Content recall method and device, equipment, storage medium and program product

By using the combination of the dual-tower model and interactive tower in the recommendation system, layered storage and multi-round scoring search, the problem of low recall efficiency in the existing technology is solved, and efficient and accurate recall in large-scale content collections are achieved.

CN118939842APending Publication Date: 2024-11-12SHUXING TECH (BEIJING) CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411069501.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-06
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

It is difficult for existing recommendation systems to improve the efficiency of content recall on the basis of ensuring accuracy during the recall stage, especially when processing tens of millions or even billions of content candidates.

Method used

The trained double tower model is adopted to obtain the evaluation values ​​of the target object feature vector and content feature vector through the object tower and the interactive tower, and to store and multiple rounds of scoring to find the target content, and quickly recall relevant content using the index of the content feature vector.

Benefits of technology

Improve the efficiency of content recall, while ensuring the accuracy of recalled content. Through the combination of layered storage and interactive towers, efficient and accurate recall in large-scale content collections are achieved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118939842A_ABST
    Figure CN118939842A_ABST
Patent Text Reader

Abstract

The embodiment of the invention discloses a content recall method and device, equipment, a storage medium and a program product, and the method comprises the steps: responding to a content recall request, and carrying out the extraction through an object tower in a trained double-tower model, so as to obtain a target object feature vector of a target object; obtaining an index of a content feature vector of each content stored in a content database; traversing the M layers, obtaining evaluation values of candidate content feature vectors and target object feature vectors in the currently traversed storage layer through an interaction tower in the trained twin-tower model, and determining a target content feature vector from the candidate content feature vectors in the currently traversed storage layer according to the evaluation values; and after the traversal is finished, sending the target content corresponding to the determined target content feature vector to a client corresponding to the target object. By adopting the embodiment of the invention, the content recall efficiency can be improved on the basis of ensuring the accuracy of the recalled content.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of computer application technology, and in particular to a content recall method, apparatus, device, storage medium and program product. Background Art

[0002] Recommendation systems are usually divided into two main stages: recall and sorting. The recall stage needs to take into account accuracy, diversity, and efficiency, and quickly find candidate content sets related to user interests through multi-way recall. The candidate content sets for the recall stage may be tens of millions or even hundreds of millions. From the tens of millions or even hundreds of millions of content candidate sets, find the target content sets of thousands or hundreds of thousands that the user may be interested in. Based on this, how to improve the efficiency of content recall while ensuring the accuracy of the recalled content is a problem that needs to be solved urgently. Summary of the invention

[0003] The technical problem to be solved by the embodiments of the present application is to provide a content recall method, apparatus, device, storage medium and program product, which can improve the efficiency of content recall while ensuring the accuracy of the recalled content.

[0004] In a first aspect, an embodiment of the present application provides a content recall method, the method comprising:

[0005] In response to the content recall request, extracting a target object feature vector of the target object using the object tower in the trained dual-tower model;

[0006] Obtaining the index of the content feature vector of each content stored in the content database; the index of any content feature vector is used to indicate the storage layer where the any content feature vector is located, the storage layer includes M layers, and M is an integer greater than 1;

[0007] Traversing the M layers, obtaining evaluation values ​​of the candidate content feature vectors and the target object feature vector in the currently traversed storage layer through the interactive tower in the trained dual-tower model, and determining the target content feature vector from the candidate content feature vectors in the currently traversed storage layer according to the evaluation values; the evaluation values ​​of the target content feature vector and the target object feature vector are greater than the evaluation values ​​of other content feature vectors and the target object feature vector in the candidate content feature vectors contained in the currently traversed storage layer except the target content feature vector; the target content feature vector is determined as the candidate content feature vector in the storage layer to be traversed next time;

[0008] After the traversal is completed, the target content corresponding to the determined target content feature vector is sent to the client corresponding to the target object.

[0009] In a second aspect, an embodiment of the present application provides a content recall device, the device comprising:

[0010] An extraction unit, configured to extract a target object feature vector of a target object by using the object tower in the trained dual-tower model in response to a content recall request;

[0011] An acquisition unit, used to acquire an index of a content feature vector of each content stored in a content database; the index of any content feature vector is used to indicate a storage layer where the content feature vector is located, and the storage layer includes M layers, where M is an integer greater than 1;

[0012] The acquisition unit is further used to traverse the M layers, obtain evaluation values ​​of the candidate content feature vectors and the target object feature vector in the currently traversed storage layer through the interactive tower in the trained dual-tower model, and determine the target content feature vector from the candidate content feature vectors in the currently traversed storage layer according to the evaluation values; the evaluation values ​​of the target content feature vector and the target object feature vector are greater than the evaluation values ​​of other content feature vectors and the target object feature vector in the candidate content feature vectors contained in the currently traversed storage layer except the target content feature vector; the target content feature vector is determined as the candidate content feature vector in the storage layer to be traversed next time;

[0013] The sending unit is used to send the target content corresponding to the determined target content feature vector to the client corresponding to the target object after the traversal is completed.

[0014] In one implementation, the acquisition unit traverses the M layers, acquires evaluation values ​​of the candidate content feature vectors in the currently traversed storage layer and the target object feature vector through the interactive tower in the trained dual-tower model, and determines the target content feature vector from the candidate content feature vectors in the currently traversed storage layer according to the evaluation values, including:

[0015] Obtaining evaluation values ​​of the candidate content feature vectors in the Mth layer and the target object feature vector through the interactive tower, and determining the target content feature vector in the Mth layer from the candidate content feature vectors in the Mth layer according to the evaluation values ​​corresponding to the Mth layer; the evaluation values ​​of the target content feature vector in the Mth layer and the target object feature vector are greater than the evaluation values ​​of other content feature vectors in the candidate content feature vectors included in the Mth layer except the target content feature vector in the Mth layer and the target object feature vector; the target content feature vector in the Mth layer is determined as the candidate content feature vector in the M-1th layer;

[0016] Obtaining evaluation values ​​of the candidate content feature vectors in the M-1th layer and the target object feature vector through the interactive tower, and determining the target content feature vector from the candidate content feature vectors in the M-1th layer according to the evaluation values ​​corresponding to the M-1th layer, until the target content feature vector in the first layer is determined;

[0017] The step of sending the target content corresponding to the determined target content feature vector to the client corresponding to the target object includes:

[0018] The target content corresponding to the target content feature vector in the first layer is sent to the client corresponding to the target object.

[0019] In one embodiment, the index of any content feature vector is also used to indicate a content feature vector that is adjacent to any content feature vector in distance;

[0020] The obtaining unit obtains evaluation values ​​of the candidate content feature vectors in the Mth layer and the target object feature vector through the interaction tower, and determines the target content feature vector in the Mth layer from the candidate content feature vectors in the Mth layer according to the evaluation values ​​corresponding to the Mth layer, including:

[0021] Obtaining, through the interactive tower, evaluation values ​​of the candidate content feature vectors in the Mth layer and the target object feature vector, as well as evaluation values ​​of the content feature vectors and the target object feature vectors that are adjacent to the candidate content feature vectors in the Mth layer and are indicated by the index of the candidate content feature vectors in the Mth layer, and determining the target content feature vector in the Mth layer from the candidate content feature vectors in the Mth layer and the content feature vectors that are adjacent to the candidate content feature vectors in the Mth layer according to the evaluation values ​​corresponding to the Mth layer;

[0022] The target content feature vector in the Mth layer and the content feature vector indicated by the index of the target content feature vector in the Mth layer and adjacent to the target content feature vector in the Mth layer are determined as candidate content feature vectors in the M-1th layer; the content feature vectors included in the Mth layer are from the M-1th layer, and the number of content feature vectors included in the M-1th layer is greater than the number of content feature vectors included in the M layer.

[0023] In one embodiment, the evaluation value includes evaluation values ​​of N dimensions, where N is an integer greater than 1; the evaluation value of any dimension is used to represent the degree of interaction between the target object and the content corresponding to each candidate content feature vector in the any dimension;

[0024] The acquisition unit acquires evaluation values ​​of the candidate content feature vectors in the currently traversed storage layer and the target object feature vector through the interactive tower in the trained dual-tower model, and determines the target content feature vector from the candidate content feature vectors in the currently traversed storage layer according to the evaluation values, including:

[0025] Obtaining evaluation values ​​of N dimensions of the candidate content feature vector and the target object feature vector in the currently traversed storage layer through the interaction tower;

[0026] According to the evaluation values ​​of the N dimensions, evaluation values ​​of the candidate content feature vectors in the currently traversed storage layer and the target object feature vector are determined, and according to the evaluation values, a target content feature vector is determined from the candidate content feature vectors in the currently traversed storage layer.

