Content screening method and device, electronic equipment, medium and product

Abstract

The invention provides a content screening method and device, electronic equipment, a medium and a product, and relates to the technical field of data pushing. In the method, when interest detection is needed, associated information of first content browsed by a target user in a first historical period and associated information of second content browsed by the target user in a second historical period are obtained; the second historical time period is before the first historical time period; the associated information comprises a description text; inputting the associated information of the first content, the associated information of the second content and the cue word into a large language model to obtain an interest migration prediction result containing a target content type interested by the target user at the current moment; the cue word is used for indicating the large language model to detect the interest migration condition of the target user from the second historical time period to the first historical time period according to the association information difference between the first content and the second content; and determining the content of the target content type as the content pushed to the target user. According to the invention, the accuracy of the determined content to be pushed to the user can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data pushing, and in particular to a content screening method and device, electronic equipment, medium and product. BACKGROUND

[0002] With the rapid development of the Internet, users can access content (such as music, video, etc.) pushed by a server through a client. In order to improve the access rate of the content, the server can determine the content to be pushed to each user according to the content type of interest of the user recorded in advance.

[0003] However, the content type of interest of the user can change, and therefore, in this case, how to improve the accuracy of the determined content to be pushed to the user becomes a problem to be solved. SUMMARY

[0004] The purpose of the embodiments of the present application is to provide a content screening method and device, electronic equipment, medium and product to improve the accuracy of the determined content to be pushed to the user. The specific technical solutions are as follows:

[0005] In a first aspect, the embodiments of the present application provide a content screening method, which comprises:

[0006] When interest detection of a target user is needed, the associated information of the first content browsed by the target user in a first historical period is obtained, and the associated information of the second content browsed by the target user in a second historical period is obtained; wherein the second historical period is before the first historical period; the associated information of one content includes the description text of the content;

[0007] The associated information of each first content, the associated information of each second content and the prompt word are input into a large language model to obtain the interest transfer prediction result of the target user; wherein the interest transfer prediction result of the target user includes the target content type of interest of the target user at the current time; the prompt word is used to instruct the large language model to detect the interest transfer of the target user from the second historical period to the first historical period according to the difference between the associated information of each first content and the associated information of each second content;

[0008] Among the selected content, the content belonging to the target content type is determined as the content to be pushed to the target user.

[0009] Optionally, before the associated information of each first content, the associated information of each second content and the prompt word are input into the large language model, the method further comprises:

[0010] obtain a similarity description text used for describing the similarity between each first content and each second content;

[0011] The inputting of the association information of each first content, the association information of each second content, and the prompt word into the large language model comprises:

[0012] The association information of each first content, the association information of each second content, the similarity description text, and the prompt word are input into the large language model to obtain the interest migration prediction result of the target user; wherein the prompt word is further used to instruct the large language model to output an interest migration prediction result representing that the interest of the target user has not changed in the case that the similarity description text represents that the similarity between each first content and second content is greater than a similarity threshold, and to output an interest migration prediction result representing that the interest of the target user has changed in the case that the similarity description text represents that the similarity between each first content and second content is not greater than the similarity threshold.

[0013] Optionally, the obtaining of the similarity description text used for describing the similarity between each first content and each second content comprises:

[0014] The association information of each first content is sorted in the order of the browsing time to obtain a first information sequence;

[0015] The first information sequence is input into a pre-trained interest prediction model to obtain a first information vector output by a feature extraction layer in the interest prediction model; wherein the interest prediction model is trained based on a sample information sequence and a true value; the sample information sequence is obtained by sorting the association information of each sample content in the order of the browsing time of a sample user to browse each sample content; the true value represents the content type that the sample user is interested in; the interest prediction model further comprises an output layer; the output layer is used to normalize the information vector output by the feature extraction layer to obtain a predicted content type;

[0016] A second information vector corresponding to a second information sequence is obtained; wherein the second information sequence is obtained by sorting the association information of each second content in the order of the browsing time; and the second information vector is obtained by inputting the second information sequence into the interest prediction model;

[0017] A description text used for describing the similarity between the first information vector and the second information vector is generated to obtain the similarity description text used for describing the similarity between each first content and each second content.

[0018] Optionally, the first historical period is a period from a first time point of the last interest detection of the target user to a current time point; and the second historical period is a period from a second time point of the latest interest detection of the target user before the first time point to the first time point.

[0019] Optionally, the prompt word further comprises: historical content types in the interest migration prediction result obtained at the first time point.

[0020] In the case that the interest of the target user migrates, the interest migration prediction result of the target user further comprises: a description text representing that the interested type of the target user is updated from the historical content type to the target content type.

[0021] Before determining, in the candidate content, the content belonging to the target content type as the content for pushing to the target user, the method further comprises:

[0022] In a pre-stored knowledge graph, a node representing the historical content type is determined as a first node, and a node representing the target content type is determined as a second node; wherein the nodes in the knowledge graph correspond to content types one by one, and the content types represented by two connected nodes are associated;

[0023] From the knowledge graph, other nodes connected to the first node are determined as third nodes, and other nodes connected to the second node are determined as fourth nodes;

[0024] A similarity between the content type represented by the third node and the content type represented by the fourth node is calculated as a to-be-used similarity;

[0025] The determining, in the candidate content, the content belonging to the target content type as the content for pushing to the target user comprises:

[0026] If the to-be-used similarity is greater than a preset threshold, the content belonging to the target content type is determined in the candidate content as the content for pushing to the target user.

[0027] Optionally, the method further comprises:

[0028] In the candidate content, the content belonging to the historical content type is determined;

[0029] The content belonging to the historical content type and the content belonging to the target content type are pushed to the target user.

[0030] Optionally, before pushing the content belonging to the historical content type and the content belonging to the target content type to the target user, the method further comprises:

[0031] obtaining the interest migration prediction results obtained respectively when the interest detection is performed on the target user for the historical specified number of times as the historical interest migration prediction results;

[0032] for any one of the historical interest migration prediction results and the interest migration prediction result of the target user, if the interest migration prediction result contains the target content type, determining that the interest migration prediction result is a to-be-utilized interest migration prediction result;

[0033] calculating the proportion of the to-be-utilized interest migration prediction result in the historical interest migration prediction results and the interest migration prediction result of the target user as the weight of the target content type;

[0034] the pushing of the content belonging to the historical content type and the content belonging to the target content type to the target user comprises:

[0035] pushing the content belonging to the historical content type and the content belonging to the target content type to the target user according to the weight of the target content type.

[0036] Optionally, the prompt words further contain the browsing time of the target user for each first content and the browsing time of the target user for each second content;

[0037] the prompt words are used to instruct the large language model to determine the importance of each content for the interest migration prediction result according to the browsing time of the content, and determine the first content and the second content with the importance greater than a preset importance threshold, and detect the interest migration of the target user from the second historical period to the first historical period according to the difference between the determined associated information of each first content and the determined associated information of each second content; the time length of the browsing time of a content from the current time and the importance of the content for the interest migration prediction result are inversely related.

[0038] Optionally, the associated information of a content further contains at least one of the following: the type of the browsing device used by the target user to browse the content, the type of the time period in which the target user browses the content, the browsing mode type of the target user to browse the content, and the time length of the target user to browse the content; the browsing mode type of the target user to browse the content is continuous watching mode or non-continuous watching mode.

[0039] In a second aspect, an embodiment of the present application provides a content screening device, the device comprising:

[0040] The first obtaining module is configured to, when interest detection of a target user is needed, obtain associated information of first content browsed by the target user in a first historical period, and obtain associated information of second content browsed by the target user in a second historical period; the second historical period is before the first historical period; the associated information of one content includes a description text of the content;

[0041] The input module is configured to input the associated information of each first content, the associated information of each second content, and a prompt word into a large language model to obtain an interest migration prediction result of the target user; the interest migration prediction result of the target user includes a target content type in which the target user is interested at a current time; the prompt word is used to instruct the large language model to detect interest migration of the target user from the second historical period to the first historical period according to differences between the associated information of each first content and the associated information of each second content;

[0042] The determining module is configured to determine, in the candidate content, content belonging to the target content type as content for pushing to the target user.

[0043] The embodiment of the application further provides an electronic device, including a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory complete mutual communication through the communication bus.

[0044] The memory is used for storing a computer program.

[0045] The processor is used for executing the program stored on the memory, and realizes the content screening method.

[0046] The embodiment of the application further provides a computer readable storage medium, the computer readable storage medium stores a computer program, and the computer program is executed by the processor to realize the content screening method.

