Search result reordering method, apparatus, device, and computer storage medium

Abstract

The application discloses a search result reordering method and device, equipment and a computer storage medium, and relates to the technical field of media content search. In the method, a suitable timing is intelligently selected to trigger a reordering process according to real-time interaction between a target object and media content during a search session, the un-consumed media content in a search result list of the search session is reordered, and the reordered search result list is presented to the target object. The method adjusts the order of each media content in the search result according to instant interest, can improve the accuracy of the result presentation of the search session, improves the content searching efficiency, and solves the problem that the interaction behavior feedback of the target object is not timely in the media content search scene. The order of the search result is adjusted in time to meet the search intention of the user.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computers, in particular to the technical field of search, and provides a search result reordering method and device, equipment and a computer storage medium. BACKGROUND

[0002] With the development of Internet technology, short videos have gradually become a common way of information acquisition. A short video refers to a video communication content with a duration of usually less than 5 minutes in a short video platform. Generally, due to the openness of the short video platform, a large number of creators upload short videos through the short video platform, so there are a large number of short videos in the short video platform. Therefore, when searching for short videos, how to select short videos that meet the search conditions from the large number of short videos and understand the search intent of the target object to place the result that the target object wants most in the front position to perform accurate search result recommendation is always a problem that the short video platform is concerned about.

[0003] The short video search technology used in the related art is usually divided into stages of recall, rough sorting, fine sorting and mixed sorting, and after completion, a short video list is returned to the client as a search result to present to the target object.

[0004] However, the large number of short video libraries may contain different types of short videos related to the search conditions. Similarly, the returned short video list may also contain different types of short videos, and the search intent of the target object may be one of the types. For example, when the target object searches for "show", short videos related to "cool effects" or "show rice" may appear, although the same search keywords are used, but belong to different types of short videos. Therefore, the mixed short videos of multiple types may make the target object unable to quickly find the short video that he wants to watch, greatly increasing the time spent in searching for the short video that he wants to watch, and the search efficiency is not high.

[0005] Similarly, in addition to short videos, the search of other forms of media content also has the above-mentioned problems. SUMMARY

[0006] The embodiments of the present application provide a search result reordering method, device, equipment and computer storage medium, which are used to improve the accuracy of search result presentation and improve the content search efficiency.

[0007] In one aspect, a search result reordering method is provided, and the method comprises:

[0008] displaying a search result page corresponding to a search session, the search result page displaying a first content sequence obtained based on a search keyword corresponding to the search session;

[0009] obtaining at least one real-time interaction behavior of a target object with respect to the search result page;

[0010] when the at least one real-time interaction behavior satisfies a reordering trigger condition, updating the search result page, the updated search result page displaying a second content sequence, the second content sequence being obtained by reordering each media content to be sorted included in the first content sequence, the media content to be sorted being other media content in the first content sequence except for target media content involved in the at least one real-time interaction behavior.

[0011] In one aspect, a search result reordering apparatus is provided, and the apparatus includes:

[0012] a page display unit configured to display a search result page corresponding to a search session, the search result page displaying a first content sequence obtained based on a search keyword corresponding to the search session;

[0013] an obtaining unit configured to obtain at least one real-time interaction behavior of a target object with respect to the search result page;

[0014] a page updating unit configured to, when the at least one real-time interaction behavior satisfies a reordering trigger condition, update the search result page, the updated search result page displaying a second content sequence, the second content sequence being obtained by reordering each media content to be sorted included in the first content sequence, the media content to be sorted being other media content in the first content sequence except for target media content involved in the at least one real-time interaction behavior.

[0015] Optionally, the page updating unit is specifically configured to:

[0016] if the at least one real-time interaction behavior includes at least one of the following interaction behaviors, update the search result page:

[0017] trigger display of a sub-result page corresponding to target media content in the search result page;

[0018] trigger switching of displayed media content in the search result page;

[0019] perform forward interaction in the sub-result page, the forward interaction representing a behavior of the target object being interested in corresponding media content;

[0020] perform negative interaction in the sub-result page, the negative interaction representing a behavior of the target object being uninterested in corresponding media content;

[0021] return from the sub-result page to the search result page.

[0022] Optionally, the page updating unit is specifically configured to:

[0023] obtain reordering trigger prediction data, the reordering trigger prediction data comprising at least one of real-time interaction behavior data of the at least one real-time interaction behavior, historical interaction behavior data of the target object, a current load condition of the terminal device, and a current network condition of the terminal device;

[0024] input the reordering trigger prediction data into a reordering trigger prediction model to obtain a prediction result indicating whether to trigger reordering;

[0025] if the prediction result indicates to trigger reordering, update the search result page.

[0026] Optionally, the page updating unit is specifically configured to:

[0027] when the at least one real-time interaction behavior meets a reordering trigger condition, initiate a reordering request to a server, the reordering request carrying an end feature corresponding to the target media content, the end feature being extracted based on real-time interaction behavior data of the at least one real-time interaction behavior and historical interaction behavior data of the target object;

[0028] receive the second content sequence returned by the server in response to the reordering request, the second content sequence being obtained by the server based on the end feature using a reordering model;

[0029] update the search result page based on the second content sequence.

[0030] Optionally, the apparatus further comprises a reordering unit configured to:

[0031] construct at least one interaction content sequence based on target media content corresponding to different interaction types in the at least one real-time interaction behavior, each interaction content sequence corresponding to one interaction type;

[0032] construct a to-be-ordered content sequence based on each to-be-ordered media content;

[0033] determine, using a reordering model, a degree of interest of the target object in each to-be-ordered media content in the to-be-ordered content sequence based on the search keyword, a trigger condition type corresponding to the at least one real-time interaction behavior, and the at least one interaction content sequence;

[0034] the page updating unit is specifically configured to obtain the second content sequence based on the obtained degrees of interest, and update the search result page based on the second content sequence.

[0035] Optionally, the reordering unit is specifically configured to:

[0036] Encode the search keyword, the trigger condition type, the at least one interactive content sequence, and the to-be-sequenced content sequence respectively to obtain corresponding encoded feature vectors;

[0037] Based on the obtained encoded feature vectors, respectively perform attention mechanism-based interaction processing on the encoded feature vectors of the to-be-sequenced content sequence to obtain a plurality of attention feature vectors;

[0038] Concatenate the plurality of attention feature vectors with the encoded feature vectors of the search keyword and the encoded feature vectors of the trigger condition type to obtain an output feature vector;

[0039] Map the output feature vector into a plurality of index vectors respectively, and determine the respective degrees of interest based on the plurality of index vectors. Each index vector corresponds to an index parameter dimension, and a value in each index vector represents the degree of interest of the target object in the corresponding to-be-sequenced media content in the corresponding index parameter dimension.

[0040] Optionally, the reordering unit is specifically configured to:

[0041] Encode any content sequence in the at least one interactive content sequence and the to-be-sequenced content sequence by the following process:

[0042] Respectively obtain a corresponding basic feature vector for each media content in the any content sequence;

[0043] Respectively obtain a position feature vector corresponding to each of the media contents based on the position of each of the media contents in the any content sequence;

[0044] Superimpose the content feature vector and the position feature vector corresponding to each of the media contents to obtain a content feature vector corresponding to each of the media contents;

[0045] Perform feature extraction on the obtained content feature vectors by using a serialization feature extraction method to obtain an encoded feature vector corresponding to the any content sequence.

[0046] Optionally, the device further comprises an end feature extraction unit configured to:

[0047] Extract an end feature corresponding to each target media content based on real-time interaction behavior data of the at least one real-time interaction behavior and historical interaction behavior data of the target object, the end feature being used to represent an interaction tendency of the target object for the search session;

[0048] The reordering unit is specifically configured to:

[0049] For any target media content, the corresponding basic feature vector is obtained through the following steps:

[0050] Based on the content category to which the any target media content belongs, a corresponding category feature vector is obtained;

[0051] Based on the terminal feature corresponding to the any target media content, feature extraction is performed to obtain a corresponding terminal feature vector;

[0052] The fine arrangement feature vector of the any target media content at the fine arrangement stage is obtained;

[0053] The category feature vector, the terminal feature vector, and the fine arrangement feature vector are concatenated to obtain a corresponding basic feature vector.

[0054] Optionally, the terminal feature extraction unit is specifically configured to:

[0055] Based on the real-time interaction behavior data, a first type of feature set of the corresponding target media content is extracted;

[0056] From the historical interaction behavior data, a respective associated data set corresponding to each target media content is determined, and the associated data set of each target media content includes historical interaction data corresponding to media content with the same attribute;

[0057] Based on the obtained each associated data set, a second type of feature set corresponding to each target media content is extracted;

[0058] Based on the obtained each first type of feature set and second type of feature set, a terminal feature of the corresponding target media content is obtained.

[0059] Optionally, the reordering unit is specifically configured to:

[0060] For the at least one interactive content sequence, the following steps are respectively performed:

[0061] For an interactive content sequence, based on the similarity between the encoding feature vector of the interactive content sequence and the encoding feature vector of the content sequence to be sorted, a corresponding attention weight is determined;

[0062] Based on the obtained attention weight and the encoding feature vector of the interactive content sequence, a corresponding attention feature vector is obtained.

[0063] Optionally, the reordering unit is specifically configured to:

[0064] construct an exposure content sequence based on the target media content displayed in the search result interface;

[0065] construct a positive feedback sequence based on the target media content corresponding to the positive interaction behavior in the at least one real-time interaction behavior and the historical interaction behavior;

[0066] construct a negative feedback sequence based on the target media content corresponding to the negative interaction behavior in the at least one real-time interaction behavior and the historical interaction behavior.

[0067] In an aspect, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any of the above methods when executing the computer program.

[0068] In an aspect, a computer storage medium is provided, having a computer program stored thereon, wherein the computer program is executable by a processor to implement the steps of any of the above methods.

[0069] In an aspect, a computer program product is provided, including a computer program stored in a computer readable storage medium. A processor of a computer device reads the computer program from the computer readable storage medium, and the processor executes the computer program, so that the computer device executes the steps of any of the above methods.

[0070] In the embodiments of the present application, after the search result page corresponding to the search session is displayed, at least one real-time interaction behavior of the target object on the search result page is obtained, and when the at least one real-time interaction behavior meets the reordering trigger condition, the search result page is updated, and the first content sequence originally displayed is updated to the second content sequence, which is obtained by reordering each media content not involved in the at least one real-time interaction behavior in the first content sequence. In this way, in a search session, the real-time interaction behavior represents the immediate interest of the target object, so that the order of each media content in the search result is adjusted according to the immediate interest, which can improve the accuracy of the result of the search session, timely meet the search intention of the user, improve the content searching efficiency, and improve the timeliness of the feedback to the interaction behavior of the target object. BRIEF DESCRIPTION OF DRAWINGS

[0071] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related art, the drawings needed in the embodiments or the related art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to the provided drawings without creative labor.

