A search request recommendation method and device, electronic equipment and storage medium

Abstract

The application relates to the computer technical field, and in particular to a search request recommendation method and device, electronic equipment and a storage medium, to improve the recommendation accuracy of a search request. The method comprises the following steps: obtaining first interaction representation information based on the interest label of a target object and a current target search request; based on the first interaction representation information, filtering a plurality of historical search requests with a similarity to the target search request reaching a first threshold from a pre-constructed historical search request set as candidate search requests; determining the probability of each candidate search request as the next search request of the target object based on each candidate search request and each historical search request related to the target object; and based on the probabilities, ranking each candidate search request, and recommending a search request to the target object based on the ranking result. The application combines the interest label of the target object and the historical search request to recommend a search request to the target object, and the recommendation accuracy can be improved.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method, apparatus, electronic device, and storage medium for recommending search requests. Background Technology

[0002] With the rapid development of internet technology, internet services are constantly increasing, and the types of data on the internet are also constantly increasing. In addition to text, internet data types also include video. Users can download and install various information applications to watch videos through terminal devices such as smartphones and tablets.

[0003] Specifically, users can enter search terms on the application interface to obtain the videos they want to watch. In the search results page, in addition to showing users some videos related to the search terms, other search terms that users may search for can also be provided, making it easier for users to search for other related content.

[0004] In related technologies, video-related search is mainly implemented based on the current search query. It retrieves other queries related to the current search query from the platform's historical search query library as candidates for related searches. However, this approach relies on filtering historical searches of a large number of objects on the platform, resulting in low relevance to the current search object and thus failing to adequately meet the search needs of the target audience. Summary of the Invention

[0005] This application provides a search request recommendation method, apparatus, electronic device, and storage medium to improve the accuracy of search request recommendations.

[0006] This application provides a search request recommendation method, including:

[0007] Based on the target object's interest tags and the current target search request, obtain the first interactive representation information corresponding to the target search request;

[0008] Based on the first interaction representation information corresponding to the target search request, multiple historical search requests that have a similarity to the target search request reaching a first threshold are selected from the pre-constructed set of historical search requests as candidate search requests related to the target object.

[0009] Based on each candidate search request and each historical search request related to the target object, the probability of the corresponding candidate search request being the next search request for the target object is determined.

[0010] After ranking each candidate search request based on its probability, a search request recommendation is made to the target object based on the ranking result.

[0011] This application provides a search request recommendation device, comprising:

[0012] An interactive representation unit is used to obtain first interactive representation information corresponding to the target search request based on the interest tags of the target object and the current target search request;

[0013] The filtering unit is used to filter out multiple historical search requests that have a similarity to the target search request reaching a first threshold from a pre-built set of historical search requests based on the first interaction representation information corresponding to the target search request, as candidate search requests related to the target object;

[0014] The determining unit is used to determine the probability of a corresponding candidate search request being the next search request for the target object based on each candidate search request and each historical search request related to the target object.

[0015] The recommendation unit is used to rank each candidate search request based on its probability and then recommend search requests to the target object based on the ranking results.

[0016] Optionally, the determining unit is specifically used for:

[0017] Based on the interest tags of the target object and each historical search request related to the target object, first interaction representation information corresponding to each historical search request is obtained;

[0018] For each candidate search request, perform the following operations:

[0019] Based on the interest tags of the target object and a candidate search request, obtain the first interaction representation information corresponding to the candidate search request;

[0020] The first interaction representation information corresponding to the candidate search request and the first interaction representation information corresponding to each of the historical search requests are respectively input into the trained search time sequence prediction model to obtain the probability that the candidate search request is the next search request for the target object.

[0021] Optionally, the device further includes:

[0022] The first training unit is used to train the search time-series prediction model in the following manner:

[0023] Obtain a first training sample set, wherein each first training sample in the first training sample set includes: a sample search request for a sample object and multiple historical search requests, and the sample search request as the true probability of the next search request for the sample object;

[0024] Based on the first training sample set, the search time-series prediction model to be trained is trained iteratively multiple times to obtain a trained search time-series prediction model; wherein, in one iteration, the following operations are performed:

[0025] The selected first training samples are input into the search time-series prediction model to be trained, and the corresponding sample search requests of the sample objects are output as the predicted probability of the next search request of the sample objects.

[0026] The parameters of the search time series prediction model are adjusted based on the difference between the predicted probability and the corresponding true probability.

[0027] Optionally, the recommendation unit is further configured to:

[0028] After sorting each candidate search request based on each probability, and before recommending a search request to the target object based on the sorting result, the first interaction representation information corresponding to each candidate search request is obtained based on the interest tags of the target object and each candidate search request.

[0029] According to the sorting result, each candidate search request is traversed sequentially, and the candidate search request with the lower probability is deduplicated among the two candidate search requests whose similarity between the corresponding first interaction representation information reaches the second threshold.

[0030] The recommendation unit is specifically used for:

[0031] Among the remaining candidate search requests after deduplication, at least one candidate search request whose probability reaches the third threshold is recommended to the target object.

[0032] Optionally, the first interaction representation information corresponding to each search request can be obtained in the following ways, where the search request is a target search request, a candidate search request, or a historical search request;

[0033] For each search request, perform the following operations:

[0034] The interest tags of the object corresponding to a search request and the search request are respectively input into the trained joint representation model to obtain the first deep representation information corresponding to the interest tags of the search object and the second deep representation information corresponding to the search request.

[0035] The first depth representation information corresponding to the interest tag of the search object and the second depth representation information corresponding to the search request are subjected to feature interaction processing to obtain the first interaction representation information corresponding to the search request.

[0036] Optionally, the device further includes:

[0037] The second training unit is used to train the joint representation model in the following manner:

[0038] Obtain a second training sample set, wherein each second training sample in the second training sample set includes: an interest tag and a sample search request of a sample object, relevant search results of the sample object for the sample search request, and real information on the sample object's satisfaction with the relevant search results;

[0039] Based on the second training sample set, the joint representation model to be trained is trained iteratively multiple times to obtain a trained joint representation model; wherein, in one iteration, the following operations are performed:

[0040] The selected second training samples are input into the joint representation model, and the corresponding sample object's satisfaction prediction information for the relevant search results is output.

[0041] The parameters of the joint representation model are adjusted based on the differences between the predicted satisfaction information and the actual satisfaction information of each search result.

[0042] Optionally, the true satisfaction information of the sample objects with the relevant search results is determined based on the feedback behavior of the sample objects towards the relevant search results.

[0043] Optionally, the second training unit is specifically used for:

[0044] The second training samples selected from the second training sample set are respectively input into the joint representation model to be trained to obtain the first depth representation information corresponding to the interest tags of the sample objects in the second training samples, the second depth representation information corresponding to the sample search requests of the sample objects, and the third depth representation information corresponding to the relevant search results of the sample search requests.

[0045] Based on the first depth representation information corresponding to the interest tags of the sample object, the second depth representation information corresponding to the search request of the sample object, and the third depth representation information, the sample object is predicted to obtain the satisfaction prediction information for the sample object.

[0046] Optionally, the second training unit is specifically used for:

[0047] The first depth representation information corresponding to the interest tag of the sample object and the second depth representation information corresponding to the search request of the sample object are subjected to feature interaction processing to obtain the first interaction representation information corresponding to the sample object; and the second depth representation information corresponding to the search request of the sample object and the third depth representation information corresponding to the relevant search results of the sample search request are subjected to feature interaction processing to obtain the second interaction representation information corresponding to the sample object.

[0048] Based on the first and second interaction representation information of the sample object, a prediction is made to obtain the satisfaction prediction information for the sample object.

[0049] An electronic device provided in this application includes a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of any of the search request recommendation methods described above.

[0050] This application provides a computer-readable storage medium including a computer program. When the computer program is run on an electronic device, the computer program is used to cause the electronic device to perform the steps of any of the search request recommendation methods described above.

[0051] This application provides a computer program product, which includes a computer program stored in a computer-readable storage medium. When a processor of an electronic device reads the computer program from the computer-readable storage medium, the processor executes the computer program, causing the electronic device to perform the steps of any of the search request recommendation methods described above.

[0052] The beneficial effects of this application are as follows:

[0053] This application provides a search request recommendation method, apparatus, electronic device, and storage medium. Because this application obtains first interaction representation information corresponding to the target search request based on the target object's interest tags and the current target search request; based on this first interaction representation information, the selected candidate search requests are more in line with the object's next search needs. Furthermore, this application combines the target object's interest tags and historical search requests to estimate the probability that each candidate search request will be the target object's next search request, and recommends search requests to the target object based on this probability. This satisfies the object's next search request, eliminates the object's search input, speeds up the object's discovery of desired video content, and improves the accuracy of search request recommendations.

