Video recommendation method and system based on large language model, electronic equipment and storage medium

Abstract

The invention provides a video recommendation method and system based on a large language model, electronic equipment and a storage medium. The method comprises the following steps: acquiring at least one target user intention tag data from a key value database according to a recommendation request initiated by a target user; screening out a first video item set from the streaming media platform according to the first content label set; screening out 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; screening out a third video entry set from the video entries according to the target user text feedback data; screening out a fourth video entry set from the video entries according to each preset recall algorithm, and fusing the fourth video entry set with the second video entry set and the third video entry set to obtain a fifth video entry set; and recommending videos to the target user according to a video recommendation list generated by a sixth video item set obtained by sorting the fifth video item set based on the second content label set.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and more specifically, to a video recommendation method, system, electronic device and storage medium based on a large language model. Background Art

[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 streaming media platforms (such as a TV), provides users with personalized video recommendations through the video recommendation system, which can not only increase user stickiness, improve the platform's activity and user satisfaction, but also have a huge driving effect on the platform's commercial value.

[0003] Traditional video recommendation systems on streaming platforms mainly use collaborative filtering, matrix decomposition and other technologies to recommend corresponding videos to users. However, this method has certain limitations in capturing and portraying users' real-time intentions. On the one hand, user behavior is casual, which means that users' current preferences may not be consistent with historical data. For example, users may be particularly interested in videos on a certain topic during a certain period of time, but after a period of time, users' interests may shift to other topics. On the other hand, traditional recommendation systems are passive recommendations and cannot interact with users to provide users with active intention feedback. Users cannot actively tell the recommendation system their needs and preferences, which makes it difficult for the recommendation system to accurately capture user intentions and recommend videos that meet their needs and preferences, resulting in a low user experience. Summary of the invention

[0004] In view of this, the present application provides a video recommendation method, system, electronic device and storage medium based on a large language model, so as to recommend videos that meet the needs and preferences of users and improve the 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 each user intention label data stored 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 text feedback data fed back by the user;

[0007] The content tags in the target user intention tag data whose intention tags are recall strategy tags in each of the target user intention tag data 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] Acquire a first video tag set of each of the first video entries, and filter out a second video entry set from each of the first video entries according to the first content tag set and the first video tag set of each of the first video entries;

[0009] Acquire target user text feedback data corresponding to the target user, and filter out a third video item set from all video items on the streaming media platform according to the target user text feedback data;

[0010] Filtering a fourth video item set from each of the video items according to each preset recall algorithm, and fusing the second video item set, the third video item set and the fourth video item set to obtain a fifth video item set;

[0011] The content tags in the target user intention tag data having the intention tag as the sorting strategy tag in each of the target user intention tag data are combined into a second content tag set;

[0012] 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;

[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, generating the user intention label data according to the user text feedback data fed back by the user includes:

[0015] Acquire user text feedback data of user feedback in real time, wherein the user text feedback data includes multiple user feedback texts;

[0016] Generate corresponding prompt words according to each of the user feedback texts and a preset prompt word template, and input the prompt words into a large language model to obtain user intention label data corresponding to each of the user feedback texts; wherein the user intention label data includes multiple content labels and multiple intention labels;

[0017] For each of the user intention label data, verify the validity of the user intention label data;

[0018] 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 a pre-built key-value database in the form of a key-value pair.

[0019] Optionally, obtaining the first video tag set of each first video entry, and filtering out the second video entry set from each first video entry according to the first content tag set and the first video tag set of each first video entry, includes:

[0020] Acquire a first video tag set for each of the first video entries from a preset video tag database;

[0021] Calculating the similarity between the first video tag set and the first content tag set of each first video entry using the Jaccard similarity coefficient;

[0022] A second video entry set is screened out from each of the first video entries according to the similarity between the first video tag set of each of the first video entries and the first content tag set.

[0023] Optionally, the acquiring target user text feedback data corresponding to the target user, and filtering out a third video item set from all video items on the streaming media platform according to the target user text feedback data, includes:

[0024] Acquire target user text feedback data of the target user, and determine the target user feedback text of the target user from the target user text feedback data of the target user;

[0025] Calculating the latent vector of the target user feedback text;

[0026] Obtaining a latent vector of each of the video entries, and calculating the cosine similarity between the latent vector of each of the video entries and the latent vector of the target user feedback text,

[0027] A third set of video items is screened out from all video items on the streaming platform according to the cosine similarity between the latent vector of each of the video items and the latent vector of the target user feedback text.

[0028] Optionally, the step of 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 includes:

[0029] Obtaining a second video tag set for each fifth video item in the fifth video item set;

[0030] For each of the fifth video entries, calculating a cross feature between 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 features between each second video tag of the fifth video entry and the second content tag set to obtain a cross feature between the fifth video entry and the second content tag;

[0031] The score of each of the fifth video entries is calculated according to the cross features of each of the fifth video entries and the second content tag set by a preset fine ranking model, and the fifth video entries are sorted according to their scores to obtain a sixth video entry set.

[0032] Optionally, before sorting the fifth video items in the fifth video item set based on the second content tag set to obtain the sixth video item set, the method further includes:

[0033] The content tags in the target user intention tag data having the intention tag as the filtering policy tag in each of the target user intention tag data are combined into a third content tag set;

[0034] Filtering the fifth video item set using the third content tag set;

[0035] Correspondingly, the fifth video items in the fifth video item set are sorted based on the second content tag set to obtain a sixth video item set, including:

[0036] The fifth video items in the filtered fifth video item set are sorted based on the second content tag set to obtain a sixth video item set.

[0037] Optionally, before generating a corresponding video recommendation list according to the sixth video item set, the method further includes:

[0038] Determine whether there is target user intention label data whose intention label is a re-dispersion strategy label among the target user intention label data;

[0039] If so, the content tags in the target user intention tag data having the intention tag as the re-dispersion strategy tag in each of the target user intention tag data are combined into a fourth content tag set;

[0040] Reordering the sixth set of video items using the fourth set of content tags;

[0041] Correspondingly, generating a corresponding video recommendation list according to the sixth video item set includes:

[0042] A corresponding video recommendation list is generated according to the reordered sixth video item set, and corresponding videos are recommended to the target user according to the video recommendation list.

