Content recommendation method, apparatus, device, storage medium, and product

Abstract

The embodiment of the application provides a content recommendation method, device, equipment, storage medium and product, including: obtaining demand expression information, historical browsing content and candidate recommendation content input by a target object, and extracting intention information in the demand expression information. The intention information is matched with the historical browsing content, and based on the matching result between the intention information and specific entities in the historical browsing content, a correlation score between the intention information and the candidate recommendation content is generated. Based on the correlation score and the candidate recommendation content, the target object is subjected to content recommendation processing. The technical scheme of the embodiment of the application can more accurately understand the real intention of the target object by matching the intention information with the historical browsing content. Further, the correlation of the content is discriminated based on the real intention of the target object, which is beneficial to improving the accuracy and effectiveness of the correlation discrimination, thereby improving the accuracy of personalized content recommendation.

Description

Technical Field

[0001] This application relates to the field of computer technology, specifically to a content recommendation method, apparatus, device, storage medium, and product. Background Technology

[0002] With the widespread adoption of smart devices and the rapid development of internet technology, various types of content have experienced explosive growth on platforms. Against this backdrop, personalized content recommendation services have emerged, which analyze users' individual preferences to accurately recommend content that meets their needs, thereby improving the user's browsing experience on the platform.

[0003] Currently, personalized content recommendation services primarily rely on keyword matching mechanisms based on user input. Platforms extract keywords from user input and compare them one-to-one with various tags in the content library to recommend content corresponding to the tag with the highest match score. However, in practice, user input may contain inaccuracies or vague descriptions. This approach may fail to accurately capture the user's true intent, leading to recommendations that deviate from the user's actual needs, resulting in low accuracy and negatively impacting the user experience on the platform.

[0004] Therefore, how to accurately recommend personalized content to users has become a pressing technical problem that needs to be solved. Summary of the Invention

[0005] This application provides a content recommendation method, apparatus, device, storage medium, and product. By matching intent information with historical browsing content, the true intent of the target object can be understood more accurately, and relevance judgment can be made based on the true intent, which helps to improve the accuracy of relevance judgment and, to a certain extent, improves the accuracy of content recommendation.

[0006] In a first aspect, embodiments of this application provide a content recommendation method, the method comprising:

[0007] Obtain the demand expression information input by the target object, the target object's historical browsing content, and the candidate recommendation content for the target object, and extract the intent information from the demand expression information;

[0008] The intent information is matched with the historical browsing content, and a relevance score between the intent information and the candidate recommended content is generated based on the matching results between the intent information and specific entities in the historical browsing content.

[0009] Based on the relevance score and the candidate recommended content, content recommendation processing is performed on the target object.

[0010] Secondly, embodiments of this application provide a content recommendation device, the device comprising:

[0011] The acquisition unit is used to acquire the demand expression information input by the target object, the historical browsing content of the target object, and the candidate recommendation content for the target object, and extract the intent information from the demand expression information;

[0012] The generation unit is used to match the intent information with the historical browsing content, and generate a relevance score between the intent information and the candidate recommended content based on the matching result between the intent information and a specific entity in the historical browsing content.

[0013] The processing unit is used to perform content recommendation processing on the target object based on the relevance score and the candidate recommended content.

[0014] In one possible implementation, the generation unit is configured to match the intent information with the historical browsing content, and based on the matching result between the intent information and specific entities in the historical browsing content, generate a relevance score between the intent information and the candidate recommended content, specifically for:

[0015] Obtain historical description information of the historical browsing content, and candidate description information of the candidate recommended content;

[0016] Extract the multimedia information of the historical browsing content to obtain historical multimedia information, and extract the multimedia information of the candidate recommended content to obtain candidate multimedia information;

[0017] The intent information, historical attribute information, historical multimedia information, candidate attribute information, and candidate multimedia information are input into the scoring generation model. The scoring generation model matches the intent information and the historical browsing content, and generates the relevance score based on the matching results.

[0018] In one possible implementation, the rating generation model includes a scheduling module and at least two processing modules; the generation unit is used to match the intent information and the historical browsing content through the rating generation model, and generate the relevance rating based on the matching result, specifically for:

[0019] Extract the intent information, the historical attribute information, the historical multimedia information, the candidate attribute information, and the feature information of the candidate multimedia information;

[0020] The scheduling module determines the target processing module that needs to be activated from the at least two processing modules based on the feature information.

[0021] The target processing module matches the intent information with the historical browsing content and generates the relevance score based on the matching results.

[0022] In one possible implementation, the acquisition unit is further configured to acquire training samples and rating labels corresponding to the training samples. The training samples include training intent information, training history browsing content, training candidate recommendation content, and thought chain information associated with the training samples. The thought chain information is used to indicate the reasoning process of determining the correlation score between the training intent information and the training candidate recommendation content as the rating label based on the training history browsing content.

[0023] The training unit is used to train the initial rating generation model based on the training samples and the rating labels to obtain the rating generation model.

[0024] In one possible implementation, the acquisition unit is used to acquire the thought chain information associated with the training sample, specifically for:

[0025] The training intent information, the training history browsing content, the training candidate recommendation content, and the rating label are input into the text generation model. The text generation model determines the inference step information of the correlation rating between the training intent information and the training candidate recommendation content based on the training history browsing content according to the set inference rules.

[0026] Obtain the reasoning step information generated by the text generation model, and determine the reasoning step information as the thought chain information.

[0027] In one possible implementation, the inference step information further includes an inference relevance score between the training intent information generated by the text generation model based on the set inference rules and the training candidate recommendation content; the acquisition unit is used to determine the inference step information as the thought chain information, specifically for:

[0028] If the reasoning relevance score is the same as the score indicated by the score label, then the reasoning step information is identified as the thought chain information.

[0029] In one possible implementation, the generation unit is further configured to input the training samples and the rating labels into the rating generation model if the number of model parameters of the rating generation model is greater than a set threshold, and generate a specified training sample and a specified rating label corresponding to the specified training sample by using the training samples and the rating labels as training sample cases through the rating generation model.

[0030] A determining unit is configured to determine the training samples and the specified training samples as distillation training samples, and to determine the scoring labels and the specified scoring labels as distillation scoring labels;

[0031] The training unit is used to train the initial first student rating model based on the distillation training samples and the distillation rating labels to obtain the first student rating model; wherein the number of model parameters of the initial first student rating model is less than that of the rating generation model.

[0032] In one possible implementation, the training unit is used to train an initial first student rating model based on the distillation training samples and the distillation rating labels to obtain a first student rating model, specifically for:

[0033] The distilled training samples are input into the initial first student rating model, and the training relevance score generated by the initial first student rating model is obtained;

[0034] Based on the difference between the training relevance score and the distillation score label, a first loss data is constructed, and the model parameters of the initial first student scoring model are adjusted based on the first loss data to obtain the first student scoring model.

[0035] In one possible implementation, the training unit is configured to adjust the model parameters of the initial first student rating model based on the first loss data to obtain the first student rating model, specifically for:

[0036] The distillation training samples are input into the rating generation model, and the reference relevance score generated by the rating generation model is obtained;

[0037] Based on the difference between the training relevance score and the reference relevance score, a second loss data is constructed;

[0038] Based on the sum of the first loss data and the second loss data, the target loss data is determined;

[0039] The model parameters of the initial first student rating model are adjusted based on the target loss data to obtain the first student rating model.

[0040] In one possible implementation, the acquisition unit is further configured to acquire a test sample and a test score corresponding to the test sample;

[0041] The processing unit is further configured to perform test processing on the first student scoring model based on the test sample and the test score;

[0042] The generation unit is used to input the intent information, the historical attribute information, the historical multimedia information, the candidate attribute information, and the candidate multimedia information into the scoring generation model, specifically for:

[0043] If the first student scoring model is determined to have passed the test based on the test sample and the test score, and the number of model parameters of the first student scoring model is less than the set threshold, then the intent information, the historical attribute information, the historical multimedia information, the candidate attribute information, and the candidate multimedia information are input into the first student scoring model.

[0044] In one possible implementation, the training unit is further configured to train an initial second student rating model based on the distillation training samples and the distillation rating labels to obtain a second student rating model if the number of model parameters of the first student rating model is greater than the set threshold; wherein the number of model parameters of the initial second student rating model is less than that of the first student rating model.

[0045] The generation unit is used to input the intent information, the historical attribute information, the historical multimedia information, the candidate attribute information, and the candidate multimedia information into the scoring generation model, specifically for:

[0046] If the model parameters of the second student scoring model are less than the set parameter threshold, then the intent information, the historical attribute information, the historical multimedia information, the candidate attribute information, and the candidate multimedia information are input into the second student scoring model.

[0047] Thirdly, embodiments of this application provide an electronic device, which includes one or more processors and a memory for storing one or more computer programs. When the one or more computer programs are executed by the one or more processors, the electronic device enables the content recommendation method described in the first aspect above.

[0048] Fourthly, embodiments of this application provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the content recommendation method described in the first aspect.

[0049] Fifthly, embodiments of this application provide a computer program product, which includes a computer program or computer instructions, and when the computer program or computer instructions are executed by a processor, they implement the content recommendation method as described in the first aspect.

[0050] In some embodiments of this application, the technical solutions involve acquiring the target object's input demand expression information, the target object's historical browsing content, and candidate recommended content for the target object, and extracting intent information from the demand expression information. Then, the intent information can be matched with the historical browsing content, and a relevance score between the intent information and the candidate recommended content is generated based on the matching result between the intent information and specific entities in the historical browsing content. Subsequently, content recommendation processing can be performed on the target object based on the relevance score and the candidate recommended content. Therefore, matching the extracted intent information with historical browsing content to more accurately understand the target object's true intent is suitable for complex situations where the target object's input demand expression information is unclear or has multiple possible interpretations. Furthermore, generating a relevance score based on the matching result between the intent understanding and specific entities allows for a clearer and more detailed understanding of the target object's true intent based on specific entities in the historical browsing content. Relevance judgment based on this true intent improves the accuracy and effectiveness of relevance judgment, thereby enhancing the accuracy of personalized content recommendations and improving the user's browsing experience on the platform. Attached Figure Description

[0051] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a schematic diagram of the architecture of a content recommendation system provided in an embodiment of this application;

[0053] Figure 2 This is a flowchart illustrating a content recommendation method provided in an embodiment of this application;

[0054] Figure 3 This is a flowchart illustrating another content recommendation method provided in an embodiment of this application;

[0055] Figure 4 This is a schematic diagram of a personalized relevance assessment process provided in an embodiment of this application;

[0056] Figure 5 This is a timing diagram of knowledge distillation provided in an embodiment of this application;

[0057] Figure 6 This is a schematic diagram of the technical framework for a personalized correlation discrimination task provided in an embodiment of this application;

[0058] Figure 7This is a schematic diagram of the structure of a content recommendation device provided in an embodiment of this application;

[0059] Figure 8 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0060] It should be noted in advance that, in order to enable those skilled in the art to better understand the technical solutions proposed in the embodiments of this application, the embodiments of this application will be described clearly and completely in conjunction with one or more accompanying drawings. Furthermore, the various drawings shown in the embodiments of this application are merely illustrative examples; for example, the execution order of each step in the drawings can be adaptively adjusted according to the actual application scenario. In addition, in the embodiments of this application, the block diagrams shown in the various drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, or in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0061] In this application embodiment, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0062] It should be noted that "multiple" in this article refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0063] With the rapid development of internet technology and the increasing demand for user information, various types of content have experienced explosive growth on platforms. Against this backdrop, personalized content recommendation services have emerged. These services analyze users' individual preferences to accurately recommend content that matches their needs, thereby improving the user's browsing experience on the platform. Specifically, personalized content recommendation services can be implemented in two scenarios: personalized content search and personalized content recommendation. Among them:

[0064] Personalized content search refers to a platform's ability to filter and sort massive amounts of information based on a user's individual needs and preferences when the user initiates a search, in order to provide highly relevant search results. Personalized content search has become a core research direction in the field of information retrieval, and it is crucial for improving user experience.

