A search result sorting method, device, equipment and storage medium
By combining search term characteristics, multimedia resource characteristics and prediction models, predicting the click-through rate and sorting of videos, the problem of inaccurate sorting of new videos in search results is solved, and the accuracy of search results is improved.
Patent Information
- Application Number
- CN202111277526.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-10-29
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2041-10-29
AI Technical Summary
In the video search scenario, the newly uploaded video lacks posterior behavior data, which leads to inaccurate in the sorting of search results, affecting the accuracy of the search results.
By obtaining the characteristics of the current search term and the characteristics of the multimedia resource, combining the pre-trained prediction model, predict the click rate of the multimedia resource, and sorting iteratively with training samples to improve prediction accuracy using training samples to improve prediction accuracy sex.
Without posterior behavior data, the click rate of multimedia resources can be accurately predicted, the accuracy of search results can be improved, and the results returned by users can be more accurate.
Smart Images

Figure CN113934872B_ABST
Abstract
Description
Technical Field
[0001] The present disclosure relates to the field of computer technology, and in particular to a search result sorting method, device, equipment and storage medium. Background Art
[0002] In the video search scenario, after the user enters the search term, the device will obtain multiple videos matching the search term from the video library, and sort the multiple videos according to their click-through rates to return the search results to the user. The click-through rates required for sorting the multiple videos are determined by the device based on the user's a posteriori behavior data.
[0003] Specifically, the user's posterior behavior data includes the search terms entered by the user in the historical time period, the identifiers of multiple videos matching the search terms, and for each of the multiple videos, the user's behavior data such as whether the user clicked, the viewing time, whether the user liked the video, and whether the user paid attention to the video. The device can determine the attention of each video based on "whether the user clicked", and determine the attention of each video based on "viewing time, whether the user liked the video, and whether the user paid attention to the video". Furthermore, the device calculates the click-through rate of each video based on the attention and satisfaction of each video.
[0004] However, for newly uploaded videos in the video library, there is no corresponding a posteriori behavior data (because the video is new and the user has not clicked it), so the device will set its click rate to 0 and its ranking will be correspondingly backward, which will result in inaccurate ranking in the search results. Summary of the invention
[0005] The present disclosure provides a search result sorting method, device, equipment and storage medium to at least solve the problem of inaccurate search result sorting in the related art. The technical solution of the present disclosure is as follows:
[0006] According to a first aspect of an embodiment of the present disclosure, a method for sorting search results is provided, comprising: obtaining a current search term and a plurality of multimedia resources related to the current search term; predicting a click-through rate of each multimedia resource based on a search term feature of the current search term, a resource feature of each multimedia resource, and a pre-trained prediction model, and sorting the plurality of multimedia resources based on the click-through rate of each multimedia resource; the search term feature of the current search term is used to identify the current search term, the prediction model is trained based on a sample search term, a plurality of sample multimedia resources related to the sample search term, and a sample operation record of each sample multimedia resource; the sample operation record is used to characterize the click-through rate of each sample multimedia resource when a user searches for the sample search term in a historical time period.
[0007] Optionally, the above-mentioned predicting the click-through rate of each multimedia resource based on the search term features of the current search term, the resource features of each multimedia resource and a pre-trained prediction model includes: predicting the attention and satisfaction of each multimedia resource based on the search term features of the current search term, the resource features of each multimedia resource and the prediction model; the attention is used to reflect the order in which the user performs click operations on different multimedia resources, and the satisfaction is used to reflect the user's feedback operation on satisfaction-type information of each multimedia resource; the click-through rate of each multimedia resource is determined based on the attention and satisfaction of each multimedia resource.
[0008] Optionally, the above-mentioned predicting the attention and satisfaction of each multimedia resource based on the search term features of the current search term, the resource features of each multimedia resource and the prediction model includes: determining the fusion features of each multimedia resource based on the search term features of the current search term and the resource features of each multimedia resource; the fusion features include the features obtained by splicing the search term features of the current search term and the resource features of each multimedia resource; the fusion features of each multimedia resource are input into the prediction model to obtain the attention and satisfaction of each multimedia resource.
[0009] Optionally, the above-mentioned method determines the fusion feature of each multimedia resource based on the search term feature of the current search term and the resource feature of each multimedia resource, including: cross-processing the search term feature of the current search term and the resource feature of each multimedia resource to obtain the cross-feature of each multimedia resource and the current search term; splicing the search term feature of the current search term, the resource feature of each multimedia resource and the cross-feature corresponding to each multimedia resource to obtain the fusion feature of each multimedia resource.
[0010] Optionally, the search term feature of the current search term is obtained by concatenating the text feature of the current search term with the embedded feature of the identifier of the current search term.
[0011] Optionally, the resource feature of each multimedia resource is obtained by concatenating at least two of a text feature of the description of each multimedia resource, an image feature of each multimedia resource, and an embedded feature of the identifier of each multimedia resource.
[0012] Optionally, the above method also includes: obtaining multiple groups of training samples; each group of training samples includes a search term feature of a sample search term, a resource feature of a sample multimedia resource matching the sample search term, and a sample attention and a sample satisfaction of the sample multimedia resource; and iteratively training the initial prediction model according to the obtained training samples to obtain a prediction model.
[0013] Optionally, the above-mentioned iterative training of the initial prediction model according to the acquired training samples to obtain a prediction model, including: predicting the estimated attention and estimated satisfaction of the sample multimedia resources according to the search term features of the sample search terms, the resource features of the sample multimedia resources and the initial prediction model; determining the attention loss of the initial prediction model according to the estimated attention of the sample multimedia resources and the sample attention of the sample multimedia resources; determining the satisfaction loss of the initial prediction model according to the estimated satisfaction of the sample multimedia resources and the sample satisfaction of the sample multimedia resources; optimizing the initial prediction model according to the attention loss of the initial prediction model and the satisfaction loss of the initial prediction model to obtain a prediction model.
[0014] Optionally, the above-mentioned determination of the attention loss of the initial prediction model based on the estimated attention of the sample multimedia resources and the sample attention of the sample multimedia resources includes: determining the estimated neglect of the sample multimedia resources based on the estimated attention of the sample multimedia resources; the estimated neglect of the sample multimedia resources is negatively correlated with the estimated attention of the sample multimedia resources; determining the sample neglect of the sample multimedia resources based on the sample attention of the sample multimedia resources; the sample neglect of the sample multimedia resources is negatively correlated with the sample attention of the sample multimedia resources; determining the attention loss of the initial prediction model based on the estimated attention of the sample multimedia resources, the estimated neglect of the sample multimedia resources, the sample attention of the sample multimedia resources and the sample neglect of the sample multimedia resources.
[0015] Optionally, the above-mentioned determination of the attention loss of the initial prediction model based on the estimated attention of the sample multimedia resources, the estimated neglect of the sample multimedia resources, the sample attention of the sample multimedia resources and the sample neglect of the sample multimedia resources includes: determining the logarithmic value of the estimated attention of each sample multimedia resource and the logarithmic value of the estimated neglect of each sample multimedia resource, and calculating the first product and the second product respectively; the first product is the product of the logarithmic value of the estimated attention of each sample multimedia resource and the sample attention of each sample multimedia resource, and the second product is the product of the logarithmic value of the estimated neglect of each sample multimedia resource and the sample neglect of each sample multimedia resource; based on the calculated first product and second product, the attention loss of the initial prediction model is calculated.
[0016] Optionally, the above-mentioned determination of the satisfaction loss of the initial prediction model based on the estimated satisfaction of the sample multimedia resources and the sample satisfaction of the sample multimedia resources includes: determining the estimated exclusion of the sample multimedia resources based on the estimated satisfaction of the sample multimedia resources; the estimated exclusion of the sample multimedia resources is negatively correlated with the estimated satisfaction of the sample multimedia resources; determining the sample exclusion of the sample multimedia resources based on the sample satisfaction of the sample multimedia resources; the sample exclusion of the sample multimedia resources and the sample satisfaction of the sample multimedia resources; determining the satisfaction loss of the initial prediction model based on the estimated satisfaction of the sample multimedia resources, the estimated exclusion of the sample multimedia resources, the sample satisfaction of the sample multimedia resources and the sample exclusion of the sample multimedia resources.
[0017] Optionally, the above-mentioned determination of the satisfaction loss of the initial prediction model based on the estimated satisfaction of the sample multimedia resources, the estimated rejection degree of the sample multimedia resources, the sample satisfaction degree of the sample multimedia resources and the sample rejection degree of the sample multimedia resources includes: determining the logarithmic value of the estimated satisfaction of each sample multimedia resource and the logarithmic value of the estimated rejection degree of each sample multimedia resource, and calculating the third product and the fourth product respectively; the third product is the product of the logarithmic value of the estimated satisfaction of each sample multimedia resource and the sample satisfaction of each sample multimedia resource, and the fourth product is the product of the logarithmic value of the estimated rejection degree of each sample multimedia resource and the sample rejection degree of each sample multimedia resource; based on the calculated third product and fourth product, the satisfaction loss of the initial prediction model is calculated.