[0054] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0055] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0056] Figure 1 This is an optional schematic diagram of an application scenario in an embodiment of this application;

[0057] Figure 2 This is an optional schematic diagram of a search interface in an embodiment of this application;

[0058] Figure 3 This is a flowchart illustrating the implementation of a search request recommendation method in an embodiment of this application.

[0059] Figure 4 This is a schematic diagram of a training method for a joint representation model in an embodiment of this application;

[0060] Figure 5 This is a schematic diagram of the structure of a joint representation model in an embodiment of this application;

[0061] Figure 6 This is a schematic diagram of a relevant search and retrieval process in one embodiment of this application;

[0062] Figure 7 This is a schematic diagram of the structure of a search time series prediction model in an embodiment of this application;

[0063] Figure 8 This is a schematic diagram illustrating a training method for a search time series prediction model in an embodiment of this application;

[0064] Figure 9A This is a simplified flowchart of an embodiment of this application;

[0065] Figure 9B This is a general logic diagram of an embodiment of this application;

[0066] Figure 9C This is a schematic diagram of a search process in an embodiment of this application;

[0067] Figure 9D This is a schematic diagram illustrating the training and usage process of a model in one embodiment of this application;

[0068] Figure 10 This is a schematic diagram of the composition structure of a search request recommendation device according to an embodiment of this application;

[0069] Figure 11 This is a schematic diagram of the hardware structure of an electronic device using an embodiment of this application;

[0070] Figure 12 This is a schematic diagram of the hardware structure of another electronic device using an embodiment of this application. Detailed Implementation

[0071] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this application. Obviously, the described embodiments are only some embodiments of the technical solutions of this application, and not all embodiments. Based on the embodiments recorded in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the technical solutions of this application.

[0072] The following describes some of the concepts involved in the embodiments of this application.

[0073] Video-related searches: When a user searches on a video site, the search results page provides other search terms that the user might have searched for, making it easier for them to find other related content, such as... Figure 2 As shown, the related searches section also recommends other search options for users to quickly select.

[0074] Interactive representation information: obtained by interactively representing at least two types of feature information. In this application embodiment, two types of interactive representation information are listed: a first interactive representation information obtained by interactively representing an object's interest tags and its search request; and a second interactive representation information obtained by interactively representing an object's search request and search results.

[0075] Faiss (Facebook AI Similarity Search) is a library for efficient similarity search and dense vector clustering. It provides efficient similarity search and clustering for dense vectors, supporting searches on the order of billions of vectors. It is an approximate nearest neighbor search library. It includes various algorithms for searching vector sets of arbitrary sizes, as well as computer programs for algorithm evaluation and parameter tuning. The size of the vector set is determined by the random-access memory (RAM).

[0076] Similarity retrieval index: An index used for similarity retrieval. In this embodiment, the index is obtained by vector compression of the first interaction information corresponding to the search request. One index uniquely identifies a search request and its corresponding first interaction information. In this embodiment, the similarity retrieval index can efficiently and quickly retrieve multiple candidate search requests related to the target search request from the historical search request set.

[0077] The embodiments of this application relate to artificial intelligence (AI), NLP (Natural Language Processing), and machine learning technologies, and are designed based on computer vision technology and machine learning (ML) in artificial intelligence.

[0078] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence.

[0079] Artificial intelligence (AI) studies the design principles and implementation methods of various intelligent machines, enabling them to perceive, reason, and make decisions. AI technology mainly includes computer vision, natural language processing, and machine learning / deep learning. With the research and advancement of AI technology, it is being researched and applied in multiple fields, such as smart homes, intelligent customer service, virtual assistants, smart speakers, intelligent marketing, autonomous driving, robotics, and smart healthcare. It is believed that with further technological development, AI will be applied in even more fields and play an increasingly important role.

[0080] Natural Language Processing (NLP) is an important field within computer science and artificial intelligence. It studies various theories and methods for enabling effective communication between humans and computers using natural language. NLP is a science that integrates linguistics, computer science, and mathematics. Therefore, research in this field involves natural language, that is, the language people use in daily life, and thus it has a close relationship with linguistic research. NLP technologies typically include text processing, semantic understanding, machine translation, question answering, and knowledge graphs. This application employs a Bidirectional Encoder Representations from Transformers (BERT) model to vectorize features such as search requests, interest tags, and search results.

[0081] Machine learning is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory, among others. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Compared to data mining, which focuses on finding patterns in large datasets, machine learning emphasizes algorithm design, enabling computers to automatically "learn" patterns from data and use these patterns to predict unknown data.

[0082] Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence. Its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, and inductive learning. The search time-series prediction model and joint representation module in this application embodiment are trained using machine learning or deep learning techniques. Based on the joint representation model in this application embodiment, the first interaction representation information corresponding to each search request can be determined, and then candidate search requests can be filtered based on the first interaction representation information; based on the search time-series prediction model in this application embodiment, each candidate search request can be predicted, the probability of an object clicking on each candidate search request in the next step can be determined, and the candidate search requests can be ranked accordingly.

[0083] Specifically, during the model training process, when training the search time series prediction model listed above using machine learning technology, it is mainly trained based on the first training sample set. After the first training sample passes through the model, it will output the search request of the sample in the first training sample as the predicted probability of the next search request of the sample object. Based on the difference between the predicted probability and the true probability, the model parameters are continuously adjusted through optimization algorithms to obtain the trained search time series prediction model.

[0084] Furthermore, when training the joint representation models listed above using machine learning techniques, the training is mainly based on the second training sample set. After the second training sample is passed through the model, it will output the satisfaction prediction information of each search result in the second training sample. Based on the difference between the satisfaction prediction information and the corresponding real satisfaction information, the model parameters can be continuously adjusted through optimization algorithms to obtain the trained joint representation model.

[0085] Taking video search as an example, users can enter search terms on the application interface to obtain the videos they want to watch. In the search results page, in addition to showing users some videos related to the search terms, other search terms that users may search for can also be provided, making it easier for users to search for other related content.

[0086] The relevant search implementation method mainly retrieves other relevant search queries based on the object's current search query. However, it lacks a comprehensive understanding of the platform object, search query, and search results, resulting in insufficient modeling of the search query. Furthermore, after retrieving relevant search candidates, it fails to model and predict based on the object's historical search behavior sequence, causing the constructed relevant search candidates to be unable to directly meet the object's next search needs.

[0087] In view of this, embodiments of this application provide a search request recommendation method, apparatus, electronic device, and storage medium. Since this application obtains first interaction representation information corresponding to the target search request based on the target object's interest tags and the current target search request, the selected candidate search requests are more in line with the object's next search needs. Furthermore, this application combines the target object's interest tags and historical search requests to estimate the probability that each candidate search request will be the target object's next search request. Recommending search requests to the target object based on this probability satisfies the object's next search request, eliminates the object's search input, speeds up the object's discovery of desired video content, and improves the accuracy of search request recommendations.

[0088] The preferred embodiments of this application are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit this application. Furthermore, the embodiments and features in the embodiments of this application can be combined with each other without conflict.

[0089] like Figure 1 The diagram shown is an application scenario illustration of an embodiment of this application. The application scenario diagram includes two terminal devices 110 and one server 120.

[0090] In this embodiment, the terminal device 110 includes, but is not limited to, mobile phones, tablets, laptops, desktop computers, e-book readers, smart voice interaction devices, smart home appliances, and in-vehicle terminals. The terminal device may have a search and recommendation-related client installed. This client can be software (such as a browser, video software, etc.), or a webpage, mini-program, etc. The server 120 is the backend server corresponding to the software, webpage, mini-program, etc., or a server specifically used for search and recommendation; this application does not impose specific limitations. The server 120 can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms.

[0091] It should be noted that the methods in the various embodiments of this application can be executed by an electronic device, which can be a server 120 or a terminal device 110. That is, the method can be executed by the server 120 or the terminal device 110 alone, or by both the server 120 and the terminal device 110. For example, when executed by both the terminal device 110 and the server 120, the server 120 can first obtain the target user's interest tags and the current target search request. Based on this information, it can determine the first interaction representation information corresponding to the target search request. Then, it can filter out multiple candidate search requests from a pre-built set of historical search requests. After predicting the probability corresponding to each candidate search request, the server 120 can notify the terminal device 110 of each candidate search request and its corresponding probability. The terminal device 110 can sort each candidate search request according to its probability and recommend search requests to the target user based on the sorting results.

[0092] Specifically, taking video search as an example, such as Figure 2 The diagram shown is a schematic representation of a video search interface as illustrated in an embodiment of this application. Users can enter their search requests in the search box. Figure 2 The user's current search query is "XG classic police and gangster movies". The corresponding recommended search queries are: police and gangster TV series, police and gangster showdown ZSJ, police and gangster movies and TV series, 2021 police and gangster movies, etc., ZF.