[0065] Personalized content recommendation refers to the platform proactively recommending content that users may be interested in based on user behavior, such as user-entered comments. It does not require users to perform explicit search operations, but automatically predicts the content that users may like and presents it to them. The accuracy of the recommendation is also crucial for optimizing the user experience.

[0066] Currently, the platform primarily relies on two methods to provide personalized content recommendation services to users:

[0067] The first method is keyword matching based on user input. The platform can extract keywords from the user's input and compare them one by one with each tag in the content library. Then, it can recommend the content corresponding to the tag with the highest matching degree to the user, such as ranking it at the top of the search results.

[0068] The second approach is based on fundamental user behavior analysis. The platform can collect user interaction data, such as clicks, browsing time, and purchase history. By analyzing this data, a personalized recommendation model can be trained to predict users' future interests and make content recommendations.

[0069] However, both of these approaches have significant limitations in understanding and predicting users' true needs, especially when faced with complex situations where user input is ambiguous and has multiple possible interpretations. For example, if a user inputs "Wu Zetian" and wants to browse content related to a game called "Wu Zetian," keyword matching based on the input might match content related to historical figures, failing to accurately capture the user's true intent. This makes it difficult to assess the relevance between recommended content and user needs, leading to recommendations that deviate from the user's actual requirements. On the other hand, if personalized recommendation models process user input, their lack of in-depth world knowledge and complex logical reasoning capabilities may prevent them from understanding the true intent indicated by "Wu Zetian," resulting in inaccurate assessments of the relevance between recommended content and user needs. This leads to low accuracy in content recommendations and negatively impacts the user experience on the platform.

[0070] Based on this, embodiments of this application provide a content recommendation scheme that can find specific entities in historical browsing content that match the intent information extracted from user input content. This allows for a more accurate understanding of the user's query intent and helps determine the user's true intent. Furthermore, a relevance score between the user's true intent and candidate content can be generated based on the matching results. Compared to directly judging relevance between user input content and candidate content, this improves the accuracy and effectiveness of relevance judgment, thereby enhancing the accuracy of content recommendation and, to some extent, increasing user satisfaction and improving the user experience on the platform.

[0071] To better understand the solutions of the embodiments of this application, the relevant terms and concepts that may be involved in the embodiments of this application will be introduced below.

[0072] 1. Large Language Model (LLM): LLM refers to a language model with an extremely large number of parameters, typically ranging from billions to hundreds of billions. It is a language model trained with a large number of parameters and a wide corpus. Its large number of parameters enables it to understand and generate natural language, allowing it to capture extremely complex semantic, grammatical, and pragmatic information in language. It can be regarded as an artificial intelligence (AI) system that can efficiently perform various complex natural language processing (NLP) tasks.

[0073] 2. Vision-Language Model (VLM): VLM is a model capable of processing and understanding multimodal data. Modality refers to data in different forms or from different information sources, such as text, images, and videos. Multimodality refers to data composed of two or more modalities. VLM can process and understand various types of data, such as text and images, enabling cross-modal understanding and generation, and is widely used in fields such as visual question answering.

[0074] 3. Supervised Fine-Tuning (SFT): SFT refers to the process of further training a pre-trained model using labeled data specific to a particular task. By introducing domain-specific knowledge or task-related nuances, it enhances the model's performance on that specific task. SFT not only improves model performance but also ensures that its output better meets the needs of real-world applications. Compared to training a completely new model, SFT requires less computational resources and time, reducing computational consumption and allowing for flexible adaptation to different tasks.

[0075] 4. Chain of Thought: Chain of Thought is a reasoning method applicable to LLM. By breaking down complex reasoning tasks, it yields a series of logically coherent and progressively advancing thought processes. It can help models solve complex problems more effectively and generate accurate answers. It can also clearly demonstrate its thought process, helping to improve the transparency and interpretability of the model, and to a certain extent, it can also enhance the model's reasoning ability.

[0076] 5. Knowledge Distillation: Knowledge distillation, or simply distillation, is a model compression technique that extracts knowledge from large or complex models and transfers the extracted knowledge to smaller or simpler models. Smaller or simpler models can maintain high performance while significantly reducing computing resources and inference time, making them more suitable for deployment in resource-constrained real-world applications, thereby achieving the goal of similar performance but lower computing costs.

[0077] (1) Teacher Model: The teacher model, also known as the teacher training model, refers to a well-trained, original, large-scale model, which usually has many parameters and a complex architecture. Although it has high accuracy, it may be difficult to deploy in practical applications due to resource constraints.

[0078] (2) Student model: A student model is a small model with a relatively simple structure and few parameters. Its goal is to imitate the behavior of the teacher model and get as close to its performance as possible, so that it can be deployed in resource-constrained environments.

[0079] 6. Query: A query is a request issued by a user or system to retrieve specific information or results from a large dataset. Queries can be in the form of text, images, audio, etc. For example, in personalized content search scenarios, a query can be the search text information carried in a user's submitted search request.

[0080] 7. Document (doc): In the personalized content service of this application embodiment, a document can be understood as an independent unit of information, such as a webpage article, a video, or any form of content block. For example, in a personalized content search scenario, a document can refer to each piece of content in the result list returned by the search engine.

[0081] Based on the above description, please refer to Figure 1 , Figure 1 This is a schematic diagram of the architecture of a content recommendation system provided in an embodiment of this application, such as... Figure 1As shown, the content recommendation system includes a user device 101, a content recommendation device 102, and a model training device 103. The user device 101 and the content recommendation device 102 can be connected directly or indirectly via wired or wireless means. The content recommendation device 102 and the model training device 103 can be connected directly or indirectly via wired or wireless means. It should be noted that... Figure 1 The number and form of the devices shown are for illustrative purposes only and do not constitute a limitation on the embodiments of this application. In some embodiments, there may be multiple user devices 101, content recommendation devices 102, and model training devices 103. In some embodiments, the content recommendation device 102 and the model training device 103 may be the same electronic device.

[0082] The user device 101 can be a device used by a target object (such as a user), and is an electronic device that can be used to install and run a platform or software with content publishing functions. Users can publish content on the platform through the user device 101, and can also browse content published by other users through the user device 101. Specifically, the user device 101 may include input devices, such as a touch screen, keyboard, microphone, etc., allowing users to input information expressing their needs on the platform, such as search information, or interactive information regarding other content, such as comments. The user device 101 may also include output devices, such as a display screen, speaker, etc., which can be used to output content on the platform, such as content recommended by the platform based on the user's input.

[0083] The content recommendation device 102 can be an electronic device that provides content publishing functions, such as a platform or software. It may be equipped with a machine learning model to determine the relevance between a user's true intent and candidate recommended content. It can also establish a communication connection with other electronic devices equipped with the same machine learning model to obtain relevance determination results. Based on the relevance determination results and candidate recommended content for the target object, the content recommendation device 102 can perform content recommendation processing for the target object. For example, it can recommend content with higher relevance in the relevance determination results to the user, or optimize its ranking order to place it higher in the search results.

[0084] The model training device 103 can be an electronic device used to train a machine learning model capable of recognizing a user's true intent and determining the correlation between that intent and candidate recommended content. The trained machine learning model can then be deployed in the content recommendation device 102 to provide personalized content recommendation services. With the rapid development of multimodal large-scale model technology, fine-tuning multimodal large-scale models for specific business scenarios has become a popular research direction. The machine learning model trained by the model training device 103 can be a multimodal large-scale model, possessing good capabilities in understanding complex personalized scenarios, recognizing personalized needs, and logical reasoning. The model training device 103 can fine-tune these large models in personalized content recommendation scenarios to achieve the ability to recognize a user's true intent and determine the correlation between that intent and candidate recommended content.

[0085] The content recommendation method provided in this application can be executed by the content recommendation device 102. The user device 101 can be a terminal device, and the dialogue information response device 102 and the model training device 103 can be terminal devices or servers. Terminal devices include, but are not limited to, smartphones (such as Android phones, iOS phones, etc.), tablets, portable personal computers, mobile internet devices (MIDs), smart voice interaction devices, smart home appliances, vehicle terminals, aircraft, wearable devices, etc., and this application embodiment does not limit this. The server can be an independent physical server, a server cluster composed of multiple physical servers, or a distributed system. It can also be a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms, and this application embodiment does not limit this.

[0086] The general process of the content recommendation method provided in this application is as follows:

[0087] Content recommendation device 102 can obtain target objects, such as Figure 1 The user inputs a request expression through user device 101, along with the target object's historical browsing content and candidate recommended content, and extracts intent information from the request expression. Content recommendation device 102 then matches the intent information with the historical browsing content and generates a relevance score between the intent information and the candidate recommended content based on the matching results between the intent information and specific entities in the historical browsing content. Subsequently, content recommendation device 102 performs content recommendation processing for the target object based on the relevance score and the candidate recommended content.

[0088] The content recommendation device 102 can match intent information and historical browsing content through a rating generation model, and generate a relevance score between intent information and candidate recommended content based on the matching results. The rating generation model can be trained by the model training device 103 based on training samples including training intent information, training historical browsing content, training candidate recommended content and thought chain information, as well as their corresponding rating labels to train the initial rating generation model.

[0089] In some embodiments, if the number of model parameters of the rating generation model is greater than a set threshold, the model training device 103 can perform knowledge distillation. Specifically, the rating generation model can be used as a teacher model to generate training samples and corresponding rating labels in batches, and the generated training samples and rating labels can be used to train student models with a smaller number of model parameters.

[0090] In some embodiments, the model training device 103 can perform knowledge distillation incrementally until the number of model parameters in the trained model is less than a set threshold. In this case, the model training device 103 can send it as a model to be deployed in an online environment to the content recommendation device 102, so that the content recommendation device 102 can deploy it to provide online services.

[0091] Therefore, based on Figure 1 The system shown can, on the one hand, improve the accuracy of relevance judgment based on the user's true intent, thereby enhancing the accuracy of content recommendation. On the other hand, knowledge distillation can leverage the capabilities of large models while conserving computational resources and improving training efficiency, thus providing users with a higher-quality personalized content recommendation experience.

[0092] In one implementation, the aforementioned demand expression information, browsing history, candidate recommendation content, and intent information can all be stored in the blockchain, preventing this information from being tampered with. Blockchain is a novel application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms. Essentially, it is a decentralized database, a chain of data blocks linked using cryptographic methods. Each data block contains information about a batch of network transactions, used to verify the validity of the information (anti-counterfeiting) and generate the next block.

