Method for training search term recommendation model, search term recommendation method and device

Abstract

The present disclosure relates to a search term recommendation model training method, a search term recommendation method and device, the search term recommendation model comprising a feature extraction network and a click rate estimation network, the training method comprising: obtaining a video sample and a sample label thereof; extracting a plurality of search terms and content information corresponding to the video sample, and inputting the plurality of search terms and the content information into the feature extraction network to obtain a plurality of search term embedding features corresponding to the plurality of search terms and a content embedding feature corresponding to the content information; determining N search term embedding features close to the content embedding feature from the plurality of search term embedding features; inputting the N search term embedding features and the content embedding feature into the click rate estimation network to obtain an estimated click rate of each search term embedding feature; determining a value of a loss function of the search term recommendation model according to the estimated click rate and the sample label; and adjusting parameters of the feature extraction network and the click rate estimation network according to the value of the loss function to obtain a trained search term recommendation model.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of search, and in particular, to a search term recommendation model training method, a search term recommendation method and device. BACKGROUND

[0002] It is a common means for users to find interested content by initiating a search using a search term. Currently, in order to assist users in searching, there are schemes for recommending search terms to users, such as recommending search terms searched by other users (e.g., users with the same interests as the current user) to the current user, or further supplementing a search term entered by the user for the user to select (e.g., after the user enters “cat”, further supplementing “cat avatar”, “cat funny video”, etc.). However, these methods are often obtained by counting high-frequency search terms of other users, and are difficult to effectively match the search needs of the current user. SUMMARY

[0003] The present disclosure provides a search term recommendation model training method, a search term recommendation method and device to at least solve the problem that the related art is difficult to effectively match the search needs of the current user, and can not solve any of the above problems.

[0004] According to a first aspect of the present disclosure, a search term recommendation model training method is provided, the search term recommendation model comprising a feature extraction network and a click rate estimation network, the training method comprising: obtaining a video sample and a sample label thereof; extracting a plurality of search terms and content information corresponding to the video sample, and inputting the feature extraction network to extract a plurality of search term embedding features corresponding to the plurality of search terms and content embedding features corresponding to the content information; determining N search term embedding features close to the content embedding features from the plurality of search term embedding features; inputting the N search term embedding features and the content embedding features into the click rate estimation network to obtain an estimated click rate of each search term embedding feature; determining a value of a loss function of the search term recommendation model according to the estimated click rate and the sample label; and adjusting parameters of the feature extraction network and the click rate estimation network according to the value of the loss function to obtain a trained search term recommendation model.

[0005] Optionally, the obtaining a video sample and a sample label thereof comprises: obtaining a candidate video set; the candidate video set comprises a plurality of candidate videos; determining an interaction behavior type of the candidate video under a search request; and filtering a positive video sample and a negative video sample from the candidate video set based on the interaction behavior type.

[0006] Optionally, the filtering, from the candidate video set, the positive video samples and the negative video samples based on the interaction behavior type comprises: determining a candidate video that is selected to be played after being displayed in response to a search request as the video sample; and adding a positive sample label or a negative sample label to the video sample based on the interaction behavior type.

[0007] Optionally, the adding of the positive sample label or the negative sample label to the video sample based on the interaction behavior type comprises: determining that the interaction behavior type is a positive interaction behavior type, and adding a positive sample label to the corresponding video sample; or determining that the interaction behavior type is not the positive interaction behavior type, and adding a negative sample label to the corresponding video sample.

[0008] Optionally, the positive interaction behavior type comprises at least one of a play duration greater than a duration threshold, a like, and a follow.

[0009] Optionally, the content information comprises text information and image information, and the feature extraction network comprises a text feature extraction network and an image feature extraction network, wherein the inputting of the plurality of search terms and the content information into the feature extraction network to obtain the plurality of search term embedding features corresponding to the plurality of search terms and the content embedding feature corresponding to the content information comprises: inputting the plurality of search terms and the text information into the text feature extraction network to obtain the plurality of search term embedding features corresponding to the plurality of search terms and a text embedding feature corresponding to the text information; and inputting the image information into the image feature extraction network to obtain an image embedding feature corresponding to the image information.

[0010] Optionally, the determining of the N search term embedding features close to the content embedding feature from the plurality of search term embedding features comprises: determining at least one search term embedding feature close to the text embedding feature from the plurality of search term embedding features, and / or determining at least one search term embedding feature close to the image embedding feature from the plurality of search term embedding features to obtain the N search term embedding features.

[0011] Optionally, the determining of the N search term embedding features close to the content embedding feature from the plurality of search term embedding features comprises: determining N search term embedding features closest to the content embedding feature from the plurality of search term embedding features; or determining N search term embedding features with a distance less than a distance threshold to the content embedding feature from the plurality of search term embedding features.

[0012] Optionally, the image feature extraction network comprises cascaded bottom convolutional layers and semantic convolutional layers, and a fully connected layer connected to the semantic convolutional layers, and the adjusting parameters of the feature extraction network and the click rate estimation network according to the value of the loss function to obtain the trained search term recommendation model comprises: adjusting parameters of the semantic convolutional layers, the fully connected layer and the click rate estimation network in the image feature extraction network according to the value of the loss function, and keeping parameters of the bottom convolutional layers of the image feature extraction network and the text feature extraction network unchanged to obtain the trained search term recommendation model.

