A video recommendation method and system based on a large language model, an electronic device, and a storage medium

Abstract

The application provides a video recommendation method and system based on a large language model, an electronic device, and a storage medium. At least one target user intention label data is obtained from a key-value database according to a recommendation request initiated by a target user. A first video item set is screened from a streaming media platform according to a first content label set. A second video item set is screened from each first video item according to the first content label set and a first video label set of each first video item. A third video item set is screened from each video item according to target user text feedback data. A fourth video item set is screened from each video item according to each preset recall algorithm, and the fourth video item set is fused with the second video item set and the third video item set to obtain a fifth video item set. A video recommendation list generated according to a sixth video item set obtained by sorting the fifth video item set based on a second content label set is used to recommend videos to the target user.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computers, more particularly, to a video recommendation method and system based on a large language model, an electronic device, and a storage medium. BACKGROUND

[0002] With the rapid development of Internet technology, streaming media platforms have become an important way for people to obtain video content. Among them, the video recommendation system, as one of the core functions of the streaming media platform (such as a certain TV), provides personalized video recommendations for users through the video recommendation system, which not only increases user stickiness, improves platform activity and user satisfaction, but also has a huge driving effect on the commercial value of the platform.

[0003] Traditional streaming media platform video recommendation systems mainly use collaborative filtering, matrix decomposition, and other technologies to recommend corresponding videos to users. However, this approach has certain limitations in capturing and depicting real-time user intentions; on the one hand, user behavior has a spontaneous component, resulting in inconsistencies between the user's current preferences and historical data. For example, a user may be particularly interested in videos on a certain topic during a certain period of time, but after a period of time, the user's interest may shift to other topics; on the other hand, traditional recommendation systems are passive recommendations and cannot interact with users to achieve active user intention feedback. Users cannot actively tell the recommendation system about their needs and preferences, making it difficult for the recommendation system to accurately capture user intentions and recommend videos that meet user needs and preferences, resulting in a low user experience. SUMMARY

[0004] In view of the above, the present application provides a video recommendation method and system based on a large language model, an electronic device, and a storage medium to recommend videos that meet user needs and preferences, and improve user experience.

[0005] The first aspect of the present application provides a video recommendation method based on a large language model, the method comprising:

[0006] When a recommendation request initiated by a target user is detected, at least one target user intention label data matching the target user is obtained from the stored user intention label data of each user in the key-value database according to the recommendation request; wherein the target user intention label includes at least one content label and at least one intention label, and the user intention label data is generated according to user feedback of user text feedback data;

[0007] The content tags in the target user intent label data in which the intent label exists are combined into a first content tag set, and a first video item set is screened from all video items on the streaming media platform according to the first content tag set; wherein the first video item set includes a plurality of first video items;

[0008] A first video tag set of each of the first video items is obtained, and a second video item set is screened from each of the first video items according to the first content tag set and the first video tag set of each of the first video items;

[0009] Target user text feedback data corresponding to the target user is obtained, and a third video item set is screened from all video items on the streaming media platform according to the target user text feedback data;

[0010] A fourth video item set is screened from each of the video items according to a plurality of preset recall algorithms, and the second video item set, the third video item set and the fourth video item set are fused to obtain a fifth video item set;

[0011] The content tags in the target user intent label data in which the intent label exists are combined into a first content tag set, and a first video item set is screened from all video items on the streaming media platform according to the first content tag set; wherein the first video item set includes a plurality of first video items;

[0012] Each of the fifth video items in the fifth video item set is sorted based on the second content tag set to obtain a sixth video item set;

[0013] A corresponding video recommendation list is generated according to the sixth video item set, and corresponding videos are recommended to the target user according to the video recommendation list.

[0014] Optionally, the user intent label data is generated according to the user feedback text feedback data, including:

[0015] The user feedback text feedback data is obtained in real time, wherein the user feedback text feedback data includes a plurality of user feedback texts;

[0016] According to each of the user feedback texts and a preset prompt word template, a corresponding prompt word is generated, and the prompt word is input into a large language model to obtain user intent label data corresponding to each of the user feedback texts; wherein the user intent label data includes a plurality of content tags and a plurality of intent labels;

[0017] For each of the user intent label data, the user intent label data is verified for validity;

[0018] If the validity verification of the user intention label data passes, the user information of the user and the user intention label data are stored in a pre-constructed key-value database in the form of key-value pairs.

[0019] Optionally, the first video tag set of each first video entry is obtained, and a second video entry set is filtered from each first video entry according to the first content tag set and the first video tag set of each first video entry, including:

[0020] The first video tag set of each first video entry is obtained from a pre-set video tag database;

[0021] The similarity of the first video tag set of each first video entry and the first content tag set is calculated by using a Jaccard similarity coefficient;

[0022] The second video entry set is filtered from each first video entry according to the similarity of the first video tag set of each first video entry and the first content tag set.

[0023] Optionally, the target user text feedback data corresponding to the target user is obtained, and a third video entry set is filtered from all video entries on the streaming media platform according to the target user text feedback data, including:

[0024] The target user text feedback data of the target user is obtained, and target user feedback text of the target user is determined from the target user text feedback data of the target user;

[0025] The latent vector of the target user feedback text is calculated;

[0026] The latent vector of each video entry is obtained, and the cosine similarity between the latent vector of each video entry and the latent vector of the target user feedback text is calculated,

[0027] The third video entry set is filtered from all video entries on the streaming media platform according to the cosine similarity between the latent vector of each video entry and the latent vector of the target user feedback text.

[0028] Optionally, each fifth video entry in the fifth video entry set is sorted based on the second content tag set to obtain a sixth video entry set, including:

[0029] The second video tag set of each fifth video entry in the fifth video entry set is obtained;

[0030] calculating, for each of the fifth video entries, a cross feature of each second video tag in the second video tag set of the fifth video entry and the second content tag set, and accumulating the cross feature of each second video tag in the fifth video entry and the second content tag set to obtain the cross feature of the fifth video entry and the second content tag;

[0031] calculating, by the pre-set fine arrangement model, a score of each of the fifth video entries according to the cross feature of the fifth video entry and the second content tag set, and ranking each of the fifth video entries according to the score of the fifth video entry to obtain a sixth video entry set.

[0032] Optionally, before the fifth video entries in the fifth video entry set are ranked based on the second content tag set to obtain a sixth video entry set, the method further comprises:

[0033] composing, into a third content tag set, content tags in the target user intent tag data in which the intent tag exists in each of the target user intent tag data;

[0034] filtering the fifth video entry set by using the third content tag set;

[0035] Correspondingly, ranking each of the fifth video entries in the fifth video entry set based on the second content tag set to obtain a sixth video entry set comprises:

[0036] ranking each of the filtered fifth video entries in the fifth video entry set based on the second content tag set to obtain a sixth video entry set.

[0037] Optionally, before the corresponding video recommendation list is generated according to the sixth video entry set, the method further comprises:

[0038] determining whether there is target user intent tag data in which the intent tag is a heavy scattering strategy tag in each of the target user intent tag data;

[0039] If there is, composing, into a fourth content tag set, content tags in the target user intent tag data in which the intent tag is a heavy scattering strategy tag in each of the target user intent tag data;

[0040] re-ranking the sixth video entry set by using the fourth content tag set;

[0041] Correspondingly, generating the corresponding video recommendation list according to the sixth video entry set comprises:

[0042] generate a corresponding video recommendation list according to the reordered sixth video item set, and recommend corresponding videos to the target user according to the video recommendation list.

[0043] The second aspect of the application provides a video recommendation system based on a large language model, the system comprising an intention execution module;

[0044] The intention execution module is configured to, when a recommendation request initiated by a target user is detected, obtain at least one target user intention label data matched with the target user from each user intention label data stored in a key-value database according to the recommendation request; wherein the target user intention label comprises at least one content label and at least one intention label, and the user intention label data is generated according to user text feedback data fed back by a user; the content labels in the target user intention label data in which the intention label is a recall strategy label are grouped into a first content label set, and a first video item set is screened from all video items on a streaming media platform according to the first content label set; wherein the first video item set comprises a plurality of first video items; a first video label set of each first video item is obtained, and a second video item set is screened from each first video item according to the first content label set and the first video label set of each first video item; target user text feedback data corresponding to the target user is obtained, and a third video item set is screened from each video item according to the target user text feedback data; a fourth video item set is screened from each video item according to each preset recall algorithm set in advance, and the second video item set, the third video item set, and the fourth video item set are fused to obtain a fifth video item set; the content labels in the target user intention label data in which the intention label is a sorting strategy label are grouped into a second content label set; each fifth video item in the fifth video item set is sorted based on the second content label set to obtain a sixth video item set; a corresponding video recommendation list is generated according to the sixth video item set, and corresponding videos are recommended to the target user according to the video recommendation list.