[0072] Figure 1 An application scenario provided for an embodiment of the present application is shown in the following figure;

[0073] Figure 2 An architecture diagram of a media content search system provided for an embodiment of the present application is shown in the following figure;

[0074] Figure 3 A flow diagram of a search result reordering method provided for an embodiment of the present application is shown in the following figure;

[0075] Figures 4a-4d A presentation page diagram of a client provided for an embodiment of the present application is shown in the following figure;

[0076] Figure 5 An extraction process diagram of an end feature provided for an embodiment of the present application is shown in the following figure;

[0077] Figure 6 An interaction diagram corresponding to an end feature provided for an embodiment of the present application is shown in the following figure;

[0078] Figure 7 A flow diagram of end-cloud collaborative execution provided for an embodiment of the present application is shown in the following figure;

[0079] Figure 8 A flow diagram of a reordering process provided for an embodiment of the present application is shown in the following figure;

[0080] Figure 9 A model structure diagram of a reordering model provided for an embodiment of the present application is shown in the following figure;

[0081] Figure 10 A processing flow diagram of a reordering model provided for an embodiment of the present application is shown in the following figure;

[0082] Figure 11 A flow diagram of feature encoding for a content sequence provided for an embodiment of the present application is shown in the following figure;

[0083] Figure 12 A flow diagram of a feature vector processing process provided for an embodiment of the present application is shown in the following figure;

[0084] Figure 13 A structure diagram of a Transformer network provided for an embodiment of the present application is shown in the following figure;

[0085] Figure 14 A structure diagram of a search result reordering apparatus provided for an embodiment of the present application is shown in the following figure;

[0086] Figure 15 A component structure diagram of a computer device provided for an embodiment of the present application is shown in the following figure;

[0087] Figure 16 A component structure diagram of another computer device to which an embodiment of the present application is applied is shown in the following figure. Detailed Implementation

[0088] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application. Unless otherwise specified, the embodiments and features in the embodiments of this application can be arbitrarily combined with each other. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0089] It is understood that in the following specific embodiments of this application, data related to the target object, such as interactive behavior data, is involved. When the various embodiments of this application are applied to specific products or technologies, relevant licenses or consents need to be obtained, and the collection, use, and processing of relevant data need to comply with the relevant laws, regulations, and standards of the relevant countries and regions. For example, when it is necessary to obtain relevant data, relevant volunteers can be recruited and relevant agreements for volunteer data authorization can be signed, and then the data of these volunteers can be used for implementation; or, it can be implemented within the authorized scope of an organization, and relevant recommendations can be made to internal members by using the data of internal members of the organization to implement the following implementation methods; or, the relevant data used in the specific implementation are all simulated data, such as simulated data generated in a virtual scene.

[0090] To facilitate understanding of the technical solutions provided in the embodiments of this application, some key terms used in the embodiments of this application will be explained below:

[0091] Media content: Media content refers to information transmitted through multimedia technology. In the embodiments of this application, media content can refer to any possible content that can be recommended to the target object when the target object is conducting a search, including but not limited to images, articles (such as news and information), audio, video (including long videos and short videos), and products.

[0092] End-to-cloud collaboration: In the search re-ranking scenario of this application embodiment, end-to-cloud collaboration refers to a re-ranking framework in which the client performs feature extraction and determines the timing of triggering re-ranking, while the cloud is responsible for the actual re-ranking logic. The two perform their respective duties and work together to complete the re-ranking process.

[0093] Search session: refers to a complete search process, including the process from triggering the search by the target object (for example, initiating the search request by clicking the search button) to exiting the search result (for example, exiting the search result page or changing the search keyword, etc.). For example, after clicking the search button, the search result page is presented, and the process continues until the search result page is exited.

[0094] Real-time interaction behavior: refers to a series of interaction behaviors generated by the target object in the process of consuming video content, including but not limited to up-sliding, down-sliding, page turning, clicking, liking, following the author, commenting, and fast sliding. Real-time interaction behavior can be defined differently in actual application scenarios. Here, the video scenario is taken as an example:

[0095] In one possible implementation, the real-time interaction behavior may, for example, be the interaction behavior generated in the consumption process of the current video, which may include, for example, the behaviors of liking, commenting, pulling the progress bar, playing the progress, forwarding, etc. in the current video. Correspondingly, the historical interaction behavior may be the interaction behavior before the real-time interaction behavior, for example, the interaction behavior generated in the consumption process of other videos before the current video in the video sequence consumed by the target object. The other videos may be other consumed videos in the search session except the current video, or may also include other consumed videos outside the search session, that is, the videos consumed before entering the search session.

[0096] In another possible implementation, the real-time interaction behavior may, for example, be the interaction behavior generated in the search session, and the historical interaction behavior may be the interaction behavior generated before the search session.

[0097] In another possible implementation, the real-time interaction behavior may refer to the interaction behavior performed within a short time from the current time, corresponding to the historical interaction behavior performed by the target object a long time ago from the current time. The time threshold can be set by the user.

[0098] Search session average duration: refers to the average duration of a search session, which is a measure of search results. In different scenarios, the representation of search session average duration is different. For example, for short videos, the short video platform expects to give the target object more interesting short videos, so the longer the target object stays on the search result page, the longer the search session average duration, which means that the short video platform's short video is more attractive, and the target object's search experience and service quality are better. For other scenarios, such as search in browsers, the target object's intention may be to search for keyword-related results, and the target object expects to search for relevant content as soon as possible, so the shorter the search session average duration, the better, which means that the target object's search efficiency is higher.

[0099] Multi-task learning model (Multi-gate Mixture-of-Experts, MMOE): a multi-task learning model can improve the learning efficiency and quality of each task by learning the relationship and difference between different tasks. The multi-task learning framework widely adopts the shared-bottom structure, and the hidden layers of the bottom are shared between different tasks. The core idea of MMoE is ensemble learning, which explicitly learns the relationship between subtasks by introducing a Mixture-of-Experts (MoE) layer in multi-task learning, and uses a threshold network to optimize each task.

[0100] Transformer: Transformer is a deep learning model that uses a self-attention mechanism to weight the importance of each part of the input data differently. It was originally used in natural language processing (NLP) and has since been extended to other fields.

[0101] Content sequence: refers to a sequence of media content, in which media content is arranged in a certain order, such as according to the target object's interest level or interaction time, etc.

[0102] Consumption: in the embodiments of the present application, consumed media content represents the target object's interactive behavior with the presented media content. For example, in the case of short videos, this can include exposing short videos to target objects, target objects clicking to watch short videos, target objects commenting on short videos, and other behaviors, all of which are a kind of consumption of short videos.

[0103] The embodiments of the present application relate to artificial intelligence and machine learning (Machine Learning, ML) technology, and are mainly designed based on machine learning in artificial intelligence.

[0104] Artificial intelligence is the theory, method, technology and application system of using digital computer or machine controlled by digital computer to simulate, extend and expand human intelligence, perceive environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology of computer science, which attempts to understand the essence of intelligence and produce a new intelligent machine that can react in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines, so that machines have the functions of perception, reasoning and decision-making.

[0105] Artificial intelligence technology is a comprehensive discipline, involving a wide range of fields, both hardware and software technologies. Artificial intelligence basic technologies generally include sensors, special artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction system, mechatronics and other technologies. Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology and machine learning / deep learning and other major directions.

[0106] Machine learning is a multi-disciplinary subject, involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. It is a subject that studies how computers simulate or implement human learning behavior to acquire new knowledge or skills, reorganize existing knowledge structure to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental approach to making computers intelligent, and its applications are widespread in various fields of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and rule-based learning.

[0107] Machine learning is the core of artificial intelligence and the fundamental approach to making computers intelligent, and its applications are widespread in various fields of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning and other technologies. Artificial neural network (ANN) is an abstraction of human brain neuron network from the perspective of information processing, which establishes a certain simple model and forms different networks according to different connection modes. Neural network is an operation model composed of a large number of nodes (or neurons) connected with each other, each node represents a certain output function, called activation function. The connection between each two nodes represents a weight value for the signal passing through the connection, which is called weight, which is equivalent to the memory of artificial neural network. The output of the network depends on the connection mode, weight value and excitation function of the network, while the network itself is usually an approximation of a certain algorithm or function in nature, or a representation of a logical strategy.

[0108] The embodiments of the present application relate to reordering and scoring media content such as short videos, thereby obtaining a new content sequence after reordering based on the scores of each media content to be reordered, and presenting it to the target object. In addition, the triggering time of the reordering is also predicted. Taking the reordering score as an example, when the media content is reordered and scored, an artificial neural network model based on deep learning needs to be used, that is, the machine learning method is used to obtain a reordering model for reordering and scoring media content. The reordering model has the ability to process and understand the search intent of the target object based on machine learning, for example, measures the long-term and short-term interests of the target object based on the relevance between the historical interaction behavior and the real-time interaction behavior of the target object, the search keyword, the reordering trigger type and the real-time interaction behavior, which expresses the search intent of the target object in this search session, thereby measuring the interest degree of the target object for each media content to be reordered, and obtaining the score of each media content to be reordered.

[0109] Specifically, the reordering score in the embodiments of the present application can be divided into two parts, including a training part and an application part. The training part relates to the technical field of machine learning. In the training part, the artificial neural network model (i.e. the reordering model mentioned later) is trained through machine learning technology, so that the artificial neural network model is trained based on the historical interaction behavior and the real-time interaction behavior, the search keyword, the reordering trigger type and other training sample data given in the embodiments of the present application, and the model parameters are continuously adjusted through the optimization algorithm until the model converges. The application part uses the artificial neural network model trained in the training part to score each media content to be reordered when reordering is triggered in the actual use process, so as to reorder each media content to be reordered. In addition, it should be noted that the artificial neural network model in the embodiments of the present application can be online training or offline training, which is not limited here. In this paper, offline training is taken as an example for illustration.

[0110] In the related art, the search result is returned to the client, and the order of each short video in the sequence is fixed. No matter what interaction behavior the target object performs in the search session, the relative order of the short video will not change. In this way, the target object may need to spend more time to find the short video he needs. In addition, if the search result page presents too many short videos that do not match the search intent of the target object, the target object may quickly exit the search session, which also reflects the poor experience of the short video platform.

[0111] Similarly, in addition to short videos, other forms of media content search also have the above-mentioned problems.

[0112] In an actual search process, the target object usually generates a large number of interactive behaviors such as likes, comments and page turning for a presented search result page, and these interactive behaviors contain various positive feedback and negative feedback, which reflect the intention and preference of the target object. However, the related art does not consider the influence brought by these interactive behaviors, so the order of the content sequence obtained by searching cannot be changed in the search process, and the search intention of the user cannot be met in time.

[0113] Based on this, the embodiment of the present application provides a search result reordering method. In the method, after displaying a search result page corresponding to a search session, at least one real-time interactive behavior of a target object for the search result page is recorded during the duration of the search session. When the at least one real-time interactive behavior meets a reordering trigger condition, the search result page is updated, and a first content sequence originally displayed is updated to a second content sequence. The second content sequence is obtained by reordering each media content not involved in the at least one real-time interactive behavior in the first content sequence. In this way, during a search session, real-time interactive behavior represents the immediate interest of the target object, so that the order of each media content in the search result is adjusted according to the immediate interest, which can improve the accuracy of the result presentation of the search session, meet the search intention of the user in time, improve the content search efficiency, and also improve the timeliness of feedback on the interactive behavior of the target object.

[0114] In addition, the embodiment of the present application records the real-time interactive behavior of the target object, and comprehensively intelligently decides whether to trigger reordering in combination with the current device condition and the historical interactive behavior of the target object, so as to improve the accuracy of the reordering trigger timing and meet the use demand of the target object.