[0018] According to a second aspect of an embodiment of the present disclosure, a search result sorting device is provided, comprising an acquisition unit, a prediction unit and a sorting unit; the acquisition unit is used to acquire a current search term and a plurality of multimedia resources related to the current search term; the prediction unit is used to predict a click-through rate of each multimedia resource based on a search term feature of the current search term, a resource feature of each multimedia resource and a pre-trained prediction model; the search term feature of the current search term is used to identify the current search term, and the prediction model is trained based on a sample search term, a plurality of sample multimedia resources related to the sample search term and a sample operation record of each sample multimedia resource; the sample operation record is used to characterize the click-through rate of each sample multimedia resource when a user searches for the sample search term in a historical time period; the sorting unit is used to sort the plurality of multimedia resources based on the click-through rate of each multimedia resource.
[0019] Optionally, the above-mentioned prediction unit is specifically used to: predict the attention and satisfaction of each multimedia resource based on the search term features of the current search term, the resource features of each multimedia resource and the prediction model; the attention is used to reflect the order in which users perform click operations on different multimedia resources, and the satisfaction is used to reflect the user's feedback operation on satisfaction-related information of each multimedia resource; the click-through rate of each multimedia resource is determined based on the attention and satisfaction of each multimedia resource.
[0020] Optionally, the above-mentioned prediction unit is specifically used to: determine the fusion features of each multimedia resource based on the search term features of the current search term and the resource features of each multimedia resource; the fusion features include the features obtained by splicing the search term features of the current search term and the resource features of each multimedia resource; the fusion features of each multimedia resource are input into the prediction model to obtain the attention and satisfaction of each multimedia resource.
[0021] Optionally, the above-mentioned prediction unit is specifically used to: cross-process the search term features of the current search term and the resource features of each multimedia resource to obtain the cross-features of each multimedia resource and the current search term; splice the search term features of the current search term, the resource features of each multimedia resource and the cross-features corresponding to each multimedia resource to obtain the fusion features of each multimedia resource.
[0022] Optionally, the search term feature of the current search term is obtained by concatenating the text feature of the current search term with the embedded feature of the identifier of the current search term.
[0023] Optionally, the resource feature of each multimedia resource is obtained by concatenating at least two of a text feature of the description of each multimedia resource, an image feature of each multimedia resource, and an embedded feature of the identifier of each multimedia resource.
[0024] Optionally, the above-mentioned sorting device also includes a training unit; the acquisition unit is also used to acquire multiple groups of training samples; each group of training samples includes a search term feature of a sample search term, a resource feature of a sample multimedia resource matching the sample search term, and a sample attention and a sample satisfaction of the sample multimedia resource; the training unit is used to iteratively train the initial prediction model according to the acquired training samples to obtain a prediction model.
[0025] Optionally, the above-mentioned training unit is specifically used to: predict the estimated attention and estimated satisfaction of the sample multimedia resources based on the search term features of the sample search terms, the resource features of the sample multimedia resources and the initial prediction model; determine the attention loss of the initial prediction model based on the estimated attention of the sample multimedia resources and the sample attention of the sample multimedia resources; determine the satisfaction loss of the initial prediction model based on the estimated satisfaction of the sample multimedia resources and the sample satisfaction of the sample multimedia resources; optimize the initial prediction model based on the attention loss of the initial prediction model and the satisfaction loss of the initial prediction model to obtain a prediction model.
[0026] Optionally, the above-mentioned training unit is specifically used to: determine the estimated neglect degree of the sample multimedia resource based on the estimated attention of the sample multimedia resource; the estimated neglect degree of the sample multimedia resource is negatively correlated with the estimated attention degree of the sample multimedia resource; determine the sample neglect degree of the sample multimedia resource based on the sample attention of the sample multimedia resource; the sample neglect degree of the sample multimedia resource is negatively correlated with the sample attention degree of the sample multimedia resource; determine the attention loss of the initial prediction model based on the estimated attention of the sample multimedia resource, the estimated neglect degree of the sample multimedia resource, the sample attention of the sample multimedia resource and the sample neglect degree of the sample multimedia resource.
[0027] Optionally, the above-mentioned training unit is specifically used to: determine the logarithmic value of the estimated attention of each sample multimedia resource and the logarithmic value of the estimated neglect of each sample multimedia resource, and calculate the first product and the second product respectively; the first product is the product of the logarithmic value of the estimated attention of each sample multimedia resource and the sample attention of each sample multimedia resource, and the second product is the product of the logarithmic value of the estimated neglect of each sample multimedia resource and the sample neglect of each sample multimedia resource; based on the calculated first product and second product, calculate the attention loss of the initial prediction model.
[0028] Optionally, the above-mentioned training unit is specifically used to: determine the estimated exclusion degree of the sample multimedia resource based on the estimated satisfaction degree of the sample multimedia resource; the estimated exclusion degree of the sample multimedia resource is negatively correlated with the estimated satisfaction degree of the sample multimedia resource; determine the sample exclusion degree of the sample multimedia resource based on the sample satisfaction degree of the sample multimedia resource; the sample exclusion degree of the sample multimedia resource and the sample satisfaction degree of the sample multimedia resource; determine the satisfaction loss of the initial prediction model based on the estimated satisfaction degree of the sample multimedia resource, the estimated exclusion degree of the sample multimedia resource, the sample satisfaction degree of the sample multimedia resource and the sample exclusion degree of the sample multimedia resource.
[0029] Optionally, the above-mentioned training unit is specifically used to: determine the logarithmic value of the estimated satisfaction of each sample multimedia resource and the logarithmic value of the estimated rejection of each sample multimedia resource, and calculate the third product and the fourth product respectively; the third product is the product of the logarithmic value of the estimated satisfaction of each sample multimedia resource and the sample satisfaction of each sample multimedia resource, and the fourth product is the product of the logarithmic value of the estimated rejection of each sample multimedia resource and the sample rejection of each sample multimedia resource; based on the calculated third product and fourth product, calculate the satisfaction loss of the initial prediction model.
[0030] According to a third aspect of an embodiment of the present disclosure, a server is provided, comprising: a processor and a memory for storing instructions executable by the processor; wherein the processor is configured to execute instructions to implement a method for sorting search results as provided in the first aspect and any possible design thereof.
[0031] According to a fourth aspect of an embodiment of the present disclosure, a computer-readable storage medium is provided. When instructions in the computer-readable storage medium are executed by a processor of a server, the server is enabled to execute a method for sorting search results as provided in the first aspect and any possible design thereof.
[0032] According to a fifth aspect of an embodiment of the present disclosure, a computer program product is provided, which includes computer instructions. When the computer instructions are executed on a server, the server executes a search result sorting method as provided in the first aspect and any possible design thereof.
[0033] The technical solution provided by the present disclosure brings at least the following beneficial effects: the click rate of each multimedia resource can be determined according to the search term features of the current search term, the resource features of each multimedia resource and the pre-trained prediction model. Since the prediction model is trained according to the sample search term, multiple sample multimedia resources related to the sample search term and the sample operation record of each sample multimedia resource, and the sample operation record is used to characterize the click rate of each sample multimedia resource when the user searches for the sample search term in the historical time period, in this way, the click rate of the multimedia resource can be predicted in the absence of a posteriori behavior data, and participate in the sorting of the multimedia resources, so that the search results returned to the user can be more accurate.
[0034] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] The drawings herein are incorporated into and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the description are used to explain the principles of the present disclosure, and do not constitute improper limitations on the present disclosure.
[0036] Figure 1 is a structural schematic diagram of a multi-search system according to an exemplary embodiment;
[0037] Figure 2 This is one of the flowchart diagrams of a method for sorting search results according to an exemplary embodiment;
[0038] Figure 3 is a second flow chart of a method for sorting search results according to an exemplary embodiment;
[0039] Figure 4 This is a third flow chart of a method for sorting search results according to an exemplary embodiment;
[0040] Figure 5 This is a fourth flow chart of a method for sorting search results according to an exemplary embodiment;
[0041] Figure 6 is a fifth flow chart of a method for sorting search results according to an exemplary embodiment;
[0042] Figure 7 is a sixth flow chart of a method for sorting search results according to an exemplary embodiment;
[0043] Figure 8 is a structural schematic diagram of a search result sorting device according to an exemplary embodiment;
[0044] Fig. 9 The figure is a schematic diagram showing the structure of a server according to an exemplary embodiment. DETAILED DESCRIPTION
[0045] In order to enable ordinary persons in the art to better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings.
[0046] It should be noted that the terms "first", "second", etc. in the specification and claims of the present disclosure and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.
[0047] In addition, in the description of the embodiments of the present disclosure, unless otherwise specified, " / " means or, for example, A / B can mean A or B. "And / or" in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships, for example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, in the description of the embodiments of the present disclosure, "multiple" means two or more than two.