[0013] According to a second aspect of the present disclosure, a search term recommendation method is provided, comprising: extracting a plurality of search terms and content information corresponding to a reference video, and inputting the search term recommendation model into a feature extraction network to extract a plurality of search term embedding features corresponding to the plurality of search terms and a content embedding feature corresponding to the content information, wherein the search term recommendation model comprises the feature extraction network and a click rate estimation network; determining N search term embedding features close to the content embedding feature from the plurality of search term embedding features; inputting the N search term embedding features and the content embedding feature into the click rate estimation network to obtain an estimated click rate of each search term embedding feature; selecting M search term embedding features from the N search term embedding features according to the estimated click rate, and taking search terms corresponding to the M search term embedding features as recommended search terms corresponding to the reference video.

[0014] Optionally, the content information comprises text information and image information, and the feature extraction network comprises a text feature extraction network and an image feature extraction network, wherein inputting the plurality of search terms and the content information into the feature extraction network to extract the plurality of search term embedding features corresponding to the plurality of search terms and the content embedding feature corresponding to the content information comprises: inputting the plurality of search terms and the text information into the text feature extraction network to obtain the plurality of search term embedding features corresponding to the plurality of search terms and a text embedding feature corresponding to the text information; and inputting the image information into the image feature extraction network to obtain an image embedding feature corresponding to the image information.

[0015] Optionally, the determining N search term embedding features close to the content embedding feature from the plurality of search term embedding features comprises: determining at least one search term embedding feature close to the text embedding feature from the plurality of search term embedding features, and / or determining at least one search term embedding feature close to the image embedding feature from the plurality of search term embedding features to obtain the N search term embedding features.

[0016] Optionally, the determining, from the plurality of search term embedding features, N search term embedding features close to the content embedding feature comprises: determining, from the plurality of search term embedding features, N search term embedding features closest to the content embedding feature; or determining, from the plurality of search term embedding features, N search term embedding features with a distance less than a distance threshold to the content embedding feature.

[0017] Optionally, the selecting, from the N search term embedding features, M search term embedding features according to the estimated click-through rates comprises: selecting, from the N search term embedding features, M search term embedding features corresponding to the maximum estimated click-through rates; or selecting, from the N search term embedding features, M search term embedding features corresponding to the estimated click-through rates greater than a click-through rate threshold.

[0018] Optionally, the method further comprises: determining an interactive user having an interactive association relationship with the reference video; and sending the recommended search term to a terminal corresponding to the interactive user, so that the terminal displays the recommended search term.

[0019] Optionally, the search term recommendation model is obtained by training using the training method.

[0020] According to a third aspect of the present disclosure, a training apparatus of a search term recommendation model is provided, the search term recommendation model comprising a feature extraction network and a click-through rate estimation network, the training apparatus comprising: an acquisition unit configured to acquire video samples and sample labels thereof; an extraction unit configured to extract a plurality of search terms and content information corresponding to the video samples, and input the feature extraction network to extract a plurality of search term embedding features corresponding to the plurality of search terms and a content embedding feature corresponding to the content information; a recall unit configured to determine, from the plurality of search term embedding features, N search term embedding features close to the content embedding feature; an estimation unit configured to input the N search term embedding features and the content embedding feature into the click-through rate estimation network to obtain an estimated click-through rate of each search term embedding feature; a calculation unit configured to determine a value of a loss function of the search term recommendation model according to the estimated click-through rates and the sample labels; and a parameter adjustment unit configured to adjust parameters of the feature extraction network and the click-through rate estimation network according to the value of the loss function to obtain a trained search term recommendation model.

[0021] Optionally, the acquisition unit is further configured to acquire a candidate video set, the candidate video set comprising a plurality of candidate videos; determine an interactive behavior type of the candidate videos under a search request; and filter positive video samples and negative video samples from the candidate video set based on the interactive behavior type.

[0022] Optionally, the obtaining unit is further configured to determine a candidate video that is selected to be played after being displayed in response to a search request as the video sample; and add a positive sample label or a negative sample label to the video sample based on the interaction behavior type.

[0023] Optionally, the obtaining unit is further configured to determine that the interaction behavior type is a positive interaction behavior type, and add a positive sample label to the corresponding video sample; or determine that the interaction behavior type is not the positive interaction behavior type, and add a negative sample label to the corresponding video sample.

[0024] Optionally, the positive interaction behavior type includes at least one of a play duration greater than a duration threshold, a like, and a follow.

[0025] Optionally, the content information includes text information and image information, the feature extraction network includes a text feature extraction network and an image feature extraction network, and the extraction unit is further configured to input the plurality of search terms and the text information into the text feature extraction network to obtain a plurality of search term embedding features corresponding to the plurality of search terms and a text embedding feature corresponding to the text information; and input the image information into the image feature extraction network to obtain an image embedding feature corresponding to the image information.

[0026] Optionally, the recall unit is further configured to determine at least one search term embedding feature close to the text embedding feature from the plurality of search term embedding features, and / or determine at least one search term embedding feature close to the image embedding feature from the plurality of search term embedding features, to obtain the N search term embedding features.

