Model training method and device, equipment, storage medium and program product

Abstract

Embodiments of the present application disclose a model training method and device, equipment, a storage medium and a program product. The method comprises: obtaining a training sample, the training sample comprising training search information and a plurality of label data; wherein one label data corresponds to one search understanding expert network in a search understanding model, and any search understanding expert network is used for error correction processing, rewriting processing, entity recognition processing, intent recognition processing or category classification processing; inputting the training sample into the search understanding model, and performing search understanding task processing on the training search information through a plurality of search understanding expert networks in the search understanding model to obtain a plurality of training task feature vectors corresponding to the training search information; and training the search understanding model according to the target of reducing the difference between each training task feature vector and the corresponding training label data to obtain a trained search understanding model. The embodiments of the present application can improve the training efficiency of the search understanding model.

Description

Technical Field

[0001] This application relates to the field of computer application technology, and in particular to a model training method, apparatus, device, storage medium, and program product. Background Technology

[0002] In current search scenarios, query understanding is a crucial link connecting user needs with content and product supply, directly impacting the accuracy of search results and the overall user experience. Query understanding permeates almost the entire search journey, typically involving multiple processes including spelling correction, query rewriting, named entity recognition (NER), intent detection, and taxonomy classification. However, currently, these multiple processes for search queries are handled by independently trained models, resulting in low model training efficiency. Summary of the Invention

[0003] The technical problem to be solved by the embodiments of this application is to provide a model training method, apparatus, device, storage medium and program product that can improve the training efficiency of search understanding models.

[0004] On the one hand, embodiments of this application provide a model training method, the method comprising: Acquire training samples, which include training search information and multiple label data corresponding to the training search information; wherein, each label data corresponds to a search understanding expert network in the search understanding model, and any search understanding expert network is used for error correction, rewriting, entity recognition, intent recognition, or category classification. The training samples are input into the search understanding model, and the training search information is processed by multiple search understanding expert networks in the search understanding model to obtain multiple training task feature vectors corresponding to the training search information; wherein, one training task feature vector is obtained through one search understanding expert network. The search understanding model is trained with the goal of reducing the difference between the feature vectors of each training task and the corresponding training label data, resulting in a trained search understanding model. The trained search understanding model is used to generate multiple search understanding information corresponding to the target search information based on the input target search information.

[0005] On the other hand, embodiments of this application provide a data retrieval method, which includes: Obtain the target search information for the target account; The trained search understanding model is used to process the target search information to obtain multiple search understanding information corresponding to the target search information; The multiple search understanding information is matched with the candidate media data to obtain the matching degree between the multiple search understanding information and the candidate media data; Based on the matching degree between the multiple search understanding information and the candidate media data, target media data is determined, and the target media data is used as target recall data.

[0006] On the other hand, embodiments of this application provide a model training apparatus, which includes: The acquisition unit is used to acquire training samples, which include training search information and multiple label data corresponding to the training search information; wherein, each label data corresponds to a search understanding expert network in the search understanding model, and any search understanding expert network is used to perform error correction processing, rewriting processing, entity recognition processing, intent recognition processing, or category classification processing. The processing unit is configured to input the training samples into the search understanding model, and process the training search information for search understanding tasks through multiple search understanding expert networks in the search understanding model to obtain multiple training task feature vectors corresponding to the training search information; wherein, one training task feature vector is obtained through one search understanding expert network. The training unit is used to train the search understanding model with the goal of reducing the difference between the feature vectors of each training task and the corresponding training label data, so as to obtain the trained search understanding model; wherein, the trained search understanding model is used to generate multiple search understanding information corresponding to the target search information based on the input target search information.

[0007] In one implementation, the processing unit performs search understanding task processing on the training search information through multiple search understanding expert networks in the search understanding model, obtaining multiple training task feature vectors corresponding to the training search information, which can be used for: Determine the processing order of each search understanding expert network; According to the processing order, the training task feature vector obtained by the previous one or more search understanding expert networks and the training search information are input into the next search understanding expert network. The next search understanding expert network then performs search understanding task processing on the training task feature vector and the training search information corresponding to the previous one or more search understanding expert networks to obtain the training task feature vector corresponding to the next search understanding expert network.

[0008] In one implementation, the plurality of search understanding expert networks include an error correction expert network, a rewriting expert network, an entity recognition expert network, an intent recognition expert network, and a category classification expert network. The processing unit, following the processing order, inputs the training task feature vector obtained through one or more previous search understanding expert networks and the training search information into a subsequent search understanding expert network. The subsequent search understanding expert network then performs search understanding task processing on the training task feature vector corresponding to the previous one or more search understanding expert networks and the training search information, obtaining the training task feature vector corresponding to the subsequent search understanding expert network, which can be used for: The training search information is corrected by the error correction expert network to obtain the training error correction feature vector. The training search information and training error correction feature vector are rewritten by the rewriting expert network to obtain the training rewritten feature vector. The entity recognition expert network performs entity recognition processing on the training search information, training error correction feature vector, and training rewritten feature vector to obtain training entity feature vector. The intent recognition expert network performs intent recognition processing on the training search information, training error correction feature vector, training rewrite feature vector, and training entity feature vector to obtain a training intent feature vector. The category classification expert network performs category classification processing on the training search information, training error correction feature vector, training rewrite feature vector, training entity feature vector, and training intent feature vector to obtain the training category feature vector.

[0009] In one implementation, the search understanding model further includes a gating network, and the training samples further include fused labeled data; the processing unit can also be used for: The gating network is used to perform feature processing on the training search information to obtain the task weights corresponding to the feature vectors of each training task. Based on the multiple training task feature vectors and the task weights corresponding to each training task feature vector, the training fusion feature vector corresponding to the training search information is obtained. The training unit trains the search understanding model with the goal of reducing the difference between the feature vectors of each training task and the corresponding training label data, resulting in a trained search understanding model that can be used for: The search understanding model is trained with the goal of reducing the difference between the training fusion feature vector and the fusion label data, resulting in the trained search understanding model.

[0010] In one embodiment, the acquisition unit can also be used for: Obtain the account information of the training account and the training published content; the training published content includes the content published by the training account during a historical time period; The processing unit inputs the training samples into the search understanding model, and performs search understanding task processing on the training search information through multiple search understanding expert networks in the search understanding model, obtaining multiple training task feature vectors corresponding to the training search information, which can be used for: The training samples, the account information of the training account, and the training published content are input into the search understanding model. The search understanding model uses multiple search understanding expert networks to process the training search information, the account information of the training account, and the training published content to obtain multiple training task feature vectors corresponding to the training search information.

[0011] In one embodiment, the processing unit can also be used for: Based on the training search information, pre-search processing is performed to obtain the pre-search media data corresponding to the training search information; The processing unit inputs the training samples into the search understanding model, and performs search understanding task processing on the training search information through multiple search understanding expert networks in the search understanding model, obtaining multiple training task feature vectors corresponding to the training search information, which can be used for: The training samples and the pre-search media data are input into the search understanding model, and the training search information and the pre-search media data are processed by multiple search understanding expert networks in the search understanding model to obtain multiple training task feature vectors corresponding to the training search information.

[0012] On the other hand, embodiments of this application provide a data retrieval device, which includes: The acquisition unit is used to acquire target search information for the target account. The processing unit is used to perform search understanding task processing on the target search information through the trained search understanding model to obtain multiple search understanding information corresponding to the target search information. A matching unit is used to match the plurality of search understanding information with candidate media data to obtain the matching degree between the plurality of search understanding information and candidate media data; The matching unit is further configured to determine target media data based on the matching degree between the plurality of search understanding information and the candidate media data, and the target media data is used as target recall data.

[0013] In one embodiment, the search understanding model includes multiple search understanding expert networks, which include an entity recognition expert network, an intent recognition expert network, and a category classification expert network; the multiple search understanding information includes entity recognition information and entity confidence corresponding to the entity recognition expert network, intent recognition information and intent confidence corresponding to the intent recognition expert network, and category classification information and category confidence corresponding to the category classification expert network. The processing unit can also be used for: If any confidence level is less than or equal to the confidence threshold, then the target search information is processed by a preset search understanding model to obtain multiple search understanding information corresponding to the target search information.

[0014] On the other hand, embodiments of this application provide a computer device, which includes a memory, a communication interface, and a processor, wherein the memory, the communication interface, and the processor are interconnected; the memory stores a computer program, and the processor calls the computer program stored in the memory to implement the above-described method.

[0015] On the other hand, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0016] On the other hand, embodiments of this application provide a computer program product, which includes a computer program stored in a computer storage medium; the processor of a computer device reads the computer program from the computer storage medium and executes the computer program, causing the computer device to perform the above-described method.

[0017] In this embodiment, training samples are obtained, including training search information and multiple label data corresponding to the training search information. Each label data corresponds to a search understanding expert network in the search understanding model. Each search understanding expert network is used for error correction, rewriting, entity recognition, intent recognition, or category classification. The training samples are input into the search understanding model, and multiple search understanding expert networks within the model process the training search information to perform search understanding tasks, resulting in multiple training task feature vectors corresponding to the training search information. Each search understanding expert network yields one training task feature vector. The search understanding model is trained with the goal of reducing the difference between each training task feature vector and its corresponding training label data, resulting in a trained search understanding model. The trained search understanding model is used to generate multiple search understanding information corresponding to the input target search information. By constructing multiple search understanding expert networks within the search understanding model and jointly training these networks, the training efficiency of the search understanding model can be improved. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the background art, the accompanying drawings used in the embodiments of this application or the background art will be described below.

[0019] Figure 1 This is a schematic diagram of the architecture of a model training system provided in an embodiment of this application; Figure 2 This is a schematic diagram of a hybrid expert network architecture provided in an embodiment of this application; Figure 3 This is a schematic flowchart of a model training method provided in an embodiment of this application; Figure 4 This is a flowchart illustrating a data retrieval method provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of a model training device provided in an embodiment of this application; Figure 6 This is a schematic diagram of the structure of a data retrieval device provided in an embodiment of this application; Figure 7 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0021] Furthermore, in the description of the embodiments of this application, the terms "first," "second," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance. Therefore, a feature specified as "first" or "second" may explicitly or implicitly include one or more features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0022] In the specific implementation of this application, user-related data is involved, such as user account information, search information, etc. When the embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data must comply with local laws, regulations and standards.

[0023] Research has found that search query understanding permeates almost the entire search chain, typically involving multiple processes such as error correction, rewriting, entity recognition, intent recognition, and category classification. However, current methods for handling these multiple search query processes involve independently training multiple models, resulting in low training efficiency. Therefore, this application's embodiments can obtain training samples, including training search information and multiple label data corresponding to the training search information. Each label data corresponds to a search understanding expert network within the search understanding model, and any search understanding expert network is used for error correction, rewriting, entity recognition, intent recognition, or category classification. The training samples are input into the search understanding model, and multiple search understanding expert networks within the model process the training search information to perform search understanding tasks, obtaining multiple training task feature vectors corresponding to the training search information. Each search understanding expert network yields one training task feature vector. The search understanding model is trained with the goal of minimizing the differences between each training task feature vector and its corresponding training label data, resulting in a trained search understanding model. This trained search understanding model is used to generate multiple search understanding information corresponding to the input target search information. By constructing multiple search understanding expert networks within the search understanding model and jointly training these networks, the training efficiency of the search understanding model can be improved.