[0027] In one embodiment, the evaluation values ​​of the N dimensions include one or more of an exposure evaluation value, a duration evaluation value, and an interaction evaluation value;

[0028] The exposure evaluation value is used to indicate exposure information obtained by the target object being exposed to the content corresponding to the candidate content feature vector in the currently traversed storage layer within a preset time period;

[0029] The duration evaluation value is used to indicate the duration information obtained when the content corresponding to the candidate content feature vector in the currently traversed storage layer is displayed on the device corresponding to the target object;

[0030] The interaction evaluation value is used to indicate the interaction information obtained by triggering the target object with respect to the content corresponding to the candidate content feature vector in the currently traversed storage layer.

[0031] In one embodiment, the acquisition unit is further configured to include:

[0032] Acquire sample data, the sample data including training content, training object behavior data and label data, the label data including a preset evaluation value, the label data being used to characterize the degree of interaction between the training object corresponding to the training object behavior data and the training content;

[0033] Calling the dual-tower model to extract features of the training content to obtain a training content feature vector of the training content;

[0034] Extracting features from the training object behavior data to obtain a training object feature vector of the training object behavior data; the training object behavior data includes behavior data generated during the interaction between the training object and the training content;

[0035] Obtaining training evaluation values ​​of the training content feature vector and the training object feature vector through the interactive tower in the dual-tower model;

[0036] The twin-tower model is trained according to the training goal of reducing the difference between the training evaluation value and the preset evaluation value to obtain the trained twin-tower model.

[0037] In one embodiment, the training evaluation value includes evaluation values ​​of N dimensions, where N is an integer greater than 1; the evaluation value of any dimension is used to characterize the degree of interaction between the training object and the training content in any dimension;

[0038] The acquiring unit acquires the training evaluation values ​​of the training content feature vector and the training object feature vector through the interactive tower in the dual-tower model, including:

[0039] Obtaining evaluation values ​​of N dimensions of the training content feature vector and the training object feature vector through the interactive tower in the dual-tower model;

[0040] The training evaluation value is determined according to the evaluation values ​​of the N dimensions.

[0041] In a third aspect, an embodiment of the present application provides a computer device, which includes a memory, a communication interface, and a processor, wherein the memory, the communication interface, and the processor are interconnected; the memory stores a computer program, and the processor calls the computer program stored in the memory to implement the content recall method of the first aspect above.

[0042] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, which stores a computer program, and when the computer program is executed by a processor, the content recall method of the first aspect is implemented.

[0043] In a fifth aspect, an embodiment of the present application provides a computer program product, which includes a computer program stored in a computer storage medium; a processor of a computer device reads the computer program from the computer storage medium, and the processor executes the computer program, so that the computer device executes the content recall method of the first aspect mentioned above.

[0044] In the embodiment of the present application, in response to a content recall request, the target object feature vector of the target object is obtained by extracting the object tower in the trained dual-tower model; the index of the content feature vector of each content stored in the content database is obtained; the index of any content feature vector is used to indicate the storage layer where any content feature vector is located, and the storage layer includes M layers, where M is an integer greater than 1; traversing the M layers, the evaluation values ​​of the candidate content feature vector and the target object feature vector in the currently traversed storage layer are obtained through the interactive tower in the trained dual-tower model, and the target content feature vector is determined from the candidate content feature vector in the currently traversed storage layer according to the evaluation value; after the traversal is completed, the target content corresponding to the determined target content feature vector is sent to the client corresponding to the target object. By storing the content feature vectors of each content in layers, and searching for the recalled content through hierarchical multi-round scoring through the interactive tower, the accuracy of the recalled content can be improved; through the index of the content feature vector of each content, when recalling the content, the target content can be quickly found according to the index, and the efficiency of content recall is improved on the basis of ensuring the accuracy of the recalled content. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the background technology, the drawings required for use in the embodiments of the present application or the background technology will be described below.

[0046] Figure 1 It is a flowchart of a content recall method provided in an embodiment of the present application;

[0047] Figure 2 This is a schematic diagram of hierarchical storage provided by an embodiment of the present application;

[0048] Figure 3 It is a schematic diagram of the architecture of a model training system provided in an embodiment of the present application;

[0049] Figure 4 It is a flowchart of a model training method provided in an embodiment of the present application;

[0050] Figure 5 is a structural diagram of a content recall device provided in an embodiment of the present application;

[0051] Figure 6 It is a structural diagram of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0052] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0053] It should be noted that, in this article, the terms "include", "comprises" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element. In addition, components, features, and elements with the same name in different embodiments of the present application may have the same meaning or different meanings, and their specific meanings need to be determined by their explanation in the specific embodiment or further combined with the context of the specific embodiment.

[0054] At the same time, in the description of the embodiments of the present application, the terms "first", "second", etc. are only used to distinguish the description and cannot be understood as indicating or implying relative importance. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more features. In the description of the embodiments of the present application, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined.

[0055] In the specific implementation of this application, user-related data, such as user behavior data, is involved. When the embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of relevant data must comply with local laws, regulations and standards.

[0056] The content recall method provided in the embodiment of the present application is applied in a content recommendation system, and can recall the content of interest to the object based on the object behavior data, wherein the object can be called a user and the content can be called a note. In the content recommendation system, content recommendation recall can be performed through a dual-tower model, wherein the model structure of the dual-tower model is based on two parallel neural networks (towers), and each tower processes the information of the object and the item (or content) respectively. Object Tower (User Tower): The input of the object tower is information about the object, such as the object identification (Identity document, id), historical behavior, personal data, etc. of the object. The goal of the object tower is to learn the feature representation of the object and map the features of the object to a low-dimensional continuous vector space; Item Tower (Item Tower): The input of the item tower is information about the item (or content), such as the item ID, attributes, text description, etc. The goal of the item tower is to learn the feature representation of the item and map the features of the item to the same low-dimensional continuous vector space. The present application adds an interactive tower on the basis of the dual-tower model. After extracting the content feature vector through the object tower, a hierarchical index of the content feature vector is constructed to obtain the nodes and node connection diagram contained in each layer, and the index and content feature vector are stored in a local content database in a hierarchical manner. After receiving the online request, the server requests the object tower to obtain the target object feature vector, finds the candidate content feature vector through the index of the content feature vector of each content, and then obtains the evaluation value of the candidate content feature vector and the target object feature vector in the currently traversed storage layer through the interactive tower, and determines the target content feature vector according to the evaluation value. After the traversal is completed, the target content corresponding to the determined target content feature vector is sent to the client corresponding to the target object.

[0057] Based on the above description, see Figure 1 , Figure 1 is a flow chart of a content recall method provided by an embodiment of the present application, such as Figure 1 The content recall method shown includes but is not limited to steps S101-S104, wherein:

[0058] S101, in response to a content recall request, extracting a target object feature vector of a target object using an object tower in a trained dual-tower model.

[0059] In this embodiment, the object tower can also be called a user tower, and the behavior data of the target object is extracted by using the object tower in the trained dual-tower model to obtain the object feature vector of the target object. The behavior data of the target object can indicate the data of the interaction between the target object and the content, such as the object's historical clicks, forwarding, comments, and favorite notes, or data such as the object ID and object portrait.

[0060] The object tower in the trained dual-tower model includes a multi-layer network. The behavior data of the target object is passed to the embedding layer of the object tower to map the behavior data of the target object into a representation form in a low-dimensional dense vector space. After the embedding layer, a multilayer perceptron (MLP) is used to further process the embedding vector of the target object. The MLP consists of multiple fully connected layers, each of which contains a set of weight parameters and activation functions for learning the complex relationship between the features of the target object. In the last layer of the MLP, the feature representation of the target object is obtained, and the target object feature vector of the target object is extracted from the feature representation of the target object.

[0061] S102, obtaining the index of the content feature vector of each content stored in the content database; the index of any content feature vector is used to indicate the storage layer where any content feature vector is located, and the storage layer includes M layers, where M is an integer greater than 1.

[0062] In this embodiment, each content may indicate each note, and the content database may indicate the note database, that is, the index of the note feature vector of each note stored in the note database may be obtained. The index may indicate the storage layer in the content database where any content feature vector is located, wherein the storage layer includes M layers, and M is an integer greater than 1. For example, see Figure 2 , Figure 2 This is a hierarchical storage diagram provided in an embodiment of the present application, wherein content feature vectors are stored through three storage layers, i.e., M is 3, two content feature vectors are stored in Layer=3, four content feature vectors are stored in Layer=2, and six content feature vectors are stored in Layer=1, i.e., in Layer 1, wherein Layer 1 indicates the bottom layer.

[0063] In an optional implementation, the index of any content feature vector is used to indicate the storage layer in which any content feature vector is located and the content feature vectors that are adjacent to any content feature vector. After the content feature vectors of each storage layer are determined, the content feature vectors that are adjacent to any content feature vector of each storage layer can be determined, and the content feature vectors that are adjacent to any content feature vector can refer to the content feature vectors that are closest to a preset number of adjacent content feature vectors. There is a correlation between any content feature vector in each layer and the content feature vectors that are adjacent to it. When searching for target content that matches the content feature vector and the target object feature vector, the evaluation values ​​of the currently searched content feature vector and the target object feature vector, as well as the evaluation values ​​of the content feature vectors and the target object feature vector that are adjacent to the currently searched content feature vector, are calculated.

