Content processing method, device and computer readable storage medium

Abstract

The present application relates to a content processing method, apparatus, computer-readable storage medium and computer equipment, the method comprising: obtaining data of content to be counted under at least two similarity evaluation dimensions; determining content coding of the content to be counted under at least two similarity evaluation dimensions according to the data of the content to be counted under at least two similarity evaluation dimensions; obtaining content similarity between each of the content to be counted according to the content coding of the content to be counted under at least two similarity evaluation dimensions; determining target similar content of each of the content to be counted according to the content similarity between each of the content to be counted; the target similar content is used to recommend to the corresponding user. The scheme provided by the present application can realize the determination of content similarity between content from multiple similarity evaluation dimensions, thereby improving the accuracy of determining content similarity between content, which is conducive to improving the accuracy of determining similar content.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a content processing method, apparatus, computer-readable storage medium, and computer device. Background Technology

[0002] With the development of computer technology, various applications for browsing content have emerged, and more and more users are choosing to browse content through applications, such as news and information. During the process of users browsing content through applications, some similar content will be sent to the corresponding applications based on user requests.

[0003] However, the current method of determining similar content is based on the content type of each piece of content. Content with the same content type is identified as similar content, such as content with the same theme. This results in a large number of similar content being identified, but the actual content similarity is low, leading to a low accuracy rate in identifying similar content. Summary of the Invention

[0004] Therefore, it is necessary to provide a content processing method, apparatus, computer-readable storage medium, and computer device to address the technical problem of low accuracy in identifying similar content.

[0005] A content processing method, comprising:

[0006] Obtain data for the content to be analyzed across at least two similarity evaluation dimensions;

[0007] Based on the data of the content to be counted under at least two similarity evaluation dimensions, determine the content code of the content to be counted under at least two similarity evaluation dimensions;

[0008] Based on the content encoding of the content to be counted under the at least two similarity evaluation dimensions, the content similarity between each of the content to be counted is obtained;

[0009] Based on the content similarity between the various items to be counted, target similar content is determined for each item to be counted; the target similar content is used to recommend to the corresponding user.

[0010] In one embodiment, obtaining the content similarity between each of the content items to be counted based on the similarity scores of each pair of content items under at least two similarity evaluation dimensions includes:

[0011] Based on the similarity scores of the pairwise content to be statistically analyzed under at least two similarity evaluation dimensions, calculate the cross-score between each of the similarity scores;

[0012] Obtain the weights of each similarity score and each cross score;

[0013] Based on the similarity scores and their corresponding weights, and the cross scores and their corresponding weights, the content similarity between the contents to be counted is calculated.

[0014] In one embodiment, the content push request also carries a push mode;

[0015] The step of pushing the target-like content to the user terminal includes:

[0016] The target similar content is sorted according to content similarity;

[0017] The sorted target similar content is pushed to the user terminal according to the push mode.

[0018] A content processing apparatus, the apparatus comprising:

[0019] The data acquisition module is used to acquire data on the content to be analyzed across at least two similarity evaluation dimensions.

[0020] The encoding determination module is used to determine the content encoding of the content to be counted in at least two similarity evaluation dimensions based on the data of the content to be counted in at least two similarity evaluation dimensions.

[0021] The similarity acquisition module is used to obtain the content similarity between each of the contents to be counted based on the content encoding of the contents to be counted under the at least two similarity evaluation dimensions;

[0022] The content determination module is used to determine the target similar content for each of the content to be counted based on the content similarity between each of the content to be counted; the target similar content is used to recommend to the corresponding user.

[0023] A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the following steps:

[0024] Obtain data for the content to be analyzed across at least two similarity evaluation dimensions;

[0025] Based on the data of the content to be counted under at least two similarity evaluation dimensions, determine the content code of the content to be counted under at least two similarity evaluation dimensions;

[0026] Based on the content encoding of the content to be counted under the at least two similarity evaluation dimensions, the content similarity between each of the content to be counted is obtained;

[0027] Based on the content similarity between the various items to be counted, target similar content is determined for each item to be counted; the target similar content is used to recommend to the corresponding user.

[0028] A computer device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the following steps:

[0029] Obtain data for the content to be analyzed across at least two similarity evaluation dimensions;

[0030] Based on the data of the content to be counted under at least two similarity evaluation dimensions, determine the content code of the content to be counted under at least two similarity evaluation dimensions;

[0031] Based on the content encoding of the content to be counted under the at least two similarity evaluation dimensions, the content similarity between each of the content to be counted is obtained;

[0032] Based on the content similarity between the various items to be counted, target similar content is determined for each item to be counted; the target similar content is used to recommend to the corresponding user.

[0033] The aforementioned content processing method, apparatus, computer-readable storage medium, and computer device, based on acquired data of the content to be statistically analyzed under at least two similarity evaluation dimensions, determine the content encoding of the content to be statistically analyzed under at least two similarity evaluation dimensions, thereby obtaining the content similarity between each piece of content to be statistically analyzed, and determining the target similar content for each piece of content to be statistically analyzed based on the content similarity between each piece of content to be statistically analyzed. This achieves the goal of comprehensively determining the content similarity between the content to be statistically analyzed through multiple similarity evaluation dimensions, making the content similarity evaluation between content more accurate, thereby improving the accuracy rate of determining content similarity, and further improving the accuracy rate of determining similar content. Simultaneously, based on the accurately determined similar content, it is beneficial to achieve precise content recall and precise recommendation, thereby improving the accuracy rate of content recall and content recommendation. Attached Figure Description

[0034] Figure 1 This is an application environment diagram of the content processing method in one embodiment;

[0035] Figure 2 This is a structural block diagram of a computer device in one embodiment;

[0036] Figure 3 This is a flowchart illustrating a content processing method in one embodiment;

[0037] Figure 4This is a flowchart illustrating the steps for obtaining data on the content to be statistically analyzed in one embodiment;

[0038] Figure 5 This is a flowchart illustrating the steps for obtaining data on the content to be statistically analyzed in another embodiment;

[0039] Figure 6 This is a flowchart illustrating the steps for determining content encoding in one embodiment;

[0040] Figure 7 This is a flowchart illustrating the steps involved in constructing a matrix in one embodiment;

[0041] Figure 8 This is a flowchart illustrating the steps of decomposing a matrix in one embodiment;

[0042] Figure 9 This is a flowchart illustrating the steps for obtaining content similarity in one embodiment;

[0043] Figure 10 This is a flowchart illustrating the steps for calculating content similarity in one embodiment;

[0044] Figure 11 This is a flowchart illustrating the steps for calculating content similarity in another embodiment;

[0045] Figure 12 This is a flowchart illustrating the steps of training an FM model in one embodiment;

[0046] Figure 13 This is a flowchart illustrating the steps of training the FM model in another embodiment;

[0047] Figure 14 This is a flowchart illustrating the steps of storing target-similar content in one embodiment;

[0048] Figure 15 This is an application environment diagram of the content push method in one embodiment;

[0049] Figure 16 This is a flowchart illustrating the steps of a content push method in one embodiment;

[0050] Figure 17 This is a schematic diagram of a news recommendation interface in one embodiment;

[0051] Figure 18 This is a flowchart illustrating the steps of pushing content to a user terminal in one embodiment;

[0052] Figure 19 This is a timing diagram of a content push method in one embodiment;

[0053] Figure 20This is a flowchart illustrating the content push method in another embodiment;

[0054] Figure 21 This is a flowchart illustrating an article push method in one embodiment;

[0055] Figure 22 This is a structural block diagram of a content processing apparatus in one embodiment. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0057] Figure 1 This is a diagram illustrating the application environment of a content processing method in one embodiment. (Refer to...) Figure 1 The application environment diagram includes server 110. Server 110, based on the acquired content to be analyzed, such as content to be analyzed 1, content to be analyzed 2, etc., obtains data for the content to be analyzed in at least two similarity evaluation dimensions; based on the data for the content to be analyzed in at least two similarity evaluation dimensions, it determines the content encoding of the content to be analyzed in at least two similarity evaluation dimensions; based on the content encoding of the content to be analyzed in at least two similarity evaluation dimensions, it obtains the content similarity between each piece of content to be analyzed; based on the content similarity between each piece of content to be analyzed, it determines the target similar content for each piece of content to be analyzed; for example, the target similar content for content to be analyzed 1, the target similar content for content to be analyzed 2, etc. The target similar content is used to recommend to the corresponding user; for example, based on the content push request from the user's terminal, the server pushes the corresponding target similar content to the corresponding user terminal for the user to view.

[0058] Figure 2 An internal structural diagram of a computer device in one embodiment is shown. Specifically, this computer device may be... Figure 1 Server 110 in the middle. For example... Figure 2As shown, the computer device includes a processor, memory, network interface, input device, and display screen connected via a system bus. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system and may also store computer programs. When executed by the processor, these computer programs enable the processor to perform content processing methods. The internal memory may also store computer programs, which, when executed by the processor, enable the processor to perform content processing methods. The display screen can be an LCD screen or an e-ink screen. The input device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0059] Those skilled in the art will understand that Figure 2 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0060] like Figure 3 As shown, in one embodiment, a content processing method is provided. This embodiment mainly applies this method to the above-mentioned... Figure 1 Let's take server 110 as an example. (Refer to...) Figure 3 The content processing method specifically includes the following steps:

[0061] S302, Obtain data for the content to be analyzed in at least two similarity evaluation dimensions.