[0043] A second aspect of the present 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 used to, when a recommendation request initiated by a target user is detected, obtain at least one target user intention label data matching the target user from each user intention label data stored 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 text feedback data fed back by the user; the content labels in the target user intention label data whose intention labels are recall strategy labels in each of the target user intention label data are formed into a first content label set, and a first video item set is filtered out from all video items on the streaming media platform according to the first content label set; wherein the first video item set includes multiple first video items; obtain a first video label set for each of the first video items, and select a first video label set from the first video items according to the first content label set and the first video label set of each of the first video items. A second set of video items is screened out from each of the first video items; target user text feedback data corresponding to the target user is obtained, and a third set of video items is screened out from each of the video items according to the target user text feedback data; a fourth set of video items is screened out from each of the video items according to each preset recall algorithm, and the second set of video items, the third set of video items and the fourth set of video items are merged to obtain a fifth set of video items; content tags in the target user intention tag data whose intention tags are sorting strategy tags in each of the target user intention tag data are formed into a second content tag set; each fifth video item in the fifth video item set is sorted based on the second content tag set to obtain a sixth set of video items; a corresponding video recommendation list is generated according to the sixth set of video items, and corresponding videos are recommended to the target user according to the video recommendation list.

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

[0046] The fourth aspect of the present application provides a computer-readable storage medium, in which computer-executable instructions are stored, 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 present application provides a video recommendation method, system, electronic device and storage medium based on a large language model, which generates and stores corresponding user intention label data based on user text feedback data fed back by the user in advance, so as to interact with the user, timely understand the user's needs and preferences, and thus 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 the 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 with the content labels in the target user intention label data with the intention label being a recall strategy label in each of the target user intention label data, and filter out 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 out videos that the target user may like from all video items on the streaming media platform according to the target user's preferences and demand content labels; in order to further improve the user's experience, a first video label set of each of the first video items can be further obtained, and a second video item set can be filtered out from each of the first video items according to the first content label set and the first video label set of each of the first video items, and the second video item set can be filtered out from each of the first video items according to the first content label set and the first video label set of each of the first video items, and the second video item set can be filtered out from each of the first video items according to the first content label set and the first video label set of each of the first video items. When the target user text feedback data corresponding to the target user is obtained, a third video item set is screened out from all video items on the streaming media platform according to the target user text feedback data, a fourth video item set is screened out from all video items on the streaming media platform according to each preset recall algorithm, and the second video item set, the third video item set and the fourth video item set are merged to obtain a fifth video item set; in order to further improve the user experience, the content tags in the target user intention tag data whose intention tags are sorting strategy tags in each of the target user intention tag data are formed into a second content tag set, so as to sort the fifth video item set based on the second content tag set to obtain a sixth video item set, that is, by re-sorting the fifth video item set according to the content tags in the target user intention tag data whose intention tags are sorting strategy tags, it can be further ensured that the target user can see the video that he wants to see more first, thereby improving the user experience; a corresponding video recommendation list is generated according to the sixth video item set, and the corresponding video is recommended to the target user according to the video recommendation list, thereby achieving the recommendation of videos that meet the needs and preferences of the target user, thereby improving the user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

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

[0050] Figure 2 A schematic diagram of a process of video recommendation based on a large language model provided in an embodiment of the present application;

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

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

[0053] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0054] In this application, relational terms such as first and second, etc. are used only 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 these entities or operations. Moreover, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the statement "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device including the element.

[0055] See also Figure 1 , showing a schematic diagram of the structure of a video recommendation system based on a large language model provided in an embodiment of the present application, the video recommendation system based on a large language model includes a user interaction module, an intention recognition module, an intention memory module, an intention execution module and a user feedback module;

[0056] An intention execution module is used to, when a recommendation request initiated by a target user is detected, obtain at least one target user intention label data matching 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 includes 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 the user; the content labels in the target user intention label data whose intention labels are recall strategy labels in each of the target user intention label data are formed into a first content label set, and a first video item set is filtered out from all video items on the streaming media platform according to the first content label set; wherein the first video item set includes multiple first video items; obtain a first video label set for each of the first video items, and select a first video label set from each of the first video items according to the first content label set and the first video label set of each of the first video items. A second set of video items is filtered out from the first video items; target user text feedback data corresponding to the target user is obtained, and a third set of video items is filtered out from all video items on the streaming media platform according to the target user text feedback data; a fourth set of video items is filtered out from each of the video items according to each preset recall algorithm, and the second set of video items, the third set of video items and the fourth set of video items are merged to obtain a fifth set of video items; content tags in the target user intention tag data whose intention tags are sorting strategy tags in each of the target user intention tag data are formed into a second content tag set; each fifth video item in the fifth video item set is sorted based on the second content tag set to obtain a sixth set of video items; a corresponding video recommendation list is generated according to the sixth set of video items, and corresponding videos are recommended to the target user according to the video recommendation list.

[0057] Optionally, a user interaction module is used to obtain user text feedback data of user feedback in real time, wherein the user text feedback data includes multiple user feedback texts;

[0058] The intention recognition module is used to generate corresponding prompt words according to each user feedback text and a preset prompt word template, and input the prompt words into the large language model to obtain user intention label data corresponding to each user feedback text; wherein the user intention label data includes multiple content labels and multiple intention labels;

[0059] The intent memory module is used to verify the validity of each user intent label data; if the validity verification of the user intent label data is passed, the user information and the user intent label data are stored in a pre-built key-value database in the form of a key-value pair.

[0060] Optionally, obtaining a first video tag set of each first video entry, and filtering out an intention execution module of a second video entry set from each first video entry according to the first content tag set and the first video tag set of each first video entry, is specifically used for:

[0061] The first video tag set of each first video entry is obtained from a preset video tag database; the similarity between the first video tag set and the first content tag set of each first video entry is calculated using the Jaccard similarity coefficient; and the second video entry set is screened out from each first video entry according to the similarity between the first video tag set and the first content tag set of each first video entry.

[0062] Optionally, the intention execution module of obtaining target user text feedback data corresponding to the target user and filtering out the third video item set from all video items on the streaming media platform according to the target user text feedback data of the target user is specifically used for:

[0063] Obtain target user text feedback data of the target user, and 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; obtain the latent vector of each video entry, and calculate the cosine similarity between the latent vector of each video entry and the latent vector of the target user feedback text, and filter out a third video entry set from all video entries on the streaming 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, sorting each fifth video item in the fifth video item set based on the second content tag set to obtain an intention execution module of the sixth video item set is specifically used to:

[0065] Obtain a second video tag set for each fifth video entry in the fifth video entry set; for each fifth video entry, calculate a cross feature between each second video tag in the second video tag set of the fifth video entry and a second content tag set, and accumulate the cross features between each second video tag of the fifth video entry and the second content tag set to obtain a cross feature between the fifth video entry and the second content tag; calculate a score for each fifth video entry according to the cross features between each fifth video entry and the second content tag set using a preset fine ranking model, and sort each fifth video entry according to the score to obtain a 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 the sixth video item set, the intention execution module is further configured to:

[0067] The content tags in the target user intention tag data having the intention tag as the filtering policy tag in each of the target user intention tag data are combined into a third content tag set; and the fifth video item set is filtered 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] Determine whether there is target user intention label data with an intention label that is a re-dispersal strategy label in each of the target user intention label data; if so, group the content labels in the target user intention label data with an intention label that is a re-dispersal strategy label in each of the target user intention label data into a fourth content label set; and use the fourth content label set to reorder the sixth video entry set.