[0064] The content feature vector of each content can be extracted through the content tower in the trained dual-tower model. The content tower refers to the item tower. The content tower in the trained dual-tower model includes a multi-layer network. The content data of each content is passed to the embedding layer of the content tower to map the content data of each content to a representation form in a low-dimensional dense vector space. The content data may include content ID, text description, label and other information. After the embedding layer, MLP is used to further process the embedding vector of each content. MLP consists of multiple fully connected layers, each of which contains a set of weight parameters and activation functions for learning the complex relationship between the features of the target object. In the last layer of MLP, the feature representation of each content is obtained, and the content feature vector of each content is extracted from the feature representation of each content.

[0065] S103, traversing M layers, obtaining evaluation values ​​of the candidate content feature vectors and the target object feature vectors in the currently traversed storage layer through the interactive tower in the trained dual-tower model, and determining the target content feature vector from the candidate content feature vectors in the currently traversed storage layer according to the evaluation values.

[0066] In this embodiment, the storage layer where the content feature vector of each content is located is determined by index, and then the traversal can be started from the first layer to the Mth layer; or the traversal can be started from the Mth layer to the first layer, and the first layer is the bottom layer. The evaluation values ​​of the candidate content feature vector and the target object feature vector in the currently traversed storage layer are obtained through the interactive tower in the trained dual-tower model, wherein the interactive tower can obtain the evaluation values ​​of the candidate content feature vector and the target object feature vector in the currently traversed storage layer through a dot product operation, or the interactive tower includes a multi-layer neural network structure to obtain the evaluation values ​​of the candidate content feature vector and the target object feature vector in the currently traversed storage layer. The evaluation values ​​output by the interactive tower can be sorted from high to low according to the evaluation values, and the content feature vector corresponding to the preset number of evaluation values ​​with the highest scores is determined as the target content feature vector in the currently traversed storage layer. The target content feature vector in the currently traversed storage layer is determined as the candidate content feature vector in the storage layer to be traversed next time, until the traversal ends.

[0067] In an optional implementation, traversing M layers, obtaining evaluation values ​​of the candidate content feature vectors and the target object feature vectors in the currently traversed storage layer through the interactive tower in the trained dual-tower model, and determining the target content feature vector from the candidate content feature vectors in the currently traversed storage layer according to the evaluation values, can be done in the following manner: obtaining evaluation values ​​of the candidate content feature vectors and the target object feature vectors in the Mth layer through the interactive tower, and determining the target content feature vector in the Mth layer from the candidate content feature vectors in the Mth layer according to the evaluation values ​​corresponding to the Mth layer; determining the target content feature vector in the Mth layer The evaluation value of the target object feature vector and the target content feature vector is greater than the evaluation values ​​of other content feature vectors and the target object feature vector in the candidate content feature vectors contained in the Mth layer except the target content feature vector in the Mth layer; the target content feature vector in the Mth layer is determined as the candidate content feature vector in the M-1th layer; the evaluation values ​​of the candidate content feature vector and the target object feature vector in the M-1th layer are obtained through the interaction tower, and the target content feature vector is determined from the candidate content feature vectors in the M-1th layer according to the corresponding evaluation values ​​of the M-1th layer, until the target content feature vector in the first layer is determined.

[0068] In this embodiment, the storage layer where the content feature vector of each content is located is determined by indexing, and then traversing starts from the Mth layer until the first layer, which is the bottom layer. The evaluation values ​​of the candidate content feature vector and the target object feature vector in the Mth layer are obtained through the interactive tower in the trained dual-tower model, wherein the interactive tower can obtain the evaluation values ​​of the candidate content feature vector and the target object feature vector in the currently traversed storage layer through a dot product operation, or the interactive tower includes a multi-layer neural network structure to obtain the evaluation values ​​of the candidate content feature vector and the target object feature vector in the currently traversed storage layer. The evaluation values ​​output by the interactive tower can be sorted from high to low according to the evaluation values, and the content feature vector corresponding to the preset number of evaluation values ​​with the highest scores is determined as the target content feature vector in the Mth layer. The target content feature vector in the Mth layer is determined as the candidate content feature vector in the M-1th layer, and then the evaluation values ​​of the candidate content feature vector and the target object feature vector in the M-1th layer are obtained through the interactive tower, and the target content feature vector is determined from the candidate content feature vectors in the M-1th layer according to the evaluation value corresponding to the M-1th layer, until the target content feature vector in the first layer (the bottom layer) is determined.

[0069] In an optional implementation, the index of any content feature vector is also used to indicate a content feature vector that is adjacent to any content feature vector in distance; obtaining evaluation values ​​of the candidate content feature vectors and the target object feature vector in the Mth layer through the interactive tower, and determining the target content feature vector in the Mth layer from the candidate content feature vectors in the Mth layer according to the evaluation values ​​corresponding to the Mth layer, can be performed in the following manner: obtaining evaluation values ​​of the candidate content feature vectors and the target object feature vector in the Mth layer through the interactive tower, and content feature vectors that are adjacent to the candidate content feature vectors in the Mth layer and the target object feature vector indicated by the index of the candidate content feature vector in the Mth layer An evaluation value of the object feature vector is obtained, and a target content feature vector in the Mth layer is determined from the candidate content feature vectors in the Mth layer and the content feature vectors that are adjacent to the candidate content feature vectors in the Mth layer according to the evaluation value corresponding to the Mth layer; the target content feature vector in the Mth layer and the content feature vectors that are adjacent to the target content feature vector in the Mth layer indicated by the index of the target content feature vector in the Mth layer are determined as candidate content feature vectors in the M-1th layer; the content feature vectors included in the Mth layer are from the M-1th layer, and the number of content feature vectors included in the M-1th layer is greater than the number of content feature vectors included in the Mth layer.

[0070] In this embodiment, after the content feature vector of each storage layer is determined, the content feature vectors adjacent to any content feature vector of each storage layer can be determined, and the content feature vectors adjacent to any content feature vector can refer to the content feature vectors with a preset number of adjacent distances to any content feature vector. There is an association relationship between any content feature vector in each layer and the content feature vectors adjacent to the content feature vectors. When searching for target content that matches the content feature vector and the target object feature vector, the evaluation values ​​of the currently searched content feature vector and the target object feature vector, as well as the evaluation values ​​of the currently searched content feature vector and the target object feature vector adjacent to the content feature vector, are calculated.

[0071] In the process of determining the target content feature vector of the Mth layer, the evaluation values ​​of the candidate content feature vector and the target object feature vector in the Mth layer are obtained through the interactive tower in the Mth layer, as well as the evaluation values ​​of the content feature vector and the target object feature vector that are adjacent to the candidate content feature vector in the Mth layer indicated by the index of the candidate content feature vector in the Mth layer, and the highest k evaluation values ​​are obtained, where k is a positive integer, for example, if k is 100, the 100 highest evaluation values ​​are obtained, and the content feature vectors corresponding to the 100 highest evaluation values ​​are determined as the target content feature vector of the next layer, that is, the M-1th layer. The content feature vectors in the M-1th layer include the content feature vectors in the Mth layer, and the number of content feature vectors in the M-1th layer is greater than the number of content feature vectors in the Mth layer.

[0072] Further, the target content feature vector and the content feature vectors that are adjacent to each target content feature vector are obtained at the M-1 layer. The content feature vectors that are adjacent to each target content feature vector may refer to the content feature vectors whose distance to the target content feature vector is less than a preset distance threshold, or refer to the P content feature vectors that are closest to the target content feature vector, where P is a preset number, for example, if P is 10, it refers to the 10 content feature vectors that are closest to the target content feature vector. By obtaining the target content feature vector at the M-1 layer and the content feature vectors that are adjacent to each target content feature vector, a candidate content feature vector is obtained. The target object feature vector and the candidate content feature vector are evaluated to obtain the k evaluation values ​​with the highest scores, for example, if k is 100, the 100 evaluation values ​​with the highest scores are obtained, and the content feature vectors corresponding to the 100 evaluation values ​​with the highest scores are determined as the target content feature vectors of the next layer. M is reduced by 1, and the operation of selecting the candidate content feature vector from the M-1 layer is triggered until the target content feature vector is determined from the first layer.

