A method and device for training and retrieving a recall model, and an electronic device

Abstract

The application discloses a method and device for training and retrieving a recall model and electronic equipment. The method comprises the following steps: inputting search side information of a training sample as first input information of a recall model to be trained into a first subnetwork of the recall model to obtain first output information, and determining search term features of the training sample according to the first output information; inputting object side information of the training sample as second input information of the recall model into a second subnetwork of the recall model to obtain second output information, and determining recall object features of the training sample according to the second output information; training the recall model based on a first loss function, and obtaining a target recall model after the training; the first loss function is constructed based on the similarity between the search term features of the training sample and corresponding recall object features, and is used as an objective function for training the recall model. By using the method, the problem of low precision of search recall is solved.

Description

Technical Field

[0001] This application relates to the field of search technology, specifically to a training method, apparatus, electronic device, and storage medium for a recall model. This application also relates to a search recall method, apparatus, electronic device, and storage medium. This application further relates to a retrieval method, apparatus, electronic device, and storage medium. Background Technology

[0002] As more and more people use the internet for searching, comprehensive platforms offering search functionality often provide a variety of industry products and / or information and connect to search requests from multiple sources, leading to increasingly diversified search needs. This places higher demands on search accuracy. Search recall is the foundation of search; it involves retrieving information related to user-inputted search terms from massive amounts of data. The depth and quality of search recall are key factors determining search accuracy; therefore, ensuring the depth and quality of search recall is particularly important.

[0003] In existing technologies, search retrieval can be performed through keyword matching, knowledge / tag matching, or by using machine learning models to determine the similarity between search terms and candidate information. However, the diversification of search needs in reality makes it difficult to guarantee the depth and quality of existing search retrieval methods. In particular, in some search scenarios, the candidate set for search retrieval is constrained by factors such as spatiotemporal attributes. In these scenarios, the candidate set obtained from search retrieval changes dynamically with these constraints, such as the user's search location and time. Existing search retrieval processes do not incorporate the impact of these constraints on search results, making it difficult to guarantee the depth and quality of retrieval, and thus resulting in insufficient accuracy.

[0004] Therefore, improving the accuracy of search retrieval is a problem that needs to be solved.

[0005] The information disclosed in the background section is only intended to enhance the understanding of the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The training method for the recall model provided in this application embodiment solves the problem of low accuracy in search and recall.

[0007] This application provides a method for training a recall model, comprising: using search-side information of training samples as first input information of the recall model to be trained, inputting it into a first sub-network of the recall model to obtain first output information, and determining the search term features of the training samples based on the first output information; using object-side information of training samples as second input information of the recall model, inputting it into a second sub-network of the recall model to obtain second output information, and determining the recall object features of the training samples based on the second output information; training the recall model based on a first loss function, and obtaining a target recall model after training; wherein, the first loss function is constructed based on the similarity between the search term features of the training samples and the corresponding recall object features, and is used as the objective function for training the recall model.

[0008] Optionally, it further includes: obtaining neighbor information of search terms in the training samples; obtaining search term neighbor features of the search terms based on the neighbor information; determining graph attention features related to the search terms and the neighbor information based on the search term neighbor features and the first output information; determining a second loss function based on the graph attention features and the search term features; and using the second loss function to train the recall model.

[0009] Optionally, the recall model further includes a third sub-network, which is a sub-network on the search side used to obtain search term neighbor features based on the neighbor information of the search terms in the training samples; obtaining the search term neighbor features based on the neighbor information includes: inputting the neighbor information into the third sub-network to obtain the search term neighbor features.

[0010] Optionally, the third sub-network has the same network structure and / or shares parameters with the first sub-network.

[0011] Optionally, training the recall model based on the first loss function includes: using the first loss function and the second loss function as objective functions for training the recall model, and training the recall model.

[0012] Optionally, obtaining the neighbor information of the search terms in the training samples includes: obtaining search logs; generating a click relationship graph based on the search logs to represent the relationship between the search terms and the recall objects of the search terms, wherein each node in the click relationship graph is a search term node corresponding to the search term information, and a click object node corresponding to each search term that has been clicked by the user; generating a bipartite graph based on neighbor aggregation of each search term node in the click relationship graph, and using the generated bipartite graph as the search click graph; sampling the neighbor nodes related to the search terms in the search click graph according to the search terms, and using the sampled data as the neighbor information.

[0013] Optionally, determining the search term features of the training sample based on the first output information includes: inputting the first output information into a first multilayer perceptron to obtain the search term features, wherein the search term features contain semantic information of the search term.

[0014] Optionally, it further includes: acquiring click data of the recall objects corresponding to the search terms of the training samples; wherein, the click data includes one or more search terms for the recall objects whose click rate meets a click rate threshold or whose click quantity meets a click quantity threshold; and generating the second input information based on the click data and the object information of the recall objects.

[0015] Optionally, determining the recall object features of the training samples based on the second output information includes: inputting the second output information into a second multilayer perceptron to obtain the recall object features, wherein the recall object features contain semantic information of the recall object corresponding to the search term.

[0016] Optionally, the first sub-network and the second sub-network have the same network structure and / or share parameters.

[0017] Optionally, it may also include: constructing the first loss function based on the similarity between the search term features of the training samples and the corresponding recall object features; the first loss function adopts one of the following loss functions: infoNCE loss function, Hinge loss function, Triplet loss function, Circle loss function.

[0018] Optionally, it also includes: obtaining the search terms input by the user on the search side and the recall objects corresponding to the search terms from the search logs, determining the recall objects that have been exposed and clicked in the recall objects; constructing positive samples based on the search terms and the exposed and clicked recall objects; generating negative samples based on the positive samples, and using the positive samples and / or the negative samples as the training samples.

[0019] This application embodiment also provides a search recall method, including: acquiring a search term input by a user, inputting the search term into a first sub-network of a target recall model to obtain a search term representation vector corresponding to the search term; wherein, the target recall model is a recall model trained using the training method of the recall model; acquiring object information of the object to be recalled, inputting the object information into a second sub-network of the target recall model to obtain an object representation vector corresponding to the object information; and determining the recall object corresponding to the search term based on the similarity between the search term representation vector and the object representation vector.

[0020] Optionally, the step of inputting the search term into the first sub-network of the target recall model to obtain the search term representation vector corresponding to the search term includes: inputting the search term into the first sub-network to obtain a first prediction output; inputting the first prediction output into the first multilayer perceptron of the target recall model, and using the output of the first multilayer perceptron as the search term representation vector.

[0021] Optionally, the step of inputting the object information into the second sub-network of the target recall model to obtain the object representation vector corresponding to the object information includes: inputting the object information into the second sub-network to obtain a second prediction output; inputting the second prediction output into the second multilayer perceptron of the target recall model, and using the output of the second multilayer perceptron as the object representation vector.

[0022] This application embodiment also provides a retrieval method, including: acquiring a search term input by a user; obtaining a search term representation vector corresponding to the search term based on a first sub-network of a target recall model; wherein the target recall model is a recall model trained using the training method of the recall model; determining merchants whose delivery range covers the user's location based on the user's location, and acquiring the address code covering the delivery range of the merchant; acquiring an object representation vector set associated with the address code, and determining several nearest neighbor vectors of the search term representation vector in the object representation vector set; determining the object representation vector retrieved for the search term representation vector based on the nearest neighbor vectors, and taking the object corresponding to the retrieved object representation vector as the recall object retrieved for the search term; the distance between the nearest neighbor vector and the search term representation vector satisfies the vector similarity condition.

[0023] Optionally, it also includes: determining the object representation vector corresponding to the object to be recalled in the object candidate set based on the second sub-network of the target recall model; generating the address code of the object to be recalled based on the delivery information related to the object representation vector; performing data aggregation on each object representation vector based on the address code; and constructing an inverted index for retrieving the corresponding recalled object based on search terms.

[0024] Optionally, the inverted index is a two-level index, including a first-level inverted index for indexing merchants to their corresponding address code sets, and a second-level inverted index for indexing address codes to their corresponding object representation vector sets.

[0025] Optionally, obtaining the address code covering the merchant's delivery range includes: retrieving at least one address code covering the merchant's delivery range based on the first-level inverted index.

[0026] Optionally, obtaining the object representation vector set associated with the address code includes: retrieving the object representation vector set corresponding to the address code based on the second-level inverted index.

[0027] Optionally, generating the address code of the object to be recalled based on the delivery information related to the object representation vector includes: obtaining the object identifier, merchant identifier, and delivery range identifier corresponding to the object representation vector; expanding the polygon corresponding to the delivery range identifier, obtaining the address code covered by the polygon, and using it as the address code of the object to be recalled.

[0028] Optionally, the step of aggregating the object representation vectors based on the address codes includes: when the number of object representation vectors corresponding to the same address code is less than the number of objects, aggregating the object representation vectors based on a linear index; otherwise, aggregating the object representation vectors based on hierarchical clustering.

[0029] Optionally, determining several nearest neighbor vectors of the search term representation vector in the object representation vector set includes: if the object representation vector set is generated based on hierarchical clustering, determining the center point layer by layer, traversing the next-level nodes contained in each layer's center point, until the similarity between the vector corresponding to each node and the search term representation vector is linearly compared, filtering out nodes whose similarity to the search term representation vector satisfies the vector similarity condition, and using the vector corresponding to the filtered nodes as the nearest neighbor vector.

[0030] Optionally, the step of determining the center point layer by layer and traversing the next-level nodes contained in each center point until the similarity between the vector corresponding to each node and the vector representing the search term is linearly compared includes: determining at least one first-level center point; for each first-level center point, traversing the second-level center points contained in that first-level center point; determining the nodes contained in each second-level center point; traversing the nodes contained in each second-level center point; and linearly comparing the similarity between the vector corresponding to each node and the vector representing the search term.