[0093] Besides video search, it can also be used to search for other content, such as news, music, and novels. For example, to search for novels, users can enter their search request in the search box. Figure 2The user's current search request is "XG police and crime genre". Similar search requests for novels can also be made, without specific limitations.

[0094] Furthermore, the embodiments of this application can be applied not only to search scenarios, but also to various other scenarios, including but not limited to cloud technology, artificial intelligence, smart transportation, and assisted driving. In these scenarios, some search recommendation-related applications can be involved. For example, in a smart transportation scenario, a user can search for music, novels, etc., based on some in-vehicle terminals. Based on the search recommendation method in the embodiments of this application, relevant search recommendations can be made to the user to reduce user operations and improve search efficiency.

[0095] In one alternative implementation, the terminal device 110 and the server 120 can communicate via a communication network.

[0096] In one alternative implementation, the communication network is a wired network or a wireless network.

[0097] It should be noted that, Figure 1 The examples shown are merely illustrative; in reality, the number of terminal devices and servers is unlimited and is not specifically limited in the embodiments of this application.

[0098] In this embodiment of the application, when there are multiple servers, the multiple servers can form a blockchain, and the servers are nodes on the blockchain; as disclosed in the search request recommendation method of this embodiment, the object data involved can be stored on the blockchain, such as object interest tags, search requests, search results, interaction representation information, etc.

[0099] It is understood that in the specific implementation of this application, data related to the object's interest tags, search requests, search results, etc. are involved. When the above embodiments of this application are applied to specific products or technologies, permission or consent from the object is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0100] The search request recommendation method provided by the exemplary embodiments of this application will be described below with reference to the accompanying drawings and the application scenarios described above. It should be noted that the application scenarios described above are only shown to facilitate understanding of the spirit and principles of this application, and the embodiments of this application are not limited in any way in this respect.

[0101] See Figure 3 The diagram shown is an implementation flowchart of a search request recommendation method provided in this application embodiment. Taking the server as the executing entity as an example, the specific implementation flow of this method includes the following steps S31-S34:

[0102] S31: The server obtains the first interaction representation information corresponding to the target search request based on the target object's interest tags and the current target search request;

[0103] In this embodiment of the application, the first interactive representation information can be in the form of a vector, matrix, etc. Taking a vector as an example, the first interactive representation information can be obtained through artificial intelligence.

[0104] An optional implementation involves obtaining the first interaction representation information corresponding to the target search request based on a trained joint representation model, which is trained using artificial intelligence. The specific process may include the following sub-steps S311-S312 (not shown in the accompanying drawings):

[0105] S311: Input the interest label of the target object and the target search request into the trained joint representation model respectively to obtain the first deep representation information corresponding to the interest label of the target object and the second deep representation information corresponding to the target search request.

[0106] In this embodiment of the application, the object refers to the user who triggered the search request, or the account used by the user, etc. This article mainly uses the user as an example for explanation.

[0107] Interest tags are words used by an object to describe its interests. For example, a user can use words like "basketball" or "games" as interest tags to describe their interests. When a user has multiple interest tags, it can be represented as an interest tag sequence, such as (basketball, games, entertainment).

[0108] Both the first and second depth representations can be vectors or matrices. Taking vectors as an example, the first depth representation is a vector obtained by representing the interest tags of the target object, while the second depth representation is a vector obtained by representing the target search request.

[0109] In this embodiment of the application, the BERT model is used as an example for illustration. The interest label of the target object and the target search request are respectively input into the trained joint representation model. Through the BERT part of the joint representation model, the depth representation information corresponding to the interest label of the target object and the target search request are obtained respectively, namely the first depth representation information and the second depth representation information, for example, denoted as vector A1 and vector B1 respectively.

[0110] In one alternative implementation, the joint representation model is trained in the following manner, see [reference]. Figure 4 As shown, this is a schematic diagram of a training method for a joint representation model in an embodiment of this application. Taking the server as the execution subject as an example, the method specifically includes the following steps:

[0111] S41: The server obtains the second training sample set, wherein each second training sample in the second training sample set includes: an interest tag and a sample search request of a sample object, relevant search results of the sample object for the sample search request, and real information on the sample object's satisfaction with the relevant search results.

[0112] In this embodiment, a related search recall can be constructed by jointly modeling user interests, search queries, and user consumption of search results; specifically, this means that a sample object's interest tags, sample search requests, and related search results for the sample object in response to the sample search requests.

[0113] Among them, the real information on the satisfaction of the sample objects with the relevant search results can be used as sample labels to represent the degree of satisfaction of the sample objects with the relevant search results, which can be obtained by analyzing the feedback behavior of the sample objects with the relevant search results.

[0114] One alternative implementation is that the true satisfaction information can be determined based on the viewing time of the sample object for the relevant search results. Based on the viewing time and the full duration of the search results, the playback completion rate of the sample object for the relevant search results can be determined.

[0115] Taking a user's video search as an example, the user's satisfaction with the search results can be divided into K levels (e.g., K=4) to jointly represent the user's personalized interests and search query through supervised learning based on classification objectives. Specifically, the user's completion rate of the videos under the current search query is divided into K intervals, such as [0, 1 / K), [1 / K, 2 / K), ..., [(K-1) / K, 1], which correspond to levels 0, 1, ..., K-1, respectively. Since a user may watch multiple videos under a single search query, this embodiment uses the video with the highest completion rate as the input video for the model. That is, when there are multiple relevant search results, the search result with the highest completion rate can be selected based on the actual satisfaction information of each search result to construct a second training sample for training the model.

[0116] It should be noted that the methods for obtaining training samples and determining the true information on satisfaction listed above are merely illustrative examples and are not actually limited to the methods described above. This application does not impose any specific limitations.

[0117] S42: The server iterates through the joint representation model to be trained based on the second training sample set to obtain a trained joint representation model; wherein, in one iteration, the following operations are performed:

[0118] S421: Input the selected second training samples into the joint representation model respectively, and output the corresponding sample object's satisfaction prediction information for the relevant search results;

[0119] S422: Adjust the parameters of the joint representation model based on the differences between the predicted satisfaction information and the corresponding actual satisfaction information of each search result.

[0120] By comparing the predicted results with the actual results and adjusting the model parameters, the final trained model can better consider the degree of satisfaction.

[0121] For example Figure 5 The diagram shown illustrates the structure of a joint representation model in an embodiment of this application. Based on this model, user interests, the user's search query, and the playback status of video results after the user's search query are jointly represented. The goal is to predict the user's satisfaction level with the search result video under the current search query.

[0122] Among them, the interest tags of the sample object can be represented as a sequence of user interest tags, the sample search request of the sample object is the user search query, and the relevant search results of the sample object for the sample search request are the video tags, title text, etc. of the video with the highest playback completion rate listed above.

[0123] In an optional implementation, step S421 can be further divided into the following sub-steps S4211-S4212 (not shown in the accompanying drawings):

[0124] S4211: The second training samples selected from the second training sample set are respectively input into the joint representation model to be trained to obtain the first depth representation information corresponding to the interest label of the sample object in the second training sample, the second depth representation information corresponding to the sample search request of the sample object, and the third depth representation information corresponding to the relevant search results of the sample search request.

[0125] like Figure 5 As shown, the user interest tag sequence in the second training sample is passed through the left part sub-model BERT in the joint representation model to obtain the corresponding user personalized interest depth representation, that is, the first depth representation information corresponding to the interest tags of the sample object; the user search query in the second training sample is passed through the middle part sub-model BERT in the joint representation model to obtain the corresponding user search query depth representation, that is, the second depth representation information corresponding to the sample search request of the sample object; the video tag and title text in the second training sample are passed through the right part sub-model BERT in the joint representation model to obtain the corresponding search result video content depth representation, that is, the third depth representation information corresponding to the relevant search results of the sample search request.

[0126] S4212: Based on the first depth representation information corresponding to the interest tags of the sample object, the second depth representation information corresponding to the search request of the sample object, and the third depth representation information, predict the sample object and obtain the satisfaction prediction information for the sample object.

[0127] Based on these three types of deep representation information, the user's satisfaction level with the search results corresponding to the search query can be predicted, which is the satisfaction prediction information in the embodiments of this application.

[0128] See Figure 5 As shown, firstly, the first depth representation information (user personalized interest depth representation) corresponding to the interest tags of the sample object and the second depth representation information (i.e., user search query depth representation) corresponding to the search request of the sample object need to be processed by feature interaction to obtain the first interaction representation information (i.e., user personalized interest and search query interaction representation) corresponding to the sample object; and secondly, the second depth representation information (i.e., user search query depth representation) corresponding to the search request of the sample object and the third depth representation information (i.e., search result video content depth representation) corresponding to the relevant search results of the sample search request need to be processed by feature interaction to obtain the second interaction representation information (i.e., search query interaction and search result video content interaction representation) corresponding to the sample object.