[0047] The embodiment of the application further provides a computer program product, including a computer program, and the computer program is executed by the processor to realize the content screening method.

[0048] In the embodiment of the application, when interest detection of a target user is needed, a description text of first content browsed by the target user in a first historical period and a description text of second content browsed by the target user in a second historical period are obtained. Since the second historical period is before the first historical period, differences between each second content and each first content can reflect interest migration of the target user from the second historical period to the first historical period.

[0049] The embodiment of the present application can identify the semantic difference between the description text of the first content and the description text of the second content through a large language model, can identify the change of the content type of the content browsed by the target user from the second historical period to the first historical period with the help of the text reasoning capability of the large language model, can obtain the change trend of the content browsed by the target user, can infer the change of the content type interested by the target user at the current moment, can accurately determine the content type interested by the target user at the current moment, and can further screen the content conforming to the content type interested by the target user at the current moment, thereby improving the accuracy of the content determined to be pushed to the user. BRIEF DESCRIPTION OF DRAWINGS

[0050] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced.

[0051] Figure 1 The flowchart of the first content screening method provided by the embodiment of the present application;

[0052] Figure 2 The flowchart of the second content screening method provided by the embodiment of the present application;

[0053] Figure 3 The flowchart of determining the similarity description text in the content screening method provided by the embodiment of the present application;

[0054] Figure 4 The flowchart of the third content screening method provided by the embodiment of the present application;

[0055] Figure 5 The flowchart of the fourth content screening method provided by the embodiment of the present application;

[0056] Figure 6 The structural diagram of the content screening device provided by the embodiment of the present application;

[0057] Figure 7 The structural diagram of the electronic device provided by the embodiment of the present application. DETAILED DESCRIPTION

[0058] The technical solutions in the embodiments of the present application will be described below with reference to the drawings in the embodiments of the present application.

[0059] The content type interested by the user (i.e., the user interest) can quickly migrate due to factors such as hot events, seasonal changes, or personal status, and in this case, how to improve the accuracy of the content determined to be pushed to the user becomes a problem to be solved.

[0060] To improve the accuracy of the determined content to be pushed to the user, the embodiment of the application provides a content screening method and device, an electronic device, a medium and a product.

[0061] The content screening method provided by the embodiment of the application can be applied to a server, the server can screen content to be pushed to a target user, and then push the screened content to a client logged in by the target user. For example, the server can be a server of a video provider or a server of a music provider. The server can be a physical server or a virtual server, and the client can be a mobile phone or a tablet computer. The target user is a user who needs to use the client to access the content provided by the server. To distinguish from the target user, the user of the server can be a technician, who can be a technician of the video provider or a technician of the music provider.

[0062] The content screening method provided by the embodiment of the application comprises:

[0063] When it is necessary to detect the interest of the target user, the associated information of the first content browsed by the target user in a first historical period is obtained, and the associated information of the second content browsed by the target user in a second historical period is obtained. The second historical period is before the first historical period. The associated information of a content includes the description text of the content.

[0064] The associated information of each first content, the associated information of each second content and the prompt word are input into the large language model to obtain the interest migration prediction result of the target user. The interest migration prediction result of the target user includes the target content type in which the target user is interested at the current moment. The prompt word is used to instruct the large language model to detect the interest migration of the target user from the second historical period to the first historical period according to the difference between the associated information of each first content and the associated information of each second content.

[0065] In the selected content, the content belonging to the target content type is determined as the content to be pushed to the target user.

[0066] The embodiment of the application can identify the semantic difference between the description text of the first content and the description text of the second content through the large language model. The change of the content type of the content browsed by the target user from the second historical period to the first historical period can be identified with the help of the text reasoning capability of the large language model, the change trend of the content browsed by the target user is obtained, and the change of the content type in which the target user is interested at the current moment is inferred. The content type in which the target user is interested at the current moment is accurately determined, and then the content conforming to the content type in which the target user is interested at the current moment is screened, thereby improving the accuracy of the determined content to be pushed to the user.

[0067] The content screening method provided by the embodiment of the present application will be described below with reference to the accompanying drawings. As shown in Figure 1 The method can include steps S101-S103.

[0068] S101, when interest detection of a target user is needed, obtaining associated information of a first content browsed by the target user in a first historical period, and obtaining associated information of a second content browsed by the target user in a second historical period.

[0069] The second historical period is before the first historical period; and the associated information of a content includes a description text of the content.

[0070] It can be understood that a user can register an account, log in a client through the registered account, and then access a server; the server can store the accounts of the registered users to verify the identities of the users, and then provide the users with contents. The target user can be a user specified by a technical staff from the registered users, or the target user can also be each user in the registered users. The moment when the interest detection of the target user is needed can be a preset moment, for example, the preset moment can be a periodic moment, such as 12 o'clock on Sunday night; or the moment when the interest detection of the target user is needed can also be a moment when the server receives a detection instruction for the interest detection of the target user issued by the technical staff.

[0071] If the interest detection of the target user is needed, the server can determine that the interest detection of the target user is needed at a current moment. The current moment is the moment when the interest detection of the target user is needed; for example, the current moment can be the preset moment, such as 12 o'clock on Sunday night; or the current moment can also be the moment when the detection instruction is received this time.

[0072] The first historical period can be a period before the current moment, and the second historical period can be a period before the first historical period.

[0073] In an implementation manner, in order to more accurately determine the content type that the target user is interested in at the current moment, the server can take a period closest to the current moment as the first historical period, and take a period closest to the first historical period as the second historical period.

[0074] Specifically, the first historical period is a period from a first time point of the last interest detection of the target user to the current time point; and the second historical period is a period from a second time point of the last interest detection of the target user before the first time point to the first time point. In the embodiment of the present application, the content browsed by the target user from the first time point to the current time point can reflect the current interest of the target user, and the content browsed by the target user from the second time point to the first time point can reflect the historical interest of the target user before a certain period of time. Subsequently, by analyzing each first content and second content through the large language model, whether the historical interest of the target user changes to the current interest can be inferred, and the interest change of the target user in the recent period of time can be determined, so as to accurately determine the content type that the target user is interested in at the current time point.

[0075] The first time point is a time point of the last interest detection of the target user, that is, a time point of the last time when the interest detection of the target user is needed. For example, the first time point can be a last preset time point, such as 12 o'clock on Sunday night, and the first time point is 12 o'clock on the last Sunday night; or the first time point can also be a time point of the last time when a detection instruction is received. The second time point is a time point of the last interest detection of the target user before the first time point, that is, a time point of the last time when the interest detection of the target user is needed. For example, the second time point can be a last last preset time point, such as 12 o'clock on Sunday night, and the second time point is 12 o'clock on the last last Sunday night; or the second time point is a time point of the last last time when a detection instruction is received.

[0076] In another implementation manner, the server can select any period of time before the current time point as the first historical period, and select any period of time before the first historical period as the second historical period. In this regard, the embodiment of the present application is only illustrative, and is not specifically limited.

[0077] The server can obtain the required description text from the pre-stored description texts of the contents. For example, the server can store the account of each user in association with the behavior log of the user, and the behavior log of each user records the identity of each content browsed by the user and the browsing time of the content. In order to obtain the required description text, the server can determine the behavior log of the target user in the behavior logs of the users according to the account of the target user, and then obtain the identity of the content whose browsing time is in the first historical period (i.e., the identity of the first content) and the identity of the content whose browsing time is in the second historical period (i.e., the identity of the second content) from the behavior log of the target user. The pre-stored description text of each content can be stored in association with the identity of the content, and the server can obtain the description text associated with the identity of the first content from the pre-stored description texts of the contents to obtain the description text of the first content browsed by the target user in the first historical period, and obtain the description text associated with the identity of the second content to obtain the description text of the second content browsed by the target user in the second historical period.

[0078] The identity (ID) of a content can be the name or serial number of the content, etc. The form of the browsing time of a user browsing a content can be a timestamp; or the form of the browsing time of a user browsing a content can be a structured format time, for example, the browsing time can be 2025-01-02-0800, indicating that the browsing time is on January 2, 2025 at 8 o'clock.

[0079] It can be understood that the description text of a content is used to describe the content, for example, if a content is a video, the description text of the content can include at least one of the following: the title of the video, the introduction of the video, and the cover description text of the video; wherein the cover description text of the video can be the description text of the content contained in the cover of the video. For another example, if a content is music, the description text of the content can include at least one of the following: the category of the music, the title of the music, and the lyrics of the music. The description text of a content can be referred to as the metadata of the content, for example, if the content is a video, the description text of the content can be referred to as video metadata; if the content is music, the description text of the content can be referred to as music metadata.