[0031] Optionally, determining the object representation vector retrieved for the search term representation vector based on the nearest neighbor vector includes: determining the delivery range of the nearest neighbor vector; the delivery range being a polygon; drawing a horizontal ray with the user's location as the starting point; if the number of intersections between the ray and each side of the polygon is odd, then the delivery range covers the user's location; otherwise, the delivery range does not cover the user's location; removing vectors from the nearest neighbor vectors whose delivery range does not cover the user's location, and using the remaining nearest neighbor vectors as the retrieved object representation vector.

[0032] Optionally, the search term representation vector and the object representation vector are vectors in the same high-dimensional space; the object representation vector is a vector generated offline using the object-side network structure of the recall model.

[0033] This application embodiment also provides a training device for a recall model, comprising: a search-side network unit, configured to use search-side information of training samples as first input information of the recall model to be trained, input the first sub-network of the recall model to obtain first output information, and determine the search term features of the training samples based on the first output information; an object-side network unit, configured to use object-side information of training samples as second input information of the recall model, input the second sub-network of the recall model to obtain second output information, and determine the recall object features of the training samples based on the second output information; and a training unit, configured to train the recall model based on a first loss function, and obtain a target recall model after training; wherein, the first loss function is constructed based on the similarity between the search term features of the training samples and the corresponding recall object features, and is used as the objective function for training the recall model.

[0034] This application embodiment also provides a search retrieval device, comprising: a search term representation unit, configured to acquire a search term input by a user, input the search term into a first sub-network of a target retrieval model, and obtain a search term representation vector corresponding to the search term; wherein the target retrieval model is a retrieval model trained using the training method of the retrieval model; an object representation unit, configured to acquire object information of an object to be recalled, input the object information into a second sub-network of the target retrieval model, and obtain an object representation vector corresponding to the object information; and a retrieval unit, configured to determine the retrieval object corresponding to the search term based on the similarity between the search term representation vector and the object representation vector.

[0035] This application embodiment also provides a retrieval device, including: a search term representation unit, used to acquire a search term input by a user, and obtain a search term representation vector corresponding to the search term based on a first sub-network of a target recall model; wherein, the target recall model is a recall model trained using the training method of the recall model; a retrieval address encoding unit, used to determine merchants whose delivery range covers the user's location based on the user's location, and obtain the address encoding of the delivery range covered by the merchant; a retrieval nearest neighbor vector unit, used to acquire the object representation vector set associated with the address encoding, and determine several nearest neighbor vectors of the search term representation vector in the object representation vector set; a recall unit, used to determine the object representation vector retrieved for the search term representation vector based on the nearest neighbor vectors, and take the object corresponding to the retrieved object representation vector as the recall object retrieved for the search term; the distance between the nearest neighbor vector and the search term representation vector satisfies the vector similarity condition.

[0036] This application also provides an electronic device, including: a memory and a processor; the memory is used to store a computer program, which, when run by the processor, executes the method provided in this application.

[0037] This application also provides a computer storage medium storing computer execution instructions, which, when executed by a processor, are used to implement the method provided in this application.

[0038] Compared with the prior art, this application has the following advantages:

[0039] This application provides a training method, apparatus, electronic device, and storage medium for a recall model. The method involves using the search-side information of training samples as the first input information to the recall model, inputting it into a first sub-network of the recall model to obtain first output information, and determining the search term features of the training samples based on the first output information. The method also involves using the object-side information of training samples as the second input information to the recall model, inputting it into a second sub-network of the recall model to obtain second output information, and determining the recall object features of the training samples based on the second output information. The recall model is trained based on a first loss function, resulting in a target recall model. The first loss function is constructed based on the similarity between the search term features of the training samples and the corresponding recall object features, and is used as the objective function for training the recall model. The target recall model and the recall model represent the text of the search term and the object to be recalled as semantic vectors in the same high-dimensional space, and recall objects based on the similarity of the semantic vectors. This target recall model and recall model have strong applicability, exhibiting high recall and precision even in scenarios where merchants' fulfillment capabilities are constrained by time and / or space.

[0040] This application provides a search retrieval method, apparatus, electronic device, and storage medium. By applying the retrieval model or target retrieval model trained using the method provided in this application, the first and second sub-networks of the trained model respectively obtain search term representation vectors and object representation vectors. Based on the similarity between the search term vectors and object representation vectors, the retrieval objects for the search term retrieval are determined. This can improve both object recall rate and recall precision.

[0041] This application provides a retrieval method, apparatus, electronic device, and storage medium. By acquiring a user-input search term, a search term representation vector corresponding to the search term is obtained based on a first sub-network of a target recall model. The target recall model is a recall model trained using the training method of the recall model. Merchants whose delivery range covers the user's location are determined based on the user's location, and the address codes covering the delivery range of the merchants are obtained. An object representation vector set associated with the address codes is obtained, and several nearest neighbor vectors of the search term representation vector in the object representation vector set are determined. An object representation vector retrieved for the search term representation vector is determined based on the nearest neighbor vectors, and the object corresponding to the retrieved object representation vector is used as the recall object retrieved for the search term. The distance between the nearest neighbor vector and the search term representation vector satisfies a vector similarity condition. This provides a location-based service-based retrieval scheme and improves retrieval efficiency. Attached Figure Description

[0042] Figure 1 This is a flowchart illustrating the training method for a recall model provided in the first embodiment of this application;

[0043] Figure 2 This is a schematic diagram of the structure of a recall model provided in the first embodiment of this application;

[0044] Figure 3 This is a schematic diagram of another recall model provided in the first embodiment of this application;

[0045] Figure 4 This is a schematic diagram of generating a search click map provided in the first embodiment of this application;

[0046] Figure 5 This is a schematic diagram of another recall model provided in the first embodiment of this application;

[0047] Figure 6 This is a flowchart of a search and recall method provided in the second embodiment of this application;

[0048] Figure 7 This is a flowchart of a retrieval method provided in the third embodiment of this application;

[0049] Figure 8 This is a schematic diagram of a hierarchical clustering-based retrieval method provided in the third embodiment of this application;

[0050] Figure 9 This is a schematic diagram of a training device for a recall model provided in the fourth embodiment of this application;

[0051] Figure 10 This is a schematic diagram of a search and recall device provided in the fifth embodiment of this application;

[0052] Figure 11 This is a schematic diagram of a retrieval device provided in the sixth embodiment of this application;

[0053] Figure 12 This is a schematic diagram of the electronic device provided in this application. Detailed Implementation

[0054] Many specific details are set forth in the following description to provide a full understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this application; therefore, this application is not limited to the specific embodiments disclosed below.

[0055] This application provides a method, apparatus, electronic device, and storage medium for training a recall model. This application also provides a search recall method, apparatus, electronic device, and storage medium. This application further provides a retrieval method, apparatus, electronic device, and storage medium. These will be described in detail in the following embodiments.

[0056] To facilitate understanding, an application scenario for the recall model is first presented. If a platform or system providing search functionality (collectively referred to as a search system) offers search users information on various industry products and / or services and connects to search requests from multiple endpoints, this places higher demands on the depth and quality of search recall. In one scenario providing search functionality, platform A provides online-to-offline services. The product form of this platform's search functionality is to provide search services across multiple endpoints and industries. Multiple endpoints include, but are not limited to, client apps connected to the platform, mini-program homepages, channel search pages, and search tabs or search cards associated with the platform in related apps. Multiple industries include, but are not limited to, catering, department store retail, and pharmaceuticals. This scenario is merely one example of the application of the recall model. The purpose of providing this scenario example is to facilitate understanding of the application of the recall model provided in this application, and it is not intended to limit the application or training method of the recall model in this application.

[0057] The recall model trained using the method described in this application is applied to the recall stage of the search function. Using this model, information related to the user's input search terms is retrieved from massive amounts of data. This retrieved information is then further processed and provided to the user. The retrieved information related to the search terms can be understood as objects related to the search terms recalled by the search system based on the received search terms. An object refers to one or more items included in the recall results. Specifically, it can be a store, a product or service, a review, or a document related to the search terms, without special limitations. The recalled objects can be attached to a category. The category to which the object is attached is the object's category. The object's category can be a hierarchical first-level or multi-level category, and the object's category is the object's context. The same object can also belong to multiple categories. For example, a store, such as a cake shop (A), can be attached to multiple categories such as Food / Mealtimes / Afternoon Tea, Food / Bread & Desserts / Bread & Cake, and Food / Hot Selling Items / Cake. In this application, the user-input search term can be understood as the search side of search recall, and the information related to the search term can be understood as the object side (or document side) of search recall.

[0058] It should be noted that the information disclosed above is only for the purpose of helping to understand this application and does not constitute prior art known to those skilled in the art.

[0059] The following combination Figures 1 to 5 The training method of the recall model provided in the first embodiment of this application will be described. Figure 1 The training methods for the recall model shown include:

[0060] Step S101: Use the search side information of the training sample as the first input information of the recall model to be trained, input it into the first sub-network of the recall model to obtain the first output information, and determine the search word features of the training sample based on the first output information.

[0061] Step S102: Use the object-side information of the training samples as the second input information of the recall model, input it into the second sub-network of the recall model to obtain the second output information, and determine the recall object features of the training samples based on the second output information;

[0062] Step S103: Train the recall model based on the first loss function to obtain the target recall model after training; wherein, the first loss function is constructed based on the similarity between the search term features of the training samples and the corresponding recall object features, and is used as the target function for training the recall model.