[0045] The third aspect of the application provides an electronic device, comprising a processor and a memory connected through a communication bus; wherein the processor is configured to call and execute a program stored in the memory; and the memory is configured to store a program, the program being used to implement the video recommendation method based on a large language model provided in the first aspect of the application.

[0046] The fourth aspect of the present application provides a computer readable storage medium, wherein the computer readable storage medium stores computer executable instructions, and the computer executable instructions are used to execute the video recommendation method based on the large language model provided in the first aspect of the present application.

[0047] The application provides a video recommendation method and system based on a large language model, an electronic device and a storage medium. The application generates and stores corresponding user intention label data according to user text feedback data fed back by a user in advance, so as to interact with the user, understand the user's needs and preferences in a timely manner, and accurately capture the user's intention label and content label according to the preferences and needs actively fed back by the target user. When a recommendation request initiated by a target user is detected, at least one target user intention label data matching the target user is obtained from each user intention label data stored in a key-value database according to the recommendation request, so as to form a first content label set by using content labels in target user intention label data in which an intention label is a recall strategy label in each target user intention label data, and filter a first video item set from all video items on a streaming media platform according to the first content label set, so as to filter a video that the target user may like from all video items on the streaming media platform according to the preference and need content label of the target user. In order to further improve the user experience, a first video label set of each first video item can be further obtained, and a second video item set can be filtered from each first video item according to the first content label set and the first video label set of each first video item. Meanwhile, target user text feedback data corresponding to the target user can be obtained, and a third video item set can be filtered from all video items on the streaming media platform according to the target user text feedback data. A fourth video item set can be filtered from all video items on the streaming media platform according to each preset recall algorithm set in advance, and the second video item set, the third video item set and the fourth video item set are fused to obtain a fifth video item set. In order to further improve the user experience, content labels in target user intention label data in which an intention label is a sorting strategy label in each target user intention label data are combined to form a second content label set, so as to sort the fifth video item set based on the second content label set to obtain a sixth video item set. That is, by reordering the fifth video item set based on the second content label set formed by content labels in target user intention label data in which an intention label is a sorting strategy label, it can be further ensured that the target user can see the video he wants to see first, thereby improving the user experience. A corresponding video recommendation list is generated according to the sixth video item set, and corresponding videos are recommended to the target user according to the video recommendation list, thereby recommending videos that meet the needs and preferences of the target user to the target user, and improving the user experience. BRIEF DESCRIPTION OF DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description only constitute a part of the embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of the provided drawings.

[0049] Figure 1 A structural schematic diagram of a video recommendation system based on a large language model provided by an embodiment of the present application;

[0050] Figure 2 A flowchart of a video recommendation based on a large language model provided by an embodiment of the present application;

[0051] Figure 3 An example diagram of a video recommendation method based on a large language model provided by an embodiment of the present application;

[0052] Figure 4 A structural schematic diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0053] The technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments only constitute a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0054] In the present application, the relationship terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the terms “include”, “contain” or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or device. Without more limitations, the element defined by the statement “including a…” does not exclude the presence of another identical element in the process, method, article or device including the element.

[0055] Referring to Figure 1 , a structural schematic diagram of a video recommendation system based on a large language model provided by an embodiment of the present application is shown, which includes a user interaction module, an intent recognition module, an intent memory module, an intent execution module and a user feedback module.

[0056] The intention execution module is configured to, when detecting a recommendation request initiated by a target user, acquire at least one target user intention label data matched with the target user from each user intention label data stored in the key-value database according to the recommendation request; the target user intention label comprises at least one content label and at least one intention label, and the user intention label data is generated according to user text feedback data fed back by a user; the content labels in the target user intention label data in which the intention label is the recall strategy label are composed into a first content label set, and a first video item set is screened from all video items on the streaming media platform according to the first content label set; the first video item set comprises a plurality of first video items; a first video label set of each first video item is acquired, and a second video item set is screened from each first video item according to the first content label set and the first video label set of each first video item; target user text feedback data corresponding to the target user is acquired, and a third video item set is screened from all video items on the streaming media platform according to the target user text feedback data; a fourth video item set is screened from each video item according to each preset recall algorithm, and the second video item set, the third video item set and the fourth video item set are fused to obtain a fifth video item set; the content labels in the target user intention label data in which the intention label is the sorting strategy label are composed into a second content label set; each fifth video item in the fifth video item set is sorted based on the second content label set to obtain a sixth video item set; a corresponding video recommendation list is generated according to the sixth video item set, and corresponding videos are recommended to the target user according to the video recommendation list.

[0057] Optionally, the user interaction module is configured to acquire user text feedback data fed back by a user in real time, wherein the user text feedback data comprises a plurality of user feedback texts.

[0058] The intention recognition module is configured to generate a corresponding prompt word according to each user feedback text and a preset prompt word template, and input the prompt word into a large language model to obtain user intention label data corresponding to each user feedback text; the user intention label data comprises a plurality of content labels and a plurality of intention labels.

[0059] The intention memory module is configured to, for each user intention label data, perform validity verification on the user intention label data; if the validity verification of the user intention label data is passed, the user information of the user and the user intention label data are stored in the key-value database in the form of a key-value pair.

[0060] Optionally, the intent execution module is configured to acquire the first video tag set of each first video entry, and filter the second video entry set from the respective first video entries according to the first content tag set and the first video tag set of each first video entry, and specifically configured to:

[0061] The intent execution module is configured to acquire the first video tag set of each first video entry from the pre-set video tag database, calculate the similarity between the first video tag set of each first video entry and the first content tag set by using the Jaccard similarity coefficient, and filter the second video entry set from the respective first video entries according to the similarity between the first video tag set of each first video entry and the first content tag set.

[0062] Optionally, the intent execution module is configured to acquire the target user text feedback data corresponding to the target user, and filter the third video entry set from all video entries on the streaming media platform according to the target user text feedback data of the target user, and specifically configured to:

[0063] The intent execution module is configured to acquire the target user text feedback data of the target user, determine the target user feedback text of the target user from the target user text feedback data of the target user, calculate the latent vector of the target user feedback text, acquire the latent vector of each video entry, calculate the cosine similarity between the latent vector of each video entry and the latent vector of the target user feedback text, and filter the third video entry set from all video entries on the streaming media platform according to the cosine similarity between the latent vector of each video entry and the latent vector of the target user feedback text.

[0064] Optionally, the intent execution module is configured to sort the respective fifth video entries in the fifth video entry set based on the second content tag set to obtain the sixth video entry set, and specifically configured to:

[0065] The intent execution module is configured to acquire the second video tag set of each fifth video entry in the fifth video entry set, calculate the cross feature of each second video tag in the second video tag set of each fifth video entry and the second content tag set for each fifth video entry, accumulate the cross feature of each second video tag of the fifth video entry and the second content tag set to obtain the cross feature of the fifth video entry and the second content tag, calculate the score of each fifth video entry according to the cross feature of each fifth video entry and the second content tag set by using the pre-set precision ranking model, and sort the respective fifth video entries according to the score of each fifth video entry to obtain the sixth video entry set.

[0066] Optionally, before sorting each fifth video item in the fifth video item set based on the second content tag set to obtain a sixth video item set, the intention execution module is further configured to:

[0067] The content tags in the target user intention tag data in which the intention tag is the filter strategy tag are grouped into a third content tag set, and the fifth video item set is filtered by using the third content tag set.

[0068] Optionally, before generating a corresponding video recommendation list according to the sixth video item set, the intention execution module is further configured to:

[0069] It is determined whether there is target user intention tag data in which the intention tag is the re-dispersal strategy tag in each target user intention tag data. If there is, the content tags in the target user intention tag data in which the intention tag is the re-dispersal strategy tag are grouped into a fourth content tag set, and the sixth video item set is re-ordered by using the fourth content tag set.

[0070] Optionally, the user feedback module is configured to: when a corresponding video is recommended to the target user according to the video recommendation list, a pre-set prompt text is displayed beside or above the recommended video to prompt the target user why the video is recommended and ask the user whether the recommendation strategy is recommended.