[0115] In terms of reordering, the embodiment of the present application provides a reordering model, which can take into account the short-term and long-term interests of the target object to reorder media content, improve the accuracy of the content sequence obtained by reordering, and help the target object to quickly find the required media content and meet the search intention of the user in time.

[0116] The following describes the application scenarios to which the technical solutions of the embodiments of the present application can be applied. It should be noted that the following application scenarios are only used to illustrate the embodiments of the present application and are not limiting. In the specific implementation process, the technical solutions provided by the embodiments of the present application can be flexibly applied according to actual needs.

[0117] The scheme provided by the embodiment of the present application can be applied to most content search scenarios, such as short video search scenarios. Figure 1As shown, an application scenario provided by the embodiment of the present application is shown, and the scenario can include a terminal device 101 and a server 102.

[0118] The terminal device 101 can be any device with the function of presenting and searching media content, such as a mobile phone, a tablet computer (PAD), a notebook computer, a desktop computer, a smart television, a smart vehicle device, a smart wearable device, and the like. The terminal device 101 can be installed with a target application, which has the function of searching and presenting media content, such as an instant messaging application, a music application, a video application, a short video application, a news application, a shopping application, and the like. The application involved in the embodiment of the present application can be a software client, a web page, a small program, or the like. The server 102 is a server corresponding to the software or the web page, the small program, or the like, and the specific type of the client is not limited. The server 102 can be a standalone physical server, a server cluster or a distributed system composed of multiple physical servers, a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content distribution networks (CDNs), and big data and artificial intelligence platforms, but is not limited thereto.

[0119] It should be noted that the search result reordering method in the embodiment of the present application can be performed by the terminal device 101 alone or by the server 102 and the terminal device 101 together. For example, after the terminal device 101 displays the search result page corresponding to the search session, at least one real-time interaction behavior of the target object on the search result page is recorded during the duration of the search session. When the at least one real-time interaction behavior meets the reordering trigger condition, a second content sequence is obtained to update the search result page, and the first content sequence originally displayed is updated to the second content sequence. That is, the reordering process is performed in the terminal device 101. Alternatively, when the terminal device 101 determines that the at least one real-time interaction behavior meets the reordering trigger condition, the terminal device 101 triggers a reordering request to the server 102. The server 102 performs reordering to obtain a second content sequence, and the terminal device 101 updates the search result page based on the second content sequence. The present application does not make specific limitations here, and the following examples are mainly described by taking the server 102 and the terminal device 101 together as an example.

[0120] Taking the server 102 and the terminal device 101 jointly performing as an example, the server 102 and the terminal device 101 can each include one or more processors, memories, and I / O interfaces for interaction, and the like. In addition, the server 102 can further be configured with a database, which can be used to store historical interaction behavior data of a target object and trained model parameters, and the like. The memories of the server 102 and the terminal device 101 can further store program instructions required for respective execution in the search result reordering method provided in the embodiments of the present application. When the program instructions are executed by the processors, the search result reordering process provided in the embodiments of the present application can be implemented. For example, the terminal device 101 can store program instructions of processes such as search result page display, determination of whether a reordering trigger condition is met, and search result page update, to be executed by the processor of the terminal device 101 to implement these processes. The server 102 can store program instructions of the reordering process to be executed by the processor of the server 102 to implement the reordering process.

[0121] In a possible implementation, the method of the embodiments of the present application can be applied to a short video search scenario. Since the timing determination of reordering and the reordering process both involve interaction behavior data of a target object, in order to implement reordering of search results to the target object, the target object can be requested to authorize collection and use permissions of data related to reordering, such as real-time interaction behavior data and historical interaction behavior data. When the target object allows, the related data is input into the reordering model provided in the embodiments of the present application to predict the interest degrees of each media content to be sorted of the target object, and each media content to be sorted is reordered and presented to the target object based on this, so as to effectively improve the accuracy of the search session result presentation, timely meet the search intention of the user, thereby improving the content searching efficiency, and also improving the timeliness of feedback to the interaction behavior of the target object.

[0122] The implementation process of the method in other search scenarios is similar to that in the short video search scenario described above, and thus will not be described in detail here.

[0123] In the embodiments of the present application, the terminal device 101 and the server 102 can be directly or indirectly connected through one or more networks 103. The network 103 can be a wired network or a wireless network. For example, the wireless network can be a mobile cellular network or a wireless fidelity (WIFI) network. Of course, other possible networks can also be used, and the embodiments of the present application do not limit this.

[0124] It should be noted that, Figure 1The number of terminal devices and servers is not limited in practice, and is not specifically limited in the embodiments of the present application.

[0125] As shown in Figure 2 An architecture diagram of a media content search system provided by the embodiments of the present application is shown, which can include a client part and a cloud part. The client part includes an end feature acquisition module and a reordering trigger module. The cloud part can include a reordering module. When the above method steps of the embodiments of the present application are executed by the terminal device 101, each module is deployed in the terminal device 101. When the above method steps of the embodiments of the present application are executed by the terminal device 101 and the server 102 together, the end feature acquisition module and the reordering trigger module can be deployed in the terminal device 101, and the reordering module can be deployed in the server 102. Figure 2 Specifically, this is exemplified.

[0126] The client presents a corresponding search result page according to the content sequence. The end feature acquisition module generates and continuously maintains the corresponding end feature of each media content in the search result page according to the real-time interaction of the target object, to assist the decision of intelligent triggering and the input feature of reordering. The reordering trigger module decides whether the current state needs to trigger a rearrangement according to the collected end feature. If rearrangement is selected, a module call is initiated to the rearrangement module in the cloud. The rearrangement module reorders the media content to be sorted according to the current interaction context, end feature and historical interaction behavior, and returns to the client to update the content sequence, so as to achieve the effect of updating the search result page.

[0127] The search result reordering method provided by the exemplary embodiments of the present application is described below in combination with the above-described application scenarios and with reference to the accompanying drawings. It should be noted that the above-described application scenarios are only shown to facilitate understanding of the spirit and principles of the present application, and the embodiments of the present application are not limited in this respect.

[0128] As shown in Figure 3 An architecture diagram of a media content search system provided by the embodiments of the present application is shown, which can include a client part and a cloud part. The client part includes an end feature acquisition module and a reordering trigger module. The cloud part can include a reordering module. When the above method steps of the present application are executed by the terminal device 101, each module is deployed in the terminal device 101. When the above method steps of the present application are executed by the terminal device 101 and the server 102 together, the end feature acquisition module and the reordering trigger module can be deployed in the terminal device 101, and the reordering module can be deployed in the server 102.

[0129] Step 301: Display a search result page corresponding to a search session. The search result page displays a first content sequence obtained based on a search keyword corresponding to the search session.

[0130] In the embodiments of the present application, when the target object wants to search for certain media content in the client, the target object can input a search keyword in the search interface provided by the client to trigger the search session. For example, after the target object inputs the search keyword in the search box and confirms to search, the search session starts, and then the client initiates a search request to the cloud. The cloud obtains a first content sequence for the search keyword based on the search keyword of the target object and the preference information of the target object through the stages of recall, coarse sorting, fine sorting, and mixed sorting, and returns the first content sequence to the client to present to the target object. It should be noted that the first content sequence can be a mixture of media content of multiple forms, or a single form of media content. Here, the form mainly refers to the carrying manner of media content information, such as video, audio, text, image, etc. The specific content sequence can be set according to the actual needs of the client.

[0131] Referring to Figure 4a , it is a schematic diagram of a search result page presented by the client. After the client receives the first content sequence, the corresponding media content is displayed to the target object in the search result page. Generally, when displayed in the search result page, the media content can be displayed according to different content forms, such as video and audio displayed in different regions; or the media content can be displayed according to the priority, such as displaying topic content first, such as the topic “showing off” in Figure 4a ; or the media content can be displayed according to the score in the fine sorting or mixed sorting stage, such as the descending order of the sorting scores of videos 1-4 in Figure 4a .

[0132] In actual application, there are multiple ways to display the first content sequence in the search result page. One way is to display all the content contained in the first content sequence, and the other way is to display only part of the content in the first content sequence. Referring to Figure 4a , the videos are displayed in a video double-column stream in the search result page. Due to the limited display area, only a few videos can be displayed, and currently only videos 1-4 are displayed. The remaining videos can be displayed gradually in the form of a sliding window or a drop-down box.

[0133] It should be noted that when displaying the media content, the related information of the media content can be displayed. Taking a short video as an example, the cover, introduction, video author (i.e. copyright owner, CP), and play count of the short video can be displayed when the short video is displayed.

[0134] Step 302: Obtain at least one real-time interaction behavior of the target object on the search result page.

[0135] In the embodiments of the present application, the search result page can include a search result main page and a sub-result page corresponding to each media content in the main page, and thus the interactive behaviors performed on the search result main page and each sub-result page can be included in the at least one real-time interactive behavior performed on the search result page.

[0136] Specifically, during the duration of a search session, the target object can perform interactive behaviors on the search result page to find the media content that the target object wants, and any interactive behavior of the target object during this process is acquired as a data basis for triggering reordering subsequently.

[0137] Among them, the real-time interactive behavior can include the following behaviors:

[0138] 1. The following behaviors performed on the search result main page:

[0139] (1) The behavior of triggering display of a sub-result page corresponding to the target media content, for example, clicking the target media content. Referring to FIG. 6, triggering display of a sub-result page can include the behavior of clicking a topic or a video to display the corresponding sub-result page. Figure 4b

[0140] (2) The behavior of switching display of media content on the search result main page. Referring to FIG. 7, it can include the behaviors of pulling up, pulling down, swiping left, swiping right, and page turning on the search result main page. Figure 4b

[0141] 2. Referring to FIG. 8, the following behaviors performed on the sub-result page: Figure 4c

[0142] (1) Adding the current media content to the target object's collection.

[0143] (2) Liking the current media content.

[0144] (3) Commenting on the current media content.

[0145] (4) Forwarding the current media content.

[0146] (5) Following the author corresponding to the current media content.

[0147] (6) Performing media content switching operations on the sub-result page, such as the behaviors of pulling up, pulling down, swiping left, and swiping right.

[0148] (7) Other non-operation behaviors, for example, the user stays on the current sub-result page for less than 3 seconds, which can be referred to as a fast swipe behavior, or the play completion ratio of the video on the current sub-result page exceeds a certain ratio threshold, which can be referred to as a high complete play behavior. ​​​

[0149] Step 303: when the at least one real-time interaction behavior meets the reordering trigger condition, updating the search result page, the updated search result page displays a second content sequence, the second content sequence is obtained by reordering each media content to be sorted contained in the first content sequence, and the media content to be sorted is other media content in the first content sequence except for the target media content involved in the at least one real-time interaction behavior.

[0150] In the embodiment of the application, it is considered that when the server generates the first content sequence, in addition to the current search keyword, the preference data of the target object and the like are also referred to for comprehensive consideration. The first content sequence obtained in this way has a certain accuracy, but the preference data mainly represents the long-term interest of the user, and the real-time interaction behavior of the target object in the current search session represents the short-term interest of the target object and can better reflect the search intention of the target object in the search session. These interaction behaviors are very necessary to be considered.