[0063] The method provided in this embodiment is used to train a recall model, which, or the target recall model obtained by training the recall model, is applied to the recall phase of the search function. The recall model provides the data foundation for the depth and quality of search recall, and needs to balance the relevance between the recalled objects and search terms, as well as recall efficiency. The recall model obtained by the training method of the recall model provided in this application is built on a semantic model based on deep learning, logically including a first sub-network and a second sub-network. The first sub-network is a sub-network on the search side for learning search-side information; the second sub-network is a sub-network on the object side for learning object-side information. The input information of the first sub-network is the first input information, which can be the search term and / or search term-related information input by the user on the search side, or data formatted after processing the search term or search term-related information. The input information of the second sub-network is the second input information, which can be the recalled object and / or recalled object-related information corresponding to the search term on the object side, or data formatted after processing the recalled object or recalled object-related information. Furthermore, the search term features are not limited to the feature information of the search term (i.e., the search text) itself entered by the user. They can also be extended features that incorporate other feature information related to the search term. The other feature information related to the search term is the search term extended information. For example, the search term features are feature information based on the search term and incorporating the search term segmentation sequence, the feature words corresponding to the search term, the content obtained by concatenating the search term and the corresponding user behavior features, etc.

[0064] As described in step S101, the search-side information of the training samples is used as the first input information of the recall model to be trained, and input into the first sub-network of the recall model to obtain the first output information. The search term features of the training samples are then determined based on the first output information. This step involves determining the search term features of the training samples through the first sub-network. Specifically, using the search-side information of the training samples as the first input information of the recall model to be trained includes: obtaining the search terms and / or extended search term information contained in the training samples as the search-side information, and constructing the first input information based on the search terms. Determining the search term features of the training samples based on the first output information includes: inputting the first output information into a first multilayer perceptron to obtain the search term features, where the search term features include semantic information of the search terms. A multilayer perceptron (MLP) is a type of perceptron.

[0065] Please refer to Figure 2The model structure shown in the figure separates the search-side network structure and the object-side network structure of the recall model, representing the search-side information and the object-side information respectively. Logically, it is a dual-tower structure, specifically including: a first sub-network 201, a second sub-network 202, first input information 203, second input information 204, a first multilayer perceptron 205, a second multilayer perceptron 206, and a first loss function 207. The first and second sub-networks can employ semantic models based on deep learning. The first input information is input into the first sub-network to obtain the first output information. This first output information is then input (i.e., passed to) the first multilayer perceptron. After activation by the first multilayer perceptron, search term features containing semantic information of the search terms are obtained. These search term features can be understood as query embeddings represented in vector form. The second input information is fed into the second sub-network to obtain the second output information. This second output information is then fed into (i.e., passed to) the second multilayer perceptron. After activation by the second multilayer perceptron, recall object features containing semantic information about the recalled objects corresponding to the search terms are obtained. These recall object features can be understood as object-side embedded feature vectors (item embeddings) represented in vector form. In the diagram, the first loss function is the objective function during the recall model training phase. It is constructed based on the search term features and recall object features. During the training phase, the search term features and recall object features of the training samples are mapped to the same high-dimensional space, progressively training the network parameters to make similar search term features and recall object features closer together, while dissimilar features are further apart. In the diagram, PE represents a processing unit (ProcessElement), a specific example of which could be a corresponding neuron.

[0066] In this embodiment, the first sub-network and the second sub-network have the same network structure and / or shared parameters, where shared parameters refer to the two sub-networks sharing the same parameters. Further, the first sub-network and the second sub-network can also be two logically distinct networks, but in practice, the same network is used to process the first input information and the second input information respectively. Preferably, the first sub-network and the second sub-network employ a BERT model. Further, the first sub-network and the second sub-network employ a specific BERT model pre-trained using domain corpus from the search system's application domain. The domain corpus is data collected for the domain in which the search system is applied, and the collected data undergoes data extraction, cleaning, and other processing steps to obtain the corpus. For example, when the search system is applied to the O2O (online-to-offline) domain, a general BERT model is pre-trained using specific O2O domain corpus to obtain a specific BERT (Bidirectional Encoder Representation from Transformers, BERT) model as the first sub-network and / or the second sub-network. The BERT model and / or the specific BERT model are language representation models, and their last layer is a pooling layer. The network nodes that receive input information before the first and / or second sub-networks are embedded layers, meaning the input to the sub-network is introduced by one of the hidden layers of the recall model. In implementation, the outputs from the last pooling layer of each of the first and / or second sub-networks are passed to the corresponding multilayer perceptrons. After activation by the multilayer perceptrons, the search term features and corresponding recall object features are obtained. The BERT model, combined with the learning method of the multilayer perceptron, can improve training performance and prediction accuracy. Preferably, the first and second output information of the first and second multilayer perceptrons are mapped to the same high-dimensional space, respectively. The resulting search term features and corresponding recall object features are embedded feature vectors representing the semantic information of the search term and the semantic information of the recall object, respectively, within the same high-dimensional space.

[0067] Search-side information (or features) is related to user behavior. To enrich the semantic expression of search-side information during the training process of the recall model and fully learn search-side features, so that the recall model can fully extract the semantics of user-input search terms when applied to online recall, this embodiment incorporates search click data related to user behavior during the learning process of search-side information to expand search term features, thereby fully mining the semantic expression of the search side. Here, search-side features can be understood as features that incorporate the search term features of the training samples and other relevant search-side information. Preferably, the search click data added to the search side in each round of training of the recall model includes the neighbor information of the search terms of the current training sample in this round of training. The neighbor information of the search terms is the information of neighboring search terms that have a neighbor relationship with the search term, determined based on the click behavior of the recall object. That is, the neighboring search terms of the search term are search terms that have a neighbor relationship with the search term. Specifically, search terms corresponding to the same recall object that have experienced a click behavior are neighbors, and these search terms have a neighbor relationship with each other. For example, the first search term and the second search term have a direct neighbor relationship and are neighbors; at the same time, the first search term and the third search term do not have a direct neighbor relationship, while the second search term and the third search term have a direct neighbor relationship, then the first search term and the third search term have an indirect neighbor relationship. Specifically, the method includes: acquiring search click data of search terms in the training samples, the search click data including the neighbor information of the search terms; acquiring search term neighbor features of the search terms based on the neighbor information, the search term neighbor features being used to determine graph attention features related to the neighbor information of the search terms; determining graph attention features related to the neighbor information of the search terms based on the search term neighbor features and the first output information; determining a second loss function based on the graph attention features and the search term features, and using the second loss function to train the recall model, specifically, using the first loss function and the second loss function as the objective function for training the recall model.

[0068] In a preferred embodiment, the recall model further includes a third sub-network, meaning the recall model comprises a first, a second, and a third sub-network. The third sub-network is used by the search side to obtain the neighbor features of the search terms based on the neighbor information of the search terms in the training samples. The neighbor information of the search terms is preferably the search click graph data corresponding to the search terms. That is, during the training process of the recall model, search click graph data is introduced on the search side, and the third sub-network is used to process the search click graph data to obtain the search neighbor features. A search click graph refers to a bipartite graph generated by establishing click relationships between nodes by treating the search terms and their corresponding recall objects as search term nodes and object nodes, respectively, and aggregating search term nodes with neighbor relationships in the click relationships. The neighbor relationships are described above. Compared to using only the click information of the search terms themselves, learning the search click graph data can expand the search click information related to the search terms, making the semantic expression on the search side richer. During the training of the recall model, for each training sample, information from one or more neighboring nodes of the search term it contains is used to recover the expression of the search term node in the currently computed training sample based on the expression of the search term's neighboring nodes, in order to learn network parameters. Furthermore, an attention mechanism can be introduced into the recall model based on a third sub-network to distinguish the contributions of different neighboring nodes, thereby fully exploring the semantic expression on the search side, expanding the search term features on the search side, and enabling the learned network parameters to provide better recall quality after training.

[0069] Preferably, the third sub-network has the same network structure and / or shares parameters with the first sub-network. Further, the third sub-network may have the same network model as the first sub-network, but is only a logical sub-network. The third sub-network is preferably a BERT network, especially a BERT network pre-trained using domain corpus corresponding to the domain of the search system.

[0070] Preferably, the second loss function is a reconstruction loss function constructed based on the search term features and the graph attention features.

[0071] Please refer to Figure 3 The recall model structure shown in the figure is in Figure 2Based on the network structure shown, search click graph data is introduced into the search-side network structure of the recall model, including: a first Bert encoder 301, a second Bert encoder 302, first Bert encoder input information 303, second Bert encoder input information 304, a first multilayer perceptron 305, a second multilayer perceptron 306, a first loss function 307, a search-side click graph 308, a third Bert encoder 309, graph attention 310, and a second loss function 311. The first Bert encoder is the first sub-network, and its input information is the input information of the first sub-network; the second Bert encoder is the second sub-network, and its input information is the input information of the second sub-network. The third Bert encoder is the third sub-network that processes the search click graph data; it is the same network model as the first and second sub-networks and shares parameters. In the figure, during each round of training using training samples, the current node is the search term node corresponding to the current search term contained in the current training sample in this round of calculation; the information contained in the search-side click graph is the search click graph data, which is the input information of the third Bert encoder. In implementation, based on the adjacency of each neighboring node of the current node in the search-side click graph, the neighboring node information in the search-side click graph is processed by PE and then input into the third Bert encoder. That is, there can be multiple logical third sub-networks (i.e., the third Bert encoder). For example, the neighbor Q1 of the current node is input into a third sub-network, and the neighbor Q2 of the current node is input into a third sub-network. In implementation, the input can be drawn from different hidden layers to their respective logical third sub-networks. In the diagram, the graph attention between the current node and its neighboring nodes is calculated based on the output of the third Bert encoder. In practice, the representation vector of each neighboring node obtained from the third Bert encoder is used to calculate attention with the representation vector of the current node, resulting in an attention coefficient for each neighboring node corresponding to the current node. Then, the representation vectors of the corresponding neighboring nodes are added together according to the attention coefficient to obtain the graph attention feature (i.e., the representation vector of the search click graph). Based on the graph attention feature and the search term features obtained after the current node has been processed by the first Bert encoder and then by the first multilayer perceptron, a reconstruction loss function is constructed, which is the second loss function. The recall model is trained based on the second loss function and the first loss function. This recall model integrates the attention information between the search terms of the training samples and their neighbors, and can be understood as a recall model that integrates neighbor attention. In the diagram, the information of the current node and its neighboring nodes corresponding to the search terms of the current training sample is input into the BERT layer of the recall model (the first Bert encoder and the third Bert encoder, respectively), resulting in the following representation vectors for the current node and its neighboring nodes:

[0072] h q =bert(q);

[0073] The representation vectors corresponding to neighboring nodes can be understood as the search term's neighbor features. Then, attention is calculated between each neighboring node and the representation vector of the current node to obtain the attention coefficient for each neighboring node corresponding to the current node. The representation vectors of the corresponding neighboring nodes are then summed according to these attention coefficients to obtain the graph representation vector of the search click graph corresponding to the current node. This graph representation vector can be understood as the graph attention features of the search term in the current training sample related to its neighbor information, as follows:

[0074] e qn =a(Wh q ·Wh n );

[0075]

[0076]

[0077] Among them, e qn The attention coefficient of each neighbor node to the current node; a qn h represents the normalized attention coefficient. graph(q) This refers to the graph attention features related to neighbor information corresponding to the search terms of the current training sample in this round of training.

[0078] Furthermore, the first loss function in the figure is a reconstruction loss function constructed based on the graph attention features and the representation vector (i.e., search term features) obtained by the MLP after encoding by the first Bert encoder. The details are as follows:

[0079]

[0080] in,

[0081] The second loss function aims to maximize the similarity between the current training sample's search term and its neighboring search terms, and minimize the distance between them, during each round of training. This second loss function is used for backpropagation to update network parameters on the search side of the recall model, such as updating network parameters in the first sub-network.

[0082] In this embodiment, as Figure 3 The training objective of the recall model that integrates neighbor attention is that the first loss function and the second loss function satisfy the convergence condition.

[0083] In this embodiment, the method further includes generating the search click graph. Generating the search click graph includes: acquiring search logs; generating a click relationship graph based on the search logs to represent the relationship between search terms and their corresponding recall objects; each node in the click relationship graph is a search term node corresponding to the search term information, and a click object node corresponding to each search term that has been clicked by the user; the click object node represents any object clicked by the user among the recall objects corresponding to the search term; generating a bipartite graph based on the click relationship graph, and using the generated bipartite graph as the search click graph. Specifically, generating the bipartite graph based on the click relationship graph includes: generating the bipartite graph by performing neighbor aggregation on each search term node in the click relationship graph; wherein search term nodes with neighbor relationships in the bipartite graph are neighbor nodes, and the edges between neighbor nodes can have weights, reflecting the degree of mutual influence between neighbor nodes through the weights. The "neighbor relationship" is consistent with the above description, meaning that search term nodes with the same recall object, and whose recall objects have been clicked by the user for the corresponding search term, have a neighbor relationship. Please refer to [reference needed]. Figure 4 The figure illustrates a search click graph, including a click relationship graph 401 obtained from search logs and a search click graph 402 obtained from aggregated neighbors. The nodes in the click relationship graph are: search term (query) nodes and recall object (product) nodes, as shown in the examples q1, q2, q3, q4, q5, and p1, p2, p3, p4; the lines between nodes indicate clicks. The aggregated search click graph is a bipartite graph, showing that q1 is a neighbor of q2 and q3, but not of q4 and q5. The nodes in the figure are merely illustrative, and the number and names of the nodes do not constitute a limitation on the method. The data from the search click graph is used as search click graph data introduced from the search side to participate in the training of the recall model. Specifically, after generating the search click graph, the recall model is trained using the search click graph data, including: sampling neighbor nodes related to the search term in the search click graph based on the search term, using the sampled data as the neighbor information of the search term, and having a third sub-network obtain the search term neighbor features of the search term based on the neighbor information; the search term neighbor features are used with the first output information to determine the graph attention features related to the search term and the neighbor information; the graph attention features are used with the search term features to determine a second loss function. The sampling of neighbor nodes related to the search term in the search click graph based on the search term includes: sampling according to the sampling weight of each search term, and / or sampling according to the number of clicks on the recalled object. The search log can refer to search record information generated by the main search system providing search functionality.

[0084] As described in step S202, the object-side information of the training samples is used as the second input information of the recall model, and input into the second sub-network of the recall model to obtain the second output information. The recall object features of the training samples are determined based on the second output information. This step involves learning the object-side information through the second sub-network, that is, learning network parameters from the object side of the recall model. In this embodiment, the network parameters of the first and second sub-networks are shared parameters, that is, they share the same network parameters. The second input information. Wherein, determining the recall object features of the training samples based on the second output information includes: inputting the second output information into a second multilayer perceptron to obtain the recall object features, wherein the recall object features contain semantic information of the recall object corresponding to the search term.

[0085] The second input information may be the recalled object and / or related information of the recalled object corresponding to the search term on the object side. For example, the second input information includes the object name, and may further include the object context. The object context may be the category to which the object belongs.

[0086] In this embodiment, to introduce interaction between the search side and the target side during the recall model training phase, thereby improving the recall rate and recall precision of the trained model, the recall rate reflects the depth and breadth of the recall, while the recall precision reflects the relevance between the recalled object and the search term. Higher relevance leads to a higher conversion rate of potential users using the search function into actual users. In practice, click data corresponding to the recalled objects (products or stores) included in the training samples is added. This click data includes one or more search terms for the recalled object whose click rate meets a click rate threshold or whose click quantity meets a click quantity threshold. Specifically, this includes: acquiring the click data of the recalled objects corresponding to the search terms in the training samples; and generating the second input information (or...) based on the click data and the object information of the recalled objects. Figure 3 The second BERT encoder input information in the sample includes the object name and object context, separated from the click data (such as one or more search terms that meet the click-through rate threshold or the number of clicks threshold) by [SEP]. It is understood that the one or more search terms corresponding to each recalled object in the sample are neighbors of the search terms in the sample.

[0087] In this embodiment, it also includes... Figure 2 or Figure 3 Based on the model shown, the click data is introduced into the object-side network structure of the recall model (such as the second sub-network). Please refer to... Figure 5 The recall model structure shown in the figure is in Figure 3The model structure obtained by introducing the click data on the object side based on the model shown includes: a recall object 501 corresponding to the search term on the object side, which, compared to Figure 3 The recall object 304 corresponding to the search term on the object side has been expanded to include "the top N search terms in the ranking of recall object clicks or click-through rates", where N is an illustrative number and is not used to limit the method.

[0088] As described in step S103, the recall model is trained based on the first loss function, and the target recall model is obtained after training; wherein, the first loss function is constructed based on the similarity between the search term features of the training samples and the corresponding recall object features, and is used as the target function for training the recall model.

[0089] Similarity refers to the degree of matching between a search term and the retrieved object or the category to which the retrieved object belongs. It reflects the relevance between the search term and the retrieved object. The higher the similarity, the higher the expected match between the retrieved object or its category and the search term, and the better the relevance. Conversely, the lower the similarity, the lower the expected match between the retrieved object or its category and the search term, and the worse the relevance.

[0090] During the training phase of the recall model, training stops when the objective function meets the convergence condition, resulting in the target recall model. The objective function for model training is constructed based on the similarity between the search term features of the training samples and the corresponding features of the recalled objects. The trained target recall model, when applied to a search system, achieves a higher degree of match between the recalled objects and the user's input search terms, better reflecting the user's search intent, thus increasing the click-through rate of the recalled objects and ultimately improving the conversion rate from potential users to actual users. Figure 2 , Figure 3 as well as Figure 5 As can be seen from this, the first loss function can be specifically determined by the output of the first multilayer perceptron and the output of the second layer perceptron.

[0091] This embodiment provides several ways to construct the first loss function. One preferred embodiment uses a contrast-based infoNCE loss function, as follows:

[0092]

[0093] Where q is the representation vector of the search term; k + k is the representation vector of the recalled objects of the positive samples. -τ is the representation vector of the recalled objects for negative samples; τ is the popularity hyperparameter. Positive samples refer to the training samples corresponding to recalled objects that have been exposed and clicked within a certain historical period. Specifically, they can be composed of the tuple <search term, clicked object>, where the clicked object refers to the recalled object corresponding to the search term that has been clicked; positive samples indicate that the user is interested in the recalled object. Negative samples refer to the training samples constructed based on objects that have not been recalled for a search term, and / or, the recalled objects that have not been exposed or have been exposed but not clicked within the recalled objects for a search term. In practice, for a search term, random samples can be drawn from all objects to be recalled or objects that have not been recalled, and the sampling results can be combined with the search term to construct negative samples.

[0094] Of course, other implementation methods can also be used to construct the first loss function. For example, in one implementation, a classification-based Hinge loss function is used, as follows:

[0095] L hinge =max(0,e q ·e p -e q ·e n +m);

[0096] Among them, e q ·e p The cosine similarity of positive samples; e q ·e n The cosine similarity is used for negative samples; positive samples are selected from <search term, recall object> pairs with sufficient exposure and high click-through rate; negative samples can be constructed based on positive samples. For example, they can be randomly sampled from objects that belong to the same level category but different level categories as the recall objects of the positive samples, and the sampled objects and the search terms of the positive samples are used to form negative samples.