[0129] Furthermore, based on the first and second interaction representation information of the sample objects, predictions are made to obtain satisfaction prediction information for the sample objects.

[0130] For example Figure 5 As shown, one possible implementation is to perform vector summation on the first interaction representation information and the second interaction representation information, and then predict the final satisfaction prediction information based on the summed vector.

[0131] In the above method, by jointly modeling the personalized interests, search queries, and playback behavior of video results of a large number of users on the video platform, the search query and user interests can be more fully represented. By combining the prediction of user satisfaction with the video results of the search query, the model's representation of users and search queries can better consider the degree of user satisfaction of the final search results, making the related search recall built by this model more in line with the user's search intent.

[0132] S312: Perform feature interaction processing on the first depth representation information corresponding to the interest tag of the target object and the second depth representation information corresponding to the target search request to obtain the first interaction representation information corresponding to the target search request.

[0133] The first depth representation information and the second depth representation information mentioned above can be represented as vector A1 and vector B1. The process of performing feature interaction processing on vector A1 and vector B1 can be understood as adding corresponding elements in the two vectors or weighted averaging, etc. The first interaction representation information corresponding to the target search request can be denoted as C1, without specific limitations here.

[0134] S32: Based on the first interaction representation information corresponding to the target search request, the server selects multiple historical search requests from the pre-built historical search request set that have a similarity to the target search request that reaches a first threshold, as candidate search requests related to the target object.

[0135] The historical search collection contains historical search requests for multiple historical objects (which may include the target object).

[0136] In this embodiment of the application, the similarity between each historical search request and the target search request is determined based on the first interaction representation information corresponding to the target search request and the first interaction representation information of each historical search request in the historical search set. In this way, multiple historical search requests with a similarity to the target search request that reaches a first threshold can be selected as candidate search requests related to the target object.

[0137] The calculation method for the first interaction representation information corresponding to each historical search request is similar to steps S311 and S312 above. For each historical search request, the following process is executed:

[0138] First, for a historical search request, the interest tags of the historical objects corresponding to the historical search request and the historical search request itself are input into the trained joint representation model to obtain the first deep representation information corresponding to the interest tags of the historical objects and the second deep representation information corresponding to the historical search request; for example, they are denoted as vector A2 and vector B2 respectively.

[0139] Secondly, the first depth representation information A2 corresponding to the interest tag of the historical object and the second depth representation information B2 corresponding to the historical search request are subjected to feature interaction processing to obtain the first interaction representation information C2 corresponding to the historical search request.

[0140] Using the same method, the first interaction representation information of each historical search request can be obtained.

[0141] In an optional implementation, step S32 specifically includes the following sub-steps S321-S322 (not shown in the figures):

[0142] S321: Obtain the similarity retrieval index corresponding to each historical search request in the historical search request set. The similarity retrieval index of each historical search request is obtained by vector compression of the first interaction representation information corresponding to the corresponding historical search request, which is determined based on the interest tags and historical search requests of the corresponding historical objects.

[0143] For example, after the joint representation model listed above has been trained, the intermediate sub-models of this model can be used to represent a large number of search queries on the video platform. The deep representations of these search queries can be used to build a similarity retrieval index (QI), which can be constructed using ElasticFaiss.

[0144] S322: Based on each similarity retrieval index and the first interaction representation information corresponding to the target search request, select multiple historical search requests whose similarity to the target search request reaches a first threshold.

[0145] For example, when a user (the target object) triggers a search, by inputting the user's interest tag sequence and the user's current search query as described above... Figure 5 The model shown has left and middle sub-models, which, after passing through a user personalized interest and search query interaction representation layer, construct a deep representation (UQ) for this user under the current search query. Then, using this user's current interaction representation (UQ), the constructed query deep similarity retrieval index is queried. Other queries in the index whose similarity meets a certain threshold (the first threshold) are used as the candidate queue for related searches for the current user under the current search term.

[0146] like Figure 6 The diagram illustrates a relevant search retrieval process in an embodiment of this application. First, based on each historical search request in the historical search request set and its corresponding first interaction representation information, Faiss training is performed to obtain a similarity retrieval index (hereinafter referred to as an index) for each historical search request. Then, based on each index and the first interaction representation information corresponding to the target search request, a Faiss query is performed to obtain the Faiss query results and determine each candidate search request.

[0147] In the above implementation, by jointly modeling the user, the search query, and the user's video playback behavior in the search results, a more complete model of the user and the search query can be achieved. This not only captures the search relevance between the user and the query, but also models the user's viewing behavior when searching for this query. As a result, when the model is used for relevant recommendations, the search candidates recommended to the user are more in line with the user's expectations.

[0148] S33: The server determines the probability of each candidate search request becoming the next search request for the target object based on each candidate search request and each historical search request related to the target object.

[0149] In this embodiment, this step mainly utilizes historical search requests related to the target object to analyze the probability that each candidate search request will be the next search request for the target object. The higher the probability value, the greater the likelihood that the target object will use that candidate search request as its search term in the next search. This process can also be achieved through artificial intelligence.

[0150] An optional implementation involves calculating each probability based on a trained search time-series prediction model, which is trained using artificial intelligence. The specific process may include the following sub-steps S331-S332 (not shown in the accompanying drawings):

[0151] S331: Based on the interest tags of the target object and each historical search request related to the target object, obtain the first interaction representation information corresponding to each historical search request;

[0152] Specifically, the calculation method for the first interaction information corresponding to the historical search requests related to the target object is similar to steps S311 and S312 above. For each historical search request related to the target object, the historical object corresponding to the historical search request is the target object, because the following processes can be executed respectively:

[0153] First, for a historical search request, the interest tags of the target object and the historical search request are respectively input into the trained joint representation model to obtain the first deep representation information corresponding to the interest tags of the target object and the second deep representation information corresponding to the historical search request. Second, the first deep representation information corresponding to the interest tags of the target object and the second deep representation information corresponding to the historical search request are subjected to feature interaction processing to obtain the first interaction representation information corresponding to the historical search request.

[0154] Using the same method, we can obtain the first interaction representation information of each historical search request related to the target object.

[0155] S332: For each candidate search request, perform the following operations:

[0156] S3321: Based on the interest tags of the target object and a candidate search request, obtain the first interaction representation information corresponding to a candidate search request;

[0157] This process is similar to the one described above, in that the candidate search request and the interest label of the target object are respectively input into the trained joint representation model to obtain the first interaction representation information corresponding to the candidate search request.

[0158] It should be noted that, in the embodiments of this application, the first interaction representation information corresponding to each search request can be calculated based on the joint representation model listed above. For the specific process, please refer to steps S311 and S312. In summary, it is obtained through the trained joint representation model. The input features of the model are: the interest tags of the search object corresponding to the search request and the search request. The output features are the first interaction representation information corresponding to the search request. The search object corresponding to the search request is: an object for which a related search is performed based on the search request. Repeated details will not be repeated.

[0159] S3322: Input the first interaction representation information corresponding to a candidate search request and the first interaction representation information corresponding to each historical search request into the trained search time-series prediction model to obtain the probability that a candidate search request will be the next search request for the target object.

[0160] In this embodiment of the application, in order to further improve the ability of the constructed relevant search candidates to meet the user's next search needs, the constructed relevant search candidates are ranked by a user search time series prediction model, and the model structure is as follows: Figure 7 As shown, it is a structural schematic diagram of a search time series prediction model in an embodiment of this application.

[0161] In this embodiment of the application, by modeling a search time sequence prediction model on the search sequences of a large number of users on the platform, it is possible to learn whether the next search query is the user's expected search query based on their historical search queries. For example... Figure 7 As shown, the model uses the user's historical search query sequence as the preceding sequence, and the current search query as the current search query (corresponding to...). Figure 7 The relevant search candidate query position is used during model usage. This position corresponds to the candidate search request (as described in S3322). The input layer vector of the query at each time step is obtained by using... Figure 5 In the model shown, the user's personalized interests and search query interaction are used as inputs. This allows the model to capture both user interests and search query features simultaneously, and it is also easier to converge compared to training a model with random initialization.

[0162] In one alternative implementation, the search time-series prediction model is trained in the following manner, see reference. Figure 8As shown, this is a schematic diagram of a training method for a search time series prediction model in an embodiment of this application. Taking the server as the execution subject as an example, it specifically includes the following steps S81-S82:

[0163] S81: The server obtains the first training sample set, wherein each first training sample in the first training sample set includes: a sample search request for a sample object and multiple historical search requests, as well as the true probability of the sample search request as the next search request for the sample object.