[0062] The content to be analyzed refers to articles whose content similarity needs to be statistically determined; for example, the content to be analyzed could be news, sports information, international news, etc. The content to be analyzed can be viewed through applications installed on the terminal; these applications could be browsers, dedicated news browsing applications, etc.

[0063] The similarity evaluation dimension refers to the evaluation angle used to measure the content similarity between content. For example, specific similarity evaluation dimensions could be user-click article angle, article-article keyword angle, user-click article keyword angle, etc. The user-click article angle describes whether a user has read an article; the article-article keyword angle describes the weight of keywords in an article; and the user-click article keyword angle describes whether a user has read an article containing those keywords. Data refers to the data of the content to be analyzed under each similarity evaluation dimension. For example, the data for the content to be analyzed under the user-click article angle refers to the relationship between the user and the content to be analyzed, such as whether the user has read the content or not.

[0064] It should be noted that, in addition to the listed user-click article perspective, article-article keyword perspective, and user-click article keyword perspective, other similarity evaluation dimensions may also be included, and this application does not impose any restrictions on them. At least two similarity evaluation dimensions may be included, which can be two or more similarity evaluation dimensions, and this application does not impose any restrictions on them.

[0065] Specifically, the server uses big data technology to collect unretrieved content from the current network as content to be statistically analyzed; it obtains multiple similarity evaluation dimensions, performs statistical analysis on the content to be statistically analyzed according to the similarity evaluation dimensions, and obtains the data of the content to be statistically analyzed under each similarity evaluation dimension.

[0066] In one embodiment, the server can also extract content from the local cache within the same time period as the content to be statistically analyzed; obtain multiple similarity evaluation dimensions; and perform statistical analysis on the content to be statistically analyzed based on the similarity evaluation dimensions to obtain the data of the content to be statistically analyzed under each similarity evaluation dimension.

[0067] S304, Based on the data of the content to be statistically analyzed in at least two similarity evaluation dimensions, determine the content encoding of the content to be statistically analyzed in at least two similarity evaluation dimensions.

[0068] Content encoding refers to article encoding, which is a real-valued vector of a specified dimension.

[0069] Specifically, the server obtains a matrix of the content to be counted in at least two similarity evaluation dimensions based on the data of the content to be counted in at least two similarity evaluation dimensions; the matrix is ​​then decomposed to obtain the content encoding of the content to be counted in each similarity evaluation dimension.

[0070] S306, based on the content encoding of the content to be counted under at least two similarity evaluation dimensions, obtain the content similarity between each piece of content to be counted.

[0071] Content similarity is used to describe the similarity between content; for example, the higher the content similarity between the content to be analyzed, the higher the similarity between the content to be analyzed.

[0072] Specifically, the server determines the similarity score of each pair of content to be counted in at least two similarity evaluation dimensions based on the content encoding of the content to be counted in at least two similarity evaluation dimensions; based on the similarity score of each pair of content to be counted in at least two similarity evaluation dimensions, the server determines the content similarity between each pair of content to be counted, thereby obtaining the content similarity between each pair of content to be counted.

[0073] S308, Based on the content similarity between each piece of content to be analyzed, determine the target similar content for each piece of content to be analyzed; the target similar content is used to recommend to the corresponding user.

[0074] Among them, target similar content refers to content that is similar to each content to be counted; for example, content with a content similarity greater than 60% and less than 80% with each content to be counted.

[0075] Specifically, the server determines the content with a content similarity greater than a first preset similarity and less than a second preset similarity with each of the content to be counted, based on the content similarity between each content to be counted, and uses this content as the target similar content for each content to be counted.

[0076] In one embodiment, after determining the target similar content for each content to be counted, the server can also receive a content push request from a user terminal, the content push request carrying a content identifier; parse the content push request to obtain the content identifier; determine the content to be counted corresponding to the content identifier; determine the target similar content corresponding to the content to be counted from the target similar content of each content to be counted; and push the target similar content to the corresponding terminal; this facilitates users browsing content that is similar to the content to be counted.

[0077] The above content processing method, based on the acquired data of the content to be analyzed across at least two similarity evaluation dimensions, determines the content encoding of the content to be analyzed across at least two similarity evaluation dimensions, thereby obtaining the content similarity between each piece of content to be analyzed, and based on the content similarity between each piece of content to be analyzed, determines the target similar content for each piece of content to be analyzed. This achieves the goal of comprehensively determining the content similarity between the content to be analyzed through multiple similarity evaluation dimensions, making the content similarity evaluation more accurate, thus improving the accuracy of content similarity determination, and further improving the accuracy of similar content determination. Simultaneously, based on the accurately determined similar content, it is beneficial to achieve precise content recall and precise recommendation, thereby improving the accuracy of content recall and content recommendation.

[0078] like Figure 4 As shown, in one embodiment, step S302, obtaining data on the content to be analyzed under at least two similarity evaluation dimensions, specifically includes the following steps:

[0079] S402, collect data on the content to be statistically analyzed and the data on user clicks on the content to be statistically analyzed within a preset time period.

[0080] Among them, the data on user clicks on the content to be counted refers to the historical data of whether the user has read the content to be counted.

[0081] Specifically, the server determines the content within a preset time period from the locally cached historical content as the content to be statistically analyzed; it collects user click data reported by multiple user terminals and determines the data of user clicks on the content to be statistically analyzed within the preset time period from the user click data; this is beneficial for obtaining the content to be statistically analyzed within the preset time period and the data of user clicks on the content to be statistically analyzed.

[0082] For example, the server selects articles within a preset time period from the historical articles cached in the local database as articles to be counted; it collects user click data reported by user terminals and counts the number of user clicks on the articles to be counted within the preset time period from the user click data.

[0083] S404: Based on the content to be statistically analyzed within a preset time period and the data of user clicks on the content to be statistically analyzed, obtain the data of the content to be statistically analyzed within a preset time period under at least two similarity evaluation dimensions.

[0084] Specifically, the server obtains multiple similarity evaluation dimensions, and performs statistical analysis on the content to be counted and the data of user clicks on the content to be counted within a preset time period based on the similarity evaluation dimensions, so as to obtain the data of the content to be counted under each similarity evaluation dimension within the preset time period.

[0085] In this embodiment, by performing statistical analysis on the content to be counted and the data of user clicks on the content to be counted within a preset time period, the data of the content to be counted under various similarity evaluation dimensions within the preset time period are obtained. This achieves the purpose of obtaining the data of the content to be counted under multiple similarity evaluation dimensions within a preset time period, which facilitates the subsequent analysis of the content similarity between the content to be counted from multiple similarity evaluation dimensions, and further improves the accuracy of determining similar content.

[0086] like Figure 5As shown, in one embodiment, if the similarity evaluation dimensions include a first similarity evaluation dimension, a second similarity evaluation dimension, and a third similarity evaluation dimension, then step S404, based on the content to be counted within a preset time period and the data of user clicks on the content to be counted, obtains the data of the content to be counted within each similarity evaluation dimension within the preset time period, specifically including the following steps:

[0087] S502, based on the data of user clicks on the content to be counted, determine the relationship between the user and the content to be counted within a preset time period, and use it as the data of the content to be counted under the first similarity evaluation dimension.

[0088] The first similarity evaluation dimension refers to the user-click article perspective, which describes the relationship between the user and the content to be analyzed. It indicates whether the user has read the content and reflects the similarity of the content from the user's perspective. The data of the content to be analyzed under the first similarity evaluation dimension can be represented by coordinates (i, j), where (i, j) indicates whether user i has read the content j. If user i has read the content j, the coordinate value of (i, j) is equal to 1; if user i has not read the content j, the coordinate value of (i, j) is equal to 0.

[0089] Specifically, the server analyzes the data of user clicks on the content to be counted within a preset time period to determine whether the user has read the content to be counted, thereby obtaining the relationship between the user and the content to be counted; the relationship between the user and the content to be counted is mapped to the corresponding coordinates, and the coordinates are used as the data of the content to be counted under the first similarity evaluation dimension within the preset time period.

[0090] For example, the server represents users as i and articles to be analyzed as j, and calculates the relationship between user i and article j. The relationship between user i and article j is mapped to the corresponding coordinates (i, j), and these coordinates (i, j) are used as the data of the article to be analyzed in the first similarity evaluation dimension within a preset time period.

[0091] S504, extract keywords from the content to be analyzed.

[0092] Keywords refer to key information used to describe the content to be analyzed; for example, in a news article describing the start of filming for a new TV series featuring a certain celebrity, keywords would include the celebrity's name, the name of the new TV series, etc. Each piece of content to be analyzed contains one or more keywords.

[0093] In practice, the server uses semantic recognition technology to identify the content to be counted within a preset time period, obtains the key information of the content to be counted, and uses the key information as the keywords of the content to be counted.

[0094] S506, obtain the weight of keywords in the content to be analyzed within a preset time period, and identify the weight as data of the content to be analyzed under the second similarity evaluation dimension.

[0095] The second similarity evaluation dimension refers to the article-article keyword perspective, which describes the weight of article keywords in the content to be analyzed, reflecting the semantic similarity between the content to be analyzed. The data of the content to be analyzed under the second similarity evaluation dimension can be represented by coordinates (i, j), where coordinates (i, j) represent the weight of keyword j in content i. If the weight of keyword j in content i is k, then the coordinate value corresponding to coordinate (i, j) is equal to k; if the weight of keyword j in content i is 0, then the coordinate value corresponding to coordinate (i, j) is equal to 0.