[0070] Optionally, a user feedback module is used to: when recommending corresponding videos to target users according to the video recommendation list, display a preset prompt text next to or above the recommended videos to prompt the target users why these videos are recommended and ask the users whether to recommend the strategy.

[0071] See also Figure 2 , shows a flow chart of a video recommendation method based on a large language model provided in an embodiment of the present application, 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 label data matching the target user is obtained from each user intention label data stored 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 based on user text feedback data fed back by the user.

[0073] In an embodiment of the present application, a corresponding key-value database can be pre-constructed, and user intention label data generated based on user text feedback data obtained in real time can be pre-stored in the key-value database. At the same time, the user text feedback data of the user can also be stored in a pre-set key-value database.

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

[0075] It should be noted that by using the large language model to accurately capture the user's real-time intention and emotion from the user feedback text, the accuracy of personalized video recommendations can be enhanced. Among them, the large language model technology is based on deep learning algorithms such as Bert and Transformer, and conducts large-scale neural network model training on a large amount of text data. It can understand and generate natural language and achieve a deep understanding of user intentions.

[0076] In some embodiments, the pre-set key-value database can be a distributed key-value storage engine. After obtaining user text feedback data of user feedback in real time, the user text feedback data of the user feedback can be stored in the distributed key-value storage engine. Moreover, when it is determined that the validity verification of the user intention label data obtained using the user feedback text in the user text feedback data of the user feedback has passed, the user intention label data is stored in the distributed key-value storage engine.

[0077] It should be noted that the user feedback text may be the user opinion text of the user's feedback on the recommended content exposed by the streaming media platform, or the user opinion text of the user's feedback on the recommended content with long clicks, or the search text entered by the user and the comment text posted by the user on 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., which is not limited in the embodiments of this application.

[0079] In some embodiments, when it is detected that a large amount of recommended content is exposed by the streaming platform but no user clicks on the recommended content exposed by the streaming platform, user opinion texts of users' feedback on the recommended content exposed by the streaming platform are collected in natural language form, that is, an active natural language feedback module is triggered, and user opinion texts of users' feedback on the recommended content exposed by the streaming platform are collected in text or language dialog boxes; when it is detected that a large amount of recommended content is exposed by the streaming platform and there are users who give long-click feedback on the recommended content exposed by the streaming platform, user opinion texts of users' feedback on the long-clicked recommended content are collected in natural language form; when it is detected that a user actively searches in the search bar of the streaming platform, the search text input by the user is collected in natural language form, that is, the active natural language feedback module is triggered to collect the search text actively input by the user; when it is detected that a user posts corresponding comments on the recommended content exposed by the streaming platform, the comment texts of users on the recommended content exposed by the streaming platform are collected in natural language form, that is, the active natural language feedback module is triggered to collect the comment text actively input by the user, and finally the user text feedback data is generated according to the collected user feedback text, search text and comment text of the user.

[0080] It should be noted that the user feedback text may be a desire to watch more videos related to a certain star, a dislike for watching videos related to a certain star, etc.; the search text may be a reality show of a certain star, a sketch of a certain sketch actor, etc.; the comment text may be that a certain sketch of a certain sketch actor is great, a certain star's acting skills in a certain TV series are really amazing, etc. The embodiments of this application do not limit the user feedback text, search text, and comment text.

[0081] In actual application, when it is detected that a large amount of recommended content is exposed on the streaming platform but no user clicks on the recommended content exposed on the streaming platform, the user can be invited to give feedback on the recommended content exposed on the streaming platform in the form of a pop-up window or prompt bar. That is, when it is detected that a large amount of recommended content is exposed on the streaming platform but no user clicks on the recommended content exposed on the streaming platform, a corresponding pop-up window or prompt bar pops up to prompt the user to enter the corresponding user feedback text in the pop-up window or prompt bar. In the process of users browsing the recommended content exposed on the streaming platform, if the user is confused or dissatisfied with the recommended content exposed on the streaming platform, the user can long-click on the recommended content that causes doubt or dissatisfaction, so as to detect that the user has long-clicked on the recommended content exposed on the streaming platform, so that the video recommendation system based on the large language model can pop up a corresponding pop-up window or prompt bar when it detects that the user has long-clicked on a certain recommended content, so as to prompt the user to enter the corresponding user feedback text in the pop-up window or prompt bar.

[0082] In some embodiments, a set of prompt word templates can be pre-set so that after obtaining the user text feedback data of the user feedback, each user feedback text in the user text feedback data is filled into the pre-set prompt word template respectively 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 that can reflect the user's recommendation preference in each user feedback text in the prompt word, and 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 the content labels and the intent labels corresponding to each user feedback text, that is, multiple user intent label data corresponding to the user are obtained. Among them, 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 filtering strategy label ('filter'), or a sorting strategy label ('rank'), or a re-shuffle strategy label ('rerank'). Correspondingly, if the intent label is a recall strategy label, the corresponding recommendation algorithm can be the Jaccard similarity coefficient; if the intent label is a filtering strategy label, the corresponding recommendation algorithm can be the filtering algorithm; if the intent label is a sorting strategy label, the corresponding recommendation algorithm can be the sorting algorithm; if the intent label is a re-shuffle strategy label, the corresponding recommendation algorithm can be the re-shuffle algorithm.

[0084] It should also be noted that the recall algorithm can be CF, youtubeDNN, Eges, Item2Vec, Swing, graph neural network and other recall algorithms, which are not limited in the embodiments of this application.

[0085] In actual application, if the large language model recognizes from the user feedback text in the input prompt word that the user is interested in specific content or clearly expresses the desire to see more similar content, it can be determined that the intent label corresponding to the user feedback text may include a recall intent label. For example, if the user feedback text in the prompt word is that the user likes a certain star's video, then the large language model can recognize from the user feedback text in the prompt word that the user is interested in the certain star's video. 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 recognizes from the user feedback text in the input prompt word that the user is dissatisfied with, or does not like, or clearly indicates that he does not want to see more types of content with a specific content, it can be determined that the intent label corresponding to the user feedback text may include a filtering intent label. For example, if the user feedback text in the prompt word indicates that the user does not want to see a video of a certain star, then the large language model can recognize from the user feedback text in the prompt word that the user is sure that he does not want to see a video of a certain star. 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 recognizes from the user feedback text in the input prompt word that the user prefers specific content over other content, or requires specific content to be ranked higher in the recommendation category table, it can be determined that the intent label corresponding to the user feedback text may include a sorting intent label. For example, if the user feedback text in the prompt word is that the user wants to see more videos of a certain star, then the large language model can recognize from the user feedback text of the prompt word that the user not only clearly expressed interest in the videos of the certain star, but also clearly expressed a higher preference for the videos of the 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 a sorting intent label.