[0073] In an optional implementation, evaluation values ​​of candidate content feature vectors and target object feature vectors in the Mth layer are obtained through an interactive tower, and the candidate target content feature vector is determined based on the evaluation values ​​of the candidate content feature vectors and the target object feature vector in the Mth layer; the evaluation values ​​of the candidate target content feature vector and the target object feature vector are greater than the evaluation values ​​of other content feature vectors in the candidate content feature vectors contained in the Mth layer except the candidate target content feature vector in the Mth layer and the target object feature vector; based on the candidate target content feature vector, content feature vectors adjacent to the candidate target content feature vector in distance are determined, and evaluation values ​​of content feature vectors adjacent to the candidate target content feature vector in distance and the target object feature vector are obtained through an interactive tower, and the target content feature vector is determined based on the evaluation value corresponding to the candidate target content feature vector and the evaluation value corresponding to the evaluation value of the content feature vector adjacent to the candidate target content feature vector in distance and the target object feature vector. That is, after obtaining a preset number of candidate target content feature vectors with the highest evaluation values ​​through the interactive tower, the evaluation values ​​of content feature vectors adjacent to the candidate target content feature vectors are obtained through the interactive tower. According to the evaluation values ​​corresponding to the candidate target content feature vectors and the evaluation values ​​of content feature vectors adjacent to the candidate target content feature vectors, the preset number of content feature vectors with the highest evaluation values ​​are taken as the target content feature vectors.

[0074] Among them, the first layer, that is, the bottom layer, contains all the content feature vectors, and for each content feature vector, it is promoted to a higher layer with a certain probability. This probability is determined by random sampling or using probability distribution, wherein not all content feature vectors are promoted to all layers, but the maximum number of layers to which each content feature vector can be promoted is limited to ensure that the hierarchy does not become too deep. This random promotion process is repeated during the construction process until the predetermined conditions are met or a certain number of layers is reached, such as M layers. In the end, each content feature vector may appear at some levels from the 1st layer to the Mth layer. This random promotion strategy helps to maintain the randomness of the distribution of content feature vectors in the hierarchy, enabling efficient neighbor search in high-dimensional space.

[0075] Optionally, when obtaining the evaluation values ​​of the candidate content feature vectors and the target object feature vectors in the Mth layer through the interactive tower, the calculation process may be accelerated by a graphics processing unit (GPU).

[0076] In an optional implementation, the Hierarchical Navigable Small World (HNSW) algorithm is called to determine the target content feature vector from each content feature vector. HNSW is an algorithm for fast approximate nearest neighbor search. Among them, the content feature vectors are stored in a high-dimensional space. The HNSW algorithm is based on the "small world" idea, that is, in a high-dimensional space, the relationship between the content feature vectors is formed into a special structure, and efficient search can be performed through a series of locally connected content feature vectors. This structure enables the rapid location of related content feature vectors in the approximate nearest neighbor search task, thereby improving retrieval efficiency. HNSW adopts a hierarchical structure, in which the content feature vectors are distributed on multiple levels, and each content feature vector has one or more links connected to adjacent layers, so that the search can be carried out in the hierarchical structure, which is more efficient. HNSW uses a special navigation strategy to find potential nearest neighbors by jumping and local search. This navigation strategy makes full use of the small world structure to improve search efficiency.

[0077] In an optional embodiment, the evaluation value includes evaluation values ​​of N dimensions, where N is an integer greater than 1; the evaluation value of any dimension is used to characterize the degree of interaction between the target object and the content corresponding to each candidate content feature vector in any dimension; the evaluation values ​​of the candidate content feature vectors and the target object feature vectors in the currently traversed storage layer are obtained through the interactive tower in the trained dual-tower model, and the target content feature vector is determined from the candidate content feature vectors in the currently traversed storage layer according to the evaluation values. This can be done in the following manner: the evaluation values ​​of the N dimensions of the candidate content feature vectors and the target object feature vectors in the currently traversed storage layer are obtained through the interactive tower; the evaluation values ​​of the candidate content feature vectors and the target object feature vectors in the currently traversed storage layer are determined according to the evaluation values ​​of the N dimensions, and the target content feature vector is determined from the candidate content feature vectors in the currently traversed storage layer according to the evaluation values.

[0078] In this embodiment, the evaluation value may include evaluation values ​​of multiple dimensions, such as exposure evaluation value, duration evaluation value, interaction evaluation value, preference evaluation value, etc. For the evaluation value of each dimension, the evaluation value of any dimension is obtained based on the target object feature vector of any dimension and each content feature vector in the Mth layer. For example, for the interaction evaluation value, the more times the target object interacts with the content, the higher the evaluation value may be. After obtaining the evaluation values ​​of N dimensions, the evaluation values ​​of the N dimensions may be weighted and summed to determine the evaluation values ​​of the candidate content feature vector and the target object feature vector in the currently traversed storage layer, and the target content feature vector is determined from the candidate content feature vectors in the currently traversed storage layer according to the evaluation value.

[0079] Optionally, the object tower may include multiple sub-towers, each sub-tower corresponds to a dimension, that is, corresponds to a target, and the characteristic representation of the behavior data of the target object is obtained through each sub-tower to obtain the target object feature vector of the target object of the dimension corresponding to each sub-tower.

[0080] In an optional embodiment, the evaluation values ​​of the candidate content feature vector and the target object feature vector are calculated in the following manner: after obtaining the target object feature vector of any dimension, the evaluation value of any dimension is obtained based on the target object feature vector and the candidate content feature vector of any dimension, and the evaluation value of any dimension is used to characterize the degree of interaction between the target object and the content corresponding to the candidate content feature vector in any dimension; after obtaining the evaluation values ​​of N dimensions, the evaluation values ​​of the target object feature vector and the candidate content feature vector are obtained based on the evaluation values ​​of the N dimensions.

[0081] In an optional embodiment, the evaluation values ​​of N dimensions include one or more of an exposure evaluation value, a duration evaluation value and an interaction evaluation value; wherein the exposure evaluation value is used to indicate exposure information obtained by the target object exposing the content corresponding to the candidate content feature vector in the currently traversed storage layer within a preset time period; the duration evaluation value is used to indicate duration information obtained by the time the content corresponding to the candidate content feature vector in the currently traversed storage layer is displayed on the device corresponding to the target object; the interaction evaluation value is used to indicate interaction information obtained by the target object triggering the content corresponding to the candidate content feature vector in the currently traversed storage layer.

[0082] In this embodiment, the evaluation values ​​of the N dimensions include one or more of an exposure evaluation value, a duration evaluation value, and an interaction evaluation value. The exposure evaluation value is used to indicate the evaluation value corresponding to the exposure information obtained by the target object when each content in the Mth layer is exposed within a preset time period. For example, the preset time period is 7 days, and the exposure information is the number of exposures or the exposure frequency, etc., that is, the exposure evaluation value represents the score (evaluation value) of the number of exposures or the exposure frequency of the content exposed to the target object in 7 days; the duration evaluation value is used to indicate the evaluation value corresponding to the duration information obtained by the time when each content in the Mth layer is displayed on the device corresponding to the target object, for example, the duration evaluation value represents the evaluation value corresponding to the duration information obtained by the time when the target object views the content through the device displaying the content; the interaction evaluation value is used to indicate the evaluation value corresponding to the interaction information obtained by the target object when each content in the Mth layer is triggered. For example, the interaction information can indicate the evaluation value corresponding to the object's like information, comment information, forwarding information, etc. for the content.

[0083] S104: After the traversal is completed, the target content corresponding to the determined target content feature vector is sent to the client corresponding to the target object.

[0084] In an optional implementation, the target content corresponding to the determined target content feature vector is sent to the client corresponding to the target object in the following manner: the target content corresponding to the target content feature vector in the first layer is sent to the client corresponding to the target object.

[0085] Optionally, the target content (recalled content) is sorted according to the evaluation value calculated based on the content feature vector corresponding to the target content and the target object feature vector. The target content can be sorted from high to low according to the evaluation value, and the sorted target content is sent to the client corresponding to the target object.

[0086] Optionally, after determining the target content of the target object, the similarity between the target object and the target content may be calculated, and the content with a similarity greater than a preset similarity threshold may be screened out as the target content that is ultimately sent to the client corresponding to the target object.

[0087] Optionally, if the target content is not sorted, the client can sort the target content after receiving the target content. The target content can be sorted according to the evaluation value calculated from the content feature vector corresponding to the target content and the target object feature vector, and the target content can be sorted from high to low according to the evaluation value, and the sorted target content can be recommended and displayed.

[0088] In the embodiment of the present application, in response to a content recall request, the target object feature vector of the target object is extracted using the object tower in the trained dual-tower model; the index of the content feature vector of each content stored in the content database is obtained; the index of any content feature vector is used to indicate the storage layer where any content feature vector is located, and the storage layer includes M layers, where M is an integer greater than 1; the M layers are traversed, and the evaluation values ​​of the candidate content feature vector and the target object feature vector in the currently traversed storage layer are obtained through the interactive tower in the trained dual-tower model, and the target content feature vector is determined from the candidate content feature vector in the currently traversed storage layer according to the evaluation value; after the traversal is completed, the target content corresponding to the determined target content feature vector is sent to the client corresponding to the target object. By upgrading the dual-tower structure to an interactive structure, the defect of the lack of interaction of the dual-tower structure is overcome, and the expressiveness and accuracy of the dual-tower model are improved. During online estimation, the object feature vector is calculated only once for each request, and the content feature vector is accessed locally. By combining the hierarchical storage index of the content feature vector and performing multiple rounds of scoring iterations during retrieval, the complexity increase can be controlled within a reasonable range, thereby improving the efficiency of content recall while ensuring the accuracy of the recalled content.