[0048] A search result sorting method provided by an embodiment of the present disclosure may be applicable to a search system. Figure 1 A schematic diagram of the structure of the search system is shown in FIG. Figure 1 As shown, the search system 10 is used to solve the problem of inaccurate search result sorting in the related art. The search system 10 includes a search result sorting device (hereinafter referred to as a sorting device for ease of description) 11 and a server 12. The sorting device 11 is connected to the server 12. The sorting device 11 and the server 12 can be connected in a wired manner or in a wireless manner, which is not limited in the embodiment of the present disclosure.
[0049] The sorting device 11 may be used to perform data interaction with the server 12 . For example, the sorting device 11 may obtain a search term and search results corresponding to the search term from the server 12 .
[0050] The sorting device 11 can also execute the search result sorting method in the embodiment of the present disclosure, for example, for the acquired search terms and the search results corresponding to the search terms, determine the attention and satisfaction of each multimedia resource in the search results, and calculate the click-through rate of each multimedia resource based on the attention and satisfaction of each multimedia resource, and sort the multimedia resources in the search results based on the determined click-through rate.
[0051] It should be noted that the multimedia resources involved in the embodiments of the present disclosure may include resources such as video, audio, graphics, and text, which are not specifically limited in the embodiments of the present disclosure. At the same time, in the subsequent description of the embodiments of the present disclosure, video is used as an example of a multimedia resource for subsequent description, and other resources such as audio and graphics can refer to the subsequent description.
[0052] The ranking device 11 may also send the ranking result of the multimedia resources in the search result to the server 12 .
[0053] The server 12 is used to receive a search request sent by a user device, determine multimedia resources related to the search term according to the search term in the search request, and send the search term and identifiers of the multimedia resources in the search results to the sorting device 11 .
[0054] At the same time, the server 12 is also used to receive the sorting results sent by the sorting device 11 and return the sorted search results to the user.
[0055] It should be noted that the sorting device 11 and the server 12 may be independent devices or integrated into the same device, and the present invention does not make any specific limitation on this.
[0056] When the sorting device 11 and the server 12 are integrated into the same device, the communication between the sorting device 11 and the server 12 is the communication between the internal modules of the device. In this case, the communication process between the two is the same as "the communication process between the sorting device 11 and the server 12 when they are independent of each other".
[0057] In the following embodiments provided by the present invention, the present invention is described by taking the example that the sorting device 11 and the server 12 are independently configured.
[0058] In practical applications, the search result sorting method provided in the embodiment of the present invention can be applied to a sorting device or to a server. The following describes the search result sorting method provided in the embodiment of the present invention by taking the application of the search result sorting method to a sorting device as an example in conjunction with the accompanying drawings.
[0059] like Figure 2 As shown, the search result sorting method provided by the embodiment of the present disclosure includes the following S201-S204.
[0060] S201. The sorting device obtains a current search term and a plurality of multimedia resources related to the current search term.
[0061] As a possible implementation manner, the sorting device obtains the current search term and a plurality of multimedia resources related to the current search term from a server.
[0062] It should be noted that the number of the above-mentioned multiple multimedia resources is two or more. The current search term is a search term in a search request sent by a user to a server through a user device.
[0063] Correspondingly, after receiving the search request, the server obtains the current search term from the search request, and determines, based on the current search term, a plurality of multimedia resources related to the current search term as search results of the current search term.
[0064] The method for the server to determine multiple multimedia resources related to the current search term may specifically be a method of determining the similarity between the search term and the multimedia resources, which will not be described in detail here.
[0065] As another possible implementation manner, the sorting device obtains the current search term from the server, and determines a plurality of multimedia resources related to the current search term according to the current search term.
[0066] The specific implementation of this step can refer to the above steps of the server, which will not be repeated here. The difference lies in the different execution entities.
[0067] S202: The sorting device determines the search term features of the current search term and the resource features of each multimedia resource respectively.
[0068] The search term feature of a search term is used to uniquely identify a search term, and the resource feature of a multimedia resource is used to uniquely identify a multimedia resource.
[0069] It can be understood that the search term feature of the current search term is used to identify the current search term, and the resource feature of the multimedia resource is used to uniquely identify the multimedia resource.
[0070] The specific implementation of this step can refer to the subsequent description of the embodiment of the present disclosure, and will not be repeated here.
[0071] S203: The sorting device predicts the click rate of each multimedia resource according to the search term features of the current search term, the resource features of each multimedia resource and the pre-trained prediction model.
[0072] The prediction model is trained based on the sample search terms, multiple sample multimedia resources related to the sample search terms, and the sample operation records of each sample multimedia resource. The sample operation records are used to characterize the click rate of each sample multimedia resource when the user searches for the sample search terms in the historical time period. The click rate of a multimedia resource is used to reflect the probability of the multimedia resource being clicked by the user.
[0073] As a possible implementation manner, the ranking device inputs the search term features of the current search term and the resource features of each multimedia resource into a pre-trained prediction model to obtain the click rate of each multimedia resource.
[0074] For the specific implementation of this step, please refer to the subsequent description of the embodiment of the present disclosure, which will not be repeated here.
[0075] S204: The sorting device sorts the multiple multimedia resources according to the click rate of each multimedia resource.
[0076] As a possible implementation manner, the sorting device sorts the multiple multimedia resources according to the click rate of each multimedia resource in the multiple multimedia resources.
[0077] The technical solution provided by the present disclosure brings at least the following beneficial effects: the click rate of each multimedia resource can be determined according to the search term features of the current search term, the resource features of each multimedia resource and the pre-trained prediction model. Since the prediction model is trained according to the sample search term, multiple sample multimedia resources related to the sample search term and the sample operation record of each sample multimedia resource, and the sample operation record is used to characterize the click rate of each sample multimedia resource when the user searches for the sample search term in the historical time period, in this way, the click rate of the multimedia resource can be predicted in the absence of a posteriori behavior data, and participate in the sorting of the multimedia resources, so that the search results returned to the user can be more accurate.
[0078] In one design, in order to determine the click rate of each multimedia resource predicted by the prediction model, such as Figure 3 As shown, S203 provided in the embodiment of the present disclosure specifically includes the following S2031-S2032.
[0079] S2031. The ranking device predicts the attention and satisfaction of each multimedia resource according to the search term features of the current search term, the resource features of each multimedia resource and a pre-trained prediction model.
[0080] Among them, attention is used to reflect the order in which users click on different multimedia resources, and satisfaction is used to reflect the user's feedback operation on satisfaction information of each multimedia resource.
[0081] It can be understood that the attention of a multimedia resource is also used to characterize the degree of attraction of the multimedia resource to users under different search terms. The attention of a multimedia resource is proportional to the order in which users click on the multimedia resource in the search results, that is, the higher the attention of a multimedia resource, the easier it is for the multimedia resource to be clicked by users first.
[0082] The above-mentioned satisfaction information may specifically include information such as playback time, whether it is liked, whether it is followed, etc. The satisfaction level of a multimedia resource is related to the user's feedback operation on the multimedia resource. That is, the higher the satisfaction level of a multimedia resource, the more likely it is that the multimedia resource will be liked, followed or played for a long time by the user.
[0083] It should be noted that the above-mentioned attention and satisfaction are both values greater than or equal to 0 and less than or equal to 1.
[0084] As a possible implementation method, the sorting device determines the fusion characteristics of each multimedia resource based on the search term characteristics of the current search term and the resource characteristics of each multimedia resource, and inputs the fusion characteristics of each multimedia resource into the prediction model to obtain the attention and satisfaction of each multimedia resource.
[0085] The fusion feature includes a fusion feature obtained by splicing the search term feature of the current search term and the resource feature of each multimedia resource.
[0086] For the specific implementation method of this step, please refer to the subsequent description of the embodiment of the present disclosure, which will not be repeated here.
[0087] S2032: The ranking device determines the click rate of each multimedia resource according to the attention and satisfaction of each multimedia resource.
[0088] As a possible implementation manner, the ranking device determines the product of the attention degree and the satisfaction degree of each multimedia resource as the click rate of each multimedia resource.
[0089] The technical solution provided by the present disclosure brings at least the following beneficial effects: considering the correlation between the search term features of the search term and the resource features of the multimedia resource and the attention and satisfaction, the attention and satisfaction of each multimedia resource can be predicted according to the search term features of the current search term, the resource features of each multimedia resource and the pre-trained prediction model, and then the click rate of each multimedia resource can be determined. In this way, the click rate predicted by the prediction model can be guaranteed to be more accurate.
[0090] In order to determine the search term features of the current search term, the embodiments of the present disclosure provide at least one of the following methods for determining the search term features of the current search term.
[0091] In one design, the search term feature of the current search term provided by the embodiment of the present disclosure may be a text feature of the current search term.
[0092] Specifically, the sorting device inputs the current search term into a preset text model, and the text model performs semantic analysis on the current search term to obtain text features of the current search term.
[0093] Exemplarily, the preset text model may be a Chinese word vector model (chinese word2vector), and the text feature of the current search term may be a text vector.
[0094] In one design, the search term feature of the current search term provided by the embodiment of the present disclosure may also be an embedded feature of the identifier of the current search term.
[0095] Specifically, after obtaining the current search term, the sorting device determines the identifier of the current search term in the above-mentioned prediction model, and inputs the identifier of the current search term into a preset word vector model, so as to embed the identifier of the current search term through the word vector model to obtain the embedded features of the identifier of the current search term output by the word vector model.