[0096] Specifically, the server analyzes the content to be statistically analyzed within a preset time period to determine the weight of the extracted keywords in the content to be statistically analyzed; the weight of the keywords in the content to be statistically analyzed is mapped to the corresponding coordinates, and the coordinates are used as the data of the content to be statistically analyzed under the second similarity evaluation dimension within the preset time period.

[0097] For example, the server represents the article to be analyzed as i and the keyword as j, and calculates the weight of keyword j in article i. The weight of keyword j in article i is mapped to the corresponding coordinate (i, j), and the coordinate (i, j) is used as the data of the article to be analyzed in the second similarity evaluation dimension within a preset time period.

[0098] S508: Based on the relationship between users and the content to be analyzed within a preset time period, and the weight of keywords in the content to be analyzed, the relationship between the weight of users and keywords in the content to be analyzed is determined, and this relationship is used as the data for the content to be analyzed in the third similarity evaluation dimension.

[0099] The third similarity evaluation dimension refers to the user-click article keyword perspective, which describes the relationship between the weight of the user and the keyword in the content to be analyzed. In other words, it indicates whether the user has read the content containing the keyword, reflecting the similarity of the keyword from the user's perspective. The data of the content to be analyzed under the third similarity evaluation dimension can be represented by coordinates (i, j), where coordinates (i, j) indicate whether user i has read the content containing keyword j. If user i has read the content containing keyword j, the coordinate value corresponding to coordinate (i, j) is equal to the weight of keyword j in the read content. If user i has not read the content containing keyword j, the coordinate value corresponding to coordinate (i, j) is equal to 0.

[0100] It should be noted that if the corresponding keyword j has different weights among the content that the user has read and is to be statistically analyzed, the maximum weight of keyword j will be used as the coordinate value corresponding to coordinate (i, j).

[0101] Specifically, the server analyzes the relationship between users and the content to be analyzed within a preset time period, as well as the weight of keywords in the content to be analyzed, to determine whether users have read the content to be analyzed containing keywords, thereby obtaining the relationship between the weight of users and keywords in the content to be analyzed; the relationship between the weight of users and keywords in the content to be analyzed is mapped to corresponding coordinates, and these coordinates are used as the data of the content to be analyzed in the third similarity evaluation dimension within the preset time period.

[0102] For example, the server represents users as i and keywords as j, and calculates the relationship between the weights of user i and keyword j in the article to be analyzed; the relationship between the weights of user i and keyword j in the article to be analyzed is mapped to the corresponding coordinates (i, j), and these coordinates (i, j) are used as the data of the article to be analyzed in the third similarity evaluation dimension within a preset time period.

[0103] In this embodiment, by analyzing the data of the content to be counted and the data of user clicks on the content to be counted within a preset time period, the data of the content to be counted under various similarity evaluation dimensions within the preset time period can be obtained. This is beneficial for subsequent analysis of the content similarity between the content to be counted from multiple similarity evaluation dimensions, and further improves the accuracy of determining similar content.

[0104] like Figure 6 As shown, in one embodiment, step S304, based on the data of the content to be statistically analyzed in at least two similarity evaluation dimensions, determines the content encoding of the content to be statistically analyzed in at least two similarity evaluation dimensions, specifically including the following steps:

[0105] S602, Based on the data of the content to be counted under at least two similarity evaluation dimensions, construct a matrix of the content to be counted under at least two similarity evaluation dimensions.

[0106] Here, a matrix refers to a matrix composed of the coordinate values ​​of the coordinates corresponding to the data of the content to be counted in the similarity evaluation dimension; for example, in the perspective of user-clicked article, with user i as the horizontal axis and the content to be counted j as the vertical axis, the matrix obtained based on the coordinate (i, j) is |(i, j)|.

[0107] Specifically, the server parses the data of the content to be analyzed under at least two similarity evaluation dimensions to obtain the coordinates of the content to be analyzed under at least two similarity evaluation dimensions; obtains the coordinate values ​​corresponding to the coordinates of the content to be analyzed under the same similarity evaluation dimension, and combines the coordinate values ​​to obtain the matrix of the content to be analyzed under that similarity evaluation dimension, thus obtaining the matrix of the content to be analyzed under at least two similarity evaluation dimensions.

[0108] For example, from the perspective of user-clicked articles, if user 1 has read article 1 but not article 2, and user 2 has read article 2 but not article 1, then (1,1) = 1, (1,2) = 0, (2,1) = 0, (2,2) = 1. The matrix composed of (1,1), (1,2), (2,1), and (2,2) is...

[0109]

[0110] This matrix is ​​identified as the matrix of Article 1 and Article 2 from the perspective of user-clicked articles. It should be noted that the process of constructing the matrix based on the data from the perspective of article-article keywords or user-clicked article keywords is the same as described above, and will not be repeated here.

[0111] S604, decompose the matrix to obtain the content encoding of the content to be counted under at least two similarity evaluation dimensions.

[0112] Specifically, considering that the matrix of the content to be analyzed under each similarity evaluation dimension is composed of the coordinate values ​​of two objects, the server performs matrix decomposition on the matrix of the content to be analyzed under each similarity evaluation dimension to obtain multiple object codes of the content to be analyzed under each similarity evaluation dimension. The content code is then selected from the multiple object codes or obtained by processing the object codes, and this content code is used as the content code of the content to be analyzed under the corresponding similarity evaluation dimension.

[0113] For example, from the perspective of user-clicked articles, the resulting object codes are user codes and article codes, both of which are real-valued vectors with a specified dimension. The article code is used as the content code of the content to be analyzed from the user-clicked article perspective. Similarly, from the perspective of user-clicked article keywords, the resulting object codes are user codes and article keyword codes, both of which are real-valued vectors with a specified dimension. The article keyword codes are weighted to obtain the corresponding article codes, which are then used as the content code of the content to be analyzed from the user-clicked article keyword perspective.

[0114] In this embodiment, by analyzing the data of the content to be counted under at least two similarity evaluation dimensions, the content encoding of the content to be counted under at least two similarity evaluation dimensions can be obtained, which facilitates the subsequent determination of the content similarity between each content to be counted based on the content encoding.

[0115] like Figure 7 As shown, in one embodiment, step S602 involves constructing a matrix of the content to be analyzed in at least two similarity evaluation dimensions based on the data of the content to be analyzed in at least two similarity evaluation dimensions. This specifically includes the following steps:

[0116] S702, construct a first matrix, a second matrix, and a third matrix based on the data of the content to be statistically analyzed in at least two similarity evaluation dimensions; the first matrix is ​​used to represent the relationship between user coding and content coding, the second matrix is ​​used to represent the relationship between content coding and keyword coding, and the third matrix is ​​used to represent the relationship between user coding and keyword coding.

[0117] Specifically, the server constructs a first matrix based on the data of the content to be analyzed under the first similarity evaluation dimension; constructs a second matrix based on the data of the content to be analyzed under the second similarity evaluation dimension; and constructs a third matrix based on the data of the content to be analyzed under the third similarity evaluation dimension.

[0118] S704 identifies the first matrix, the second matrix, and the third matrix as matrices of the content to be statistically analyzed under at least two similarity evaluation dimensions.

[0119] In this embodiment, a matrix of the content to be counted under at least two similarity evaluation dimensions is constructed based on the data of the content to be counted under at least two similarity evaluation dimensions. This facilitates the subsequent decomposition of the matrix to obtain the content encoding of the content to be counted under at least two similarity evaluation dimensions.

[0120] In one embodiment, if the matrix is ​​a first matrix or a second matrix, then in step S604, the matrix is ​​decomposed to obtain the content code of the content to be counted under at least two similarity evaluation dimensions, including: decomposing the matrix to obtain the content code, which is used as the content code of the content to be counted under the corresponding similarity evaluation dimension.

[0121] Specifically, the server decomposes the first matrix to obtain user codes and content codes, using the content codes as the content codes for the content to be analyzed under the first similarity evaluation dimension; it then decomposes the second matrix to obtain content codes and keyword codes, using the content codes as the content codes for the content to be analyzed under the second similarity evaluation dimension. This achieves the goal of obtaining the content codes for the content to be analyzed under both the first and second similarity evaluation dimensions.

[0122] like Figure 8 As shown, in another embodiment, if the matrix is ​​a third matrix, then in step S604, the matrix is ​​decomposed to obtain the content encoding of the content to be counted under at least two similarity evaluation dimensions, specifically including the following steps:

[0123] S802, decompose the matrix to obtain the keyword codes in the content to be counted.

[0124] Specifically, the server decomposes the third matrix to obtain user codes and keyword codes in the content to be analyzed.

[0125] S804, obtain the weight of the keyword in the content to be analyzed, and use it as the weight corresponding to the keyword code.

[0126] S806, based on keyword coding and corresponding weights, obtains content coding, which serves as the content coding of the content to be analyzed under the corresponding similarity evaluation dimension.

[0127] For example, if the content to be analyzed contains keywords 1, 2, and 3, and the weights of keywords 1, 2, and 3 are w1, w2, and w3 respectively, and the corresponding keyword codes are v1, v2, and v3 respectively, then the content code of the content to be analyzed under the corresponding similarity evaluation dimension is: w1×v1+w2×v2+w3×v3.

[0128] In this embodiment, when the matrix is ​​a third matrix, the content code corresponding to the content to be counted can be obtained by weighted combination of the keyword codes obtained from matrix decomposition. This achieves the purpose of processing the keyword codes corresponding to the content to be counted in order to obtain the content code corresponding to a single content to be counted.