[0027] Optionally, the recall unit is further configured to determine N search term embedding features closest to the content embedding feature from the plurality of search term embedding features; or determine N search term embedding features with a distance less than a distance threshold from the content embedding feature from the plurality of search term embedding features.

[0028] Optionally, the image feature extraction network includes cascaded bottom convolution layers and semantic convolution layers, and a fully connected layer connected to the semantic convolution layers, and the parameter adjustment unit is further configured to adjust parameters of the semantic convolution layers, the fully connected layer, and the click rate estimation network in the image feature extraction network according to the value of the loss function, and keep the bottom structure of the image feature extraction network and the parameters of the text feature extraction network unchanged, to obtain the trained search term recommendation model.

[0029] According to a fourth aspect of the present disclosure, a search term recommendation apparatus is provided, comprising: an extraction unit configured to extract a plurality of search terms corresponding to a reference video and content information, and input the plurality of search term embedding features corresponding to the plurality of search terms and the content embedding feature corresponding to the content information into a feature extraction network of a search term recommendation model, wherein the search term recommendation model comprises the feature extraction network and a click rate estimation network; a recall unit configured to determine N search term embedding features close to the content embedding feature from the plurality of search term embedding features; an estimation unit configured to input the N search term embedding features and the content embedding feature into the click rate estimation network to obtain an estimated click rate of each search term embedding feature; and a recommendation unit configured to select M search term embedding features from the N search term embedding features according to the estimated click rate, and take the search terms corresponding to the M search term embedding features as recommended search terms corresponding to the reference video.

[0030] Optionally, the content information comprises text information and image information, the feature extraction network comprises a text feature extraction network and an image feature extraction network, and the extraction unit is further configured to input the plurality of search terms and the text information into the text feature extraction network to obtain the plurality of search term embedding features corresponding to the plurality of search terms and a text embedding feature corresponding to the text information, and input the image information into the image feature extraction network to obtain an image embedding feature corresponding to the image information.

[0031] Optionally, the recall unit is further configured to determine at least one search term embedding feature close to the text embedding feature from the plurality of search term embedding features, and / or determine at least one search term embedding feature close to the image embedding feature from the plurality of search term embedding features, to obtain the N search term embedding features.

[0032] Optionally, the recall unit is further configured to determine N search term embedding features closest to the content embedding feature from the plurality of search term embedding features, or determine N search term embedding features with a distance less than a distance threshold to the content embedding feature from the plurality of search term embedding features.

[0033] Optionally, the recommendation unit is further configured to select M search term embedding features with the largest corresponding estimated click rate from the N search term embedding features, or select M search term embedding features with a corresponding estimated click rate greater than a click rate threshold from the N search term embedding features.

[0034] Optionally, the search term recommendation apparatus further comprises a determination unit configured to determine an interactive user having an interactive association relationship with the reference video; and a sending unit configured to send the recommended search term to a terminal corresponding to the interactive user, so that the terminal displays the recommended search term.

[0035] Optionally, the search term recommendation model is obtained by using the training method described above.

[0036] According to a fifth aspect of the present disclosure, an electronic device is provided, comprising at least one processor; at least one memory storing computer executable instructions, wherein the computer executable instructions, when executed by the at least one processor, cause the at least one processor to perform the training method of the search term recommendation model or the search term recommendation method according to the present disclosure.

[0037] According to a sixth aspect of the present disclosure, a computer readable storage medium is provided, when instructions in the computer readable storage medium are executed by at least one processor, the at least one processor is caused to perform the training method of the search term recommendation model or the search term recommendation method according to the present disclosure.

[0038] According to a seventh aspect of the present disclosure, a computer program product is provided, comprising computer instructions, when the computer instructions are executed by at least one processor, the training method of the search term recommendation model or the search term recommendation method according to the present disclosure is implemented.

[0039] The technical solutions provided by the embodiments of the present disclosure at least bring the following beneficial effects:

[0040] According to the training method of the search term recommendation model, the search term recommendation method and the apparatus according to the embodiments of the present disclosure, the search term recommendation model comprises a feature extraction network and a click rate estimation network. The feature extraction network is used to extract search term embedding features and content embedding features of a reference video, so as to facilitate semantic comparison between multiple search terms corresponding to the reference video and the content of the reference video, and further to recall search terms with similar semantics to the content of the reference video, so as to realize sufficient understanding and mining of the content of the reference video, and help to provide rich search terms. The click rate estimation network is used to process search term embedding features and content embedding features corresponding to the recalled search terms, and estimate the click rate of the recalled search terms, that is, estimate the degree of closeness of the search terms to user search intent, so as to facilitate sorting of the recalled search terms according to the click rate. Since the search term recommendation model is trained by using multiple video samples and sample labels reflecting the interaction of these video samples, the search intent of the user can be further mined based on semantic comparison based on embedding features and combined with the interaction of the video samples, which helps to recommend search terms that are more close to the user's intent.

[0041] It should be understood that the general description above and the detailed description below are only exemplary and explanatory, and are not limiting of the present disclosure. BRIEF DESCRIPTION OF DRAWINGS

[0042] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and serve to explain the principles of the present disclosure, and do not limit the present disclosure.