[0096] It can be understood that the word vector model can obtain the word vector of the current search word identifier through word embedding, so that the search word identifier can be converted into a computer-processable vector form, which improves the processability and expression ability of the search word.
[0097] Exemplarily, the preset word vector model may be a word embedding model, and the embedding feature of the identifier of the current search word may be an embedding vector.
[0098] In one design, in order to make the search term feature of the current search term fully reflect the features of the current search term, the search term feature of the current search term provided by the embodiment of the present disclosure can also be obtained by splicing the text feature of the current search term with the embedded feature of the identifier of the current search term.
[0099] Specifically, after determining the text features of the current search term and the embedded features of the identifier of the current search term, the sorting device concatenates and merges the text features of the current search term with the embedded features of the identifier of the current search term to obtain the search term features of the current search term.
[0100] It should be noted that, in the process of concatenating and merging text features and embedded features, the disclosed embodiment does not limit the order of the text features and embedded features in the search term features of the current search term.
[0101] Exemplarily, when the text feature is a text vector and the embedded feature is an embedded vector, the text feature of the current search term may be [a, b, c], and the embedded feature of the identifier of the current search term may be [d, e, f, g]. Then, the search term feature of the current search term obtained by concatenating and merging the text feature of the current search term with the embedded feature of the identifier of the current search term may be [a, b, c, d, e, f, g].
[0102] It should be noted that the above method of determining the search term features of the current search term can also be applied to the step of determining the search term features of the sample search term in the subsequent training process of the prediction model, and will not be further described in the embodiments of the present disclosure.
[0103] The technical solution provided by the present disclosure brings at least the following beneficial effects: the text features of the current search term are spliced with the embedded features of the identifier of the current search term to obtain the search term features of the current search term, which can ensure that the determined search term features can uniquely and comprehensively identify the current search term.
[0104] In order to determine the resource characteristics of multimedia resources, the embodiments of the present disclosure provide at least one of the following methods for determining the resource characteristics of multimedia resources.
[0105] In one design, the resource features of the multimedia resources provided by the embodiments of the present disclosure may be text features of the description of the multimedia resources.
[0106] Specifically, the sorting device inputs the description of the multimedia resource into a preset text model, and the text model performs semantic analysis on the description of the multimedia resource to obtain text features of the description of the multimedia resource.
[0107] It should be noted that the description of the multimedia resource may be a description in the content of the multimedia resource or a description on the cover of the multimedia resource.
[0108] Exemplarily, the preset text model may be a Chinese word vector model (chinese word2vector), and the text features of the description of the multimedia resource may be text vectors.
[0109] In one design, the resource features of the multimedia resources provided by the embodiments of the present disclosure may also be image features of the multimedia resources.
[0110] Specifically, the sorting device inputs the multimedia resources into a preset convolutional neural network, and the convolutional neural network processes the multimedia resources to obtain image features of the multimedia resources.
[0111] It can be understood that the image features of the multimedia resources are image features of the image frames included in the multimedia resources.
[0112] In one design, the resource features of the multimedia resources provided by the embodiments of the present disclosure may also be embedded features of the identifier of the multimedia resources.
[0113] Specifically, after acquiring the multimedia resources, the sorting device determines the identifier of the multimedia resource in the above-mentioned prediction model, and inputs the identifier of the multimedia resource into a preset word vector model, so as to embed the identifier of the multimedia resource through the word vector model to obtain the embedded features of the identifier of the multimedia resource output by the word vector model.
[0114] It can be understood that the word vector model can obtain the word vector of the multimedia resource identifier through word embedding, so that the search word identifier can be converted into a computer-processable vector form, which improves the processability and expression ability of the search word.
[0115] Exemplarily, the preset word vector model may be a word embedding model, and the embedding feature of the identifier of the multimedia resource may be an embedding vector.
[0116] In one design, in order to make the resource characteristics of multimedia resources fully reflect the characteristics of multimedia resources, the resource characteristics of multimedia resources provided by the embodiment of the present disclosure can also be obtained by splicing at least two of the text characteristics of the description of each multimedia resource, the image characteristics of each multimedia resource, and the embedded characteristics of the identification of each multimedia resource.
[0117] Specifically, after determining the text features of the description of the multimedia resource, the image features of the multimedia resource, and the embedded features of the identifier of the multimedia resource, the sorting device splices and merges at least two of the text features of the description of the multimedia resource, the image features of the multimedia resource, and the embedded features of the identifier of the multimedia resource to obtain the resource features of the multimedia resource.
[0118] It should be noted that in the process of splicing and merging text features, and / or image features, and / or embedded features, the embodiment of the present disclosure does not limit the order of text features, image features and embedded features in the resource features of the multimedia resources.
[0119] Exemplarily, when the text features are text vectors, and the image features and embedded features are embedded vectors, the text features of the description of the above multimedia resources can be [1, 2, 3, 4], the image features of the above multimedia features can be [3, 1, 4, 2], and the embedded features of the identification of the above multimedia resources can be [3, 5, 3, 2]. Then, the resource features of the multimedia resources obtained by splicing and merging the text features of the description of the multimedia resources, the image features of the multimedia resources, and the embedded features of the identification of the multimedia resources can be [1, 2, 3, 4, 3, 1, 4, 2, 3, 5, 3, 2].
[0120] The technical solution provided by the present disclosure brings at least the following beneficial effects: since the resource features of a multimedia resource are obtained by splicing and merging at least two of the text features of the description of the multimedia resource, the image features of the multimedia resource, and the embedded features of the identification of the multimedia resource, it can be ensured that the determined resource features can uniquely and comprehensively identify the multimedia resource.
[0121] In one design, in order to predict the attention and satisfaction of each multimedia resource, Figure 4 As shown, S2031 provided in the embodiment of the present disclosure may specifically include the following S301-S302.
[0122] S301. The sorting device determines the fusion feature of each multimedia resource according to the search term feature of the current search term and the resource feature of each multimedia resource.
[0123] The first fusion feature includes a fusion feature obtained by splicing the search term feature of the current search term and the resource feature of each multimedia resource.
[0124] As a possible implementation manner, the sorting device combines the search term feature of the current search term and the resource feature of each multimedia resource to obtain the fusion feature of each multimedia resource.
[0125] The specific implementation method of splicing and merging the search term features and the resource features in this step can refer to the splicing and merging of feature vectors in the above-mentioned embodiment of the present disclosure, and will not be repeated here. The difference is that the objects of splicing are different.
[0126] In some embodiments, Figure 5 As shown, the embodiment of the present disclosure also shows another implementation method of this step S301, which is specifically as follows S3011-S3012.
[0127] S3011. The sorting device cross-processes the search term feature of the current search term and the resource feature of each multimedia resource to obtain cross-features of each multimedia resource and the current search term.
[0128] Among them, the cross feature is used to reflect the correlation between the current search term and each multimedia resource.
[0129] Exemplarily, the intersection feature may be a combination intersection between the search term feature of the current search term and the resource feature of each multimedia resource.
[0130] The specific implementation of this step can refer to the implementation of the combination of feature vectors in the prior art, which will not be described in detail here.
[0131] S3012: The sorting device combines the search term feature of the current search term, the resource feature of each multimedia resource, and the cross feature corresponding to each multimedia resource to obtain a fusion feature of each multimedia resource.
[0132] As a possible implementation method, the implementation method of this step can specifically refer to the above-mentioned splicing method for feature vectors in the embodiment of the present disclosure. The difference is that the objects of splicing are different, which will not be repeated here.
[0133] The technical solution provided by the present disclosure brings at least the following beneficial effects: cross-processing the search term features of the current search term and the resource features of the multimedia resources, and splicing the search term features, resource features and the cross-features obtained by the cross-processing into fusion features of the multimedia resources. Since the cross-features obtained by the cross-processing reflect the correlation between the search term and the multimedia resources, the fusion features obtained in this way also include the correlation between the current search term and the multimedia resources, which can ensure that the determined fusion features are more comprehensive.
[0134] S302: The sorting device inputs the fusion features of each multimedia resource into the prediction model to obtain the attention and satisfaction of each multimedia resource.
[0135] As a possible implementation manner, the ranking device inputs the fusion features of each multimedia resource into the prediction model respectively to obtain the attention and satisfaction of each multimedia resource output by the prediction model.
[0136] The technical solution provided by the present disclosure brings at least the following beneficial effects: by inputting the fusion features of each multimedia resource into the prediction model, the attention and satisfaction of each multimedia resource can be obtained. Since the fusion features include the fusion features obtained by splicing the search term features of the current search term and the resource features of each multimedia resource, that is, the fusion features include both the search term features of the current search term and the resource features of the multimedia resources, the impact of the attention and satisfaction on the attention and satisfaction can be considered from multiple dimensions of the search terms and multimedia resources, thereby making the predicted attention and satisfaction more accurate.
[0137] In one design, since the prediction model is pre-trained, in order to be able to train the prediction model, such as Figure 6 As shown, the search result sorting method provided by the embodiment of the present disclosure also includes the following S401-S402.