[0024] Among them, the search understanding model can be a generative large model. A generative large model is a model based on a deep learning architecture, with massive parameters (such as more than one billion parameters) and trained in unsupervised or supervised manner. It can generate corresponding generative information based on the input information.

[0025] In this embodiment, the search understanding model can generate search understanding information corresponding to the input search information. The search information may include at least one of the following: search text information, search image information, and search video information. Optionally, if the search information includes search image information and / or search video information, the search understanding model can perform image semantic understanding on the search image information and / or search video information to obtain image understanding information, and then generate corresponding search understanding information based on the image understanding information.

[0026] Optionally, the search understanding model can be a Large Language Model (LLM). A Large Language Model is an artificial intelligence model based on a deep learning architecture (such as the Transformer architecture), pre-trained on a large amount of text data, with hundreds of billions or even trillions of parameters, and possessing powerful general language understanding and generation capabilities. It determines the complex grammatical structures, semantic relationships, and contextual logic in natural language through probabilistic modeling.

[0027] In this embodiment, the search understanding model can be used to perform search understanding, that is, to generate search understanding information corresponding to the input search information. The input search information includes search text information.

[0028] Optionally, the search understanding model may include multiple search understanding expert networks. An expert network refers to multiple independent sub-models or networks, each focusing on processing input data of a specific type or domain.

[0029] In this embodiment, the search understanding expert network is a sub-network (or module) with independent parameters and a dedicated structure for performing a specific search understanding task. Specifically, the search understanding model may include, but is not limited to, at least two of the following: an error correction expert network, a rewriting expert network, an entity recognition expert network, an intent recognition expert network, and a category classification expert network. The error correction expert network is used to perform error correction processing; the rewriting expert network is used to perform rewriting processing; the entity recognition expert network is used to perform entity recognition processing; the intent recognition expert network is used to perform intent recognition processing; and the category classification expert network is used to perform category classification processing.

[0030] Optionally, multiple search understanding expert networks employ a Mixture-of-Experts (MoE) architecture. MoE is a neural network architecture that improves overall model performance by integrating the predictions of multiple models or "experts." MoE assigns input data to different expert sub-models and then merges the outputs of all sub-models to generate the final result. This assignment can be dynamically adjusted based on the characteristics of the input data, ensuring that each expert handles the data type or task type they are best at, thus achieving more efficient and accurate predictions. MoE also includes a gating network, which determines which expert(s) should process each input data. It calculates the weights or importance of each expert based on the characteristics of the input data and then assigns the input data to the appropriate experts according to these weights. The gating network is typically a neural network, and its output is processed by a softmax activation function to ensure that the sum of the weights of all experts is 1.

[0031] In this embodiment, the multiple search understanding expert networks include, but are not limited to, at least two of the following: error correction expert networks, rewriting expert networks, entity recognition expert networks, intent recognition expert networks, and category classification expert networks. Each search understanding expert network generates a corresponding task feature vector based on the input search information. The gating network generates task weights corresponding to each task feature vector based on the input search information. Then, based on each task feature vector and its corresponding task weight, a fused feature vector is generated, and search understanding information is generated based on the fused feature vector.

[0032] Based on the above description, please refer to Figure 1 , Figure 1 This is a schematic diagram of the architecture of a model training system provided in an embodiment of this application. Training samples can be obtained, including training search information and multiple label data corresponding to the training search information. Each label data corresponds to a search understanding expert network in the search understanding model. Any search understanding expert network is used for error correction, rewriting, entity recognition, intent recognition, or category classification.

[0033] The training samples are input into the search understanding model, and the training search information is processed by multiple search understanding expert networks in the search understanding model to obtain multiple training task feature vectors corresponding to the training search information; wherein, one training task feature vector is obtained through one search understanding expert network.

[0034] Optionally, the account information of the training account and the training published content can be obtained, and the training samples, the account information of the training account, and the training published content can be input into the search understanding model. Through multiple search understanding expert networks in the search understanding model, the training search information, the account information of the training account, and the training published content are processed for search understanding tasks to obtain multiple training task feature vectors corresponding to the training search information.

[0035] Optionally, after inputting the training samples, the account information of the training account, and the training published content into the search understanding model, the search information, the account information of the training account, and the training published content can be processed through the self-attention layer and the normalization layer to obtain the feature vector of the context representation. Through multiple search understanding expert networks in the search understanding model, the feature vector of the context representation is processed by the search understanding task to obtain multiple training task feature vectors corresponding to the training search information.

[0036] The search understanding model is trained with the goal of reducing the difference between the feature vectors of each training task and the corresponding training label data, resulting in a trained search understanding model. The trained search understanding model is used to generate multiple search understanding information corresponding to the target search information based on the input target search information.

[0037] Among them, several search understanding expert networks adopt a hybrid expert network architecture; please refer to [link to relevant documentation]. Figure 2 , Figure 2This is a schematic diagram of a hybrid expert network architecture provided in an embodiment of this application. The feature vectors representing the context are input into a rewriting expert network, an entity recognition expert network, and a category classification expert network. The rewriting expert network rewrites the feature vectors representing the context to obtain training rewritten feature vectors. The entity recognition expert network performs entity recognition processing on the feature vectors representing the context and the training rewritten feature vectors to obtain training entity feature vectors. The category classification expert network performs category classification processing on the feature vectors representing the context, the training rewritten feature vectors, and the training entity feature vectors to obtain training category feature vectors. A gating network performs feature processing on the feature vectors representing the context to obtain task weights corresponding to each training task feature vector. Based on the training rewritten feature vectors, training entity feature vectors, training category feature vectors, and the task weights corresponding to each training task feature vector, a training fusion feature vector corresponding to the training search information is obtained. The training samples also include fused label data. Therefore, the search understanding model can be trained with the goal of reducing the difference between the training fusion feature vectors and the fusion label data to obtain the trained search understanding model. Optionally, the trained fused feature vector is then passed through a hidden layer for feature transformation (such as feature alignment, feature integration, etc.) to obtain the transformed trained fused feature vector. Optionally, the multiple search understanding expert networks may also include an error correction expert network and an intent recognition expert network. The error correction expert network performs error correction processing on the context representation feature vector to obtain a trained error-corrected feature vector; the intent recognition expert network performs intent recognition processing on the context representation feature vector, the trained error-corrected feature vector, the trained rewritten feature vector, and the trained entity feature vector to obtain a trained intent feature vector.

[0038] The transformed training fusion feature vector output by the hybrid expert network is normalized through a normalization layer to obtain a normalized training search feature vector. This normalized feature vector is then used to determine classification task information and generation task information. The classification task information and generation task information can be search understanding information corresponding to the training search information. Generation task information may include error correction and rewriting information. Classification task information may include entity recognition information, intent recognition information, and category classification information. Optionally, feature parsing can be performed on the normalized training search feature vector to determine the classification task information and generation task information.

[0039] In this embodiment, multiple search understanding expert networks employ a hybrid expert network (MoE) architecture, which is better compatible with system response time (RT) sensitive scenarios. RT-sensitive scenarios are those with extremely high requirements for response latency and the need for rapid result return, such as search query understanding scenarios. After a user initiates a search, results need to be obtained extremely quickly; excessive latency will severely impact the user experience. The MoE architecture is suitable for such scenarios because it requires fewer activation parameters for online inference. Specifically: 1. The MoE architecture splits the model into multiple independent "expert" sub-networks. During each inference, the gating network only selects the parameters of a few experts to participate in the calculation based on the input content. Unlike dense large models with the same parameter scale, it does not need to call all parameters to complete the inference. The actual computational load is much lower than that of dense models of the same scale, and the inference speed is greatly improved. It can effectively reduce the response latency and meet the RT requirements.

[0040] For example, with the same model having hundreds of billions of parameters, a dense model has to perform calculations on all hundreds of billions of parameters in each inference, resulting in very high latency; however, with the MoE architecture, assuming there are 64 experts and only 2 experts are activated to participate in the calculation each time, the actual number of parameters called is only a few billion, which can improve the inference speed by several times and significantly shorten the response time.

[0041] 2. Additional Advantages: MoE consumes less computing power during inference, can handle higher inference concurrency on the same hardware, and maintains stable low latency even during peak traffic periods, avoiding issues like numerous request timeouts and RT spikes. This makes it more suitable for high-concurrency, RT-sensitive businesses like search. In query understanding scenarios, which are the front-end modules of the search chain and have extremely high latency requirements, the MoE architecture can leverage the capabilities of a large model to improve the understanding of each subtask without slowing down the overall response speed, perfectly matching the needs of such scenarios.

[0042] Based on the above description, the embodiments of this application can be applied in data retrieval scenarios. Specifically, the target object corresponding to the target account can input target search information through a client. The client sends the target search information to the server. The server processes the target search information using a trained search understanding model to obtain multiple search understanding information corresponding to the target search information. The server matches the multiple search understanding information with candidate media data to obtain the matching degree between the multiple search understanding information and the candidate media data. The candidate media data includes, but is not limited to, published content, products, live broadcasts, and advertising data. Published content refers to content information pre-published by users on a network platform, which may include text, images, audio, and video information, and can be displayed in the form of notes, articles, video files, etc. The specific content information included in the published content can be adjusted accordingly for different scenarios, and the content information in the published content can be determined by the publisher. The server determines the target media data based on the matching degree between the multiple search understanding information and the candidate media data, and the target media data is used as target retrieval data. The server sends the target retrieval data to the client, and the client renders and displays the search results page based on the target retrieval data.

[0043] Based on the above description, please refer to Figure 3 , Figure 3 This is a flowchart illustrating a model training method provided in an embodiment of this application, as shown below. Figure 3 The model training method shown includes, but is not limited to, steps S301 to S303, wherein: S301, Obtain training samples. The training samples include training search information and multiple label data corresponding to the training search information. Among them, each label data corresponds to a search understanding expert network in the search understanding model. Any search understanding expert network is used for error correction, rewriting, entity recognition, intent recognition, or category classification.

[0044] The training samples are used to train the search understanding model. The training samples include training search information and multiple labeled data corresponding to that information. Each labeled data corresponds to a search understanding expert network within the search understanding model. Each expert network is used for error correction, rewriting, entity recognition, intent recognition, or category classification. Specifically, the search understanding expert network performs these processes on the training search information, generating training error correction feature vectors, training rewriting feature vectors, training entity feature vectors, training intent feature vectors, or training category feature vectors. The multiple labeled data include error correction label data corresponding to the training error correction feature vectors, rewriting label data corresponding to the training rewriting feature vectors, entity label data corresponding to the training entity feature vectors, intent label data corresponding to the training intent feature vectors, and category label data corresponding to the training category feature vectors.