[0129] like Figure 9 As shown, in one embodiment, step S306, based on the content encoding of the content to be counted under at least two similarity evaluation dimensions, obtains the content similarity between each piece of content to be counted, specifically including the following steps:

[0130] S902, based on the content codes of the content to be counted under at least two similarity evaluation dimensions, calculate the similarity scores of each pair of content to be counted under at least two similarity evaluation dimensions.

[0131] The similarity score refers to the cosine similarity between the content to be counted, which is used to describe the content similarity between any two pieces of content to be counted under the corresponding similarity evaluation dimension. It can be represented by the cosine similarity between the content to be counted.

[0132] For example, if the content encoding of article 1 in the first similarity evaluation dimension is v1, and the content encoding of article 2 in the first similarity evaluation dimension is v2, then the similarity score between article 1 and article 2 in the first similarity evaluation dimension is:

[0133] It should be noted that the similarity scores of the pairs of items to be statistically analyzed under other similarity evaluation dimensions can be calculated based on the above method, which will not be repeated here.

[0134] S904: Based on the similarity scores of each pair of content to be counted under at least two similarity evaluation dimensions, the content similarity between each content to be counted is obtained.

[0135] In this embodiment, the content similarity between each content to be counted is obtained by using the similarity scores of each pair of content to be counted under at least two similarity evaluation dimensions. This achieves the goal of calculating the content similarity between each content to be counted through multiple similarity evaluation dimensions, thereby improving the accuracy of determining the content similarity between content. This is beneficial for accurately determining the similar content of the content to be counted in the future, and further improves the accuracy of determining similar content.

[0136] like Figure 10 As shown, in one embodiment, step S904, based on the similarity scores of each pair of content to be counted under at least two similarity evaluation dimensions, obtains the content similarity between each content to be counted, specifically including the following steps:

[0137] S1002, calculate the cross-score between each similarity score based on the similarity scores of each pair of items to be statistically analyzed in at least two similarity evaluation dimensions.

[0138] The cross score refers to the intersection of two similarity scores. For example, the cross score of similarity score a and similarity score b is a×b.

[0139] S1004, obtain the weights of each similarity score and each cross score.

[0140] The weights of each similarity score and each cross score are obtained based on training with historical content.

[0141] S1006. Based on each similarity score and its corresponding weight, and each cross score and its corresponding weight, the content similarity between each item to be counted is calculated.

[0142] Specifically, the server calculates the product of each similarity score and its corresponding weight, as well as the product of each cross score and its corresponding weight. The products of each similarity score and its corresponding weight, and the products of each cross score and its corresponding weight are added together to obtain the content similarity between each pair of content to be counted, thus obtaining the content similarity between each piece of content to be counted.

[0143] In one embodiment, the content similarity between the various pieces of content to be analyzed can also be calculated in the following way:

[0144]

[0145] Where w0 is the initial weight, n is the number of similarity scores, and X i For similarity scores, w i For similarity score X i The corresponding weight, X i X j w represents the crossover score between similarity scores. ij Cross-score X i X j The corresponding weight, y(X), represents the content similarity between each pair of items to be counted.

[0146] Furthermore, in order to facilitate the calculation of content similarity between the content to be counted and to improve the accuracy of determining the content similarity between the content to be counted, step S1002, based on the similarity scores of each pair of content to be counted under each similarity evaluation dimension, calculates the cross-score between each similarity score, and further includes: discretizing the similarity scores of each pair of content to be counted under at least two similarity evaluation dimensions to obtain discrete values ​​of the similarity scores of each pair of content to be counted under at least two similarity evaluation dimensions; and calculating the discrete value of the cross-score between each discrete value of the similarity score.

[0147] For example, if Article 1 and Article 2 have similarity scores of 0.6, 0.7, and 0.5 across the three similarity evaluation dimensions, and their scores are rounded to the nearest ten using the following ten-fold discretization formula, the resulting discrete similarity scores are 6, 7, and 5. Therefore, the cross-score between the discrete similarity score of 6 and the discrete similarity score of 7 is 6 × 7.

[0148] The formula for the ten-order separation and dispersion is shown below:

[0149] round(max(X×10,0));

[0150] Where X is the similarity score, and round is the rounding operation. It should be noted that the similarity score with a value less than 0 has a discrete value of 0. Therefore, the similarity score of two articles under one similarity evaluation dimension has a total of 11 values, ranging from 0 to 10.

[0151] like Figure 11 As shown, in another embodiment, step S904, based on the similarity scores of each pair of content to be counted under at least two similarity evaluation dimensions, obtains the content similarity between each content to be counted, specifically including the following steps:

[0152] S1102, input the similarity scores of each pair of items to be statistically analyzed in at least two similarity evaluation dimensions into the pre-trained FM model.

[0153] The FM model refers to the Factorization Machines model, also known as the factor decomposition machine model. Compared to the linear regression model, the FM model introduces a binary cross term, which considers the impact of the cross combination of two similarity scores on the regression accuracy.

[0154] S1104. Using the FM model, confirm the cross scores between each similarity score; obtain the weights of each similarity score and each cross score; and calculate the content similarity between each piece of content to be analyzed based on each similarity score and its corresponding weight, and each cross score and its corresponding weight.

[0155] In this embodiment, the server inputs the similarity scores of each pair of content to be counted under each similarity evaluation dimension into a pre-trained FM model. The FM model calculates the content similarity between each content to be counted, fully considering the cross-score of the similarity scores, which further improves the accuracy of determining the content similarity between the content to be counted.

[0156] like Figure 12 As shown, in one embodiment, the FM model is trained through the following steps:

[0157] S1202, obtain the similarity scores of sample content in at least two similarity evaluation dimensions and the corresponding first weights, the cross scores between each similarity score and the corresponding second weights, and the actual content similarity between each sample content.

[0158] Here, sample content refers to historical articles, and the actual content similarity can be represented by the true co-occurrence (i.e., ground truth) of the two articles in the current time period. The ground truth can be represented by support. For example, if in 1000 click sequences, article 1 and article 2 are clicked together 600 times, then the support is 600 / 1000 = 0.6. Therefore, the ground truth of the two articles in the current time period is 0.6, meaning the actual content similarity between the two articles is 0.6.

[0159] S1204. The FM model is trained based on the similarity scores of the sample content in at least two similarity evaluation dimensions and the corresponding first weights, the cross scores between the similarity scores and the corresponding second weights, to obtain the trained FM model.

[0160] Specifically, the server trains the FM model using the similarity scores of different sample contents across at least two similarity evaluation dimensions, the corresponding first weights, the cross-scores between the similarity scores, and the corresponding second weights.

[0161] S1206, obtain the prediction error between the content similarity output by the trained FM model and the corresponding actual content similarity.

[0162] S1208 When the prediction error is greater than or equal to the preset threshold, the first weight and the second weight are adjusted according to the prediction error, and the FM model is repeatedly trained according to the adjusted first weight and the second weight until the prediction error obtained from the trained FM model is less than the preset threshold.

[0163] Specifically, when the measurement error is greater than or equal to a preset threshold, the server continuously adjusts the first and second weights in the FM model to repeatedly train the FM model until the prediction error obtained from the trained FM model is less than the preset threshold, and the current FM model is used as the trained FM model.

[0164] In this embodiment, by repeatedly training the FM model, the server can improve the accuracy of the content similarity predicted by the FM model, and further improve the accuracy of determining the content similarity between content.

[0165] In one embodiment, such as Figure 13 As shown, another method for training the FM model is provided, which includes the following steps:

[0166] S1302, collect user click data and article information within a preset time period.

[0167] Specifically, the server extracts user-clicked article data and article information within a preset time period from a pre-established user-clicked article database and an article information cache database. The user-clicked article database stores user-clicked article data from multiple users, while the article information cache database caches information from multiple articles.

[0168] S1304. Based on the user click data and article information within a preset time period, obtain the data of the article to be statistically analyzed from the perspectives of user-click article, user-click article keyword, and article-article keyword, and obtain the user-click article data, user-click article keyword data, and article-article keyword data of the article to be statistically analyzed within the preset time period.

[0169] Among them, the user-click article perspective, the user-click article keyword perspective, and the article-article keyword perspective refer to the similarity evaluation dimensions.

[0170] S1306, Encode the user-click article data using matrix factorization to obtain the article code; Encode the user-click article keyword data using matrix factorization to obtain the article keyword code; Perform weighted calculation on the article keyword codes to obtain the article code; Encode the article-article keyword data using matrix factorization to obtain the article code.

[0171] Among these, the article codes obtained from the user-click article perspective better reflect the similarity of articles from the user's point of view, embodying article similarity based on user interests, and appearing more abstract. The article keyword codes obtained from the user-click article keyword perspective reflect the similarity of article keywords from the user's point of view; since the number of article keyword codes is on the order of millions, it is more granular than the previous article code, better reflecting the details that users like. Evaluating article similarity based on article codes weighted by article keyword codes reflects users' detailed interests, thus reducing abstraction. The article codes obtained from the article-article keyword perspective do not consider user information, but reflect article similarity information from a semantic perspective.

[0172] In this way, user-click article data, user-click article keyword data, and article-article keyword data are encoded separately through matrix factorization, so that the encoded data reflects the similarity of the articles to be counted among users, articles, and article keywords in different similarity evaluation dimensions.