[0097] For example, in one implementation, a Triplet loss function can be used. The feedback data includes anchor samples, positive samples, and negative samples. The training process optimizes the distance between the anchor sample and the positive sample to be less than the distance between the anchor sample and the negative sample, as follows:

[0098]

[0099] Where d(q,p) represents the distance between vectors q and p; d(q,n) represents the distance between vectors q and n; (q,p) represents the positive sample consisting of the search term and the recalled objects that have been exposed and clicked; (q,n) represents the negative sample consisting of the search term and other recalled objects, or the negative sample generated by modifying the category to which the clicked recalled objects belong.

[0100] For example, in one implementation, the Circle loss function is used, which adds weights to negative samples based on the Triplet loss function, thereby controlling the gradient contributions of positive and negative samples, as follows:

[0101]

[0102] Among them, s p It is the intra-class similarity, s n It represents inter-class similarity, γ is the scaling factor, and α is the inter-class similarity. p and α n It is an adaptive pacing weighting factor.

[0103] In this embodiment, the method further includes constructing training samples for training the recall model based on search logs before training; wherein the training samples include the search-side information and the object-side information of the recall objects corresponding to the search-side information. Specifically, the search terms input by the user on the search side and the recall objects corresponding to the search terms are obtained from the search logs, and the training samples are constructed based on the search terms and the recall objects. The constructed training samples include positive samples and / or negative samples. This application does not limit the generation method of positive and negative samples. Positive samples can be generated based on search logs or sample pairs consisting of manually annotated search terms and clicked objects can be used as positive samples. Negative samples can be generated based on positive samples. For example, the category information to which the recall object of the positive sample belongs can be modified, and then the modified sample can be combined with the search terms of the positive sample to form a negative sample; or, other objects besides the recall objects of the positive sample can be sampled, and the sampled objects can be combined with the search terms of the positive sample to form a negative sample. Of course, negative samples can also be constructed based on manually annotated information.

[0104] In practice, an unlabeled domain corpus can be used to pre-train a BERT model. This pre-trained BERT model serves as the logical first and second sub-networks, constructing the initial recall model to be trained. Then, positive and / or negative samples generated from search logs or manually labeled are used to train the recall model. Further, a search click graph is introduced on the search side and / or information interaction information between the search side and the object side is introduced on the object side to train the target recall model. This target recall model represents the text of the search term and the object to be recalled as semantic vectors in the same high-dimensional space, and recalls objects based on the similarity of the semantic vectors. This target recall model has strong domain applicability, exhibiting high recall and precision even in scenarios where merchants' fulfillment capabilities are constrained by time and / or space. The recall rate reaches 0.95, and the average recall depth is also improved. The trained target recall model also has a superior F1 score, which is an indicator of model precision.

[0105] It should be noted that, unless otherwise specified, the features given in this embodiment and other embodiments of this application can be combined with each other, and steps S101 and S102 or similar terms do not limit the steps to be performed in a specific order.

[0106] The method provided in this embodiment has been described above. The method uses the search-side information of the training samples as the first input information to the recall model to be trained, inputs it into the first sub-network of the recall model to obtain first output information, and determines the search term features of the training samples based on the first output information. It also uses the object-side information of the training samples as the second input information to the recall model, inputs it into the second sub-network of the recall model to obtain second output information, and determines the recall object features of the training samples based on the second output information. The recall model is trained based on a first loss function, resulting in a target recall model. The first loss function is constructed based on the similarity between the search term features of the training samples and the corresponding recall object features, and is used as the objective function for training the recall model. The target recall model and the recall model represent the text of the search terms and the objects to be recalled as semantic vectors in the same high-dimensional space, and recall objects based on the similarity of the semantic vectors. This target recall model and the recall model have strong applicability, exhibiting high recall and precision even in scenarios where merchants' fulfillment capabilities are constrained by time and / or space.

[0107] Based on the above embodiments, the second embodiment of this application provides a search and recall method. The following is combined with... Figure 6 The method will be described below. For the same parts, please refer to the description of the corresponding parts in the above embodiments, which will not be repeated here. Figure 6 The search recall methods shown include:

[0108] Step S601: Obtain the search term input by the user, input the search term into the first sub-network of the target recall model, and obtain the search term representation vector corresponding to the search term; wherein, the target recall model is a recall model trained using the method provided in this application;

[0109] Step S602: Obtain object information of the object to be recalled, input the object information into the second sub-network of the target recall model, and obtain the object representation vector corresponding to the object information;

[0110] Step S603: Based on the similarity between the search term representation vector and the object representation vector, determine the recall object corresponding to the search term.

[0111] The target recall model provided in this application is used as a recall model applied to search recall. The target recall model refers to the trained recall model. The user-inputted search term is fed into the target recall model. The target recall model then recalls objects corresponding to the search term from the objects to be recalled, based on the user-inputted search term, as the recalled objects. The objects to be recalled can be understood as objects within the recall scope, often involving massive amounts of data. An example is the entire product or service offerings of a platform or system deploying search recall functionality, which can be understood as objects within the recall scope.

[0112] As described in the above embodiments of the training method for the recall model, the first sub-network is located in the search-side network structure of the target recall model, and the second sub-network is located in the object-side network structure of the target recall model. In practical applications, the search-side network structure of the target recall model, such as the first sub-network and other PEs and / or perceptrons connected to it, can be deployed on the online data side for processing user-input search terms online and predicting the corresponding search term representation vectors in real time for the user-input search terms. Specifically, the search-side network structure can be deployed to an RTP (Real-time Predict) system, which receives user-input search terms and uses the search-side network of the target recall model to predict vectors that meet the scoring thresholds in real time as the search term representation vectors. The object-side network structure of the target recall model, such as the second sub-network and other PEs and / or perceptrons that predict its vectors, can be deployed on the offline data side in the form of tasks to predict the object representation vectors of the recalled objects offline. Preferably, the obtained object representation vectors are used to form a retrieval library, and incremental offline prediction is performed on the retrieval library. Then, based on the search term representation vector, one or more objects whose similarity meets the preset conditions are queried from the object representation vector obtained by offline prediction. These objects are used as the recall objects corresponding to the search term, thereby achieving the goal of finding recall objects of interest to the user based on vector similarity.

[0113] In this embodiment, the output of the first sub-network is passed to the first multilayer perceptron, and activated to obtain a search term representation vector corresponding to the search term predicted by the search term. Specifically, inputting the search term into the first sub-network of the target recall model to obtain the search term representation vector includes: inputting the search term into the first sub-network to obtain a first prediction output; inputting the first prediction output into the first multilayer perceptron of the target recall model, and using the output of the first multilayer perceptron as the search term representation vector. The output of the second sub-network is passed to the second multilayer perceptron, and activated to obtain an object representation vector corresponding to an object within the recall range. Specifically, inputting the object information into the second sub-network of the target recall model to obtain an object representation vector corresponding to the object information includes: inputting the object information into the second sub-network to obtain a second prediction output; inputting the second prediction output into the second multilayer perceptron of the target recall model, and using the output of the second multilayer perceptron as the object representation vector.

[0114] This concludes the description of the method provided in this embodiment. The method applies the target recall model trained using the method described in this application. The first and second sub-networks of the trained model respectively obtain the search term representation vector and the object representation vector. Based on the similarity between the search term vector and the object representation vector, the recall objects for the search term are determined. This can improve both the object recall rate and the recall precision.

[0115] Based on the above embodiments, the third embodiment of this application provides a retrieval method. The following is in conjunction with... Figures 7 to 8 The method will be described below. For the same parts, please refer to the description of the corresponding parts in the above embodiments, which will not be repeated here. Figure 7 The retrieval method shown includes steps S701 to S704.

[0116] Step S701: Obtain the search term input by the user, and obtain the search term representation vector corresponding to the search term based on the first sub-network of the target recall model; wherein, the target recall model is a recall model trained by the training method of the recall model provided in this application.

[0117] The retrieval method provided in this embodiment is based on vector similarity. Specifically, based on the search term representation vector of the user-input search term, it retrieves object representation vectors within the recall range that satisfy the vector similarity condition of the search term representation vector. The object represented by the object representation vector is the recalled object. This method can recall as many objects as possible that are related to the search intent represented by the user-input search term. The target recall model logically includes a search-side network structure and an object-side network structure. The first sub-network is located in the search-side network structure, and the second sub-network is located in the object-side network structure. In implementation, the search-side network structure is partially deployed online to predict search term representation vectors online; the object-side network structure is partially deployed offline to predict object representation vectors offline. New objects added to the object pool are added incrementally to the vector retrieval library containing object representation vectors. The object pool can be understood as the full set of objects within the recall range of the search system; the vector retrieval library is the range of object representation vectors found by online vector similarity retrieval based on the search representation vectors predicted online. Offline prediction of object representation vectors and incremental updates to the vector retrieval library can provide higher performance retrieval.

[0118] This step involves using a target recall model to predict the search term representation vector corresponding to the search term input by the user on the search side. This search term representation vector contains the semantic information of the search term, implying the user's search intent. In subsequent steps, this search term representation vector is used to determine the object representation vectors that are similar to it, thereby recalling the objects corresponding to the search term. In this embodiment, the first sub-network and the second sub-network have the same network structure and share the same network parameters. Their predicted search term representation vectors and object representation vectors are vectors in the same high-dimensional space; the object representation vector is a vector generated offline using the object-side network structure of the recall model.

[0119] In this embodiment, the method further includes constructing an index for retrieving object representation vectors similar to the search representation vector. Specifically, constructing the index includes: determining the object representation vector corresponding to the object to be recalled within the object candidate set based on the second sub-network of the target recall model; generating the address code of the object to be recalled based on the delivery information related to the object representation vector; performing data aggregation on each object representation vector based on the address code; and constructing an inverted index for retrieving the corresponding recalled object based on the search term.