[0071] Referring to Figure 2 , a flowchart of a video recommendation method based on a large language model is shown, the video recommendation method based on a large language model is applied to a video recommendation system based on a large language model, and the video recommendation method based on a large language model specifically includes the following steps:

[0072] S201: When a recommendation request initiated by a target user is detected, at least one target user intention tag data matched with the target user is obtained from each user intention tag data stored in a key-value database according to the recommendation request; wherein the target user intention tag includes at least one content tag and at least one intention tag, and the user intention tag data is generated according to user text feedback data fed back by the user.

[0073] In the embodiment of the application, a corresponding key-value database can be constructed in advance, and user intention tag data generated according to user text feedback data of the user obtained in real time can be stored in the key-value database in advance. Meanwhile, the user text feedback data of the user can also be stored in the pre-set key-value database.

[0074] Optionally, the process of generating user intent label data according to the real-time acquired user text feedback data and storing the user intent label data into the pre-constructed key-value database can be specifically: acquiring user text feedback data of user feedback in real time, wherein the user text feedback data includes multiple user feedback texts; generating corresponding prompt words according to each user feedback text and a pre-set prompt word template, and inputting the prompt words into a large language model to obtain user intent label data corresponding to each user feedback text; wherein the user intent label data includes multiple content labels and multiple intent labels; for each user intent label data, performing validity verification on the user intent label data; if the validity verification of the user intent label data passes, storing the user information of the user and the user intent label data in the pre-constructed key-value database in the form of a key-value pair.

[0075] It should be noted that the real-time user intent and emotion accurately captured from the user feedback text by using the large language model can enhance the accuracy of personalized recommendation of the video. The large language model technology is based on Bert, Transformer and other deep learning algorithms, and can understand and generate natural language by training a large-scale neural network model on a large amount of text data, thereby achieving deep understanding of user intent.

[0076] In some embodiments, the pre-set key-value database can be a distributed key-value storage engine. After acquiring the user text feedback data of user feedback in real time, the user text feedback data of the user feedback can be stored into the distributed key-value storage engine, and in the case that the validity verification of the user intent label data obtained by using the user feedback text in the user text feedback data of the user feedback passes, the user intent label data can be stored into the distributed key-value storage engine.

[0077] It should be noted that the user feedback text can be user opinion text feedback by the user for the recommended content exposed by the streaming media platform, or user opinion text feedback by the user for the long-clicked recommended content, or search text input by the user and comment text published by the user for the recommended content exposed by the streaming media platform.

[0078] It should also be noted that the large language model can be GPT-4, Claude-3, GLM-4, etc., and the embodiments of the present application are not limited herein.

[0079] In some embodiments, when it is detected that the streaming media platform exposes a large amount of recommended content but there is no user click on the recommended content exposed by the streaming media platform, the user opinion text in natural language form is collected for the user feedback on the recommended content exposed by the streaming media platform, that is, the active natural language feedback module is triggered, and the user opinion text for the user feedback on the recommended content exposed by the streaming media platform is collected through a text or language dialogue box; when it is detected that the streaming media platform exposes a large amount of recommended content and there is user long click feedback on the recommended content exposed by the streaming media platform, the user opinion text in natural language form is collected for the user feedback on the long click recommended content; when it is detected that the user actively searches in the search bar of the streaming media platform, the search text input by the user in natural language form is collected, that is, the active natural language feedback module is triggered, and the search text input by the user is collected; when it is detected that the user publishes a comment on the recommended content exposed by the streaming media platform, the comment text published by the user on the recommended content exposed by the streaming media platform in natural language form is collected, that is, the active natural language feedback module is triggered, and the comment text input by the user is collected. Finally, the user feedback text, the search text and the comment text of the user are collected to generate the user feedback data of the user.

[0080] It should be noted that the user feedback text can be that the user wants to watch more videos related to a star, does not like to watch videos related to a star, and the like; the search text can be a reality show of a star, a sketch of a sketch actor, and the like; the comment text can be that a sketch of a sketch actor is really great, the acting of a star in a TV series is really great, and the like. In this embodiment of the present application, the user feedback text, the search text and the comment text are not limited.

[0081] In actual application, when it is detected that the streaming media platform exposes a large amount of recommended content but there is no user click on the recommended content exposed by the streaming media platform, the user can be invited to feedback corresponding opinions on the recommended content exposed by the streaming platform through the form of a pop-up window or a prompt bar, that is, when it is detected that the streaming media platform exposes a large amount of recommended content but there is no user click on the recommended content exposed by the streaming media platform, a corresponding pop-up box or prompt bar is popped up to prompt the user to input corresponding user feedback text in the pop-up box or prompt bar. In the process of browsing the recommended content exposed by the streaming media platform, if the user has doubts or dissatisfaction with the recommended content exposed by the streaming media platform, the user can long click the recommended content with doubts or dissatisfaction, so that the user long clicks the recommended content exposed by the streaming media platform is detected, and the corresponding pop-up box or prompt bar is popped up based on the video recommendation system of the large language model to prompt the user to input the corresponding user feedback text in the pop-up box or prompt bar.

[0082] In some embodiments, a set of prompt word templates can be pre-set, so that after obtaining the user feedback text feedback data of the user feedback, each user feedback text in the user feedback text feedback data is filled into the pre-set prompt word template to obtain the corresponding prompt word, and the prompt word is input into the large language model; the large language model can identify the structured user intent data reflecting the user recommendation preference in each user feedback text in the prompt word, to obtain the user intent label data corresponding to each user feedback text, wherein the user intent label data includes at least one intent label and at least one content label, and finally the user intent label data corresponding to each user feedback text is generated according to each content label and each intent label corresponding to each user feedback text, that is, the multiple user intent label data corresponding to the user is obtained. The user intent label data includes at least one content label describing the user feedback text and an intent label describing the recommendation algorithm strategy.

[0083] It should be noted that the intent label can be a recall strategy label ('recall'), or a filter strategy label ('filter'), or a ranking strategy label ('rank'), or a re-ranking strategy label ('rerank'). Correspondingly, if the intent label is a recall strategy label, the corresponding recommendation algorithm can be a Jaccard similarity coefficient; if the intent label is a filter strategy label, the corresponding recommendation algorithm can be a filter algorithm; if the intent label is a ranking strategy label, the corresponding recommendation algorithm can be a ranking algorithm; and if the intent label is a re-ranking strategy label, the corresponding recommendation algorithm can be a re-ranking algorithm.

[0084] It should also be noted that the recall algorithm can be a CF, youtubeDNN, Eges, Item2Vec, Swing, graph neural network, etc. The embodiments of the present application are not limited here.

[0085] In actual application process, when the large language model identifies from the user feedback text in the input prompt word that the user is interested in a specific content or explicitly indicates that he wants to see more similar content, it can be determined that the intent label corresponding to the user feedback text can include a recall intent label, for example, the user feedback text in the prompt word is that the user likes a video of a certain star, then the large language model can identify from the user feedback text in the prompt word that the user is interested in the video of the certain star, at this time it can be determined that the intent label corresponding to the user feedback text is a recall intent label.

[0086] If the large language model identifies from the user feedback text in the input prompt word that the user is not satisfied with a specific content, or does not like it, or explicitly indicates that he does not want to see more of this type of content, it can be determined that the intent label corresponding to the user feedback text can include a filtering intent label. For example, if the user feedback text in the prompt word is that he does not want to see a video of a certain star, the large language model can identify from the user feedback text in the prompt word that the user determines to indicate that he does not want to see a video of a certain star, and at this time it can be determined that the intent label corresponding to the user feedback text is a filtering intent label.

[0087] If the large language model identifies from the user feedback text in the input prompt word that the user's preference for a specific content is higher than that of other content, or requires a specific content to have a higher ranking in the recommended category table, it can be determined that the intent label corresponding to the user feedback text can include an ordering intent label. For example, if the user feedback text in the prompt word is that he wants to see more videos of a certain star, the large language model can identify from the user feedback text in the prompt word that the user not only explicitly indicates interest in videos of a certain star, but also explicitly indicates a higher preference for videos of a certain star. At this time, it can be determined that the intent label corresponding to the user feedback text includes a recall intent label and an ordering intent label.

[0088] If the large language model identifies from the user feedback text in the input prompt word that the user is not satisfied with the current ordering result of a large number of recommended contents currently exposed by the streaming platform, and wants to re-adjust the order of the recommended contents, it can be determined that the intent label corresponding to the user feedback text can include a reshuffle intent label.