[0164] For example, this model can be learned by analyzing the search query sequences of a large number of users on a video platform. For instance, a user's historical search sequence might be: q0, q1, q2, ..., q m , ..., q t It can use any length of preceding text as a historical search query sequence, such as q0, q1, q2, ..., q m (i.e., multiple historical search requests for the sample object), its next search query q m+1 As a positive example of the search query at the "current" moment on the right (i.e., the sample search request of the sample object, corresponding to...) Figure 7 (Positions of candidate queries in the relevant search), and randomly sample other queries as negative examples. Figure 7 The search time-series prediction model shown is trained so that the model has the input user's historical search query sequence and a certain current related search term candidate, and outputs the probability that the related search term candidate is the user's next search query.

[0165] Based on this, when the historical search requests related to the target object and a candidate search request are input into the trained search time-series prediction model, the probability of the candidate search request becoming the next search query for the target object can be predicted.

[0166] S82: Based on the first training sample set, the server iteratively trains the search time-series prediction model to be trained multiple times to obtain a trained search time-series prediction model; wherein, in one iteration, the following operations are performed:

[0167] S821: Input the selected first training sample into the search time-series prediction model to be trained, and output the corresponding sample search request of the sample object as the predicted probability of the sample object's next search request.

[0168] S822: Adjust the parameters of the search time series prediction model based on the difference between the predicted probability and the corresponding true probability.

[0169] By using the search time-series prediction model, the probability of each candidate search request in the relevant search candidate recall queue constructed in step S32 above is calculated to obtain the probability of the next search. Based on this probability, the relevant search candidate list can be sorted.

[0170] In the above implementation, the relevant search candidates recommended to the user are further sorted by combining user search time-series modeling, which further improves the user's click / play of relevant search candidates and corresponding video search results, so that the user can directly click on relevant search candidates to proceed to the next search and play, thereby improving the user's search experience.

[0171] S34: After ranking each candidate search request based on its probability, the server recommends search requests to the target object based on the ranking results.

[0172] For example Figure 2 As shown, when a user performs a video search, the target search request entered in the search box is "XG classic police and gangster movies". The five candidate search requests recommended to the user are: police and gangster TV series, police and gangster showdown ZSJ, police and gangster movies and TV series, 2021 police and gangster movies, etc., ZF.

[0173] It should be noted that the above implementation method proposes a relevant search implementation method that integrates user search results and temporal modeling. By using a large amount of user search queries and user consumption data of video search results for search queries, the method jointly models users, search queries, and corresponding videos. This allows the model to take into account the modeling representation of users and the platform's historical queries, as well as users' viewing expectations for the corresponding search results. Furthermore, the constructed relevant search candidates are further ranked through a user-personalized search temporal prediction model, enabling the relevant search candidates to more accurately capture users' next search needs and improve users' search experience on video sites.

[0174] In addition, considering that the related search candidate list constructed in the above steps may contain candidate search requests with high repetition, when these candidate search requests are displayed to the user at the same time, it will waste the exposure position of related search candidates. Therefore, it is necessary to deduplicate the related search candidate list constructed in the above steps, that is, to perform diversity post-processing based on the depth representation of related search candidates.

[0175] See Figure 9A As shown, this is a simplified flowchart of an embodiment of this application, taking the server as the executing entity as an example, namely:

[0176] S91: The server jointly models user interests, search queries, and user consumption of search results to build relevant search recall;

[0177] S92: The server ranks relevant search candidates based on a search time-series prediction model;

[0178] S93: The server deduplicates relevant search candidates.

[0179] In this application embodiment, an optional deduplication method is as follows:

[0180] First, based on the target object's interest tags and each candidate search request, the first interaction representation information corresponding to each candidate search request is obtained;

[0181] Specifically, this can be achieved through the above Figure 5 The left and middle sub-models of the listed joint representation model, after passing through the user personalized interest and search query interaction representation layer, construct a deep representation of this target object under each candidate search request.

[0182] Then, according to the sorting results, each candidate search request is traversed in turn, and the candidate search requests with lower probability are deduplicated among the two candidate search requests whose similarity between the corresponding first interaction representation information reaches the second threshold.

[0183] For example, in step S33, when the server sorts the candidate search requests based on their probabilities, it can sort them in descending order of probability value. This way, when iterating through the candidate search requests according to the sorting result, the requests can be traversed sequentially from front to back, with the candidate search request with the lower probability being the one ranked later. Alternatively, the requests can be sorted in ascending order of probability value. This way, when iterating through the candidate search requests according to the sorting result, the requests can be traversed sequentially from back to front, with the candidate search request with the lower probability being the one ranked earlier.

[0184] Specifically, when the similarity of the depth representations of two related search candidates is greater than a certain threshold, the two candidates can be considered to be too repetitive. In this embodiment, cosine similarity can be used to calculate the depth representation similarity between two related search candidates. That is, when the cosine similarity between the first interaction representation information corresponding to the two candidate search requests reaches the second threshold, deduplication can be performed.

[0185] Finally, among the remaining candidate search requests after deduplication, at least one candidate search request whose probability reaches the third threshold is recommended to the target object.

[0186] The relevant search candidate list constructed in the above process can be traversed from front to back. If the similarity between a candidate and other candidates reaches a certain threshold, the candidate is removed from the candidate list. After deduplication, based on the probability corresponding to each candidate search request, the Top N candidates that meet a certain threshold are selected as relevant searches and displayed to the user.

[0187] For example, if there are 7 candidate search requests corresponding to the target object, named candidate search requests 1-7, the relevant search candidate list after sorting them from highest to lowest probability is as follows:

[0188] Candidate search request 1, candidate search request 3, candidate search request 4, candidate search request 7, candidate search request 5, candidate search request 2, candidate search request 6.

[0189] Next, the first interaction representation information corresponding to each of the seven candidate search requests is calculated. That is, using the joint representation model described above, the interest tags of the target object and each candidate search request are interactively represented to obtain the corresponding first interaction representation information. Then, the cosine similarity between the first interaction representation information of candidate search request 1 and candidate search request 3 is calculated. If the similarity is greater than the second threshold, candidate search request 3 is removed. The cosine similarity between the first interaction representation information of candidate search request 1 and candidate search request 4 is further calculated. If the similarity is not greater than the second threshold, candidate search request 4 is retained. The cosine similarity between the first interaction representation information of candidate search request 4 and candidate search request 7 is further calculated. Based on the similarity, it is determined whether to remove duplicates, and so on.

[0190] Assuming that candidate search request 3 and candidate search request 6 are ultimately removed, the updated list of related search candidates is as follows:

[0191] Candidate search request 1, candidate search request 4, candidate search request 7, candidate search request 5, candidate search request 2.

[0192] Finally, when making a search request to the target object, the search candidates with a probability of reaching the third threshold in the updated list can be recommended to the target object, and recommendations can be made to the target object based on the probability.

[0193] For example Figure 2 The diagram shows a search interface related to video search as illustrated in this application embodiment. The target object's current search request is "XG classic police and gangster movies," and the corresponding recommended search requests include: police and gangster TV series, police and gangster showdown ZSJ, police and gangster movies and TV series, 2021 police and gangster movies…, ZF.

[0194] The target audience can view search results on the current interface and quickly search for other video content through these recommended search requests. For example, if the target audience clicks on "Police vs. Criminals ZSJ", video content related to that search term will be quickly recommended to them without them having to re-enter the search term "Police vs. Criminals ZSJ".

[0195] In summary, existing search implementation methods primarily rely on the user's current search query to recall other search queries as relevant search candidates. However, they fail to jointly model the platform user, search query, and search results, resulting in recommended relevant search candidates that do not adequately meet the user's search needs. This application proposes a relevant search implementation method that integrates user search performance with temporal modeling. (See reference...) Figure 9B As shown, it is a general logical diagram of the method.

[0196] In summary, by first comprehensively modeling users, search queries, and corresponding video search results from a large number of users on the platform, a joint representation model is obtained. This model can simultaneously construct representations for users and historical queries on the platform, and can reflect users' viewing expectations for the search results corresponding to the queries.

[0197] Based on this, firstly, when the target object's interest tags and the target object's current target search request are used as inputs to the joint representation model, the first interaction representation information corresponding to the target search request output by the model is obtained; and when the interest tags and historical search requests of each historical object in the historical search request set are used as inputs to the joint representation model, the first interaction representation information corresponding to each historical search request output by the model is obtained. Based on this, a list of candidate search requests related to the target object (also known as a list of related search candidates) is recalled.

[0198] After recalling a list of relevant search candidates related to user interests, the user-personalized search time-series prediction model is used to rank and predict the relevant search candidates.

[0199] Specifically, based on the first interaction representation information corresponding to each historical search request related to the target object, and the first interaction representation information corresponding to each candidate search request, the probability corresponding to each candidate search request is predicted by the search time series prediction model, and the ranking prediction is performed based on the probability value.

[0200] Then, the relevant search candidate results are deduplicated, and the Top N candidate search requests with the probability reaching the third threshold are selected from the removed relevant search candidate list and displayed to the target audience. This allows the relevant search candidates to better meet the user's next search needs, and the search results to better match the user's search intent. Users can complete further searches by directly clicking on relevant search candidates, thus improving the user's search experience on video sites.

