the target account, the display of search controls and recommended content includes: After logging into the content application through the target account, a recommendation page including the recommended content is displayed; In response to a drag-and-drop operation on the recommended page, the search control and the recommended content are displayed.

13. The method according to claim 11, characterized in that, The method further includes: In response to a selection operation on the recommended content, the detailed content of the recommended content is displayed, including search keywords, which are determined based on the target account's historical viewing content and the target account's object characteristic data; In response to a triggering action for the search keyword, the user is redirected to a search page for the search keyword, the search page including input controls and search results for the search keyword.

14. The method according to claim 11, characterized in that, If the recommended content includes the target recommended content, then the method includes: Obtain multiple recommended content vectors and the historical search sequence of the target account, wherein the historical search sequence includes multiple historical search contents that the target account has searched, sorted by time; Based on the historical search sequence, a virtual interest vector for the target account is obtained through prediction. Calculate the similarity between the virtual interest vector and each of the recommended content vectors; The recommended content corresponding to the recommended content vector that satisfies the similarity condition is determined as supplementary recommended content; Based on the supplementary recommendations, target recommendations for the target account are determined.

15. A search term recommendation device, characterized in that, The device includes: an acquisition unit, a determination unit, a calculation unit, a filtering unit, and a recommendation unit; The acquisition unit is used to acquire the historical viewing content viewed through the target account and the object feature data corresponding to the target account; The determining unit is used to determine multiple candidate keywords from the historical viewing content; The calculation unit is used to determine the search tendency parameters of each candidate keyword based on the object feature data. The search tendency parameters are used to characterize the probability that the target account wants to search for the candidate keywords. The filtering unit is used to determine the candidate keywords corresponding to the search tendency parameters that meet the tendency conditions as the target search terms; The recommendation unit is used to recommend the target search term to the target account.

16. A training device for a recognition model, characterized in that, The device includes: an acquisition unit and a training unit; The acquisition unit is used to acquire a first sample pair and a second sample pair. The first sample pair includes the sample content where the sample keyword is located and the sample importance parameter of the sample keyword. The second sample pair includes the feature data of a first sample object, the sample understanding difficulty parameter of the sample keyword, and the sample relevance parameter of the sample keyword. The sample keyword is the keyword marked by the first sample object during the process of viewing the sample content, as represented by the feature data of the first sample object. The training unit is used to perform a first-stage training based on the first sample pair and a second-stage training based on the second sample pair to obtain a target recognition model. The first phase of training is as follows: The first initial recognition model is used to extract keywords from the sample content included in the first sample pair to obtain predicted keywords and the predicted important parameters of the predicted keywords. Based on the differences between the predicted keywords and the sample keywords included in the first sample pair, and the differences between the predicted importance parameters and the sample importance parameters included in the first sample pair, the model parameters of the first initial recognition model are adjusted to obtain the second initial recognition model. The second phase of training is as follows: Based on the second sample pair, a correlation analysis is performed on the feature data of the first sample object and the sample keywords using a third initial recognition model to determine the prediction comprehension difficulty parameter and the prediction correlation parameter of the sample keywords; Based on the difference between the predicted understanding difficulty parameter of the sample keywords and the sample understanding difficulty parameter, and the difference between the predicted relevance parameter of the sample keywords and the sample relevance parameter, the model parameters of the third initial recognition model are adjusted to obtain the fourth initial recognition model. If the first stage of training occurs before the second stage of training, then the third initial recognition model is the second initial recognition model, and the fourth initial recognition model is the target recognition model; if the second stage of training occurs before the first stage of training, then the first initial recognition model is the fourth initial recognition model, and the second initial recognition model is the target recognition model.

17. A search term recommendation device, characterized in that, The device includes: a first display unit and a second display unit; The first display unit is used to display an input control and a target search term after logging into the content application through the target account. The input control is used to search for the input content, and the target search term is determined based on the target account's historical viewing content and the target account's object feature data. The second display unit is used to display the content obtained by searching based on the target search term if a selection operation for the target search term is obtained.

18. A computer device, characterized in that, The computer device includes a processor and memory: The memory is used to store computer programs and to transfer the computer programs to the processor; The processor is configured to perform the method according to any one of claims 1-5, or the method according to claim 6 or claim 7, or the method according to any one of claims 8-14, based on the computer program.

19. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a computer program for performing the method of any one of claims 1-5, or the method of claim 6 or 7, or the method of any one of claims 8-14.

20. A computer program product comprising a computer program, characterized in that, When it is run on a computer device, it causes the computer device to perform the method of any one of claims 1-5, or the method of claim 6 or 7, or the method of any one of claims 8-14.

Images