[0089] In some embodiments, for each user feedback text in the user text feedback data of the user, after generating a corresponding prompt word according to a pre-set prompt word template and each user feedback text, inputting the prompt word into the large language model, and making the large language model analyze each user feedback text in the input prompt word, at least one content label is obtained to describe the preference of each user feedback text. Wherein, the at least one content label can include: a time label, a geographical label, a video type label, a theme keyword label, a label indicating whether the preference expressed by the text is a positive emotion or a negative emotion, an episode label, a segment label, a star label, a video effect label, a plot label, and the like. The embodiments of the present application are not limited herein.

[0090] In actual application process, for positive or negative emotion label: if the large language model identifies that the preference tendency expressed by the user feedback text in the prompt word input by the user is positive, the large language model can feed back the corresponding positive emotion label in the form of IS POSITIVE DATA, and if the large language model identifies that the preference tendency expressed by the user feedback text is negative, the large language model can feed back the corresponding positive emotion label in the form of IS POSITIVE DATA, wherein IS POSITIVE DATA is an integer, which can be 0 or 1, 0 represents negative emotion, and 1 represents positive emotion; specifically, the large language model can extract label information describing user emotional tendency from the user feedback text in the prompt word input by the user, and if the label information describing user emotional tendency is positive, the large language model outputs 1, and if the label information describing user emotional tendency is negative, the large language model outputs 0.

[0091] For example, the user feedback text is: 'I want to see I am a great detective', at this time the large language model can identify that the preference tendency expressed by the user feedback text is positive from the prompt word generated by the user feedback text, at this time the large language model can output 1; or the user feedback text is: 'Why do you recommend me so many romance TV series', at this time the large language model can identify that the preference tendency expressed by the user feedback text is negative from the prompt word generated by the user feedback text, at this time the large language model can output 0.

[0092] For content label extraction within a fixed content label range: if the large language model identifies time, geographic location, video type focused by the text, and theme keyword focused by the text from the user feedback text in the prompt word input, the large language model can select corresponding geographic label, video type label, theme keyword label and the like from the pre-set content label range, and feed back the corresponding geographic label, video type label, theme keyword label in the form of CATEGORY DATA, and if the large language model cannot identify the corresponding content label from the user feedback text, it can feed back an empty value list; wherein CATEGORY DATA is in the form of a string list. The pre-set content label range can be ['geography', 'theme keyword', 'variety', 'TV series','movie', 'children', 'animation', 'news', 'documentary','music', 'live', 'entertainment', 'life', 'game', 'education', 'fashion'], 'TV series','movie', 'children', 'animation', 'news', 'documentary','music', 'live', 'entertainment', 'life', 'game', 'education', 'fashion' are video types, and the pre-set content label range is not limited in the embodiment of the application.

[0093] For example, the user feedback text is: 'Han Han, no future', the large language model can identify the video type "movie" that the text focuses on from the prompt words generated according to the user feedback text, at this time, the large language model can select the corresponding video type label from the pre-set content label range, and feedback the corresponding video type label in the form of CATEGORY_DATA. Or, the user feedback text is: 'T1 Saudi Cup double kill LPL team to win the championship', the large language model can identify the video types "game" and "live" that the text focuses on from the prompt words generated according to the user feedback text, at this time, the large language model can select the corresponding video type label from the pre-set content label range, and feedback the corresponding video type label in the form of CATEGORY_DATA.

[0094] For time labels: If the large language model extracts all key label information related to time from the user feedback text in the input prompt words, and feeds back the corresponding content label in the form of TIME_DATA, where TIME_DATA is a string class format; If there is no explicit time information in the user feedback text, an empty list can be fed back.

[0095] For example, the user feedback text is: 'I want to watch the latest episode of Singer 2024', the large language model can extract the key label information "2024" related to time from the prompt words generated according to the user feedback text, at this time the large language model can generate and feed back the corresponding time label in the form of TIME_DATA according to the extracted key label information; Or the user feedback text is: 'What is the latest variety show today', the large language model can extract the key label information "today" related to time from the prompt words generated according to the user feedback text, at this time the large language model can generate and feed back the corresponding time label in the form of TIME_DATA according to the extracted key label information. Or, the user feedback text is: 'I want to watch the performance of user 1 in Singer 2024', the large language model can extract the key label information "2024" related to time and the theme keyword "user 1" from the prompt words generated according to the user feedback text, at this time the large language model can generate and feed back the corresponding time label in the form of TIME_DATA according to the extracted key label information, and feed back the corresponding keyword label in the form of CATEGORY_DATA.

[0096] In some embodiments, after outputting the corresponding user intent label data according to the input prompt words by the large language model, the user intent label data output by the large language model can be further verified for effectiveness, so as to store the user intent label data in the pre-constructed key-value database in the case that the effectiveness verification of the user intent label data is determined to pass.

[0097] In the embodiment of the present application, the label format and label content of the user intent label data can be verified. If the label format and label content of the user intent label data are verified, it is determined that the validity verification of the user intent label data is passed, otherwise it is determined that the validity verification of the user intent label data is not passed.

[0098] It should be noted that the present application can use the JSON Schema verification technology to check the label format of the user intent label data. Specifically, a strict schema is defined to constrain the data structure, ensuring that the field type, necessity and value range of the output data meet the expected data format. If the data format is found to be incorrect, an exception handling process is triggered, and the user intent label data with data format error is saved as log information to a specified file for technical personnel to check and optimize. The expected data format can be JSON format.

[0099] Specifically, after the large language model outputs the corresponding user intent label data according to the user feedback text in the input prompt word, it can be determined whether the label format of the user intent label data is JSON format, and whether there are content labels and intent labels in the user intent label data. If the label format of the user intent label data is JSON format, and there are content labels and intent labels in the label content of the user intent label data, it can be determined that the label format and label content of the user intent label data are verified, and then it can be determined that the validity verification of the user intent label data is passed. If the label format of the user intent label is not JOSN format, and / or, there are no content labels or intent labels in the label content of the user intent label data, it can be determined that the label format and / or label content of the user intent label data is not verified, and then it can be determined that the validity verification of the user intent label data is not passed. The content labels and intent labels in the user intent label data are labels in the pre-set label library.

[0100] It should be noted that if the content labels and intent labels in the user intent label data are not labels in the pre-set label library, it means that the content labels and intent labels in the user intent label data may belong to sensitive labels, and it can be further determined whether the content labels and intent labels in the user intent label data belong to sensitive labels. If they belong, the corresponding exception handling process can be triggered. If they do not belong, the user intent label data can be stored in the key-value database.

[0101] It should be noted that in the case where the validity verification of the user intent label data fails to pass, a corresponding exception handling process can be triggered to use the OSS object storage service to record and store the user intent label data that fails to pass the validity verification into a specified file of the distributed file system, to ensure the safe persistence of the data, and to provide a query and analysis interface for the operation personnel or technical personnel to check or update the corresponding label library, wherein the distributed file system can be an OSS service such as Amazon S3, Google Cloud Storage, and Aliyun OSS, which are not limited in the embodiments of the present application; in the case where the validity verification of the user intent label data passes, the user intent label data and the user text feedback data can be stored using a distributed key-value storage engine (Redis) with the user ID of the corresponding user as the key value.

[0102] It should also be noted that when writing new user intent label data or user text feedback data into the distributed key-value storage engine, the old data can be automatically overwritten with conflicting key values to ensure real-time updating of the data and timely reflection of the latest preferences of the user by the system, thereby improving the timeliness of video recommendation.

[0103] In the process of specifically executing step S201, the video recommendation system based on the large language model can detect in real time whether there is a recommendation request initiated by a target user (for the sake of distinction, the user who initiates the recommendation request is referred to as a target user), and when detecting a recommendation request initiated by a target user, at least one target user intent label data matched with the target user can be obtained from each user intent label data stored in the key-value database by the recommendation engine according to the user ID of the target user carried in the recommendation request (for the sake of distinction, the user intent label data matched with the target user is referred to as target user intent label data).

[0104] S202: The content labels in the target user intent label data in which the intent label is the recall strategy label are grouped into a first content label set, and a first video item set is selected from all video items on the streaming media platform according to the first content label set.

[0105] In the process of specifically performing step S202, after obtaining at least one target user intent label data matched by the target user UK, at least one target user intent label data in which the intent label is the recall strategy label can be filtered out from each target user intent label data by a recommendation engine, and the content labels in each target user intent label data in which the intent label is the recall strategy label are composed into a first content label set; according to each content label in the first content label set, a plurality of first video entries are filtered out from all video entries on the streaming media platform, and each first video entry filtered out is composed into a first video entry set.