[0093] It is understood that the embodiments described in this application are as follows: Figure 1The content recommendation system shown is intended to more clearly illustrate the technical solutions of the embodiments of this application and does not constitute a limitation on the technical solutions provided in the embodiments of this application. As those skilled in the art will know, with the evolution of system architecture and the emergence of new business scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0094] Based on the above content recommendation system, this application provides a content recommendation method. The content recommendation method described in this application can be executed by an electronic device, which can be... Figure 1 The content recommendation device 102 in the content recommendation system shown. Please refer to [link / reference]. Figure 2 , Figure 2 This is a flowchart illustrating a content recommendation method provided in an embodiment of this application. The content recommendation method includes the following steps S201-S203:

[0095] S201. Obtain the demand expression information input by the target object, the historical browsing content of the target object, and the candidate recommended content for the target object, and extract the intent information from the demand expression information.

[0096] In this application embodiment, the target object can refer to a user, specifically a user (user) on a platform with content publishing functionality. Users can publish content on the platform and also browse content within the platform. Request expression information can refer to the search information carried by the user when performing a search operation on the platform, expressing the user's desire to further browse content; this can be understood as a query. Request expression information can also refer to the input content carried by the user during interactive operations on the platform regarding previously viewed content. Interactive operations can be, for example, comment operations, and the input content can be comment information entered by the user regarding viewed content, expressing that the user is interested in or concerned about the content and may subsequently want to browse content with the same theme or elements.

[0097] The target user's browsing history can refer to content viewed on the platform, which can be in various formats, such as text, images, videos, and audio. For example, the target user's browsing history might include recently watched videos. The candidate recommended content for the target user can refer to content the platform prepares to recommend to the user based on their needs and interests. For example, the platform might retrieve content based on a user's search query. Or, the platform might filter content from a vast amount of content based on the user's browsing history and interactions.

[0098] Specifically, the user can input demand expression information through the user device. The user device can send the demand expression information to the content recommendation device that provides the platform, and the user device can also send the account identifier of the user in the platform. Furthermore, the content recommendation device can obtain the user's historical browsing content from the database based on the account identifier. For example, the content recommendation device can obtain video files and the like that the user has historically browsed.

[0099] In some embodiments, in the personalized content recommendation scenario, the content recommendation device can obtain candidate recommended content for the target object from the database based on the account identifier. Specifically, for example, it can also obtain candidate video files, web page files of graphic and text content, etc.

[0100] In some embodiments, in the personalized content search scenario, the content recommendation device can perform retrieval processing on the massive content of the platform based on the demand expression information, and use the retrieved content as the candidate recommended content.

[0101] In order to more accurately understand the user's true intention, the content recommendation device can extract the intention information in the demand expression, which can be understood as the user's query. The intention information can be the information with specific meaning or referring to specific things in the demand expression information, such as words or phrases. The intention information can be understood as the entity information in the demand expression information. An entity refers to a specific object or concept mentioned in the demand expression information, which can be physically existent, such as a person (e.g., the nickname of a specific internet celebrity), a place, a thing, or can also be abstractly existent, such as a concept, an event, such as an internet buzzword. The entity information is the word or noun phrase that describes the entity, that is, the entity word. The process of extracting entity words from the demand expression information can be called Named Entity Recognition (NER). For example, the intention information extracted from "The mysterious bird is赋予了特乐且神圣的意义." can be "mysterious bird".

[0102] In some embodiments, the content recommendation device can use a rule-based method. For example, the preset vocabulary includes all known entity words, and it can match the demand expression information with the preset vocabulary to determine whether the demand expression information contains these entity words, so as to obtain the matched intention information.

[0103] In some embodiments, the content recommendation device can also use a machine learning method, extract the features of the demand expression information based on a language model such as the Bidirectional Encoder Representations from Transformers (BERT), and identify the entity words therein, so as to obtain the intention information.

[0104] In some embodiments, the content recommendation device may invoke relevant analysis tools that provide entity word extraction to extract entity words from the demand expression information in order to obtain intent information.

[0105] Furthermore, after extracting intent information, the content recommendation device can match it with historical browsing content to determine the user's true intent, and then make a relevance judgment on the candidate recommended content.

[0106] S202. Match the above intent information with the above historical browsing content, and generate a relevance score between the above intent information and the above candidate recommended content based on the matching result between the above intent information and the specific entity in the above historical browsing content.

[0107] In this embodiment, a specific entity can refer to an entity that has appeared in the historical browsing content, such as a specific person, an internet slang term, or an item. The relevance score is a score used to evaluate the relevance (association) between intent information and candidate recommended content. The content recommendation device can determine whether to recommend the candidate content to the user based on the relevance score, and can also determine the ranking position of the candidate recommended content in the search results based on the relevance score.

[0108] Matching intent information with browsing history involves searching the browsing history for the existence of the specific entity referred to by the intent information (entity word). This is because the user's input of their needs may be ambiguous and have multiple possible interpretations, thus the extracted intent information may also be ambiguous and have multiple meanings. For example, the intent information "apple" could refer to fruit, electronic products, or movies, etc. Therefore, matching intent information with browsing history can clarify the user's true intent.

[0109] Specifically, based on the matching, two matching results can be obtained: one is used to indicate that there is no specific entity in the historical browsing content that matches the intent information, and the other is used to indicate that there is a specific entity in the historical browsing content that matches the intent information, and to indicate the specific entity that matches.

[0110] In one implementation, if the matching result indicates that there is no specific entity in the historical browsing content that matches the intent information, the content recommendation device can generate a relevance score based on the matching result. This can be based on the meaning of the intent information itself, i.e., its general meaning, such as apples referring to a type of fruit. The relevance between the intent information and the candidate recommended content can be judged. This can be understood as the user's true intent being a general need. The relevance score between the intent information and the candidate recommended content is thus obtained.

[0111] In another implementation, if the matching result indicates that a specific entity in the historical browsing content matches the intent information, the content recommendation device can generate a relevance score based on the matching result by judging the relevance between the specific entity and the candidate recommended content, thereby obtaining the relevance score.

[0112] In one possible implementation, the process of a content recommendation device matching and generating a relevance score could involve first acquiring historical description information of previously viewed content, as well as candidate description information of candidate recommended content. The content recommendation device can also extract multimedia information from the historically viewed content to obtain historical multimedia information, and extract multimedia information from the candidate recommended content to obtain candidate multimedia information. Then, the content recommendation device can input intent information, historical attribute information, historical multimedia information, candidate attribute information, and candidate multimedia information into a score generation model. The score generation model then matches the intent information with the historically viewed content and generates a relevance score based on the matching results.

[0113] The descriptive information can be understood as information used to describe the attributes of the historically viewed content. For example, if the historically viewed content and the candidate recommended content are video content, the historical descriptive information and the candidate descriptive information can be textual feature information such as video title, category, tag, entity, and publishing account. The historical descriptive information and the candidate descriptive information can be obtained from the platform's content database or obtained through other means; this application does not limit this.

[0114] Multimedia information can refer to information in at least one modality within the content, such as image information and text information. For example, if the historical browsing content and candidate recommendation content are video content, the historical multimedia information and candidate multimedia information can include video frames and video text information. Video text information can refer to video information extracted from the video. For instance, Optical Character Recognition (OCR) can be performed on each video frame in the historical browsing content and candidate recommendation content to identify the text characters in each video frame. As another example, Automatic Speech Recognition (ASR) can be performed on the historical browsing content and candidate recommendation content to convert the audio content in the video into text characters.

[0115] In some embodiments, video frames can be all video frames in the content, or a portion of video frames, such as a set number of video frames, for example, four video frames. Video frames can be extracted in the order of video playback time, randomly, or at intervals; this application does not limit this.

[0116] Furthermore, content recommendation devices can combine historical descriptive information and historical multimedia information to summarize video content, and combine candidate descriptive information and candidate multimedia information to summarize video content. For example, semantic analysis based on machine learning models can be used to summarize video content. This process can be called content understanding, which can provide background information for subsequent personalized relevance determination for users.

[0117] It's important to note that this rating generation model can be a large-scale multimodal machine learning model, i.e., a multimodal model with a very large number of parameters. This is because while personalized relevance judgment can be performed based on a model with a smaller number of parameters combined with user profiles, for example, information such as the user's browsing history (title, content, etc.), the user's expressed needs, and candidate recommendations can be input into a BERT model trained based on business expectations. The BERT model can extract features from this information and encode them. Then, BERT can merge the encoded feature vector with the query information to form the final embedding vector. Afterwards, based on the relevance between the BERT model's output vector (embedded vector) and the vector encoded from the query information, a relevance score can be determined between the query information and the candidate recommendations. This relevance score can then be used to optimize the quality of search results and the user experience.

[0118] However, due to the limited number of parameters in the BERT model, its knowledge of the world is limited. This may result in limited ability to understand browsing history, candidate recommendations, and discover users' true intentions. It is difficult to deeply understand users' true query intentions and cannot perform complex logical reasoning, resulting in low effectiveness and accuracy in judgment.

[0119] Therefore, the rating generation model in this application embodiment can be, for example, a multimodal LLM or VLM. This is because these models have a large number of parameters, possess in-depth world knowledge, and have complex logical reasoning capabilities, especially when used to process non-textual data such as images and videos, where they perform well. Since they perform well in handling complex tasks, they can be used for matching to perform personalized relevance judgment based on the user's true intent.

[0120] Specifically, the rating generation model can extract feature information from intent information, historical attribute information, historical multimedia information, candidate attribute information, and candidate multimedia information, and match intent information and historical browsing content based on the extracted feature information to determine the user's true intent, and generate a relevance rating based on the matching result (i.e. the user's true intent).

[0121] In some embodiments, because multimodal large-scale models have slow inference speeds and consume a lot of computational resources, model pruning and model quantization can be used to optimize them in order to improve inference speed and reduce resource consumption. Model pruning refers to removing weights from the scoring generation model that have a smaller impact on the model output, thereby reducing the number of model parameters and computational complexity. Model quantization refers to converting high-precision values ​​(such as 32-bit floating-point numbers) in the scoring generation model to lower-precision values ​​(such as 8-bit integers) since low-precision data is processed faster than high-precision data. This not only reduces the model's storage requirements but also accelerates the computation process, thereby improving the model's inference speed.

[0122] However, both of these methods have some drawbacks. For model pruning, the pruning process selectively removes connections or neurons according to certain rules, which may lead to an irregular remaining network structure, making it difficult to map onto the hardware architecture and increasing the difficulty of hardware implementation. For model quantization, quantization reduces the number of bits required to represent each value, which may result in information loss. This loss of precision may affect model performance, especially in applications with extremely high accuracy requirements. Furthermore, model quantization involves complex calibration processes and specific hardware support.

[0123] Based on this, the rating generation model in this application embodiment can be a machine learning model including a scheduling module and at least two processing modules. In specific processing, the content push device can extract intent information, historical attribute information, historical multimedia information, candidate attribute information, and feature information of the candidate multimedia information through the rating generation model. The scheduling module then determines the target processing module to be activated from at least two processing modules based on the feature information. Furthermore, the target processing module matches the intent information with the historical browsing content and generates a relevance rating based on the matching results.