[0173] S1308 uses cosine similarity calculation to calculate the similarity score between pairs of articles based on the article codes of the articles to be analyzed from the perspectives of user-clicked articles, user-clicked article keywords, and article-article keywords.

[0174] For example, the article codes for article 1 to be analyzed are v from the perspectives of user-click article, user-click article keyword, and article-article keyword, respectively. 11 v 12 v 13 The article codes for article 2, which need to be statistically analyzed, are v in the user-click article perspective, user-click article keyword perspective, and article-article keyword perspective. 21 v 22 v 23 Then, the article similarity scores between Article 1 and Article 2 under the perspectives of user-click article, user-click article keyword, and article-article keyword are respectively...

[0175] S1310: Discretize the similarity scores of each pair of articles to be statistically analyzed from the perspectives of user-clicked articles, user-clicked articles keywords, and articles-article keywords to obtain discrete values ​​of similarity scores from these perspectives. Using the true co-occurrence of articles within a preset time period as the ground truth, and using the discrete values ​​of similarity scores from these perspectives as features, train the FM model to be trained to obtain the trained FM model.

[0176] Co-occurrence refers to the joint appearance of two elements in an observation sequence. For example, in natural language processing, two co-occurring words are usually analyzed and processed as a group. True co-occurrence within a preset time period describes the true article similarity between two articles. It refers to the proportion of shared clicks between the two articles to the total number of clicks. For example, if the total number of clicks is 1000, and article 1 and article 2 were shared 600 times, then the true article similarity between article 1 and article 2 within the preset time period is 600 / 1000 = 0.6, meaning the true similarity between article 1 and article 2 within the preset time period is 0.6.

[0177] Discretizing the similarity scores of each pair of articles to be analyzed from the perspectives of user-clicked articles, user-clicked articles keywords, and articles-article keywords means taking the rounded value after multiplying the similarity scores of each pair of articles to be analyzed by 10, and using it as the discrete value of the similarity scores from the perspectives of user-clicked articles, user-clicked articles keywords, and articles-article keywords.

[0178] For example, suppose the discrete values ​​of the article similarity scores between article 1 and article 2 under the user-click article perspective, user-click article keyword perspective, and article-article keyword perspective are x1', x2', and x3', respectively. Input x1', x2', and x3' as features into the FM model to be trained, and obtain the article similarity between article 1 and article 2 output by the FM model. If the difference between the article similarity and the corresponding real article similarity is greater than or equal to a preset threshold, then the FM model to be trained is retrained based on the difference until the difference obtained from the FM model is less than the preset threshold. Then, the current FM model is taken as the trained FM model.

[0179] Furthermore, the trained FM model can be used to obtain the article similarity between any two articles, which facilitates subsequent article retrieval based on user needs and recommendation to the corresponding users.

[0180] In this embodiment, online co-occurrence data reflects users' true preferences for different articles. Using co-occurrence data as benchmark data effectively reflects users' actual behavior. Simultaneously, the FM algorithm is used to train the model, considering not only linear relationships between different features but also binary relationships, resulting in better model robustness. Thus, by considering article similarity from different perspectives, the similarity evaluation between articles is more accurate, further improving the accuracy of identifying similar articles; simultaneously, it provides users with better article recall, enabling the retrieval of a large number of high-quality articles that users are interested in.

[0181] In one embodiment, step S308, which determines the target similar content of each content to be counted based on the content similarity between each content to be counted, specifically includes: taking the content to be counted that has a content similarity greater than a first preset similarity and less than a second preset similarity with each content to be counted as the target similar content of each content to be counted.

[0182] For example, based on the content similarity between other articles and article 1, articles with a content similarity greater than 0.65 and less than 0.85 with article 1 are identified, and these articles are considered as the target similar content of article 1.

[0183] In this embodiment, the goal of determining the target similar content of each content to be counted is achieved based on the content similarity between each content to be counted, thereby further improving the accuracy of similar content determination.

[0184] like Figure 14 As shown, in one embodiment, after determining the target similar content of each content to be counted in step S208, the method further includes a step of storing the target similar content of the content to be counted, specifically including the following steps:

[0185] S1402, Obtain the content identifier of the content to be analyzed.

[0186] Content identifiers refer to the identifier information used to identify the content to be counted, such as content number and content name.

[0187] Specifically, the server assigns a corresponding content number to the content to be counted, such as 1, 2, 3, etc., and uses the content number as the content identifier of the content to be counted.

[0188] Furthermore, the server can also obtain the content name of the content to be counted and use the content name as the content identifier of the content to be counted.

[0189] S1404: Recall the target similar content of the content to be counted to the recall file corresponding to the content identifier.

[0190] The recall file is a file that stores the target similar content of the content to be counted, and there is a one-to-one correspondence between the file and the content identifier of the content to be counted. For example, recall file 1 stores the target similar content of content to be counted 1, recall file 2 stores the target similar content of content to be counted 2, and recall file 3 stores the target similar content of content to be counted 3.

[0191] Taking the content to be counted as an example, the server will recall the top N articles that are similar to a single article to the recall file corresponding to the article's article identifier.

[0192] It should be noted that servers can also recall content in other ways, such as recall based on user profiles or collaborative filtering.

[0193] In one embodiment, the server may also obtain the user profile of the target user, determine the preference weight of the target user for each piece of content to be counted based on the user profile of the target user, take the content to be counted with a preference weight greater than a preset weight as the content to be pushed associated with the target user, and recall the content to be pushed associated with the target user to the recall file corresponding to the user identifier of the target user.

[0194] User profiles are user preference descriptions abstracted from historical user behavior data of target users. They can be used to characterize the degree of preference of target users for various content categories. User identifiers are used to identify user information of target users, and can be user names, user accounts, etc.

[0195] S1406, Store the recall file.

[0196] Specifically, the server stores the recall file in a Redis database. Of course, the server can also use other caching methods to store the recall file.

[0197] Furthermore, after storing the target similar content of the content to be counted in the form of a recall file, when a user requests the content, the server can also extract the corresponding content from the recall file and push the content to the corresponding user.

[0198] In this embodiment, the goal of storing the target similar content of the content to be counted is achieved, which facilitates the accurate push of the target similar content to the corresponding user according to the user's request, and further improves the accuracy of content push.

[0199] In one embodiment, this application also provides a content push method, which can be applied to, for example... Figure 15 The application environment diagram shown is for reference. Figure 15 This content recommendation method is applied to a content recommendation system. The system includes a user terminal 1510 and a server 1520; the user terminal 1510 and server 1520 are connected via a network. The user terminal can be a desktop terminal or a mobile terminal; a mobile terminal can be at least one of a mobile phone, tablet, or laptop. The server can be a standalone server or a server cluster consisting of multiple servers.

[0200] like Figure 16 As shown, in one embodiment, a content push method is provided. This embodiment mainly applies this method to the above-mentioned... Figure 15 Let's take server 1520 as an example. (Refer to...) Figure 16 The content push method specifically includes the following steps:

[0201] S1602, Receive a content push request sent by the user terminal; the content push request carries a content identifier.

[0202] Specifically, when a user clicks the refresh button on the content recommendation interface of their user terminal, the user terminal generates a content push request based on the content identifier of the content the user has already read, and sends the content push request to the corresponding server.

[0203] S1604, Extract target-similar content from the recall file corresponding to the content identifier.

[0204] The recall file corresponding to the content identifier stores the target similar content of the content corresponding to that content identifier.

[0205] Specifically, the server finds the recall file corresponding to the content identifier based on the correspondence between the content identifier and the recall file, and extracts the target similar content of the content corresponding to the content identifier from the recall file.

[0206] Furthermore, in addition to extracting target similar content corresponding to content identifiers from pre-stored recall files, the server can also recall target similar content corresponding to content identifiers in real time.

[0207] For example, the server acquires data on the content to be analyzed on the network in real time under at least two similarity evaluation dimensions; determines the content code of the content to be analyzed under at least two similarity evaluation dimensions based on the data of the content to be analyzed under at least two similarity evaluation dimensions; obtains the content similarity between each piece of content to be analyzed based on the content code of the content to be analyzed under at least two similarity evaluation dimensions; determines the target similar content of the content corresponding to the content identifier based on the content similarity between each piece of content to be analyzed; and recalls the target similar content of the content corresponding to the content identifier.

[0208] S1606, push similar content to the target to the user terminal.

[0209] Specifically, the server pushes similar content to the target terminal so that the target similar content corresponding to the content identifier can be displayed through the content recommendation interface on the user terminal. For example, the target similar content can be displayed in order of content similarity from high to low.

[0210] Please refer to Figure 17Let's take the "Tencent News" application as an example. If a user has the "Tencent News" application installed on their device, and they've read news item 1 on the news recommendation page of the "Tencent News" application, clicking the refresh button will trigger a news push request, which will be sent to the server via the user's device. Based on the news push request, the server extracts similar news items from the recall file of news item 1 and pushes these similar news items to the "Tencent News" application. This allows the application's news recommendation page to display similar news items such as "News 1," "News 2," and "News 3," etc.

[0211] It should be noted that the content recommendation method of this application can be applied not only to news recommendation scenarios, but also to other content recommendation scenarios, and this application does not limit the specific applications.

[0212] In this embodiment, the goal of pushing relevant target-like content to the corresponding user based on the content push request is achieved, making it easier for users to view content they are interested in, while improving the accuracy of content push.

[0213] like Figure 18 As shown, in one embodiment, the content push request also carries a push mode. Therefore, step S1606, pushing similar content to the user terminal, specifically includes the following steps:

[0214] S1802, Sort the target similar content according to content similarity.