[0088] If the large language model recognizes from the user feedback text in the input prompt word that the user is dissatisfied with the current ranking results of a large amount of recommended content currently exposed on the streaming platform and hopes to readjust the order of the recommended content, it can be determined that the intent tag corresponding to the user feedback text may include the re-dispersed intent tag.

[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 preset prompt word template and each user feedback text, the prompt word is input into the large language model, so that the large language model analyzes each user feedback text in the input prompt word, and obtains at least one content tag for describing the preference of each user feedback text. Among them, at least one content tag may include: time tag, geographic tag, video type tag, theme keyword tag, whether the preference tendency expressed by the text is positive emotion or negative emotion tag, drama tag, episode tag, star tag, video effect tag, plot tag, etc., which are not limited in this embodiment of the present application.

[0090] In actual application, for positive or negative emotion labels: if the large language model identifies from the user feedback text in the input prompt word that the preference tendency expressed by the user feedback text is positive, then the large language model can feedback the corresponding positive emotion label in the form of IS_POSITIVE_DATA; if it identifies that the preference tendency expressed by the user feedback text is negative, then the large language model can feedback the corresponding positive emotion label in the form of IS_POSITIVE_DATA, where IS_POSITIVE_DATA is an integer number, 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 the user's emotional tendency from the user feedback text in the input prompt word, if the label information expressing the user's emotional tendency is positive, then the large language model outputs 1, if the label information expressing the user's emotional tendency is negative, then the large language model outputs 0.

[0091] For example, the user feedback text is: 'I want to watch I am a Great Detective'. At this time, the large language model can identify from the prompt words generated by the user feedback text that the preference tendency expressed by the user feedback text is positive, and the large language model can output 1; or the user feedback text is: 'Why do you recommend so many romantic TV series to me'. At this time, the large language model can identify from the prompt words generated by the user feedback text that the preference tendency expressed by the user feedback text is negative, and the large language model can output 0.

[0092] For content tag extraction within a fixed content tag range: If the large language model recognizes the time, geographic location, video type that the text focuses on, and topic keywords that the text focuses on from the user feedback text in the input prompt word, then the large language model can select the corresponding geographic tags, video type tags, topic keyword tags, etc. from the pre-set content tag range, and feedback the corresponding geographic tags, video type tags, and topic keyword tags in the form of CATEGORY_DATA. If the large language model is unable to recognize the corresponding content tags from the user feedback text, it can feedback an empty value list; where CATEGORY_DATA is in the string list format. The range of pre-set content tags can be ['geography', 'topic keywords', 'variety show', 'TV series', movie', 'children', 'anime', 'news', 'documentary', 'music', 'live broadcast', 'entertainment', 'life', 'game', 'education', 'fashion'], 'TV series', 'movie ... are video types, and the embodiment of the present application does not limit the range of pre-set content tags.

[0093] For example, if the user feedback text is: 'Han Han, we will never meet again', the large language model can identify the video type "movie" that the text focuses on from the prompt words generated based on the user feedback text. At this time, the large language model can select the corresponding video type tag from the pre-set content tag range and feedback the corresponding video type tag in the form of CATEGORY_DATA. Alternatively, if the user feedback text is: 'T1 wins the Saudi Cup by defeating the LPL team twice', the large language model can identify the video types that the text focuses on as "games" and "live broadcasts" from the prompt words generated based on the user feedback text. At this time, the large language model can select the corresponding video type tag from the pre-set content tag range and feedback the corresponding video type tag in the form of CATEGORY_DATA.

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

[0095] For example, if the user feedback text is: 'I want to watch the latest episode of Singer 2024', the large language model can extract the key tag 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 feedback the corresponding time tag in the form of TIME_DATA based on the extracted key tag information; or if the user feedback text is: 'What are the latest variety shows today', the large language model can extract the key tag 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 feedback the corresponding time tag in the form of TIME_DATA based on the extracted key tag information. Or, if 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 tag information "2024" related to time and the topic keyword "User 1" from the prompt words generated according to the user feedback text. At this time, the large language model can generate and feedback the corresponding time tag in the form of TIME_DATA based on the extracted key tag information, and feedback the corresponding keyword tag in the form of CATEGORY_DATA.

[0096] In some embodiments, after using a large language model to output corresponding user intent label data based on the input prompt word, the user intent label data output by the large language model can be further verified for validity, so that if it is determined that the user intent label data has passed the validity verification, the user intent label data can be stored in a pre-built key-value database.

[0097] In an embodiment of the present application, the label format and label content of the user intention label data can be verified; if the label format and label content of the user intention label data are both verified, it is determined that the validity verification of the user intention label data has passed; otherwise, it is determined that the validity verification of the user intention label data has failed.

[0098] It should be noted that this application can use JSON Schema verification technology to verify the label format of user intent label data. Specifically, by defining a set of strict patterns (schemas) to constrain the data structure, ensure that the field type, necessity and value range of the output data conform to the expected data format. If a data format error is found, the exception handling process is triggered, and the user intent label data with data format errors is saved as log information in a specified file for technical personnel to check and adjust. Among them, the expected data format can be in JSON format.

[0099] Specifically, after using a large language model to output corresponding user intention label data according to the user feedback text in the input prompt word, it can be determined whether the label format of the user intention label data is in JSON format, and whether there are content labels and intent labels in the user intention label data; if the label format of the user intention label data is in JSON format, and there are content labels and intent labels in the label content of the user intention label data, it can be determined that both the label format and label content of the user intention label data have been verified, and then it can be determined that the validity verification of the user intention label data has been passed; if the label format of the user intention label is not in JOSN format, and / or there are no content labels or intent labels in the label content of the user intention label data, it can be determined that the label format and / or label content verification of the user intention label data have not passed, and then it can be determined that the validity verification of the user intention label data has not passed. Among them, the content labels and intent labels in the user intention label data are labels in a pre-set label library.

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

[0101] It should be noted that, when it is determined that the validity verification of the user intention label data has failed, the corresponding exception handling process can be triggered to use the OSS object storage service to record and store the user intention label data that has failed the validity verification in a designated file of the distributed file system to ensure the security and persistence of the data, and provide a query and analysis interface for operators or technical personnel to verify or update the corresponding label library, wherein the distributed file system can be Amazon S3, Google Cloud Storage, Alibaba Cloud OSS and other OSS services, which are not limited in the embodiments of this application; when it is determined that the validity verification of the user intention label data has passed, the user ID of the corresponding user can be used as the key value, and the distributed key-value storage engine (Redis) can be used to store the user intention label data and user text feedback data.

[0102] It should also be noted that when new user intent label data or user text feedback data is written to the distributed key-value storage engine, the conflicting key-value pairs in the old data can be automatically overwritten to ensure real-time data updates, ensure that the system promptly reflects the user's latest preferences, and improve the timeliness of video recommendations.