[0043] Figure 1 is a flowchart illustrating a training method of a search term recommendation model according to an exemplary embodiment of the present disclosure.

[0044] Figure 2 is a network structure diagram of a ResNet model in the related art.

[0045] Figure 3 is a flowchart illustrating a search term recommendation method according to an exemplary embodiment of the present disclosure.

[0046] Figure 4 is a block diagram of a training device of a search term recommendation model according to an exemplary embodiment of the present disclosure.

[0047] Figure 5 is a block diagram of a search term recommendation device according to an exemplary embodiment of the present disclosure.

[0048] Figure 6 is a block diagram of an electronic device according to an exemplary embodiment of the present disclosure. DETAILED DESCRIPTION

[0049] In order to make the ordinary person skilled in the art better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be described clearly and completely below in conjunction with the drawings.

[0050] It should be noted that the terms "first", "second", etc. in the specification and claims of the present disclosure and the above drawings are used to distinguish similar objects, and do not necessarily describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, 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 embodiments do not represent all embodiments consistent with the present disclosure. Rather, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0051] It should be noted that "at least one of a plurality" appearing in the present disclosure represents three types of parallel cases including "any one of the plurality", "a combination of any multiple of the plurality", and "all of the plurality". For example, "including at least one of A and B" includes the following three parallel cases: (1) including A; (2) including B; (3) including A and B. For another example, "performing at least one of step one and step two" represents the following three parallel cases: (1) performing step one; (2) performing step two; (3) performing step one and step two.

[0052] For a short video distribution platform, pushing search terms or content meeting the viewing interests of users to the users helps to enhance the stickiness of the platform and the users, and is crucial for the long-term development of the platform. Search discovery refers to recommending relevant search terms or content to users based on the historical interaction behaviors of the users, such as search history and viewing history. Taking recommending search terms based on the viewing history of a user as an example, a same video corresponds to multiple search terms, that is, when the user respectively inputs these search terms, the video can be searched, and thus the multiple search terms corresponding to the video viewed by the user can be recommended to the user. Although this method can expand search terms, the number of search terms corresponding to a video is usually large, and a part of the search terms needs to be selected for recommendation, and it is often difficult to ensure that the selected search terms are exactly the search terms that the user is interested in.

[0053] Therefore, an exemplary embodiment of the present disclosure proposes to establish a search term recommendation model, which can be based on a video (hereinafter referred to as a reference video) recently interacted by a user, fully understand and mine the content of the reference video, and further recommend rich and appropriate search terms. Specifically, the search term recommendation model includes a feature extraction network and a click rate estimation network. The feature extraction network is used to extract search term embedding features and content embedding features of the reference video. The embedding features are feature vectors capable of reflecting the semantics of corresponding elements, which facilitates semantic comparison between the multiple search terms corresponding to the reference video and the content of the reference video, and further recalls search terms with similar semantics to the content of the reference video, so as to fully understand and mine the content of the reference video and help to provide rich search terms. The click rate estimation network is used to process search term embedding features and content embedding features corresponding to the recalled search terms, and estimate the click rate of the recalled search terms, that is, estimate the degree of closeness of the search terms to the search intent of the user, so as to sort the recalled search terms according to the click rate. Since the search term recommendation model is trained by using multiple video samples and sample labels capable of reflecting the interaction of the video samples, the search intent of the user can be further mined based on semantic comparison based on embedding features and the interaction of the video samples, which helps to recommend search terms more close to the search intent of the user.

[0054] Next, exemplary embodiments of the present disclosure are described from the perspective of training and application of the search term recommendation model, respectively.

[0055] Figure 1 is a flowchart illustrating a training method of a search term recommendation model according to an exemplary embodiment of the present disclosure.

[0056] Referring to Figure 1 At step 101, a video sample and its sample label are obtained. The sample label, as a learning target of the search term recommendation model, can reflect whether the video sample meets the search intent, so that the search term recommendation model obtained through training has the possibility to predict that the search term corresponding to the reference video is clicked and used by the user (i.e., the click rate described below), thus helping to mine the search intent of the user, and thus obtaining a search term that is more in line with the user's intent. It can be understood that one video sample corresponds to multiple search terms, and at the same time, one search term also corresponds to multiple video samples, that is, the search term and the video sample are in a many-to-many relationship, so the final training result will be jointly affected by the multiple video samples used, and the learning target of the multiple search terms will not be the same because they correspond to the same video sample.

[0057] Step 101 can specifically include: obtaining a candidate video set, the candidate video set including multiple candidate videos that can be used as video samples; determining the interaction behavior type of the candidate videos under a search request to reflect whether the candidate videos meet the search intent; and based on the interaction behavior type, screening positive video samples and negative video samples from the candidate video set, that is, determining the sample label of the candidate video (i.e., the video sample) according to whether the candidate video meets the search intent. It should be understood that a video can be used as a candidate video as long as it has been displayed in response to a search request, and there is no need to particularly require the search term used in the search request, because the user's search has its corresponding search term, so the interaction behavior type of the candidate video under the search request is sufficient to reflect whether the candidate video meets the corresponding search intent, so that the model does not need to be trained separately for each user, which can improve the universal applicability of the trained model and improve the model training efficiency.