[0120] The step of generating the address code of the object to be recalled based on the delivery information related to the object representation vector includes: obtaining the object identifier, merchant identifier, and delivery range identifier corresponding to the object representation vector; expanding the polygon corresponding to the delivery range identifier, obtaining the address code covered by the polygon, and using it as the address code of the object to be recalled; the generated address code can be understood as an address code based on the delivery range. In implementation, the code generated by the Geohash encoding method of map partitioning algorithm can be used. The Geohash encoding method encodes two-dimensional spatial latitude and longitude data, such as a map or a region obtained by its segmentation, into a string to represent it, such as using Base32 encoding. Preferably, generating the address code based on the delivery range is as follows: generating the Geohash code based on the delivery information of the object to be recalled includes: obtaining the corresponding merchant identifier based on the product identifier, obtaining the merchant's delivery circle identifier, expanding the delivery range corresponding to the merchant's delivery circle identifier into a polygon, enumerating the address codes covered by the polygon, and using it as the address code generated based on the delivery range of the object to be recalled.

[0121] The process of aggregating object representation vectors based on address encoding involves organizing these vectors using the address encoding as a key. Specifically, this includes: aggregating the object representation vectors based on a linear index when the number of vectors corresponding to the same address encoding is less than a threshold; otherwise, aggregating the vectors based on hierarchical clustering. A linear index organizes the set of index items into a linear structure. Hierarchical clustering divides a set of data into several categories based on the similarity and differences of the index item data. Data within the same category exhibits high similarity, while data between different categories shows low similarity and low cross-category correlation. In implementation, a threshold for the number of objects can be preset, such as 10,000. If the number of items (i.e., the number of object vectors) is less than 10,000, linear indexing is used; otherwise, hierarchical clustering is performed. The hierarchical clustering process includes: initially assigning each object representation vector to a first class, calculating the distance between every two first classes, where the distance represents the similarity between the two vectors; finding the two closest first classes among all classes and assigning them to a second class, thus reducing the total number of classes by one; recalculating the similarity between the newly assigned second class and each first class; and so on, to achieve the classification of all object representation vectors associated with an address encoding.

[0122] The construction of the inverted index for retrieving corresponding recall objects based on search terms includes: forming a first index term (first-level key) based on the product identifier, merchant identifier, and merchant's delivery range identifier to organize each object representation vector into a Geohash generated based on the delivery range; using the Geohash as a second index term (second-level key) to organize each object representation vector, the constructed two-level index is the inverted index. The two-level index structure of delivery range -> Geohash -> object representation vector (i.e., merchant product) hierarchical clustering is an example of the inverted index constructed in this application, in order to reduce index redundancy. Preferably, the inverted index is a two-level index, including a first-level inverted index for indexing merchants to their corresponding address encoding set (as described above, delivery range -> Geohash), and a second-level inverted index for indexing addresses to their corresponding object representation vector set (as described above, Geohash -> object representation vector).

[0123] Step S702: Determine the merchants whose delivery range covers the user's location based on the user's location, and obtain the address code of the merchant's delivery range coverage.

[0124] This step involves determining one or more merchants (such as all merchants corresponding to an object pool) based on the user's location, thereby obtaining the address codes covered by each merchant. The user's location refers to a location selected by the user or a location where the user allows it. Obtaining the address codes covering the merchant's delivery range includes retrieving at least one address code covering the merchant's delivery range based on the first-level inverted index.

[0125] Step S703: Obtain the object representation vector set associated with the address encoding, and determine several nearest neighbor vectors of the search term representation vector in the object representation vector set.

[0126] This embodiment employs ANN (Approximate Nearest Neighbor) retrieval to accelerate the search process. This step involves determining several nearest neighbor vectors for the search term representation vector. Specifically, obtaining the object representation vector set associated with the address code includes retrieving the object representation vector set corresponding to the address code based on the second-level inverted index. Determining several nearest neighbor vectors for the search term representation vector in the object representation vector set includes: if the object representation vector set is generated based on hierarchical clustering, determining the centroid layer by layer, traversing the next-level nodes contained in each centroid, until the similarity between the vector corresponding to each node and the search term representation vector is linearly compared, and selecting nodes whose similarity to the search term representation vector satisfies the vector similarity condition. The vectors corresponding to the selected nodes are then used as the nearest neighbor vectors. The process of determining the center point layer by layer, traversing the next-level nodes contained in each center point, until the similarity between the vectors corresponding to each node and the vector representing the search term is linearly compared, includes: determining at least one first-level center point; for each first-level center point, traversing the second-level center points contained in that first-level center point; determining the nodes contained in each second-level center point; traversing the nodes contained in each second-level center point; and linearly comparing the similarity between the vectors corresponding to each node and the vector representing the search term. In implementation, for each Geohash object representation vector set, a BBF (Best Bin First) search strategy is used to first find the nearest few first-level center points, then sequentially traverse and examine their second-level center points, and finally linearly compare all nodes (i.e., all vectors) under these second-level center points to complete the selection of the top-k nearest neighbor vectors. Please refer to [reference needed]. Figure 8 The diagram illustrates a retrieval method based on hierarchical clustering. The diagram exemplifies two layers of centroids, but it is not limited to two layers. Primary centroids include C0, C1, etc. Taking C0 as an example, it contains secondary centroids. Each secondary centroid contains one or more nodes aggregated into that category, until each node is an object representation vector. The top-K nearest neighbors of the query point are selected. The query point represents the search representation vector that needs to be retrieved for the corresponding object representation vector.

[0127] Step S704: Determine the object representation vector retrieved for the search term representation vector based on the nearest neighbor vector, and use the object corresponding to the retrieved object representation vector as the recall object retrieved for the search term; the distance between the nearest neighbor vector and the search term representation vector satisfies the vector similarity condition.

[0128] This step involves determining the retrieved object representation vector, where several nearest neighbor vectors determined in the previous steps can be used as the retrieved object representation vector. Alternatively, these nearest neighbor vectors can be further filtered. Determining the retrieved object representation vector based on the nearest neighbor vectors for the search term representation vector includes: determining the delivery range of the nearest neighbor vectors; the delivery range being a polygon; drawing a horizontal ray from the user's location as the starting point; if the number of intersections between the ray and each edge of the polygon is odd, then the delivery range covers the user's location; otherwise, the delivery range does not cover the user's location; removing vectors from the nearest neighbor vectors whose delivery range does not cover the user's location, and using the remaining nearest neighbor vectors as the retrieved object representation vector. This filters out merchants whose delivery range does not include the user's location from the several nearest neighbor vectors, removing the influence of spatial constraints on the retrieved recall object set, and achieving recall accuracy from a global search perspective.

[0129] In this embodiment, the method further includes: using the object corresponding to the retrieved object representation vector as the first recall object retrieved for the search term; obtaining the second recall object that matches the search term based on text matching; and re-ranking the recall results for the first recall object and the second recall object, and using the re-ranked result as a candidate set of search results provided to the user.

[0130] The retrieval method provided in this embodiment improves the average depth of object representation vector recall, while significantly reducing the scene RT99 line (response time percentile), resulting in high recall rate and retrieval efficiency.

[0131] The retrieval method provided in this embodiment can meet diverse and precise search needs, and can be applied to search services across multiple platforms and industries. Furthermore, its target recall model fully mines semantics, learns the influence between search term neighbors, and introduces information interaction between the search side and the target side, thus exhibiting strong generalization ability and achieving good recall and precision even in scenarios with forced constraints based on spatiotemporal attributes. In addition, the two-level index structure based on delivery range facilitates the removal of information corresponding to merchants not covered by the delivery range based on user location, thereby providing efficient and accurate retrieval services and overcoming the impact of spatial constraints on recall accuracy.

[0132] The method provided in this embodiment has been described above. The method uses the training method of the recall model provided in this application to train a target recall model. The first and second sub-networks of the target recall model are used to obtain the search term representation vector and the object representation vector, respectively. Based on the similarity between the search term vector and the object representation vector, the method recalls the recall object for the search term. This can find the recall object corresponding to the search intent reflected by the user's input search term as much as possible, thereby improving the object recall rate and recall precision.

[0133] Corresponding to the first embodiment, the fourth embodiment of this application provides a training apparatus for a recall model; for relevant parts, please refer to the description of the corresponding method embodiment. Figure 9 The training device for the recall model shown in the figure includes:

[0134] The search-side network unit 901 is used to take the search-side information of the training samples as the first input information of the recall model to be trained, input the first sub-network of the recall model to obtain the first output information, and determine the search word features of the training samples based on the first output information.

[0135] The object-side network unit 902 is used to take the object-side information of the training samples as the second input information of the recall model, input it into the second sub-network of the recall model to obtain the second output information, and determine the recall object features of the training samples based on the second output information.

[0136] Training unit 903 is used to train the recall model based on a first loss function, and obtain a target recall model after training; wherein, the first loss function is constructed based on the similarity between the search term features of the training samples and the corresponding recall object features, and is used as the target function for training the recall model.

[0137] Optionally, the search-side network unit 901 is specifically used for: acquiring the neighbor information of the search terms of the training samples; acquiring the search term neighbor features of the search terms based on the neighbor information; determining the graph attention features related to the search terms and the neighbor information based on the search term neighbor features and the first output information; determining a second loss function based on the graph attention features and the search term features; and using the second loss function to train the recall model.

[0138] Optionally, the recall model further includes a third sub-network, which is a search-side sub-network used to obtain search term neighbor features based on the neighbor information of the search terms in the training samples; the search-side network unit 901 is specifically used to: input the neighbor information into the third sub-network to obtain the search term neighbor features.