[0080] S102, input the association information of each first content, the association information of each second content, and the prompt word into the large language model to obtain the interest transfer prediction result of the target user.

[0081] The interest migration prediction result of the target user includes a target content type that the target user is interested in at the current time; and the prompt word is used to instruct the large language model to detect the interest migration of the target user from the second historical period to the first historical period according to the difference between the association information of each first content and the association information of each second content.

[0082] It can be understood that the text input by the server to the large language model can include two parts: the association information of each first content and the association information of each second content, and the prompt word. For example, the server can fill in the association information of each first content and the association information of each second content into a preset prompt word template to obtain the text input to the large language model.

[0083] The large language model can be a language model trained by using large-scale text data. The large language model has rich language knowledge, understanding ability and generation ability, can recognize the semantics of the input content, and can generate text for responding to the input content.

[0084] The large language model can recognize the semantics of the description text of the first content and the semantics of the description text of the second content, and according to the instruction of the prompt word, analyze the difference between the association information of each first content and the association information of each second content according to the recognized semantics, and detect the interest migration of the target user from the second historical period to the first historical period.

[0085] The prompt word is also used to represent that each first content is content whose browsing time is in the first historical period, and each second content is content whose browsing time is in the second historical period. The prompt word can also include a task description text for instructing the large language model to complete a task, for example, the task description text can be "detecting the interest of the target user and outputting a target content type that the target user is currently interested in". Correspondingly, the current interest migration prediction result output by the large language model according to the prompt word can be the text "the target content type that the target user is currently interested in is science fiction type".

[0086] It can be understood that, in order to instruct the large language model to more accurately detect the interest, on the basis of the prompt word including the association information of each first content and the association information of each second content, the prompt word can also include other information.

[0087] In the first implementation manner, the prompt word also includes a browsing time at which the target user browses each first content and a browsing time at which the target user browses each second content; the prompt word is also used to instruct the large language model to detect the interest based on the input description text of the specified content; and a time length between the browsing time of the specified content and the current time is less than a preset time length.

[0088] It can be understood that, as the content type of interest of the target user can change at any time, if the length of time between the browsing time of the target user browsing a content and the current time is large, that is, the content is browsed a long time ago, the accuracy of the content type of interest of the target user that can be reflected by the content is small.

[0089] The prompt word can also include the browsing time of the target user browsing each first content and the browsing time of the target user browsing each second content. The large language model can calculate the length of time between the browsing time of each content and the current time, and determine the importance of the content for the interest migration prediction result. The greater the importance, the greater the influence of the associated information of the content on the interest migration prediction result. Specifically, the inverse of the calculated length of time between the browsing time of each content and the current time can be taken as the importance of the content, or the negative of the length of time between the browsing time of each content and the current time can be taken as the importance of the content. In this regard, the embodiments of the present application are only illustrative and are not specifically limited.

[0090] Then the server determines the associated information of the specified content with an importance greater than the importance threshold value through the large language model, and then determines the content type of interest of the target user at the current time according to the semantics of the description text of the specified content.

[0091] The task description text in the prompt word can be "analyze the content type of interest of the target user at the current time according to the content browsed by the target user in [the specified time length]". For example, if the specified time length is 3 days, the task description text in the prompt word can be "analyze the content type of interest of the target user at the current time according to the content browsed by the target user in [three days]". Correspondingly, the current interest migration prediction result output by the large language model can be "the target user's completion rate of science fiction content in the past three days increases by 3 times, and the target user may be interested in science fiction content at the current time". The above-mentioned importance threshold value can be inversely related to the specified time length, for example, the specified time length can be the inverse or negative of the importance threshold value.

[0092] The specified time length can be a time length specified by a technician, or the specified time length can be calculated according to a time decay weighting formula. Specifically, the time decay weighting formula can be T=(ln 10)÷α; T represents the specified time length, and α represents a preset weight parameter, for example, if α=0.5 / day, T=4.6, and the specified time length is 4.6 days. It should be noted that the specified time length is not associated with the preset time, the first time, and the second time.

[0093] In the embodiment of the present application, the prompt word can also include a browsing time, and the prompt word can instruct the large language model to analyze the description text of the content close to the current time, avoid analyzing the description text of the content far from the current time, reduce the historical noise interference caused by the sparse behavior of the target user in browsing the content a long time ago, analyze the more real-time interest of the target user, improve the real-time performance of the user interest detection, and further improve the accuracy of the target content type.

[0094] In the second implementation manner, the prompt word further includes a similarity description text; and the prompt word is further used to instruct the large language model to perform interest detection based on the input similarity description text.

[0095] It can be understood that the similarity description text can represent the similarity of each first content and each second content. The specific acquisition manner of the similarity description text and the manner of performing interest detection based on the similarity description text will be described in detail in subsequent embodiments. The large language model can combine the semantics of the association information of each content and the semantics of the similarity description text to perform reasoning, avoid the lack of semantic information leading to insufficient understanding of the large language model, and output accurate interest transfer prediction results through rich semantic information.

[0096] In the third implementation manner, the prompt word can also simultaneously include the browsing time of the target user in browsing each first content, the browsing time of the target user in browsing each second content, and the similarity description text. The prompt word can instruct the large language model to perform interest detection based on the input similarity description text and the description text of the specified content. The large language model can analyze the more real-time interest of the target user while combining the browsing time, and ensure the output of accurate interest transfer prediction results by combining more rich semantic information.

[0097] In one implementation manner, the association information of one content further includes at least one of the following: a type of a browsing device used by the target user to browse the content, a type of a time period in which the target user browses the content, a browsing manner type of the target user in browsing the content, and a duration of the target user in browsing the content; and the browsing manner type of the target user in browsing the content is a continuous watching manner or a non-continuous watching manner.

[0098] It can be understood that the type of the browsing device used by the target user to browse one content can be a mobile phone or a television, etc.; the type of the time period in which the target user browses one content can be daytime or nighttime, etc. The browsing manner type of the target user in browsing one content is a continuous watching manner or a non-continuous watching manner. The continuous watching manner means that the target user does not have a pause behavior exceeding a preset duration in the process of browsing the content, for example, the preset duration can be 2 hours. The non-continuous watching manner means that the target user has a pause behavior exceeding a preset duration in the process of browsing the content.

[0099] In the prior art, the server can count the click rate of the user on the pushed content in each historical time period (i.e., taking each historical time period as a sliding window to count the user behavior index), and then, for a content type, calculate the year-on-year decline rate of the click rate of the user on the content of the content type in the last two historical time periods. If the calculated year-on-year decline rate reaches a threshold, the server can determine that the user is no longer interested in the content of the content type. However, if the user browses less content in a historical time period, the browsed content will be affected by the user's accidental behavior (for example, the user accidentally clicks on the content that is not interested), resulting in that the calculated year-on-year decline rate of the click rate cannot truly reflect the change of the content type that the user is interested in, and the calculated year-on-year decline rate of the click rate will incorrectly determine that the user's interest has changed. That is, the prior art cannot identify the accidental behavior fluctuation of the user.

[0100] In the embodiment of the present application, the duration of the target user browsing a content can reflect whether the target user is interested in the content, specifically, the duration of the target user browsing a content is positively correlated with the degree of interest of the target user in the content. The server inputs the duration of the target user browsing a content into a large language model, and the large language model can determine whether the duration of the target user browsing a content belongs to normal browsing behavior, for example, if the duration of the target user browsing a content is less than 5 seconds, it can be determined that the browsing of the target user for the content is accidental behavior; if the duration of the target user browsing a content is greater than 5 seconds, it can be determined that the target user is browsing content according to the content type that the target user is interested in, which belongs to the normal browsing behavior of the target user. The large language model can use the associated information of the content browsed by the normal browsing behavior for interest detection, and does not use the associated information of the content browsed by the accidental behavior for interest detection.

[0101] The embodiment of the present application identifies whether the target user browsing a content belongs to accidental behavior through the duration of the target user browsing a content, avoids the influence of accidental behavior on identifying the content type that the target user is interested in, and improves the accuracy of the determined content type that the target user is currently interested in.