[0215] Specifically, the server sorts the target similar content in descending order of content similarity to obtain the sorted target similar content.

[0216] S1804: The sorted target similar content is pushed to the user terminal according to the push mode.

[0217] Among them, the content push mode is used to represent the content layout mode of similar content, such as graphic view mode, text view mode, video view mode, image view mode, etc.

[0218] Specifically, the server converts the sorted target similar content into content corresponding to the push mode, and pushes this content to the user terminal so that the sorted target similar content can be displayed on the content recommendation page of the user terminal.

[0219] In this embodiment, based on multiple similarity evaluation dimensions, target similar content is filtered out from a massive amount of content and pushed to the user terminal, which helps to improve the accuracy of content push.

[0220] like Figure 19 As shown, Figure 19 This is a sequence diagram of a content push method in one embodiment. User terminal 1510 initiates a content push request and sends it to server 1520. Server 1520 parses the content push request to obtain a content identifier; extracts target similar content from the recall file corresponding to the content identifier; and pushes the target similar content to user terminal 1510. User terminal 1510 displays the target similar content, making it convenient for users to view content of interest, further improving the accuracy of content push.

[0221] like Figure 20 As shown, in one embodiment, another content push method is provided. This embodiment mainly applies this method to the above-mentioned... Figure 15 Let's take server 1520 as an example. (Refer to...) Figure 20 The content push method specifically includes the following steps:

[0222] S2002, determine the target similar content and content identifiers of the content to be counted.

[0223] Specifically, determining the target similar content of the content to be statistically analyzed can be done with reference to this application, such as... Figure 3 The embodiment of the content processing method shown will not be described in detail here. The server determines the content identifier of the content to be counted based on the content name or content number of the content to be counted.

[0224] S2004 retrieves similar content to the target content to be counted and sends it to the recall file corresponding to the content identifier.

[0225] S2006: Obtain the user profile and user identifier of the target user, and determine the content to be pushed to the target user based on the user profile.

[0226] Specifically, the server can determine the target user's preference for each content category based on the target user's historical user behavior data, thereby obtaining the target user's user profile; obtain the target user's username or user account as the target user's user identifier; determine the target user's preference weight for each piece of content to be analyzed based on the target user's user profile; and select the content to be analyzed with a preference weight greater than a preset weight as the content to be pushed to the target user.

[0227] S2008 retrieves the content to be pushed to the target user and sends it to the recall file corresponding to the user identifier.

[0228] S2010, Receive content push request sent by user terminal; the content push request carries content identifier and user identifier.

[0229] S2012, extract target-similar content from the recall file corresponding to the content identifier, and extract content to be pushed from the recall file corresponding to the user identifier.

[0230] S2014 pushes similar content and content to be pushed to the user's terminal.

[0231] Specifically, the server inputs the target similar content and the content to be pushed to the rank server and the rerank server. The rank server and the rerank server perform fine sorting on the obtained target similar content and the content to be pushed, and distribute them to the user terminals corresponding to the target users according to certain rules, so that the target users can selectively read the obtained content according to their own interests.

[0232] Furthermore, the server can update the user profile corresponding to the target user based on the reading behavior feedback of the target user, and provide training data for content recall based on multiple similarity evaluation dimensions, which helps to improve the accuracy of content push.

[0233] In this embodiment, combining content recall based on multiple similarity evaluation dimensions and content recall based on user profiles helps to push more accurate content to users, further improving the accuracy of content push.

[0234] like Figure 21 As shown, Figure 21 This is a flowchart illustrating an article push method in one embodiment. In one embodiment, an article push method is provided; this embodiment primarily applies this method to the above-mentioned method. Figure 15 Let's take server 1520 as an example.

[0235] Reference Figure 21The article push method specifically includes the following: The server recalls articles to be counted based on a multi-angle recall algorithm; for example, the server determines the target similar articles and article identifiers of the articles to be counted; and recalls the target similar articles to the articles to be counted to the recall file corresponding to the article identifiers. The server recalls articles to be counted based on other recall algorithms; for example, the server obtains the user profile and user identifier of the target user, determines the articles to be pushed associated with the target user based on the user profile, and recalls the articles to be pushed associated with the target user to the recall file corresponding to the user identifiers. The server stores the recall files corresponding to article identifiers and the recall files corresponding to user identifiers in a Redis database. When a user requests an article, the server receives an article push request from the user's terminal. The server parses the article push request to obtain the article identifier and the user identifier. In the Redis database, the server extracts target similar articles from the recall files corresponding to the content identifiers and extracts articles to be pushed from the recall files corresponding to the user identifiers. The target similar articles and articles to be pushed are input into the rank server and the rerank server. The rank server and the rerank server perform fine sorting on the obtained target similar articles and articles to be pushed, and distribute them to the user terminals corresponding to the target users according to certain rules, so that the target users can read the recommended articles.

[0236] Furthermore, the user terminal statistically reports and feeds back the target user's reading behavior of recommended articles to the server; based on the feedback of the target user's reading behavior of recommended articles, the server updates the user profile corresponding to the target user and provides training data for the multi-angle recall algorithm, further improving the accuracy of content push.

[0237] In this embodiment, a recall algorithm based on multi-angle encoding of article click co-occurrence makes the similarity evaluation between articles more accurate, thereby improving the accuracy and coverage of recalled articles. Simultaneously, it reduces the input to the rank server, thus alleviating its pressure and improving its resource utilization. Furthermore, the recommended articles output by the rerank server are more representative of user interests and meet user needs, thereby improving the accuracy of article pushes.

[0238] It should be understood that, although Figure 3-14 The steps in flowcharts 16, 18, and 20 are shown sequentially as indicated by the arrows; however, these steps are not necessarily executed in the exact order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are performed, and they can be executed in other orders. Figure 3-14At least some of the steps in 16, 18, and 20 may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but may be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but may be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0239] like Figure 22 As shown, in one embodiment, a content processing apparatus 2200 is provided, which includes: a data acquisition module 2202, an encoding determination module 2204, a similarity acquisition module 2206, and a content determination module 2208, wherein:

[0240] The data acquisition module 2202 is used to acquire data on the content to be analyzed in at least two similarity evaluation dimensions.

[0241] The encoding determination module 2204 is used to determine the content encoding of the content to be counted in at least two similarity evaluation dimensions based on the data of the content to be counted in at least two similarity evaluation dimensions.

[0242] The similarity acquisition module 2206 is used to obtain the content similarity between each piece of content to be counted based on the content encoding of the content to be counted under at least two similarity evaluation dimensions.

[0243] The content determination module 2208 is used to determine the target similar content for each piece of content to be counted based on the content similarity between each piece of content to be counted; the target similar content is used to recommend to the corresponding user.

[0244] In one embodiment, the data acquisition module 2202 is further configured to collect data on the content to be statistically analyzed and the data on user clicks on the content to be statistically analyzed within a preset time period; and based on the data on the content to be statistically analyzed and the data on user clicks on the content to be statistically analyzed within the preset time period, to obtain data on the content to be statistically analyzed under at least two similarity evaluation dimensions within the preset time period.

[0245] In one embodiment, the similarity evaluation dimensions include a first similarity evaluation dimension, a second similarity evaluation dimension, and a third similarity evaluation dimension. The data acquisition module 2202 is further configured to determine the relationship between the user and the content to be analyzed within a preset time period based on the data of user clicks on the content to be analyzed, and use this relationship as data for the content to be analyzed under the first similarity evaluation dimension; extract keywords from the content to be analyzed; obtain the weight of the keywords in the content to be analyzed within a preset time period, and identify the weight as data for the content to be analyzed under the second similarity evaluation dimension; and confirm the relationship between the weight of the user and the keywords in the content to be analyzed based on the relationship between the user and the content to be analyzed within a preset time period, as well as the weight of the keywords in the content to be analyzed, and use this relationship as data for the content to be analyzed under the third similarity evaluation dimension.

[0246] In one embodiment, the encoding determination module 2204 is further configured to construct a matrix of the content to be counted in at least two similarity evaluation dimensions based on the data of the content to be counted in at least two similarity evaluation dimensions; decompose the matrix to obtain the content encoding of the content to be counted in at least two similarity evaluation dimensions.

[0247] In one embodiment, the encoding determination module 2204 is further configured to construct a first matrix, a second matrix, and a third matrix based on the data of the content to be counted under at least two similarity evaluation dimensions; the first matrix is ​​used to represent the relationship between user encoding and content encoding, the second matrix is ​​used to represent the relationship between content encoding and keyword encoding, and the third matrix is ​​used to represent the relationship between user encoding and keyword encoding; the first matrix, the second matrix, and the third matrix are respectively identified as matrices of the content to be counted under each similarity evaluation dimension.

[0248] In one embodiment, if the matrix is ​​a first matrix or a second matrix, the encoding determination module 2204 is further used to decompose the matrix to obtain the content encoding, which serves as the content encoding of the content to be counted in the corresponding similarity evaluation dimension.

[0249] In one embodiment, if the matrix is ​​a third matrix, the encoding determination module 2204 is further used to decompose the matrix to obtain the keyword codes in the content to be counted; obtain the weights of the keywords in the content to be counted as the weights corresponding to the keyword codes; and obtain the content codes based on the keyword codes and their corresponding weights as the content codes of the content to be counted under the corresponding similarity evaluation dimension.