[0139] Optionally, the third sub-network has the same network structure and / or shares parameters with the first sub-network.

[0140] Optionally, the training unit 903 is specifically used to: use the first loss function and the second loss function as target functions for training the recall model, and train the recall model.

[0141] Optionally, the search-side network unit 901 is specifically configured to: acquire search logs; generate a click relationship graph representing the relationship between search terms and their recall objects based on the search logs, wherein each node in the click relationship graph is a search term node corresponding to the search term information, and a click object node corresponding to each search term that has been clicked by the user; generate a bipartite graph based on neighbor aggregation of each search term node in the click relationship graph, and use the generated bipartite graph as the search click graph; sample the neighbor nodes related to the search term in the search click graph based on the search term, and use the sampled data as the neighbor information.

[0142] Optionally, the search-side network unit 901 is specifically used to: input the first output information into the first multilayer perceptron to obtain the search term features, wherein the search term features include search term semantic information.

[0143] Optionally, the object-side network unit 902 is specifically used to: acquire click data of the recall object corresponding to the search terms of the training sample; wherein, the click data includes one or more search terms for the recall object whose click rate meets a click rate threshold or whose click quantity meets a click quantity threshold; and generate the second input information based on the click data and the object information of the recall object.

[0144] Optionally, the object-side network unit 902 is specifically used to: input the second output information into the second multilayer perceptron to obtain the recall object features, wherein the recall object features contain semantic information of the recall object corresponding to the search term.

[0145] Optionally, the first sub-network and the second sub-network have the same network structure and / or share parameters.

[0146] Optionally, the training unit 903 is specifically used to: construct the first loss function based on the similarity between the search term features of the training samples and the corresponding recall object features; the first loss function adopts one of the following loss functions: infoNCE loss function, Hinge loss function, Triplet loss function, Circle loss function.

[0147] Optionally, the device further includes a sample construction unit, which is configured to: obtain search terms input by the user on the search side and recall objects corresponding to the search terms from the search log; determine the recall objects that have been exposed and clicked among the recall objects; construct positive samples based on the search terms and the exposed and clicked recall objects; generate negative samples based on the positive samples; and use the positive samples and / or the negative samples as the training samples.

[0148] Corresponding to the second embodiment, the fifth embodiment of this application provides a search and recall device; for relevant parts, please refer to the description of the corresponding method embodiment. Figure 10 The search and recall device shown in the figure includes:

[0149] The search term representation unit 1001 is used to acquire the search term input by the user, input the search term into the first sub-network of the target recall model, and obtain the search term representation vector corresponding to the search term; wherein, the target recall model is a recall model trained using the training method of the recall model;

[0150] The object representation unit 1002 is used to acquire object information of the object to be recalled, input the object information into the second sub-network of the target recall model, and obtain the object representation vector corresponding to the object information.

[0151] The recall unit 1003 is used to determine the recall object corresponding to the search term based on the similarity between the search term representation vector and the object representation vector.

[0152] Optionally, the search term representation unit 1001 is specifically used to: input the search term into the first sub-network to obtain a first prediction output; input the first prediction output into the first multilayer perceptron of the target recall model, and use the output of the first multilayer perceptron as the search term representation vector.

[0153] Optionally, the object representation unit 1002 is specifically used to: input the object information into the second sub-network to obtain a second prediction output; input the second prediction output into the second multilayer perceptron of the target recall model, and use the output of the second multilayer perceptron as the object representation vector.

[0154] Corresponding to the third embodiment, the sixth embodiment of this application provides a retrieval device; for related parts, please refer to the description of the corresponding method embodiment. Figure 11 The retrieval device shown in the figure includes:

[0155] The search term representation unit 1101 is used to acquire the search term input by the user and obtain the search term representation vector corresponding to the search term based on the first sub-network of the target recall model; wherein, the target recall model is a recall model trained using the training method of the recall model;

[0156] The address encoding unit 1102 is used to determine the merchants whose delivery range covers the user's location based on the user's location, and obtain the address encoding of the merchant's delivery range.

[0157] The nearest neighbor vector unit 1103 is used to obtain the object representation vector set associated with the address code and determine several nearest neighbor vectors of the search term representation vector in the object representation vector set;

[0158] The recall unit 1104 is used to determine the object representation vector retrieved for the search term representation vector based on the nearest neighbor vector, and to take the object corresponding to the retrieved object representation vector as the recall object retrieved for the search term; the distance between the nearest neighbor vector and the search term representation vector satisfies the vector similarity condition.

[0159] Optionally, the apparatus further includes an index building unit, which is configured to: determine the object representation vector corresponding to the object to be recalled in the object candidate set based on the second sub-network of the target recall model; generate the address code of the object to be recalled based on the delivery information related to the object representation vector; perform data aggregation on each object representation vector based on the address code; and construct an inverted index for retrieving the corresponding recalled object based on search terms.

[0160] Optionally, the inverted index is a two-level index, including a first-level inverted index for indexing merchants to their corresponding address code sets, and a second-level inverted index for indexing address codes to their corresponding object representation vector sets.

[0161] Optionally, the retrieval address encoding unit 1102 is specifically used to: retrieve at least one address encoding that covers the delivery range of the merchant based on the first-level inverted index.

[0162] Optionally, the nearest neighbor vector unit 1103 is specifically used to: retrieve the object representation vector set corresponding to the address code based on the second-level inverted index.

[0163] Optionally, the index construction unit is specifically used to: obtain the object identifier, merchant identifier, and delivery range identifier corresponding to the object representation vector; expand the polygon corresponding to the delivery range identifier, and obtain the address code covered by the polygon as the address code of the object to be recalled.

[0164] Optionally, the index building unit is specifically used to: aggregate the object representation vectors based on a linear index when the number of object representation vectors corresponding to the same address encoding is less than the object number threshold; otherwise, aggregate the object representation vectors based on hierarchical clustering.

[0165] Optionally, the nearest neighbor vector retrieval unit 1103 is specifically used for: if the object representation vector set is generated based on hierarchical clustering, then determining the center point layer by layer, traversing the next-level nodes contained in each layer's center point, until the similarity between the vector corresponding to each node and the search term representation vector is linearly compared, filtering out nodes whose similarity to the search term representation vector satisfies the vector similarity condition, and using the vector corresponding to the filtered node as the nearest neighbor vector.

[0166] Optionally, the nearest neighbor vector retrieval unit 1103 is specifically used to determine at least one primary centroid, and for each primary centroid, traverse the secondary centroids contained in that primary centroid; determine the nodes contained in each secondary centroid, traverse the nodes contained in each secondary centroid, and linearly compare the similarity between the vector corresponding to each node and the vector representing the search term.

[0167] Optionally, the recall unit 1104 is specifically used to: determine the delivery range of the nearest neighbor vector; the delivery range is a polygon; draw a horizontal ray with the user's location as the starting point; if it is determined that the number of intersections between the ray and each side of the polygon is odd, then the delivery range covers the user's location; otherwise, the delivery range does not cover the user's location.

[0168] Vectors whose delivery range does not cover the user's location are removed from the nearest neighbor vectors, and the remaining nearest neighbor vectors are used as the retrieved object representation vectors.

[0169] Optionally, the search term representation vector and the object representation vector are vectors in the same high-dimensional space; the object representation vector is a vector generated offline using the object-side network structure of the recall model.

[0170] Based on the above embodiments, the seventh embodiment of this application provides an electronic device. For relevant parts, please refer to the corresponding descriptions in the above embodiments. Figure 12 The electronic device shown in the figure includes a memory 1201 and a processor 1202; the memory is used to store a computer program, which, after being run by the processor, executes the method provided in the embodiments of this application.

[0171] Based on the above embodiments, the eighth embodiment of this application provides a computer storage medium. For relevant parts, please refer to the corresponding descriptions in the above embodiments. The schematic diagram of the computer storage medium is similar. Figure 12The memory in the figure can be understood as the storage medium. The computer storage medium stores computer execution instructions, which, when executed by a processor, are used to implement the method provided in the embodiments of this application.

[0172] In a typical configuration, an electronic device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory. Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0173] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0174] Although this application discloses preferred embodiments as described above, it is not intended to limit this application. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of this application. Therefore, the scope of protection of this application should be determined by the scope defined in the claims of this application.

Claims

1. A training method for a recall model, characterized in that, include: The search-side information of the training samples is used as the first input information of the recall model to be trained. The first sub-network of the recall model is input to obtain the first output information. The search word features of the training samples are determined based on the first output information. The object-side information of the training samples is used as the second input information of the recall model, and the second sub-network of the recall model is input to obtain the second output information. The recall object features of the training samples are determined based on the second output information. The neighbor information of the search terms in the training samples is input into the third sub-network of the recall model. The search term neighbor features of the search terms are obtained through the third sub-network, and the graph attention features related to the search terms and the neighbor information are determined based on the search term neighbor features and the first output information. The neighbor information of the search terms is obtained by sampling the search click graph data corresponding to the search terms. The search click graph data is generated by aggregating the search term nodes with neighbor relationships in the click relationship between the search terms and the recall objects. The recall model is trained based on a first loss function and a second loss function, and a target recall model is obtained after training. The first loss function is constructed based on the similarity between the search term features of the training samples and the corresponding recall object features, and the second loss function is determined based on the search term features and the graph attention features. The first loss function and the second loss function are used as the target functions for training the recall model. This also includes: The search terms input by the user are obtained, and the search term representation vector corresponding to the search terms is obtained based on the first sub-network of the target recall model; Based on the user's location, determine the merchants whose delivery range covers the user's location, and obtain the address codes of the merchants whose delivery range covers the location; Obtain the object representation vector set associated with the address encoding, and determine several nearest neighbor vectors of the search term representation vector in the object representation vector set; Based on the nearest neighbor vector, the object representation vector retrieved for the search term representation vector is determined, and the object corresponding to the retrieved object representation vector is used as the recall object retrieved for the search term; the distance between the nearest neighbor vector and the search term representation vector satisfies the vector similarity condition.