[0151] Therefore, the embodiment of the application determines whether the reordering needs to be triggered by the real-time interaction behavior of the target object in the search session, and when the reordering needs to be triggered, the other media content in the first content sequence which is not involved in the at least one real-time interaction behavior is reordered to obtain the second content sequence, and the search result page is updated based on this.

[0152] Taking the reordering triggered by the click behavior as an example, when the target object clicks to play the video 2, it is determined that the reordering trigger condition is met, and then the sub-result page corresponding to the video 2 is entered, and when returning from the search result page to the search result page, the updated search result page is displayed. Referring to FIG. 4, in the original search result page, the videos 1-4 are displayed in sequence, that is, the next video of the video 2 in the video list is the video 3, which is quite different from the content of the current video 2, so that the reordering of the search result is triggered once in the process of the target object clicking the video 2 for browsing. When the target object ends the consumption of the video 2 and returns to the search result page, the updated search result page is displayed, as shown in FIG. 5. Figure 4d Figure 4d As shown in FIG. 5, the next video of the video 2 has been replaced by another video 7 which has higher relevance with the video 2, and the original video 3 is moved to the next position. Of course, this is an example of video, and the same is true for other forms of media content, so the other forms of media content are not described again.

[0153] ​For example, when the preloading of the next page is triggered on the search result page, if the first content sequence contains a large amount of media content and the search result page cannot display all the media content, the media content can be displayed in pages. When the target object finishes browsing the media content on the current page and wants to flip to the next page, it means that the target object is not interested in the media content on the current page. At this time, the reordering can be triggered, and the subsequent media content can be reordered, and then the subsequent page can be displayed in the new order. For example, the video sequence of the next page in the first content sequence is video 5678, and after reordering, the video sequence in the second content sequence can be changed to 8697, but since the video 5678 has not been displayed, the change is essentially imperceptible to the target object.

[0154] Of course, when the reordering is triggered by other trigger types, the process is similar to the above process. As can be seen, adjusting the search results of the media content according to the user's intention can help improve the accuracy of the search session result presentation, timely meet the user's search intention, improve the content search efficiency, and also improve the timeliness of the feedback of the target object's interactive behavior.

[0155] In the embodiments of the present application, considering whether to trigger reordering and the data basis required for the reordering process, the end features of each media content in the first content sequence are extracted during the search session. The end feature extraction can be performed on all media content in the first content sequence or on part of the media content in the first content sequence. For example, considering that the end feature mainly comes from a series of operations generated in the interactive behavior of the target object, the object of extracting the end feature is mainly the media content consumed by the target object, such as the media content exposed to the target object.

[0156] In the embodiments of the present application, the end feature of each media content can be collected and maintained by the end feature collection module shown in Figure 2 Specifically, the end feature collection module can extract the end feature corresponding to each target media content based on the real-time interactive behavior data of at least one real-time interactive behavior during the search session and the historical interactive behavior data of the target object. The end feature is generated in the consumption process of the target object to the media content and can express the interactive tendency of the target object for the search session.

[0157] According to the category of the interactive behavior, the end feature can include a first feature set and a second end feature set, and then, as shown in Figure 5

[0158] ​At least one real-time interaction behavior in the search session duration is directly able to reflect the interest of the target object in the search session, and thus a first feature set of the corresponding target media content, also called direct feature, can be extracted from real-time interaction behavior data generated by the real-time interaction behavior.

[0159] Here, the real-time interaction behavior is an interaction behavior performed for the current media content, and the historical interaction behavior is an interaction behavior performed for other consumed media content before the current media content in the search session. The direct feature is obtained by the target object in the current page browsing behavior, as shown in Figure 6 The lower half of the interaction diagram, the like operation, comment, attention, and play progress of the target object in the play page (i.e., the sub-result page) will generate direct features, which are used to describe the consumption state of the current video. The consumption state can directly indicate whether the target object is interested in the current video, and can be used as an important input feature in the reordering process. As shown in Table 1 below, the direct features provided by the embodiments of the present application include the following feature list:

[0160] Feature name Feature ID Feature meaning is_cur_play 0 Whether the current video is played (0 / 1) is_click 1 Whether the result page is clicked (0 / 1) is_quickSlash 2 Whether the play page is fast-slid (0 / 1) is_dianZan 3 Whether the play page is liked (0 / 1) is_guanZhu 4 Whether the play page is followed (0 / 1) is_comment 5 Whether the play page is commented (0 / 1) complete_rate 6 Play page play completion degree (float value between 0-1)

[0161] Table 1

[0162] As shown in Figure 6 , the interaction diagram corresponding to the end feature. The feature ID in Table 1 above corresponds to the ID identified in Figure 6 . Taking a video as an example, as shown in Figure 6 , for a video, the feature ID 0 represents whether the video is the current play video. If it is the current play video, the corresponding feature value is 1, otherwise the feature value is 0. The feature ID 1 represents whether the video is clicked in the search result page. If it is clicked, the feature value is 1, otherwise the feature value is 0. Other features are similar.

[0163] As shown in Figure 6 , when the target object performs a certain real-time interaction behavior, the end feature corresponding to the media content it targets can be updated. For example Figure 6 , the video 1 is clicked, and the feature 0 of the video 1 is updated to 1, and the feature 1 is updated to 1.

[0164] In the embodiments of this application, it is also considered that in addition to real-time interaction behavior, historical interaction behavior also reflects the interests of the target object, and therefore historical interaction behavior data can also be combined to maintain the end features of each media content. Specifically, for each media content, a set of associated data corresponding to each target media content can be determined from the historical interaction behavior data, and the set of associated data of each target media content includes historical interaction behavior data corresponding to media content with the same attribute. The same attribute may, for example, include the same author of the media content, the same type (here taking the first-level type of a video as an example, such as game videos), and the like, and of course, other same attributes are also included, and the embodiments of this application do not limit this.

[0165] Further, based on the obtained each set of associated data, a second set of features corresponding to each target media content, which can also be referred to as generalized features, is extracted, so that based on the obtained each first set of features and second set of features, the end features of the corresponding target media content are obtained.

[0166] The generalized features are obtained by associating the behavior generated in the historical consumption process of the target object with the current video. Since in the direct features, in addition to the playback completion degree, which is necessary for each video, the other behaviors are relatively sparse (according to statistical conclusions, the proportion is about 10%), which means that most of the direct features will be 0, and the amount of information that can be provided is less. Therefore, it can be considered to generalize the direct features to a certain extent to increase the coverage of the features. Considering that in the short video scene, CP and CATE1 are two important attributes for short videos, these two attributes are selected for generalization, and of course, other attributes can also be selected. Here, CATE1 is taken as an example to explain the specific generalization process, as shown in FIG. 6, in the upper half of the dashed line of the interaction graph, the target object has consumed a game video in the past and has generated a like behavior, and when the target object consumes another game video, the generalized feature corresponding to the video will be set to 1. Similarly, for attention or comment behaviors, similar generalization operations are also performed, and the generalization method of CP is similar to that of CATE1. Figure 6

[0167] In the embodiments of this application, an end feature platform can be developed to provide a series of feature extraction interfaces, and not to be responsible for the specific implementation of the feature logic. In actual application, the function of feature extraction can be quickly realized through the script issuing mode according to the own needs, so as to realize the improvement of development and maintenance efficiency.

[0168] Referring to Table 2 shown below, the feature list included in the generalized features provided by the embodiments of this application can include the following features:

[0169] ​

[0170]

[0171] Table 2

[0172] Referring to Figure 6 As shown in Table 2, similar to Table 1, the feature IDs in Table 2 also correspond to the IDs identified in Figure 6 For example, in Table 2, feed represents a video, CP represents an author, cate1 represents a first-level type of the video, result page represents a search result main page, and play page represents a video play page. Referring to Figure 6 As shown in Table 2, for a video, feature ID 7 represents whether the author of the video is clicked in the search result page. If the author is clicked, the corresponding feature value is 1, otherwise the feature value is 0. Feature ID 8 represents whether a same-type video of the video is clicked in the search result page. If the same-type video is clicked, the corresponding feature value is 1, otherwise the feature value is 0. Other features are similar.

[0173] As can be seen, Table 1 and Table 2 above have a total of 21 dimensions of terminal features, which include direct features generated by direct interaction behaviors of the current video and generalized features obtained by generalizing interaction behaviors of other videos to the current video through certain dimensions (such as first / second-level classification, video author, etc.). Through the design and processing of the generalized features, the coverage rate of the terminal features is effectively improved to about 40%, which greatly improves the use efficiency of the features.

[0174] After having the data, it can be determined whether to trigger reordering in the process of a search session. In an embodiment of the present application, the determination of whether to trigger reordering can be performed by a reordering triggering module as shown in Figure 2 The reordering triggering module determines whether to trigger reordering once at the current time based on real-time interaction behaviors of the target object in the current search session.

[0175] In a possible implementation, the determination of whether to trigger reordering can be based on whether the real-time interaction behaviors meet a reordering triggering condition. When the reordering triggering condition is met, reordering is triggered once.

[0176] In practical applications, the design of the reordering trigger condition needs to consider two aspects. On the one hand, it is necessary to adjust the sequence of subsequent content in a timely manner when the target object exhibits obvious positive or negative interaction behavior. In this way, for positive interaction behavior, the consumption time can be prolonged to help the target object find the content that the target object wants to watch, and for negative interaction behavior, the potential interest of the target object can be explored in a timely manner to avoid the target object quickly exiting the current search session. On the other hand, the triggering frequency caused by the reordering trigger condition also needs to be considered. Because triggering reordering, whether the cloud performs the reordering process or the client performs the reordering process, will have certain computational consumption, especially for end reordering, the computational ability of different models differs greatly, and too frequent reordering can affect user experience. Of course, if the trigger condition is too strict, it will lead to insufficient coverage of reordering and difficulty in achieving business goals.

[0177] Based on the above considerations, the embodiments of the present application design the following trigger conditions, that is, when the real-time interaction behavior meets the reordering trigger condition, the search result page is updated, including:

[0178] (1) Triggering the display of a sub-result page corresponding to the target media content in the search result page. That is, when the target object operates on a specific target media content in the search result page and enters the corresponding sub-result page to view the details of the target media content, it can indicate that the target object currently has great interest in the target media content, which reflects the current interest of the target object. Therefore, the order of media content that is not closely related to the target media content should be moved backward, and the order of media content that is closely related to the target media content should be moved forward.

[0179] (2) Triggering the display of media content in the search result page. Such behavior can include triggering preloading page turning and other behaviors in the search result page. When the target object performs these behaviors, it usually indicates that the target object is not satisfied with the currently displayed media content, and the order of the media content should be adjusted in a timely manner.

[0180] (3) Positive interaction in the sub-result page. Positive interaction: behavior indicating that the target object is interested in the corresponding media content. Such behavior can include attention, forwarding, sharing, commenting, liking, and high complete playback behavior in the sub-result page. High complete playback behavior means that the browsing degree of the media content exceeds a threshold value, and the threshold value can be set by the user, for example, for a video, the threshold value can be set according to the duration of the video.

[0181] (4) Negative interaction on the sub-result page. Negative interaction refers to a behavior of the target object indicating that the target object is not interested in the corresponding media content. Such behaviors can include triggering preloading of a flip page and continuous fast sliding behavior, which both indicate that the target object is not interested in the currently displayed media content. Continuous fast sliding behavior refers to a situation where the target object stays on a continuous M sub-result pages for less than N seconds. In an embodiment, M can be set to 2 and N can be set to 3.