[0250] In one embodiment, the similarity acquisition module 2206 is further configured to calculate the similarity score of each pair of content to be counted in at least two similarity evaluation dimensions based on the content encoding of the content to be counted in at least two similarity evaluation dimensions; and obtain the content similarity between each pair of content to be counted based on the similarity score of each pair of content to be counted in at least two similarity evaluation dimensions.

[0251] In one embodiment, the similarity acquisition module 2206 is further configured to calculate the cross score between each similarity score based on the similarity scores of each pair of content to be counted under at least two similarity evaluation dimensions; obtain the weight of each similarity score and the weight of each cross score; and calculate the content similarity between each piece of content to be counted based on each similarity score and its corresponding weight, and each cross score and its corresponding weight.

[0252] In one embodiment, the similarity acquisition module 2206 is further configured to input the similarity scores of each pair of content to be counted under each similarity evaluation dimension into a pre-trained FM model; through the FM model, confirm the cross scores between each similarity score; obtain the weights of each similarity score and the weights of each cross score; and calculate the content similarity between each content to be counted based on each similarity score and its corresponding weight, and each cross score and its corresponding weight.

[0253] In one embodiment, the content processing device 2200 further includes a model training module.

[0254] The model training module is used to obtain the similarity scores of sample content in at least two similarity evaluation dimensions and their corresponding first weights, the cross scores between each similarity score and their corresponding second weights, and the actual content similarity between each sample content; based on the similarity scores of sample content in at least two similarity evaluation dimensions and their corresponding first weights, the cross scores between each similarity score and their corresponding second weights, the FM model is trained to obtain the trained FM model; the prediction error between the content similarity output by the trained FM model and the corresponding actual content similarity is obtained; when the prediction error is greater than or equal to a preset threshold, the first weights and second weights are adjusted according to the prediction error, and the FM model is repeatedly trained according to the adjusted first weights and second weights until the prediction error obtained from the trained FM model is less than the preset threshold.

[0255] In one embodiment, the content determination module 2208 is further configured to, based on the content similarity between each content to be counted, identify the content to be counted that has a content similarity greater than a first preset similarity and less than a second preset similarity as the target similar content of each content to be counted.

[0256] In one embodiment, the content processing device 2200 further includes a recall file storage module.

[0257] The recall file storage module is used to obtain the content identifier of the content to be counted; recall the target similar content of the content to be counted to the recall file corresponding to the content identifier; and store the recall file.

[0258] In one embodiment, the content processing device 2200 further includes a content push module.

[0259] The content push module is used to receive content push requests sent by user terminals; the content push request carries a content identifier; extract target similar content from the recall file corresponding to the content identifier; and push the target similar content to the user terminal.

[0260] In one embodiment, the content push request also carries a push mode; the content push module is further used to sort the target similar content according to content similarity; and push the sorted target similar content to the user terminal according to the push mode.

[0261] In this embodiment, based on the acquired data of the content to be statistically analyzed under various similarity evaluation dimensions, the content encoding of the content to be statistically analyzed under each similarity evaluation dimension is determined, thereby obtaining the content similarity between each piece of content to be statistically analyzed. Based on the content similarity between each piece of content to be statistically analyzed, the target similar content for each piece of content to be statistically analyzed is determined. This achieves the goal of comprehensively determining the content similarity between the content to be statistically analyzed through multiple similarity evaluation dimensions, making the content similarity evaluation more accurate, thus improving the accuracy rate of content similarity determination, and further improving the accuracy rate of similar content determination. Simultaneously, based on the accurately determined similar content, it is beneficial to achieve precise content recall and precise recommendation, thereby improving the accuracy rate of content recall and content recommendation.

[0262] In one embodiment, the content processing apparatus provided in this application can be implemented as a computer program, and the computer program can be implemented as follows: Figure 2 It runs on the computer device shown. The computer device's memory can store the various program modules that make up the content processing device, for example, Figure 22 The data acquisition module 2202, encoding determination module 2204, similarity acquisition module 2206, and content determination module 2208 are shown. The computer program comprised of these modules causes the processor to execute the steps of the content processing methods in the various embodiments of this application described in this specification.

[0263] For example, Figure 2 The computer device shown can be used as follows Figure 22The data acquisition module 2202 in the content processing device 2200 acquires data on the content to be analyzed in at least two similarity evaluation dimensions, and the encoding determination module 2204 determines the content encoding of the content to be analyzed in at least two similarity evaluation dimensions based on the data. The computer device can obtain the content similarity rows between each piece of content to be analyzed based on the content encoding of the content to be analyzed in at least two similarity evaluation dimensions through the similarity acquisition module 2206, and determine the target similar content for each piece of content to be analyzed based on the content similarity between each piece of content to be analyzed through the content determination module 2208; the target similar content is used to recommend to the corresponding user.

[0264] In one embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the content processing method described above. The steps of the content processing method here can be steps from the content processing methods of the various embodiments described above.

[0265] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, causes the processor to perform the steps of the content processing method described above. The steps of the content processing method here may be steps from the content processing methods of the various embodiments described above.

[0266] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0267] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0268] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A content processing method, comprising: Obtain data on the content to be counted under at least two similarity evaluation dimensions; The content to be counted is an article; The similarity evaluation dimension refers to an evaluation angle used to measure the content similarity between articles; Determining content codes of the content to be counted under the at least two similarity evaluation dimensions according to data of the content to be counted under the at least two similarity evaluation dimensions; Obtaining content similarities between the contents to be counted according to content codes of the contents to be counted under the at least two similarity evaluation dimensions; The content similarity is calculated based on the similarity scores of the two contents to be counted in the at least two similarity evaluation dimensions and the cross scores between the similarity scores; The similarity score is obtained based on the content encoding; Determining target similar content for each of the contents to be counted according to the content similarities between the contents to be counted; The target similar content is used to recommend to the corresponding user.

2. The method according to claim 1, characterized in that The step of obtaining data of the content to be counted under at least two similarity evaluation dimensions includes: Collect the content to be counted within a preset time period and the data of users clicking on the content to be counted; According to the content to be counted within the preset time period and the data of users clicking on the content to be counted, data of the content to be counted in at least two similarity evaluation dimensions within the preset time period are obtained.

3. The method according to claim 2, characterized in that The similarity evaluation dimensions include a first similarity evaluation dimension, a second similarity evaluation dimension and a third similarity evaluation dimension; The obtaining, according to the content to be counted within the preset time period and the data of users clicking on the content to be counted, data of the content to be counted in at least two similarity evaluation dimensions within the preset time period includes: Determine, according to the data of the user clicking on the content to be counted, the relationship between the user and the content to be counted within a preset time period as the data of the content to be counted under the first similarity evaluation dimension; Extracting keywords from the content to be counted; Obtaining the weight of the keyword in the content to be counted within the preset time period, and identifying the weight as data of the content to be counted under the second similarity evaluation dimension; According to the relationship between the user and the content to be counted within the preset time period, and the weight of the keyword in the content to be counted, the relationship between the user and the weight of the keyword in the content to be counted is confirmed as the data of the content to be counted under the third similarity evaluation dimension.

4. The method according to claim 1, characterized in that: The step of determining the content codes of the content to be counted in at least two similarity evaluation dimensions according to the data of the content to be counted in at least two similarity evaluation dimensions comprises: According to the data of the content to be counted in at least two similarity evaluation dimensions, constructing a matrix of the content to be counted in at least two similarity evaluation dimensions; The matrix is ​​decomposed to obtain content codes of the content to be counted under the at least two similarity evaluation dimensions.

5. The method according to claim 4, characterized in that According to the data of the content to be counted in at least two similarity evaluation dimensions, constructing a matrix of the content to be counted in at least two similarity evaluation dimensions, including: A first matrix, a second matrix and a third matrix are constructed according to the data of the content to be counted under at least two similarity evaluation dimensions respectively; the first matrix is ​​used to represent the relationship between the user code and the content code, the second matrix is ​​used to represent the relationship between the content code and the keyword code, and the third matrix is ​​used to represent the relationship between the user code and the keyword code; Respectively identifying the first matrix, the second matrix, and the third matrix as matrices of the content to be counted under the at least two similarity evaluation dimensions; If the matrix is ​​the first matrix or the second matrix, decomposing the matrix to obtain content codes of the content to be counted under the at least two similarity evaluation dimensions includes: The matrix is ​​decomposed to obtain content coding as the content coding of the content to be counted under the corresponding similarity evaluation dimension.

6. The method according to claim 5, characterized in that If the matrix is ​​the third matrix, decomposing the matrix to obtain content codes of the content to be counted under the at least two similarity evaluation dimensions includes: Decomposing the matrix to obtain keyword codes in the content to be counted; Obtaining the weight of the keyword in the content to be counted as the weight corresponding to the keyword code; According to the keyword codes and the corresponding weights, content codes are obtained as the content codes of the content to be counted under the corresponding similarity evaluation dimension.

7. The method according to claim 1, characterized in that The obtaining the content similarity between the respective contents to be counted according to the content coding of the contents to be counted in the at least two similarity evaluation dimensions comprises: Calculating similarity scores of two pieces of content to be counted in at least two similarity evaluation dimensions according to content codes of the content to be counted in at least two similarity evaluation dimensions; Calculate the cross score between the similarity scores of the two pieces of content to be counted in at least two similarity evaluation dimensions; The content similarity between the contents to be counted is obtained according to the similarity scores of the two contents to be counted in at least two similarity evaluation dimensions and the cross scores of the similarity scores.