[0124] In one implementation, the scoring generation model can be a machine learning model with a dynamic network architecture, such as a dynamic neural network model. This model can automatically adjust its own structure or parameters based on the characteristics of the input data. Adjusting its own structure can be understood as activating some or all of the network layers for processing, and adjusting its own parameters can be understood as reducing or increasing the weights of the parameters of the relevant network layers to make their impact on the result smaller or larger.

[0125] Specifically, the rating generation model can include a network structure for extracting multimodal features, such as intent information, historical attribute information, historical multimedia information, candidate attribute information, and feature information of candidate multimedia information, like the vector representation information of each part. The scheduling module included in this rating generation model can refer to the gating unit of a dynamic neural network model. This gating unit can be used to determine which network layers (i.e., at least two processing modules) of the dynamic neural network model need to be activated based on the extracted feature information; this is the defect target processing module. The scheduling module (i.e., the gating unit) can also adjust the model parameters in the activated network layers (i.e., the target processing module) to allow processing by the target processing module, such as matching and generating relevance scores.

[0126] In another implementation, the rating generation model can be a Mixture of Experts (MOE) model, which can improve efficiency and performance by assigning tasks to different "expert" subnetworks. Each subnetwork can be used to process a specific type of input or task, and a gating mechanism determines which subnetworks should participate in the computation.

[0127] Specifically, the rating generation model can include a network structure for extracting multimodal features, such as intent information, historical attribute information, historical multimedia information, candidate attribute information, and candidate multimedia information. The scheduling module of this rating generation model can refer to the gating unit of a hybrid expert model, and each sub-"expert" network can refer to at least two processing modules. The scheduling module (gating unit) can calculate the activation weights of each sub-network based on the extracted feature information. These activation weights can be calculated based on the complexity or task type represented by the feature information. Therefore, based on the activation weights calculated according to factors such as complexity and task type, the "expert" sub-networks that need to be activated in the hybrid expert model can be selected, i.e., the target processing module can be determined. Then, processing can be performed by the target processing module, such as matching and generating relevance scores.

[0128] Therefore, both implementation methods can reduce unnecessary computation while ensuring the processing performance of the rating generation model, thereby reducing the consumption of computing resources and improving the inference speed of the model. They are suitable for personalized content recommendation systems with diverse and complex user needs and can flexibly cope with different types and levels of difficulty.

[0129] In another implementation, because large multimodal models consume a lot of resources and have a slow inference speed, they are limited in applications that provide real-time personalized content recommendation services. Therefore, their knowledge can be distilled into a model with a smaller number of parameters, such as the BERT model with 60 million parameters, to transfer the capabilities of the large model to the small model. This balances the resource consumption and the ability to handle complex tasks, thus overcoming the problems of high resource consumption and slow inference speed caused by large models.

[0130] Furthermore, deploying the rating generation model, or a smaller model derived from knowledge distillation based on the rating generation model, in an online environment can generate relevance scores. Subsequently, content recommendation devices can personalize content recommendations for users based on these relevance scores and candidate content.

[0131] S203. Based on the above relevance scores and the above candidate recommended content, perform content recommendation processing on the above target objects.

[0132] In this embodiment, the relevance score can be a specific score, such as a score between 1 and 5. For example, a score of 5 indicates very high relevance, with no deviation between the intent information and the candidate recommended content, while a score of 1 indicates that the candidate recommended content is seriously deviated from or missing from the intent information and is completely unrelated to it. The aforementioned candidate recommended content can be multiple items, and a corresponding relevance score can be generated for each candidate recommended content. Then, content recommendation processing can be performed on the target object based on the relevance scores of each candidate recommended content.

[0133] In one implementation, in a personalized content search scenario, multiple candidate recommendations can be multiple search results. The content recommendation device can sort the relevance scores of each search result from high to low, and pull a set number of search results according to this sorting for content recommendation processing. For example, the set number of search results can be displayed in the user interface in order of score from high to low.

[0134] In another implementation, in a personalized content recommendation scenario, multiple candidate recommended content can be content prepared to be recommended to the user for browsing. The content recommendation device can remove recommended content with a relevance score lower than a set score threshold, or it can recommend content to the user in descending order of relevance score. For example, if the user is continuously browsing video content, the user can browse each video content in descending order of relevance score.

[0135] Therefore, by comprehensively processing information from multiple media types such as text, images, audio, and video, this application embodiment can more comprehensively understand the true intent in the user's input expression of needs, and can also better understand historical browsing content and candidate recommended content. It can more accurately understand the relationship between the user's true intent, historical browsing content, and candidate recommended content, and achieve more standardized, effective, and accurate personalized relevance judgment. This can overcome the inaccuracy of the model in recognizing the user's true intent, improve the accuracy of relevance judgment, better meet user needs, and to a certain extent improve the user experience of searching and browsing on the platform.

[0136] In some embodiments of this application, the technical solutions involve acquiring the target object's input demand expression information, the target object's historical browsing content, and candidate recommended content for the target object, and extracting intent information from the demand expression information. Then, the intent information can be matched with the historical browsing content, and a relevance score between the intent information and the candidate recommended content is generated based on the matching result between the intent information and specific entities in the historical browsing content. Subsequently, content recommendation processing can be performed on the target object based on the relevance score and the candidate recommended content. Therefore, matching the extracted intent information with historical browsing content to more accurately understand the target object's true intent is suitable for complex situations where the target object's input demand expression information is unclear or has multiple possible interpretations. Furthermore, generating a relevance score based on the matching result between the intent understanding and specific entities allows for a clearer and more detailed understanding of the target object's true intent based on specific entities in the historical browsing content. Relevance judgment based on this true intent improves the accuracy and effectiveness of relevance judgment, thereby enhancing the accuracy of personalized content recommendations and improving the user's browsing experience on the platform.

[0137] This application also provides a content recommendation method. The content recommendation method described in this application can be executed by an electronic device, which can be... Figure 1 The content shown recommends device 102 or model training device 103. Please refer to [link / reference]. Figure 3 , Figure 3 This is a flowchart illustrating another content recommendation method provided in this application embodiment. The processing method of the image generation model includes the following steps S301-S302:

[0138] S301. Obtain training samples and corresponding rating labels for the training samples. The training samples include training intent information, training history browsing content, training candidate recommendation content, and thought chain information associated with the training samples. The thought chain information is used to indicate the reasoning process of determining the correlation score between the training intent information and the training candidate recommendation content as the rating label based on the training history browsing content.

[0139] In this embodiment, a training sample refers to the data used to train the rating generation model. A training sample can be one of the samples in the training dataset used to train the rating generation model. The rating label is the relevance score label corresponding to the training sample. Specifically, the training sample may include training intent information, training history browsing content, training candidate recommendation content, and thought chain information associated with the training sample.

[0140] The training intent information, similar to intent information, can refer to entity information with a specific meaning or referring to a specific matter, and can be extracted from the information expressed in the training requirements. Training history browsing content can refer to reference content used during the training process for subsequent matching and relevance scoring; this can be text / image content, video content, etc., on the platform. Training candidate recommendation content can refer to content used during the training process for subsequent determination of relevance with intent information; this can also be text / image content, video content, etc., on the platform. The thought chain information associated with the training sample can refer to text information containing a detailed reasoning process; that is, it describes the detailed reasoning process used to determine the relevance score between the training intent information and the training candidate recommendation content based on the training history browsing content, which is the score indicated by the score label. This can include the reasoning basis and process for each step.

[0141] In some embodiments, the training samples and corresponding rating labels can be obtained by the model training device. For example, the model training device can obtain them from a database storing the training dataset, such as a sample database. The training data in the sample database can be collected and stored after removing privacy information from a large amount of user-related data on the platform and then annotating the data. Alternatively, it can be manually constructed and annotated according to the required data format and then stored. This application does not limit this.

[0142] It should be noted that, in order to improve the inference performance of the trained rating generation model, the data format can be the same in the training and inference phases. For example, if the training history browsing content includes only one item in the training phase, the target object's history browsing content in the inference phase can also include only one item. Similarly, if the training candidate recommendation content includes only one item in the training phase, the candidate recommendation content for the target object in the inference phase can also include only one item.

[0143] In one possible implementation, the rating labels corresponding to the training samples can be annotated (generated) by annotators, generated using deep learning methods such as a rating label generation model, or annotated in other ways; this application does not limit this. To make the annotated rating labels more accurate and interpretable, embodiments of this application provide a personalized relevance evaluation rule and process, including specific evaluation rules, such as the degree of relevance corresponding to different scores, and an evaluation process, such as the steps that annotators and the label generation model can follow to generate rating labels. This personalized relevance evaluation rule and process may also include annotated evaluation cases to facilitate better annotation by annotators and the rating label generation model based on these cases.

[0144] Understandably, since the mind chain information depends on the score indicated by the generated rating tags, the corresponding rating tags can be generated first based on the training intent information, training history browsing content, and training candidate recommendation content. Then, the mind chain information can be generated based on the training intent information, training history browsing content, training candidate recommendation content, and rating tags.

[0145] In some embodiments, descriptive information of the training history browsing content and training candidate recommendation content can be obtained separately, and multimedia information of the training history browsing content and training candidate recommendation content can be extracted to generate more accurate rating tags. For ease of description, taking video content as an example, the model training device can obtain descriptive information such as the video title, category, tags, entities, and publishing account of the video content. The model training device can also extract text characters from the text and image content, such as performing OCR on each video frame and ASR on the audio content of the video to obtain and parse the main text fields in the video content. The model training device can also extract images from the video content, such as extracting a set number of video frames, thereby obtaining the descriptive information and multimedia information (i.e., Doc information) of the training history browsing content and training candidate recommendation content.

[0146] In other words, the model training device can convert the browsing content of the training history and the recommended content of the training candidates into structured text information, and download a set number of video images to obtain descriptive information and multimedia information, and use the descriptive information and multimedia information as the basis for data annotation.

[0147] Specifically, the personalized relevance assessment process used to generate rating labels can encompass the following six steps: training content understanding of historical browsing content, training intent information parsing and matching, determining true intent, rewriting training intent information based on true intent, training evidence assessment of whether candidate recommended content meets the requirements, and generating relevance judgment conclusions. Among these:

[0148] 1) Content understanding of training history browsing content: Based on the attribute information and multimedia information included in the training history browsing content, semantic analysis is performed on the history browsing content, and then the video content of the training history browsing video is summarized to provide background information for subsequent personalized judgment.

[0149] 2) Parsing and matching training intent information: Lexical analysis, syntactic analysis, and semantic understanding can be performed on the intent information to parse the training intent information. Combined with the training history browsing content, the intent information is matched with the training history browsing videos to find specific entities that match the intent information in the training history browsing videos, so as to further determine the specific needs corresponding to the intent information.

[0150] 3) Determining the True Intent: By combining the training history browsing content and the analysis of the training intent information, the true intent information can be determined based on the matching results between the browsing history videos and the intent information. If the matching result indicates that a specific entity in the browsing history content matches the intent information, the true intent indicated by the intent information can be determined based on the specific entity. If the matching result indicates that no specific entity in the browsing history content matches the intent information, the most general and common meaning of the intent information is determined as the true intent.