[0102] The browsing behavior data of the target user browsing the content can be collectively referred to as the type of browsing equipment used by the target user to browse the content, the type of time period in which the target user browses the content, the type of way in which the target user browses the content, and the length of time for which the target user browses the content. The type of browsing equipment used by the target user to browse the content and the type of way in which the target user browses the content can be regarded as spatial dimension information. The type of time period in which the target user browses the content and the length of time for which the target user browses the content can be regarded as time dimension information. The behavior log of the target user can record the browsing behavior data of each content browsed by the target user. The server can obtain the browsing behavior data of each content from the behavior log of the target user.

[0103] It can be understood that the large language model can combine the browsing behavior data containing the time dimension information and the spatial dimension information and the description text representing the semantic dimension information to infer, determine the content type currently interested by the target user through rich information, and improve the accuracy of the determined content type currently interested by the target user.

[0104] For example, the browsing equipment used by the target user to browse a second content is a computer, and the description text of the second content represents that the second content belongs to the job skill type. The browsing equipment used by the target user to browse a first content is a mobile phone, and the description text of the first content represents that the first content belongs to the cooking tutorial type. The large language model can infer that the target user used to learn job skills using a computer in the past and used to learn cooking using a mobile phone at present. The large language model can determine, based on its own language knowledge, that it is reasonable behavior to learn job skills using a computer and to learn cooking using a mobile phone. Therefore, the accuracy of updating the content type interested by the target user from the job skill type to the cooking tutorial type is high.

[0105] The time period type in which the target user browses a second content is daytime, and the description text of the second content represents that the second content belongs to the knowledge education type. The time period type in which the target user browses a first content is nighttime, and the description text of the first content represents that the first content belongs to the humor entertainment type. The large language model can infer that the target user used to learn knowledge in the daytime in the past and used to browse easy content at night at present. The large language model can determine, based on its own language knowledge, that it is reasonable behavior to learn knowledge in the daytime and to browse easy content at night. Therefore, the accuracy of updating the content type interested by the target user from the knowledge education type to the humor entertainment type is high.

[0106] The browsing manner type of the target user browsing a second content is continuous watching manner, and the description text of the second content represents that the second content belongs to short video type; the browsing manner type of the target user browsing a first content is non-continuous watching manner, and the description text of the first content represents that the first content belongs to long video type; the large language model can infer that the target user used to continuously watch short videos and now watches long videos in a non-continuous manner, and the large language model can determine, in combination with the language knowledge of the large language model, that continuously watching short videos is a reasonable behavior and watching long videos in a non-continuous manner is a reasonable behavior, so the accuracy of updating the content type interested by the target user from short videos to long videos is higher.

[0107] S103, in the candidate content, determine the content belonging to the target content type as the content for pushing to the target user.

[0108] It can be understood that the candidate content can be each content stored in advance, each content can be stored in association with the content type to which the content belongs, and the server can search for the content of the target content type in each candidate content. After determining the content, the server can generate a to-be-pushed content queue containing the determined content, and push each content in the to-be-pushed content queue to the client logged in by the account of the target user.

[0109] The embodiment of the application can identify the semantic difference between the description text of the first content and the description text of the second content by means of the large language model, can identify the change of the content type of the content browsed by the target user from the second historical period to the first historical period by means of the text reasoning capability of the large language model, can obtain the change trend of the content browsed by the target user, can infer the change of the content type interested by the target user at the current moment, can accurately determine the content type interested by the target user at the current moment, and can further screen the content conforming to the content type interested by the target user at the current moment, thereby improving the accuracy of the content determined to be pushed to the user.

[0110] In an embodiment, as shown in Figure 2 based on the content screening method shown in Figure 1 the method further includes step S201, and step S102 includes step S1021.

[0111] S201, obtaining similarity description text for describing the similarity between each first content and each second content.

[0112] It can be understood that the server can calculate the similarity between each first content and each second content, and generate similarity description text for describing the similarity.

[0113] In an implementation, the server can generate a first information sequence containing the association information of each first content and a second information sequence containing the association information of each second content, and then compare the similarity of the first information sequence and the second information sequence to obtain a similarity description text describing the similarity of each first content and each second content. This implementation will be described in detail in subsequent embodiments.

[0114] In another implementation, the server can count the frequency of each text in the association information of each first content and the frequency of each text in the association information of each second content, determine the text in the association information of each first content with a frequency greater than a preset frequency threshold (i.e., the text belonging to the first content) and the text in the association information of each second content with a frequency greater than the preset frequency threshold (i.e., the text belonging to the second content), and compare the similarity of the text belonging to the first content and the text belonging to the second content to obtain a similarity description text describing the similarity of each first content and each second content. This embodiment of the present application is only illustrative and is not limited.

[0115] S1021, input the association information of each first content, the association information of each second content, the similarity description text, and the prompt word into the large language model to obtain an interest transfer prediction result of the target user.

[0116] The prompt word is further used to instruct the large language model to output an interest transfer prediction result representing that the interest of the target user has not changed when the similarity description text represents that the similarity of each first content and second content is greater than a similarity threshold, and output an interest transfer prediction result representing that the interest of the target user has changed when the similarity description text represents that the similarity of each first content and second content is not greater than the similarity threshold.

[0117] It can be understood that the similarity description text can represent whether each first content and each second content is similar. If the similarity description text represents that the similarity of each first content and second content is not greater than the similarity threshold, the large language model can determine that each first content and each second content is not similar, and thus can determine that each first content and each second content has a difference, and the interest of the target user has changed. If the similarity description text represents that the similarity of each first content and second content is greater than the similarity threshold, the large language model can determine that each first content and each second content is similar, and thus can determine that each first content and each second content has no difference, and the interest of the target user has not changed.

[0118] In an implementation manner, the interest migration prediction result output by the large language model can further include a reason for judging whether the interest of the target user changes, so as to help the technical personnel understand the judgment basis of whether the interest of the target user changes, provide more information for the technical personnel, and improve the experience of the technical personnel. For example, the reason can include the description text representing the low or high similarity.

[0119] The embodiment of the present application determines whether each first content is similar to each second content by using the similarity description text to assist the large language model, and then accurately determines whether the interest of the user changes.

[0120] Figure 3 The flowchart for determining the similarity description text in the content screening method provided by the embodiment of the present application is shown in FIG. 2. Figure 3 As shown in FIG. 2, step S201 includes steps S2011-S2014.

[0121] S2011, the associated information of each first content is sorted in the order of the browsing time, and a first information sequence is obtained.

[0122] It can be understood that, since the associated information of each first content in the first information sequence is arranged in the order of the browsing time, the first information sequence can also represent the change of each first content browsed by the target user over time, and the change can be analyzed by the interest prediction model to extract a vector representing the change of the interest of the user over time.

[0123] In an implementation manner, the associated information of the first content can include the description text of the first content. In another implementation manner, the associated information of the first content can include the description text of the first content and the browsing behavior data, so that the first information vector of the first content generated subsequently includes rich features, and the first information vector can more accurately represent the first content, thereby improving the accuracy of the similarity description text.

[0124] S2012, the first information sequence is input into a pre-trained interest prediction model, and a first information vector output by a feature extraction layer of the interest prediction model is obtained.

[0125] The interest prediction model is trained based on a sample information sequence and a true value; the sample information sequence is obtained by sorting the associated information of each sample content in the order of the browsing time of the sample user browsing each sample content; the true value represents the content type interested by the sample user; the interest prediction model further includes an output layer; and the output layer is used for normalizing the information vector output by the feature extraction layer to obtain a predicted content type.

[0126] It can be understood that the interest prediction model can be a model constructed based on a Transfomer network structure, the feature extraction layer in the interest prediction model can analyze the dependency relationship between each first content in the first information sequence based on the self-attention mechanism of the Transfomer network structure, that is, analyze the change of each first content over time implied in the first information sequence, and then extract an information vector representing the change of each first content over time. The output layer in the interest prediction model can normalize the information vector output by the feature extraction layer to obtain the predicted content type. For example, the interest prediction model can be a recurrent neural network (RNN) model or a long short-term memory network (LSTM) model.

[0127] In the embodiment of the application, the information vector output by the feature extraction layer in the interest prediction model can be obtained to compare the similarity of each first content and each second content subsequently, provide more rich semantic information for large language model reasoning, and improve the accuracy of interest transfer prediction results.

[0128] The sample information sequence includes the association information of a preset number of sample contents, each sample content is browsed by a sample user, and the association information of each sample content in the sample information sequence is arranged in the order of the browsing time of the sample user browsing each sample content.