[0138] S401. The sorting device obtains multiple groups of training samples.
[0139] Each group of training samples includes a search term feature of a sample search term, a resource feature of a sample multimedia resource matching the sample search term, and a sample attention and a sample satisfaction of the sample multimedia resource.
[0140] As a possible implementation method, the sorting device obtains from the server sample search terms requested by different users within a historical period of time, as well as search term features of each sample search term, resource features of sample multimedia resources matching each sample search term, and sample attention and sample satisfaction of the sample multimedia resources.
[0141] It should be noted that the sample attention of each sample multimedia resource obtained in this step can be determined based on the user's click operation on the sample multimedia resource in the search results corresponding to the sample search term in the historical time. The sample satisfaction of each sample multimedia resource can be determined based on the user's playback time, whether to like, whether to pay attention to, and other operations on the sample multimedia resource in the historical time.
[0142] In this step, the implementation manner in which the sorting device obtains the search term features of the sample search term can refer to the implementation manner in which the sorting device obtains the search term features of the current search term in the above embodiment of the present disclosure, which will not be repeated here. The difference is that the objects of the obtained search term features are different.
[0143] At the same time, in this step, the implementation method of the sorting device obtaining the resource characteristics of the sample multimedia resources can refer to the implementation method of the sorting device obtaining the resource characteristics of multiple multimedia resources in the above embodiment of the present disclosure, which will not be repeated here. The difference is that the objects of the obtained multimedia resource characteristics are different.
[0144] S402: The sorting device iteratively trains the initial prediction model according to the acquired training samples to obtain a prediction model.
[0145] As a possibility, the sorting device iteratively trains the initial prediction model according to the acquired training samples until the training is fit to obtain a trained prediction model.
[0146] The specific implementation of this step can be referred to the subsequent description of the embodiment of the present disclosure, and will not be repeated here.
[0147] The technical solution provided by the present disclosure brings at least the following beneficial effects: the initial prediction model can be iteratively trained according to the sample data to obtain a trained prediction model. Since the student set in the sample data is the search term features of the sample search terms and the resource features of the sample multimedia resources, and the teacher set is the sample attention and sample satisfaction of the multimedia resources. In this way, the trained prediction model includes the fitting relationship between the search terms, multimedia resources and attention and satisfaction, which can ensure that the predicted attention and satisfaction are more accurate.
[0148] In one design, in order to be able to train a fitted prediction model, such as Figure 7 As shown, S402 provided in the embodiment of the present disclosure specifically includes the following S4021-S4024.
[0149] S4021. The sorting device predicts the estimated attention and the estimated satisfaction of the sample multimedia resource according to the search term features of the sample search term, the resource features of the sample multimedia resource and the initial prediction model.
[0150] Among them, the estimated attention is the predicted attention, and the estimated satisfaction is the predicted satisfaction.
[0151] As a possible implementation method, the sorting device can splice the search term features of the sample search terms and the resource features of the sample multimedia resources to obtain sample fusion features, and input the sample fusion features into the initial prediction model to predict the estimated attention and the estimated satisfaction.
[0152] The specific implementation method of this step can refer to the description of the above-mentioned splicing to obtain the fusion features in the implementation example of the present disclosure, and will not be repeated here.
[0153] As another possible implementation method, the sorting device may cross-process the search term features of the sample search terms and the resource features of the sample multimedia resources to obtain sample cross-features, and splice and merge the search term features of the sample search terms, the resource features of the sample multimedia resources, and the sample cross-features to obtain sample fusion features, and input the sample fusion features into the initial prediction model to predict the estimated attention and the estimated satisfaction.
[0154] The specific implementation method of this step can refer to the description of the above-mentioned splicing to obtain the fusion features in the implementation example of the present disclosure, and will not be repeated here.
[0155] S4022: The sorting device determines the attention loss of the initial prediction model according to the estimated attention of the sample multimedia resource and the sample attention of the sample multimedia resource.
[0156] As a possible implementation method, the sorting device can determine the estimated neglect degree of the sample multimedia resources and the sample neglect degree of the sample multimedia resources, and further determine the attention loss of the initial prediction model based on the estimated attention degree of the sample multimedia resources, the estimated neglect degree of the sample multimedia resources, the sample attention degree of the sample multimedia resources and the sample neglect degree of the sample multimedia resources.
[0157] The estimated neglect degree of the sample multimedia resource is negatively correlated with the estimated attention degree of the sample multimedia resource. The sample neglect degree of the sample multimedia resource is negatively correlated with the sample attention degree of the sample multimedia resource.
[0158] For the specific implementation of this step, please refer to the subsequent description of the embodiment of the present disclosure, which will not be repeated here.
[0159] S4023: The sorting device determines the satisfaction loss of the initial prediction model according to the estimated satisfaction of the sample multimedia resources and the sample satisfaction of the sample multimedia resources.
[0160] As a possible implementation method, the sorting device can determine the estimated exclusion degree of the sample multimedia resources and the sample exclusion degree of the sample multimedia resources, and further determine the satisfaction loss of the initial prediction model based on the estimated satisfaction degree of the sample multimedia resources, the estimated exclusion degree of the sample multimedia resources, the sample satisfaction degree of the sample multimedia resources and the sample exclusion degree of the sample multimedia resources.
[0161] Among them, the estimated repulsion of sample multimedia resources is negatively correlated with the estimated satisfaction of sample multimedia resources. Sample repulsion of sample multimedia resources and sample satisfaction of sample multimedia resources.
[0162] For the specific implementation of this step, please refer to the subsequent description of the embodiment of the present disclosure, which will not be repeated here.
[0163] It should be noted that, in the specific implementation process of the embodiment of the present disclosure, S4022 may be executed first and then S4023, S4023 may be executed first and then S4022, or S4022 and S4023 may be executed simultaneously, and the embodiment of the present disclosure is not limited to this.
[0164] S4024. The sorting device optimizes the initial prediction model according to the attention loss of the initial prediction model and the satisfaction loss of the initial prediction model to obtain a prediction model.
[0165] As a possible implementation method, the sorting device determines the sum of the attention loss of the initial prediction model and the satisfaction loss of the initial prediction model as the overall loss of the initial prediction model, and performs back propagation optimization on the initial prediction model based on the overall loss to obtain the prediction model.
[0166] It should be noted that the method of optimizing the initial prediction model according to the overall loss in this step can refer to the description of the stochastic gradient descent (SGD) method in back propagation, which will not be repeated here.
[0167] The technical solution provided by the present disclosure brings at least the following beneficial effects: the initial prediction model can be back-propagated and optimized through corresponding attention loss and satisfaction loss, thereby ensuring the accuracy of the initial prediction model during the fitting process.
[0168] In some designs, in order to determine the attention loss of the initial prediction model, the above S4022 provided in the embodiment of the present disclosure specifically includes the following 501-S503.
[0169] S501: The sorting device determines the estimated neglect degree of the sample multimedia resource according to the estimated attention degree of the sample multimedia resource.
[0170] The estimated neglect degree of the sample multimedia resource is negatively correlated with the estimated attention degree of the sample multimedia resource. The estimated neglect degree is used to indicate the predicted probability that the user does not click on the sample multimedia resource.
[0171] As a possible implementation manner, the sorting device may use the difference between 1 and the estimated attention degree as the estimated neglect degree of the sample multimedia resource.
[0172] S502: The sorting device determines the sample neglect degree of the sample multimedia resource according to the sample attention degree of the sample multimedia resource.
[0173] The sample neglect degree of the sample multimedia resource is negatively correlated with the sample attention degree of the sample multimedia resource. The sample neglect degree of the sample multimedia resource is used to indicate the probability that the user does not pay attention to the sample multimedia resource.
[0174] As a possible implementation manner, the sorting device may determine the difference between 1 and the sample attention degree of the sample multimedia resource as the sample neglect degree of the sample multimedia resource.
[0175] S503: The sorting device determines the attention loss of the initial prediction model according to the estimated attention of the sample multimedia resources, the estimated neglect of the sample multimedia resources, the sample attention of the sample multimedia resources, and the sample neglect of the sample multimedia resources.
[0176] As a possible implementation, the ranking device determines the logarithmic value of the estimated attention and the logarithmic value of the estimated neglect of each sample multimedia resource, and calculates the first product and the second product respectively.
[0177] The first product is the product of the logarithm of the estimated attention of each sample multimedia resource and the sample attention, and the second product is the product of the logarithm of the estimated neglect of each sample multimedia resource and the sample neglect.
[0178] Furthermore, the sorting device calculates the attention loss of the initial prediction model based on the calculated first product and the second product.
[0179] In some embodiments, the attention loss of the initial prediction model satisfies the following formula 1:
[0180]
[0181] Among them, loss attractive is the attention loss of the initial prediction model, is the sample attention of the ith multimedia resource in the sample multimedia resources, p is the estimated attention of the ith sample multimedia resource, is the first product, is the sample neglect degree of the ith sample multimedia resource, 1-p is the estimated neglect degree of the ith sample multimedia resource, is the second product.