[0182] (5) Returning to the search result page from the sub-result page. Similarly, such a behavior also indicates that the target object is not interested in the currently displayed media content.

[0183] In the above trigger conditions, various behaviors can strongly guarantee the current interest tendency of the target object. For example, the above-mentioned click-to-play, positive interaction, or high complete play behavior are all positive feedback behaviors of the user, and continuous fast sliding and returning are negative feedback behaviors. Different trigger types can flexibly describe the intention and state changes of the target in the search session process, thereby providing beneficial signals and data basis for subsequent reordering work.

[0184] In another possible embodiment, considering that the above trigger conditions mainly rely on artificial rules to design trigger logic, in actual application, a model can be used to replace the trigger logic, that is, a reordering trigger prediction model can be trained, and factors such as current consumption context, historical consumption, current network environment, and mobile phone load can be combined to comprehensively decide whether to trigger reordering, so as to realize higher-level intelligent decision-making.

[0185] Specifically, reordering trigger prediction data is obtained, the reordering trigger prediction data is input into a trained reordering trigger prediction model, a prediction result indicating whether to trigger reordering is obtained, if the prediction result indicates to trigger reordering, reordering is performed, and a search result page is updated.

[0186] The reordering trigger prediction data can include one or a combination of the following data:

[0187] (1) Real-time interaction behavior data of real-time interaction behavior

[0188] (2) Historical interaction behavior data of the target object

[0189] (3) Current load condition of the terminal device

[0190] (4) Current network condition of the terminal device

[0191] Of course, other possible data can also be included, and the embodiments of the present application do not limit this.

[0192] In the embodiments of the present application, the process of reordering can be performed by the client itself, or can be performed by the cloud (server) with rich computing resources to improve the efficiency of ordering.

[0193] When performed by the cloud, referring to Figure 7 , a flowchart of end-cloud collaborative execution provided by the embodiments of the present application is shown.

[0194] Step 701: The client collects real-time interaction behaviors of the target object in the search session.

[0195] Step 702: The client determines that the reordering trigger condition is met based on the real-time interaction behaviors.

[0196] Step 703: The client initiates a reordering request to the server.

[0197] The reordering request can carry the end features corresponding to the target media content, the end features being extracted based on the real-time interaction behavior data of at least one real-time interaction behavior and the historical interaction behavior data of the target object, i.e., the direct features and generalized features extracted above.

[0198] Step 704: The server performs a reordering process to obtain a second content sequence. Specifically, the server reorders the videos in the search results that have not been consumed by the target object based on the context and historical interests of the target object consumption to obtain a second content sequence, and returns it to the client. Specifically, the second content sequence can be obtained by the server based on the end features using a trained reordering model.

[0199] Step 705: The server returns the second content sequence to the client.

[0200] Step 706: The client updates the search result page based on the second content sequence.

[0201] In the embodiments of the present application, the process of reordering can be performed by Figure 2 , the reordering module can perform the reordering process based on a trained reordering model, referring to Figure 8 , a reordering flowchart provided by the embodiments of the present application is shown, which includes the following steps:

[0202] Step 801: Based on the target media content corresponding to different interaction types in at least one real-time interaction behavior, at least one interaction content sequence is constructed, and each interaction content sequence corresponds to one interaction type.

[0203] In the embodiments of the present application, the interaction type can include exposure, positive interaction, negative interaction, etc., and the following content sequences can be constructed:

[0204] (1) Exposure content sequence, which is constructed based on the target media content displayed in the search result interface, for example, can refer to the play exposure sequence generated by the target object in a search session.

[0205] The consumption information in the current search session reflects the intention and interest of the target object in this search session, which is crucial to the sorting effect. The exposure content sequence will increase an element every time the target object consumes a media content (such as display on the result page or the target object clicks to play, etc.).

[0206] In a possible implementation, the exposure content sequence maintains a window with a maximum length of M (for example, 10) to record the latest M media contents. Each media content in the sequence can be composed of several features: category features, posterior statistical features, fine ranking advanced features, and terminal features. The features of each media content are combined into a feature vector with a total feature dimension of h. The features of the entire sequence are combined into an M*h matrix as the input of the re-ranking model.

[0207] (2) Positive feedback sequence, which is constructed based on at least one real-time interaction behavior and the target media content corresponding to the positive interaction behavior in the historical interaction behavior, refers to the sequence of media contents with positive feedback collected from the consumption history of the target object.

[0208] From the perspective of depicting the interest of the target object, it can be divided into long-term interest and short-term interest. The exposure content sequence described above is a depiction of the short-term interest of the target object because the exposure content sequence is generated in the current search session. The long-term interest needs to continue to trace the historical consumption of the target object forward. For some high-consumption target objects, the number of historical consumption is often very large, and it is difficult to achieve the expected effect by inputting all the data into the model, but it may introduce more noise. Therefore, it can be focused on the consumption history of the target object with explicit or implicit positive feedback behavior. The explicit positive feedback includes, for example, the explicit interaction behaviors of the target object to the media content, such as clicking, commenting, following the author, and liking. The implicit positive feedback includes, for example, the high complete play behavior of the target object. By screening these two types of behaviors, the long-term interest of the target object in the positive feedback dimension can be obtained.

[0209] In a possible implementation, the positive feedback sequence also maintains a window with a maximum length of M. Each media content in the sequence is composed of several features: category features, posterior statistical features, fine ranking advanced features, and terminal features. The features of each media content are combined into a feature vector with a total feature dimension of h. The features of the entire sequence are combined into an M*h matrix as the input of the re-ranking model.

[0210] (3) The negative feedback sequence is constructed based on at least one real-time interactive behavior and the target media content corresponding to negative interactive behaviors in historical interactive behaviors.

[0211] Similar to the design principles of positive feedback sequences, negative feedback is also an important component of a target's long-term interests. Negative feedback is divided into explicit and implicit negative feedback. Explicit negative feedback behaviors include unliking and unfollowing, while implicit negative feedback behaviors can include swiping.

[0212] In one possible implementation, the negative feedback sequence maintains a window of maximum length M. Each media content in the sequence consists of several features: category features, posterior statistical features, high-level ranking features, and end features. These features of each media content are combined into a feature vector with a total feature dimension of h. The features of the entire sequence are then combined into an M*h matrix, which is used as the input to the re-ranking model.

[0213] Step 802: Construct a sequence of content to be sorted based on each media content to be sorted.

[0214] In this embodiment, the sequence of content to be sorted is a video sequence that has not yet been consumed in the current search session. For example, it could be media content that has never been displayed or media content that the target object has not interacted with. The goal of reordering is to re-sort this part of the video, hoping to increase the playback duration of the entire sequence.

[0215] In one possible implementation, the sequence of content to be sorted maintains a window of length 2M. Each media in the sequence consists of several features: category features, posterior statistical features, and high-level ranking features. When the target object continues to consume content and the length of the sequence to be sorted is less than 20, a pre-fetching process can be triggered. The client requests a page from the backend server to retrieve the media content and corresponding features of the new page, which are then appended to the end of the sequence to be sorted as supplementary data. The feature vectors of the sequence to be sorted have a total feature dimension of h2, and the entire sequence is constructed into a 2M*h2 matrix, which serves as the input to the re-ranking model.

[0216] Step 803: Using a re-ranking model, based on search keywords, the trigger condition type corresponding to at least one real-time interactive behavior, and at least one interactive content sequence, determine the target object's level of interest in each unranked media content in the unranked content sequence.

[0217] Step 804: Based on the obtained levels of interest, obtain the second content sequence.

[0218] See Figure 9 and Figure 10 As shown, Figure 9A model structure diagram of a reordering model is provided for an embodiment of the present application. The reordering model includes an input layer, an encoding layer, an interaction layer, a task layer, and an output layer. Figure 10 A processing flow diagram of the reordering model is provided for an embodiment of the present application. The layers of the reordering model are introduced below in combination with the processing flow shown in the figure. Figure 10 The input data of the input layer of the reordering model includes the following:

[0219] The input data of the input layer of the reordering model includes the following:

[0220] (1) Query Embedding, a Query text vector corresponding to a search keyword of a current search session, representing the semantic information of the Query. In the search scenario, the search Query is the main embodiment of the user's intention, and it is necessary to be input. Specifically, a Bert (a text encoding model) or other encoders can be selected as a text encoder to encode the Query text into a fixed-length Query Embedding as the input of the reordering model.

[0221] (2) Trigger type, a corresponding trigger type in the above reordering trigger module, different trigger types will affect the sorting strategy. Since different trigger types represent different behavior accumulation of the target object, some belong to positive behavior and some belong to negative behavior, and the expectation of subsequent reordering may be different. For example, fast sliding triggers, the target object is more likely to stop fast sliding or even exit the behavior; for positive interaction, it is more desirable to recommend some videos similar to the interacted video to meet the interest of the target object.

[0222] In an implementation, considering that the trigger type is a discrete distribution, an embedding lookup method is adopted to enter the model.

[0223] (3) Exposure sequence

[0224] (4) Positive feedback sequence

[0225] (5) Negative feedback sequence

[0226] (6) To-be-sorted content sequence

[0227] Step 1001: Encode the search keyword, trigger condition type, at least one interactive content sequence, and to-be-sorted content sequence respectively to obtain corresponding encoding feature vectors.

[0228] In the embodiment of the present application, the process of step 1001 can be implemented by the encoding layer of the reordering model. The encoding layer is used for embedding and sequence encoding of the input data to extract the feature information related to reordering in the input data. Based on the type of input data, multiple encoding methods can be included for encoding different types of input data.

[0229] Referring to Table 3, the encoding methods for each input data are shown as follows:

[0230]

[0231]

[0232] Table 3

[0233] The trigger type is a discrete value. The embedding lookup method can be used to convert the discrete value into a vector for subsequent neural network processing. Of course, other methods of converting discrete values into vectors can also be used.

[0234] For each content sequence, the feature encoding can be performed by the steps shown in Table 4. Figure 11 Figure 11 Table 4 is a flowchart for feature encoding of the content sequence. Since the encoding processes of the above-mentioned various content sequences are similar, hereinafter, one of the content sequences A will be taken as an example for introduction. The content sequence A can be any content sequence.

[0235] Step 10011: For each media content in the content sequence A, the corresponding basic feature vector is obtained.

[0236] The basic feature vector is the feature of each target media content input by the input layer, which can be composed of the following features:

[0237] (1) The category feature vector is obtained based on the content category to which the target media content belongs;

[0238] (2) The end feature vector is obtained by feature extraction based on the end feature corresponding to the target media content;

[0239] (3) The fine arrangement feature vector is obtained based on the fine arrangement feature vector of the target media content in the fine arrangement stage.

[0240] Of course, other possible feature vectors can also be included. The category feature vector, end feature vector and fine arrangement feature vector are spliced to obtain the basic feature vector of the target media content. The process of obtaining the basic feature vector for the media content to be sorted is similar. ​

[0241] Step 10012: Based on the position of each media content in any content sequence, obtain the position feature vector corresponding to each media content.

[0242] Considering that most serialization extraction methods cannot take into account the positional information of each media content in the sequence, in order to solve the defect that serialization extraction methods cannot distinguish positions, a positional feature vector can be added to the media content at each position. The positional feature vector can usually be fixed, corresponding to the position, and can express the current position in the sequence, so no training is required.