[0089] Based on the above description, see Figure 3 , Figure 3 It is a schematic diagram of the architecture of a model training system provided by an embodiment of the present application. Exemplarily, sample data is obtained, the sample data includes training content, training object behavior data and label data, the label data includes a preset evaluation value, and the label data is used to characterize the degree of interaction between the training object corresponding to the training object behavior data and the training content; the content tower in the double-tower model is called to extract features of the training content to obtain a training content feature vector of the training content; the object tower in the double-tower model is called to extract features of the training object behavior data to obtain a training object feature vector of the training object behavior data; the training object behavior data includes the behavior data generated during the interaction between the training object and the training content; the training evaluation value of the training content feature vector and the training object feature vector is obtained through the interaction tower in the double-tower model; the double-tower model is trained according to reducing the difference between the training evaluation value and the preset evaluation value as the training goal, that is, the double-tower model is trained according to reducing the loss value as the training goal, and the trained double-tower model is obtained.

[0090] Based on the above description, see Figure 4 , Figure 4 is a flow chart of a model training method provided in an embodiment of the present application, such as Figure 4 The model training method shown includes but is not limited to steps S401-S405, wherein:

[0091] S401, obtaining sample data, the sample data including training content, training object behavior data and label data, the label data including a preset evaluation value, and the label data is used to characterize the degree of interaction between the training object corresponding to the training object behavior data and the training content.

[0092] In this embodiment, the training object behavior data may indicate the object data used to train the twin tower model. The object data may include object id, object portrait, object historical behavior and other information. The object historical behavior includes the object's clicks, forwarding, comments and collections of content. The training content may indicate the content used for training. The training content may include content id, content type, content keywords, content category information, etc. The preset evaluation value may be set with different score values ​​according to the degree of interaction between the object and the content. For example, if the degree of interaction between the object and the content is weak, the preset evaluation value may be set to 60 points. If the degree of interaction between the object and the content is strong, the preset evaluation value may be set to 80 points. For example, if the object clicks on the content and does not have further interaction with the content, it may be determined that the degree of interaction between the object and the content is strong.

[0093] S402, calling the dual-tower model to extract features of the training content, and obtaining a training content feature vector of the training content.

[0094] Among them, the training content can be feature extracted through the content tower of the dual-tower model to obtain a training content feature vector of the training content. For example, the data of the training content is passed to the embedding layer to obtain the corresponding embedding vector. The embedding vector is processed through the MLP or other network layers to extract representative content features. Finally, the training content feature vector of the training content is obtained from the output layer of the network or a specific pooling layer (such as average pooling or maximum pooling).

[0095] S403, extracting features from the training object behavior data to obtain a training object feature vector of the training object behavior data; the training object behavior data includes behavior data generated during the interaction between the training object and the training content.

[0096] In this embodiment, the training object behavior data is feature extracted through the object tower of the dual-tower model to obtain the training object feature vector of the training object behavior data. The training object behavior data may indicate the object data used to train the dual-tower model. The object data may include information such as object ID, object portrait, object historical behavior, etc. The object historical behavior includes the object's clicks, forwarding, comments and collections of notes. For example, the training object behavior data is passed to the embedding layer in the object tower to obtain its corresponding embedding vector, and the object embedding vector is further processed by MLP to capture the complex features of the training object, which involves multiple fully connected layers, each with weights and activation functions. Finally, the training object feature vector of the training object is obtained from the output layer of the network or a specific pooling layer (such as average pooling or maximum pooling).

[0097] S404, obtaining training evaluation values ​​of the training content feature vector and the training object feature vector through the interactive tower in the dual-tower model.

[0098] In this embodiment, after the training content feature vector and the training object feature vector are extracted through the object tower and the content tower of the dual-tower model, the training content feature vector and the training object feature vector can be spliced ​​together, and the training evaluation value of the spliced ​​training content feature vector and the training object feature vector can be obtained through the interaction tower in the dual-tower model, wherein the interaction tower obtains the target evaluation value according to the degree of interaction between the training content feature vector and the training object feature vector, for example, the target evaluation value is 85 points.

[0099] In an optional embodiment, the training evaluation value includes evaluation values ​​of N dimensions, where N is an integer greater than 1; the evaluation value of any dimension is used to characterize the degree of interaction between the training object and the training content in any dimension; the training evaluation value of the training content feature vector and the training object feature vector is obtained through the interactive tower in the dual-tower model, which can be done in the following way: the evaluation values ​​of N dimensions of the training content feature vector and the training object feature vector are obtained through the interactive tower in the dual-tower model; and the training evaluation value is determined based on the evaluation values ​​of the N dimensions.

[0100] In this embodiment, the training object feature vector used to train the dual-tower model may include N-dimensional training object feature vectors, such as exposure feature vectors, duration feature vectors, interaction feature vectors, preference feature vectors, etc. According to the number of feature vector dimensions, for example, N=4-dimensional feature vectors, 4 sub-towers may be designed in the interaction tower, each sub-tower corresponding to a feature vector of one dimension, and each sub-tower may include a deep neural network. For example, the interaction tower corresponding to the exposure feature vector calculates the evaluation values ​​of the user exposure feature vector and the training content feature vector, and the remaining three sub-towers obtain the evaluation values ​​of the corresponding dimensions in the same way. According to the preset weights, the training evaluation values ​​of each dimension are summed up according to the weights to obtain the training evaluation value. For example, the weight corresponding to the exposure feature vector is 20%, the weight corresponding to the duration feature vector is 20%, the weight corresponding to the interaction feature vector is 30%, and the weight corresponding to the preference feature vector is 30%. The interaction tower calculates the evaluation value corresponding to the exposure feature vector as 80 points, the evaluation value corresponding to the duration feature vector is 85 points, the evaluation value corresponding to the interaction feature vector is 90 points, and the evaluation value corresponding to the preference feature vector is 80 points according to the degree of interaction between the object and the content. The final training evaluation value is 84 points. Through feature vectors of multiple dimensions, each indicator can be taken into account, avoiding the loss of accuracy caused by single-dimensional features.

[0101] In an optional embodiment, the evaluation values ​​of N dimensions include one or more of an exposure evaluation value, a duration evaluation value and an interaction evaluation value; wherein the exposure evaluation value is used to indicate the exposure information obtained by the training object when the content corresponding to each training content feature vector is exposed within a preset time period; the duration evaluation value is used to indicate the duration information obtained by the time when the content corresponding to each training content feature vector is displayed on the device corresponding to the training object; the interaction evaluation value is used to indicate the interaction information obtained by the training object when the content corresponding to each training content feature vector is triggered.

[0102] In this embodiment, the training evaluation values ​​of N dimensions include one or more of exposure evaluation value, duration evaluation value and interaction evaluation value. The exposure evaluation value is used to indicate the evaluation value corresponding to the exposure information obtained by the training object when each training content is exposed within a preset time period. For example, the preset time period is 7 days, and the exposure information is the number of exposures or exposure frequency, etc., that is, the exposure evaluation value represents the score (evaluation value) of the number of exposures or exposure frequency of the content exposed to the training object in 7 days; the duration evaluation value is used to indicate the evaluation value corresponding to the duration information obtained by the time each training content is displayed on the device corresponding to the training object, for example, the duration evaluation value represents the evaluation value corresponding to the duration information obtained by the time the training object watches the content through the device displaying the content; the interaction evaluation value is used to indicate the evaluation value corresponding to the interaction information obtained by the training object when each training content is triggered. For example, the interaction information can indicate the evaluation value corresponding to the like information, comment information, forwarding information, etc. of the training object for the content.

[0103] S405, training the twin-tower model according to reducing the difference between the training evaluation value and the preset evaluation value as a training goal to obtain a trained twin-tower model.

[0104] In this embodiment, a plurality of preset evaluation values ​​are designed according to the degree of interaction between the training object and the training content, for example, the preset evaluation value of the weak interaction degree is 70 points; the preset evaluation value of the relatively weak interaction degree is 75 points, for example, the training object only clicks on the training content; the preset evaluation value of the relatively strong interaction degree is 85 points, for example, the training object clicks and likes the training content; the preset evaluation value of the strong interaction degree is 90 points, for example, the training object clicks, likes and forwards the training content. The twin tower model obtains a training evaluation value according to the training object feature vector and the training content feature vector, for example, 82 points. The label data, that is, the preset evaluation value for the training object and the training content is 85 points, that is, the degree of interaction is relatively strong, and the difference between the training evaluation value and the preset evaluation value is 3 points; by obtaining the training evaluation values ​​and preset evaluation values ​​of multiple training objects and training contents, calculating the mean square error of the training evaluation value and the preset evaluation value as the loss value, the twin tower model is trained in the direction of reducing the loss value, and the trained twin tower model is obtained.