[0045] Optionally, the search understanding model can be a generative large model, which can generate multiple training task feature vectors corresponding to the input training search information, and determine the search understanding information corresponding to the training search information based on the multiple training task feature vectors. The training search information may include at least one of the following: training search text information, training search image information, and training search video information. Optionally, if the training search information includes training search image information and / or training search video information, then the search understanding model can perform image semantic understanding on the training search image information and / or training search video information to obtain image understanding information, and then generate multiple corresponding training task feature vectors based on the image understanding information. Optionally, the search understanding model can be a large language model, which can generate multiple training task feature vectors corresponding to the input training search information, and determine the search understanding information corresponding to the training search information based on the multiple training task feature vectors. The input training search information includes training search text information.

[0046] Optionally, the search understanding model may include multiple search understanding expert networks. Any one of these expert networks can be used for error correction, rewriting, entity recognition, intent recognition, or category classification.

[0047] Specifically, the search understanding model may include, but is not limited to, at least two of the following: error correction expert network, rewriting expert network, entity recognition expert network, intent recognition expert network, and category classification expert network. The error correction expert network is used to perform error correction processing; the rewriting expert network is used to perform rewriting processing; the entity recognition expert network is used to perform entity recognition processing; the intent recognition expert network is used to perform intent recognition processing; and the category classification expert network is used to perform category classification processing.

[0048] The error correction expert network is used to perform error correction processing, that is, to identify and correct information errors and non-standard expressions in the training search information. Error correction processing includes, but is not limited to: spelling correction, grammar and standardization correction. Spelling correction is used to correct spelling errors, such as correcting the training search information "spring women's coat" to "spring women's outerwear"; grammar and standardization correction is used to adjust expressions according to standard language habits or domain norms, such as optimizing the colloquial "buy clothes store" to the standard "clothing store".

[0049] The rewriting expert network is used to perform rewriting processes, which generate synonymous, expanded, or normalized training search information that is more suitable for search engine retrieval or understanding, while maintaining the search intent of the original training search information. Rewriting processes include, but are not limited to: synonym substitution and expansion, and query simplification or generalization. Specifically, synonym substitution and expansion are used to replace or expand the training search information with synonyms, such as rewriting "sports shoe prices" to "latest sports shoe price quotes"; query simplification or generalization is used to simplify or generalize the training search information, such as simplifying "what are the latest mobile phone models released in 2026" to "new mobile phone models".

[0050] The entity recognition expert network is used to perform entity recognition processing, which involves detecting, identifying, and classifying named entities or keywords with specific semantic types from the training search information. Entity recognition processing includes, but is not limited to, entity boundary recognition and entity type classification. Entity boundary recognition is used to mark the start and end positions of entities; entity type classification is used to categorize the identified entities into predefined entity categories, such as "person," "brand," "product model," "location," "movie title," etc.

[0051] The intent recognition expert network is used to perform intent recognition processing, that is, to determine the fundamental purpose or intent category of the search. The identified intents can include, but are not limited to, any of the following: navigation intents, informational intents, and transactional intents. Navigational intents refer to specific websites or pages that the user hopes to reach. Informational intents refer to knowledge or information that the user desires to obtain. Transactional intents refer to the user's intention to perform actions such as purchasing or downloading.

[0052] The category classification expert network is used to perform category classification processing, mapping training search information to preset categories. Category classification includes, but is not limited to, any of the following: product categories, content categories, and live streaming categories. Specifically, product categories indicate the classification of products; for example, if the training search information is "Bluetooth wireless noise-canceling headphones," then this training search information can be classified into the category "Electronics > Audio Equipment > Headphones > Bluetooth Noise-Canceling Headphones." Content categories indicate the classification of content; for example, if the training search information is "Football Schedule," then this training search information can be classified into the category "Sports > Football."

[0053] Optionally, the search understanding expert network may further include a word segmentation expert network, which performs word segmentation processing on the training search information. Specifically, the word segmentation expert network identifies the boundaries of basic units (i.e., "words") with independent semantic or syntactic functions in the training search information, thereby converting the training search information into a normalized sequence composed of word or sub-word units, i.e., multiple tokens. A token refers to the smallest text unit that the model processes and understands. Optionally, the output data of the word segmentation expert network can be input into at least one of an error correction expert network, a rewriting expert network, an entity recognition expert network, an intent recognition expert network, and a category classification expert network to perform search understanding task processing.

[0054] Optionally, the search understanding expert network may also include a term analysis expert network, which is used to determine the importance of entities in the training search information. Specifically, the term analysis expert network determines the importance weights of each entity in the training search information, so that when the intent recognition expert network performs intent recognition processing and the category classification expert network performs category classification processing, the importance weights of each entity are combined to determine the intent and category corresponding to the training search information.

[0055] Optionally, multiple search understanding expert networks adopt a MoE architecture, which includes a gating network. These multiple search understanding expert networks include, but are not limited to, at least two of an error correction expert network, a rewriting expert network, an entity recognition expert network, an intent recognition expert network, and a category classification expert network. Each search understanding expert network generates a corresponding training task feature vector based on the input training search information. The gating network generates task weights corresponding to each training task feature vector based on the input training search information. Furthermore, based on the training task feature vectors and their corresponding task weights, a training fusion feature vector is generated, and search understanding information corresponding to the training search information is generated based on this training fusion feature vector.

[0056] In one alternative implementation, the account information of the training account and the training published content can also be obtained; the training published content includes the content published by the training account during historical time periods.

[0057] To enhance the personalization and context-awareness of the search understanding model, account information and training content can be obtained. Account information may include details such as age, gender, and interests. Training content includes posts made by the training account over a historical period, such as clicks, views, likes, favorites, and comments within a 30-day timeframe. Furthermore, the training content, account information, and sample data can be input together into the search understanding model for training.

[0058] In this embodiment, the training content, account information, and sample data are input together into the search understanding model for training, which can improve the training effect and accuracy of the search understanding model.

[0059] Optionally, the training content can include the content posted by the training account during a target historical time period. The target historical time period can refer to a period of time whose time interval with the current system time is less than or equal to a preset time interval, such as the most recent period (within the last 7 days). In other words, the training account's recent interactive content can be obtained, and this content, account information, and sample data can be input into the search understanding model for training.

[0060] In this embodiment of the application, the published content of the target interaction within the target time period of the training account is obtained, that is, the published content of the interaction in the most recent period. The published content of the interaction in the most recent period can better represent the user's interests and preferences, thereby improving the training effect of the search understanding model and improving the accuracy of the search understanding model in understanding search information.

[0061] Optionally, training content may include content associated with training search information.

[0062] The published content associated with the training search information can refer to published content related to the search behavior corresponding to the training search information, or published content semantically associated with the training search information. Published content associated with the search behavior corresponding to the training search information can be understood as: published content generated by an interactive behavior (such as clicking, browsing, or saving) after any account inputs search information that is the same as or similar to the training search information; this published content is considered as published content associated with the training search information. Published content semantically associated with the training search information can be understood as: obtaining the search semantics of the training search information and the content semantics of candidate published content, calculating the semantic similarity between the search semantics and the content semantics of each candidate published content, thereby determining the published content associated with the training search information from the candidate published content. Candidate published content can be obtained from a published content database. The semantic similarity between the search semantics and the content semantics of each candidate published content can be sorted from largest to smallest, and the candidate published content that meets the preset ranking (such as the top K, where K is a positive integer) can be used as the published content associated with the training search information; or, the candidate published content whose semantic similarity between the search semantics and the content semantics of each candidate published content meets the preset similarity threshold can be used as the published content associated with the training search information.

[0063] Optionally, the training content can include content that is associated with the training search information and meets a preset level of interaction.

[0064] Once the published content associated with the training search information is determined, the interaction volume of each published content associated with the training search information can be obtained. The interaction volume of any published content can be determined based on the interaction operations of each account on the published content (such as likes, favorites, comments, etc.), and published content with an interaction volume greater than or equal to the preset interaction volume is used as training published content.

[0065] Optionally, the training content published may include content published by the training account during historical time periods that is associated with the training search information. For a detailed description of the content published that is associated with the training search information, please refer to the relevant descriptions above, which will not be repeated here.

[0066] In one alternative implementation, pre-search processing can also be performed based on the training search information to obtain pre-search media data corresponding to the training search information.

[0067] Pre-search processing refers to the search processing performed on the training search information using pre-defined search rules or models before it is input into the search understanding model. Pre-search media data refers to the media data obtained by pre-searching the training search information. This media data includes, but is not limited to, published content, product data, live stream data, and advertising data.

[0068] Optionally, the search understanding model includes a feature encoding layer, also known as an embedding layer or contextualized query encoder. This layer encodes the training search information to obtain encoded training search features. Then, through multiple search understanding expert networks in the search understanding model, the encoded training search features are processed for search understanding tasks to obtain multiple training task feature vectors corresponding to the encoded training search features.

[0069] Specifically, the feature encoding layer maps each word or sub-token in the input training search information into a high-dimensional, learnable embedding vector and performs position encoding, that is, generates a specific position encoding vector for each position in the sequence and adds it to the word embedding vector to obtain the encoded training search features.

[0070] Optionally, at least one of the following can be used: account information of the training account, training published content, pre-search media data, and training search information. This results in encoded training search features. Multiple search understanding expert networks in the search understanding model are then used to process the encoded training search features for search understanding tasks, resulting in multiple training task feature vectors corresponding to the encoded training search features.

[0071] Optionally, the training search information can be subjected to self-attention and normalization processing to obtain the feature vector of the context representation, and the feature vector of the context representation can be input into the search understanding expert network for search understanding task processing.

[0072] Self-attention processing refers to encoding the tokens corresponding to the training search information using self-attention, resulting in self-attention-encoded training search features. These features represent the distance relationships between tokens. Normalization processing, following self-attention processing, involves standardizing the training search features (e.g., normalization) to obtain a contextual feature vector. Normalization standardizes the features at each position in the self-attention-processed training search features, stabilizing their mean and variance. This accelerates the convergence of the search understanding model and improves its training stability. The contextual feature vector is then input into the search understanding expert network for search understanding task processing.

[0073] Optionally, the encoded training search features can be subjected to self-attention and normalization processing to obtain the context representation feature vector, and the context representation feature vector can be input into the search understanding expert network for search understanding task processing.

[0074] Optionally, self-attention and normalization processing can be performed on at least one of the following: account information of the training account, training published content, pre-search media data, and training search information, to obtain a feature vector representing the context. This feature vector representing the context is then input into the search understanding expert network for search understanding task processing.