[0058] Optionally, when the positive video samples and the negative video samples are filtered from the candidate video set based on the interaction behavior type, the candidate video that is selected to be played after being displayed in response to the search request can be determined as the video sample, that is, the candidate video is filtered first; and then the positive sample label or the negative sample label is added to the video sample based on the interaction behavior type. Since the candidate video that has positive interaction (that is, the positive video sample) under the search request because it best meets the search intent is much less than the candidate video that has no positive interaction (that is, the negative video sample) because it does not meet the search intent or meets the search intent to a low degree, the click rate value output by the trained model is extremely small, and a sufficient number of decimal places need to be retained to enable comparison, which is not convenient for application and also increases the memory burden corresponding to the click rate and the number of samples. The candidate video that is not selected to be played has a high probability of not meeting the search intent, and the number of such candidate videos is much greater than that of the candidate videos that are selected to be played. Based on this, by filtering out the candidate videos that are not selected to be played first, the number of negative video samples can be greatly reduced, the proportion of positive video samples in the video samples can be increased, the training calculation amount can be reduced, the memory burden corresponding to the click rate can be reduced, and the training efficiency and the model application efficiency can be improved.

[0059] Optionally, when the positive sample label or the negative sample label is added to the video sample based on the interaction behavior type, the positive sample label can be added to the corresponding video sample when it is determined that the interaction behavior type is the positive interaction behavior type; and the negative sample label can be added to the corresponding video sample when it is determined that the interaction behavior type is not the positive interaction behavior type. That is, whether the video sample is a positive video sample or not can be determined according to whether the interaction behavior type is a positive interaction behavior type, and the objective and specific positive interaction behavior type that can be detected can be used to represent that the abstract user search intent is met. Specifically, the positive interaction behavior can include, but is not limited to, a playback duration greater than a duration threshold (hereinafter referred to as long playback), a like, and a follow, that is, as long as any one or more of the positive interaction behaviors such as long playback, like, and follow exist in a video sample, the video sample is marked as a positive video sample; and if there is no positive interaction behavior after being selected to be played, that is, short playback (the playback duration is less than or equal to the duration threshold) and no like and no follow, the video sample is marked as a negative video sample.

[0060] At step 102, a plurality of search terms and content information corresponding to the video sample are extracted and input into a feature extraction network to extract search term embedding features corresponding to the plurality of search terms and content embedding features corresponding to the content information. This step extracts embedding features for later use, which facilitates comparison of whether the search terms and the content of the video sample are semantically close. It can be understood that the search terms corresponding to a video sample refer to the current video sample that can be searched under the search terms, so a video sample can correspond to multiple search terms, and accordingly multiple search term embedding features can be obtained, while the content information corresponding to the video sample is unique, and accordingly one content embedding feature can be obtained. Here, unique refers to the case where multiple search terms exist in parallel, and once the content to be selected is determined (for example, the text information and image information described below are selected), there is no multiple content information in parallel for a video sample.

[0061] The content information is used to reflect the specific content of the video sample. In one example, the content information includes text information and image information. The text information can include hashtags (topic labels, referring to keywords marked with "#" by content producers when publishing content), labels, cover texts, titles, etc. of the video sample, and can also include text obtained by converting the audio of the video sample into speech text, and can also include other extractable text, which is not limited in the present disclosure. In addition, since users interact based on video images, the most intuitive understanding is video images, and there is a lot of information loss from video images to text. For example, for a video of "a dog happily running around a table", the meaning is relatively complex, and the related text information is often only entities such as dogs and tables, so image information is added here to supplement the lost information and facilitate a more comprehensive understanding of the video content. Of course, in other examples, the content information can include only text information or only image information.

[0062] Corresponding to the example where the content information includes text information and image information, the content embedding features include text embedding features and image embedding features, the feature extraction network includes a text feature extraction network for extracting search term embedding features corresponding to the search terms (the search terms are essentially text) and text embedding features corresponding to the text information, for example, a BERT (Bidirectional Encoder Representations from Transformers) network can be used, and the feature extraction network further includes an image feature extraction network for extracting image embedding features corresponding to the image information, for example, a CNN (Convolutional Neural Network) can be used, and the specific image embedding features extracted are before the classification layer of the CNN. Figure 2 is a network structure diagram of a ResNet (Residual Neural Network) model.Figure 2 The left side is a 19-layer VGG network (VGG-19, Visual Geometry Group, super-resolution test sequence), which includes 16 convolutional layers (conv, the first 3*3 represents the size of the convolution kernel, and the number after represents the number of channels of the convolutional layer) and 3 fully connected layers. The middle "pool, / 2" represents a halving pooling layer. The middle is a 34-layer plain network, and the right side is a 34-layer ResNet. Both of them include 33 convolutional layers and 1 fully connected layer. The ResNet network is modified on the basis of the VGG-19 network and adds a residual unit through a short circuit mechanism. The main changes are that the ResNet directly uses a convolution with a step of 2 for down-sampling, and replaces the fully connected layer fc4096 with a global pooling layer (avg pool). An important design principle of the ResNet is that when the size of the feature map is reduced by half, the number of feature maps is doubled, which maintains the complexity of the network layer. It can be seen from Figure 2 that the ResNet adds a short circuit mechanism between every two layers (the arc with an arrow on the right side), which forms residual learning. The dashed line indicates that the number of feature maps has changed. Referring to Figure 2 , the output of the last layer fc4096 or the last layer avgpool in the figure can be extracted when extracting image embedding features. Specifically, the feature extraction network can be a pre-trained network, and the CNN can be trained based on a public data set such as ImageNet.