[0105] In an embodiment of the present application, sample data is obtained, the sample data includes training content, training object behavior data and label data, the label data includes a preset evaluation value, and the label data is used to characterize the degree of interaction between the training object corresponding to the training object behavior data and the training content; the double-tower model is called to extract features of the training content to obtain a training content feature vector of the training content; the training object behavior data is extracted to obtain a training object feature vector of the training object behavior data; the training object behavior data includes the behavior data generated during the interaction between the training object and the training content; the training evaluation value of the training content feature vector and the training object feature vector is obtained through the interactive tower in the double-tower model; the double-tower model is trained according to the training goal of reducing the difference between the training evaluation value and the preset evaluation value, and a trained double-tower model is obtained. By upgrading the double-tower structure to an interactive structure, that is, adding an interactive tower to the double-tower model, the defect of the double-tower structure lacking interaction is overcome, and the expressiveness and accuracy of the double-tower model are improved.

[0106] An embodiment of the present application further provides a computer storage medium, in which program instructions are stored. When the program instructions are executed, they are used to implement the corresponding methods described in the above embodiments.

[0107] See also Figure 5 , Figure 5 It is a structural diagram of a content recall device provided in an embodiment of the present application.

[0108] In one implementation of the content recall device of the embodiment of the present application, the content recall device includes the following structure:

[0109] An extraction unit 501 is used to extract a target object feature vector of a target object using the object tower in the trained dual-tower model in response to a content recall request;

[0110] The acquisition unit 502 is used to acquire the index of the content feature vector of each content stored in the content database; the index of any content feature vector is used to indicate the storage layer where the any content feature vector is located, and the storage layer includes M layers, where M is an integer greater than 1;

[0111] The acquisition unit 502 is further used to traverse the M layers, obtain evaluation values ​​of the candidate content feature vectors and the target object feature vector in the currently traversed storage layer through the interactive tower in the trained dual-tower model, and determine the target content feature vector from the candidate content feature vectors in the currently traversed storage layer according to the evaluation values; the evaluation values ​​of the target content feature vector and the target object feature vector are greater than the evaluation values ​​of other content feature vectors and the target object feature vector in the candidate content feature vectors contained in the currently traversed storage layer except the target content feature vector; the target content feature vector is determined as the candidate content feature vector in the storage layer to be traversed next time;

[0112] The sending unit 503 is used to send the target content corresponding to the determined target content feature vector to the client corresponding to the target object after the traversal is completed.

[0113] In one implementation, the acquisition unit 502 traverses the M layers, acquires evaluation values ​​of the candidate content feature vectors in the currently traversed storage layer and the target object feature vector through the interactive tower in the trained dual-tower model, and determines the target content feature vector from the candidate content feature vectors in the currently traversed storage layer according to the evaluation values, including:

[0114] Obtaining evaluation values ​​of the candidate content feature vectors in the Mth layer and the target object feature vector through the interactive tower, and determining the target content feature vector in the Mth layer from the candidate content feature vectors in the Mth layer according to the evaluation values ​​corresponding to the Mth layer; the evaluation values ​​of the target content feature vector in the Mth layer and the target object feature vector are greater than the evaluation values ​​of other content feature vectors in the candidate content feature vectors included in the Mth layer except the target content feature vector in the Mth layer and the target object feature vector; the target content feature vector in the Mth layer is determined as the candidate content feature vector in the M-1th layer;

[0115] Obtaining evaluation values ​​of the candidate content feature vectors in the M-1th layer and the target object feature vector through the interactive tower, and determining the target content feature vector from the candidate content feature vectors in the M-1th layer according to the evaluation values ​​corresponding to the M-1th layer, until the target content feature vector in the first layer is determined;

[0116] The step of sending the target content corresponding to the determined target content feature vector to the client corresponding to the target object includes:

[0117] The target content corresponding to the target content feature vector in the first layer is sent to the client corresponding to the target object.

[0118] In one embodiment, the index of any content feature vector is also used to indicate a content feature vector that is adjacent to any content feature vector in distance;

[0119] The acquiring unit 502 acquires evaluation values ​​of the candidate content feature vectors in the Mth layer and the target object feature vector through the interaction tower, and determines the target content feature vector in the Mth layer from the candidate content feature vectors in the Mth layer according to the evaluation values ​​corresponding to the Mth layer, including:

[0120] Obtaining, through the interactive tower, evaluation values ​​of the candidate content feature vectors in the Mth layer and the target object feature vector, as well as evaluation values ​​of the content feature vectors and the target object feature vectors that are adjacent to the candidate content feature vectors in the Mth layer and are indicated by the index of the candidate content feature vectors in the Mth layer, and determining the target content feature vector in the Mth layer from the candidate content feature vectors in the Mth layer and the content feature vectors that are adjacent to the candidate content feature vectors in the Mth layer according to the evaluation values ​​corresponding to the Mth layer;

[0121] The target content feature vector in the Mth layer and the content feature vector indicated by the index of the target content feature vector in the Mth layer and adjacent to the target content feature vector in the Mth layer are determined as candidate content feature vectors in the M-1th layer; the content feature vectors included in the Mth layer are from the M-1th layer, and the number of content feature vectors included in the M-1th layer is greater than the number of content feature vectors included in the M layer.

[0122] In one embodiment, the evaluation value includes evaluation values ​​of N dimensions, where N is an integer greater than 1; the evaluation value of any dimension is used to represent the degree of interaction between the target object and the content corresponding to each candidate content feature vector in the any dimension;

[0123] The acquisition unit 502 acquires evaluation values ​​of the candidate content feature vectors in the currently traversed storage layer and the target object feature vector through the interactive tower in the trained dual-tower model, and determines the target content feature vector from the candidate content feature vectors in the currently traversed storage layer according to the evaluation values, including:

[0124] Obtaining evaluation values ​​of N dimensions of the candidate content feature vector and the target object feature vector in the currently traversed storage layer through the interaction tower;

[0125] According to the evaluation values ​​of the N dimensions, evaluation values ​​of the candidate content feature vectors in the currently traversed storage layer and the target object feature vector are determined, and according to the evaluation values, a target content feature vector is determined from the candidate content feature vectors in the currently traversed storage layer.

[0126] In one embodiment, the evaluation values ​​of the N dimensions include one or more of an exposure evaluation value, a duration evaluation value, and an interaction evaluation value;

[0127] The exposure evaluation value is used to indicate exposure information obtained by the target object being exposed to the content corresponding to the candidate content feature vector in the currently traversed storage layer within a preset time period;

[0128] The duration evaluation value is used to indicate the duration information obtained when the content corresponding to the candidate content feature vector in the currently traversed storage layer is displayed on the device corresponding to the target object;

[0129] The interaction evaluation value is used to indicate the interaction information obtained by triggering the target object with respect to the content corresponding to the candidate content feature vector in the currently traversed storage layer.

[0130] In one implementation, the acquiring unit 502 is further configured to include:

[0131] Acquire sample data, the sample data including training content, training object behavior data and label data, the label data including a preset evaluation value, the label data being used to characterize the degree of interaction between the training object corresponding to the training object behavior data and the training content;

[0132] Calling the dual-tower model to extract features of the training content to obtain a training content feature vector of the training content;

[0133] Extracting features from the training object behavior data to obtain a training object feature vector of the training object behavior data; the training object behavior data includes behavior data generated during the interaction between the training object and the training content;

[0134] Obtaining training evaluation values ​​of the training content feature vector and the training object feature vector through the interactive tower in the dual-tower model;

[0135] The twin-tower model is trained according to the training goal of reducing the difference between the training evaluation value and the preset evaluation value to obtain the trained twin-tower model.

[0136] In one embodiment, the training evaluation value includes evaluation values ​​of N dimensions, where N is an integer greater than 1; the evaluation value of any dimension is used to characterize the degree of interaction between the training object and the training content in any dimension;

[0137] The acquiring unit 502 acquires the training evaluation values ​​of the training content feature vector and the training object feature vector through the interactive tower in the dual-tower model, including:

[0138] Obtaining evaluation values ​​of N dimensions of the training content feature vector and the training object feature vector through the interactive tower in the dual-tower model;

[0139] The training evaluation value is determined according to the evaluation values ​​of the N dimensions.

[0140] In the embodiment of the present application, the extraction unit 501 responds to the content recall request and obtains the target object feature vector of the target object by using the object tower in the trained dual-tower model; the acquisition unit 502 obtains the index of the content feature vector of each content stored in the content database; the index of any content feature vector is used to indicate the storage layer where any content feature vector is located, and the storage layer includes M layers, where M is an integer greater than 1; traverse the M layers, obtain the evaluation values ​​of the candidate content feature vector and the target object feature vector in the currently traversed storage layer through the interactive tower in the trained dual-tower model, and determine the target content feature vector from the candidate content feature vector in the currently traversed storage layer according to the evaluation value; after the traversal, the sending unit 503 sends the target content corresponding to the determined target content feature vector to the client corresponding to the target object. By storing the content feature vectors of each content in layers and searching for the recalled content by hierarchical multi-round scoring through the interactive tower, the accuracy of the recalled content can be improved; through the index of the content feature vector of each content, when recalling the content, the target content can be quickly found according to the index, and the efficiency of content recall can be improved on the basis of ensuring the accuracy of the recalled content.