[0075] S302, the training samples are input into the search understanding model, and the training search information is processed by multiple search understanding expert networks in the search understanding model to obtain multiple training task feature vectors corresponding to the training search information; wherein, one training task feature vector is obtained through one search understanding expert network.

[0076] This can be achieved by using multiple search understanding expert networks within the search understanding model to process the training search information in parallel, meaning multiple search understanding expert networks process the training search information simultaneously; or, multiple search understanding expert networks within the search understanding model can process the training search information sequentially. The multiple training task feature vectors include training error correction feature vectors corresponding to the error correction expert network, training rewrite feature vectors corresponding to the rewrite expert network, training entity feature vectors corresponding to the entity recognition expert network, training intent feature vectors corresponding to the intent recognition expert network, and training category feature vectors corresponding to the category classification expert network.

[0077] In one optional implementation, the processing order of each search understanding expert network can be determined; and according to the processing order, the training task feature vector and training search information obtained by the previous search understanding expert network are input into the next search understanding expert network, and the next search understanding expert network performs search understanding task processing on the training task feature vector and training search information corresponding to the previous search understanding expert network to obtain the training task feature vector corresponding to the next search understanding expert network.

[0078] The processing order of each search understanding expert network can be pre-set; for example, the error correction expert network can be set to process before the rewriting expert network, and the rewriting expert network can be set to process before the entity recognition expert network, etc. Alternatively, the processing order of each search understanding expert network can be determined based on the search understanding task corresponding to each network. For example, if error correction precedes rewriting, then the error correction expert network will process before the rewriting expert network; similarly, if rewriting precedes entity recognition, then the rewriting expert network will process before the entity recognition expert network. Based on the processing order of each search understanding expert network, the training task feature vector and training search information obtained from the previous network are input into the next network. The next network then processes the training task feature vector and training search information from the previous network to obtain the corresponding training task feature vector.

[0079] In one optional implementation, the processing order of each search understanding expert network can be determined; and according to the processing order, the training task feature vector and training search information obtained by the previous one or more search understanding expert networks are input into the next search understanding expert network, and the next search understanding expert network performs search understanding task processing on the training task feature vector and training search information corresponding to the previous one or more search understanding expert networks to obtain the training task feature vector corresponding to the next search understanding expert network.

[0080] If, before the search understanding expert network corresponding to the next processing order, multiple search understanding expert networks have processed the training search information to obtain the training task feature vector, then the training task feature vectors and training search information obtained by the multiple search understanding expert networks preceding the search understanding expert network corresponding to the next processing order can be input into the next search understanding expert network. The next search understanding expert network then processes the training task feature vectors and training search information corresponding to the previous one or more search understanding expert networks to obtain the training task feature vector corresponding to the next search understanding expert network.

[0081] In this embodiment, the training task feature vectors and training search information obtained from one or more previous search understanding expert networks are input into a subsequent search understanding expert network. The subsequent search understanding expert network then processes the training task feature vectors and training search information obtained from the previous one or more search understanding expert networks to perform search understanding task processing. This allows downstream processing to consume the upstream processing results, reduces error propagation, and improves the accuracy of search understanding for each search understanding expert network.

[0082] In one optional implementation, the multiple search understanding expert networks include an error correction expert network, a rewriting expert network, an entity recognition expert network, an intent recognition expert network, and a category classification expert network. The error correction expert network can then perform error correction processing on the training search information to obtain a training error-correcting feature vector; the rewriting expert network can rewrite the training search information and the training error-correcting feature vector to obtain a training rewriting feature vector; the entity recognition expert network can perform entity recognition processing on the training search information and the training rewriting feature vector to obtain a training entity feature vector; the intent recognition expert network can perform intent recognition processing on the training search information and the training entity feature vector to obtain a training intent feature vector; and the category classification expert network can perform category classification processing on the training search information and the training intent feature vector to obtain a training category feature vector.

[0083] The search understanding expert networks include error correction expert networks, rewriting expert networks, entity recognition expert networks, intent recognition expert networks, and category classification expert networks. The processing order of the error correction expert networks is before that of the rewriting expert networks, entity recognition expert networks, intent recognition expert networks, and category classification expert networks; the processing order of the rewriting expert networks is before that of the entity recognition expert networks, intent recognition expert networks, and category classification expert networks; the processing order of the entity recognition expert networks is before that of the intent recognition expert networks and category classification expert networks; and the processing order of the intent recognition expert networks is before that of the category classification expert networks. Then, an error correction expert network can be used to correct the training search information, resulting in a training error-corrected feature vector, which indicates the corrected training search information. A rewriting expert network can be used to rewrite the training search information and the training error-corrected feature vector, resulting in a training rewritten feature vector, which indicates the rewritten training search information. An entity recognition expert network can be used to perform entity recognition on the training search information and the training rewritten feature vector, resulting in a training entity feature vector, which indicates the entity corresponding to the training search information. An intent recognition expert network can be used to perform intent recognition on the training search information and the training entity feature vector, resulting in a training intent feature vector, which indicates the intent corresponding to the training search information. Finally, a category classification expert network can be used to perform category classification on the training search information and the training intent feature vector, resulting in a training category feature vector, which indicates the category corresponding to the training search information.

[0084] In one optional implementation, the multiple search understanding expert networks include an error correction expert network, a rewriting expert network, an entity recognition expert network, an intent recognition expert network, and a category classification expert network. The error correction expert network can then perform error correction processing on the training search information to obtain a training error-correcting feature vector; the rewriting expert network can rewrite the training search information and the training error-correcting feature vector to obtain a training rewriting feature vector; the entity recognition expert network can perform entity recognition processing on the training search information, the training error-correcting feature vector, and the training rewriting feature vector to obtain a training entity feature vector; the intent recognition expert network can perform intent recognition processing on the training search information, the training error-correcting feature vector, the training rewriting feature vector, and the training entity feature vector to obtain a training intent feature vector; and the category classification expert network can perform category classification processing on the training search information, the training error-correcting feature vector, the training rewriting feature vector, the training entity feature vector, and the training intent feature vector to obtain a training category feature vector.

[0085] In this embodiment, the rewriting expert network performs rewriting processing based on training search information and training error correction feature vectors; the entity recognition expert network performs entity recognition processing based on training search information, training error correction feature vectors, and training rewritten feature vectors; the intent recognition expert network performs intent recognition processing based on training search information, training error correction feature vectors, training rewritten feature vectors, and training entity feature vectors; and the category classification expert network performs category classification processing based on training search information, training error correction feature vectors, training rewritten feature vectors, training entity feature vectors, and training intent feature vectors. By allowing downstream fields to directly consume multiple results from upstream, error propagation is reduced, and the accuracy of the search understanding model's search understanding is improved.

[0086] In one alternative implementation, the search understanding model further includes a gating network, which can then be used to perform feature processing on the training search information to obtain the task weights corresponding to the feature vectors of each training task; based on the feature vectors of multiple training tasks and the task weights corresponding to the feature vectors of each training task, a training fusion feature vector corresponding to the training search information is obtained.

[0087] The gating network determines the task weights corresponding to each training task feature vector based on the training search information. The task weights can range from 0 to 1. Based on multiple training task feature vectors and their corresponding task weights, the feature vectors are weighted and fused to obtain a training fusion feature vector corresponding to the training search information. Specifically, each training task feature vector is multiplied by its corresponding task weight, and then all weighted feature vectors are summed or concatenated to obtain the training fusion feature vector. This training fusion feature vector is used to comprehensively represent the understanding of the training search information by each training task feature vector. Optionally, when determining the task weights for each training task feature vector, the gating network can apply load balancing regularization to ensure a balanced distribution of the determined task weights.

[0088] In one optional implementation, the account information of the training account and the training published content can also be obtained; the training published content includes the content published by the training account during a historical time period; the training samples, the account information of the training account, and the training published content are input into the search understanding model, and the search understanding task is processed by multiple search understanding expert networks in the search understanding model to obtain multiple training task feature vectors corresponding to the training search information.

[0089] For details regarding the account information of the training accounts and the specific descriptions of the training published content, please refer to the relevant descriptions above, which will not be repeated here. Through multiple search understanding expert networks in the search understanding model, search understanding tasks are processed on the training search information, the account information of the training accounts, and the training published content. In other words, each search understanding expert network processes the search understanding task based on the training search information, combined with the account information of the training accounts and the training published content. This improves the model's training performance and search understanding accuracy, supplements semantic clues for long-tail or cross-domain queries, and enhances retrieval precision and recall coverage for complex long-tail queries.

[0090] In one optional implementation, pre-search processing can be performed based on the training search information to obtain pre-search media data corresponding to the training search information; the training samples and pre-search media data are input into the search understanding model, and the training search information and pre-search media data are processed by multiple search understanding expert networks in the search understanding model to obtain multiple training task feature vectors corresponding to the training search information.

[0091] Pre-search processing refers to the search processing performed on the training search information before it is input into the search understanding model, using pre-defined search rules or a search model. Pre-search media data refers to the media data obtained by pre-searching the training search information. This media data includes, but is not limited to, published content, product data, live stream data, and advertising data. Through multiple search understanding expert networks within the search understanding model, search understanding tasks are processed on both the training search information and the pre-search media data. In other words, each search understanding expert network performs search understanding tasks based on the training search information and the pre-search media data, which improves the model's training performance and search understanding accuracy, supplements semantic clues for long-tail or cross-domain queries, and enhances retrieval precision and recall coverage for complex long-tail queries.

[0092] Optionally, at least one of the following can be used: account information of the training account, training published content, pre-search media data, and training search information. This results in encoded training search features. The encoded training search features are then input into a search understanding model. Multiple search understanding expert networks within the search understanding model process the encoded training search features using search understanding tasks to obtain multiple training task feature vectors corresponding to the training search information.

[0093] S303, with the goal of reducing the difference between the feature vectors of each training task and the corresponding training label data, the search understanding model is trained to obtain the trained search understanding model; wherein, the trained search understanding model is used to generate multiple search understanding information corresponding to the target search information based on the input target search information.

[0094] Specifically, based on the feature vectors of each training task and the corresponding training label data, the task loss value corresponding to each training task feature vector is determined. The search understanding model is then trained with the goal of reducing the loss value of each task, resulting in a trained search understanding model. This trained model is used to generate multiple search understanding messages corresponding to the input target search information. The task loss value corresponding to each training task feature vector can be determined using a preset loss function, such as the cross-entropy loss function or the negative log-likelihood loss function, etc., which are not limited here.

[0095] Optionally, if training is performed on any one of multiple search understanding expert networks, the goal can be to reduce the difference between the training task feature vector and the corresponding training label data of any search understanding expert network, thereby obtaining the trained search understanding expert network. The trained search understanding expert network is then used for error correction, rewriting, entity recognition, intent recognition, or category classification.