[0151] 4) Rewriting Training Intent Information Based on Real Intent: If the matching result indicates that a specific entity in the historical browsing content matches the intent information, the training intent information can be rewritten based on the entity information of that specific entity so that the rewritten training intent information can clearly indicate the real intent. If the matching result indicates that no specific entity in the historical browsing content matches the intent information, then there is no need to rewrite the training intent information.

[0152] 5) Evidence assessment of whether the training candidate recommended content meets the requirements: Evaluate the correlation between the training intention information (or the rewritten training intention information) of the training candidate recommended content, specifically by matching the correlation with the text content (such as descriptive information) and image content (such as extracted video frames) in the training candidate recommended content.

[0153] The following scoring criteria can be used in the personalized relevance assessment rules:

[0154] 5 points: The training candidate recommendations are a perfect match with the intent information / rewritten intent information, and there is no topic bias.

[0155] 4 points: The training candidate recommendations are basically matched with the intent information / rewritten intent information, but there are some topic mismatches or content mismatches, but there is no significant bias overall.

[0156] 3 points: Some content in the training candidate recommendations matches the intent information / rewritten intent information, but there is a significant topic deviation; or the training candidate recommendations can match some information in the intent information / rewritten intent information, and there is no significant topic deviation.

[0157] 2 points: The training candidate recommendations are only partially relevant, or only domain relevant, and almost do not match the intent information / rewritten intent information; or the training candidate recommendations match the general meaning corresponding to the intent information well, but have a significant deviation from the true intent.

[0158] 1 point: The training candidate recommendation content does not match the intent information / rewritten intent information at all, the core words are seriously deviated or lost, and the training candidate recommendation content is completely unrelated to the intent information / rewritten intent information.

[0159] 6) Generation of relevance discrimination conclusions: Based on the above personalized relevance evaluation process and evaluation rules, the final relevance discrimination conclusions can be generated to determine the matching level between the training candidate recommended videos and the real needs indicated by the training intent information.

[0160] Therefore, annotators (such as professionals) or rating label generation models can finely annotate training samples according to a predetermined personalized relevance assessment process, rules, and cases to generate high-quality personalized relevance rating labels. For example, the aforementioned personalized relevance assessment process, rules, and cases can be organized into specific prompts and input into the rating label generation model. Based on the content of the prompts, the model can generate corresponding rating labels for the input training samples. Furthermore, the model training device can store the rating labels in a database, such as the sample database mentioned above.

[0161] In some embodiments, in order to ensure the quality of the generated rating labels, quality control can also be carried out. Specifically, it can be checked by random sampling and cross-validation. For example, some training samples and corresponding rating labels can be randomly selected, and cross-validation and checking can be carried out by other personnel besides those who generated the rating labels. This allows the annotation results to be reviewed and corrected to ensure the accuracy and reliability of the data.

[0162] In some embodiments, multi-task learning (MTL) methods can be used to generate rating labels. That is, the rating label generation model described above can be a multi-task learning model, which can be called a personalized multi-task learning model. It can simultaneously learn to perform multiple related but not identical tasks. These tasks may have the same input but different outputs, or related inputs and outputs.

[0163] For example, this personalized multi-task learning model can simultaneously handle the parsing and matching of training intent information, the determination of true intent, and the evaluation of evidence that training candidate recommendations meet the requirements. Its core is that part or all of the underlying structure of the personalized multi-task learning model is shared. This part can be used to capture common features or patterns across all tasks. By sharing the underlying representation, the model can effectively transfer information and patterns learned from one task to other related tasks, thereby improving learning efficiency and enabling the model to better capture the commonalities and interrelationships between tasks, thus enhancing the model's understanding of the correlations between different tasks.

[0164] Based on this, after obtaining the rating labels, the model training device can generate thought chain information based on training samples including training intent information, training history browsing content, training candidate recommendation content, and rating labels. Since this thought chain information can be used to evaluate the demand satisfaction evidence of candidate recommendation content, it can also be understood as evidence chain information.

[0165] Specifically, the model training device can input training intent information, training history browsing content, training candidate recommendation content, and rating labels into the text generation model. The text generation model, according to the set reasoning rules, determines the reasoning step information, which is the correlation score between the training intent information and the training candidate recommendation content, based on the training history browsing content, and sets the score label. Then, the model training device can obtain the reasoning step information generated by the text generation model and determine the reasoning step information as the thought chain information.

[0166] The text generation model can refer to a multimodal large model, such as a multimodal LLM or VLM. The thought chain information is textual information that can be used to indicate the detailed reasoning process of determining the relevance score between training intent information and training candidate recommendation content based on the training history browsing content, given a rating label. This process can include the basis and process of each step. Specifically, training intent information, training history browsing content (descriptive information and multimedia information), training candidate recommendation content (descriptive information and multimedia information), and rating labels can be input into the text generation model. The prompt words indicate the set reasoning rules used to generate the thought chain information, thereby obtaining the reasoning step information generated by the text generation model, i.e., the thought chain information.

[0167] Table 1 is an example of prompt words for generating thought chain information provided in the embodiments of this application. Table 1 uses video content with training history browsing content and candidate recommendation content as an example for explanation.

[0168] Table 1

[0169]

[0170]

[0171]

[0172]

[0173] As shown in Table 1, the prompts input into the text generation model can be written according to the technical approach (reasoning steps) of the thought chain. Since the thought chain connects a series of related concepts, viewpoints, questions, or ideas to form an ordered chain, using thought chain information as training data for training the rating generation model can enhance the model's logical reasoning ability and ensure that the model understands the correlation between the user's true intent and historical browsing content and candidate recommended content. The input information shown at the end of Table 1 can include slots for various input text generation models, which can be filled based on the acquired information. It should be noted that the specific content in Table 1 is only used to assist in understanding the content recommendation method provided in this application and does not constitute an actual limitation of this application.

[0174] Furthermore, after obtaining the thought chain information, it can be validated with data. This involves verifying the rationality of each step in the thought chain, such as whether the summarized video content is correct, and whether the interpretation of the relevance of the rating tags is accurate. Subsequently, the model training device can store the validated thought chain information and training samples in a database, such as the aforementioned sample database.

[0175] In some embodiments, the inference step information generated by the text generation model may further include an inference relevance score generated by the text generation model based on a set of inference rules, that is, the above-mentioned relevance level, which can be used to indicate the relevance between the training intention information and the training candidate recommendation content. After the manual verification is passed, the model inference device may further verify based on the scoring label and the inference relevance score. If the inference relevance score is the same as the score indicated by the scoring label, it can be determined that the chain of thought information can be used for model training, and then it can be correspondingly stored in the database together with other data in the training sample and the scoring label.

[0176] Please refer to Figure 4 , Figure 4 FIG. is a schematic diagram of a personalized relevance evaluation process provided by an embodiment of the present application. The personalized relevance evaluation process may refer to the above-mentioned process of generating scoring labels, or may correspond to each inference step in the chain of thought information. Figure 4 Taking the training historical browsing content and the training candidate recommendation content both being video content as an example for explanation.

[0177] Figure 4 The two upper boxes respectively include the video information of the training historical browsing content and the video information of the training candidate recommendation content. The video information of the training historical browsing content includes descriptive information such as video title, category, label, entity, and publishing account, and may also include multimedia information, such as the main content extracted by ASR and OCR, and video frames extracted from the video content, such as the above 3 key frames.Similarly, the video information of the training candidate recommendation content also includes descriptive information and multimedia information. The demand expression information input by the user can be the text part between the two boxes. For example, "The black bird is given special and sacred meaning. Especially in the historical period before the Tang Dynasty, crows were often regarded as a symbol of good luck in folk traditional culture, indicating future hope", and the intention information extracted from the demand expression information can be "black bird".

[0178] Furthermore, the training historical browsing content and the training intention information can be matched to determine whether the entity indicated by the training intention information is a specific entity in the training historical browsing content. If the matching result indicates that there is a specific entity in the training historical browsing content that matches the training intention information, the training intention information can be rewritten. For example, "black bird" can be rewritten as "black bird crow". Conversely, if the matching result indicates that there is no specific entity in the training historical browsing content that matches the training intention information, the original training intention information, that is, the original query, can be maintained.

[0179] Furthermore, the relevance between the rewritten training intent information and the training candidate recommended content can be determined based on the rewritten training intent information. For example, the rewritten training intent information can be matched with the descriptive information and multimedia information in the training candidate recommended content to confirm whether the video content meets the user's real intent needs, thereby obtaining the relevance score label, i.e., the personalized relevance level.

[0180] S302. Based on the above training samples and the above rating labels, the initial rating generation model is trained to obtain the above rating generation model.

[0181] In this embodiment, the initial rating generation model can be a multimodal model with an extremely large number of parameters (e.g., a super-large model with 7 billion parameters). This is because a large number of parameters enables powerful multimodal information processing capabilities and the ability to identify the user's true intent, thus enabling the completion of the content recommendation task proposed in this embodiment. The initial rating generation model can be a pre-trained model selected from multiple models with large parameter numbers. The selection can be based on factors such as task type and ease of transfer learning. For example, a mixed-mode large model with good multimodal information processing, semantic understanding, and logical reasoning capabilities can be selected.

[0182] In some embodiments, the model training device may preprocess the training samples and corresponding rating labels before training. Specifically, this may include converting their format to the input format of the pre-trained model, such as converting them to JavaScript Object Notation (JSON) format. It may also include verifying whether the rating labels are consistent with the final judgment conclusion in the thought chain information, and downloading images, such as video frames, to ensure that all data is suitable for model training.

[0183] Furthermore, the model training device can train the initial rating generation model based on training samples and rating labels. Specifically, training samples can be input into the initial rating generation model, i.e., business data from the personalized content recommendation service can be injected to obtain the relevance scores generated by the initial rating generation model. Then, loss data is constructed based on the difference between the relevance scores and the rating labels, which is used to adjust the model parameters of the initial rating generation model. The model training device can then determine the initial rating generation model after adjusting the model parameters as the rating generation model. It can be understood that this process of training the pre-trained model with labeled (rating labels) training data can be called the SFT process, which implicitly trains the understanding and integration of information and the ability to achieve personalized relevance discrimination into the pre-trained model.

[0184] When adjusting the model parameters of the initial rating generation model, Low-Rank Adaptation (LoRA) can be used to fine-tune the model parameters, freezing most of the model parameters in the initial rating generation model, i.e., not updating these model parameters. Specifically, a low-rank matrix can be introduced to adjust the model weights instead of directly updating all parameters. Before training, the network layers that need to apply LoRA can be specified, and during training, only the model parameters of the specified networks are updated. This maintains the original semantic understanding and logical reasoning capabilities of the large model and injects business domain knowledge.

[0185] In some embodiments, after training is completed, a test set can be obtained, such as test samples and corresponding test score labels, to evaluate the trained score generation model. Professionals can also verify the model and continuously iterate the model to ensure high-precision performance of the model in personalized relevance discrimination tasks.

[0186] Deploying models with a large number of parameters to personalized content recommendation services consumes significant computational resources. Therefore, model training devices can preset a threshold for the number of model parameters, allowing deployment only when the number of parameters falls below this threshold. Thus, through knowledge distillation, the core capabilities of a fully trained, high-parameter model can be transferred to a model with fewer parameters, achieving a balance between performance and efficiency.