[0243] Step 10013: Overlay the content feature vectors and position feature vectors corresponding to each media content to obtain the content feature vectors corresponding to each media content.

[0244] See Figure 12 As shown, the content of a media can be superimposed with the location feature vector corresponding to its location to obtain the corresponding content feature vector.

[0245] Specifically, the superposition can be performed by summation or pooling, and this application embodiment does not limit this.

[0246] Step 10014: Using a serialization feature extraction method, extract features from each obtained content feature vector to obtain the encoded feature vector corresponding to any content sequence. See [link to relevant documentation]. Figure 12 As shown, the obtained content feature vector is processed by serialization feature extraction to obtain the final encoded feature vector.

[0247] In this embodiment, to integrate sequence information into each media content embedding, a serialization encoding method is required. Commonly used serialization encoding methods include recurrent neural networks (RNNs), long short-term memory networks (LSTMs), and Transformers. It should be noted that different sequences can use the same or different serialization encoding methods.

[0248] This example assumes all content sequences use the same serialization encoding method. Considering that the Transformer is more direct in modeling item-item relationships, avoids the problem of weak long-distance dependencies, and has better parallelism, it is chosen as the encoder. However, in the reordering scenario, the backend service faces a large number of requests, making model performance crucial. Based on this consideration, a single-layer Transformer was ultimately chosen as the sequence encoder. Of course, in practical applications, a multi-layer encoder structure can also be used.

[0249] See Figure 13As shown, it is a structural diagram of the Transformer network, the Transformer network is composed of multiple encoders and decoders, the multiple encoders are in a cascaded structure, and the embodiment of the application can only use a one-layer encoder structure, that is Figure 13 The encoder structure shown in the dashed box, when the content feature vector after superposition is input to the encoder, then after processing by a multi-attention layer, after processing by a residual & normalization layer, the final encoding feature vector is obtained by inputting to a feed forward layer and processing by a residual & normalization layer.

[0250] Step 1002: Based on the obtained encoding feature vectors, the encoding feature vectors of the to-be-ordered content sequence are respectively processed by an attention mechanism-based interaction to obtain multiple attention feature vectors.

[0251] In the embodiment of the application, the process of step 1002 can be realized by an interaction layer of the reordering model, and the function of the interaction layer is to make the media content in the to-be-ordered content sequence "focus" on the short-term and long-term interest points through the element interaction between sequences, because the historical behavior of the target object will greatly affect the current behavior logic.

[0252] Specifically, the to-be-ordered content sequence can be processed by attention interaction with the exposure content sequence, the positive feedback sequence and the negative feedback sequence respectively.

[0253] In one possible implementation, taking one of the interactive content sequences A as an example, which can be any of the at least one interactive content sequence described above, the attention interaction between the interactive content sequence A and the to-be-ordered content sequence can include the following steps:

[0254] Based on the similarity between the encoding feature vector of the interactive content sequence A and the encoding feature vector of the to-be-ordered content sequence, the corresponding attention weight is determined, and then based on the obtained attention weight and the encoding feature vector of the interactive content sequence A, the corresponding attention feature vector is obtained. The encoding feature vector of the sequence is obtained by splicing the encoding feature vectors of the media contents included therein.

[0255] In a possible implementation, the attention interaction process can be implemented in the manner of Transformer attention. Taking the attention interaction process of the to-be-ordered content sequence and the exposure content sequence as an example, each element of the to-be-ordered content sequence is taken as a query, and the elements in the exposure content sequence are taken as keys and values, which are input into a layer of Transformer network. The dimension of the finally obtained output vector is unchanged, but the output vector already contains the attention information of the exposure content sequence. Similarly, the to-be-ordered content sequence respectively performs attention calculation with the positive feedback sequence and the negative feedback sequence to obtain the corresponding output vectors.

[0256] Step 1003: The plurality of attention feature vectors are concatenated with the encoding feature vector of the search keyword and the encoding feature vector of the trigger condition type to obtain an output feature vector.

[0257] Through the above process, three attention feature vectors with historical information are obtained, and the three attention feature vectors are concatenated with the Query Embedding vector and the encoding feature vector corresponding to the trigger type as the output of the interaction layer.

[0258] Step 1004: The output feature vector is respectively mapped into a plurality of index vectors; a value in each index vector represents the interest degree of the target object in the corresponding to-be-ordered media content in the corresponding index parameter dimension.

[0259] In the embodiment of the application, the process of step 1004 can be implemented by a task layer of a reordering model. The task layer is configured to predict the values of a plurality of indexes based on the output feature vector.

[0260] Taking a video as an example, it is considered that the main goal of video reordering is to improve the search session average play duration, and the goal cannot be directly predicted. Therefore, a plurality of model optimization goals related to the goal can be designed to complete the optimization of the goal.

[0261] Specifically, three prediction tasks are designed for the media content in the to-be-ordered content sequence, and each prediction task predicts the value of an index:

[0262] (1) A task of predicting whether to complete playing, which is a binary classification task. Here, the complete playing specifically refers to that the playing degree exceeds a specific threshold, and the threshold is related to the video duration.

[0263] (2) A task of predicting whether to fast slide, which is also a binary classification task, and is used to predict whether the playing degree of the video is less than 3 seconds.

[0264] (3) a task of predicting the playing time length, which is a multi-classification task. Since the playing time length of a video is a discrete numerical value, it is obviously more difficult to directly predict the discrete numerical value, and thus the time length can be processed in buckets, each bucket corresponding to a time length segment, and the playing time length corresponding to the video is predicted.

[0265] Among the three tasks, the task of predicting the playing time length focuses on optimizing the playing time length of the current video, and the tasks of predicting whether to finish playing and predicting whether to fast-forward focus on optimizing the length of the playing sequence. In order to simultaneously optimize the three targets, a multi-task learning method is used for learning.

[0266] In a possible implementation, the MMOE algorithm can be used to implement, that is, the output feature vector output by the above interaction layer is used as the input of MMOE, and the MMOE model outputs the prediction probability values of the above three targets. Of course, the MMOE algorithm can also use other possible multi-task deep learning models to implement, such as the ESSM model, the PLE model, and the DUPN model, etc., and the embodiments of the present application do not limit this.

[0267] Step 1005: determining the degree of interest of each based on the plurality of index parameter values.

[0268] The output of the task layer is a score vector corresponding to the content sequence to be sorted, which includes the score value of each media content in the content sequence to be sorted corresponding to each task, that is, the index parameter value of each index, and then the final score of the target object for each media content to be sorted, that is, the degree of interest, can be determined based on the obtained plurality of index parameter values.

[0269] It should be noted that the reordering model corresponds to a training phase and a prediction phase, and the processes performed by each layer before the output layer in the two phases are the same, the difference lies in that when in the training phase, after obtaining the plurality of index parameter values, the model loss of the entire model can be calculated based on the labeled supervision values of the index parameter values, so as to optimize and update the training parameters in the model. For example, a cross-entropy loss function or the like can be used as the loss function of the reordering model, and the output layer calculates the cross-entropy loss of the three tasks respectively, and the cumulative loss is used as the overall loss to optimize the model.

[0270] When in the prediction phase, the final score of the target object for each media content to be sorted can be represented as:

[0271] final_score = playtime * (finish - quick)

[0272] Wherein, the finish represents the probability of predicting the complete play, the quick represents the probability of predicting the fast sliding, the playtime represents the barrel number of predicting the maximum play probability, and then, the final_score is used to sort the media content to obtain the second content sequence.

[0273] To sum up, in the embodiment of the application, according to the real-time interaction between the target object and the media content, an intelligent triggering scheme is designed to automatically select a suitable opportunity to trigger reordering and send a request to the cloud server. After calculation, the server returns the reordered search result list to the client and presents it to the target object. This scheme solves the problem of not timely feedback of the interaction behavior of the target object in the media content search scene, and timely adjusts the search result order, which can effectively improve the search average time length.

[0274] The technical scheme of the embodiment of the application provides a complete end-cloud integrated reordering scheme, including feature engineering, triggering conditions and reordering model design, which has universality and portability in video, audio and text search scenes. Through end-cloud cooperation, the search result is adjusted in real time according to the feedback of the target object, realizing the personalized effect of thousands of people with thousands of models. Moreover, through internal experiments, the above-mentioned effects are verified on the business indicators: in the experiment, the control group does not perform cloud rearrangement, and the experimental group performs cloud rearrangement. The experimental results are shown in Table 4 as follows:

[0275] Control group Experimental group Duration improvement Search average play duration 203.214848 seconds 206.843972 seconds 2.3981％

[0276] Table 4

[0277] It can be seen that the experimental group has achieved a relative improvement of 2.398% in the main indicator of search average play time, achieving a large benefit compared with the control group.

[0278] In addition, in the design of the end feature, the coverage of the end feature is effectively improved through the design of the direct feature and the generalization feature. The rearrangement model generates a short-term interest sequence and a long-term interest sequence through historical consumption behavior, and interacts with the content sequence to be sorted through the Transformer attention mechanism, so that different media content can pay attention to different interests, improving the personalized effect.

[0279] Please refer to Figure 14 Based on the same inventive concept, the embodiment of the application also provides a search result reordering device 140, which comprises:

[0280] A page display unit 1401 is configured to display a search result page corresponding to a search session, and the search result page displays a first content sequence obtained based on a search keyword corresponding to the search session;

[0281] The acquisition unit 1402 is configured to acquire at least one real-time interaction behavior of the target object on the search result page.

[0282] The page updating unit 1403 is configured to update the search result page when the at least one real-time interaction behavior meets a reordering triggering condition, and the updated search result page displays a second content sequence, the second content sequence being obtained by reordering each media content to be sorted included in the first content sequence, the media content to be sorted being other media content than the target media content involved in the at least one real-time interaction behavior in the first content sequence.

[0283] Optionally, the page updating unit 1403 is specifically configured to:

[0284] If the at least one real-time interaction behavior includes at least one of the following interaction behaviors, the search result page is updated.

[0285] Triggering display of a sub-result page corresponding to the target media content in the search result page;

[0286] Triggering switching of displayed media content in the search result page;

[0287] Performing positive interaction in the sub-result page, the positive interaction representing a behavior of the target object being interested in the corresponding media content;

[0288] Performing negative interaction in the sub-result page, the negative interaction representing a behavior of the target object being uninterested in the corresponding media content;

[0289] Returning from the sub-result page to the search result page.

[0290] Optionally, the page updating unit 1403 is specifically configured to:

[0291] Acquiring reordering triggering prediction data, the reordering triggering prediction data including at least one of real-time interaction behavior data of the at least one real-time interaction behavior, historical interaction behavior data of the target object, a current load condition of the terminal device, and a current network condition of the terminal device;

[0292] Inputting the reordering triggering prediction data into a reordering triggering prediction model to obtain a prediction result indicating whether to trigger reordering;

[0293] If the prediction result indicates triggering reordering, the search result page is updated.