[0096] In one optional implementation, the search understanding model further includes a gating network, and the training samples also include fused label data. The gating network can then be used to perform feature processing on the training search information to obtain the task weights corresponding to the feature vectors of each training task. Based on the multiple training task feature vectors and the task weights corresponding to each training task feature vector, a training fused feature vector corresponding to the training search information is obtained. The search understanding model is then trained with the goal of reducing the difference between the training fused feature vector and the fused label data to obtain the trained search understanding model.

[0097] In this process, the gating network determines the task weights corresponding to the feature vectors of each training task and obtains the training fusion feature vectors corresponding to the training search information. Then, with the goal of minimizing the difference between the training fusion feature vectors and the fusion label data, the fusion loss value can be determined based on these two data points. The search understanding model is then trained to minimize this fusion loss value, resulting in the trained search understanding model. The task loss value corresponding to the training fusion feature vectors can be determined using a pre-defined loss function, such as the cross-entropy loss function or the negative log-likelihood loss function, which is not limited here.

[0098] In this embodiment, training samples are obtained, including training search information and multiple label data corresponding to the training search information. Each label data corresponds to a search understanding expert network in the search understanding model. Each search understanding expert network is used for error correction, rewriting, entity recognition, intent recognition, or category classification. The training samples are input into the search understanding model, and multiple search understanding expert networks within the model process the training search information to perform search understanding tasks, resulting in multiple training task feature vectors corresponding to the training search information. Each search understanding expert network yields one training task feature vector. The search understanding model is trained with the goal of reducing the difference between each training task feature vector and its corresponding training label data, resulting in a trained search understanding model. The trained search understanding model is used to generate multiple search understanding information corresponding to the input target search information. By constructing multiple search understanding expert networks within the search understanding model and jointly training these networks, the training efficiency of the search understanding model can be improved.

[0099] Based on the above description, please refer to Figure 4 , Figure 4 This is a flowchart illustrating a data retrieval method provided in an embodiment of this application, such as... Figure 4 The data retrieval method shown includes, but is not limited to, steps S401-S404, wherein: S401, Obtain target search information for the target account.

[0100] The target account is the account that inputs the target search information, indicating a data search need. The target object corresponding to the target account can input the target search information through a client, which then sends the search information to the server. Specifically, the target account can input the target search information on the search page of an online platform (such as a content platform, e-commerce platform, live streaming platform, etc.) to obtain the target account's search information.

[0101] Optionally, the target search information may include at least one of the following: target search text information, target search image information, and target search video information. For example, the target search information is the target search text information "women's coat".

[0102] Optionally, the target search information can be segmented. Specifically, the boundaries of basic units (i.e., "words") with independent semantic or grammatical functions in the target search information can be identified, thereby converting the target search information into a normalized sequence composed of words or sub-word units, i.e., multiple tokens. These multiple tokens are the smallest text units that the trained search understanding model processes and understands.

[0103] In one alternative implementation, the account information of the target account and the target published content can also be obtained; the target published content includes the content published by the target account during a historical time period.

[0104] The target account information may include details such as age, gender, and interests. The target posted content includes content posted by the target account within a historical timeframe, such as clicks, views, likes, favorites, and comments within a 30-day period. Furthermore, the target posted content, account information, and target search information can be input together into the trained search understanding model for search understanding task processing, which can improve the accuracy of search understanding.

[0105] In this embodiment, the target published content, account information, and target search information are input together into the trained search understanding model for search understanding task processing, which can improve the accuracy of search understanding.

[0106] Optionally, the target published content can include content published by the target account within a target historical time period. The target historical time period can refer to a period of time whose time interval with the current system time is less than or equal to a preset time interval, such as the most recent period (within 7 days). In other words, it is possible to obtain the published content published by the target account within a recent period, and this published content, account information, and target search information can be input together into the trained search understanding model for search understanding task processing, thereby improving the accuracy of search understanding.

[0107] Optionally, the target content may include content associated with the target search information.

[0108] The published content associated with the target search information can refer to published content related to the search behavior corresponding to the target search information, or published content semantically associated with the target search information. Published content associated with the search behavior corresponding to the target search information can be understood as published content that generates an interactive behavior (such as clicking, browsing, or saving) after any account enters search information that is the same as or similar to the target search information; this published content is considered as published content associated with the target search information. Published content semantically associated with the target search information can be understood as follows: the search semantics of the target search information can be obtained, and the content semantics of candidate published content can be obtained. The semantic similarity between the search semantics and the content semantics of each candidate published content can be calculated, thereby determining the published content associated with the target search information from the candidate published content. Candidate published content can be obtained from a published content database. The semantic similarity between the search semantics and the content semantics of each candidate published content can be sorted from largest to smallest, and the candidate published content that meets the preset ranking (such as the top K, where K is a positive integer) can be used as the published content associated with the target search information; or, the candidate published content whose semantic similarity between the search semantics and the content semantics of each candidate published content meets the preset similarity threshold can be used as the published content associated with the target search information.

[0109] Optionally, the target content to be published may include content that is associated with the target search information and meets a preset level of interaction.

[0110] Once the published content associated with the target search information is identified, the interaction volume of each published content associated with the target search information can be obtained. The interaction volume of any published content can be determined based on the interaction operations of each account on the published content (such as likes, favorites, comments, etc.), and published content with an interaction volume greater than or equal to the preset interaction volume is selected as the target published content.

[0111] Optionally, the target posted content may include content posted by the target account during a historical time period that is associated with the target search information. For a detailed description of the posted content associated with the target search information, please refer to the relevant descriptions above, which will not be repeated here.

[0112] In one alternative implementation, pre-search processing can also be performed based on the target search information to obtain pre-search media data corresponding to the target search information.

[0113] Pre-search processing refers to the search processing performed on the target search information using pre-defined search rules or models before it is input into the trained search understanding model. Pre-search media data refers to the media data obtained after pre-search processing of the target search information. This media data includes, but is not limited to, published content, product data, live stream data, and advertising data.

[0114] Optionally, feature encoding can be performed on at least one of the target account's account information, the target published content, the pre-search media data, and the target search information to obtain encoded target search features. Furthermore, the encoded target search features can be processed using a trained search understanding model to obtain multiple search understanding pieces of information corresponding to the target search information.

[0115] Optionally, self-attention and normalization processing can be applied to the target search information to obtain a feature vector representing the context. Alternatively, self-attention and normalization processing can be applied to the encoded target search features to obtain a feature vector representing the context. Alternatively, self-attention and normalization processing can be applied to at least one of the target account information, target published content, and pre-search media data, along with the target search information, to obtain a feature vector representing the context. Furthermore, the trained search understanding model can be used to process the feature vector representing the context using a search understanding task to obtain multiple search understanding pieces of information corresponding to the target search information.

[0116] S402 uses the trained search understanding model to perform search understanding tasks on the target search information, and obtains multiple search understanding information corresponding to the target search information.

[0117] The trained search understanding model includes multiple search understanding expert networks, including an error correction expert network, a rewriting expert network, an entity recognition expert network, an intent recognition expert network, and a category classification expert network. The search understanding information includes error correction information from the error correction expert network; rewriting information from the rewriting expert network; entity recognition information from the entity recognition expert network; intent recognition information from the intent recognition expert network; and category classification information from the category classification expert network. For example, if the target search information is "spring women's outerwear," then the error correction information from the error correction expert network could be "spring women's coats"; the rewriting information from the rewriting expert network could be "which spring women's coat is best to buy?"; the entity recognition information from the entity recognition expert network could be "spring, women's clothing, coats, styles"; the intent recognition information from the intent recognition expert network could be "clothing, shopping"; and the category classification information from the category classification expert network could be "women's clothing > coats".

[0118] Optionally, the target search information can be processed by the trained search understanding model to obtain multiple target task feature vectors corresponding to the target search information; wherein, a target task feature vector is obtained through a search understanding expert network; and feature parsing is performed on each target task feature vector to obtain multiple search understanding information corresponding to the target search information.

[0119] The multiple target task feature vectors include target error correction feature vectors corresponding to the error correction expert network, target rewriting feature vectors corresponding to the rewriting expert network, target entity feature vectors corresponding to the entity recognition expert network, target intent feature vectors corresponding to the intent recognition expert network, and target category feature vectors corresponding to the category classification expert network. Feature parsing is performed on each target task feature vector to obtain multiple search understanding information corresponding to the target search information. Optionally, the trained search understanding model may include a feature parser (or decoder, post-processing module) capable of performing feature parsing on each target task feature vector to obtain multiple search understanding information corresponding to the target search information.

[0120] Optionally, the trained search understanding model also includes a gating network. The gating network can then be used to perform feature processing on the target search information to obtain the task weights corresponding to the feature vectors of each target task. Based on the multiple target task feature vectors and the task weights corresponding to each target task feature vector, the target fusion feature vector corresponding to the target search information can be obtained.

[0121] Correspondingly, feature parsing can be performed on the target fusion feature vector to obtain multiple search understanding information corresponding to the target search information.

[0122] Optionally, the parsing order of each target task feature vector can be determined; and feature parsing can be performed on each target task feature vector according to the parsing order to obtain multiple search understanding information corresponding to the target search information.

[0123] The parsing order of the feature vectors for each target task can be pre-set; for example, the parsing order of the target error correction feature vector can be set before the target rewriting feature vector, and the parsing order of the target error correction feature vector can be set before the target entity feature vector, and so on. Alternatively, the parsing order of the feature vectors for each target task can be determined based on each search and understanding task. For example, if error correction processing precedes rewriting processing, then the parsing order of the target error correction feature vector can be before the target rewriting feature vector; similarly, if rewriting processing precedes entity recognition processing, then the processing order of the target rewriting feature vector can be before the target entity feature vector.

[0124] Optionally, following the parsing order, after obtaining the search understanding information through parsing the feature vectors of the previous one or more target tasks, when parsing the feature vector of the next target task, the feature vector of the next target task can be parsed based on the feature vector of the next target task and the search understanding information obtained from parsing the feature vectors of the previous one or more target tasks, so as to obtain the search understanding information corresponding to the feature vector of the next target task.

[0125] In this embodiment of the application, when parsing the feature vector of the next target task, the search understanding information obtained by parsing the feature vector of the next target task and the feature vector of the previous one or more target tasks can be used to parse the feature vector of the next target task, so as to obtain the search understanding information corresponding to the feature vector of the next target task. This allows downstream fields to directly consume upstream results, reduce error propagation, and improve the accuracy of feature parsing.

[0126] Optionally, following the parsing order, after obtaining search understanding information through parsing the feature vectors of one or more target tasks, when parsing the feature vector of the next target task, the feature vector of the next target task can be parsed based on the feature vector of the next target task and the search understanding information obtained from parsing the feature vectors of one or more target tasks, to obtain the search understanding information corresponding to the feature vector of the next target task; and based on the preset task elements and the parsed search understanding information, structured search understanding information can be obtained.