[0106] In the embodiments of the present application, the video entry set of all video entries on the streaming media platform is V = {V1, V2,..., Vn}, and the key-value index T of each video entry is maintained in the pre-constructed key-value database, where n is a positive integer greater than 1, the key-value index of each video entry is a video label of each video entry, each video entry includes a plurality of video labels, and the specific formula is as follows:

[0107] Ti←V T1,...,Ta ={V T1 ,...,V Ta} (1)

[0108] As can be seen from formula 1, for each video entry in the video entry set of the streaming media platform, the video entry has a plurality of video labels, and the video label of the video entry is a video label in the pre-set video label set {T1, T2,..., Tm}, a is greater than or equal to 2, and m is greater than or equal to i.

[0109] It should be noted that for the video entries in the video entry set, each video entry can also be deeply embedded by natural language processing (NLP) technology such as a BRET pre-training model to obtain an implicit vector of each video entry, and an implicit vector set of the video entry set is constructed according to the implicit vectors of each video entry, as shown in formula (2):

[0110] E = {E V1 ,E V2 ,...,E Vn} (2)

[0111] Wherein, E is the implicit vector set of the video entry set, E V1 is the implicit vector of the video entry 1, E V2 is the implicit vector of the video entry 2, and E Vn is the implicit vector of the video entry n.

[0112] In the embodiments of the present application, assuming that a video item set of all video items on the streaming media platform is V={V1, V2,..., Vn}, the video recommendation system based on the large language model filters at least one target user intent label data in which the intent label is the recall strategy label from each target user intent label data through the recommendation engine, and forms a first content label set Y={Y1, Y2,..., Yr} from the content labels in each target user intent label data in which the intent label is the recall strategy label, where r is greater than or equal to 2; using each content label in the first content label set, recalling a first video item set V UK1 ={V1 UK1 ,V2 UK1 ,......,VN UK1} from the video item set V that is likely to coincide with the target user's preference, where N is greater than 1 and less than or equal to n, and each first video item includes at least one content label in the first content label set in the video label of the first video item, as shown in the following formula:

[0113] UK←Vi UK1 ={V1 T1 ,...,V Tj} (3)

[0114] where j is greater than 1 and less than or equal to m, UK is the target user, and Vi UK1 is any first video item in the first video item set V UK1 , and from formula (3) it can be seen that the first video label set of each first video item in the first video item set includes a plurality of video labels {V1 T1 ,...,V Tj}, and at least one content label in the first content label set is included in the plurality of video labels of the first video item.

[0115] S203: Obtain the first video label set of each first video item, and filter a second video item set from each first video item according to the first content label set and the first video label set of each first video item.

[0116] In the process of specifically performing step S203, after obtaining the first video item set, the video recommendation system based on the large language model can perform association calculation on the first content label set and the first video label set of each first video item in the first video item set to filter the second video item set from the first video item set.

[0117] Specifically, for each first video item in the first video item set, a first video tag set of the first video item can be obtained, and a similarity between the first video tag set of each first video item and the first content tag set can be calculated, so as to filter a second video item set from each first video item according to the similarity between the first video tag set of each first video item and the first content tag set.

[0118] Optionally, the process of filtering the second video item set from each first video item according to the first content tag set and the first video tag set of each first video item can be specifically: calculating the similarity between the first video tag set of each first video item and the first content tag set by using the Jaccard similarity coefficient; and filtering the second video item set from each first video item according to the similarity between the first video tag set of each first video item and the first content tag set. The similarity between the first video tag set of each first video item and the first content tag set is calculated by using the Jaccard similarity coefficient, and the formula (4) is as follows.

[0119]

[0120] wherein S1(Vi UK1 is the similarity between the first video tag set of the i th first video item and the first content tag set, Y is the first content tag set, is the first video tag set of the i th first video item.

[0121] In some embodiments, after obtaining the similarity between the first video tag set of each first video item and the first content tag set, each first video item can be sorted from high to low according to the similarity, and the first M video items in the sequence obtained by sorting are taken as the second video item set.

[0122] S204: Obtain target user text feedback data corresponding to a target user, and filter a third video item set from each video item according to the target user text feedback data.

[0123] In the process of specifically performing step S204, the target user text feedback data corresponding to the target user can be obtained from the key-value database according to the user ID of the target user, and each target user feedback text in the target user text feedback data of the target user can be deeply embedded by using a natural language processing (NLP) technology, such as a BRET pre-training model, to obtain a hidden vector of each target user feedback text, and a third video item set can be filtered from each video item according to the hidden vector of each target user feedback text.

[0124] Optionally, the process of obtaining the target user text feedback data corresponding to the target user and filtering the third video item set from the video items according to the target user text feedback data can be specifically: obtaining the target user text feedback data of the target user, and calculating the latent vector of each target user feedback text in the target user text feedback data of the target user; obtaining the latent vector of each video item, and calculating the cosine similarity between the latent vector of each video item and the latent vector of each target user feedback text, and filtering the third video item set from the video items according to the cosine similarity between the latent vector of each video item and the latent vector of each target user feedback text.

[0125] In actual application, the target user text feedback data of the target user can be obtained from the key-value database according to the user ID of the target user; for each target user feedback text in the target user text feedback data of the target user, the target user feedback text is deeply embedded by natural language processing technology (such as a pre-trained model of BERT) to capture the semantic features and context association of the target user feedback text, and the latent vector of the target user feedback text is obtained; the latent vector of each video item is obtained from the latent vector set of the pre-set video item set; for each video item, the cosine similarity between the latent vector of the video item and the latent vector of each target user feedback text can be calculated, and the plurality of cosine similarities of the video item are obtained; for each video item, if each cosine similarity of the video item is zero, the cosine similarity between the latent vector of the video item and the latent vector of the target user feedback text recently fed back by the target user is taken as the final cosine similarity of the video item; if there is a cosine similarity that is not zero in each cosine similarity of the video item, and there are a plurality of non-zero cosine similarities, the cosine similarity corresponding to the target user feedback text recently fed back by the user can be determined as the final cosine similarity of the video item from each target user feedback text corresponding to each non-zero cosine similarity; if there is a cosine similarity that is not zero in each cosine similarity of the video item, and there is only one non-zero cosine similarity, the non-zero cosine similarity can be determined as the final cosine similarity of the video item.

[0126] After determining the final cosine similarity of each video item, the video items can be sorted from high to low according to the final cosine similarity of each video item, and the first M video items in the sorted result form the third video item set. The cosine similarity between the latent vector of the video item and the latent vector of the target user feedback text is calculated as shown in formula (5):

[0127]

[0128] wherein S2(Vi) is a cosine similarity between the i-th video entry hidden vector and the target user feedback text hidden vector, E Vi is the i-th video entry hidden vector, is the i-th video entry hidden vector.

[0129] S205: Filtering a fourth video entry set from each video entry according to each preset recall algorithm, and fusing the second video entry set, the third video entry set and the fourth video entry set to obtain a fifth video entry set.

[0130] In the embodiments of the present application, a plurality of preset recall algorithms can be preset, wherein the preset recall algorithms can be CF, youtubeDNN, Eges, Item2Vec, Swing, graph neural network and the like, which are not limited in the embodiments of the present application.

[0131] In the process of specifically executing step S205, after obtaining the second video entry set and the third video entry set, the fourth video entry set can be filtered from each video entry by using each preset recall algorithm, so as to fuse the obtained second video entry set, the third video entry set and the fourth video entry set to obtain a fused fifth video entry set.

[0132] S206: The content labels in the target user intent label data in which the intent label is the sorting strategy label are grouped to obtain a second content label set.

[0133] In the process of specifically executing step S206, at least one target user intent label data in which the intent label is the sorting strategy label can be filtered from each target user intent label data, and the content labels in each target user intent label data in which the intent label is the sorting strategy label are grouped to obtain a second content label set H={H1, H2,..., Hb}, wherein b is greater than or equal to 2.

[0134] S207: Sorting each fifth video entry in the fifth video entry set based on the second content label set to obtain a sixth video entry set.

[0135] In the specific execution step S207, after obtaining the fifth video item set, for each fifth video item in the fifth video item set (for ease of distinction, each video item in the fifth video item set is referred to as the fifth video item), the second video tag set of the fifth video item can be obtained from the pre-set video tag set, so as to call the pre-set sorting algorithm to sort each fifth video item according to the second video tag set and the second content tag set of each fifth video item, and obtain the sixth video item set.