[0103] In the specific process of 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 the target user). When a recommendation request initiated by the target user is detected, the recommendation engine can obtain at least one target user intention label data matching the target user from each user intention label data stored in the key-value database according to the user ID of the target user carried in the recommendation request (for the sake of distinction, the user intention label data matching the target user is referred to as the target user intention label data).

[0104] S202: The content tags in the target user intention tag data whose intention tags are recall strategy tags in each target user intention tag data are combined into a first content tag set, and a first video item set is filtered out from all video items on the streaming media platform according to the first content tag set.

[0105] In the process of specifically executing step S202, after obtaining at least one target user intention label data matching the target user UK, the recommendation engine can filter out at least one target user intention label data having an intention label as a recall strategy label from each target user intention label data, and the content labels in each target user intention label data having an intention label as a recall strategy label are combined into a first content label set; based on each content label in the first content label set, multiple first video entries are filtered out from all video entries on the streaming media platform, and each filtered first video entry is combined into a first video entry set.

[0106] In the embodiment 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, and the key value index of each video entry is a video tag of each video entry, and each video entry includes multiple video tags, as shown in the following formula:

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

[0108] It can be seen from Formula 1 that for each video entry in the video entry set of the streaming media platform, the video entry has multiple video tags, and the video tags of the video entries are video tags in the pre-set video tag 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 items in the video item set, it is also possible to perform deep embedding on each video item through natural language processing (NLP) technology, such as BRET and other pre-trained models, to obtain a latent vector for each video item, and construct a latent vector set of the video item set based on the latent vectors of each video item, as shown in formula (2):

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

[0111] Among them, E is the latent vector set of the video item set, E V1 is the latent vector of video entry 1, E V2 is the latent vector of video entry 2,...,E Vn is the latent vector of video entry n.

[0112] In an embodiment of the present application, assuming that the 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 uses the recommendation engine to filter out at least one target user intention label data whose intention label is a recall strategy label from each target user intention label data, and the content labels in each target user intention label data whose intention label is a recall strategy label are combined into a first content label set Y = {Y1, Y2, ..., Yr}, where r is greater than or equal to 2; using each content label in the first content label set, recall all first video item sets V that may overlap with the target user preferences from the video item set V UK1 = {V1 UK1 ,V2 UK1 ,......,VN UK1}, wherein N is greater than 1 and less than or equal to n, and the video tag of each first video entry includes at least one content tag in the first content tag set, as shown in the following formula:

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

[0114] Among them, j is greater than 1 and less than or equal to m, UK is the target user, Vi UK1 is the first video entry set V UK1 Any first video entry in the set of first video entries, it can be seen from formula (3) that the first video tag set of each first video entry in the set of first video entries is Include multiple video tags {V1 T1 ,...,V Tj}, and the multiple video tags of the first video entry include at least one content tag in the first content tag set.

[0115] S203: Acquire a first video tag set of each first video entry, and filter out a second video entry set from each first video entry according to the first content tag set and the first video tag set of each first video entry.

[0116] In the specific process of executing step S203, after obtaining the first video item set, the video recommendation system based on the large language model can perform an association calculation between the first content tag set and the first video tag set of each first video item in the first video item set, so as to filter out the second video item set from the first video item set.

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

[0118] Optionally, the process of selecting 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 may specifically be: using the Jaccard similarity coefficient to calculate the similarity between the first video tag set of each first video item and the first content tag set; and 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. The method of calculating the similarity between the first video tag set of each first video item and the first content tag set using the Jaccard similarity coefficient is shown in formula (4).

[0119]

[0120] Among them, S1(Vi UK1 ) is the similarity between the first video tag set and the first content tag set of the i-th first video entry, Y is the first content tag set, is the first video tag set of the i-th first video entry.

[0121] In some embodiments, after obtaining the similarity between the first video tag set and the first content tag set of each first video entry, the first video entries can be sorted from high to low according to the similarity, and the first M first video entries in the sorted sequence are used as the second video entry set.

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

[0123] In the specific process of executing 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 through natural language processing (NLP) technology, such as pre-trained models such as BRET, each target user feedback text in the target user text feedback data of the target user is deeply embedded to obtain a latent vector of each target user feedback text, and a third video entry set is screened out from each video entry according to the latent vector of each target user feedback text.

[0124] Optionally, the process of obtaining target user text feedback data corresponding to the target user, and filtering out a third video item set from each video item based on the target user text feedback data can specifically be: 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 out a third video item set from each video item based on 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 through natural language processing technology (such as pre-trained models such as BERT) to capture the semantic features and contextual associations of the target user feedback text, and obtain the latent vector of the target user feedback text; the latent vector of each video item is obtained from the latent vector set of the preset 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 to obtain multiple cosine similarities of the video item; for each For a video entry, if all cosine similarities of the video entry are zero, the cosine similarity between the latent vector of the video entry and the latent vector of the target user feedback text most recently fed back by the target user is taken as the final cosine similarity of the video entry; if there is a non-zero cosine similarity among the cosine similarities of the video entry, and there are multiple non-zero cosine similarities, the cosine similarity corresponding to the target user feedback text most recently fed back by the user can be determined as the final cosine similarity of the video entry from the target user feedback texts corresponding to the non-zero cosine similarities; if there is a non-zero cosine similarity among the cosine similarities of the video entry, 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 entry.

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

[0127]

[0128] Where S2(Vi) is the cosine similarity between the latent vector of the i-th video item and the latent vector of the target user feedback text, E Vi is the latent vector of the i-th video entry, is the latent vector of the sth target user feedback text in the target user text feedback data.

[0129] S205: Filter out a fourth video item set from each video item according to each preset recall algorithm, and merge the second video item set, the third video item set and the fourth video item set to obtain a fifth video item set.

[0130] In an embodiment of the present application, multiple preset recall algorithms can be pre-set, wherein the preset recall algorithm can be CF, youtubeDNN, Eges, Item2Vec, Swing, graph neural network and other recall algorithms, which are not limited in the embodiment of the present application.

[0131] In the specific process of executing step S205, after obtaining the second video item set and the third video item set, the fourth video item set can be screened out from the various video items using the pre-set preset recall algorithms, so as to merge the obtained second video item set, the third video item set and the fourth video item set to obtain a merged fifth video item set.

[0132] S206: The content tags in the target user intention tag data whose intention tags are sorting strategy tags in each target user intention tag data are combined into a second content tag set.

[0133] In the process of specifically executing step S206, at least one target user intention label data whose intention label is a sorting strategy label can be screened out from each target user intention label data, and the content labels in each target user intention label data whose intention label is a sorting strategy label are combined into a second content label set H = {H1, H2, ..., Hb}, where b is greater than or equal to 2.