[0063] Referring back to Figure 1 , in step 103, N search term embedding features close to the content embedding feature are determined from the plurality of search term embedding features. This step is a step of recalling search term embedding features. As described above, one content embedding feature and a plurality of search term embedding features can be extracted from one video sample, and the embedding feature is a feature vector that can reflect the semantics of the corresponding element. For a video sample, by respectively calculating the distances between the plurality of search term embedding features extracted in step 102 and the content embedding feature of the video sample, the semantic closeness of the corresponding search terms and the content information of the video sample can be understood, and the N search terms with closer distances are recalled as candidate search terms. N search terms close to the semantic content of the video sample can be obtained, so as to realize sufficient understanding and mining of the video content, and help to provide rich search terms.

[0064] Optionally, step 103 specifically comprises: determining N search term embedding features closest to the content embedding feature from the plurality of search term embedding features; or determining N search term embedding features with a distance less than a distance threshold to the content embedding feature from the plurality of search term embedding features. That is, when selecting N search term embedding features, the distance between the search term embedding features and the content embedding feature needs to be determined first, and N search term embedding features with the smallest distance can be selected according to the distance size. At this time, the distance between each search term embedding feature and the content embedding feature needs to be calculated, or N search term embedding features with a distance small enough (less than the distance threshold) can be selected according to the size relationship between the distance and the distance threshold, so as to realize flexibility in specific operation.

[0065] Optionally, as described above, the content embedding feature comprises a text embedding feature and an image embedding feature, and step 103 can specifically determine N search term embedding features close to at least one of the text embedding feature and the image embedding feature from the search term embedding features based on at least one of the text embedding feature and the image embedding feature. That is, the search term embedding features can be recalled only by using the text embedding feature (i.e., only the distance between the text embedding feature and the search term embedding feature is calculated), or the search term embedding features can be recalled only by using the image embedding feature (i.e., only the distance between the image embedding feature and the search term embedding feature is calculated), or the search term embedding features can be recalled by using both of the above embedding features (i.e., the distances between both the text embedding feature and the image embedding feature and the search term embedding feature are calculated). For the third scheme, it can be further subdivided into two schemes. One is to consider the distances between the text embedding feature and the image embedding feature and the search term embedding feature at the same time, for example, to calculate the statistical value of the two distances after calculating the two distances respectively, for example, to sum the two distances by weighting, specifically, to calculate the average distance, and then to determine N search term embedding features closer according to the statistical value, so as to take into account the semantics of the text information and the image information. The other is to recall a plurality of search term embedding features by using the text embedding feature and the image embedding feature respectively, and then to obtain N search term embedding features by combining these search term embedding features, which can appropriately take into account the semantics of the text information and the image information, and can also simplify the calculation strategy and reduce the calculation load.

[0066] At step 104, the N search term embedding features and the content embedding features are input into the click rate estimation network to obtain the estimated click rate of each search term embedding feature. The click rate refers to the ratio of the number of clicks to the number of displays of a certain content, which can reflect the degree of attention of a certain content. In terms of search term embedding features, it can reflect the probability that the search term corresponding to a search term embedding feature is used by a user who has interacted with the corresponding video sample. By inputting each search term embedding feature and the content embedding feature of the corresponding video sample into the click rate estimation network, the content information of the search term and the video sample can be considered at the same time, and then the estimated click rate is obtained. The positive video samples used for training are video samples with positive interactions, so that the trained search term recommendation model can fully learn the relationship between the search term, the content of the video sample, and the interaction habits of the user who has interacted with the video sample. Based on semantic comparison based on embedding features, the search intent of the user can be further mined, the click rate with higher accuracy can be estimated, and the search term that better meets the user's intent can be recommended.

[0067] Specifically, after the recall of the search term embedding features is completed at step 103, an inverted index can be constructed for the content embedding features, the search term embedding features, and the search terms in a manner similar to traditional retrieval methods, so that the search term embedding features and the content embedding features corresponding to each search term are determined to avoid input errors and ensure the reliability of the model.

[0068] Still taking an example in which the content embedding features include text embedding features and image embedding features, the three types of features input into the click rate estimation network are search term embedding features, text embedding features, and image embedding features. The click rate estimation network can use a three-tower network configured with three fully connected layers corresponding to the above three types of features to respectively receive the input of the above three types of features. The embedding features obtained from the three fully connected layers can be spliced, and the spliced features are then input into a fully connected layer, and then the estimated click rate is obtained through softmax.