[0127] The preset task elements are used to construct structured search understanding information with the parsed search understanding information. Each preset task element corresponds to a feature vector of a specific target task. Preset task elements can be, for example, preset task characters or preset task identifiers. For instance, the preset task element corresponding to the target error correction feature vector is "[CORRECT]", the target rewrite feature vector is "[REWRITE]", the target entity feature vector is "[ENTITIES]", the target intent feature vector is "[INTENT]", and the target category feature vector is "[TAXONOMY]". If the target search information is "spring women's coats", then based on the preset task elements and the parsed search understanding information, the structured search understanding information obtained could be: [CORRECT]Spring Women's Coats[ / CORRECT] [REWRITE]Which Spring Women's Coats to Buy in 2026? [ / REWRITE] [ENTITIES]type=Spring,text=Women's Clothing,type=Coats,text=Styles[ / ENTITIES] [INTENT]Theme=Shopping,Scene=Clothing[ / INTENT] [TAXONOMY]Women's Clothing>Coats[ / TAXONOMY].

[0128] In one optional implementation, the search understanding model includes multiple search understanding expert networks, including an entity recognition expert network, an intent recognition expert network, and a category classification expert network; multiple search understanding information includes entity recognition information and entity confidence scores corresponding to the entity recognition expert network, intent recognition information and intent confidence scores corresponding to the intent recognition expert network, and category classification information and category confidence scores corresponding to the category classification expert network; then, if any confidence score is less than or equal to a confidence score threshold, the target search information is processed by the preset search understanding model to obtain multiple search understanding information corresponding to the target search information.

[0129] The search understanding information includes both generation task information and classification task information. Generation task information includes error correction and rewriting information; classification task information can include entity recognition information, intent recognition information, and category classification information. Specifically, the entity recognition expert network, intent recognition expert network, and category classification expert network, when determining the corresponding target task feature vectors, also determine the confidence level of the corresponding target task feature vectors. Furthermore, when performing feature parsing on the target entity feature vectors, target intent feature vectors, and target category feature vectors, they will obtain the entity recognition information and entity confidence level corresponding to the entity recognition expert network, the intent recognition information and intent confidence level corresponding to the intent recognition expert network, and the category classification information and category confidence level corresponding to the category classification expert network. Each confidence level is used to characterize the credibility of the corresponding search understanding information. If any confidence level is less than or equal to a confidence threshold (e.g., any confidence level p=0.4, confidence threshold 0.6), or any confidence level p=0.4 less than confidence threshold 0.6), then the target search information will be processed using a pre-defined search understanding model to obtain multiple search understanding messages corresponding to the target search information. This means the old model (pre-defined search understanding model) is rolled back to process the target search information, ensuring that even with low confidence levels, the pre-defined search understanding model provides a fallback, guaranteeing a better data retrieval experience. Optionally, the pre-defined search understanding model can be a pre-trained artificial intelligence (AI) model or neural network model. This model can be obtained through supervised or unsupervised training. This application does not limit the model structure or training method of the pre-defined search understanding model.

[0130] Optionally, if any confidence level is less than or equal to the confidence threshold, a search understanding task can be performed on the target search information using preset search rules to obtain multiple search understanding information corresponding to the target search information. The preset search rules are pre-defined rules for performing search understanding tasks on the search information. That is, when any confidence level is less than or equal to the confidence threshold, the old rule (preset search rule) can be rolled back to perform search understanding tasks on the target search information. This ensures that even with low confidence levels, the preset search rules can provide a fallback, guaranteeing a better data retrieval experience.

[0131] In one optional implementation, the processing order of each search understanding expert network is determined; and according to the processing order, the target task feature vector and target search information obtained by the previous one or more search understanding expert networks are input into the next search understanding expert network. The next search understanding expert network then performs search understanding task processing on the target task feature vector and target search information corresponding to the previous one or more search understanding expert networks, resulting in the target task feature vector corresponding to the next search understanding expert network. Feature parsing is then performed on each target task feature vector to obtain multiple search understanding information corresponding to the target search information.

[0132] For detailed implementation methods of the embodiments of this application, please refer to the above-mentioned descriptions, which will not be repeated here.

[0133] In one optional implementation, the multiple search understanding expert networks include an error correction expert network, a rewriting expert network, an entity recognition expert network, an intent recognition expert network, and a category classification expert network. The error correction expert network can then perform error correction processing on the target search information to obtain a target error-corrected feature vector. The rewriting expert network can rewrite the target search information and the target error-corrected feature vector to obtain a target rewritten feature vector. The entity recognition expert network can perform entity recognition processing on the target search information and the target rewritten feature vector to obtain a target entity feature vector. The intent recognition expert network can perform intent recognition processing on the target search information and the target entity feature vector to obtain a target intent feature vector. The category classification expert network can perform category classification processing on the target search information and the target intent feature vector to obtain a target category feature vector. Feature parsing is then performed on each target task feature vector to obtain multiple search understanding information corresponding to the target search information.

[0134] For detailed implementation methods of the embodiments of this application, please refer to the above-mentioned descriptions, which will not be repeated here.

[0135] In one optional implementation, the multiple search understanding expert networks include an error correction expert network, a rewriting expert network, an entity recognition expert network, an intent recognition expert network, and a category classification expert network. The error correction expert network corrects the target search information to obtain a target error-corrected feature vector; the rewriting expert network rewrites the target search information and the target error-corrected feature vector to obtain a target rewritten feature vector; the entity recognition expert network performs entity recognition on the target search information, the target error-corrected feature vector, and the target rewritten feature vector to obtain a target entity feature vector; the intent recognition expert network performs intent recognition on the target search information, the target error-corrected feature vector, the target rewritten feature vector, and the target entity feature vector to obtain a target intent feature vector; and the category classification expert network performs category classification on the target search information, the target error-corrected feature vector, the target rewritten feature vector, the target entity feature vector, and the target intent feature vector to obtain a target category feature vector. Feature parsing is then performed on each target task feature vector to obtain multiple search understanding information corresponding to the target search information.

[0136] For detailed implementation methods of the embodiments of this application, please refer to the above-mentioned descriptions, which will not be repeated here.

[0137] In one optional implementation, the search understanding model further includes a gating network. The gating network can then be used to perform feature processing on the target search information to obtain the task weights corresponding to each target task feature vector. Based on multiple target task feature vectors and their corresponding task weights, a target fusion feature vector corresponding to the target search information is obtained. Feature parsing is then performed on the target fusion feature vector to obtain multiple search understanding pieces of information corresponding to the target search information.

[0138] For detailed implementation methods of the embodiments of this application, please refer to the above-mentioned descriptions, which will not be repeated here.

[0139] In one optional implementation, the account information of the target account and the target published content can also be obtained; the target published content includes the content published by the target account during a historical time period; the target sample, the account information of the target account, and the target published content are input into the search understanding model, and multiple search understanding expert networks in the search understanding model are used to process the target search information, the account information of the target account, and the target published content through search understanding tasks to obtain multiple target task feature vectors corresponding to the target search information. Feature parsing is performed on each target task feature vector to obtain multiple search understanding information corresponding to the target search information.

[0140] For detailed implementation methods of the embodiments of this application, please refer to the above-mentioned descriptions, which will not be repeated here.

[0141] In one optional implementation, pre-search processing can be performed based on the target search information to obtain pre-search media data corresponding to the target search information. The target sample and the pre-search media data are then input into a search understanding model, and multiple search understanding expert networks within the model perform search understanding task processing on the target search information and the pre-search media data to obtain multiple target task feature vectors corresponding to the target search information. Feature parsing is then performed on each target task feature vector to obtain multiple search understanding information corresponding to the target search information.

[0142] For detailed implementation methods of the embodiments of this application, please refer to the above-mentioned descriptions, which will not be repeated here.

[0143] S403, match multiple search understanding information with candidate media data to obtain the matching degree between multiple search understanding information and candidate media data.

[0144] Candidate media data includes, but is not limited to, published content, products, live streams, and advertising data. The server can calculate the matching degree (or similarity) between multiple search understanding pieces of information and candidate media data, such as by calculating the cosine similarity, Euclidean distance, or Pearson correlation coefficient between the search understanding pieces and candidate media data. A higher matching degree indicates a higher relevance between the multiple search understanding pieces of information and any candidate media data, and vice versa. In other words, a higher matching degree means that the media data is more likely to be what the target account wants to search for.

[0145] Optionally, the matching degree between multiple search understanding information and any candidate media data can be determined based on the matching degree between each search understanding information and any candidate media data, as well as the search information weight corresponding to each search understanding information.

[0146] In this way, the matching degree between multiple search understanding information and any candidate media data can be determined by weighted fusion based on the matching degree between each search understanding information and any candidate media data, as well as the search information weight corresponding to each search understanding information.

[0147] S404. Based on the matching degree between multiple search understanding information and candidate media data, target media data is determined and used as target recall data.

[0148] The server can determine the target media data based on the matching degree between multiple search understanding information and candidate media data. The target media data is used as the target recall data, and the server sends the target recall data to the client corresponding to the target account for rendering and display.

[0149] Optionally, the matching degree between multiple search understanding information and candidate media data can be sorted from largest to smallest, and the candidate media data that meets the preset ranking (such as the top M, where M is a positive integer) can be used as the target media data; or, the candidate media data whose matching degree between multiple search understanding information and candidate media data meets the preset matching degree threshold can be used as the target media data.

[0150] In this embodiment, target search information of a target account is obtained; a trained search understanding model is used to process the target search information to obtain multiple search understanding information corresponding to the target search information; these multiple search understanding information are matched with candidate media data to obtain the matching degree between the multiple search understanding information and the candidate media data; based on the matching degree between the multiple search understanding information and the candidate media data, target media data is determined, and the target media data is used as target recall data. Processing the target search information using a trained search understanding model to obtain multiple search understanding information corresponding to the target search information, determining the corresponding target media data based on these multiple search understanding information, and using the target media data as target recall data can improve the accuracy of the recalled data.

[0151] This application also provides a computer storage medium storing program instructions, which, when executed, are used to implement the corresponding methods described in the above embodiments.

[0152] This application provides a computer program product, which includes a computer program stored in a computer storage medium. The processor of a computer device reads the computer program from the computer storage medium and executes the computer program, causing the computer device to perform the corresponding methods described in the above embodiments.

[0153] See also Figure 5 , Figure 5 This is a schematic diagram of the structure of a model training device provided in an embodiment of this application.

[0154] In one implementation of the model training device according to the embodiments of this application, the model training device includes the following structure.

[0155] The acquisition unit 501 is used to acquire training samples, which include training search information and multiple label data corresponding to the training search information; wherein, a label data corresponds to a search understanding expert network in the search understanding model, and any search understanding expert network is used to perform error correction processing, rewriting processing, entity recognition processing, intent recognition processing or category classification processing. Processing unit 502 is used to input the training samples into the search understanding model, and to process the training search information into search understanding tasks through multiple search understanding expert networks in the search understanding model, so as to obtain multiple training task feature vectors corresponding to the training search information; wherein, one training task feature vector is obtained through one search understanding expert network. Training unit 503 is used to train the search understanding model with the goal of reducing the difference between the feature vectors of each training task and the corresponding training label data, so as to obtain the trained search understanding model; wherein, the trained search understanding model is used to generate multiple search understanding information corresponding to the target search information based on the input target search information.