[0141] See also Figure 6 , Figure 6 6 is a schematic diagram of the structure of a computer device provided in an embodiment of the present application. The computer device in the embodiment of the present application includes a power supply module and other structures, and includes a processor 601, a memory 602, and a communication interface 603. The processor 601, the memory 602, and the communication interface 603 can exchange data, and the processor 601 implements the corresponding content recall method.

[0142] The memory 602 may include a volatile memory, such as a random-access memory (RAM); the memory 602 may also include a non-volatile memory, such as a flash memory, a solid-state drive (SSD), etc.; the memory 602 may also include a combination of the above-mentioned types of memory.

[0143] The processor 601 may be a central processing unit (CPU). The processor 601 may also be a combination of a CPU and a GPU. In a computer device, multiple CPUs and GPUs may be included as needed to perform corresponding content recall. In one embodiment, the memory 602 is used to store program instructions. The processor 601 may call program instructions to implement various methods involved in the embodiments of the present application.

[0144] In a first possible implementation, the processor 601 of the computer device calls the program instructions stored in the memory 602, and is used to respond to the content recall request, extract the target object feature vector of the target object by using the object tower in the trained dual-tower model; obtain the index of the content feature vector of each content stored in the content database; the index of any content feature vector is used to indicate the storage layer where the any content feature vector is located, and the storage layer includes M layers, where M is an integer greater than 1; traverse the M layers, and obtain the candidate content feature vector in the currently traversed storage layer through the interactive tower in the trained dual-tower model and the target object feature vector. An evaluation value of a feature vector is obtained, and a target content feature vector is determined from the candidate content feature vectors in the currently traversed storage layer according to the evaluation value; the evaluation values ​​of the target content feature vector and the target object feature vector are greater than the evaluation values ​​of other content feature vectors other than the target content feature vector in the candidate content feature vectors contained in the currently traversed storage layer and the target object feature vector; the target content feature vector is determined as the candidate content feature vector in the storage layer to be traversed next time; after the traversal is completed, the target content corresponding to the determined target content feature vector is sent to the client corresponding to the target object.

[0145] In one implementation, the processor 601 traverses the M layers, obtains evaluation values ​​of the candidate content feature vectors in the currently traversed storage layer and the target object feature vector through the interactive tower in the trained dual-tower model, and determines the target content feature vector from the candidate content feature vectors in the currently traversed storage layer according to the evaluation values, and may perform the following operations:

[0146] Obtaining evaluation values ​​of the candidate content feature vectors in the Mth layer and the target object feature vector through the interactive tower, and determining the target content feature vector in the Mth layer from the candidate content feature vectors in the Mth layer according to the evaluation values ​​corresponding to the Mth layer; the evaluation values ​​of the target content feature vector in the Mth layer and the target object feature vector are greater than the evaluation values ​​of other content feature vectors in the candidate content feature vectors included in the Mth layer except the target content feature vector in the Mth layer and the target object feature vector; the target content feature vector in the Mth layer is determined as the candidate content feature vector in the M-1th layer;

[0147] Obtaining evaluation values ​​of the candidate content feature vectors in the M-1th layer and the target object feature vector through the interactive tower, and determining the target content feature vector from the candidate content feature vectors in the M-1th layer according to the evaluation values ​​corresponding to the M-1th layer, until the target content feature vector in the first layer is determined;

[0148] The step of sending the target content corresponding to the determined target content feature vector to the client corresponding to the target object includes:

[0149] The target content corresponding to the target content feature vector in the first layer is sent to the client corresponding to the target object.

[0150] In one embodiment, the index of any content feature vector is also used to indicate a content feature vector that is adjacent to any content feature vector in distance;

[0151] The processor 601 obtains evaluation values ​​of the candidate content feature vectors in the Mth layer and the target object feature vector through the interaction tower, and determines the target content feature vector in the Mth layer from the candidate content feature vectors in the Mth layer according to the evaluation values ​​corresponding to the Mth layer, and may perform the following operations:

[0152] Obtaining, through the interactive tower, evaluation values ​​of the candidate content feature vectors in the Mth layer and the target object feature vector, as well as evaluation values ​​of the content feature vectors and the target object feature vectors that are adjacent to the candidate content feature vectors in the Mth layer and are indicated by the index of the candidate content feature vectors in the Mth layer, and determining the target content feature vector in the Mth layer from the candidate content feature vectors in the Mth layer and the content feature vectors that are adjacent to the candidate content feature vectors in the Mth layer according to the evaluation values ​​corresponding to the Mth layer;

[0153] The target content feature vector in the Mth layer and the content feature vector indicated by the index of the target content feature vector in the Mth layer and adjacent to the target content feature vector in the Mth layer are determined as candidate content feature vectors in the M-1th layer; the content feature vectors included in the Mth layer are from the M-1th layer, and the number of content feature vectors included in the M-1th layer is greater than the number of content feature vectors included in the M layer.

[0154] In one embodiment, the evaluation value includes evaluation values ​​of N dimensions, where N is an integer greater than 1; the evaluation value of any dimension is used to represent the degree of interaction between the target object and the content corresponding to each candidate content feature vector in the any dimension;

[0155] The processor 601 obtains evaluation values ​​of the candidate content feature vectors in the currently traversed storage layer and the target object feature vector through the interactive tower in the trained dual-tower model, and determines the target content feature vector from the candidate content feature vectors in the currently traversed storage layer according to the evaluation values, and may perform the following operations:

[0156] Obtaining evaluation values ​​of N dimensions of the candidate content feature vector and the target object feature vector in the currently traversed storage layer through the interaction tower;

[0157] According to the evaluation values ​​of the N dimensions, evaluation values ​​of the candidate content feature vectors in the currently traversed storage layer and the target object feature vector are determined, and according to the evaluation values, a target content feature vector is determined from the candidate content feature vectors in the currently traversed storage layer.

[0158] In one embodiment, the evaluation values ​​of the N dimensions include one or more of an exposure evaluation value, a duration evaluation value, and an interaction evaluation value;

[0159] The exposure evaluation value is used to indicate exposure information obtained by the target object being exposed to the content corresponding to the candidate content feature vector in the currently traversed storage layer within a preset time period;

[0160] The duration evaluation value is used to indicate the duration information obtained when the content corresponding to the candidate content feature vector in the currently traversed storage layer is displayed on the device corresponding to the target object;

[0161] The interaction evaluation value is used to indicate the interaction information obtained by triggering the target object with respect to the content corresponding to the candidate content feature vector in the currently traversed storage layer.

[0162] In one implementation, the processor 601 may further perform the following operations:

[0163] Acquire sample data, the sample data including training content, training object behavior data and label data, the label data including a preset evaluation value, the label data being used to characterize the degree of interaction between the training object corresponding to the training object behavior data and the training content;

[0164] Calling the dual-tower model to extract features of the training content to obtain a training content feature vector of the training content;

[0165] Extracting features from the training object behavior data to obtain a training object feature vector of the training object behavior data; the training object behavior data includes behavior data generated during the interaction between the training object and the training content;

[0166] Obtaining training evaluation values ​​of the training content feature vector and the training object feature vector through the interactive tower in the dual-tower model;

[0167] The twin-tower model is trained according to the training goal of reducing the difference between the training evaluation value and the preset evaluation value to obtain the trained twin-tower model.

[0168] In one embodiment, the training evaluation value includes evaluation values ​​of N dimensions, where N is an integer greater than 1; the evaluation value of any dimension is used to characterize the degree of interaction between the training object and the training content in any dimension;

[0169] The processor 601 obtains the training evaluation values ​​of the training content feature vector and the training object feature vector through the interactive tower in the dual-tower model, and may perform the following operations:

[0170] Obtaining evaluation values ​​of N dimensions of the training content feature vector and the training object feature vector through the interactive tower in the dual-tower model;

[0171] The training evaluation value is determined according to the evaluation values ​​of the N dimensions.

[0172] In the embodiment of the present application, the processor 601 responds to the content recall request, and uses the object tower in the trained dual-tower model to extract the target object feature vector of the target object; obtains the index of the content feature vector of each content stored in the content database; the index of any content feature vector is used to indicate the storage layer where any content feature vector is located, and the storage layer includes M layers, where M is an integer greater than 1; traverses the M layers, obtains the evaluation values ​​of the candidate content feature vector and the target object feature vector in the currently traversed storage layer through the interactive tower in the trained dual-tower model, and determines the target content feature vector from the candidate content feature vector in the currently traversed storage layer according to the evaluation value; after the traversal is completed, the target content corresponding to the determined target content feature vector is sent to the client corresponding to the target object. By storing the content feature vectors of each content in layers, and searching for the recalled content through hierarchical multi-round scoring through the interactive tower, the accuracy of the recalled content can be improved; through the index of the content feature vector of each content, when recalling the content, the target content can be quickly found according to the index, and the efficiency of content recall is improved on the basis of ensuring the accuracy of the recalled content.