[0187] Specifically, the model training device compares the number of model parameters in the rating generation model with a set threshold. If the number of model parameters exceeds the threshold, the device inputs training samples and rating labels into the rating generation model. The model then uses these training samples and labels as training examples to generate specified training samples and corresponding rating labels. These training samples and specified training samples are then designated as distillation training samples, and the rating labels and specified rating labels are designated as distillation rating labels. Based on these distillation training samples and distillation rating labels, the initial first student rating model is trained to obtain the first student rating model.

[0188] Because this rating generation model possesses strong logical reasoning and semantic understanding capabilities, the training device can input training samples and rating labels into it. The model can then use these samples and labels as training examples to generate a batch of training samples and corresponding rating labels, resulting in a specified set of training samples and corresponding rating labels. For instance, the number of training samples and corresponding rating labels can be in the thousands, while the rating generation model can generate tens of thousands of specified training samples and corresponding rating labels. Therefore, all training samples can be used as distilled training samples, and all rating labels can be used as distilled rating labels to train an initial first-student rating model with fewer parameters, reducing the workload of data labeling and improving training efficiency. This first-student rating model can refer to a multimodal model with fewer parameters than the rating generation model.

[0189] Specifically, during training, the model training device can input distilled training samples into the initial first student rating model and obtain the training relevance score generated by the initial first student rating model. Then, based on the difference between the training relevance score and the distilled score label, a first loss data is constructed. The model parameters of the initial first student rating model are then adjusted based on the first loss data to obtain the final first student rating model. Here, the first loss data can be understood as the true loss data. The model parameters of the initial first student model can be adjusted based on the true loss data, thus determining the adjusted initial first student model as the final first student rating model.

[0190] In the knowledge distillation process, the rating generation model can be used as a teacher model and incorporated into the training of the initial first student model. Specifically, the model training device can input distilled training samples into the rating generation model and obtain the reference relevance score generated by the model. Then, based on the difference between the training relevance score and the reference relevance score, a second loss data is constructed. Furthermore, the model training device can determine the target loss data based on the sum of the first and second loss data, and adjust the model parameters of the initial first student rating model based on the target loss data to obtain the first student rating model. Here, the second loss data can be understood as distillation loss data. The model parameters of the initial first student model can be adjusted based on the sum of the true loss data and the distillation loss data, and the adjusted initial first student model is determined as the first student rating model.

[0191] Furthermore, after knowledge distillation, the model training device can test the first student model. Specifically, the model training device can acquire test samples and corresponding test scores, and then perform testing processing on the first student scoring model based on the test samples and test scores. If the first student scoring model passes the test based on the test samples and test scores, and the number of model parameters of the first student scoring model is less than a set threshold, the model training device can deploy the first student scoring model in an online environment for real-time relevance discrimination. Specifically, the intent information, historical attribute information, historical multimedia information, candidate attribute information, and candidate multimedia information obtained during the inference process can be input into the first student scoring model for relevance discrimination.

[0192] The test samples and their corresponding test scores can contain the same data types as the training samples and their corresponding score labels. Testing the first student scoring model can refer to inputting test samples into the first student scoring model, performing relevance discrimination through the model to output a test relevance score, and calculating the accuracy of the test relevance score based on the test score to evaluate the training effect of the first student scoring model. Alternatively, it can involve testing the first student scoring model using other methods; the above methods are merely examples and are not limited in this application.

[0193] In other words, if the first student rating model passes the test and the number of parameters is small enough, it can be deployed in an online environment. Conversely, if the first student rating model fails the test, further iterative training is needed to optimize its relevance discrimination ability. If the number of parameters in the first student rating model exceeds a set threshold, the model training device can train the initial second student rating model based on distilled training samples and distilled rating labels to obtain the second student rating model.

[0194] The initial second student model has fewer parameters than the first student scoring model. This is because a single-step distillation may result in a significant loss of model capabilities. Therefore, multi-step distillation can be performed to ensure that the core capabilities of the model with a large number of parameters are gradually transferred to the model with a smaller number of parameters. This ensures that the smaller model maintains high accuracy while having faster inference speed and lower resource consumption.

[0195] Furthermore, the model training device can also acquire test samples and corresponding test scores to test the second student scoring model. If the model parameters of the second student scoring model are less than the set parameter threshold and pass the test, it can be deployed in an online environment, and the intent information, historical attribute information, historical multimedia information, candidate attribute information, and candidate multimedia information obtained during the reasoning process can be input into the second student scoring model for real-time correlation discrimination.

[0196] In some embodiments, during each step of knowledge distillation, the teacher model can use high-quality training data generated and output by the teacher model based on training samples and rating labels as training sample cases to train the student model. For example, the first student rating model mentioned above can generate training samples and corresponding rating labels for the training process of the second student rating model. When constructing loss data, distillation loss data and true loss data can also be constructed for the second student rating model; details can be found in the detailed description of the first step of distillation, which will not be repeated here.

[0197] Please refer to the following: Figure 5 , Figure 5 This is a timing diagram of knowledge distillation provided in an embodiment of this application, such as... Figure 5 As shown, Figure 5 The process of two-step distillation is illustrated and explained. For example... Figure 5 As shown, the teacher model in the upper left corner could be a multimodal model with 7 billion parameters, which can be trained based on training samples and corresponding rating labels to fine-tune the model parameters. After training, this teacher model can generate a large amount of data related to the relevance judgment task in the personalized content recommendation service, which may include specified training samples, specified rating labels, and thought chain information. In the first step of distillation, Figure 5 The student model shown in the upper right corner can be a model with 4 billion parameters. This student model can be trained based on the training data generated by the teacher model with 70 parameters, as well as the original training data, so that the student model with 4 billion parameters can capture the knowledge and patterns learned by the super-large model.

[0198] In the specific training process, distillation loss data can be constructed based on the difference between the scores output by the teacher model and the scores output by the student model, and true loss data can be constructed based on the difference between the scores output by the student model and the score labels. The parameters of the student model can be adjusted based on the total loss (target loss data) obtained by summing the two. In other words, during training, the output of the large model (teacher model) can be used as the target value of the small model (student model) to achieve the goal and minimize the difference between the two.

[0199] After training a model with 4 billion parameters, since the number of model parameters still exceeds a set threshold, knowledge distillation can be further performed, for example... Figure 5The second distillation step, as shown, involves a model with 4 billion parameters on the left (as in the first student rating model mentioned above) and a model with 1.4 billion parameters on the right (as in the second student rating model mentioned above). The 1.4 billion parameter model can be trained using training samples and corresponding rating labels generated from the 4 billion parameter model, along with previous training samples and rating labels. This second distillation step further reduces the model's size while maintaining its performance.

[0200] Since a model with 1.4 billion parameters is not suitable for deployment in an online environment, a third distillation step can be performed. Similar to the first two steps, the 1.4 billion parameter model can be used as a teacher model to generate training samples and corresponding rating labels to train a student model with 400 million parameters. The specific training principles can be found in the implementation details of the first two distillation steps, which will not be elaborated here. This allows for further reduction in model size while ensuring high accuracy in the relevance judgment task of personalized content recommendation services, thereby improving inference speed.

[0201] Furthermore, the 400 million parameter model can be used as a teacher model for a fourth distillation step. For example, a BERT model with 60 million parameters can be trained based on the generated training samples and corresponding rating labels. Since the BERT model has a small enough number of parameters, it can be deployed in an online environment. Thus, through multiple distillation steps, the core capabilities of the ultra-large parameter model (knowledge of the business domain in personalized content recommendation services) can be transferred to a model with a smaller number of parameters, achieving a balance between performance and efficiency. This process not only significantly reduces the computational resource consumption of deploying the model in an online environment but also ensures the model's high-precision performance in personalized relevance discrimination tasks.

[0202] In some embodiments, after deploying a small model (such as the BERT model mentioned above) in an online environment, the performance of the model can be obtained through A / B testing and user feedback. For example, user interaction data such as click-through rate can be obtained for content recommended after relevance judgment, and compared with the interaction data of content recommended after deploying other models without deploying the BERT model, so as to continuously optimize the model performance and meet the needs of personalized content search and personalized content recommendation scenarios.

[0203] Please refer to the following: Figure 6 , Figure 6 This is a schematic diagram of the technical framework for a personalized relevance discrimination task provided in an embodiment of this application, such as... Figure 6As shown, training history browsing content, training intent information, and training candidate recommendation content can first be obtained from the database. Then, based on the relevance evaluation rules and evaluation process specified in the embodiments of this application, scoring labels are generated. Specifically, professionals or scoring label generation models can perform fine annotation on the training samples according to the established evaluation rules, evaluation process, and evaluation cases to generate high-quality relevance labels (i.e., scoring labels), which provide accurate guidance for model training. The scoring label can be a specific score from 1 to 5 points, which can be used to indicate the relevance between the training candidate recommendation content and the training intent information.

[0204] Furthermore, training samples and labels can be input into the text generation model, enabling it to generate thought chain information based on the training samples and rating labels. This allows the construction of a personalized relevance dataset containing thought chain information. Subsequently, this dataset can be used to fine-tune and train a parameter-rich, high-parameter multimodal model, giving it powerful multimodal information processing capabilities, personalized true intent recognition capabilities, and relevance discrimination capabilities. Finally, for deployment in an online environment, a large amount of training data can be generated based on the high-parameter multimodal model to train a low-parameter multimodal model, which serves as the online running model.

[0205] The embodiments of this application are as follows: Figure 2 Building upon the previous examples, by acquiring high-quality relevance scoring labels and generating thought chain information containing detailed reasoning processes, accurate guidance can be provided for model training. This enhances the model's logical reasoning ability, ensuring it can better understand the user's true intent and infer the correlation between the user and candidate recommended content. This, in turn, improves the accuracy of relevance judgment in personalized content recommendation services. Furthermore, training with a massive model, such as one with 7 billion parameters, followed by stepwise knowledge distillation and transfer learning techniques, can further improve the model's generalization ability and the accuracy of relevance judgment. During each model training phase, high-quality training data is generated in batches to ensure rapid model convergence, thereby shortening training time and improving training efficiency. Ultimately, a model with 60 million parameters can be deployed in a personalized content recommendation service (online environment) to quickly respond to user intent information with fast reasoning speed, thus providing content recommendation results in real time.

[0206] The methods of the embodiments of this application have been described in detail above. In order to facilitate better implementation of the above solutions of the embodiments of this application, the apparatus of the embodiments of this application is provided below.

[0207] Please see Figure 7 , Figure 7 This is a schematic diagram of the structure of a content recommendation device provided in an embodiment of this application. Figure 7 The content recommendation device shown can be mounted in a computer device, specifically a server. Figure 7 The content recommendation device shown can be used to perform the above. Figure 2 and Figure 3 Some or all of the functionality described in the method embodiments. Please refer to [link / reference]. Figure 7 The content recommendation device 70 includes:

[0208] The acquisition unit 701 is used to acquire the demand expression information input by the target object, the historical browsing content of the target object, and the candidate recommendation content for the target object, and extract the intent information from the demand expression information;

[0209] The generation unit 702 is used to match the intent information with the historical browsing content, and generate a relevance score between the intent information and the candidate recommended content based on the matching result between the intent information and a specific entity in the historical browsing content.

[0210] The processing unit 703 is used to perform content recommendation processing on the target object based on the relevance score and the candidate recommendation content.