2. The method according to claim 1, characterized in that, The third subnetwork has the same network structure and / or shares parameters with the first subnetwork.

3. The method according to claim 1, characterized in that, Also includes: Obtain search logs and generate a click relationship graph based on the search logs to represent the relationship between search terms and the recall objects of the search terms. Each node of the click relationship graph is a search term node corresponding to the search term information and a click object node corresponding to each search term that has been clicked by the user. A bipartite graph is generated by aggregating the neighbors of each search term node in the click relationship graph, and the generated bipartite graph is used as the search click graph. Based on the search term, neighboring nodes related to the search term in the search click graph are sampled, and the sampled data is used as the neighbor information of the search term in the training sample.

4. The method according to claim 1, characterized in that, Determining the search term features of the training samples based on the first output information includes: The first output information is input into the first multilayer perceptron to obtain the search term features, which contain semantic information of the search term.

5. The method according to claim 1, characterized in that, Also includes: Obtain click data of the recall objects corresponding to the search terms of the training samples; wherein, the click data includes one or more search terms for the recall objects whose click rate meets a click rate threshold or whose click quantity meets a click quantity threshold; The second input information is generated based on the click data and the object information of the recalled object.

6. The method according to claim 1 or 5, characterized in that, The step of determining the recall object features of the training samples based on the second output information includes: The second output information is input into the second multilayer perceptron to obtain the recall object features, which contain semantic information of the recall object corresponding to the search term.

7. The method according to claim 1, characterized in that, The first sub-network and the second sub-network have the same network structure and / or share parameters.

8. The method according to claim 1, characterized in that, Also includes: The first loss function is constructed based on the similarity between the search term features of the training samples and the corresponding recall object features; The first loss function adopts one of the following loss functions: infoNCE loss function, Hinge loss function, Triplet loss function, and Circle loss function.

9. The method according to claim 1, characterized in that, Also includes: Obtain the search terms entered by the user on the search side from the search logs and the recall objects corresponding to the search terms, and determine the recall objects that have been exposed and clicked from the recall objects; Construct positive samples based on the search terms and the exposed and clicked recall objects; Negative samples are generated based on the positive samples, and the positive samples and / or the negative samples are used as the training samples.

10. The method according to claim 1, characterized in that, Also includes: Based on the second sub-network of the target recall model, determine the object representation vector corresponding to the object to be recalled in the object candidate set; The address code of the object to be recalled is generated based on the delivery information related to the object representation vector; Data aggregation is performed on the representation vectors of each object based on the address encoding; Construct an inverted index for retrieving corresponding recall objects based on search terms.

11. The method according to claim 10, characterized in that, The inverted index is a two-level index, including a first-level inverted index for indexing merchants to their corresponding address code sets, and a second-level inverted index for indexing address codes to their corresponding object representation vector sets.

12. The method according to claim 11, characterized in that, The step of obtaining the address codes covering the merchant's delivery range includes: At least one address code covering the delivery range of the merchant is retrieved based on the first-level inverted index.

13. The method according to claim 11, characterized in that, The step of obtaining the object representation vector set associated with the address encoding includes: The object representation vector set corresponding to the address code is retrieved based on the second-level inverted index.

14. The method according to claim 10, characterized in that, The step of generating the address code of the object to be recalled based on the delivery information related to the object representation vector includes: Obtain the object identifier, merchant identifier, and delivery range identifier corresponding to the object representation vector; Expand the polygon corresponding to the delivery range identifier, obtain the address code covered by the polygon, and use it as the address code of the object to be recalled.

15. The method according to claim 10, characterized in that, The data aggregation based on the address encoding of each object representation vector includes: When the number of object representation vectors corresponding to the same address encoding is less than the object number threshold, the object representation vectors are aggregated based on linear indexing; otherwise, the object representation vectors are aggregated based on hierarchical clustering.

16. The method according to claim 15, characterized in that, Determining several nearest neighbor vectors of the search term representation vector in the object representation vector set includes: If the object representation vector set is generated based on hierarchical clustering, then the center point is determined layer by layer, and the next level nodes contained in each layer center point are traversed until the similarity between the vector corresponding to each node and the search term representation vector is linearly compared. Nodes whose similarity with the search term representation vector satisfies the vector similarity condition are selected, and the vectors corresponding to the selected nodes are taken as the nearest neighbor vectors.

17. The method according to claim 16, characterized in that, The process of determining the center point layer by layer, traversing the next-level nodes contained in each center point, until a linear comparison of the similarity between the vector corresponding to each node and the vector representing the search term, includes: Determine at least one primary center point, and for each primary center point, traverse the secondary center points contained within that primary center point; Determine the nodes contained in each secondary centroid, traverse the nodes contained in the secondary centroids, and linearly compare the similarity between the vector corresponding to each node and the vector representing the search term.

18. The method according to claim 1, characterized in that, Determining the object representation vector retrieved for the search term representation vector based on the nearest neighbor vector includes: Determine the delivery range of the nearest neighbor vector; the delivery range is a polygon. Draw a horizontal ray starting from the user's location; If the number of intersections between the ray and each edge of the polygon is odd, then the delivery range covers the user's location; otherwise, the delivery range does not cover the user's location. Vectors whose delivery range does not cover the user's location are removed from the nearest neighbor vectors, and the remaining nearest neighbor vectors are used as the retrieved object representation vectors.

19. The method according to claim 1, characterized in that, The search term representation vector and the object representation vector are vectors in the same high-dimensional space; the object representation vector is a vector generated offline using the object-side network structure of the recall model.

20. A search recall method, characterized in that, include: The user inputs a search term, and the search term is input into the first sub-network of the target recall model to obtain a search term representation vector corresponding to the search term; wherein the target recall model is a recall model trained by the method of any one of claims 1 to 9; Obtain object information of the object to be recalled, input the object information into the second sub-network of the target recall model, and obtain the object representation vector corresponding to the object information; Based on the similarity between the search term representation vector and the object representation vector, the recall object corresponding to the search term is determined.

21. The method according to claim 20, characterized in that, The step of inputting the search term into the first sub-network of the target recall model to obtain the search term representation vector corresponding to the search term includes: The search term is input into the first sub-network to obtain the first prediction output; The first prediction output is input into the first multilayer perceptron of the target recall model, and the output of the first multilayer perceptron is used as the search term representation vector.

22. The method according to claim 20, characterized in that, The step of inputting the object information into the second sub-network of the target recall model to obtain the object representation vector corresponding to the object information includes: The object information is input into the second sub-network to obtain the second prediction output; The second prediction output is input into the second multilayer perceptron of the target recall model, and the output of the second multilayer perceptron is used as the object representation vector.

23. A training device for a recall model, characterized in that, include: The search-side network unit is used to take the search-side information of the training samples as the first input information of the recall model to be trained, input the first sub-network of the recall model to obtain the first output information, and determine the search word features of the training samples based on the first output information. The object-side network unit is used to take the object-side information of the training samples as the second input information of the recall model, input it into the second sub-network of the recall model to obtain the second output information, and determine the recall object features of the training samples based on the second output information. The training unit is used to train the recall model based on a first loss function and a second loss function, and obtain the target recall model after training; wherein, the first loss function is constructed based on the similarity between the search term features of the training samples and the corresponding recall object features, the second loss function is determined according to the search term features and graph attention features, and the first loss function and the second loss function are used as the target functions for training the recall model; The method further includes: inputting the neighbor information of the search terms of the training samples into the third sub-network of the recall model, obtaining the search term neighbor features of the search terms through the third sub-network, and determining the graph attention features related to the search terms and the neighbor information based on the search term neighbor features and the first output information; wherein, the neighbor information of the search terms is obtained by sampling the search click graph data corresponding to the search terms, and the search click graph data is generated by aggregating the search term nodes with neighbor relationships in the click relationship between the search terms and the recall objects; The device further includes: The search term representation unit is used to obtain the search term input by the user and obtain the search term representation vector corresponding to the search term based on the first sub-network of the target recall model; The address encoding unit is used to determine the merchants whose delivery range covers the user's location based on the user's location, and to obtain the address encoding of the merchant's delivery range. The nearest neighbor vector unit is used to obtain the object representation vector set associated with the address code and determine several nearest neighbor vectors of the search term representation vector in the object representation vector set; The recall unit is used to determine the object representation vector retrieved for the search term representation vector based on the nearest neighbor vector, and to take the object corresponding to the retrieved object representation vector as the recall object retrieved for the search term; the distance between the nearest neighbor vector and the search term representation vector satisfies the vector similarity condition.

24. A search and recall device, characterized in that, include: The search term representation unit is used to acquire the search term input by the user, input the search term into the first sub-network of the target recall model, and obtain the search term representation vector corresponding to the search term; wherein, the target recall model is a recall model trained by the method of any one of claims 1 to 9; An object representation unit is used to acquire object information of the object to be recalled, input the object information into the second sub-network of the target recall model, and obtain an object representation vector corresponding to the object information. The recall unit is used to determine the recall object corresponding to the search term based on the similarity between the search term representation vector and the object representation vector.

25. An electronic device, characterized in that, include: A memory and a processor; the memory is used to store a computer program, which, when executed by the processor, performs the method according to any one of claims 1-22.

26. A computer storage medium, characterized in that, The device stores computer execution instructions, which, when executed by a processor, are used to implement the method described in any one of claims 1-22.

Images