[0134] S207: Sort the fifth video items in the fifth video item set based on the second content tag set to obtain a sixth video item set.

[0135] In the specific process of executing step S207, after obtaining the fifth video entry set, for each fifth video entry in the fifth video entry set (for ease of distinction, each video entry in the fifth video entry set is referred to as the fifth video entry), the second video tag set of the fifth video entry can be obtained from the preset video tag set, so as to call the preset sorting algorithm to sort each fifth video entry according to the second video tag set and the second content tag set of each fifth video entry to obtain a sixth video entry 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 specifically be: 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-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 cross-features of each second video tag of the fifth video item and the second content tag set to obtain the cross-features of the fifth video item and the second content tag; calculating the score of each fifth video item according to the cross-features of each fifth video item and the second content tag set by using a preset refined 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 through a second content tag set composed of content tags in the target user intention tag data whose intention tag is a sorting strategy tag, it can further ensure that users can give priority to watching the videos they want to watch more, 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 is detected can also be obtained, that is, the time point at which the target user initiates the recommendation request; so that in the subsequent process of calculating the cross-feature between each second video tag in the second video tag set of the fifth video item and the second content tag, a pre-set sorting algorithm can be called to calculate the cross-feature between 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 item, and the cross-feature between each second video tag and the second content tag set is accumulated to obtain the cross-feature between the second video item and the second content tag. The specific process is shown in formula (6):

[0139]

[0140] Among them, VUK2 represents the fifth video entry in the fifth set of video entries, represents the second video tag in the second video tag set of the fifth video item, g represents a second video tag in the second video tag set, t represents the time point when the recommendation request is initiated, Indicates the timestamp of tag g in the second content tag set. Function Indicates whether the tag g exists in the second content tag set H: Function p(δt) is a time penalty function used to increase the weight of recent user intention labels: p(δt) = exp(-0.035·δt) The numerical feature f(V UK2 , UK) will be used by the refined ranking model to increase the weight of the content that the user is interested in in the refined ranking model. Among them, the refined ranking model can be Wide&Deep, DIN, DIEN, DeepFM, which is not limited in this embodiment of the application.

[0141] It should be noted that the sorting algorithm usually builds feature engineering first, then trains the refined ranking model, and after the trained refined ranking model is deployed online, the ranking score of each recommended result (fifth video item) in the recommended candidate set (fifth video item set) is predicted in real time. Among them, feature engineering usually includes ID features and user's historical behavior data (such as viewing history, likes, comments, searches, payment ability, etc.), user characteristics (such as age, gender, geographic location, etc.), video characteristics (such as title, tag, category, etc.) and other information.

[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, a refined ranking model can be used to calculate the score of each fifth video item based on 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 the fifth video items in the fifth video item set from high to low according to the scores of the fifth video items to obtain a sixth video item set.

[0143] Furthermore, in an embodiment of the present application, before calculating the cross-features between each fifth video item in the fifth video item set and the second content tag set, the fifth video item set may be filtered first, so as to subsequently calculate the cross-features between each fifth video item in the filtered fifth video item set and the second content tag set.

[0144] As a method of an embodiment of the present application, at least one target user intention label data having an intention label as a filtering policy label can be screened out from each target user intention label data, and the content labels in each target user intention label data having an intention label as a filtering policy label are combined 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 using the third content label set.

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

[0146] As another embodiment of the present application, corresponding filtering rules may be pre-set, and a filtering algorithm may be called to filter the fifth video item set using the filtering rules.

[0147] Specifically, the target user text feedback data of the target user can be obtained through the recommendation engine, and the user feedback texts indicating that the user is dissatisfied with the feature content, or does not like it, or clearly indicates that he does not want to see more types of content can be identified from each user feedback text in the target user text feedback data, and the corresponding filtering rules can be constructed based on the user feedback texts indicating that the user is dissatisfied with the feature content, or does not like it, or clearly indicates that he does not want to see more types of content, so that the filtering algorithm can be called to use the filtering rules to filter out the fifth video items that the user is dissatisfied with, does not like, or clearly indicates that he does not want to see from the fifth video item set.

[0148] It should be noted that, by analyzing the target user text feedback data of the target user through the filtering algorithm, user feedback text with negative feedback from the target user is determined, such as "I don't want to watch it recently", "I don't like a certain type of result", "I don't like videos with comment counts below a certain threshold", "I don't like videos with false content", etc., and corresponding filtering rules are constructed based on the user feedback text with negative user feedback, so as to use the filtering rules in combination with the machine learning model to conduct in-depth analysis of the negative user feedback text, and dynamically schedule the filtering algorithm to filter the fifth video item set, so as to intelligently eliminate video items that the target user may not like or do not want to see from the fifth video item set.

[0149] It should also be noted that by filtering the fifth video entry set through a third content tag set or filtering rules formed by content tags in the target user intention tag data with an intention tag as a filtering policy tag, it can effectively reduce the situation where users see videos they don't like, and can further improve the user experience.

[0150] S208: Generate a corresponding video recommendation list according to the sixth video item set, and recommend corresponding videos to the target user according to the video recommendation list.

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

[0152] Furthermore, in an embodiment of the present application, before generating a corresponding video recommendation list based on the sixth video item set, it is also possible to further determine whether there is target user intention label data with an intention label that is a re-dispersal strategy label in the intention label data of each target user; if so, the content labels in the target user intention label data with an intention label that is a re-dispersal strategy label in the intention label data of each target user are combined into a fourth content label set; the sixth video item set is reordered using the fourth content label set, so as to generate a corresponding video recommendation list based on 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 are target user intention label data whose intention labels are re-dispersed strategy labels in each target user intention label data, the content labels in the target user intention label data whose intention labels are re-dispersed strategy labels in each target user intention label data can be combined into a fourth content label set D = {D1, D2, ..., Dx}, where x is greater than or equal to 2.

[0154] As one embodiment of the present application, it is determined that at least one video item V matching the content tag in the fourth content tag set is selected from the sixth video item set. UK3 = {V1 UK3 ,...,Vz UK3}, where z is greater than 1; if so, it can be further determined whether there is a content tag in the fourth content tag set indicating that the corresponding type of video is placed on the first screen. If so, a weighted scattering algorithm is called to perform weighted calculation on each video entry that matches the content tag indicating that the corresponding type of video is placed on the first screen to obtain a corresponding score, 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 that matches the content tag that places the corresponding type of video on the first screen is placed at the front of each video entry in the sixth video set, that is, all video entries that match the content tag that places the corresponding type of video on the first screen are inserted into the first screen.

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

[0156] Among them, V UK3 For video entries that match content tags in the fourth content tag set, score raw is the original score of the video entry. If the score ranges from {0,1}, then p can be the weighted weight for implementing the insert first screen function; wherein, p can be set according to the weighting requirements, which is not limited in this embodiment of the present application.