[0211] In one possible implementation, the generation unit 702 is configured to match the intent information with the historical browsing content, and based on the matching result between the intent information and specific entities in the historical browsing content, generate a relevance score between the intent information and the candidate recommended content, specifically for:

[0212] Obtain historical description information of the historical browsing content, and candidate description information of the candidate recommended content;

[0213] Extract the multimedia information of the historical browsing content to obtain historical multimedia information, and extract the multimedia information of the candidate recommended content to obtain candidate multimedia information;

[0214] The intent information, historical attribute information, historical multimedia information, candidate attribute information, and candidate multimedia information are input into the scoring generation model. The scoring generation model matches the intent information and the historical browsing content, and generates the relevance score based on the matching results.

[0215] In one possible implementation, the rating generation model includes a scheduling module and at least two processing modules; the generation unit 702 is used to match the intent information and the historical browsing content through the rating generation model, and generate the relevance rating based on the matching result, specifically for:

[0216] Extract the intent information, the historical attribute information, the historical multimedia information, the candidate attribute information, and the feature information of the candidate multimedia information;

[0217] The scheduling module determines the target processing module that needs to be activated from the at least two processing modules based on the feature information.

[0218] The target processing module matches the intent information with the historical browsing content and generates the relevance score based on the matching results.

[0219] In one possible implementation, the acquisition unit 701 is further configured to acquire training samples and rating labels corresponding to the training samples. The training samples include training intent information, training history browsing content, training candidate recommendation content, and thought chain information associated with the training samples. The thought chain information is used to indicate the reasoning process of determining the correlation score between the training intent information and the training candidate recommendation content as the rating label based on the training history browsing content.

[0220] Training unit 704 is used to train the initial rating generation model based on the training samples and the rating labels to obtain the rating generation model.

[0221] In one possible implementation, the acquisition unit 701 is used to acquire the thought chain information associated with the training sample, specifically for:

[0222] The training intent information, the training history browsing content, the training candidate recommendation content, and the rating label are input into the text generation model. The text generation model determines the inference step information of the correlation rating between the training intent information and the training candidate recommendation content based on the training history browsing content according to the set inference rules.

[0223] Obtain the reasoning step information generated by the text generation model, and determine the reasoning step information as the thought chain information.

[0224] In one possible implementation, the inference step information further includes an inference relevance score between the training intent information generated by the text generation model based on the set inference rules and the training candidate recommendation content; the acquisition unit 701 is used to determine the inference step information as the thought chain information, specifically for:

[0225] If the reasoning relevance score is the same as the score indicated by the score label, then the reasoning step information is identified as the thought chain information.

[0226] In one possible implementation, the generation unit 702 is further configured to input the training sample and the rating label into the rating generation model if the number of model parameters of the rating generation model is greater than a set threshold, and generate a specified training sample and a specified rating label corresponding to the specified training sample by using the training sample and the rating label as training sample cases through the rating generation model.

[0227] The determining unit 705 is used to determine the training sample and the specified training sample as distillation training samples, and to determine the scoring label and the specified scoring label as distillation scoring labels;

[0228] The training unit 704 is used to train the initial first student rating model based on the distillation training samples and the distillation rating labels to obtain the first student rating model; wherein the number of model parameters of the initial first student rating model is less than that of the rating generation model.

[0229] In one possible implementation, the training unit 704 is used to train an initial first student rating model based on the distillation training samples and the distillation rating labels to obtain a first student rating model, specifically for:

[0230] The distilled training samples are input into the initial first student rating model, and the training relevance score generated by the initial first student rating model is obtained;

[0231] Based on the difference between the training relevance score and the distillation score label, a first loss data is constructed, and the model parameters of the initial first student scoring model are adjusted based on the first loss data to obtain the first student scoring model.

[0232] In one possible implementation, the training unit 704 is configured to adjust the model parameters of the initial first student rating model based on the first loss data to obtain the first student rating model, specifically for:

[0233] The distillation training samples are input into the rating generation model, and the reference relevance score generated by the rating generation model is obtained;

[0234] Based on the difference between the training relevance score and the reference relevance score, a second loss data is constructed;

[0235] Based on the sum of the first loss data and the second loss data, the target loss data is determined;

[0236] The model parameters of the initial first student rating model are adjusted based on the target loss data to obtain the first student rating model.

[0237] In one possible implementation, the acquisition unit 701 is further configured to acquire a test sample and a test score corresponding to the test sample;

[0238] The processing unit is further configured to perform test processing on the first student scoring model based on the test sample and the test score;

[0239] The generation unit 702 is used to input the intent information, the historical attribute information, the historical multimedia information, the candidate attribute information, and the candidate multimedia information into the scoring generation model, specifically for:

[0240] If the first student scoring model is determined to have passed the test based on the test sample and the test score, and the number of model parameters of the first student scoring model is less than the set threshold, then the intent information, the historical attribute information, the historical multimedia information, the candidate attribute information, and the candidate multimedia information are input into the first student scoring model.

[0241] In one possible implementation, the training unit 704 is further configured to train an initial second student rating model based on the distillation training samples and the distillation rating labels to obtain a second student rating model if the number of model parameters of the first student rating model is greater than the set threshold; wherein the number of model parameters of the initial second student rating model is less than that of the first student rating model.

[0242] The generation unit 702 is used to input the intent information, the historical attribute information, the historical multimedia information, the candidate attribute information, and the candidate multimedia information into the scoring generation model, specifically for:

[0243] If the model parameters of the second student scoring model are less than the set parameter threshold, then the intent information, the historical attribute information, the historical multimedia information, the candidate attribute information, and the candidate multimedia information are input into the second student scoring model.

[0244] According to one embodiment of this application, Figure 2 and Figure 3 The steps involved in the content recommendation method shown can be derived from... Figure 7 The content shown is recommended to be executed by each unit in the device. For example, Figure 2 Step S201 shown can be performed by Figure 7 The acquisition unit 701 shown is executed, and step S702 can be performed by... Figure 7 The generation unit 702 shown is executed, and step S703 can be performed by... Figure 7The processing unit 703 shown executes this. For example, Figure 3 Step S301 shown can be performed by Figure 7 The acquisition unit 701 shown is executed, and step S302 can be performed by... Figure 7 The training unit 704 shown is executed. Figure 7 The various units in the content recommendation device shown can be individually or entirely merged into one or more other units, or some of the units can be further divided into two or more functionally smaller units. This achieves the same operation without affecting the technical effects of the embodiments of this application. The above units are based on logical function division. In practical applications, the function of one unit can also be implemented by two or more units, or the function of multiple units can be implemented by one unit. In other embodiments of this application, the content recommendation device may also include other units. In practical applications, these functions can also be implemented with the assistance of other units, and can be implemented collaboratively by two or more units.

[0245] According to another embodiment of this application, a general-purpose computing device, such as a computer device including processing elements and storage elements such as a central processing unit (CPU), random access memory (RAM), and read-only memory (ROM), can perform operations such as... Figure 3 The computer program (including program code) for each step involved in the corresponding method shown, to construct such... Figure 7 The content recommendation apparatus shown herein, and the content recommendation method for implementing the embodiments of this application, are described. A computer program may be recorded on, for example, a computer-readable recording medium, loaded onto the aforementioned computing device via the computer-readable recording medium, and executed therein.

[0246] Based on the same inventive concept, the principle and beneficial effects of the content recommendation device provided in the embodiments of this application are similar to the principle and beneficial effects of the content recommendation method in the embodiments of this application. For details, please refer to the principle and beneficial effects of the method implementation. For the sake of brevity, these will not be repeated here.

[0247] Please see Figure 8 , Figure 8 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. The computer device may be a server. Figure 8As shown, the computer device includes at least a processor 801, a communication interface 802, and a memory 803. The processor 801, communication interface 802, and memory 803 can be connected via a bus or other means. The processor 801 (or Central Processing Unit, CPU) is the computing and control core of the computer device. It can parse various instructions within the computer device and process various data. For example, the CPU can parse power-on / off commands issued by an object to the computer device and control the computer device to perform power-on / off operations; it can also transmit various interactive data between internal structures of the computer device, and so on. The communication interface 802 may optionally include standard wired interfaces or wireless interfaces (such as Wi-Fi, mobile communication interfaces, etc.), and can be used to send and receive data under the control of the processor 801; the communication interface 802 can also be used for data transmission and interaction within the computer device. The memory 803 is the storage device in the computer device, used to store programs and data. It can be understood that the memory 803 here can include the computer device's built-in memory, or it can include extended memory supported by the computer device. The memory 803 provides storage space for storing the operating system of the computer device, which may include, but is not limited to, Android, iOS, Windows Phone, etc. This application does not limit this.

[0248] This application embodiment also provides a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the processing system of the computer device. Furthermore, the storage space also stores computer programs suitable for loading and execution by the processor 801. It should be noted that the computer-readable storage medium here can be high-speed RAM memory or non-volatile memory, such as at least one disk storage device; optionally, it can also be at least one computer-readable storage medium located remotely from the aforementioned processor.

[0249] In one embodiment, processor 801 performs the following operations by running a computer program stored in memory 803:

[0250] Obtain the demand expression information input by the target object, the target object's historical browsing content, and the candidate recommendation content for the target object, and extract the intent information from the demand expression information;

[0251] The intent information is matched with the historical browsing content, and a relevance score between the intent information and the candidate recommended content is generated based on the matching results between the intent information and specific entities in the historical browsing content.

[0252] Based on the relevance score and the candidate recommended content, content recommendation processing is performed on the target object.

[0253] In one possible implementation, the processor 801 executes a computer program stored in memory 803 to match the intent information with the historical browsing content, and generates a relevance score between the intent information and the candidate recommended content based on the matching result between the intent information and specific entities in the historical browsing content, specifically for:

[0254] Obtain historical description information of the historical browsing content, and candidate description information of the candidate recommended content;

[0255] Extract the multimedia information of the historical browsing content to obtain historical multimedia information, and extract the multimedia information of the candidate recommended content to obtain candidate multimedia information;

[0256] The intent information, historical attribute information, historical multimedia information, candidate attribute information, and candidate multimedia information are input into the scoring generation model. The scoring generation model matches the intent information and the historical browsing content, and generates the relevance score based on the matching results.

[0257] In one possible implementation, the rating generation model includes a scheduling module and at least two processing modules; the processor 801 executes a computer program stored in memory 803 to match the intent information and the historical browsing content using the rating generation model, and generates the relevance score based on the matching result, specifically for:

[0258] Extract the intent information, the historical attribute information, the historical multimedia information, the candidate attribute information, and the feature information of the candidate multimedia information;

[0259] The scheduling module determines the target processing module that needs to be activated from the at least two processing modules based on the feature information.

[0260] The target processing module matches the intent information with the historical browsing content and generates the relevance score based on the matching results.

[0261] In one possible implementation, the processor 801, by running a computer program in memory 803, is also used to perform:

[0262] Obtain training samples and corresponding rating labels for the training samples. The training samples include training intent information, training history browsing content, training candidate recommended content, and thought chain information associated with the training samples. The thought chain information is used to indicate the reasoning process of determining the correlation score between the training intent information and the training candidate recommended content as the rating label based on the training history browsing content.

[0263] The initial rating generation model is trained based on the training samples and the rating labels to obtain the rating generation model.