[0156] In one embodiment, the processing unit 502 performs search understanding task processing on the training search information through multiple search understanding expert networks in the search understanding model, obtaining multiple training task feature vectors corresponding to the training search information, which can be used for: Determine the processing order of each search understanding expert network; According to the processing order, the training task feature vector obtained by the previous one or more search understanding expert networks and the training search information are input into the next search understanding expert network. The next search understanding expert network then performs search understanding task processing on the training task feature vector and the training search information corresponding to the previous one or more search understanding expert networks to obtain the training task feature vector corresponding to the next search understanding expert network.

[0157] In one implementation, the plurality of search understanding expert networks include an error correction expert network, a rewriting expert network, an entity recognition expert network, an intent recognition expert network, and a category classification expert network. The processing unit 502, according to the processing order, inputs the training task feature vector obtained through one or more previous search understanding expert networks and the training search information into a subsequent search understanding expert network, and performs search understanding task processing on the training task feature vector and the training search information corresponding to the previous one or more search understanding expert networks through the subsequent search understanding expert network, to obtain the training task feature vector corresponding to the subsequent search understanding expert network, which can be used for: The training search information is corrected by the error correction expert network to obtain the training error correction feature vector. The training search information and training error correction feature vector are rewritten by the rewriting expert network to obtain the training rewritten feature vector. The entity recognition expert network performs entity recognition processing on the training search information, training error correction feature vector, and training rewritten feature vector to obtain training entity feature vector. The intent recognition expert network performs intent recognition processing on the training search information, training error correction feature vector, training rewrite feature vector, and training entity feature vector to obtain a training intent feature vector. The category classification expert network performs category classification processing on the training search information, training error correction feature vector, training rewrite feature vector, training entity feature vector, and training intent feature vector to obtain the training category feature vector.

[0158] In one implementation, the search understanding model further includes a gating network, and the training samples further include fused labeled data; the processing unit 502 can also be used for: The gating network is used to perform feature processing on the training search information to obtain the task weights corresponding to the feature vectors of each training task. Based on the multiple training task feature vectors and the task weights corresponding to each training task feature vector, the training fusion feature vector corresponding to the training search information is obtained. The training unit 503 trains the search understanding model with the goal of reducing the difference between the feature vectors of each training task and the corresponding training label data, resulting in a trained search understanding model that can be used for: The search understanding model is trained with the goal of reducing the difference between the training fusion feature vector and the fusion label data, resulting in the trained search understanding model.

[0159] In one embodiment, the acquisition unit 501 can also be used for: Obtain the account information of the training account and the training published content; the training published content includes the content published by the training account during a historical time period; The processing unit 502 inputs the training samples into the search understanding model, and performs search understanding task processing on the training search information through multiple search understanding expert networks in the search understanding model, obtaining multiple training task feature vectors corresponding to the training search information, which can be used for: The training samples, the account information of the training account, and the training published content are input into the search understanding model. The search understanding model uses multiple search understanding expert networks to process the training search information, the account information of the training account, and the training published content to obtain multiple training task feature vectors corresponding to the training search information.

[0160] In one embodiment, the processing unit 502 can also be used for: Based on the training search information, pre-search processing is performed to obtain the pre-search media data corresponding to the training search information; The processing unit 502 inputs the training samples into the search understanding model, and performs search understanding task processing on the training search information through multiple search understanding expert networks in the search understanding model, obtaining multiple training task feature vectors corresponding to the training search information, which can be used for: The training samples and the pre-search media data are input into the search understanding model, and the training search information and the pre-search media data are processed by multiple search understanding expert networks in the search understanding model to obtain multiple training task feature vectors corresponding to the training search information.

[0161] In this embodiment, the acquisition unit 501 acquires training samples, which include training search information and multiple label data corresponding to the training search information. Each label data corresponds to a search understanding expert network in the search understanding model. Each search understanding expert network is used for error correction, rewriting, entity recognition, intent recognition, or category classification. The processing unit 502 inputs the training samples into the search understanding model and processes the training search information using multiple search understanding expert networks within the model to obtain multiple training task feature vectors corresponding to the training search information. Each search understanding expert network is used to obtain one training task feature vector. The training unit 503 trains the search understanding model with the goal of reducing the difference between each training task feature vector and the corresponding training label data, resulting in a trained search understanding model. The trained search understanding model is used to generate multiple search understanding information corresponding to the input target search information. By constructing multiple search understanding expert networks within the search understanding model and jointly training these networks, the training efficiency of the search understanding model can be improved.

[0162] See also Figure 6 , Figure 6 This is a schematic diagram of the structure of a data retrieval device provided in an embodiment of this application.

[0163] In one implementation of the data retrieval device according to the embodiments of this application, the data retrieval device includes the following structure.

[0164] Acquisition unit 601 is used to acquire target search information for the target account; Processing unit 602 is used to perform search understanding task processing on the target search information through the trained search understanding model to obtain multiple search understanding information corresponding to the target search information; The matching unit 603 is used to match the plurality of search understanding information with candidate media data to obtain the matching degree between the plurality of search understanding information and candidate media data; The matching unit 603 is further configured to determine target media data based on the matching degree between the plurality of search understanding information and the candidate media data, and the target media data is used as target recall data.

[0165] In one embodiment, the search understanding model includes multiple search understanding expert networks, which include an entity recognition expert network, an intent recognition expert network, and a category classification expert network; the multiple search understanding information includes entity recognition information and entity confidence corresponding to the entity recognition expert network, intent recognition information and intent confidence corresponding to the intent recognition expert network, and category classification information and category confidence corresponding to the category classification expert network. The processing unit 602 can also be used for: If any confidence level is less than or equal to the confidence threshold, then the target search information is processed by a preset search understanding model to obtain multiple search understanding information corresponding to the target search information.

[0166] In this embodiment, the acquisition unit 601 acquires the target search information of the target account; the processing unit 602 performs search understanding task processing on the target search information using a trained search understanding model to obtain multiple search understanding information corresponding to the target search information; the matching unit 603 matches the multiple search understanding information with candidate media data to obtain the matching degree between the multiple search understanding information and the candidate media data; based on the matching degree between the multiple search understanding information and the candidate media data, the target media data is determined and used as target recall data. By performing search understanding task processing on the target search information using a trained search understanding model to obtain multiple search understanding information corresponding to the target search information, and determining the corresponding target media data based on the multiple search understanding information, and using the target media data as target recall data, the accuracy of the recall data can be improved.

[0167] See also Figure 7 , Figure 7This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. The computer device in this embodiment includes a power supply module and other structures, and includes a processor 701, a memory 702, and a communication interface 703. The processor 701, the memory 702, and the communication interface 703 can exchange data, and the processor 701 can implement corresponding model training or data retrieval methods.

[0168] The memory 702 may include volatile memory, such as random-access memory (RAM); the memory 702 may also include non-volatile memory, such as flash memory, solid-state drive (SSD), etc.; the memory 702 may also include a combination of the above types of memory.

[0169] Processor 701 may be a central processing unit (CPU). Processor 701 may also be a combination of CPU and GPU. In a computer device, multiple CPUs and GPUs may be included as needed for corresponding model training or data retrieval. In one embodiment, memory 702 is used to store program instructions. Processor 701 can invoke program instructions to implement the various methods described above in the embodiments of this application.

[0170] The communication interface 703 may include a display screen, microphone, or speaker, etc.

[0171] In one possible implementation, the processor 701 of the computer device calls program instructions stored in the memory 702 to acquire training samples, the training samples including training search information and multiple label data corresponding to the training search information; wherein, each label data corresponds to a search understanding expert network in the search understanding model, and any search understanding expert network is used for error correction processing, rewriting processing, entity recognition processing, intent recognition processing, or category classification processing; the training samples are input into the search understanding model, and the multiple search understanding expert networks in the search understanding model perform search understanding task processing on the training search information to obtain multiple training task feature vectors corresponding to the training search information; wherein, one training task feature vector is obtained through one search understanding expert network; the search understanding model is trained with the goal of reducing the difference between each training task feature vector and the corresponding training label data to obtain a trained search understanding model; wherein, the trained search understanding model is used to generate multiple search understanding information corresponding to the input target search information.

[0172] In one embodiment, the processor 701 processes the training search information through multiple search understanding expert networks in the search understanding model to obtain multiple training task feature vectors corresponding to the training search information, and can perform the following operations: Determine the processing order of each search understanding expert network; According to the processing order, the training task feature vector obtained by the previous one or more search understanding expert networks and the training search information are input into the next search understanding expert network. The next search understanding expert network then performs search understanding task processing on the training task feature vector and the training search information corresponding to the previous one or more search understanding expert networks to obtain the training task feature vector corresponding to the next search understanding expert network.

[0173] In one implementation, the plurality of search understanding expert networks include an error correction expert network, a rewriting expert network, an entity recognition expert network, an intent recognition expert network, and a category classification expert network. The processor 701, following the processing order, inputs the training task feature vector obtained through one or more previous search understanding expert networks and the training search information into a subsequent search understanding expert network. The subsequent search understanding expert network then performs search understanding task processing on the training task feature vector and the training search information corresponding to the previous one or more search understanding expert networks to obtain the training task feature vector corresponding to the subsequent search understanding expert network. The processor 701 can perform the following operations: The training search information is corrected by the error correction expert network to obtain the training error correction feature vector. The training search information and training error correction feature vector are rewritten by the rewriting expert network to obtain the training rewritten feature vector. The entity recognition expert network performs entity recognition processing on the training search information, training error correction feature vector, and training rewritten feature vector to obtain training entity feature vector. The intent recognition expert network performs intent recognition processing on the training search information, training error correction feature vector, training rewrite feature vector, and training entity feature vector to obtain a training intent feature vector. The category classification expert network performs category classification processing on the training search information, training error correction feature vector, training rewrite feature vector, training entity feature vector, and training intent feature vector to obtain the training category feature vector.

[0174] In one implementation, the search understanding model further includes a gating network, and the training samples further include fused labeled data; the processor 701 can also perform the following operations: The gating network is used to perform feature processing on the training search information to obtain the task weights corresponding to the feature vectors of each training task. Based on the multiple training task feature vectors and the task weights corresponding to each training task feature vector, the training fusion feature vector corresponding to the training search information is obtained. The processor 701 trains the search understanding model with the goal of reducing the difference between the feature vectors of each training task and the corresponding training label data, resulting in a trained search understanding model that can perform the following operations: The search understanding model is trained with the goal of reducing the difference between the training fusion feature vector and the fusion label data, resulting in the trained search understanding model.