[0136] Optionally, the process of sorting each fifth video item in the fifth video item set based on the second content tag set to obtain the sixth video item set can be specifically as follows: obtaining the second video tag set of each fifth video item in the fifth video item set; for each fifth video item, calculating the cross features of each second video tag in the second video tag set of the fifth video item with the second content tag set, and accumulating the cross features of each second video tag in the second content tag set of the fifth video item to obtain the cross features of the fifth video item and the second content tag; calculating the score of each fifth video item based on the cross features of each fifth video item and the second content tag set using a pre-set fine ranking model, and sorting each fifth video item according to the score of each fifth video item to obtain the sixth video item set.

[0137] It should be noted that by reordering the fifth video item set using a second set of content tags composed of content tags from the target user intent tag data that has intent tags as sorting strategy tags, it is possible to further ensure that users can see the videos they want to see first, thereby improving the user experience.

[0138] In some embodiments, when a recommendation request initiated by a target user is detected, the time point at which the recommendation request was detected can also be obtained, i.e., the time point at which the target user initiated the recommendation request; so that in the subsequent process of calculating the cross features of each second video tag and the second content tag in the second video tag set of the fifth video entry, a pre-set sorting algorithm can be called to calculate the cross features of each second video tag and the second content tag set according to the time point of the recommendation request, the second content tag set, and each second video tag in the second video tag set of the second video entry, and the cross features of each second video tag and the second content tag set are accumulated to obtain the cross features of the second video entry and the second content tag, as shown in formula (6):

[0139]

[0140] Among them, VUK2 denotes a fifth video item in the fifth video item set, denotes a second video tag in the second video tag set of the fifth video item, g denotes one second video tag in the second video tag set, and t denotes a time point when the recommendation request is initiated, denotes a timestamp of the tag g in the second content tag set. The function denotes whether the tag g exists in the second content tag set H: The function p(δt) is a time penalty function for promoting the weight of the recent user intention tag: p(δt) = exp(-0.035·δt) The numerical feature f(V UK2 , UK) will be used by the fine ranking model to promote the weight of the content of interest of the user in the fine ranking model. Wherein, the fine ranking model can be Wide & Deep, DIN, DIEN, DeepFM, and the embodiments of the present application are not limited here.

[0141] It should be noted that the sorting algorithm usually first constructs a feature engineering, then trains a fine ranking model, and after deploying the trained fine ranking model online, the sorting score of each recommendation result (fifth video item) in the recommendation candidate set (fifth video item set) is predicted in real time. Wherein, the feature engineering usually has ID features and historical behavior data of users (such as viewing records, likes, comments, searches, and payment capabilities), user features (such as age, gender, and geographic location), video features (such as title, tag, and category), and the like.

[0142] In some embodiments, after calculating the cross features of each fifth video item in the fifth video item set and the second content tag set, the fine ranking model can be used to calculate the score of each fifth video item according to the cross features of each fifth video item and the second content tag set, the features of each fifth video, the ID features of the target user, the historical behavior data of the target user, and the user features of the target user, so as to reorder each fifth video item in the fifth video item set according to the scores of the fifth video items from high to low, and obtain a sixth video item set.

[0143] Further, in the embodiments of the present application, before calculating the cross features of each fifth video item in the fifth video item set and the second content tag set, the fifth video item set can be filtered first, so as to subsequently calculate the cross features of each fifth video item in the filtered fifth video item set and the second content tag set.

[0144] As a manner of the embodiments of the present application, at least one target user intention label data in which the intention label is the filtering strategy label can be filtered out from each target user intention label data, and the content labels in each target user intention label data in which the intention label is the filtering strategy label are composed into a third content label set G = {G1, G2, …, Gc}, where c is greater than or equal to 2, and the second video item set / third video item set is filtered by using the third content label set.

[0145] Specifically, for each fifth video item in the fifth video item set, a filtering algorithm is called to check whether there is a second video label in the second video label set of the fifth video item that is the same as the content label in the third content label set, that is, whether there is a second video label in the second video label set of the fifth video item that coincides with the third content label set, if there is, the fifth video item can be deleted from the fifth video item set, otherwise, the fifth video item is retained.

[0146] As another manner of the embodiments of the present application, a corresponding filtering rule can be set in advance, and a filtering algorithm is called to filter the fifth video item set by using the filtering rule.

[0147] Specifically, the target user text feedback data of the target user can be obtained by using the recommendation engine, and the user feedback text in which the user is not satisfied with the feature content, or does not like, or explicitly expresses that he does not want to see more types of content is identified from each user feedback text in the target user text feedback data, and the corresponding filtering rule is constructed according to the user feedback text in which the user is not satisfied with the feature content, or does not like, or explicitly expresses that he does not want to see more types of content, so that the filtering algorithm can filter out the fifth video item that the user is not satisfied with, or does not like, or explicitly expresses that he does not want to see from the fifth video item set by using the filtering rule.

[0148] It should be noted that the target user text feedback data of the target user is analyzed by using the filtering algorithm, and the user feedback text in which the target user has negative feedback is determined, such as the user feedback text “I don't want to see in recent period of time”, “I don't like a certain type of result”, “I don't like a video with a comment number lower than a certain threshold”, “I don't like a video with false content”, and the corresponding filtering rule is constructed according to the user feedback text in which the user has negative feedback, so that the filtering algorithm is dynamically scheduled to filter the fifth video item set by using the filtering rule combined with the deep analysis of the user feedback negative user feedback text by using the machine learning model, so as to intelligently eliminate the video item that the target user may not like or does not want to see from the fifth video item set.

[0149] It should be noted that, by the presence of the intention label for filtering the content label in the target user intention label data of the filtering strategy label to form the third content label set or the filtering rule, the fifth video item set is filtered, which can effectively reduce the case that the user sees the video that the user does not like, and can further improve the experience of the user.

[0150] In the specific implementation process of step S208, after obtaining the sixth video item set, the corresponding video recommendation list can be generated according to the sixth video item set, and the corresponding video is recommended to the target user according to the video recommendation list.

[0151] In the specific implementation process of step S208, after obtaining the sixth video item set, the corresponding video recommendation list can be generated according to the sixth video item set, and the corresponding video is recommended to the target user according to the video recommendation list.

[0152] Further, in the embodiment of the present application, before generating the corresponding video recommendation list according to the sixth video item set, it can be further judged whether there is target user intention label data in which the intention label is the heavy scattering strategy label in each target user intention label data; if there is, the content label in the target user intention label data in which the intention label is the heavy scattering strategy label in each target user intention label data is composed into a fourth content label set; the sixth video item set is reordered by using the fourth content label set, so as to generate the corresponding video recommendation list according to the reordered sixth video item set, and recommend the corresponding video to the target user according to the video recommendation list.

[0153] It should be noted that, in the case that there is target user intention label data in which the intention label is the heavy scattering strategy label in each target user intention label data, the content label in the target user intention label data in which the intention label is the heavy scattering strategy label in each target user intention label data can be composed into a fourth content label set D={D1, D2,..., Dx}, wherein x is greater than or equal to 2.

[0154] As a way of the embodiment of the present application, at least one video item V UK3 ={V1 UK3 ,...,Vz UK3, wherein z is greater than 1; if there is, it can be further judged whether there is a content label in the fourth content label set indicating that the video of the corresponding type is placed on the front screen, if there is, a weighted re-dispersing algorithm is called to perform weighted calculation on each video entry matched with the content label indicating that the video of the corresponding type is placed on the front screen, and then a corresponding score is obtained, as shown in formula (7); finally, each sixth video in the sixth video entry set is reordered according to the final score, so that each video entry matched with the content label indicating that the video of the corresponding type is placed on the front screen is placed at the forefront of each video entry in the sixth video set, that is, each video entry matched with the content label indicating that the video of the corresponding type is placed on the front screen is inserted into the front screen.

[0155] f(V UK3 , UK) = p + score raw (7)

[0156] wherein V UK3 is a video entry matched with a content label in the fourth content label set, score raw is the original score of the video entry, and if the score range is {0, 1}, p can be a weighted weight for realizing the insertion of the front screen function; wherein p can be set according to the weighting requirement, and the present embodiment is not limited herein.