[0264] In one possible implementation, the processor 801 executes a computer program stored in memory 803 to acquire thought chain information associated with the training samples, specifically for:

[0265] The training intent information, the training history browsing content, the training candidate recommendation content, and the rating label are input into the text generation model. The text generation model determines the inference step information of the correlation rating between the training intent information and the training candidate recommendation content based on the training history browsing content according to the set inference rules.

[0266] Obtain the reasoning step information generated by the text generation model, and determine the reasoning step information as the thought chain information.

[0267] In one possible implementation, the inference step information further includes an inference relevance score between the training intent information generated by the text generation model based on the set inference rules and the training candidate recommendation content; the processor 801 executes a computer program in the memory 803 to determine the inference step information as the thought chain information, specifically for:

[0268] If the reasoning relevance score is the same as the score indicated by the score label, then the reasoning step information is identified as the thought chain information.

[0269] In one possible implementation, the processor 801, by running a computer program in memory 803, is also used to perform:

[0270] If the number of model parameters of the rating generation model is greater than a set threshold, then the training sample and the rating label are input into the rating generation model, and the rating generation model generates a specified training sample and a specified rating label corresponding to the specified training sample by using the training sample and the rating label as training sample cases.

[0271] The training samples and the designated training samples are determined as distillation training samples, and the scoring labels and the designated scoring labels are determined as distillation scoring labels;

[0272] Based on the distillation training samples and the distillation rating labels, an initial first student rating model is trained to obtain a first student rating model; wherein, the number of model parameters of the initial first student rating model is less than that of the rating generation model.

[0273] In one possible implementation, the processor 801 executes a computer program stored in memory 803 to train an initial first student rating model based on the distillation training samples and the distillation rating labels, thereby obtaining a first student rating model. Specifically, this is used to:

[0274] The distilled training samples are input into the initial first student rating model, and the training relevance score generated by the initial first student rating model is obtained;

[0275] Based on the difference between the training relevance score and the distillation score label, a first loss data is constructed, and the model parameters of the initial first student scoring model are adjusted based on the first loss data to obtain the first student scoring model.

[0276] In one possible implementation, the processor 801 executes a computer program stored in memory 803 to adjust the model parameters of the initial first student rating model based on the first loss data to obtain the first student rating model, specifically for:

[0277] The distillation training samples are input into the rating generation model, and the reference relevance score generated by the rating generation model is obtained;

[0278] Based on the difference between the training relevance score and the reference relevance score, a second loss data is constructed;

[0279] Based on the sum of the first loss data and the second loss data, the target loss data is determined;

[0280] The model parameters of the initial first student rating model are adjusted based on the target loss data to obtain the first student rating model.

[0281] In one possible implementation, the processor 801, by running a computer program in memory 803, is also used to perform:

[0282] Obtain the test sample and the corresponding test score;

[0283] The first student scoring model is tested based on the test samples and the test scores.

[0284] The step of inputting the intent information, the historical attribute information, the historical multimedia information, the candidate attribute information, and the candidate multimedia information into the scoring generation model includes:

[0285] If the first student scoring model is determined to have passed the test based on the test sample and the test score, and the number of model parameters of the first student scoring model is less than the set threshold, then the intent information, the historical attribute information, the historical multimedia information, the candidate attribute information, and the candidate multimedia information are input into the first student scoring model.

[0286] In one possible implementation, the processor 801, by running a computer program in memory 803, is also used to perform:

[0287] If the number of model parameters of the first student rating model is greater than the set threshold, then the initial second student rating model is trained based on the distillation training samples and the distillation rating labels to obtain the second student rating model; wherein the number of model parameters of the initial second student rating model is less than that of the first student rating model.

[0288] The step of feeding the intent information, the historical attribute information, the historical multimedia information, the candidate attribute information, and the candidate multimedia information into the scoring generation model includes:

[0289] If the model parameters of the second student scoring model are less than the set parameter threshold, then the intent information, the historical attribute information, the historical multimedia information, the candidate attribute information, and the candidate multimedia information are input into the second student scoring model.

[0290] It should be understood that, in the embodiments of this application, the processor 801 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0291] This application provides a computer-readable storage medium storing a computer program, which includes program instructions. When the program instructions are executed by a processor, they can perform the steps described in all the above embodiments.

[0292] This application also provides a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. When the computer instructions are executed by the processor of a computer device, they perform the methods described in all the above embodiments.

[0293] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0294] The above description discloses only one preferred embodiment of the present invention, and should not be construed as limiting the scope of the present invention. Those skilled in the art will understand that all or part of the processes of the above embodiments can be implemented, and equivalent changes made in accordance with the claims of the present invention are still within the scope of the invention.

[0295] It should also be noted that when the above embodiments of this application are applied to specific products or technologies, if it is necessary to obtain user data, the user's permission or consent must be obtained, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

Claims

1. A content recommendation method, characterized in that, include: Obtain the demand expression information input by the target object, the target object's historical browsing content, and the candidate recommendation content for the target object, and extract the intent information from the demand expression information; The intent information is matched with the historical browsing content, and a relevance score between the intent information and the candidate recommended content is generated based on the matching results between the intent information and specific entities in the historical browsing content. Based on the relevance score and the candidate recommended content, content recommendation processing is performed on the target object.

2. The method according to claim 1, characterized in that, The step of matching the intent information with the historical browsing content, and generating a relevance score between the intent information and the candidate recommended content based on the matching result between the intent information and specific entities in the historical browsing content, includes: Obtain historical description information of the historical browsing content, and candidate description information of the candidate recommended content; Extract the multimedia information of the historical browsing content to obtain historical multimedia information, and extract the multimedia information of the candidate recommended content to obtain candidate multimedia information; The intent information, historical attribute information, historical multimedia information, candidate attribute information, and candidate multimedia information are input into the scoring generation model. The scoring generation model matches the intent information and the historical browsing content, and generates the relevance score based on the matching results.

3. The method according to claim 2, characterized in that, The scoring generation model includes a scheduling module and at least two processing modules; The step of matching the intent information and the historical browsing content using the scoring generation model, and generating the relevance score based on the matching result, includes: Extract the intent information, the historical attribute information, the historical multimedia information, the candidate attribute information, and the feature information of the candidate multimedia information; The scheduling module determines the target processing module that needs to be activated from the at least two processing modules based on the feature information. The target processing module matches the intent information with the historical browsing content and generates the relevance score based on the matching results.

4. The method according to claim 2, characterized in that, The method further includes: Obtain training samples and corresponding rating labels. The training samples include training intent information, training history browsing content, training candidate recommendation content, and thought chain information associated with the training samples. The thought chain information is used to indicate the reasoning process of determining the correlation score between the training intent information and the training candidate recommendation content as the rating label based on the training history browsing content. The initial rating generation model is trained based on the training samples and the rating labels to obtain the rating generation model.

5. The method according to claim 4, characterized in that, The step of obtaining the thought chain information associated with the training samples includes: The training intent information, the training history browsing content, the training candidate recommendation content, and the rating label are input into the text generation model. The text generation model determines the inference step information of the correlation rating between the training intent information and the training candidate recommendation content based on the training history browsing content according to the set inference rules. Obtain the reasoning step information generated by the text generation model, and determine the reasoning step information as the thought chain information.

6. The method according to claim 5, characterized in that, The inference step information also includes an inference relevance score between the training intent information generated by the text generation model based on the set inference rules and the training candidate recommendation content. The step of determining the reasoning step information as the thought chain information includes: If the reasoning relevance score is the same as the score indicated by the score label, then the reasoning step information is identified as the thought chain information.

7. The method according to claim 4, characterized in that, The method further includes: If the number of model parameters of the rating generation model is greater than a set threshold, then the training sample and the rating label are input into the rating generation model, and the rating generation model generates a specified training sample and a specified rating label corresponding to the specified training sample by using the training sample and the rating label as training sample cases. The training samples and the designated training samples are determined as distillation training samples, and the scoring labels and the designated scoring labels are determined as distillation scoring labels; Based on the distillation training samples and the distillation rating labels, an initial first student rating model is trained to obtain a first student rating model; wherein, the number of model parameters of the initial first student rating model is less than that of the rating generation model.

8. The method according to claim 7, characterized in that, The step of training the initial first student rating model based on the distillation training samples and the distillation rating labels to obtain the first student rating model includes: The distilled training samples are input into the initial first student rating model, and the training relevance score generated by the initial first student rating model is obtained; Based on the difference between the training relevance score and the distillation score label, a first loss data is constructed, and the model parameters of the initial first student scoring model are adjusted based on the first loss data to obtain the first student scoring model.

9. The method according to claim 8, characterized in that, The step of adjusting the model parameters of the initial first student rating model based on the first loss data to obtain the first student rating model includes: The distillation training samples are input into the rating generation model, and the reference relevance score generated by the rating generation model is obtained; Based on the difference between the training relevance score and the reference relevance score, a second loss data is constructed; Based on the sum of the first loss data and the second loss data, the target loss data is determined; The model parameters of the initial first student rating model are adjusted based on the target loss data to obtain the first student rating model.

10. The method according to claim 7, characterized in that, The method further includes: Obtain the test sample and the corresponding test score; The first student scoring model is tested based on the test samples and the test scores. The step of inputting the intent information, the historical attribute information, the historical multimedia information, the candidate attribute information, and the candidate multimedia information into the scoring generation model includes: If the first student scoring model is determined to have passed the test based on the test sample and the test score, and the number of model parameters of the first student scoring model is less than the set threshold, then the intent information, the historical attribute information, the historical multimedia information, the candidate attribute information, and the candidate multimedia information are input into the first student scoring model.

11. The method according to claim 7, characterized in that, The method further includes: If the number of model parameters of the first student rating model is greater than the set threshold, then the initial second student rating model is trained based on the distillation training samples and the distillation rating labels to obtain the second student rating model; wherein the number of model parameters of the initial second student rating model is less than that of the first student rating model. The step of feeding the intent information, the historical attribute information, the historical multimedia information, the candidate attribute information, and the candidate multimedia information into the scoring generation model includes: If the model parameters of the second student scoring model are less than the set parameter threshold, then the intent information, the historical attribute information, the historical multimedia information, the candidate attribute information, and the candidate multimedia information are input into the second student scoring model.

12. A content recommendation device, characterized in that, include: The acquisition unit is used to acquire the demand expression information input by the target object, the historical browsing content of the target object, and the candidate recommendation content for the target object, and extract the intent information from the demand expression information; The generation unit is used to match the intent information with the historical browsing content, and generate a relevance score between the intent information and the candidate recommended content based on the matching result between the intent information and a specific entity in the historical browsing content. The processing unit is used to perform content recommendation processing on the target object based on the relevance score and the candidate recommended content.

13. An electronic device, characterized in that, include: One or more processors; A memory for storing one or more computer programs that, when executed by one or more processors, cause the electronic device to implement the content recommendation method according to any one of claims 1-11.

14. A computer-readable medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the content recommendation method according to any one of claims 1-11.

15. A computer program product, characterized in that, The computer program product includes a computer program stored in a computer-readable storage medium, and the processor of the electronic device reads from the computer-readable storage medium and executes the computer program, causing the electronic device to perform the content recommendation method according to any one of claims 1-11.

Images