[0157] For example, when the fourth content tag set contains a content tag "Put celebrity i's videos on the first screen", a weighted scattering algorithm can be used to force extraction of video entries that match the content tag "results for celebrity i" from the sixth video entry set, and all the extracted video entries can be inserted into the first screen, i.e., the extracted video entries can be placed at the front of all video entries in the sixth video entry set. In this way, when making subsequent video recommendations, users can see videos related to celebrity i first, which can further improve the user experience.

[0158] As another embodiment of the present application, it is determined that at least one video item matching the content tag in the fourth content tag set is selected from the sixth video item set; if so, it is further determined whether there is a content tag in the fourth content tag set indicating that a video of the corresponding type is placed at a specified position; if so, the pit position designated re-scattering algorithm is called to determine the indication that the score of the video item corresponding to the specified position N in the sixth video item set will be obtained. N and the score of the video entry at the previous position of the specified position N-1 ; and the scores of the corresponding types of video entries are intervened as {score N ,score N-1} to put the video entry of the corresponding type into the specified position N, so that when the video is recommended later, the user can see the video he wants to watch at the specified position, which can further improve the user experience.

[0159] It should be noted that in addition to the weighted heavy scattering algorithm and the pit-specified heavy scattering algorithm, the heavy scattering algorithm can be a variety of algorithms such as star scattering, video type scattering, long and short video scattering, video and multi-function module scattering, etc., which are not limited in this embodiment of the present application.

[0160] Furthermore, in an embodiment of the present application, when recommending corresponding videos to the target user according to the video recommendation list, a pre-set prompt text can also be displayed next to or above the recommended video to prompt the target user why these videos are recommended, and ask the user whether to recommend the strategy, so that the target user can know more clearly why they see these recommended videos; and the target user can also choose whether to continue to accept this interest-based (user feedback text) video recommendation.

[0161] The present application provides a video recommendation method based on a large language model, which generates and stores corresponding user intention label data based on user text feedback data fed back by the user in advance, so as to interact with the user and timely understand the user's needs and preferences, so as to accurately capture the user's intention label and content label according to the preferences and needs actively fed back by the target user, that is, by using a large language model, it is possible to understand the user's language expression, capture the user's intention and emotion, and provide the user with a more intelligent service; when a recommendation request initiated by the 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 the key-value database according to the recommendation request, so as to form a first content label set with the content labels in the target user intention label data whose intention labels are recall strategy labels in each target user intention label data, and filter out 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 out videos that the target user may like from all video items on the streaming media platform according to the target user's preferences and demand content labels; in order to further improve the user experience, a first video label set of each first video item can be further obtained , and according to the first content tag set and the first video tag set of each first video item, a second video item set is screened out from each first video item, and at the same time, the target user text feedback data corresponding to the target user can be obtained, and a third video item set is screened out from all video items on the streaming media platform according to the target user text feedback data; in order to further improve the user experience, the content tags in the target user intention tag data with the intention tag as the sorting strategy tag in each of the target user intention tag data are combined into a second content tag set, so as to sort the fifth video item set based on the second content tag set to obtain a sixth video item set, that is, by re-sorting the fifth video item according to the content tags in the target user intention tag data with the intention tag as the sorting strategy tag to form a second content tag set, the sixth video item set can be obtained, which can further ensure that the target user can give priority to the video he wants to see more, thereby improving the user experience; a corresponding video recommendation list is generated according to the sixth video item set, and the corresponding video is recommended to the target user according to the video recommendation list, thereby achieving the recommendation of videos that meet the needs and preferences of the target user, thereby improving the user experience.

[0162] In order to better understand the video recommendation method based on a large language model provided in the embodiment of the present application, the following is an example. Figure 3 As shown:

[0163] When the user interaction module detects that a large amount of recommended content is exposed on the streaming platform but no user clicks on the recommended content exposed on the streaming platform, and / or detects that a large amount of recommended content is exposed on the streaming platform and there are users who give long-click feedback on the recommended content exposed on the streaming platform, the active natural language feedback module is triggered to collect the user opinion text of the user feedback on the recommended content exposed on the streaming platform through text or language dialog boxes. And when it is detected that the user actively searches in the search bar of the streaming platform, and / or it is detected that the user makes corresponding comments on the recommended content exposed on the streaming platform, the active natural language feedback module is triggered to collect the search text and / or comment text actively input by the user; finally, the user text feedback data generated based on the collected user feedback text, search text and comment text of the user is sent to the intention recognition module.

[0164] After obtaining the user text feedback data of the user feedback, the intention recognition module fills each user feedback text in the user text feedback data into a preset prompt word template, obtains a corresponding prompt word, and inputs the prompt word into the large language model; the large language model recognizes each user feedback text in the prompt word, and obtains user intention label data corresponding to each user feedback text, wherein the user intention label data includes at least one intention label and at least one content label; at least one intention label may include a recall strategy label ('recall'), or a filtering strategy label ('filter'), or a sorting strategy label ('rank'), or a re-ranking strategy label ('rerank'), and at least one content label may include: a time label, a geographic label, a video type label, a theme keyword label, a positive or negative emotion label expressing a preference tendency in the text, a celebrity label, etc.

[0165] The intent memory module verifies the label format and label content of the obtained user intent label data; if the label format verification of the user intent label data passes, and the label content verification of the user intent label data passes, 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 key-value pairs; if the label format of the user intent label is not in JOSN format, it can be determined that the label format verification of the user intent label data fails, and the user intent label data can be recorded and stored in a designated file of the distributed file system; and / or, there is no content label or intent label in the label content of the user intent label data, and the user intent label data can be recorded and stored in a designated file of the distributed file system.

[0166] The intention execution module is pre-set with multiple preset recall algorithms, filtering algorithms, sorting algorithms and re-dispersion algorithms; when a recommendation request initiated by a target user through a corresponding client is detected, at least one target user intention label data matching the target user is obtained from each 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 having an intention label as a recall strategy label is screened out from each target user intention label data through the recommendation engine, and the content labels in each target user intention label data having an intention label as a recall strategy label are formed into a first content label set, and the Jaccard similarity coefficient is used to compare the first content label set with the first video item set. An association calculation is performed on the first video tag set of each first video entry to filter out a second video entry set from the first video entry set; 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, and the latent vector of the target user feedback text is calculated, and a third video entry set is filtered out from each video entry according to the latent vector of the target user feedback text; a fourth video entry set is filtered out from each of the video entries according to each preset recall algorithm, and the second video entry set, the third video entry set and the fourth video entry set are merged to obtain a fifth video entry set.

[0167] The filtering algorithm is called to filter the fifth video entry by forming a third content tag set or a pre-set filtering rule based on the content tags in the target user intention tag data whose intention tag is a filtering policy tag. The sorting algorithm is called to sort each fifth video entry in the filtered fifth video entry set according to the second video tag set and the second content tag set of each fifth video entry in the filtered fifth video entry set to obtain a sixth video entry set.