[0201] For example, Figure 9C This is a schematic diagram of a search process in one embodiment of this application. The target object's interest tags include: entertainment, action, drama, etc. The target object's current input search request is "XG classic police and gangster movies". After processing by the joint representation model and search time-series prediction model listed above, and after deduplication, five candidate search requests are obtained, as shown in the figure. Based on this result, a search interface can be displayed to the target object. In addition to some videos matched based on "XG classic police and gangster movies", the search interface displays these five candidate search requests in the "Related Searches" module. The target object can quickly click on these candidate search requests to conduct further searches, improving the accuracy of search request recommendations.

[0202] The following is combined with Figure 9D The training and usage process of the model in the embodiments of this application is summarized below.

[0203] See Figure 9D The diagram shown is a flowchart illustrating the training and usage process of a model in one embodiment of this application. Taking a server as the executing entity as an example, the specific implementation flow of this method is as follows:

[0204] Step S901: The server iterates through the joint representation model to be trained based on the second training sample set to obtain the trained joint representation model.

[0205] Step S902: Based on the trained joint representation model, the server obtains the first interaction representation information corresponding to each historical search request in the historical search request set, and determines the corresponding similarity retrieval index based on each first interaction representation information;

[0206] Step S903: The server iterates through the search time series prediction model to be trained based on the first training sample set to obtain the trained search time series prediction model.

[0207] Step S904: The server inputs the interest tags of the target object and the current target search request into the trained joint representation model to obtain the first interaction representation information corresponding to the target search request;

[0208] Step S905: Based on the similarity retrieval index of each historical search request in the historical search request set, the server selects multiple historical search requests whose similarity to the target search request reaches a first threshold as candidate search requests related to the target object.

[0209] Step S906: The server obtains the probability that each candidate search request will be the next search request for the target object based on the trained search time-series prediction model.

[0210] Step S907: The server sorts the candidate search requests according to the order of probability from largest to smallest;

[0211] Step S908: Based on the interest tags of the target object and each candidate search request, the server obtains the first interaction representation information corresponding to each candidate search request.

[0212] Step S909: The server iterates through each candidate search request according to the sorting result, and removes duplicates from the candidate search requests ranked lower among the two candidate search requests whose similarity between the corresponding first interaction representation information reaches the second threshold.

[0213] Step S910: The server recommends at least one candidate search request whose probability reaches the third threshold from the remaining candidate search requests after deduplication to the target object.

[0214] Based on the same inventive concept, embodiments of this application also provide a search request recommendation device. For example... Figure 10 As shown, this is a structural schematic diagram of a search request recommendation device 1000, which may include:

[0215] The interaction representation unit 1001 is used to obtain the first interaction representation information corresponding to the target search request based on the interest tags of the target object and the current target search request.

[0216] The filtering unit 1002 is used to filter out multiple historical search requests that have a similarity to the target search request reaching a first threshold from a pre-built set of historical search requests based on the first interaction representation information corresponding to the target search request, as candidate search requests related to the target object.

[0217] The determining unit 1003 is used to determine the probability of a corresponding candidate search request being the next search request for the target object based on each candidate search request and each historical search request related to the target object.

[0218] Recommendation unit 1004 is used to rank each candidate search request based on its probability and then recommend search requests to the target object based on the ranking results.

[0219] Optionally, the filtering unit 1002 is specifically used for:

[0220] Obtain the similarity retrieval index corresponding to each historical search request in the historical search request set. The similarity retrieval index of each historical search request is obtained by vector compression of the first interaction representation information corresponding to the corresponding historical search request, which is determined based on the interest tags and historical search requests of the corresponding historical objects.

[0221] Based on each similarity retrieval index and the first interaction representation information corresponding to the target search request, multiple historical search requests with a similarity to the target search request reaching a first threshold are selected.

[0222] Optionally, the determining unit 1003 is specifically used for:

[0223] Based on the interest tags of the target object and each historical search request related to the target object, the first interaction representation information corresponding to each historical search request is obtained;

[0224] For each candidate search request, perform the following operations:

[0225] Based on the interest tags of the target object and a candidate search request, obtain the first interaction representation information corresponding to the candidate search request;

[0226] The first interaction representation information corresponding to a candidate search request and the first interaction representation information corresponding to each historical search request are respectively input into the trained search time sequence prediction model to obtain the probability that a candidate search request will be the next search request for the target object.

[0227] Optionally, the device also includes:

[0228] The first training unit 1005 is used to train the search time-series prediction model in the following manner:

[0229] Obtain the first training sample set, wherein each first training sample in the first training sample set includes: a sample search request for a sample object and multiple historical search requests, and the true probability of the sample search request as the next search request for the sample object;

[0230] Based on the first training sample set, the search time series prediction model to be trained is trained iteratively multiple times to obtain a trained search time series prediction model; wherein, in one iteration, the following operations are performed:

[0231] The selected first training sample is input into the search time-series prediction model to be trained, and the corresponding sample search request of the sample object is output as the predicted probability of the sample object's next search request.

[0232] The parameters of the search time series prediction model are adjusted based on the difference between the predicted probability and the corresponding true probability.

[0233] Optionally, the recommended unit 1004 is also used for:

[0234] After sorting each candidate search request based on each probability, and before recommending search requests to the target object based on the sorting results, the first interaction representation information corresponding to each candidate search request is obtained based on the target object's interest tags and each candidate search request.

[0235] According to the sorting results, each candidate search request is traversed sequentially, and the candidate search requests with lower probability are deduplicated among the two candidate search requests whose similarity between the corresponding first interaction representation information reaches the second threshold.

[0236] Recommended unit 1004 is specifically used for:

[0237] Among the remaining candidate search requests after deduplication, at least one candidate search request whose probability reaches the third threshold is recommended to the target object.

[0238] Optionally, the first interaction representation information corresponding to each search request can be obtained in the following ways: the search request is a target search request, a candidate search request, or a historical search request.

[0239] For each search request, perform the following operations:

[0240] Input the interest tags of the search object corresponding to a search request and the search request into the trained joint representation model to obtain the first deep representation information corresponding to the interest tags of the search object and the second deep representation information corresponding to the search request.

[0241] The first depth representation information corresponding to the interest tags of the search object and the second depth representation information corresponding to the search request are subjected to feature interaction processing to obtain the first interaction representation information corresponding to the search request.

[0242] Optionally, the device also includes:

[0243] The second training unit 1006 is used to train the joint representation model in the following manner:

[0244] Obtain a second training sample set, wherein each second training sample in the second training sample set includes: an interest tag and a sample search request for a sample object, relevant search results for the sample object in response to the sample search request, and real information on the sample object's satisfaction with the relevant search results;

[0245] Based on the second training sample set, the joint representation model to be trained is trained iteratively multiple times to obtain the trained joint representation model; wherein, in one iteration, the following operations are performed:

[0246] The selected second training samples are input into the joint representation model, and the corresponding sample object's satisfaction prediction information for the relevant search results is output.

[0247] Based on the differences between the satisfaction prediction information and the corresponding actual satisfaction information of each search result, the parameters of the joint representation model are adjusted.

[0248] Optionally, the true information on the sample objects' satisfaction with the relevant search results is determined based on the sample objects' feedback behavior towards the relevant search results.

[0249] Optionally, the second training unit 1006 is specifically used for:

[0250] The second training samples selected from the second training sample set are respectively input into the joint representation model to be trained to obtain the first deep representation information corresponding to the interest tags of the sample objects in the second training samples, the second deep representation information corresponding to the sample search requests of the sample objects, and the third deep representation information corresponding to the relevant search results of the sample search requests.

[0251] Based on the first depth representation information corresponding to the interest tags of the sample object, the second depth representation information corresponding to the search request of the sample object, and the third depth representation information, the sample object is predicted to obtain the satisfaction prediction information for the sample object.

[0252] Optionally, the second training unit 1006 is specifically used for:

[0253] The first depth representation information corresponding to the interest tag of the sample object and the second depth representation information corresponding to the search request of the sample object are subjected to feature interaction processing to obtain the first interaction representation information corresponding to the sample object; and the second depth representation information corresponding to the search request of the sample object and the third depth representation information corresponding to the relevant search results of the sample search request are subjected to feature interaction processing to obtain the second interaction representation information corresponding to the sample object.

[0254] Based on the first and second interaction representation information of the sample objects, predictions are made to obtain the satisfaction prediction information for the sample objects.