[0157] For example, when there is a content label "place the video of star i on the front screen" in the fourth content label set, a weighted re-dispersing algorithm can be used to forcibly extract the video entry matched with the content label "star i" from the sixth video entry set, and the extracted video entry is inserted into the front screen, that is, the extracted video entry is placed at the forefront of each video entry in the sixth video entry set, so that when the video recommendation is performed subsequently, the user can first see the video related to star i, and the user experience can be further improved.

[0158] As another way of the present embodiment, at least one video entry matched with a content label in the fourth content label set is selected from the sixth video entry set; if there is, it can be further judged whether there is a content label in the fourth content label set indicating that the video of the corresponding type is placed at a specified position, if there is, a position-specific re-dispersing algorithm is called to determine the score score N of the video entry corresponding to the specified position N in the sixth video entry set and the score score N-1 of the video entry of the previous position of the specified position; and the score of the video entry of the corresponding type is intervened as {score N , score N-1between the two, to achieve the corresponding type of video entry into the designated location N, so that in subsequent video recommendation, the user can see the designated location of the video they want to see, can further improve the user experience.

[0159] It should be noted that, in addition to the weighted major algorithm, the pit designated algorithm, the star algorithm, the video type algorithm, the long and short video algorithm, the video and multi-functional module algorithm, etc. are not limited in this embodiment.

[0160] Further, in the embodiment of the application, when the corresponding video is recommended to the target user according to the video recommendation list, the pre-set prompt text can be displayed beside or above the recommended video to prompt the target user why the video is recommended and ask the user whether the recommendation strategy is recommended, so that the target user can know more clearly why they can see the recommended video; and the target user can also choose whether to continue to accept the video recommendation based on the interest (user feedback text).

[0161] The application provides a video recommendation method based on a large language model. The corresponding user intention label data is generated and stored in advance according to the user text feedback data fed back by the user, so as to realize interaction with the user, understand the needs and preferences of the user in a timely manner, and accurately capture the intention label and content label of the user according to the preferences and needs actively fed back by the target user, that is, by using the large language model, the language expression of the user is understood, the intention and emotion of the user are captured, and more intelligent services are provided for the user; when detecting a recommendation request initiated by the target user, at least one target user intention label data matched with the target user is obtained from each user intention label data stored in the key-value database according to the recommendation request, so as to form a first content label set by taking the content label in the target user intention label data in which the intention label is the recall strategy label as the target user intention label data, and filter a first video item set from all video items on the streaming media platform according to the first content label set, so as to filter the video that the target user may like from all video items on the streaming media platform according to the preference and demand content label of the target user; in order to further improve the experience of the user, a first video label set of each first video item can be further obtained, and a second video item set is filtered from each first video item according to the first content label set and the first video label set of each first video item, and the target user text feedback data corresponding to the target user can also be obtained, and a third video item set is filtered from all video items on the streaming media platform according to the target user text feedback data; in order to further improve the experience of the user, the content label in the target user intention label data in which the intention label is the sorting strategy label is taken as the target user intention label data to form a second content label set, so as to sort the fifth video item set based on the second content label set to obtain a sixth video item set, that is, by reordering the fifth video item according to the second content label set formed by the content label in the target user intention label data in which the intention label is the sorting strategy label to obtain the sixth video item set, it can be further ensured that the target user can preferentially see the video that he wants to see more, thereby improving the experience of the user; a corresponding video recommendation list is generated according to the sixth video item set, and corresponding videos are recommended to the target user according to the video recommendation list, thereby realizing the recommendation of videos that meet the needs and preferences of the target user to the target user, and improving the experience of the user.

[0162] In order to better understand the video recommendation method based on the large language model provided by the embodiments of the application, the following is illustrated by way of example, as shown in Figure 3

[0163] ​The user interaction module triggers the active natural language feedback module to collect user opinion texts of user feedback on the recommended content exposed by the streaming platform through a text or language dialogue box when it detects that the streaming platform exposes a large amount of recommended content but there is no user click on the recommended content exposed by the streaming platform, and / or detects that the streaming platform exposes a large amount of recommended content and there is user long-click feedback on the recommended content exposed by the streaming platform. The active natural language feedback module is triggered to collect search texts and / or comment texts actively input by the user when it detects that the user actively searches in the search bar of the streaming platform and / or detects that the user posts corresponding comments on the recommended content exposed by the streaming platform. Finally, the user text feedback data generated according to the user feedback texts, search texts and comment texts collected from the user are sent to the intent recognition module.

[0164] The intent recognition module fills each user feedback text in the user text feedback data into a pre-set prompt word template to obtain a corresponding prompt word after obtaining the user text feedback data of user feedback, and inputs the prompt word into a large language model. The large language model identifies each user feedback text in the prompt word to obtain user intent label data corresponding to each user feedback text, wherein the user intent label data includes at least one intent label and at least one content label. The at least one intent label can include a recall strategy label ('recall'), or a filter strategy label ('filter'), or a ranking strategy label ('rank'), or a re-ranking strategy label ('rerank'), and the at least one content label can include a time label, a geographical label, a video type label, a theme keyword label, an emotion label of a positive or negative preference expressed by the text, a star label, and the like.

[0165] The intent memory module checks the label format and label content of the obtained user intent label data. If the label format of the user intent label data passes the check and the label content of the user intent label data passes the check, it is determined that the validity verification of the user intent label data passes, and then the user intent label data and the user ID of the user can be stored in the Redis KV storage cluster in the form of a key-value pair. If the label format of the user intent label is not in the JSON format, it can be determined that the label format check of the user intent label data fails, and the user intent label data can be recorded and stored in a specified file of a distributed file system. And / or, the label content of the user intent label data does not include a content label or an intent label, and the user intent label data can be recorded and stored in a specified file of a distributed file system.

[0166] The intention execution module is pre-provided with a plurality of preset recall algorithms, filtering algorithms, sorting algorithms and re-dispersing algorithms; when a recommendation request initiated by the target user through the corresponding client is detected, at least one target user intention label data matched with the target user is obtained from the user intention label data stored in the key-value database according to the user ID of the target user carried in the recommendation request through the recommendation engine; at least one target user intention label data with an intention label being a recall strategy label is filtered out from the target user intention label data through the recommendation engine, and the content labels in each target user intention label data with the intention label being the recall strategy label are composed into a first content label set; the first content label set is associated with the first video label set of each first video item in the first video item set by using the Jaccard similarity coefficient, so as to filter out a second video item set from the first video item set; the target user text feedback data corresponding to the target user is obtained from the key-value database according to the user ID of the target user, and the target user feedback text is filtered out from each user feedback text in the target user text feedback data, the latent vector of the target user feedback text is calculated, and a third video item set is filtered out from each video item according to the latent vector of the target user feedback text; a fourth video item set is filtered out from each video item according to the pre-set preset recall algorithm, and the second video item set, the third video item set and the fourth video item set are fused to obtain a fifth video item set.

[0167] The filtering algorithm is called to filter the fifth video item through the content labels in the target user intention label data with the intention label being the filtering strategy label or the pre-set filtering rule. The sorting algorithm is called to sort each fifth video item in the filtered fifth video item set according to the second video label set and the second content label set of each fifth video item in the filtered fifth video item set, to obtain a sixth video item set.

[0168] In the case that there is target user intention label data with the intention label being the re-dispersing strategy label in each target user intention label data, the re-dispersing algorithm can also be called to compose a fourth content label set from the content labels in the target user intention label data with the intention label being the re-dispersing strategy label in each target user intention label data; the sixth video item set is re-ordered by using the fourth content label set, and finally a corresponding video recommendation list is generated according to the re-ordered sixth video item set, and the corresponding video is recommended to the target user according to the video recommendation list.

[0169] The user feedback module can pop up the corresponding video push reason at a regular time when recommending the corresponding video to the target user according to the video recommendation list, that is, display the pre-set prompt text beside or above the recommended video to prompt the target user why the video is recommended, and ask the user whether to recommend the strategy.

[0170] The embodiment of the application provides an electronic device, such as Figure 4 As shown in the figure, the electronic device includes a processor 401 and a memory 402, the memory 402 is used to store program codes and data of the video recommendation based on a large language model, and the processor 401 is used to call program instructions in the memory to execute the steps shown in the method for video recommendation based on a large language model in the above embodiment.

[0171] The embodiment of the application provides a storage medium, the storage medium includes a storage program, wherein, when the program runs, the device where the storage medium is located executes the method for video recommendation based on a large language model shown in the above embodiment.