[0129] The true value represents the content type that the sample user is interested in, and the true value can be the content type that the sample user is interested in, which is labeled by a technician or a sample user. For example, a technician can manually analyze the association information of the specified true value content to determine the content type to which the specified true value content belongs as the true value, wherein the specified true value content can be the first content browsed by the sample user after browsing each sample content. Specifically, the technician can obtain the sample content browsed by the sample user within a continuous period of time, take the content type of the last browsed content in the obtained sample content as the true value, and construct the sample information sequence with the association information of the other contents in the obtained sample content except the last browsed content. For example, the sample information sequence includes: the association information of the first sample content browsed by the sample user, the association information of the second sample content browsed by the sample user, and the association information of the third sample content browsed by the sample user within a continuous period of time; the specified true value content is the fourth sample content browsed by the sample user, and the true value is the content type of the fourth sample content browsed by the sample user. For example, the sample user can actively label the content type of the first content browsed after browsing each sample content to obtain the true value.

[0130] In an implementation, the interest prediction model is trained by the following steps.

[0131] A1, obtaining a sample information sequence, a true value, and an initial structure of the interest prediction model.

[0132] A2, inputting the sample information sequence into the interest prediction model of the initial structure to obtain a predicted content type.

[0133] A3, calculating a loss value based on the predicted content type and the true value.

[0134] A4, adjusting network parameters of the interest prediction model of the initial structure based on the loss value until a preset convergence condition is reached to obtain the interest prediction model.

[0135] It can be understood that the electronic device for training the interest prediction model can be a server, or other electronic devices that are not servers, such as other electronic devices that are specifically used for training the interest prediction model.

[0136] The electronic device can obtain pre-stored association information of each sample content, and then sort the association information of each sample content according to the chronological order of the browsing time of each sample content to obtain a sample information sequence. The electronic device can obtain pre-labeled true values and pre-constructed interest prediction models of the initial structure.

[0137] Then the electronic device can input the sample information sequence into the interest prediction model of the initial structure to obtain a predicted content type output by the interest prediction model of the initial structure; then calculate a loss value based on the predicted content type and the true value; adjust the network parameters of the interest prediction model of the initial structure based on the loss value until a preset convergence condition is reached to obtain the interest prediction model. Since the network parameters of the feature extraction layer are also adjusted in the process of adjusting the network parameters of the interest prediction model, the prediction result of the trained interest prediction model is closer to the true value, that is, the trained interest prediction model can output accurate prediction results, and the feature extraction layer in the trained interest prediction model can accurately extract the first information vector of the first information sequence, and the first information vector can accurately represent the characteristics of the content browsed by the target user in the first historical period.

[0138] In an implementation, the interest prediction model further includes an encoding layer, which can encode the browsing behavior data in the association information of each first content and the description text of the content to obtain a feature value representing the association information of the first content, facilitating subsequent feature extraction by the feature extraction layer. Illustratively, the output of the encoding layer can be used as the input of the feature extraction layer.

[0139] Exemplarily, the encoding layer can use a One-Hot algorithm to encode the browsing behavior data in the associated information of the first content to obtain a feature value of the browsing behavior data; the encoding layer includes a Contrastive Language-Image Pretraining (CLIP) unit, and the encoding layer can use the CLIP unit to encode the description text of the first content and the cover picture of the first content to obtain a feature value of the description text of the first content.

[0140] S2013, obtain a second information vector corresponding to a second information sequence.

[0141] The second information sequence is obtained by sorting the associated information of each second content according to the chronological order of the browsing time; and the second information vector is obtained by inputting the second information sequence into the interest prediction model.

[0142] It can be understood that the second information vector can be obtained in the same manner as the first information vector; in an implementation manner, the second information vector can be obtained according to the above steps S2011-S2012.

[0143] In another implementation manner, the first historical period is a period from a first time to a current time, the second historical period is a period from a second time to the first time, the first time is a time when the interest of the target user is last detected, the second time is a time when the interest of the target user is last but one detected, and the server can store the first information vector obtained by detecting the interest of the target user each time. When the second information vector of the second historical period is needed, the first information vector obtained by detecting the interest of the target user last time can be directly obtained to obtain the second information vector, so that the second information vector is not generated again according to the steps S2011-S2012, and the running resources are saved.

[0144] S2014, generate a similarity description text for describing the similarity between the first information vector and the second information vector to obtain a similarity description text for describing the similarity between each first content and each second content.

[0145] It can be understood that the server can calculate the similarity between the first information vector and the second information vector. For example, the server can calculate the cosine similarity between the first information vector and the second information vector, or calculate the Euclidean distance between the first information vector and the second information vector. The similarity between the first information vector and the second information vector can represent the similarity between each first content and each second content. The server can generate a description text containing the calculated similarity, which is used to describe the similarity between the first information vector and the second information vector. For example, the description text can be "the similarity between the first information vector and the second information vector is 10%".

[0146] In another implementation, the server can input the description text of the first information vector and the description text of the second information vector into the large language model as a similarity description text for describing the similarity between the first information vector and the second information vector, and calculate the similarity between the first information vector and the second information vector through the large language model. The description text of the first information vector can include the first information vector and a text for describing that the first information vector is obtained using the associated information of each first content; and the description text of the second information vector can include the second information vector and a text for describing that the second information vector is obtained using the associated information of each second content.

[0147] In the embodiment of the application, the similarity between the first information vector and the second information vector represents the similarity between each first content and each second content, and the similarity description text is obtained. The large language model can combine the semantics of the associated information of each content and the semantics of the similarity description text to perform reasoning, and output an accurate interest transfer prediction result through rich semantic information.

[0148] In an embodiment, when the first historical period is a period from the first time to the current time, and the second historical period is a period from the second time to the first time, the first time is the last time when interest detection needs to be performed on the target user, and the second time is the time when interest detection needs to be performed on the target user. The prompt word can also include the historical content type in the interest transfer prediction result obtained at the first time. That is, the target content type (the target content type at the first time) in the interest transfer prediction result obtained by performing interest detection at the first time is used as the historical content type. In the case where the interest of the target user has migrated, the interest transfer prediction result of the target user also includes a description text representing that the interested type of the target user is updated from the historical content type to the target content type; that is, the interest transfer prediction result also includes a description text representing that the interest of the target user is updated from the target content type at the first time to the target content type at the current time.

[0149] For example, Figure 4As shown, the method further comprises steps S401-S403, and step S103 comprises step S1031.

[0150] S401, in the pre-stored knowledge graph, determine a node representing a historical content type as a first node, and determine a node representing a target content type as a second node.

[0151] Among them, the nodes in the knowledge graph correspond one-to-one to content types, and the content types represented by two connected nodes are associated.

[0152] S402, from the knowledge graph, determine other nodes connected to the first node as third nodes, and determine other nodes connected to the second node as fourth nodes.

[0153] S403, calculate the similarity between the content types represented by the third nodes and the fourth nodes as the to-be-utilized similarity.

[0154] S1031, if the to-be-utilized similarity is greater than a preset threshold, determine the content belonging to the target content type in the candidate content as the content for pushing to the target user.

[0155] It can be understood that the current interest migration prediction result also includes a description text representing that the target user's interested type is updated from the historical content type to the target content type, which can represent the target user's interest migration. For example, the description text can be "the target user's interest may be converted from content types A and B to content types C and D", or the description text can be "the target user's interest may be converted from content types A, B and C to content types A, C and D". Since the prompt word contains the association information of each second content (which can reflect the historical interest) and the association information of each first content (which can reflect the current interest), the large language model can analyze whether the target user's interest changes and the specific change direction by combining the association information of each second content and the association information of each first content.

[0156] In an implementation manner, the current interest migration prediction result can further include a description text representing the reason for the large language model to obtain the result (i.e., a thinking chain representing the reasoning process of the large language model). For example, the interest migration prediction result output by the large language model can be "the target user's interest may be converted from content types A and B to content types C and D". Since the interest migration prediction result further includes a description text representing the reason for the large language model to obtain the result, a technical personnel can understand why the target user's interest may migrate by viewing the interest migration prediction result, improve the explainability of the interest migration prediction result, and improve the experience of the technical personnel.

[0157] It can be understood that if the interest prediction model outputs a predicted content type, which is only a result representing the content type, the specific user interest migration direction cannot be determined only according to the interest prediction model, and there is no description text representing the user interest migration and the reason for obtaining the result in the interest migration prediction result of the large language model, and the explainability is low. The embodiment of the present application can output the description text representing the user interest migration and the reason for obtaining the result through the large language mode, improve the explainability, and the technical personnel can more intuitively understand the interest of the target user and improve the experience of the technical personnel.