[0294] Optionally, the page updating unit 1403 is specifically configured to:

[0295] When the at least one real-time interaction behavior meets the reordering trigger condition, a reordering request is initiated to a server, the reordering request carrying an end feature corresponding to the target media content, the end feature being extracted based on real-time interaction behavior data of the at least one real-time interaction behavior and historical interaction behavior data of the target object, and the target object being a user of the terminal device;

[0296] The second content sequence returned by the server in response to the reordering request is received, the second content sequence being obtained by the server based on the end feature and using a reordering model;

[0297] Based on the second content sequence, the search result page is updated.

[0298] Optionally, the apparatus further includes a reordering unit 1404 configured to:

[0299] Based on the target media content corresponding to different interaction types in the at least one real-time interaction behavior, at least one interaction content sequence is constructed, each interaction content sequence corresponding to one interaction type;

[0300] Based on each media content to be sorted, a content sequence to be sorted is constructed;

[0301] Using the reordering model, based on the search keyword, the trigger condition type corresponding to the at least one real-time interaction behavior, and the at least one interaction content sequence, a degree of interest of the target object in each media content to be sorted in the content sequence to be sorted is determined;

[0302] The page updating unit 1403 is specifically configured to obtain the second content sequence based on the obtained degrees of interest, and update the search result page based on the second content sequence.

[0303] Optionally, the reordering unit 1404 is specifically configured to:

[0304] The search keyword, the trigger condition type, the at least one interaction content sequence, and the content sequence to be sorted are respectively encoded to obtain corresponding encoded feature vectors;

[0305] Based on the obtained encoded feature vectors, the encoded feature vectors of the content sequence to be sorted are respectively processed based on an attention mechanism to obtain a plurality of attention feature vectors;

[0306] The plurality of attention feature vectors are concatenated with the encoded feature vectors of the search keyword and the encoded feature vectors of the trigger condition type to obtain an output feature vector;

[0307] map the output feature vectors into a plurality of index vectors respectively, and determine the respective degrees of interest based on the plurality of index vectors; each index vector corresponds to an index parameter dimension, and a value in each index vector represents a degree of interest of the target object in corresponding to-be-sequenced media content in the corresponding index parameter dimension.

[0308] Optionally, the reordering unit 1404 is specifically configured to:

[0309] Any content sequence in the encoding of the at least one interactive content sequence and the to-be-sequenced content sequence is encoded by the following process:

[0310] Respectively for each media content in any content sequence, obtain a corresponding basic feature vector;

[0311] Respectively based on the position of each media content in any content sequence, obtain a position feature vector corresponding to each media content;

[0312] Based on the content feature vector and the position feature vector corresponding to each media content, perform superposition processing to obtain a content feature vector corresponding to each media content;

[0313] Using a serialization feature extraction method, perform feature extraction on the obtained content feature vectors to obtain an encoding feature vector corresponding to any content sequence.

[0314] Optionally, the apparatus further includes an end feature extraction unit 1405 configured to:

[0315] Based on real-time interactive behavior data of at least one real-time interactive behavior and historical interactive behavior data of the target object, extract an end feature corresponding to each target media content, the end feature being used to represent an interactive tendency of the target object for the search session;

[0316] Then the reordering unit 1404 is specifically configured to:

[0317] For any target media content, obtain a corresponding basic feature vector by the following steps:

[0318] Based on a content category to which any target media content belongs, obtain a category feature vector;

[0319] Based on the end feature corresponding to any target media content, perform feature extraction to obtain an end feature vector;

[0320] Obtain a fine arrangement feature vector of any target media content at the fine arrangement stage;

[0321] Perform splicing processing on the category feature vector, the end feature vector, and the fine arrangement feature vector to obtain a corresponding basic feature vector.

[0322] Optionally, the end feature extraction unit 1405 is specifically configured to:

[0323] extract a first feature set of each target media content based on real-time interaction behavior data;

[0324] determine a respective associated data set of each target media content from historical interaction behavior data, the associated data set of each target media content containing historical interaction data corresponding to media content having the same attribute as the target media content;

[0325] extract a second feature set corresponding to each target media content based on the obtained associated data set of each target media content;

[0326] obtain the end feature of the corresponding target media content based on the obtained first feature set and second feature set of each target media content.

[0327] Optionally, the reordering unit 1404 is specifically configured to:

[0328] for at least one interactive content sequence, the following steps are performed respectively:

[0329] for one interactive content sequence, determine the corresponding attention weight based on the similarity between the encoding feature vector of one interactive content sequence and the encoding feature vector of the content sequence to be sorted;

[0330] obtain the corresponding attention feature vector based on the obtained attention weight and the encoding feature vector of one interactive content sequence.

[0331] Optionally, the reordering unit 1404 is specifically configured to:

[0332] construct an exposure content sequence based on the target media content displayed in the search result interface;

[0333] construct a positive feedback sequence based on the target media content corresponding to the positive interaction behavior in at least one real-time interaction behavior and historical interaction behavior;

[0334] construct a negative feedback sequence based on the target media content corresponding to the negative interaction behavior in at least one real-time interaction behavior and historical interaction behavior.

[0335] Through the above device, the order of each media content in the search result is adjusted according to the instant interest, which can improve the accuracy of the current search session result presentation, timely meet the search intention of the user, improve the content search efficiency, and also improve the timeliness of the feedback of the target object interaction behavior.

[0336] The apparatus can be used to perform the method shown in the embodiments of the present application, therefore, the functions that can be achieved by each functional module of the apparatus can refer to the description of the foregoing embodiments, and will not be described in detail.

[0337] Please refer to Figure 15 , based on the same technical concept, the embodiments of the present application also provide a computer device. In an embodiment, the computer device can be a server as shown in Figure 1 or a device corresponding to the cloud as shown in Figure 2 , the computer device includes a memory 1501, a communication module 1503 and one or more processors 1502, as shown in Figure 15 .

[0338] The memory 1501 is used to store computer programs executed by the processor 1502. The memory 1501 can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system and programs required for running instant messaging functions, etc.; the data storage area can store various instant messaging information and operation instruction sets, etc.

[0339] The memory 1501 can be a volatile memory such as a random-access memory (RAM); the memory 1501 can also be a non-volatile memory such as a read-only memory, a flash memory, a hard disk drive (HDD) or a solid-state drive (SSD); or the memory 1501 can be any other medium capable of carrying or storing desired program codes in the form of instructions or data structures and capable of being accessed by a computer, but not limited to this. The memory 1501 can be a combination of the above-mentioned memories.

[0340] The processor 1502 can include one or more central processing units (CPUs) or digital processing units, etc. The processor 1502 is used to implement the above-mentioned search result reordering method when invoking the computer programs stored in the memory 1501.

[0341] The communication module 1503 is used to communicate with terminal devices and other servers.

[0342] The specific connection medium between the above-mentioned memory 1501, communication module 1503 and processor 1502 is not limited in the embodiments of the present application. In the embodiments of the present application, the memory 1501 and the processor 1502 are connected through a bus 1504, and the bus 1504 is connected to the memory 1501 and the processor 1502 through a bus interface 1505. Figure 15 ​Figure 15 The diagram uses thick lines to describe the connections between other components; these are for illustrative purposes only and should not be considered limiting. The 1504 bus can be divided into address bus, data bus, control bus, etc. For ease of description, Figure 15 It is described using only a thick line, but does not indicate that there is only one bus or one type of bus.

[0343] The memory 1501 stores a computer storage medium, which stores computer-executable instructions. The computer-executable instructions are used to implement the search result reordering method of the embodiments of this application, and the processor 1502 is used to execute the search result reordering methods of the above embodiments.

[0344] In another embodiment, the computer device can also be a terminal device, such as... Figure 1 The terminal device shown. In this embodiment, the structure of the computer device can be as follows. Figure 16 As shown, it includes components such as: communication component 1610, memory 1620, display unit 1630, camera 1640, sensor 1650, audio circuit 1660, Bluetooth module 1670, processor 1680, etc.

[0345] The communication component 1610 is used to communicate with the server. In some embodiments, it may include a Circuit-Based Wireless Fidelity (WiFi) module, which is a short-range wireless transmission technology. Computer devices can use WiFi modules to help users send and receive information.

[0346] The memory 1620 can be used to store software programs and data. The processor 1680 executes various functions of the terminal device and data processing by running the software programs or data stored in the memory 1620. The memory 1620 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 1620 stores an operating system that enables the terminal device to run. In this application, the memory 1620 may store the operating system and various applications, and may also store code that executes the search result reordering method of the embodiments of this application.

[0347] The display unit 1630 can also be used to display information input by a user or information provided to a user, as well as a graphical user interface (GUI) of various menus of the terminal device. Specifically, the display unit 1630 can include a display screen 1632 disposed on the front of the terminal device. Among them, the display screen 1632 can be configured in the form of a liquid crystal display, a light-emitting diode, etc. The display unit 1630 can be used to display various search result pages or sub-result pages in the embodiments of the present application.

[0348] The display unit 1630 can also be used to receive inputted digital or character information, and generate signal input related to user settings and function control of the terminal device. Specifically, the display unit 1630 can include a touch screen 1631 disposed on the front of the terminal device, which can collect touch operations of a user thereon or therearound, such as clicking buttons, dragging scroll boxes, etc.

[0349] Among them, the touch screen 1631 can be overlaid on the display screen 1632, or the touch screen 1631 can be integrated with the display screen 1632 to realize the input and output functions of the terminal device, and after integration, it can be simply referred to as a touch display screen. The display unit 1630 in the present application can display application programs and corresponding operation steps.

[0350] The camera 1640 can be used to capture still images, and a user can publish comments through the images taken by the camera 1640. The camera 1640 can be one or multiple. An object generates an optical image through a lens and projects it to a photosensitive element. The photosensitive element can be a charge coupled device (CCD) or a complementary metal-oxide-semiconductor (CMOS) phototransistor. The photosensitive element converts the optical signal into an electrical signal, and then transmits the electrical signal to the processor 1680 to convert it into a digital image signal.

[0351] The terminal device can also include at least one sensor 1650, such as an acceleration sensor 1651, a distance sensor 1652, a fingerprint sensor 1653, a temperature sensor 1654. The terminal device can also be configured with a gyroscope, a barometer, a hygrometer, a thermometer, an infrared sensor, a light sensor, a motion sensor, and other sensors.

[0352] The audio circuit 1660, the speaker 1661, and the microphone 1662 can provide an audio interface between the user and the terminal device. The audio circuit 1660 can convert the received audio data into an electrical signal, transmit the electrical signal to the speaker 1661, and convert the electrical signal into a sound signal output by the speaker 1661. The terminal device can also be configured with a volume button for adjusting the volume of the sound signal. On the other hand, the microphone 1662 converts the collected sound signal into an electrical signal, which is received by the audio circuit 1660 and converted into audio data, which is output to the communication component 1610 for transmission to, for example, another terminal device, or to the memory 1620 for further processing.

[0353] The Bluetooth module 1670 is used to interact with other Bluetooth devices having a Bluetooth module through a Bluetooth protocol. For example, the terminal device can establish a Bluetooth connection with a wearable computer device (e.g., a smart watch) also having a Bluetooth module through the Bluetooth module 1670, and thus interact with data.