[0182] The technical solution provided by the present disclosure brings at least the following beneficial effects: through the above formula, a specific implementation method for determining the attention loss of the initial prediction model can be given, which can ensure that the attention loss can be calculated.
[0183] In some designs, in order to determine the satisfaction loss of the initial prediction model, the above S4023 provided in the embodiment of the present disclosure specifically includes the following 601-S603.
[0184] S601: The sorting device determines the estimated exclusion degree of the sample multimedia resources according to the estimated satisfaction degree of the sample multimedia resources.
[0185] The estimated repulsion degree of the sample multimedia resource is negatively correlated with the estimated satisfaction degree of the sample multimedia resource. The estimated repulsion degree is used to indicate the predicted probability that the user is not satisfied with the sample multimedia resource.
[0186] As a possible implementation manner, the sorting device may use the difference between 1 and the estimated satisfaction level as the estimated rejection level of the sample multimedia resource.
[0187] S602: The sorting device determines the sample exclusion degree of the sample multimedia resource according to the sample satisfaction degree of the sample multimedia resource.
[0188] The sample rejection degree of the sample multimedia resource is negatively correlated with the sample satisfaction degree of the sample multimedia resource. The sample rejection degree of the sample multimedia resource is used to indicate the probability that the user is not satisfied with the sample multimedia resource.
[0189] As a possible implementation manner, the sorting device may determine the difference between 1 and the sample satisfaction degree of the sample multimedia resource as the sample exclusion degree of the sample multimedia resource.
[0190] S603: The sorting device determines the satisfaction loss of the initial prediction model according to the estimated satisfaction of the sample multimedia resources, the estimated repulsion of the sample multimedia resources, the sample satisfaction of the sample multimedia resources, and the sample repulsion of the sample multimedia resources.
[0191] As a possible implementation, the ranking device determines the logarithmic value of the estimated satisfaction and the logarithmic value of the estimated repulsion of each sample multimedia resource, and calculates the third product and the fourth product respectively.
[0192] The third product is the product of the logarithm of the estimated satisfaction of each sample multimedia resource and the sample satisfaction, and the fourth product is the product of the logarithm of the estimated repulsion of each sample multimedia resource and the sample repulsion.
[0193] Furthermore, the ranking device calculates the satisfaction loss of the initial prediction model based on the calculated third product and fourth product.
[0194] In some embodiments, the attention loss of the initial prediction model satisfies the following formula 2:
[0195]
[0196] Among them, loss satisfy is the satisfaction loss of the initial prediction model, is the sample satisfaction of the i-th sample multimedia resource in the sample multimedia resources, q is the estimated satisfaction of the i-th sample multimedia resource, is the third product, is the sample exclusion degree of the ith multimedia resource, 1-q is the estimated exclusion degree of the ith multiple media resources, is the fourth product.
[0197] The technical solution provided by the present disclosure brings at least the following beneficial effects: through the above formula, a specific implementation method for determining the satisfaction loss of the initial prediction model can be given, which can ensure that the satisfaction loss can be calculated.
[0198] Figure 8 FIG. 1 is a schematic diagram of a structure of a search result sorting device according to an exemplary embodiment. Figure 8As shown, the search result ranking device 70 provided in the embodiment of the present disclosure includes an acquisition unit 701 , a prediction unit 702 and a ranking unit 703 .
[0199] The acquisition unit 701 is used to acquire a current search term and a plurality of multimedia resources related to the current search term.
[0200] The prediction unit 702 is used to predict the click rate of each multimedia resource based on the search term feature of the current search term, the resource feature of each multimedia resource and the pre-trained prediction model. The search term feature of the current search term is used to identify the current search term, and the prediction model is trained based on the sample search term, multiple sample multimedia resources related to the sample search term and the sample operation record of each sample multimedia resource. The sample operation record is used to characterize the click rate of each sample multimedia resource when the user searches for the sample search term in the historical time period.
[0201] The sorting unit 703 is used to sort the multiple multimedia resources according to the click rate of each multimedia resource.
[0202] Optional, such as Figure 8 As shown, the prediction unit 702 provided in the embodiment of the present disclosure is specifically used for:
[0203] According to the search term features of the current search term, the resource features of each multimedia resource and the prediction model, the attention and satisfaction of each multimedia resource are predicted. The attention is used to reflect the order in which users click on different multimedia resources, and the satisfaction is used to reflect the user's feedback operation on the satisfactory information of each multimedia resource.
[0204] Determine the click rate of each multimedia resource based on the attention and satisfaction of each multimedia resource.
[0205] Optional, such as Figure 8 As shown, the prediction unit 702 provided in the embodiment of the present disclosure is specifically used for:
[0206] According to the search term feature of the current search term and the resource feature of each multimedia resource, the fusion feature of each multimedia resource is determined. The fusion feature includes a feature obtained by splicing the search term feature of the current search term and the resource feature of each multimedia resource.
[0207] The fusion features of each multimedia resource are input into the prediction model to obtain the attention and satisfaction of each multimedia resource.
[0208] Optional, such as Figure 8 As shown, the prediction unit 702 provided in the embodiment of the present disclosure is specifically used for:
[0209] The search term feature of the current search term and the resource feature of each multimedia resource are cross-processed to obtain cross-features of each multimedia resource and the current search term.
[0210] The search term feature of the current search term, the resource feature of each multimedia resource, and the cross feature corresponding to each multimedia resource are concatenated to obtain the fusion feature of each multimedia resource.
[0211] Optional, such as Figure 8 As shown, the search term feature of the current search term provided by the embodiment of the present disclosure is obtained by concatenating the text feature of the current search term with the embedded feature of the identifier of the current search term.
[0212] Optional, such as Figure 8 As shown, the resource feature of each multimedia resource provided by the embodiment of the present disclosure is obtained by splicing at least two of the text feature of the description of each multimedia resource, the image feature of each multimedia resource, and the embedded feature of the identifier of each multimedia resource.
[0213] Optional, such as Figure 8 As shown, the sorting device 70 provided in the embodiment of the present disclosure further includes a training unit 704 .
[0214] The acquisition unit 701 is further used to acquire multiple groups of training samples. Each group of training samples includes a search term feature of a sample search term, resource features of sample multimedia resources matching the sample search term, and sample attention and sample satisfaction of the sample multimedia resources.
[0215] The training unit 704 is used to iteratively train the initial prediction model according to the acquired training samples to obtain a prediction model.
[0216] Optional, such as Figure 8 As shown, the training unit 704 provided in the embodiment of the present disclosure is specifically used for:
[0217] According to the search term features of the sample search terms, the resource features of the sample multimedia resources and the initial prediction model, the estimated attention and the estimated satisfaction of the sample multimedia resources are predicted.
[0218] The attention loss of the initial prediction model is determined according to the estimated attention of the sample multimedia resource and the sample attention of the sample multimedia resource.
[0219] The satisfaction loss of the initial prediction model is determined according to the estimated satisfaction of the sample multimedia resources and the sample satisfaction of the sample multimedia resources.
[0220] According to the attention loss of the initial prediction model and the satisfaction loss of the initial prediction model, the initial prediction model is optimized to obtain a prediction model.
[0221] Optional, such as Figure 8 As shown, the training unit 704 provided in the embodiment of the present disclosure is specifically used for:
[0222] According to the estimated attention degree of the sample multimedia resource, the estimated neglect degree of the sample multimedia resource is determined. The estimated neglect degree of the sample multimedia resource is negatively correlated with the estimated attention degree of the sample multimedia resource.
[0223] According to the sample attention degree of the sample multimedia resource, the sample neglect degree of the sample multimedia resource is determined. The sample neglect degree of the sample multimedia resource is negatively correlated with the sample attention degree of the sample multimedia resource.
[0224] The attention loss of the initial prediction model is determined according to the estimated attention of the sample multimedia resource, the estimated neglect of the sample multimedia resource, the sample attention of the sample multimedia resource, and the sample neglect of the sample multimedia resource.
[0225] Optional, such as Figure 8 As shown, the training unit 704 provided in the embodiment of the present disclosure is specifically used for:
[0226] Determine the logarithmic value of the estimated attention of each sample multimedia resource and the logarithmic value of the estimated neglect of each sample multimedia resource, and calculate a first product and a second product respectively. The first product is the product of the logarithmic value of the estimated attention of each sample multimedia resource and the sample attention of each sample multimedia resource, and the second product is the product of the logarithmic value of the estimated neglect of each sample multimedia resource and the sample neglect of each sample multimedia resource.
[0227] According to the calculated first product and second product, the attention loss of the initial prediction model is calculated.
[0228] Optional, such as Figure 8 As shown, the training unit 704 provided in the embodiment of the present disclosure is specifically used for:
[0229] According to the estimated satisfaction of the sample multimedia resource, the estimated repulsion degree of the sample multimedia resource is determined. The estimated repulsion degree of the sample multimedia resource is negatively correlated with the estimated satisfaction degree of the sample multimedia resource.