[0158] In order to facilitate the management of content, the technical personnel will establish a knowledge graph according to the content types of each content. Among them, the nodes in the knowledge graph correspond one by one to the content types, for example, one node can represent the content type as a costume drama, and another node can represent the content type as a historical documentary. The content types represented by two connected nodes are associated, for example, the node representing the content type as a costume drama can be connected with the node representing the content type as history, and the costume drama type is associated with the history type, and the node representing the content type as a historical documentary can be connected with the node representing the content type as history.

[0159] The historical content type is different from the target content type, and the historical content type can be inferred by the large language model from the associated information of each second content. The server can determine whether the historical content type is associated with the target content type according to the content types associated with the historical content type and the content types associated with the target content type, to verify whether the interest migration of the target user is reasonable.

[0160] The server can determine whether the historical content type is associated with the target content type through the knowledge graph. Specifically, the server can determine a node representing the historical content type as a first node and a node representing the target content type as a second node in the pre-stored knowledge graph; then determine other nodes connected with the first node as third nodes (i.e., determine the content types associated with the historical content type) and other nodes connected with the second node as fourth nodes (i.e., determine the content types associated with the target content type) in the knowledge graph.

[0161] The similarity between the content types represented by the third nodes and the fourth nodes (i.e., the to-be-utilized similarity) can represent the similarity between the content types associated with the historical content type and the content types associated with the target content type, and reflect whether the user interest has migrated. The to-be-utilized similarity can also be referred to as the interest migration confidence.

[0162] In an implementation manner, the server can input the prompt word containing the content type represented by the third node and the content type represented by the fourth node into the large language model, the prompt word instructing the large language model to identify the similarity of the input text, to obtain the similarity between the content type represented by the third node and the content type represented by the fourth node.

[0163] In another implementation manner, the server can convert the content type represented by the third node into a semantic vector, and convert the content type represented by the fourth node into a semantic vector, and calculate the similarity of the two semantic vectors. Specifically, if the content type represented by the third node is multiple and the content type represented by the fourth node is multiple, the server can splice the content type represented by each third node and then convert it into a first semantic vector, splice the content type represented by each fourth node and then convert it into a second semantic vector, and calculate the similarity of the first semantic vector and the second semantic vector; the server can also convert the content type represented by each third node into a semantic vector of the third node, convert the content type represented by each fourth node into a semantic vector of the fourth node, calculate the semantic vector of each third node and the semantic vector of each fourth node, obtain multiple similarities, and then calculate the average of the multiple similarities to obtain the similarity between the content type represented by each third node and the content type represented by each fourth node.

[0164] If the similarity to be used is greater than the preset threshold, the server can determine that the similarity between the content type associated with the historical content type and the content type associated with the target content type is high, the historical content type and the target content type are associated, the interest migration of the target user is reasonable, and is not an accidental behavior of sudden change. The server can determine the content for pushing to the target user according to the target content type in the interest migration prediction result.

[0165] If the similarity to be used is not greater than the preset threshold, the server can determine that the similarity between the content type associated with the historical content type and the content type associated with the target content type is not high, the historical content type and the target content type are not associated, and the interest migration of the target user can be unreasonable. The server will not determine the content for pushing to the target user according to the target content type in the interest migration prediction result. For example, the server will send a prompt message to the technical personnel to prompt the technical personnel to detect the association information of each first content and manually analyze the interest when the similarity to be used is not greater than the preset threshold.

[0166] For example, the first node represents the costume drama type, the second node represents the historical documentary type, and the third node represents the history type. The first node is connected to the third node, and the second node is connected to the third node. The content type associated with the history content type and the content type associated with the target content type are the same content type, and the similarity is 100%. The interest migration of the target user is reasonable. For another example, the first node represents the costume drama type, the second node represents the science fiction type, the third node represents the history type, and the fourth node represents the science and technology popularization type. The first node is connected to the third node, and the second node is connected to the fourth node. The similarity between the history type represented by the third node and the science and technology popularization type represented by the fourth node is low. The interest migration of the target user may be unreasonable.

[0167] In the embodiment of the application, the content type associated with the history content type and the content type associated with the target resource are determined by the knowledge graph to determine whether the user interest migration is reasonable. The structured information of each node and the content type represented by the knowledge graph can be used to accurately determine the content type associated with the content type, avoid the influence of the interest migration prediction result of unreasonable user interest migration on the accuracy of the content to be pushed, and improve the user experience.

[0168] In an embodiment, based on the content screening method shown in Figure 1 As shown in Figure 5 The method further includes steps S501-S502.

[0169] S501, in the selected content, determine the content belonging to the history content type.

[0170] S502, push the content belonging to the history content type and the content belonging to the target content type to the target user.

[0171] It can be understood that the server can determine the content belonging to the history content type, and the specific determination manner is similar to the determination of the content belonging to the target content type. The target content type reflects the potential interest of the target user inferred by the large language model, and the history content type reflects the historical interest of the target user. When pushing the content containing the history content type and the target content type to the target user, the type of the pushed content can be avoided. In the exploration of the potential interest of the target user, the historical interest of the target user is satisfied, and the user experience is improved.

[0172] In an implementation manner, the server can push the content belonging to the history content type and the content belonging to the target content type to the target user according to a preset weight. For example, the preset weight can be 50%.

[0173] In another implementation, the server can more intelligently calculate the weight of the target content type, and push according to the calculated weight. The specific implementation steps are as follows:

[0174] C1, obtaining the interest migration prediction results obtained respectively when the target user is detected for a specified number of times in the past as historical interest migration prediction results.

[0175] It can be understood that each time the target user is detected for interest, the server will generate an interest migration prediction result. The server can obtain a specified number of interest migration prediction results generated before the current time as historical interest migration prediction results. The specified number can be 2, 3, or 5, etc.

[0176] C2, for any one of the historical interest migration prediction results and the target user's interest migration prediction results, if the interest migration prediction result contains the target content type, it is determined that the interest migration prediction result is a to-be-utilized interest migration prediction result.

[0177] C3, calculating the proportion of the to-be-utilized interest migration prediction result in the historical interest migration prediction results and the target user's interest migration prediction results as the weight of the target content type.

[0178] It can be understood that the server can identify whether any one of the historical interest migration prediction results and the current interest migration prediction result contains the target content type. If any one of the interest migration prediction results contains the target content type, it is determined that the target user has historically been interested in the content type that appears in the target content type. For example, if an interest migration prediction result contains a description text representing that the content type currently interested by the target user is the target content type, it is determined that the interest migration prediction result contains the target content type. If the target content type is multiple content types, the server can identify that the interest migration prediction result contains one of the multiple content types, that is, it is determined that the interest migration prediction result contains the target content type.

[0179] The server can take the interest migration prediction result as a to-be-utilized interest migration prediction result, count the number of to-be-utilized interest migration prediction results, and the total number of historical interest migration prediction results and current interest migration prediction results, calculate the ratio of the number of to-be-utilized interest migration prediction results to the total number, and obtain the proportion of the to-be-utilized interest migration prediction result in the historical interest migration prediction results and the current interest migration prediction results. For example, the specified number is 3, the number of to-be-utilized interest migration prediction results is 2, and the proportion of to-be-utilized interest migration prediction results in historical detection and current interest migration prediction results is 2 / (3+1)=0.5.

[0180] C4, according to the weight of the target content type, push the content belonging to the historical content type and the content belonging to the target content type to the target user.

[0181] It can be understood that the proportion of the interest migration prediction result to be used in the historical detection and the current interest migration prediction result is positively correlated with the degree of interest of the target user to the target content type.

[0182] The server can push multiple contents to the target user according to the weight of the target content type, and the proportion of the content belonging to the target content type in the multiple contents is the weight of the target content type, and the proportion of the content belonging to the historical content type is one minus the weight of the target content type.

[0183] If the interest migration prediction result obtained by the interest detection before the current time contains the target content type, the degree of interest of the target user to the target content type is increasing, and the server can calculate a higher weight of the target content type, and the proportion of the content of the target content type pushed to the target user becomes larger to adapt to the current interest of the target user. If the interest migration prediction result obtained by the interest detection before the current time does not contain the target content type, the degree of interest of the target user to the target content type is decreasing, and the server can calculate a smaller weight of the target content type, and the proportion of the content of the target content type pushed to the target user becomes smaller, and the number of contents with a smaller degree of interest pushed to the target user is reduced.