[0354] The processor 1680 is the control center of the terminal device, which connects all parts of the terminal device through various interfaces and lines, and performs various functions of the terminal device and processes data by running or executing software programs stored in the memory 1620 and calling data stored in the memory 1620. In some embodiments, the processor 1680 can include one or more processing units; the processor 1680 can also integrate an application processor and a baseband processor, where the application processor mainly processes the operating system, user interface, and application programs, and the baseband processor mainly processes wireless communication. It can be understood that the above-mentioned baseband processor can also not be integrated into the processor 1680. In this application, the processor 1680 can run the operating system, the application program, the user interface display and the touch response, and the search result reordering method of the embodiments of the application. In addition, the processor 1680 is coupled with the display unit 1630.

[0355] Based on the same inventive concept, the embodiments of the present application also provide a storage medium storing a computer program, which, when executed on a computer, causes the computer to perform the steps of the search result reordering method according to the various exemplary embodiments of the present application described above in the specification.

[0356] In some possible implementations, various aspects of the search result reordering method provided by the present application can also be implemented in the form of a computer program product, which includes a computer program for causing a computer device to perform the steps of the search result reordering method according to the various exemplary embodiments of the present application described above in the specification when the program product is executed on the computer device, for example, the computer device can perform the steps of the embodiments.

[0357] The program product of the embodiments of the present application can employ any combination of one or more computer-readable media. The computer-readable media can be a computer- readable storage medium or a computer-readable signal medium. The computer-readable storage medium can be, for example but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer-readable storage medium include an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0358] The program product of the embodiments of the present application can employ a portable compact disc read-only memory (CD-ROM) and include a computer program, and can be run on a computer device. However, the program product of the present application is not limited thereto, and in the present application, the computer-readable storage medium can be any tangible medium containing or storing a program, which includes a computer program that can be used by or in conjunction with a command execution system, apparatus, or device.

[0359] The computer-readable signal medium can include a data signal traveling in baseband or a carrier wave traveling in a packet stream, in which a computer program is embodied. Such a program can take any of a variety of forms, including but not limited to a source code, an object code, an interpreted code, or any suitable combination of the foregoing. The computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can be used to carry or transmit a program for use by or in connection with a command execution system, apparatus, or device.

[0360] The computer program embodied on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, or the like, or any suitable combination of the foregoing.

[0361] The computer program for performing the operations of the present application can be written in any combination of one or more programming languages, including an object-oriented programming language such as Java, C++, or the like, and a conventional procedural programming language such as the "C" language or the like.

[0362] It should be noted that although several units or sub-units of the apparatus are mentioned in the above detailed description, such a division is merely exemplary and not mandatory. In fact, according to the embodiments of the present 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 into multiple units.

[0363] Moreover, although the operations of the method(s) herein can be described in a particular, sequential order, this is not intended to be a requirement or a limitation. Rather, additional steps can be provided before, after, or in between the described steps, and the method(s) can be implemented without some or all of the described steps. Further, steps can be executed in an order other than that described.

[0364] Those skilled in the art will appreciate that embodiments of the present application can be devised for a method, a system, or a computer program product. Accordingly, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer readable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) embodying computer readable program code.

[0365] While the preferred embodiments of the application have been described, additional variations and modifications can be employed by those skilled in the art without departing from the spirit and scope of the application. Therefore, it should be understood that the appended claims are intended to cover all such modifications and variations as falling within the scope of the application. Accordingly, the application is intended to embrace all such alterations, modifications, and variations that fall within the scope of the appended claims. In addition, while a particular feature of the application can have been disclosed with respect to only one of several embodiments, such feature can be combined with one or more other features of the same or different embodiments as can be desired and advantageous for any given or

[0366] Obviously, numerous modifications and variations of the present application are possible in light of the above teachings. It is therefore to be understood that within the scope of the appended claims and their equivalents, the application can be practiced otherwise than as specifically described.

Claims

1. A method of reordering search results, characterized by, The method comprises: displaying a search result page corresponding to a search session, the search result page displaying a first content sequence obtained based on a search keyword corresponding to the search session; obtaining at least one real-time interaction behavior of a target object on the search result page; when the at least one real-time interaction behavior meets a reordering trigger condition, updating the search result page, the updated search result page displaying a second content sequence, the second content sequence being obtained by reordering each media content contained in the first content sequence, the media content being other media content than target media content involved in the at least one real-time interaction behavior in the first content sequence; wherein, when the at least one real-time interaction behavior meets the reordering trigger condition, updating the search result page comprises: obtaining reordering trigger prediction data, the reordering trigger prediction data comprising at least one of real-time interaction behavior data of the at least one real-time interaction behavior, historical interaction behavior data of the target object, a current load condition of a terminal device, and a current network condition of the terminal device; inputting the reordering trigger prediction data into a reordering trigger prediction model to obtain a prediction result indicating whether to trigger reordering; if the prediction result indicates to trigger reordering, updating the search result page.

2. The method of claim 1, wherein, The method further comprises: if the at least one real-time interaction behavior comprises at least one of the following interaction behaviors, updating the search result page: triggering display of a sub-result page corresponding to target media content in the search result page; triggering switching of displayed media content in the search result page; performing positive interaction in the sub-result page, the positive interaction representing a behavior of the target object being interested in corresponding media content; performing negative interaction in the sub-result page, the negative interaction representing a behavior of the target object not being interested in corresponding media content; returning to the search result page from the sub-result page.

3. The method of claim 1, wherein, The method further comprises: when the at least one real-time interaction behavior meets the reordering trigger condition, initiating a reordering request to a server, the reordering request carrying an end feature corresponding to the target media content, the end feature being extracted based on real-time interaction behavior data of the at least one real-time interaction behavior and historical interaction behavior data of the target object; receiving the second content sequence returned by the server in response to the reordering request, the second content sequence being obtained by the server based on the end feature using a reordering model; updating the search result page based on the second content sequence.

4. The method of claim 1, wherein, The method further comprises: Based on the target media content corresponding to different interaction types in the at least one real-time interactive behavior, at least one interactive content sequence is constructed, and each interactive content sequence corresponds to one interaction type; Based on each media content to be sorted, construct a sequence of content to be sorted; A reordering model is used to determine the target object's level of interest in each unordered media content in the unordered content sequence based on the search keywords, the trigger condition type corresponding to the at least one real-time interactive behavior, and the at least one interactive content sequence. Based on the obtained levels of interest, the second content sequence is obtained, and the search results page is updated based on the second content sequence.

5. The method of claim 4, wherein, The re-ranking model, based on the search keywords, the trigger condition type corresponding to the at least one real-time interaction, and the at least one sequence of interactive content, determines the target object's level of interest in each piece of media content in the sequence to be ranked, including: Encode the search keywords, the trigger condition types, the at least one interactive content sequence, and the content sequence to be sorted to obtain corresponding encoded feature vectors; Based on the obtained encoding feature vectors, the encoding feature vectors of the content sequence to be sorted are subjected to interactive processing based on an attention mechanism to obtain multiple attention feature vectors. The multiple attention feature vectors are concatenated with the encoded feature vectors of the search keywords and the encoded feature vectors of the trigger condition types to obtain the output feature vector; The output feature vectors are mapped to multiple indicator vectors, and the degree of interest is determined based on the multiple indicator vectors. Each indicator vector corresponds to an indicator parameter dimension, and a value in each indicator vector represents the degree of interest of the target object in the corresponding unsorted media content on the corresponding indicator parameter dimension.

6. The method of claim 5, wherein, Encoding any content sequence in the encoding of at least one interactive content sequence and the content sequence to be sorted is performed through the following process: For each media content in any given content sequence, obtain the corresponding basic feature vector; Based on the position of each media content in any content sequence, obtain the position feature vector corresponding to each media content; The content feature vector and position feature vector corresponding to each media content are superimposed to obtain the content feature vector corresponding to each media content. A serialization feature extraction method is used to extract features from each of the obtained content feature vectors to obtain the encoded feature vector corresponding to any content sequence.

7. The method of claim 6, wherein, The method further includes: Based on the real-time interaction behavior data of the at least one real-time interaction behavior and the historical interaction behavior data of the target object, terminal features corresponding to each target media content are extracted. The terminal features are used to characterize the interaction tendency of the target object in the search session. The step of obtaining the corresponding basic feature vector for each media content in any content sequence includes: obtaining a category feature vector corresponding to the content category to which the any target media content belongs; performing feature extraction on the terminal feature corresponding to the any target media content to obtain a terminal feature vector; obtaining a fine arrangement feature vector of the any target media content at a fine arrangement stage; performing splicing processing on the category feature vector, the terminal feature vector and the fine arrangement feature vector to obtain a corresponding basic feature vector.

8. The method of claim 7, wherein, The terminal feature corresponding to each target media content is extracted based on the real-time interaction behavior data of the at least one real-time interaction behavior and the historical interaction behavior data of the target object, and includes: extracting a first feature set of the corresponding target media content based on the real-time interaction behavior data; determining a corresponding association data set of each target media content from the historical interaction behavior data, the association data set of each target media content containing historical interaction data corresponding to media content with the same attribute; extracting a second feature set corresponding to each target media content based on the obtained association data set; obtaining the terminal feature of the corresponding target media content based on the obtained first feature set and the second feature set.

9. The method of claim 5, wherein, The obtained encoding feature vectors are respectively subjected to attention mechanism-based interaction processing on the encoding feature vectors of the content sequence to be sorted to obtain a plurality of attention feature vectors, and include: The following steps are respectively performed for the at least one interactive content sequence: For an interactive content sequence, the attention weight corresponding to the interactive content sequence is determined based on the similarity between the encoding feature vector of the interactive content sequence and the encoding feature vector of the content sequence to be sorted; The attention feature vector corresponding to the interactive content sequence is obtained based on the obtained attention weight and the encoding feature vector of the interactive content sequence.

10. The method of claim 4, wherein, The at least one interactive content sequence is constructed based on the target media content corresponding to different interaction types in the at least one real-time interaction behavior, and includes: An exposure content sequence is constructed based on the target media content displayed in the search result interface; A positive feedback sequence is constructed based on the target media content corresponding to the positive interaction behavior in the at least one real-time interaction behavior and the historical interaction behavior; A negative feedback sequence is constructed based on the target media content corresponding to the negative interaction behavior in the at least one real-time interaction behavior and the historical interaction behavior.

11. A search result reordering apparatus characterized by comprising: The device includes: A page display unit configured to display a search result page corresponding to a search session, the search result page displaying a first content sequence obtained based on a search keyword corresponding to the search session; An acquisition unit configured to acquire at least one real-time interaction behavior of a target object performed on the search result page; The page updating unit is configured to: acquire reordering trigger prediction data, the reordering trigger prediction data including at least one of real-time interaction behavior data of the at least one real-time interaction behavior, historical interaction behavior data of the target object, a current load condition of the terminal device, and a current network condition of the terminal device; input the reordering trigger prediction data into a reordering trigger prediction model to obtain a prediction result indicating whether to trigger reordering; and if the prediction result indicates to trigger reordering, update the search result page, and display a second content sequence on the updated search result page, the second content sequence being obtained by reordering each media content included in the first content sequence, the media content being other media content than the target media content involved in the at least one real-time interaction behavior in the first content sequence. 12.A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein, the processor executes the computer program to implement the steps of the method in any one of claims 1 to 10. 13.A computer storage medium having a computer program stored thereon, wherein, the computer program is executed by a processor to implement the steps of the method in any one of claims 1 to 10. 14.A computer program product comprising a computer program, wherein, the computer program is executed by a processor to implement the steps of the method in any one of claims 1 to 10.

Images