[0168] In the case that target user intention label data whose intention label is a re-dispersion strategy label exist in each target user intention label data, the re-dispersion algorithm can also be called to combine the content labels in the target user intention label data whose intention label is a re-dispersion strategy label in each target user intention label data into a fourth content label set; the sixth video item set is reordered using the fourth content label set, and finally a corresponding video recommendation list is generated according to the reordered sixth video item set, and the corresponding video is recommended to the target user according to the video recommendation list.

[0169] When the user feedback module recommends corresponding videos to the target user according to the video recommendation list, the corresponding video push reason can be popped up at a fixed time, that is, a preset prompt text is displayed next to or above the recommended video to prompt the target user why these videos are recommended and ask the user whether to recommend the strategy.

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

[0171] An embodiment of the present application provides a storage medium, which includes a stored program, wherein when the program is running, the device where the storage medium is located is controlled to execute the video recommendation method based on the large language model shown in the above embodiment.

[0172] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system or system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The system and system embodiments described above are merely schematic, in which the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0173] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0174] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

[0175] The above are only preferred implementations of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A video recommendation method 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 intention label data matching the target user is obtained from each user intention label data stored 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 text feedback data fed back by the user; The content tags in the target user intention tag data whose intention tags are recall strategy tags in each of the target user intention tag data 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; Acquire a first video tag set of each of the first video entries, and filter out a second video entry set from each of the first video entries according to the first content tag set and the first video tag set of each of the first video entries; Acquire target user text feedback data corresponding to the target user, and filter out a third video item set from all video items on the streaming media platform according to the target user text feedback data; Filtering a fourth video item set from each of the video items according to each preset recall algorithm, and fusing the second video item set, the third video item set and the fourth video item set to obtain a fifth video item set; The content tags in the target user intention tag data having the intention tag as the sorting strategy tag in each of the target user intention tag data are combined into a second content tag set; 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; 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 according to claim 1, characterized in that The generating the user intention label data according to the user text feedback data fed back by the user includes: Acquire user text feedback data of user feedback in real time, wherein the user text feedback data includes multiple user feedback texts; Generate corresponding prompt words according to each of the user feedback texts and a preset prompt word template, and input the prompt words into a large language model to obtain user intention label data corresponding to each of the user feedback texts; wherein the user intention label data includes multiple content labels and multiple intention labels; For each of the user intention label data, verify the validity of 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 a pre-built key-value database in the form of a key-value pair.

3. The method according to claim 1, characterized in that The obtaining of the first video tag set of each first video entry, and filtering out a second video entry set from each first video entry according to the first content tag set and the first video tag set of each first video entry, includes: Acquire a first video tag set for each of the first video entries from a preset video tag database; Calculating the similarity between the first video tag set and the first content tag set of each first video entry using the Jaccard similarity coefficient; A second video entry set is screened out from each of the first video entries according to the similarity between the first video tag set of each of the first video entries and the first content tag set.

4. The method according to claim 3, characterized in that The step of obtaining target user text feedback data corresponding to the target user, and filtering out a third video item set from all video items on the streaming media platform according to the target user text feedback data, includes: Acquire target user text feedback data of the target user, and calculate the latent vector of each target user feedback text in the target user text feedback data of the target user; Obtaining a latent vector of each of the video entries, and calculating the cosine similarity between the latent vector of each of the video entries and the latent vector of each of the target user feedback texts; A third set of video items is screened out from all video items on the streaming platform according to the cosine similarity between the latent vector of each of the video items and the latent vector of the target user feedback text.

5. The method according to claim 1, characterized in that The step of 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 includes: Obtaining a second video tag set for each fifth video item in the fifth video item set; For each of the fifth video entries, calculating a cross feature between 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 features between each second video tag of the fifth video entry and the second content tag set to obtain a cross feature between the fifth video entry and the second content tag; The score of each of the fifth video entries is calculated according to the cross features of each of the fifth video entries and the second content tag set by a preset fine ranking model, and the fifth video entries are sorted according to their scores to obtain a sixth video entry set.

6. The method according to claim 1, characterized in that Before sorting the fifth video items in the fifth video item set based on the second content tag set to obtain the sixth video item set, the method further includes: The content tags in the target user intention tag data having the intention tag as the filtering policy tag in each of the target user intention tag data are combined into a third content tag set; Filtering the fifth video item set using the third content tag set; Correspondingly, the fifth video items in the fifth video item set are sorted based on the second content tag set to obtain a sixth video item set, including: The fifth video items in the filtered fifth video item set are sorted based on the second content tag set to obtain a sixth video item set.

7. The method according to claim 1, characterized in that Before generating a corresponding video recommendation list according to the sixth video item set, the method further includes: Determine whether there is target user intention label data whose intention label is a re-dispersion strategy label among the target user intention label data; If so, the content tags in the target user intention tag data having the intention tag as the re-dispersion strategy tag in each of the target user intention tag data are combined into a fourth content tag set; Reordering the sixth set of video items using the fourth set of content tags; Correspondingly, generating a corresponding video recommendation list according to the sixth video item set includes: A corresponding video recommendation list is generated according to the reordered sixth video item set, and corresponding videos are recommended to the target user according to the video recommendation list.

8. A video recommendation system based on a large language model, characterized in that: The system includes an intention execution module; The intention execution module is used to, when a recommendation request initiated by a target user is detected, obtain at least one target user intention label data matching the target user from each user intention label data stored 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 the target user text feedback data fed back by the user; the content labels in the target user intention label data whose intention labels are recall strategy labels in each of the target user intention label data are formed into a first content label set, and a first video item set is filtered out from all video items on the streaming media platform according to the first content label set; wherein the first video item set includes multiple first video items; a first video label set is obtained for each of the first video items, and according to the first content label set and the first video label set of each of the first video items, A second video item set is screened out from each of the first video items; target user text feedback data corresponding to the target user is obtained, and a third video item set is screened out from each of the video items according to the target user text feedback data; a fourth video item set is screened out from each of the video items according to each preset recall algorithm, and the second video item set, the third video item set and the fourth video item set are merged to obtain a fifth video item set; content tags in the target user intention tag data whose intention tags are sorting strategy tags in each of the target user intention tag data are formed into a second content tag set; each fifth video item in the fifth video item set is sorted based on the second content tag 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.

9. An electronic device, characterized in that: include: A processor and a memory, wherein the processor and the memory are connected via a communication bus; wherein the processor is used to call and execute a program stored in the memory; The memory is used to store a program, and the program is used to implement the video recommendation method based on a large language model as described in 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 used to execute the video recommendation method based on a large language model as described in any one of claims 1-7.

Images