[0175] In one embodiment, the processor 701 may also perform the following operations: Obtain the account information of the training account and the training published content; the training published content includes the content published by the training account during a historical time period; The processor 701 inputs the training samples into the search understanding model, and processes the training search information through multiple search understanding expert networks in the search understanding model to obtain multiple training task feature vectors corresponding to the training search information. The processor can perform the following operations: The training samples, the account information of the training account, and the training published content are input into the search understanding model. The search understanding model uses multiple search understanding expert networks to process the training search information, the account information of the training account, and the training published content to obtain multiple training task feature vectors corresponding to the training search information.

[0176] In one embodiment, the processor 701 may also perform the following operations: Based on the training search information, pre-search processing is performed to obtain the pre-search media data corresponding to the training search information; The processor 701 inputs the training samples into the search understanding model, and processes the training search information through multiple search understanding expert networks in the search understanding model to obtain multiple training task feature vectors corresponding to the training search information. The processor can perform the following operations: The training samples and the pre-search media data are input into the search understanding model, and the training search information and the pre-search media data are processed by multiple search understanding expert networks in the search understanding model to obtain multiple training task feature vectors corresponding to the training search information.

[0177] In this embodiment, the processor 701 acquires training samples, which include training search information and multiple label data corresponding to the training search information. Each label data corresponds to a search understanding expert network in the search understanding model. Each search understanding expert network is used for error correction, rewriting, entity recognition, intent recognition, or category classification. The training samples are input into the search understanding model, and the multiple search understanding expert networks in the model process the training search information to obtain multiple training task feature vectors corresponding to the training search information. Each search understanding expert network generates one training task feature vector. The search understanding model is trained with the goal of reducing the difference between each training task feature vector and the corresponding training label data, resulting in a trained search understanding model. The trained search understanding model is used to generate multiple search understanding information corresponding to the input target search information. By constructing multiple search understanding expert networks in the search understanding model and jointly training these networks, the training efficiency of the search understanding model can be improved.

[0178] In another possible implementation, the processor 701 of the computer device calls program instructions stored in the memory 702 to obtain target search information of the target account; performs search understanding task processing on the target search information through the trained search understanding model to obtain multiple search understanding information corresponding to the target search information; matches the multiple search understanding information with candidate media data to obtain the matching degree between the multiple search understanding information and the candidate media data; and determines the target media data based on the matching degree between the multiple search understanding information and the candidate media data, and the target media data is used as target recall data.

[0179] In one embodiment, the search understanding model includes multiple search understanding expert networks, which include an entity recognition expert network, an intent recognition expert network, and a category classification expert network; the multiple search understanding information includes entity recognition information and entity confidence corresponding to the entity recognition expert network, intent recognition information and intent confidence corresponding to the intent recognition expert network, and category classification information and category confidence corresponding to the category classification expert network. The processor 701 can also perform the following operations: If any confidence level is less than or equal to the confidence threshold, then the target search information is processed by a preset search understanding model to obtain multiple search understanding information corresponding to the target search information.

[0180] In this embodiment, processor 701 acquires target search information of a target account; processes the target search information using a trained search understanding model to obtain multiple search understanding information corresponding to the target search information; matches the multiple search understanding information with candidate media data to obtain the matching degree between the multiple search understanding information and the candidate media data; and determines the target media data based on the matching degree between the multiple search understanding information and the candidate media data, using the target media data as target recall data. Processing the target search information using a trained search understanding model to obtain multiple search understanding information corresponding to the target search information, determining the corresponding target media data based on the multiple search understanding information, and using the target media data as target recall data can improve the accuracy of the recall data.

[0181] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM or random access memory (RAM), magnetic disks, or optical disks.

[0182] The above-disclosed embodiments are merely some of the embodiments of this application, and should not be construed as limiting the scope of this application. Those skilled in the art can understand that all or part of the processes for implementing the above embodiments, and equivalent changes made in accordance with the claims of this application, still fall within the scope of this application.

Claims

1. A model training method, characterized in that, The method includes: Acquire training samples, which include training search information and multiple label data corresponding to the training search information; wherein, each label data corresponds to a search understanding expert network in the search understanding model, and any search understanding expert network is used for error correction, rewriting, entity recognition, intent recognition, or category classification. The training samples are input into the search understanding model, and the training search information is processed by multiple search understanding expert networks in the search understanding model to obtain multiple training task feature vectors corresponding to the training search information; wherein, one training task feature vector is obtained through one search understanding expert network. The search understanding model is trained with the goal of reducing the difference between the feature vectors of each training task and the corresponding training label data, resulting in a trained search understanding model. The trained search understanding model is used to generate multiple search understanding information corresponding to the target search information based on the input target search information.

2. The method as described in claim 1, characterized in that, The process of performing search understanding task processing on the training search information through multiple search understanding expert networks in the search understanding model to obtain multiple training task feature vectors corresponding to the training search information includes: Determine the processing order of each search understanding expert network; According to the processing order, the training task feature vector obtained by the previous one or more search understanding expert networks and the training search information are input into the next search understanding expert network. The next search understanding expert network then performs search understanding task processing on the training task feature vector and the training search information corresponding to the previous one or more search understanding expert networks to obtain the training task feature vector corresponding to the next search understanding expert network.

3. The method as described in claim 2, characterized in that, The multiple search understanding expert networks include an error correction expert network, a rewriting expert network, an entity recognition expert network, an intent recognition expert network, and a category classification expert network. The process involves, according to the processing order, inputting the training task feature vector obtained through one or more previous search understanding expert networks and the training search information into a subsequent search understanding expert network. The subsequent search understanding expert network then performs search understanding task processing on the training task feature vector corresponding to the previous one or more search understanding expert networks and the training search information to obtain the training task feature vector corresponding to the subsequent search understanding expert network, including: The training search information is corrected by the error correction expert network to obtain the training error correction feature vector. The training search information and training error correction feature vector are rewritten by the rewriting expert network to obtain the training rewritten feature vector. The entity recognition expert network performs entity recognition processing on the training search information, training error correction feature vector, and training rewritten feature vector to obtain training entity feature vector. The intent recognition expert network performs intent recognition processing on the training search information, training error correction feature vector, training rewrite feature vector, and training entity feature vector to obtain a training intent feature vector. The category classification expert network performs category classification processing on the training search information, training error correction feature vector, training rewrite feature vector, training entity feature vector, and training intent feature vector to obtain the training category feature vector.

4. The method as described in claim 1, characterized in that, The search understanding model further includes a gating network, and the training samples further include fused labeled data; the method further includes: The gating network is used to perform feature processing on the training search information to obtain the task weights corresponding to the feature vectors of each training task. Based on the multiple training task feature vectors and the task weights corresponding to each training task feature vector, the training fusion feature vector corresponding to the training search information is obtained. The search understanding model is trained with the goal of reducing the difference between the feature vectors of each training task and the corresponding training label data, resulting in a trained search understanding model, including: The search understanding model is trained with the goal of reducing the difference between the training fusion feature vector and the fusion label data, resulting in the trained search understanding model.

5. The method as described in claim 1, characterized in that, The method further includes: Obtain the account information of the training account and the training published content; the training published content includes the content published by the training account during a historical time period; The step involves inputting the training samples into the search understanding model, and then using multiple search understanding expert networks within the search understanding model to process the training search information into search understanding tasks, thereby obtaining multiple training task feature vectors corresponding to the training search information, including: The training samples, the account information of the training account, and the training published content are input into the search understanding model. The search understanding model uses multiple search understanding expert networks to process the training search information, the account information of the training account, and the training published content to obtain multiple training task feature vectors corresponding to the training search information.

6. The method as described in claim 1, characterized in that, The method further includes: Based on the training search information, pre-search processing is performed to obtain the pre-search media data corresponding to the training search information; The step involves inputting the training samples into the search understanding model, and then using multiple search understanding expert networks within the search understanding model to process the training search information into search understanding tasks, thereby obtaining multiple training task feature vectors corresponding to the training search information, including: The training samples and the pre-search media data are input into the search understanding model, and the training search information and the pre-search media data are processed by multiple search understanding expert networks in the search understanding model to obtain multiple training task feature vectors corresponding to the training search information.

7. A data retrieval method, characterized in that, The method includes: Obtain the target search information for the target account; The trained search understanding model is used to process the target search information to obtain multiple search understanding information corresponding to the target search information; The multiple search understanding information is matched with the candidate media data to obtain the matching degree between the multiple search understanding information and the candidate media data; Based on the matching degree between the multiple search understanding information and the candidate media data, target media data is determined, and the target media data is used as target recall data.

8. The method as described in claim 7, characterized in that, The search understanding model includes multiple search understanding expert networks, including an entity recognition expert network, an intent recognition expert network, and a category classification expert network; the multiple search understanding information includes entity recognition information and entity confidence corresponding to the entity recognition expert network, intent recognition information and intent confidence corresponding to the intent recognition expert network, and category classification information and category confidence corresponding to the category classification expert network. The method further includes: If any confidence level is less than or equal to the confidence threshold, then the target search information is processed by a preset search understanding model to obtain multiple search understanding information corresponding to the target search information.

9. A model training device, characterized in that, The device includes: The acquisition unit is used to acquire training samples, which include training search information and multiple label data corresponding to the training search information; wherein, each label data corresponds to a search understanding expert network in the search understanding model, and any search understanding expert network is used to perform error correction processing, rewriting processing, entity recognition processing, intent recognition processing, or category classification processing. The processing unit is configured to input the training samples into the search understanding model, and process the training search information for search understanding tasks through multiple search understanding expert networks in the search understanding model to obtain multiple training task feature vectors corresponding to the training search information; wherein, one training task feature vector is obtained through one search understanding expert network. The training unit is used to train the search understanding model with the goal of reducing the difference between the feature vectors of each training task and the corresponding training label data, so as to obtain the trained search understanding model; wherein, the trained search understanding model is used to generate multiple search understanding information corresponding to the target search information based on the input target search information.

10. A data retrieval device, characterized in that, The device includes: The acquisition unit is used to acquire target search information for the target account. The processing unit is used to perform search understanding task processing on the target search information through the trained search understanding model to obtain multiple search understanding information corresponding to the target search information. A matching unit is used to match the plurality of search understanding information with candidate media data to obtain the matching degree between the plurality of search understanding information and candidate media data; The matching unit is further configured to determine target media data based on the matching degree between the plurality of search understanding information and the candidate media data, and the target media data is used as target recall data.

11. A computer device, characterized in that, The computer device includes a memory, a communication interface, and a processor, wherein the memory, the communication interface, and the processor are interconnected; the memory stores a computer program, and the processor calls the computer program stored in the memory to implement the method according to any one of claims 1 to 8.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 8.

13. A computer program product, characterized in that, The computer program product includes a computer program stored in a computer storage medium; a processor of a computer device reads the computer program from the computer storage medium and executes the computer program, causing the computer device to perform the method as described in any one of claims 1 to 8.

Images