[0230] According to the sample satisfaction of the sample multimedia resource, the sample exclusion degree of the sample multimedia resource is determined. The sample exclusion degree of the sample multimedia resource and the sample satisfaction degree of the sample multimedia resource.
[0231] The satisfaction loss of the initial prediction model is determined according to the estimated satisfaction of the sample multimedia resources, the estimated repulsion of the sample multimedia resources, the sample satisfaction of the sample multimedia resources, and the sample repulsion of the sample multimedia resources.
[0232] Optional, such as Figure 8 As shown, the training unit 704 provided in the embodiment of the present disclosure is specifically used for:
[0233] Determine the logarithm of the estimated satisfaction of each sample multimedia resource and the logarithm of the estimated repulsion of each sample multimedia resource, and calculate the third product and the fourth product respectively. The third product is the product of the logarithm of the estimated satisfaction of each sample multimedia resource and the sample satisfaction of each sample multimedia resource, and the fourth product is the product of the logarithm of the estimated repulsion of each sample multimedia resource and the sample repulsion of each sample multimedia resource.
[0234] According to the calculated third product and fourth product, the satisfaction loss of the initial prediction model is calculated.
[0235] Regarding the device in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.
[0236] Fig. 9 Schematic diagram of the structure of a server provided by the present disclosure. Fig. 9 The server 80 may include at least one processor 801 and a memory 803 for storing processor executable instructions. The processor 801 is configured to execute instructions in the memory 803 to implement the search result sorting method in the above embodiment.
[0237] In addition, the server 80 may further include a communication bus 802 and at least one communication interface 804 .
[0238] The processor 801 may be a central processing unit (CPU), a microprocessing unit, an ASIC, or one or more integrated circuits for controlling the execution of the program of the disclosed solution.
[0239] The communication bus 802 may include a pathway for transmitting information between the above-mentioned components.
[0240] The communication interface 804 uses any transceiver or other device for communicating with other devices or communication networks, such as Ethernet, radio access network (RAN), wireless local area networks (WLAN), etc.
[0241] The memory 803 may be a read-only memory (ROM) or other types of static storage devices that can store static information and instructions, a random access memory (RAM) or other types of dynamic storage devices that can store information and instructions, or an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical disc, laser disc, optical disc, digital versatile disc, Blu-ray disc, etc.), a magnetic disk storage medium or other magnetic storage device, or any other medium that can be used to carry or store the desired program code in the form of an instruction or data structure and can be accessed by a computer, but is not limited thereto. The memory may exist independently and be connected to the processing unit through a bus. The memory may also be integrated with the processing unit.
[0242] The memory 803 is used to store instructions for executing the solution of the present disclosure, and the execution is controlled by the processor 801. The processor 801 is used to execute the instructions stored in the memory 803, so as to realize the functions in the method of the present disclosure.
[0243] As an example, combining Figure 7 The functions implemented by the acquisition unit 501, the prediction unit 502, the determination unit 503, the sorting unit 504 and the training unit 505 in the search result sorting device 50 are similar to those Fig. 9 The function of processor 801 in is the same.
[0244] In a specific implementation, as an embodiment, the processor 801 may include one or more CPUs, such as Fig. 9 CPU0 and CPU1 in.
[0245] In a specific implementation, as an embodiment, the server 80 may include multiple processors, such as Fig. 9 801 and processor 807 in the embodiment of the present invention. Each of these processors may be a single-CPU processor or a multi-CPU processor. The processor herein may refer to one or more devices, circuits, and / or processing cores for processing data (e.g., computer program instructions).
[0246] In a specific implementation, as an embodiment, the server 80 may further include an output device 805 and an input device 806. The output device 805 communicates with the processor 801 and may display information in a variety of ways. For example, the output device 805 may be a liquid crystal display (LCD), a light emitting diode (LED) display device, a cathode ray tube (CRT) display device, or a projector. The input device 806 communicates with the processor 801 and may receive user input in a variety of ways. For example, the input device 806 may be a mouse, a keyboard, a touch screen device, or a sensor device.
[0247] Those skilled in the art will understand that Fig. 9 The structure shown in the figure does not constitute a limitation on the server 80, and the server 80 may include more or less components than those shown in the figure, or combine certain components, or adopt a different component arrangement.
[0248] In addition, the present disclosure also provides a computer-readable storage medium. When the instructions in the computer-readable storage medium are executed by a processor of a server, the server can execute the search result ranking method provided in the above embodiment.
[0249] In addition, the present disclosure also provides a computer program product, including computer instructions. When the computer instructions are executed on a server, the server executes the search result sorting method provided in the above embodiment.
[0250] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. The present disclosure is intended to cover any variations, uses or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The description and examples are to be considered exemplary only, and the true scope and spirit of the present disclosure are indicated by the claims.
Claims
1. A method for sorting search results, characterized in that: include: Acquire a current search term and a plurality of multimedia resources related to the current search term; According to the search term features of the current search term, the resource features of each multimedia resource and the prediction model, the attention and satisfaction of each multimedia resource are predicted; the attention is used to reflect the order in which users click on different multimedia resources, and the satisfaction is used to reflect the user's feedback operation on the satisfaction-type information of each multimedia resource; the search term features of the current search term are used to identify the current search term, and the prediction model is trained based on sample search terms, multiple sample multimedia resources related to the sample search terms, and sample operation records of each sample multimedia resource; The sample operation record is used to represent the click rate of each sample multimedia resource when the user searches for the sample search term in a historical time period; Determining the click rate of each multimedia resource according to the attention and satisfaction of each multimedia resource; The multiple multimedia resources are sorted according to the click rate of each multimedia resource.
2. The search result sorting method according to claim 1, characterized in that: The step of predicting the attention and satisfaction of each multimedia resource according to the search term feature of the current search term, the resource feature of each multimedia resource and the prediction model comprises: Determine, according to the search term feature of the current search term and the resource feature of each multimedia resource, a fusion feature of each multimedia resource; the fusion feature includes a feature obtained by splicing the search term feature of the current search term and the resource feature of each multimedia resource; The fusion features of each multimedia resource are input into the prediction model to obtain the attention degree and the satisfaction degree of each multimedia resource.
3. The search result sorting method according to claim 2, characterized in that: The determining the fusion feature of each multimedia resource according to the search term feature of the current search term and the resource feature of each multimedia resource includes: Cross-processing the search term feature of the current search term and the resource feature of each multimedia resource to obtain a cross feature of each multimedia resource and the current search term; The search term feature of the current search term, the resource feature of each multimedia resource, and the cross feature corresponding to each multimedia resource are spliced to obtain the fusion feature of each multimedia resource.
4. The method for sorting search results according to any one of claims 1 to 3, characterized in that: The search term feature of the current search term is obtained by concatenating the text feature of the current search term with the embedded feature of the identifier of the current search term.
5. The method for sorting search results according to any one of claims 1 to 3, characterized in that: The resource feature of each multimedia resource is obtained by concatenating at least two of the text feature of the description of each multimedia resource, the image feature of each multimedia resource, and the embedded feature of the identifier of each multimedia resource.
6. The method for sorting search results according to any one of claims 1 to 3, characterized in that: The method further comprises: Acquire multiple groups of training samples; each group of training samples includes a search term feature of a sample search term, a resource feature of a sample multimedia resource matching the sample search term, and a sample attention degree and a sample satisfaction degree of the sample multimedia resource; The initial prediction model is iteratively trained according to the acquired training samples to obtain the prediction model.
7. The search result ranking method according to claim 6, characterized in that: The iterative training of the initial prediction model according to the acquired training samples to obtain the prediction model includes: Predicting an estimated attention degree and an estimated satisfaction degree of the sample multimedia resource according to the search term features of the sample search term, the resource features of the sample multimedia resource and the initial prediction model; Determining the attention loss of the initial prediction model according to the estimated attention of the sample multimedia resource and the sample attention of the sample multimedia resource; Determining the satisfaction loss of the initial prediction model according to the estimated satisfaction of the sample multimedia resource and the sample satisfaction of the sample multimedia resource; The initial prediction model is optimized according to the attention loss of the initial prediction model and the satisfaction loss of the initial prediction model to obtain the prediction model.
8. The search result ranking method according to claim 7, characterized in that: The determining the attention loss of the initial prediction model according to the estimated attention of the sample multimedia resource and the sample attention of the sample multimedia resource comprises: Determining an estimated neglect degree of the sample multimedia resource according to the estimated attention degree of the sample multimedia resource; the estimated neglect degree of the sample multimedia resource is negatively correlated with the estimated attention degree of the sample multimedia resource; Determining a sample neglect degree of the sample multimedia resource according to the sample attention degree of the sample multimedia resource; the sample neglect degree of the sample multimedia resource is negatively correlated with the sample attention degree of the sample multimedia resource; The attention loss of the initial prediction model is determined according to the estimated attention of the sample multimedia resource, the estimated neglect of the sample multimedia resource, the sample attention of the sample multimedia resource and the sample neglect of the sample multimedia resource.