[0069] At step 105, the value of the loss function of the search term recommendation model is determined according to the estimated click rate and the sample label. The search term recommendation model is essentially a model that first extracts search term embedding features and content embedding features, then recalls search term embedding features according to content embedding features, and finally estimates whether the search term corresponding to each recalled search term embedding feature will be clicked by a user (the model outputs the estimated click rate), that is, determines whether to recommend the search term corresponding to the recalled search term embedding feature to the user. Therefore, it belongs to a classification problem, and the cross-entropy loss function can be used as the loss function.

[0070] At step 106, the parameters of the feature extraction network and the click rate estimation network are adjusted according to the value of the loss function to obtain the trained search term recommendation model. Specifically, the model can be learned by using SGD (Stochastic Gradient Descent) back propagation gradient.

[0071] Optionally, still taking an example in which the content embedding features include text embedding features and image embedding features, the feature extraction network includes a text feature extraction network and an image feature extraction network, the image feature extraction network can include cascaded bottom convolutional layers and semantic convolutional layers, and a fully connected layer connected to the semantic convolutional layers, and step 106 specifically includes: adjusting the parameters of the semantic convolutional layers, the fully connected layer in the image feature extraction network, and the click rate estimation network according to the value of the loss function, and keeping the parameters of the bottom convolutional layers of the image feature extraction network and the text feature extraction network unchanged, to obtain the trained search term recommendation model. That is, when training the model, for the pre-trained image feature extraction network, the structure of the bottom convolutional layers can be kept unchanged and only the high-level semantic convolutional layers and the fully connected layer are updated, which helps to reduce the training load. For the text feature extraction network, it can be pre-trained and not updated when training the model.

[0072] Figure 3 FIG. 1 is a flowchart illustrating a search term recommendation method according to an example embodiment of the present disclosure.

[0073] Referring to Figure 3 At step 301, a plurality of search terms and content information corresponding to a reference video are extracted and input into the feature extraction network of the search term recommendation model to extract search term embedding features corresponding to the plurality of search terms and content embedding features corresponding to the content information. The search term recommendation model includes a feature extraction network and a click rate estimation network. The reference video can be a video that the current user has interacted with actively most recently. This step extracts embedding features in preparation, which can facilitate subsequent comparison of whether the content of the search term is semantically close to the reference video.

[0074] The content information is used to reflect the specific content of the reference video. In an example, the content information optionally includes text information and image information. The text information can include hashtags, tags, cover texts, titles, etc. of the reference video, and can also include text obtained by performing speech-to-text conversion on the audio of the reference video, and can also include other text that can be extracted, which is not limited in the present disclosure. In addition, since the user interacts based on the video image, it is most intuitive to understand the video image, and there is a lot of information loss from the video image to the text. For example, for a video of "a dog happily running around the table", the meaning is relatively complex, and the related text information is often only the entity of dog and table, so the image information is added here to supplement the lost information and facilitate a more comprehensive understanding of the video content. Of course, in other examples, the content information can only include text information or only include image information.

[0075] Corresponding to the example in which the content information includes text information and image information, the content embedding feature includes a text embedding feature and an image embedding feature, and the feature extraction network includes a text feature extraction network for extracting a search word embedding feature corresponding to a search word (the search word is essentially text) and a text embedding feature corresponding to the text information, and an image feature extraction network for extracting an image embedding feature corresponding to the image information.

[0076] Optionally, the search word recommendation model according to the example embodiments of the present disclosure can be trained using the training method described above, thus having all the beneficial technical effects of the training method described above, which will not be repeated here.

[0077] In step 302, N search word embedding features close to the content embedding feature are determined from the plurality of search word embedding features. This step is a step of recalling search word embedding features. One content embedding feature and a plurality of search word embedding features can be extracted from one reference video, and the embedding feature is a feature vector that can reflect the semantics of the corresponding element. For a reference video, by calculating the distance between the plurality of search word embedding features extracted in step 301 and the content embedding feature of the reference video, the semantic closeness of the corresponding search words to the content information of the reference video can be understood, and the N search words with closer distance are recalled as candidate search words, which can obtain N search words close to the content semantics of the reference video, thereby achieving a full understanding and mining of the video content, and helping to provide rich search words.

[0078] Optionally, step 302 specifically comprises: determining N search term embedding features closest to the content embedding feature from the plurality of search term embedding features; or determining N search term embedding features with a distance less than a distance threshold to the content embedding feature from the plurality of search term embedding features. That is, when selecting N search term embedding features, the distance between the search term embedding features and the content embedding feature needs to be determined first, and N search term embedding features with the smallest distance can be selected according to the distance size. At this time, the distance between each search term embedding feature and the content embedding feature needs to be calculated, or N search term embedding features with a distance small enough (less than the distance threshold) can be selected according to the size relationship between the distance and the distance threshold, so as to realize flexibility in specific operation.