[0184] In the embodiment of the application, the server can push the content belonging to the historical content type and the content belonging to the target content type to the target user according to the proportion of the interest migration prediction result to be used in the history in the current interest migration prediction result, which conforms to the current interest change trend of the target user, smooth updates the content to be pushed, avoids sudden switching of the content type of the content to be pushed, and avoids the occurrence of content with a low degree of interest of the target user, thereby improving the user experience.

[0185] In one implementation mode, the server can update the user portrait of the target user according to the current interest migration prediction result. For example, the target user's interested content type can contain a historical content type, and if the similarity to be used is greater than a preset threshold, the server adds the target content type to the target user's interested content type; then the interest migration prediction results obtained by a specified number of interest detections are acquired, the number of interest migration prediction results containing the target content type is counted, and if the counted number reaches a threshold, the historical content type in the target user's interested content type is deleted to update the user portrait of the target user.

[0186] In the embodiment of the present application, after adding the target content type in the current interest migration prediction result to the content type of interest to the target user, the server does not immediately delete the historical content type in the content type of interest to the target user, but deletes the historical content type in the content type of interest to the target user after the number of target content types appearing in the statistical historical interest migration prediction result reaches a threshold, so as to update the user portrait in a smoothing strategy, avoid user portrait mutation, ensure stable update of the user portrait in the case of gradual change of user interest, and improve the accuracy of the user portrait.

[0187] Figure 6 A structural schematic diagram of a content screening device provided by the embodiment of the present application is shown in FIG. 1. Figure 6 As shown in the figure, the device comprises:

[0188] The first acquisition module 610 is configured to acquire the associated information of the first content browsed by the target user in a first historical period and the associated information of the second content browsed by the target user in a second historical period when interest detection of the target user is needed, wherein the second historical period is before the first historical period; the associated information of one content includes the description text of the content.

[0189] The input module 620 is configured to input the associated information of each first content, the associated information of each second content and a prompt word into a large language model to obtain the interest migration prediction result of the target user, wherein the interest migration prediction result of the target user includes a target content type of interest to the target user at the current moment; the prompt word is used to instruct the large language model to detect the interest migration of the target user from the second historical period to the first historical period according to the difference between the associated information of each first content and the associated information of each second content.

[0190] The determination module 630 is configured to determine the content belonging to the target content type in the candidate content as the content for pushing to the target user.

[0191] Optionally, the apparatus further comprises a second obtaining module configured to obtain a similarity description text for describing the similarity between each first content and each second content; and the input module 620 is specifically configured to input the association information of each first content, the association information of each second content, the similarity description text, and a prompt word into the large language model to obtain the interest shift prediction result of the target user; wherein the prompt word is further used to instruct the large language model to output an interest shift prediction result representing that the interest of the target user has not changed when the similarity description text represents that the similarity between each first content and second content is greater than a similarity threshold, and output an interest shift prediction result representing that the interest of the target user has changed when the similarity description text represents that the similarity between each first content and second content is not greater than the similarity threshold.

[0192] Optionally, the second obtaining module comprises: a sorting unit configured to sort the association information of each first content according to the chronological order of the browsing time to obtain a first information sequence; an input unit configured to input the first information sequence into a pre-trained interest prediction model to obtain a first information vector output by a feature extraction layer in the interest prediction model; wherein the interest prediction model is trained based on a sample information sequence and a true value; the sample information sequence is obtained by sorting the association information of each sample content according to the chronological order of the browsing time of the sample user to browse each sample content; the true value represents the content type that the sample user is interested in; the interest prediction model further comprises an output layer; the output layer is configured to normalize the information vector output by the feature extraction layer to obtain a predicted content type; an obtaining unit configured to obtain a second information vector corresponding to a second information sequence; wherein the second information sequence is obtained by sorting the association information of each second content according to the chronological order of the browsing time; and the second information vector is obtained by inputting the second information sequence into the interest prediction model; and a generating unit configured to generate a description text for describing the similarity between the first information vector and the second information vector to obtain a similarity description text for describing the similarity between each first content and each second content.

[0193] Optionally, the first historical period is a period from a first time point at which the interest of the target user is detected last time to a current time point; and the second historical period is a period from a second time point at which the interest of the target user is detected most recently before the first time point to the first time point.

[0194] Optionally, the prompt word further comprises: a historical content type in the interest migration prediction result obtained at the first time; in the case that the interest of the target user migrates, the interest migration prediction result of the target user further comprises: a description text representing that the interested type of the target user is updated from the historical content type to the target content type; the device further comprises: a node determination module, configured to determine, in a pre-stored knowledge graph, a node representing the historical content type as a first node, and determine a node representing the target content type as a second node; wherein the nodes in the knowledge graph correspond to content types one by one, and the content types represented by two connected nodes are associated; the node determination module is configured to determine, from the knowledge graph, other nodes connected to the first node as third nodes, and determine other nodes connected to the second node as fourth nodes; a calculation module is configured to calculate a similarity between the content types represented by the third nodes and the content types represented by the fourth nodes as a to-be-utilized similarity; the determination module 630 is specifically configured to, if the to-be-utilized similarity is greater than a preset threshold, determine, in the to-be-selected content, content belonging to the target content type as content for pushing to the target user.

[0195] Optionally, the device further comprises: a content determination module, configured to determine, in the to-be-selected content, content belonging to the historical content type; and a pushing module, configured to push the content belonging to the historical content type and the content belonging to the target content type to the target user.

[0196] Optionally, the device further comprises: a third acquisition module, configured to acquire historical interest migration prediction results obtained respectively when a specified number of times of interest detection on the target user are needed, as historical interest migration prediction results; a result determination module, configured to, for any one of the historical interest migration prediction results and the interest migration prediction result of the target user, if the interest migration prediction result contains the target content type, determine that the interest migration prediction result is a to-be-utilized interest migration prediction result; and a weight determination module, configured to calculate a proportion of the to-be-utilized interest migration prediction result in the historical interest migration prediction results and the interest migration prediction result of the target user as a weight of the target content type.

[0197] The pushing module is specifically configured to push the content belonging to the historical content type and the content belonging to the target content type to the target user according to the weight of the target content type.

[0198] Optionally, the prompt word further comprises a browsing time of the target user browsing each first content and a browsing time of the target user browsing each second content; the prompt word is used to instruct the large language model to determine the importance of each content for the interest migration prediction result according to the browsing time of the content, and determine the first content and the second content with an importance greater than a preset importance threshold, and detect the interest migration of the target user from the second historical period to the first historical period according to the determined association information of each first content and the difference between the determined association information of each second content; the time length of the browsing time of a content from the current time and the importance of the content for the interest migration prediction result are inversely related.

[0199] Optionally, the association information of a content further comprises at least one of the following: a type of a browsing device used by the target user to browse the content, a type of a time period in which the target user browses the content, a browsing mode type of the target user browsing the content, and a time length of the target user browsing the content; the browsing mode type of the target user browsing the content is continuous watching or non-continuous watching.

[0200] The embodiment of the present application can identify the semantic difference between the description text of the first content and the description text of the second content through the large language model, can identify the change of the content type of the content browsed by the target user from the second historical period to the first historical period by means of the text reasoning ability of the large language model, can obtain the change trend of the content browsed by the target user, and can infer the change of the content type interested by the target user at the current time, so as to accurately determine the content type interested by the target user at the current time, and then to filter out the content conforming to the content type interested by the target user at the current time, thereby improving the accuracy of the content determined to be pushed to the user.

[0201] The embodiment of the present application further provides an electronic device, as shown in the figure, Figure 7 The electronic device comprises a processor 701, a communication interface 702, a memory 703 and a communication bus 704,

[0202] The processor 701, the communication interface 702 and the memory 703 can communicate with each other through the communication bus 704, the memory 703 is used to store a computer program, and the processor 701 is used to execute the program stored in the memory 703 to realize the above-mentioned content filtering method.

[0203] The communication bus mentioned above can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The communication bus can be divided into an address bus, a data bus, a control bus, etc. For the convenience of representation, only one thick line is used in the figure, but it does not mean that there is only one bus or only one type of bus. The communication interface is used for communication between the terminal and other devices. The memory can include a Random Access Memory (RAM) and can also include a non-volatile memory, for example, at least one disk memory. Optionally, the memory can also be at least one storage device located away from the processor.

[0204] The processor mentioned above can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component.