8. The method according to claim 7, characterized in that The obtaining of the content similarity between the contents to be counted according to the similarity scores of the two contents to be counted in at least two similarity evaluation dimensions and the cross scores of the similarity scores comprises: Obtaining the weight of each of the similarity scores and the weight of each of the crossover scores; The content similarity between the contents to be counted is calculated based on the similarity scores and the corresponding weights, and the cross scores and the corresponding weights.

9. The method according to claim 1, characterized in that: The obtaining the content similarity between the respective contents to be counted according to the content coding of the contents to be counted under the at least two similarity evaluation dimensions comprises: Calculating similarity scores of two pieces of content to be counted in at least two similarity evaluation dimensions according to content codes of the content to be counted in at least two similarity evaluation dimensions; Inputting similarity scores of the pairwise statistically analyzed contents under at least two similarity evaluation dimensions into a pre-trained FM model; Through the FM model, the cross-scores between the similarity scores are confirmed; the weights of the similarity scores and the cross-scores are obtained; and the content similarities between the contents to be counted are calculated based on the similarity scores and the corresponding weights, the cross-scores and the corresponding weights.

10. The method according to claim 9, characterized in that The FM model is trained in the following way: Obtaining similarity scores of sample contents under at least two similarity evaluation dimensions and corresponding first weights, cross scores between the similarity scores and corresponding second weights, and actual content similarities between the sample contents; Training the FM model according to similarity scores of the sample content in at least two similarity evaluation dimensions and corresponding first weights, cross scores between the similarity scores and corresponding second weights, to obtain a trained FM model; Obtaining a prediction error between the content similarity output by the trained FM model and the corresponding actual content similarity; When the prediction error is greater than or equal to a preset threshold, the first weight and the second weight are adjusted according to the prediction error, and the FM model is repeatedly trained according to the adjusted first weight and second weight until the prediction error obtained according to the trained FM model is less than the preset threshold.

11. The method according to claim 1, characterized in that Determining target similar content of each of the to-be-counted contents according to the content similarity between each of the to-be-counted contents includes: According to the content similarities between the contents to be counted, the contents to be counted whose content similarities with the contents to be counted are greater than the first preset similarity and less than the second preset similarity are used as target similar contents of the contents to be counted.

12. The method according to any one of claims 1 to 11, characterized in that After determining the target similar content of each of the contents to be counted, it includes: Obtaining the content identifier of the content to be counted; Recalling the target similar content of the content to be counted into a recall file corresponding to the content identifier; The recall file is stored.

13. The method according to claim 12, characterized in that The method further comprises: Receiving a content push request sent by a user terminal; the content push request carries a content identifier; Extracting target similar content from the recalled file corresponding to the content identifier; The target similar content is pushed to the user terminal.

14. The method according to claim 13, characterized in that The content push request also carries a push mode; The pushing the target similar content to the user terminal comprises: Sorting the target similar contents according to content similarity; The sorted target similar content is pushed to the user terminal according to the push mode.

15. A content processing device, characterized in that: The device comprises: A data acquisition module is used to acquire data of the content to be counted under at least two similarity evaluation dimensions; the content to be counted is an article; the similarity evaluation dimension refers to an evaluation angle used to measure the content similarity between articles. A coding determination module, configured to determine content coding of the content to be counted in at least two similarity evaluation dimensions according to data of the content to be counted in at least two similarity evaluation dimensions; A similarity acquisition module, configured to obtain content similarities between the contents to be counted according to content codes of the contents to be counted under the at least two similarity evaluation dimensions; the content similarities are calculated based on similarity scores of the two contents to be counted under the at least two similarity evaluation dimensions and cross scores between the similarity scores; the similarity scores are obtained based on the content codes; The content determination module is used to determine target similar content of each of the contents to be counted according to the content similarity between each of the contents to be counted; the target similar content is used to recommend to the corresponding user.

16. The device according to claim 15, characterized in that The data acquisition module is also used to collect the content to be counted within a preset time period and the data of users clicking on the content to be counted; based on the content to be counted within the preset time period and the data of users clicking on the content to be counted, the data of the content to be counted in at least two similarity evaluation dimensions within the preset time period is acquired.

17. The device according to claim 16, characterized in that The similarity evaluation dimensions include a first similarity evaluation dimension, a second similarity evaluation dimension and a third similarity evaluation dimension; The data acquisition module is further used to determine, based on the data of the user clicking on the content to be counted, the relationship between the user and the content to be counted within a preset time period as the data of the content to be counted under the first similarity evaluation dimension; Extracting keywords from the content to be counted; The weight of the keyword in the content to be counted within the preset time period is obtained, and the weight is identified as the data of the content to be counted under the second similarity evaluation dimension; based on the relationship between the user and the content to be counted within the preset time period, and the weight of the keyword in the content to be counted, the relationship between the user and the weight of the keyword in the content to be counted is confirmed as the data of the content to be counted under the third similarity evaluation dimension.

18. The device according to claim 15, characterized in that The coding determination module is further used to construct a matrix of the content to be counted under the at least two similarity evaluation dimensions according to the data of the content to be counted under the at least two similarity evaluation dimensions; The matrix is ​​decomposed to obtain content codes of the content to be counted under the at least two similarity evaluation dimensions.

19. The device according to claim 18, characterized in that The coding determination module is further used to construct a first matrix, a second matrix and a third matrix according to the data of the content to be counted under at least two similarity evaluation dimensions; the first matrix is ​​used to represent the relationship between the user code and the content code, the second matrix is ​​used to represent the relationship between the content code and the keyword code, and the third matrix is ​​used to represent the relationship between the user code and the keyword code; Respectively identifying the first matrix, the second matrix, and the third matrix as matrices of the content to be counted under the at least two similarity evaluation dimensions; If the matrix is ​​the first matrix or the second matrix, the coding determination module is further used to decompose the matrix to obtain content coding as the content coding of the content to be counted under the corresponding similarity evaluation dimension.

20. The device according to claim 19, characterized in that If the matrix is ​​the third matrix, the code determination module is further used to decompose the matrix to obtain the keyword codes in the content to be counted; obtain the weight of the keyword in the content to be counted as the weight corresponding to the keyword code; According to the keyword codes and the corresponding weights, content codes are obtained as the content codes of the content to be counted under the corresponding similarity evaluation dimension.

21. The device according to claim 15, characterized in that The similarity acquisition module is further used to calculate the similarity scores of the two contents to be counted in at least two similarity evaluation dimensions according to the content codes of the contents to be counted in at least two similarity evaluation dimensions; and calculate the cross score between the similarity scores according to the similarity scores of the two contents to be counted in at least two similarity evaluation dimensions; The content similarity between the contents to be counted is obtained according to the similarity scores of the two contents to be counted in at least two similarity evaluation dimensions and the cross scores of the similarity scores.

22. The device according to claim 21, characterized in that The similarity acquisition module is further used to obtain the weight of each similarity score and the weight of each cross score; and calculate the content similarity between each content to be counted according to each similarity score and the corresponding weight, each cross score and the corresponding weight.

23. The device according to claim 15, characterized in that The similarity acquisition module is further used to calculate the similarity scores of the contents to be counted in pairs under at least two similarity evaluation dimensions according to the content encoding of the contents to be counted in at least two similarity evaluation dimensions; input the similarity scores of the contents to be counted in pairs under at least two similarity evaluation dimensions into a pre-trained FM model; confirm the cross-scores between the similarity scores through the FM model; obtain the weights of the similarity scores and the weights of the cross-scores; and calculate the content similarities between the contents to be counted according to the similarity scores and the corresponding weights, and the cross-scores and the corresponding weights.

24. The device according to claim 23, characterized in that The device also includes: a model training module, which is used to obtain the similarity scores of the sample content under at least two similarity evaluation dimensions and the corresponding first weights, the cross scores between each of the similarity scores and the corresponding second weights, and the actual content similarities between each of the sample contents; train the FM model according to the similarity scores of the sample content under at least two similarity evaluation dimensions and the corresponding first weights, the cross scores between each of the similarity scores and the corresponding second weights to obtain a trained FM model; obtain the prediction error between the content similarity output by the trained FM model and the corresponding actual content similarity; when the prediction error is greater than or equal to a preset threshold, adjust the first weight and the second weight according to the prediction error, and repeatedly train the FM model according to the adjusted first weight and second weights until the prediction error obtained according to the trained FM model is less than the preset threshold.

25. The device according to claim 15, characterized in that The content determination module is further configured to, according to the content similarities between the various contents to be counted, respectively select the contents to be counted whose content similarities with the various contents to be counted are greater than a first preset similarity and less than a second preset similarity as target similar contents for the various contents to be counted.

26. The device according to any one of claims 15 to 25, characterized in that The device further comprises: a recall file storage module, which is used to obtain the content identifier of the content to be counted; recall the target similar content of the content to be counted into the recall file corresponding to the content identifier; and store the recall file.

27. The device according to claim 26, characterized in that The device also includes: a content push module, which is used to receive a content push request sent by a user terminal; the content push request carries a content identifier; extracts target similar content from a recall file corresponding to the content identifier; and pushes the target similar content to the user terminal.

28. The device according to claim 27, characterized in that The content push request also carries a push mode; the content push module is further used to sort the target similar contents according to content similarity; and push the sorted target similar contents to the user terminal according to the push mode.

29. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 14.

30. A computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method according to any one of claims 1 to 14.

Images