[0255] Because this application obtains the first interaction representation information corresponding to the target search request based on the target object's interest tags and the current target search request, the candidate search requests selected based on this first interaction representation information are more in line with the object's next search needs. In addition, this application combines the target object's interest tags and historical search requests to estimate the probability of each candidate search request becoming the target object's next search request. Based on this probability, the application recommends search requests to the target object, which can satisfy the object's next search request, save the object's search input, speed up the object's discovery of desired video content, and improve the accuracy of search request recommendations.

[0256] For ease of description, the above sections are divided into modules (or units) according to their functions and described separately. Of course, in implementing this application, the functions of each module (or unit) can be implemented in one or more software or hardware components.

[0257] Having introduced the search request recommendation method and apparatus according to exemplary embodiments of this application, we will now introduce an electronic device according to another exemplary embodiment of this application.

[0258] Those skilled in the art will understand that various aspects of this application can be implemented as a system, method, or program product. Therefore, various aspects of this application can be specifically implemented in the following forms: a completely hardware implementation, a completely software implementation (including firmware, microcode, etc.), or a combination of hardware and software implementations, collectively referred to herein as a "circuit," "module," or "system."

[0259] Based on the same inventive concept as the above-described method embodiments, this application also provides an electronic device. In one embodiment, the electronic device may be a server, such as... Figure 1 The server 120 is shown. In this embodiment, the structure of the electronic device can be as follows: Figure 11 As shown, it includes a memory 1101, a communication module 1103, and one or more processors 1102.

[0260] The memory 1101 is used to store computer programs executed by the processor 1102. The memory 1101 may mainly include a program storage area and a data storage area. The program storage area may store the operating system and programs required to run instant messaging functions, etc.; the data storage area may store various instant messaging information and operation instruction sets, etc.

[0261] Memory 1101 may be volatile memory, such as RAM; memory 1101 may also be non-volatile memory, such as read-only memory, flash memory, hard disk drive (HDD), or solid-state drive (SSD); or memory 1101 may be any other medium capable of carrying or storing a desired computer program having the form of instructions or data structures and accessible by a computer, but is not limited thereto. Memory 1101 may be a combination of the above-described memories.

[0262] Processor 1102 may include one or more central processing units (CPUs) or digital processing units, etc. Processor 1102 is used to implement the above-described search request recommendation method when calling computer programs stored in memory 1101.

[0263] The communication module 1103 is used to communicate with terminal devices and other servers.

[0264] This application embodiment does not limit the specific connection medium between the memory 1101, communication module 1103, and processor 1102. This application embodiment... Figure 11 The memory 1101 and the processor 1102 are connected via a bus 1104, and the bus 1104 is in Figure 11 The diagram uses thick lines to describe the connections between other components; these are for illustrative purposes only and should not be considered limiting. Bus 1104 can be divided into address bus, data bus, control bus, etc. For ease of description, Figure 11 It is described using only a thick line, but does not indicate that there is only one bus or one type of bus.

[0265] The memory 1101 stores a computer storage medium, which stores computer-executable instructions for implementing the search request recommendation method of this application embodiment. The processor 1102 is used to execute the above-described search request recommendation method, such as... Figure 3 As shown.

[0266] In another embodiment, the electronic device may also be other electronic devices, such as... Figure 1 The terminal device 110 is shown. In this embodiment, the electronic device can be structured as follows: Figure 12 As shown, it includes components such as: communication component 1210, memory 1220, display unit 1230, camera 1240, sensor 1250, audio circuit 1260, Bluetooth module 1270, processor 1280, etc.

[0267] The communication component 1210 is used to communicate with the server. In some embodiments, it may include a Circuit-Wireless Fidelity (WiFi) module, which is a short-range wireless transmission technology. Electronic devices can use the WiFi module to help users send and receive information.

[0268] The memory 1220 can be used to store software programs and data. The processor 1280 executes various functions of the terminal device 110 and performs data processing by running the software programs or data stored in the memory 1220. The memory 1220 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. The memory 1220 stores an operating system that enables the terminal device 110 to run. In this application, the memory 1220 may store the operating system and various application programs, and may also store a computer program that executes the search request recommendation method of the embodiments of this application.

[0269] The display unit 1230 can also be used to display information input by the user or information provided to the user, as well as a graphical user interface (GUI) of various menus of the terminal device 110. Specifically, the display unit 1230 may include a display screen 1232 disposed on the front of the terminal device 110. The display screen 1232 may be configured as a liquid crystal display, a light-emitting diode, or the like. The display unit 1230 can be used to display information from embodiments of this application. Figure 2 The search interface shown, etc.

[0270] The display unit 1230 can also be used to receive input digital or character information and generate signal inputs related to user settings and function control of the terminal device 110. Specifically, the display unit 1230 may include a touch screen 1231 disposed on the front of the terminal device 110, which can collect touch operations of the user on or near it, such as clicking the search button, dragging the scroll box, etc.

[0271] The touchscreen 1231 can be placed on top of the display screen 1232, or the touchscreen 1231 and the display screen 1232 can be integrated to realize the input and output functions of the terminal device 110. After integration, it can be referred to as a touch display screen. In this application, the display unit 1230 can display the application program and the corresponding operation steps.

[0272] Camera 1240 can be used to capture still images, which users can then post comments on via an application. There can be one or multiple cameras 1240. An object is projected onto a photosensitive element through a lens, generating an optical image. This photosensitive element can be a charge-coupled device (CCD) or a complementary metal-oxide-semiconductor (CMOS) phototransistor. The photosensitive element converts the light signal into an electrical signal, which is then transmitted to the processor 1280 for conversion into a digital image signal.

[0273] The terminal device may also include at least one sensor 1250, such as an accelerometer 1251, a proximity sensor 1252, a fingerprint sensor 1253, and a temperature sensor 1254. The terminal device may also be equipped with other sensors such as a gyroscope, barometer, hygrometer, thermometer, infrared sensor, light sensor, and motion sensor.

[0274] Audio circuitry 1260, speaker 1261, and microphone 1262 provide an audio interface between the user and terminal device 110. Audio circuitry 1260 converts received audio data into electrical signals, which are then transmitted to speaker 1261, where they are converted into sound signals for output. Terminal device 110 may also be equipped with volume buttons for adjusting the volume of the sound signal. On the other hand, microphone 1262 converts collected sound signals into electrical signals, which are received by audio circuitry 1260, converted into audio data, and then output to communication component 1210 for transmission to, for example, another terminal device 110, or to memory 1220 for further processing.

[0275] Bluetooth module 1270 is used to interact with other Bluetooth devices that also have Bluetooth modules via the Bluetooth protocol. For example, a terminal device can establish a Bluetooth connection with a wearable electronic device (such as a smartwatch) that also has a Bluetooth module through Bluetooth module 1270, thereby exchanging data.

[0276] The processor 1280 is the control center of the terminal device, connecting various parts of the terminal through various interfaces and lines. It executes various functions and processes data by running or executing software programs stored in the memory 1220 and calling data stored in the memory 1220. In some embodiments, the processor 1280 may include at least one processing unit; the processor 1280 may also integrate an application processor and a baseband processor. In this application, the processor 1280 can run an operating system, applications, user interface display and touch response, as well as the search request recommendation method of this embodiment. Furthermore, the processor 1280 is coupled to the display unit 1230.

[0277] In some possible implementations, various aspects of the search request recommendation method provided in this application can also be implemented in the form of a program product, which includes a computer program. When the program product is run on an electronic device, the computer program causes the electronic device to perform the steps in the search request recommendation method according to the various exemplary embodiments of this application described above. For example, the electronic device can perform actions such as... Figure 3 The steps are shown in the figure.

[0278] The program product may employ any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0279] The program product of the embodiments of this application may employ a portable compact disc read-only memory (CD-ROM) and include a computer program, and may run on an electronic device. However, the program product of this application is not limited thereto. In this document, the readable storage medium may be any tangible medium that contains or stores a program that may be used by or in conjunction with a command execution system, apparatus, or device.

[0280] A readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying a readable computer program. This propagated data signal may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable signal medium may also be any readable medium other than a readable storage medium, capable of sending, propagating, or transmitting a program for use by or in conjunction with a command execution system, apparatus, or device.

[0281] Computer programs contained on readable media may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0282] Computer programs for performing the operations of this application can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as C or similar languages. The computer program can execute entirely on the user's electronic device, partially on the user's electronic device, as a standalone software package, partially on the user's electronic device and partially on a remote electronic device, or entirely on a remote electronic device or server. In cases involving remote electronic devices, the remote electronic device can be connected to the user's electronic device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external electronic device (e.g., via the Internet using an Internet service provider).

[0283] It should be noted that although several units or sub-units of the device have been mentioned in the detailed description above, this division is merely exemplary and not mandatory. In fact, according to embodiments of this application, the features and functions of two or more units described above can be embodied in one unit. Conversely, the features and functions of one unit described above can be further divided and embodied by multiple units.

[0284] Furthermore, although the operations of the method of this application are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0285] 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 a computer-usable computer program.