[0205] In another embodiment provided by the present application, a computer readable storage medium is also provided, and the computer readable storage medium stores a computer program. The computer program is executed by a processor to implement the content screening method in any of the above embodiments. In another embodiment provided by the present application, a computer program product containing instructions is also provided, and when the computer program product is executed on a computer, the computer executes the content screening method in any of the above embodiments.

[0206] In the embodiments described above, all or some of the steps can be implemented by software, hardware or firmware, or any combination thereof. When implemented in software, all or some of the steps can be implemented in the form of one or more computer programs which are executable on a computer. The computer program can be stored in the main memory of a computer or in some other form of computer readable media, which can be accessed by the computer. The computer readable media can include computer storage media and communication media. The computer storage media includes any media that can be accessed by the computer, such as a read only memory (ROM), a random access memory (RAM), a magnetic disk storage medium, an optical storage medium, and the like. The communication media includes computer readable instructions, data structures, program modules and the like.

[0207] It should be noted that, in the present document, the terms such as first and second are used only to differentiate one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Also, the terms "comprising", "containing" or any other variant thereof are intended to cover a non-exclusive inclusion, such that a process, method, article or apparatus that comprises a list of elements does not only include those elements, but can also include other elements not expressly listed or inherent to such process, method, article or apparatus. Without more limitations, an element defined by the statement "comprising a" does not exclude the presence of additional identical elements in the process, method, article or apparatus that includes the element.

[0208] Each of the embodiments in the present specification is described in a related manner, and the same or similar parts between the embodiments can be referred to each other. Each embodiment focuses on the difference from other embodiments. In particular, for the device embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the part of the method embodiments.

[0209] The above merely provides the preferred embodiments of the application, and not intended to limit the protection scope of the application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the application shall fall within the protection scope of the application.

Claims

1. A content filtering method, characterized in that, The method includes: When it is necessary to detect the interests of a target user, the association information of the first content viewed by the target user in a first historical time period is obtained, and the association information of the second content viewed by the target user in a second historical time period is obtained; wherein, the second historical time period is before the first historical time period; the association information of a content includes: the descriptive text of the content; The association information of each first content, the association information of each second content, and the prompt words are input into the large language model to obtain the interest migration prediction result of the target user; wherein, the interest migration prediction result of the target user includes the type of target content that the target user is interested in at the current time; the prompt words are used to instruct the large language model to detect the interest migration of the target user from the second historical time period to the first historical time period based on the difference between the association information of each first content and the association information of each second content. Among the candidate content, content belonging to the target content type is determined as the content to be pushed to the target user.

2. The method according to claim 1, characterized in that, Before inputting the association information of each first piece of content, the association information of each second piece of content, and the prompt words into the large language model, the method further includes: Obtain similarity description text used to describe the similarity between each first piece of content and each second piece of content; The step of inputting the association information of each first piece of content, the association information of each second piece of content, and the prompt words into the large language model includes: The association information of each first content, the association information of each second content, the similarity description text, and the prompt words are input into the large language model to obtain the interest migration prediction result of the target user. The prompt words are further used to instruct the large language model to output an interest migration prediction result indicating that the target user's interest has not changed when the similarity description text indicates that the similarity between each first content and the second content is greater than a similarity threshold, and to output an interest migration prediction result indicating that the target user's interest has changed when the similarity description text indicates that the similarity between each first content and the second content is not greater than a similarity threshold.

3. The method according to claim 2, characterized in that, The step of obtaining similarity description text used to describe the similarity between each first piece of content and each second piece of content includes: The related information of each first content is sorted according to the order of browsing time to obtain the first information sequence; The first information sequence is input into a pre-trained interest prediction model to obtain a first information vector output by the feature extraction layer of the interest prediction model. The interest prediction model is trained based on a sample information sequence and ground truth values. The sample information sequence is obtained by sorting the association information of each sample content according to the order in which the sample user browses each sample content. The ground truth values ​​represent the content types that the sample user is interested in. The interest prediction model also includes an output layer. The output layer is used to normalize the information vector output by the feature extraction layer to obtain the predicted content type. Obtain the second information vector corresponding to the second information sequence; wherein, the second information sequence is obtained by sorting the associated information of each second content according to the order of browsing time; the second information vector is obtained by inputting the second information sequence into the interest prediction model; Generate descriptive text to describe the similarity between the first information vector and the second information vector, and obtain similarity descriptive text to describe the similarity between each first content and each second content.

4. The method according to claim 1, characterized in that, The first historical time period is the period from the first moment when the target user's interest was last detected to the current moment; the second historical time period is the period from the second moment when the target user's interest was last detected before the first moment to the first moment.

5. The method according to claim 4, characterized in that, The prompt words also include: the historical content type in the interest migration prediction results obtained at the first moment; In the event of a shift in the target user's interests, the target user's interest shift prediction result further includes: descriptive text representing that the target user's interest type has been updated from the historical content type to the target content type; Before determining, from the candidate content, content belonging to the target content type as content to be pushed to the target user, the method further includes: In the pre-stored knowledge graph, a node representing the historical content type is identified as the first node, and a node representing the target content type is identified as the second node; wherein, the nodes in the knowledge graph correspond one-to-one with the content types, and the content types represented by two connected nodes are associated. From the knowledge graph, other nodes connected to the first node are identified as the third node, and other nodes connected to the second node are identified as the fourth node; Calculate the similarity between the content type represented by the third node and the content type represented by the fourth node, and use this similarity as the similarity to be utilized; The step of determining, from the candidate content, content belonging to the target content type, as content to be pushed to the target user, includes: If the similarity to be utilized is greater than a preset threshold, then among the candidate content, content belonging to the target content type is determined as content to be pushed to the target user.

6. The method according to claim 5, characterized in that, The method further includes: Among the candidate content, identify the content that belongs to the aforementioned historical content type; Content belonging to the historical content type and content belonging to the target content type are pushed to the target user.

7. The method according to claim 6, characterized in that, Before pushing content belonging to the historical content type and content belonging to the target content type to the target user, the method further includes: Obtain the interest migration prediction results obtained from each of the specified number of historical interest detection tests on the target user, and use them as historical interest migration prediction results. For any interest migration prediction result between historical interest migration prediction results and the target user's interest migration prediction results, if the interest migration prediction result contains the target content type, then the interest migration prediction result is determined to be an interest migration prediction result to be utilized. Calculate the proportion of the interest migration prediction result to be utilized in the historical interest migration prediction results and the interest migration prediction results of the target user, and use it as the weight of the target content type; The step of pushing content belonging to the historical content type and content belonging to the target content type to the target user includes: Based on the weight of the target content type, content belonging to the historical content type and content belonging to the target content type are pushed to the target user.

8. The method according to any one of claims 1-7, characterized in that, The prompt also includes the browsing time of the target user when browsing each piece of first content, and the browsing time when browsing each piece of second content; The prompt words are used to instruct the large language model to determine the importance of each content to the interest migration prediction result based on the browsing time of each content, and to determine the first content and the second content whose importance is greater than a preset importance threshold. Based on the differences in the correlation information of each first content and the correlation information of each second content, the model detects the interest migration of the target user from the second historical time period to the first historical time period. The time elapsed between the browsing time of a content and the current time is inversely correlated with the importance of the content to the interest migration prediction result.

9. The method according to any one of claims 1-7, characterized in that, The associated information of a piece of content also includes at least one of the following: the type of browsing device used by the target user to browse the content, the type of time period during which the target user browses the content, the type of method by which the target user browses the content, and the duration of the target user browsing the content; The target user can view the content either continuously or non-continuously.

10. A content filtering device, characterized in that, The device includes: The first acquisition module is used to acquire, when it is necessary to detect the interests of a target user, the association information of a first content viewed by the target user in a first historical time period, and the association information of a second content viewed by the target user in a second historical time period; wherein, the second historical time period is before the first historical time period; the association information of a content includes: the descriptive text of the content; The input module is used to input the association information of each first content, the association information of each second content, and the prompt words into the large language model to obtain the interest migration prediction result of the target user; wherein, the interest migration prediction result of the target user includes the type of target content that the target user is interested in at the current time; the prompt words are used to instruct the large language model to detect the interest migration of the target user from the second historical time period to the first historical time period based on the difference between the association information of each first content and the association information of each second content; The determination module is used to determine, from the candidate content, the content that belongs to the target content type, as the content to be pushed to the target user.

11. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the steps of the method described in any one of claims 1-9.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the method described in any one of claims 1-9.

13. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method according to any one of claims 1-9.

Images