[0173] Those skilled in the art can understand that to implement all or part of the processes in the above-mentioned embodiments, the processes can be completed by computer programs to instruct related hardware, and the programs can be stored in computer-readable storage media. When the programs are executed, they can include the processes of the above-mentioned method embodiments. The aforementioned storage media include: ROM or random access memory RAM, magnetic disk or optical disk and other media that can store program codes.

[0174] The above disclosure is only part of the embodiments of the present application, which certainly cannot be used to limit the scope of rights of the present application. Ordinary technicians in this field can understand that all or part of the processes of implementing the above embodiments and making equivalent changes according to the claims of this application are still within the scope of the present invention.

Claims

1. A content recall method, characterized in that: include: In response to the content recall request, extracting a target object feature vector of the target object using the object tower in the trained dual-tower model; Obtaining the index of the content feature vector of each content stored in the content database; the index of any content feature vector is used to indicate the storage layer where the any content feature vector is located, the storage layer includes M layers, and M is an integer greater than 1; Traversing the M layers, obtaining evaluation values ​​of the candidate content feature vectors and the target object feature vector in the currently traversed storage layer through the interactive tower in the trained dual-tower model, and determining the target content feature vector from the candidate content feature vectors in the currently traversed storage layer according to the evaluation values; the evaluation values ​​of the target content feature vector and the target object feature vector are greater than the evaluation values ​​of other content feature vectors and the target object feature vector in the candidate content feature vectors contained in the currently traversed storage layer except the target content feature vector; the target content feature vector is determined as the candidate content feature vector in the storage layer to be traversed next time; After the traversal is completed, the target content corresponding to the determined target content feature vector is sent to the client corresponding to the target object.

2. The method according to claim 1, characterized in that The traversing the M layers, obtaining evaluation values ​​of the candidate content feature vectors in the currently traversed storage layer and the target object feature vector through the interactive tower in the trained dual-tower model, and determining the target content feature vector from the candidate content feature vectors in the currently traversed storage layer according to the evaluation values, includes: Obtaining evaluation values ​​of the candidate content feature vectors in the Mth layer and the target object feature vector through the interactive tower, and determining the target content feature vector in the Mth layer from the candidate content feature vectors in the Mth layer according to the evaluation values ​​corresponding to the Mth layer; the evaluation values ​​of the target content feature vector in the Mth layer and the target object feature vector are greater than the evaluation values ​​of other content feature vectors in the candidate content feature vectors included in the Mth layer except the target content feature vector in the Mth layer and the target object feature vector; the target content feature vector in the Mth layer is determined as the candidate content feature vector in the M-1th layer; Obtaining evaluation values ​​of the candidate content feature vectors in the M-1th layer and the target object feature vector through the interactive tower, and determining the target content feature vector from the candidate content feature vectors in the M-1th layer according to the evaluation values ​​corresponding to the M-1th layer, until the target content feature vector in the first layer is determined; The step of sending the target content corresponding to the determined target content feature vector to the client corresponding to the target object includes: The target content corresponding to the target content feature vector in the first layer is sent to the client corresponding to the target object.

3. The method according to claim 2, characterized in that The index of any content feature vector is also used to indicate a content feature vector that is adjacent to any content feature vector in distance; The step of obtaining evaluation values ​​of the candidate content feature vectors in the Mth layer and the target object feature vector through the interactive tower, and determining the target content feature vector in the Mth layer from the candidate content feature vectors in the Mth layer according to the evaluation values ​​corresponding to the Mth layer, includes: Obtaining, through the interactive tower, evaluation values ​​of the candidate content feature vectors in the Mth layer and the target object feature vector, as well as evaluation values ​​of the content feature vectors and the target object feature vectors that are adjacent to the candidate content feature vectors in the Mth layer and are indicated by the index of the candidate content feature vectors in the Mth layer, and determining the target content feature vector in the Mth layer from the candidate content feature vectors in the Mth layer and the content feature vectors that are adjacent to the candidate content feature vectors in the Mth layer according to the evaluation values ​​corresponding to the Mth layer; The target content feature vector in the Mth layer and the content feature vector indicated by the index of the target content feature vector in the Mth layer and adjacent to the target content feature vector in the Mth layer are determined as candidate content feature vectors in the M-1th layer; the content feature vectors included in the Mth layer are from the M-1th layer, and the number of content feature vectors included in the M-1th layer is greater than the number of content feature vectors included in the M layer.

4. The method according to claim 1, characterized in that The evaluation value includes evaluation values ​​of N dimensions, where N is an integer greater than 1; the evaluation value of any dimension is used to characterize the degree of interaction between the target object and the content corresponding to each candidate content feature vector in the any dimension; The step of obtaining evaluation values ​​of the candidate content feature vectors in the currently traversed storage layer and the target object feature vector through the interactive tower in the trained dual-tower model, and determining the target content feature vector from the candidate content feature vectors in the currently traversed storage layer according to the evaluation values, includes: Obtaining evaluation values ​​of N dimensions of the candidate content feature vector and the target object feature vector in the currently traversed storage layer through the interaction tower; According to the evaluation values ​​of the N dimensions, evaluation values ​​of the candidate content feature vectors in the currently traversed storage layer and the target object feature vector are determined, and according to the evaluation values, a target content feature vector is determined from the candidate content feature vectors in the currently traversed storage layer.

5. The method according to claim 4, characterized in that The evaluation values ​​of the N dimensions include one or more of an exposure evaluation value, a duration evaluation value, and an interaction evaluation value; The exposure evaluation value is used to indicate exposure information obtained by the target object being exposed to the content corresponding to the candidate content feature vector in the currently traversed storage layer within a preset time period; The duration evaluation value is used to indicate the duration information obtained when the content corresponding to the candidate content feature vector in the currently traversed storage layer is displayed on the device corresponding to the target object; The interaction evaluation value is used to indicate the interaction information obtained by triggering the target object with respect to the content corresponding to the candidate content feature vector in the currently traversed storage layer.

6. The method according to claim 1, characterized in that The method further comprises: Acquire sample data, the sample data including training content, training object behavior data and label data, the label data including a preset evaluation value, the label data being used to characterize the degree of interaction between the training object corresponding to the training object behavior data and the training content; Calling the dual-tower model to extract features of the training content to obtain a training content feature vector of the training content; Extracting features from the training object behavior data to obtain a training object feature vector of the training object behavior data; the training object behavior data includes behavior data generated during the interaction between the training object and the training content; Obtaining training evaluation values ​​of the training content feature vector and the training object feature vector through the interactive tower in the dual-tower model; The twin-tower model is trained according to the training goal of reducing the difference between the training evaluation value and the preset evaluation value to obtain the trained twin-tower model.

7. The method according to claim 6, characterized in that The training evaluation value includes evaluation values ​​of N dimensions, where N is an integer greater than 1; the evaluation value of any dimension is used to characterize the degree of interaction between the training object and the training content in any dimension; The step of obtaining the training evaluation values ​​of the training content feature vector and the training object feature vector through the interactive tower in the dual-tower model includes: Obtaining evaluation values ​​of N dimensions of the training content feature vector and the training object feature vector through the interactive tower in the dual-tower model; The training evaluation value is determined according to the evaluation values ​​of the N dimensions.

8. A content recall device, characterized in that: include: An extraction unit, configured to extract a target object feature vector of a target object by using the object tower in the trained dual-tower model in response to a content recall request; An acquisition unit, used to acquire an index of a content feature vector of each content stored in a content database; the index of any content feature vector is used to indicate a storage layer where the content feature vector is located, and the storage layer includes M layers, where M is an integer greater than 1; The acquisition unit is further used to traverse the M layers, obtain evaluation values ​​of the candidate content feature vectors and the target object feature vector in the currently traversed storage layer through the interactive tower in the trained dual-tower model, and determine the target content feature vector from the candidate content feature vectors in the currently traversed storage layer according to the evaluation values; the evaluation values ​​of the target content feature vector and the target object feature vector are greater than the evaluation values ​​of other content feature vectors and the target object feature vector in the candidate content feature vectors contained in the currently traversed storage layer except the target content feature vector; the target content feature vector is determined as the candidate content feature vector in the storage layer to be traversed next time; The sending unit is used to send the target content corresponding to the determined target content feature vector to the client corresponding to the target object after the traversal is completed.

9. A computer device, characterized in that: The computer device includes a memory, a communication interface and a processor, wherein the memory, the communication interface and the processor are interconnected; the memory stores a computer program, and the processor calls the computer program stored in the memory to implement the content recall method described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the content recall method according to any one of claims 1 to 7 is implemented.

11. A computer program product, characterized in that The computer program product includes a computer program, which is stored in a computer storage medium; a processor of a computer device reads the computer program from the computer storage medium, and the processor executes the computer program, so that the computer device executes the content recall method as described in any one of claims 1 to 7.