[0172] Each embodiment in the specification is described in a progressive manner, and the same and similar parts of each embodiment can be referred to each other. Each embodiment focuses on the difference from other embodiments. In particular, for the system or system embodiment, since it is basically similar to the method embodiment, it is described more simply, and the relevant part can be referred to the part of the method embodiment. The above-described system and system embodiment are only illustrative, and the units described as separate components can be or can not be physically separated, and the components shown as units can be or can not be physical units, that is, they can be located in one place or distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the embodiment according to actual needs. Those skilled in the art can understand and implement without creative labor.

[0173] The professional person can further realize that the units and algorithm steps of each example described in combination with the embodiments disclosed in the present application can be realized by electronic hardware, computer software or combination of the two. In order to clearly show the interchangeability of hardware and software, the components and steps of each example have been described in the above description. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. The professional person can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the application.

[0174] The above description of disclosed embodiments enables one of ordinary skill in the art to make or use the application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein can be applied to other embodiments without departing from the spirit or scope of the application. Therefore, the present application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0175] The above is only the preferred embodiment of the present application, it should be pointed out that, for those skilled in the art, without departing from the principles of the present application, can also make a number of improvements and refinements, these improvements and refinements should also be considered as the protection scope of the present application.

Claims

1. A method for video recommendation based on a large language model, characterized in that, The method comprises: When a recommendation request initiated by a target user is detected, at least one target user intent label data matching the target user is obtained from the stored user intent label data of each user in the key-value database according to the recommendation request; wherein the target user intent label comprises at least one content label and at least one intent label, and the user intent label data is generated according to user feedback of user text feedback data; The content labels in the target user intent label data in which the intent label is the recall strategy label are composed into a first content label set, and a first video item set is selected from all video items on the streaming media platform according to the first content label set; wherein the first video item set comprises a plurality of first video items; A first video label set of each first video item is obtained, and a second video item set is selected from each first video item according to the first content label set and the first video label set of each first video item; Target user text feedback data corresponding to the target user is obtained, and a third video item set is selected from all video items on the streaming media platform according to the target user text feedback data; A fourth video item set is selected from each video item according to each preset recall algorithm, and the second video item set, the third video item set and the fourth video item set are fused to obtain a fifth video item set; The content labels in the target user intent label data in which the intent label is the sorting strategy label are composed into a second content label set; Each fifth video item in the fifth video item set is sorted based on the second content label set to obtain a sixth video item set; A corresponding video recommendation list is generated according to the sixth video item set, and corresponding videos are recommended to the target user according to the video recommendation list.

2. The method of claim 1, wherein, The user intent label data is generated according to the user feedback of the user text feedback data, comprising: Real-time acquisition of user feedback of user text feedback data, wherein the user text feedback data comprises a plurality of user feedback texts; According to each user feedback text and a pre-set prompt word template, a corresponding prompt word is generated, and the prompt word is input into a large language model to obtain user intent label data corresponding to each user feedback text; wherein the user intent label data comprises a plurality of content labels and a plurality of intent labels; For each user intent label data, the user intent label data is verified for validity; If the validity verification of the user intent label data is passed, the user information of the user and the user intent label data are stored in the pre-constructed key-value database in the form of key-value pairs.

3. The method of claim 1, wherein, The method further comprises: obtaining the first video tag set of each first video item from a pre-set video tag database; calculating the similarity between the first video tag set of each first video item and the first content tag set by using the Jaccard similarity coefficient; selecting the second video item set from each first video item according to the similarity between the first video tag set of each first video item and the first content tag set.

4. The method of claim 3, wherein, The method further comprises: obtaining the target user text feedback data corresponding to the target user, and selecting the third video item set from all video items on the streaming media platform according to the target user text feedback data. obtaining the target user text feedback data of the target user, and calculating the latent vector of each target user feedback text in the target user text feedback data of the target user; obtaining the latent vector of each video item, and calculating the cosine similarity between the latent vector of each video item and the latent vector of each target user feedback text; 5. The method of claim 1, wherein, selecting the third video item set from all video items on the streaming media platform according to the cosine similarity between the latent vector of each video item and the latent vector of the target user feedback text. The method further comprises: obtaining the second video tag set of each fifth video item in the fifth video item set; for each fifth video item, calculating the intersection feature of each second video tag in the second video tag set of the fifth video item and the second content tag set, and accumulating the intersection feature of each second video tag of the fifth video item and the second content tag set to obtain the intersection feature of the fifth video item and the second content tag; 6. The method of claim 1, wherein, calculating the score of each fifth video item according to the intersection feature of each fifth video item and the second content tag set by using a pre-set fine arrangement model, and sorting each fifth video item according to the score of each fifth video item to obtain the sixth video item set. The method further comprises: composing the content tags in the target user intent tag data in which the existing intent tag in each target user intent tag data is the filtering strategy tag into a third content tag set; filtering the fifth video item set by using the third content tag set; correspondingly, the method further comprises: obtaining the target user text feedback data corresponding to the target user, and selecting the third video item set from all video items on the streaming media platform according to the target user text feedback data. obtaining the target user text feedback data of the target user, and calculating the latent vector of each target user feedback text in the target user text feedback data of the target user; obtaining the latent vector of each video item, and calculating the cosine similarity between the latent vector of each video item and the latent vector of each target user feedback text; selecting the third video item set from all video items on the streaming media platform according to the cosine similarity between the latent vector of each video item and the latent vector of the target user feedback text. The method further comprises: obtaining the second video tag set of each fifth video item in the fifth video item set; for each fifth video item, calculating the intersection feature of each second video tag in the second video tag set of the fifth video item and the second content tag set, and accumulating the intersection feature of each second video tag of the fifth video item and the second content tag set to obtain the intersection feature of the fifth video item and the second content tag; calculating the score of each fifth video item according to the intersection feature of each fifth video item and the second content tag set by using a pre-set fine arrangement model, and sorting each fifth video item according to the score of each fifth video item to obtain the sixth video item set. The method further comprises: composing the content tags in the target user intent tag data in which the existing intent tag in each target user intent tag data is the filtering strategy tag into a third content tag set; filtering the fifth video item set by using the third content tag set; correspondingly, the method further comprises: Sort each fifth video item in the filtered fifth video item set based on the second content tag set to obtain a sixth video item set.

7. The method of claim 1, wherein, Before the generating of the corresponding video recommendation list according to the sixth video item set, the method further comprises: determining whether there is target user intent tag data in each target user intent tag data, whose intent tag is the re-dispersal strategy tag; if there is, forming a fourth content tag set by using the content tags in each target user intent tag data, whose intent tag is the re-dispersal strategy tag; re-sorting the sixth video item set by using the fourth content tag set; correspondingly, the generating of the corresponding video recommendation list according to the sixth video item set comprises: generating the corresponding video recommendation list according to the re-sorted sixth video item set, and recommending the corresponding video to the target user according to the video recommendation list.

8. A large language model-based video recommendation system, characterized in that, The system comprises an intent execution module; The intent execution module is configured to, when detecting a recommendation request initiated by a target user, acquire at least one target user intent tag data matched with the target user from each user intent tag data stored in a key-value database according to the recommendation request; wherein the target user intent tag comprises at least one content tag and at least one intent tag, and the user intent tag data is generated according to target user text feedback data fed back by a user; form a first content tag set by using the content tags in each target user intent tag data, whose intent tag is the recall strategy tag, and filter a first video item set from all video items on a streaming media platform according to the first content tag set; wherein the first video item set comprises a plurality of first video items; acquire a first video tag set of each first video item, and filter a second video item set from each first video item according to the first content tag set and the first video tag set of each first video item; acquire target user text feedback data corresponding to the target user, and filter a third video item set from each video item according to the target user text feedback data; filter a fourth video item set from each video item according to each preset recall algorithm set in advance, and fuse the second video item set, the third video item set and the fourth video item set to obtain a fifth video item set; form a second content tag set by using the content tags in each target user intent tag data, whose intent tag is the sorting strategy tag; sort each fifth video item in the fifth video item set based on the second content tag set to obtain a sixth video item set; and generate the corresponding video recommendation list according to the sixth video item set, and recommend the corresponding video to the target user according to the video recommendation list.

9. An electronic device, comprising: It comprises: A processor and a memory connected by a communication bus, wherein the processor is configured to invoke and execute a program stored in the memory; The memory is configured to store a program, and the program is configured to implement the method for video recommendation based on a large language model according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer executable instructions, and the computer executable instructions are configured to execute the method for video recommendation based on a large language model according to any one of claims 1-7.

Images