9. The search result sorting method according to claim 8, characterized in that: The determining the attention loss of the initial prediction model according to the estimated attention of the sample multimedia resource, the estimated neglect of the sample multimedia resource, the sample attention of the sample multimedia resource, and the sample neglect of the sample multimedia resource comprises: Determine the logarithmic value of the estimated attention of each sample multimedia resource and the logarithmic value of the estimated neglect of each sample multimedia resource, and calculate a first product and a second product respectively; the first product is the product of the logarithmic value of the estimated attention of each sample multimedia resource and the sample attention of each sample multimedia resource, and the second product is the product of the logarithmic value of the estimated neglect of each sample multimedia resource and the sample neglect of each sample multimedia resource; The attention loss of the initial prediction model is calculated based on the calculated first product and the calculated second product.
10. The search result ranking method according to claim 7, characterized in that: The determining of the satisfaction loss of the initial prediction model according to the estimated satisfaction of the sample multimedia resource and the sample satisfaction of the sample multimedia resource includes: Determining an estimated degree of repulsion of the sample multimedia resource according to the estimated degree of satisfaction of the sample multimedia resource; the estimated degree of repulsion of the sample multimedia resource is negatively correlated with the estimated degree of satisfaction of the sample multimedia resource; Determining the sample exclusion degree of the sample multimedia resource according to the sample satisfaction degree of the sample multimedia resource; the sample exclusion degree of the sample multimedia resource and the sample satisfaction degree of the sample multimedia resource; The satisfaction loss of the initial prediction model is determined according to the estimated satisfaction of the sample multimedia resource, the estimated repulsion of the sample multimedia resource, the sample satisfaction of the sample multimedia resource and the sample repulsion of the sample multimedia resource.
11. The search result sorting method according to claim 10, characterized in that: The determining of the satisfaction loss of the initial prediction model according to the estimated satisfaction of the sample multimedia resource, the estimated repulsion of the sample multimedia resource, the sample satisfaction of the sample multimedia resource, and the sample repulsion of the sample multimedia resource comprises: Determine the logarithmic value of the estimated satisfaction of each sample multimedia resource and the logarithmic value of the estimated repulsion of each sample multimedia resource, and calculate a third product and a fourth product respectively; the third product is the product of the logarithmic value of the estimated satisfaction of each sample multimedia resource and the sample satisfaction of each sample multimedia resource, and the fourth product is the product of the logarithmic value of the estimated repulsion of each sample multimedia resource and the sample repulsion of each sample multimedia resource; The satisfaction loss of the initial prediction model is calculated based on the calculated third product and the calculated fourth product.
12. A search result sorting device, characterized in that: It includes an acquisition unit, a prediction unit and a sorting unit; The acquisition unit is used to acquire a current search term and a plurality of multimedia resources related to the current search term; The prediction unit is used to predict the attention and satisfaction of each multimedia resource based on the search term features of the current search term, the resource features of each multimedia resource and the prediction model; the attention is used to reflect the order in which users perform click operations on different multimedia resources, and the satisfaction is used to reflect the user's feedback operation on the satisfaction-type information of each multimedia resource; the search term features of the current search term are used to identify the current search term, and the prediction model is trained based on sample search terms, multiple sample multimedia resources related to the sample search terms, and sample operation records of each sample multimedia resource; the sample operation records are used to characterize the click rate of each sample multimedia resource when users search for the sample search terms in a historical time period; The prediction unit is further used to determine the click rate of each multimedia resource according to the attention and satisfaction of each multimedia resource; The sorting unit is used to sort the multiple multimedia resources according to the click rate of each multimedia resource.
13. The search result ranking device according to claim 12, characterized in that: The prediction unit is specifically used for: Determine, according to the search term feature of the current search term and the resource feature of each multimedia resource, a fusion feature of each multimedia resource; the fusion feature includes a feature obtained by splicing the search term feature of the current search term and the resource feature of each multimedia resource; The fusion features of each multimedia resource are input into the prediction model to obtain the attention degree and the satisfaction degree of each multimedia resource.
14. The search result ranking device according to claim 13, characterized in that: The prediction unit is specifically used for: Cross-processing the search term feature of the current search term and the resource feature of each multimedia resource to obtain a cross feature of each multimedia resource and the current search term; The search term feature of the current search term, the resource feature of each multimedia resource, and the cross feature corresponding to each multimedia resource are spliced to obtain the fusion feature of each multimedia resource.
15. The search result ranking device according to any one of claims 12 to 14, characterized in that: The search term feature of the current search term is obtained by concatenating the text feature of the current search term with the embedded feature of the identifier of the current search term.
16. The search result ranking device according to any one of claims 12 to 14, characterized in that: The resource feature of each multimedia resource is obtained by concatenating at least two of the text feature of the description of each multimedia resource, the image feature of each multimedia resource, and the embedded feature of the identifier of each multimedia resource.
17. The search result ranking device according to any one of claims 12 to 14, characterized in that: The sorting device also includes a training unit; The acquisition unit is further used to acquire multiple groups of training samples; each group of training samples includes a search term feature of a sample search term, a resource feature of a sample multimedia resource matching the sample search term, and a sample attention degree and a sample satisfaction degree of the sample multimedia resource; The training unit is used to iteratively train the initial prediction model according to the acquired training samples to obtain the prediction model.
18. The search result ranking device according to claim 17, characterized in that: The training unit is specifically used for: Predicting an estimated attention degree and an estimated satisfaction degree of the sample multimedia resource according to the search term features of the sample search term, the resource features of the sample multimedia resource and the initial prediction model; Determining the attention loss of the initial prediction model according to the estimated attention of the sample multimedia resource and the sample attention of the sample multimedia resource; Determining the satisfaction loss of the initial prediction model according to the estimated satisfaction of the sample multimedia resource and the sample satisfaction of the sample multimedia resource; The initial prediction model is optimized according to the attention loss of the initial prediction model and the satisfaction loss of the initial prediction model to obtain the prediction model.
19. The search result ranking device according to claim 18, characterized in that: The training unit is specifically used for: Determining an estimated neglect degree of the sample multimedia resource according to the estimated attention degree of the sample multimedia resource; the estimated neglect degree of the sample multimedia resource is negatively correlated with the estimated attention degree of the sample multimedia resource; Determining a sample neglect degree of the sample multimedia resource according to the sample attention degree of the sample multimedia resource; The sample neglect degree of the sample multimedia resource is negatively correlated with the sample attention degree of the sample multimedia resource; The attention loss of the initial prediction model is determined according to the estimated attention of the sample multimedia resource, the estimated neglect of the sample multimedia resource, the sample attention of the sample multimedia resource and the sample neglect of the sample multimedia resource.
20. The search result ranking device according to claim 19, characterized in that: The training unit is specifically used for: Determine the logarithmic value of the estimated attention of each sample multimedia resource and the logarithmic value of the estimated neglect of each sample multimedia resource, and calculate a first product and a second product respectively; the first product is the product of the logarithmic value of the estimated attention of each sample multimedia resource and the sample attention of each sample multimedia resource, and the second product is the product of the logarithmic value of the estimated neglect of each sample multimedia resource and the sample neglect of each sample multimedia resource; The attention loss of the initial prediction model is calculated based on the calculated first product and the calculated second product.
21. The search result ranking device according to claim 18, characterized in that: The training unit is specifically used for: Determining an estimated degree of repulsion of the sample multimedia resource according to the estimated degree of satisfaction of the sample multimedia resource; the estimated degree of repulsion of the sample multimedia resource is negatively correlated with the estimated degree of satisfaction of the sample multimedia resource; Determining the sample exclusion degree of the sample multimedia resource according to the sample satisfaction degree of the sample multimedia resource; the sample exclusion degree of the sample multimedia resource and the sample satisfaction degree of the sample multimedia resource; The satisfaction loss of the initial prediction model is determined according to the estimated satisfaction of the sample multimedia resource, the estimated repulsion of the sample multimedia resource, the sample satisfaction of the sample multimedia resource and the sample repulsion of the sample multimedia resource.
22. The search result ranking device according to claim 21, characterized in that: The training unit is specifically used for: Determine the logarithmic value of the estimated satisfaction of each sample multimedia resource and the logarithmic value of the estimated repulsion of each sample multimedia resource, and calculate a third product and a fourth product respectively; the third product is the product of the logarithmic value of the estimated satisfaction of each sample multimedia resource and the sample satisfaction of each sample multimedia resource, and the fourth product is the product of the logarithmic value of the estimated repulsion of each sample multimedia resource and the sample repulsion of each sample multimedia resource; The satisfaction loss of the initial prediction model is calculated based on the calculated third product and the calculated fourth product.
23. A server, characterized in that: include: A processor and a memory for storing instructions executable by the processor; wherein the processor is configured to execute instructions to implement the search result sorting method according to any one of claims 1 to 11.
24. A computer-readable storage medium, characterized in that: When the instructions in the computer-readable storage medium are executed by a processor of a server, the server is enabled to execute the search result ranking method according to any one of claims 1 to 11.
25. A computer program product comprising instructions, characterized in that The computer program product comprises computer instructions, and when the computer instructions are executed on a server, the server is caused to execute the search result ranking method according to any one of claims 1 to 11.