[0286] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, executed by the processor of the electronic device, produce a mechanism for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0287] These computer program commands may also be stored in a computer-readable storage medium that can direct electronic devices to operate in a particular manner, such that the commands stored in the computer-readable storage medium produce an article of manufacture including a command device, which is implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0288] These computer program commands can also be loaded onto electronic devices, causing a series of operational steps to be performed on the electronic device to produce a computer-implemented process, thereby providing the commands executed on the electronic device for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0289] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0290] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for recommending search requests, characterized in that, The method includes: Based on the first interaction representation information corresponding to the current target search request of the target object, multiple historical search requests with a similarity of up to a first threshold with the target search request are selected from the pre-constructed historical search request set as candidate search requests related to the target object. Based on the first interaction representation information of each candidate search request and each historical search request related to the target object, the probability of the corresponding candidate search request being the next search request for the target object is determined. The first interaction representation information of each search request is obtained by interacting the interest tags of the corresponding search object and the search request using a joint representation model. The joint representation model is used to jointly model the interest tags of the sample object, the sample search request, and the satisfaction of the sample object with the relevant search results for the sample search request, so that the representation of the interest tags and search request by the joint representation model can consider the degree of satisfaction of the final search results with the object. The search request can be a target search request, a candidate search request, or a historical search request. After ranking each candidate search request based on its probability, a search request recommendation is made to the target object based on the ranking result.

2. The method as described in claim 1, characterized in that, The first interaction representation information corresponding to the current target search request of the target object is used to filter out multiple historical search requests from a pre-built set of historical search requests that have a similarity to the target search request reaching a first threshold, including: Obtain the similarity retrieval index corresponding to each historical search request in the historical search request set. The similarity retrieval index of each historical search request is obtained by vector compression of the first interaction representation information corresponding to the corresponding historical search request, which is determined based on the interest tags and historical search requests of the corresponding historical objects. Based on each similarity retrieval index and the first interaction representation information corresponding to the target search request, multiple historical search requests that have a similarity to the target search request reaching a first threshold are selected.

3. The method as described in claim 1, characterized in that, The step of determining the probability that a corresponding candidate search request will be the next search request for the target object based on the first interaction representation information of each candidate search request and each historical search request related to the target object includes: Based on the interest tags of the target object and each historical search request related to the target object, first interaction representation information corresponding to each historical search request is obtained; For each candidate search request, perform the following operations: Based on the interest tags of the target object and a candidate search request, obtain the first interaction representation information corresponding to the candidate search request; The first interaction representation information corresponding to the candidate search request and the first interaction representation information corresponding to each of the historical search requests are respectively input into the trained search time sequence prediction model to obtain the probability that the candidate search request is the next search request for the target object.

4. The method as described in claim 3, characterized in that, The search time-series prediction model was trained in the following manner: Obtain a first training sample set, wherein each first training sample in the first training sample set includes: a sample search request for a sample object and multiple historical search requests, and the sample search request as the true probability of the next search request for the sample object; Based on the first training sample set, the search time-series prediction model to be trained is trained iteratively multiple times to obtain a trained search time-series prediction model; wherein, in one iteration, the following operations are performed: The selected first training samples are input into the search time-series prediction model to be trained, and the corresponding sample search requests of the sample objects are output as the predicted probability of the next search request of the sample objects. The parameters of the search time series prediction model are adjusted based on the difference between the predicted probability and the corresponding true probability.

5. The method as described in claim 1, characterized in that, After ranking each candidate search request based on its probability, and before recommending search requests to the target object based on the ranking results, the process further includes: Based on the interest tags of the target object and each candidate search request, the first interaction representation information corresponding to each candidate search request is obtained respectively; According to the sorting result, each candidate search request is traversed sequentially, and the candidate search request with the lower probability is deduplicated among the two candidate search requests whose similarity between the corresponding first interaction representation information reaches the second threshold. The step of making a search request recommendation to the target object based on the ranking results includes: Among the remaining candidate search requests after deduplication, at least one candidate search request whose probability reaches the third threshold is recommended to the target object.

6. The method as described in claim 1, characterized in that, The first interaction representation information corresponding to each search request is obtained through the following methods: For each search request, perform the following operations: The interest tags of the search object corresponding to a search request and the search request are respectively input into the trained joint representation model to obtain the first deep representation information corresponding to the interest tags of the search object and the second deep representation information corresponding to the search request. The first depth representation information corresponding to the interest tag of the search object and the second depth representation information corresponding to the search request are subjected to feature interaction processing to obtain the first interaction representation information corresponding to the search request.

7. The method as described in claim 1, characterized in that, The joint representation model is trained in the following manner: Obtain a second training sample set, wherein each second training sample in the second training sample set includes: an interest tag and a sample search request of a sample object, relevant search results of the sample object for the sample search request, and real information on the sample object's satisfaction with the relevant search results; Based on the second training sample set, the joint representation model to be trained is trained iteratively multiple times to obtain a trained joint representation model; wherein, in one iteration, the following operations are performed: The selected second training samples are input into the joint representation model, and the corresponding sample object's satisfaction prediction information for the relevant search results is output. The parameters of the joint representation model are adjusted based on the differences between the predicted satisfaction information and the actual satisfaction information of each search result.

8. The method as described in claim 7, characterized in that, The true information regarding the satisfaction level of the sample objects with the relevant search results is determined based on the feedback behavior of the sample objects towards the relevant search results.

9. The method as described in claim 7, characterized in that, The step of inputting the second training samples selected from the second training sample set into the joint representation model to be trained, and outputting the corresponding satisfaction prediction information of the sample object for the relevant search results, includes: The second training samples selected from the second training sample set are respectively input into the joint representation model to be trained to obtain the first depth representation information corresponding to the interest tags of the sample objects in the second training samples, the second depth representation information corresponding to the sample search requests of the sample objects, and the third depth representation information corresponding to the relevant search results of the sample search requests. Based on the first depth representation information corresponding to the interest tags of the sample object, the second depth representation information corresponding to the search request of the sample object, and the third depth representation information, the sample object is predicted to obtain the satisfaction prediction information for the sample object.

10. The method as described in claim 9, characterized in that, The method of predicting the satisfaction prediction information of the sample object based on the first depth representation information corresponding to the interest tag of the sample object, the second depth representation information corresponding to the search request of the sample object, and the third depth representation information includes: The first depth representation information corresponding to the interest tag of the sample object and the second depth representation information corresponding to the search request of the sample object are subjected to feature interaction processing to obtain the first interaction representation information corresponding to the sample object; and the second depth representation information corresponding to the search request of the sample object and the third depth representation information corresponding to the relevant search results of the sample search request are subjected to feature interaction processing to obtain the second interaction representation information corresponding to the sample object. Based on the first and second interaction representation information of the sample object, a prediction is made to obtain the satisfaction prediction information for the sample object.

11. A search request recommendation device, characterized in that, include: The filtering unit is used to filter out multiple historical search requests that have a similarity of up to a first threshold with the target search request from a pre-built set of historical search requests based on the first interaction representation information corresponding to the current target search request of the target object, as candidate search requests related to the target object; The determining unit is configured to determine the probability of a corresponding candidate search request being the next search request for the target object based on the first interaction representation information of each candidate search request and each historical search request related to the target object; wherein, the first interaction representation information of each search request is obtained by interacting the interest tags of the corresponding search object and the search request based on a joint representation model; the joint representation model is used to jointly model the interest tags of the sample object, the sample search request, and the satisfaction of the sample object with the relevant search results for the sample search request, so that the representation of the interest tags and search request by the joint representation model can take into account the degree of satisfaction of the final search results with the object; the search request is a target search request, a candidate search request, or a historical search request; The recommendation unit is used to rank each candidate search request based on its probability and then recommend search requests to the target object based on the ranking results.

12. The apparatus as claimed in claim 11, characterized in that, The filtering unit is specifically used for: Obtain the similarity retrieval index corresponding to each historical search request in the historical search request set. The similarity retrieval index of each historical search request is obtained by vector compression of the first interaction representation information corresponding to the corresponding historical search request, which is determined based on the interest tags and historical search requests of the corresponding historical objects. Based on each similarity retrieval index and the first interaction representation information corresponding to the target search request, multiple historical search requests that have a similarity to the target search request reaching a first threshold are selected.

13. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program that, when executed by the processor, causes the processor to perform the steps of any of the methods described in claims 1 to 10.

14. A computer-readable storage medium, characterized in that, It includes a computer program that, when run on an electronic device, causes the electronic device to perform the steps of any of the methods described in claims 1 to 10.

15. A computer program product, characterized in that, The method includes a computer program stored in a computer-readable storage medium; when a processor of an electronic device reads the computer program from the computer-readable storage medium, the processor executes the computer program, causing the electronic device to perform the steps of any one of claims 1 to 10.

Images