[0079] Optionally, as described above, the content embedding feature comprises a text embedding feature and an image embedding feature, and step 302 can specifically determine N search term embedding features close to at least one of the text embedding feature and the image embedding feature from the search term embedding features based on at least one of the text embedding feature and the image embedding feature. That is, the search term embedding features can be recalled only by using the text embedding feature (i.e., only the distance between the text embedding feature and the search term embedding feature is calculated), or the search term embedding features can be recalled only by using the image embedding feature (i.e., only the distance between the image embedding feature and the search term embedding feature is calculated), or the search term embedding features can be recalled by using both of the above embedding features (i.e., the distances between both the text embedding feature and the image embedding feature and the search term embedding feature are calculated). For the third scheme, it can be further subdivided into two schemes. One is to consider the distances between the text embedding feature and the image embedding feature and the search term embedding feature at the same time, for example, to calculate the statistical value of the two distances after calculating the two distances respectively, for example, to sum the two distances by weighting, specifically, to calculate the average distance, and then to determine N search term embedding features closer according to the statistical value, so as to take into account the semantics of the text information and the image information. The other is to recall a plurality of search term embedding features by using the text embedding feature and the image embedding feature respectively, and then to obtain N search term embedding features by combining these search term embedding features, which can appropriately take into account the semantics of the text information and the image information, and can also simplify the calculation strategy and reduce the calculation load.

[0080] At step 303, the N search term embedding features and the content embedding features are input into the click rate estimation network to obtain an estimated click rate of each search term embedding feature. By inputting each search term embedding feature and the content embedding feature of the reference video into the click rate estimation network, the content information of the search term and the reference video can be considered simultaneously, and then the estimated click rate is obtained. The positive video samples used for training of the search term recommendation model are video samples with which the user has positive interaction, so that the trained search term recommendation model can sufficiently learn the relationship among the search term, the content of the reference video, and the interaction habit of the user who has interacted with the reference video, can further mine the search intention of the user on the basis of semantic comparison based on embedding features, estimate a click rate with higher accuracy, and help recommend search terms that are more in line with the user's intention.

[0081] Specifically, after the search term recall is completed at step 302, an inverted index can be constructed for the content embedding features, the search term embedding features, and the search terms in a manner similar to traditional retrieval methods, so that the search term embedding features and the content embedding features corresponding to each search term are determined, to avoid input errors and ensure model reliability.

[0082] Still taking an example in which the content embedding features include text embedding features and image embedding features, the search term embedding features, the text embedding features, and the image embedding features are specifically input into the click rate estimation network. The click rate estimation network can adopt a three-tower network and be configured with full connection layers corresponding to the above three types of features to respectively receive the input of the above three types of features. The embedding features obtained from the three full connection layers can be spliced, and the spliced features are then input into a full connection layer, and then the estimated click rate is obtained through softmax.

[0083] At step 304, M search term embedding features are selected from the N search term embedding features according to the estimated click rate, and the search terms corresponding to the M search term embedding features are taken as the recommended search terms corresponding to the reference video. Wherein, M≤N. By selecting the M search term embedding features in combination with the estimated click rate, and correspondingly obtaining the M recommended search terms, the recommendation range can be further narrowed from the N search term embedding features recalled, which helps to recommend search terms that are rich and in line with the user's intention.

[0084] Optionally, the search term recommendation method according to an example embodiment of the present disclosure further includes: determining an interactive user having an interactive association relationship with the reference video; and sending the recommended search terms to a terminal corresponding to the interactive user, so that the terminal displays the recommended search terms. By sending the recommended search terms obtained at step 304 to the terminal corresponding to the interactive user, the interactive user can use the recommended search terms to initiate a search, and then search for a video that is more in line with his search intention.

[0085] Figure 4is a block diagram illustrating a training apparatus of a search term recommendation model according to an example embodiment of the present disclosure.

[0086] Referring to Figure 4 The training apparatus 400 of the search term recommendation model includes an acquisition unit 401, an extraction unit 402, a recall unit 403, an estimation unit 404, a calculation unit 405, and a parameter tuning unit 406.

[0087] The acquisition unit 401 can acquire video samples and their sample labels. The sample labels, as learning objectives of the search term recommendation model, can reflect whether the video samples meet the search intent, so that the trained search term recommendation model has the ability to predict the possibility (i.e., click rate) of the search term corresponding to the reference video being clicked and used by the user, thus helping to mine the user's search intent and thus obtaining search terms that are more in line with the user's intent. It can be understood that one video sample corresponds to multiple search terms, and at the same time, one search term also corresponds to multiple video samples, that is, the search term and the video sample are in a many-to-many relationship, so the final training result will be jointly affected by the multiple video samples used, and the learning objectives of the multiple search terms corresponding to the same video sample will not be the same.

[0088] The acquisition unit 401 can specifically perform the following operations: acquire a candidate video set, the candidate video set including multiple candidate videos that can be used as video samples; determine the interaction behavior type of the candidate videos under a search request to reflect whether the candidate videos meet the search intent; and filter out positive video samples and negative video samples from the candidate video set based on the interaction behavior type, that is, determine the sample labels of the candidate videos (i.e., video samples) according to whether the candidate videos meet the search intent. It should be understood that a video can be used as a candidate video as long as it has been displayed in response to a search request, and there is no special requirement for the search term used in the search request. Because the user's search has its corresponding search term, the interaction behavior type of the candidate video under the search request is sufficient to reflect whether the candidate video meets the corresponding search intent, so it is not necessary to train the model for each user individually, which can improve the universal applicability of the trained model and improve the model training efficiency.

[0089] Optionally, when the acquisition unit 401 filters out the positive video samples and the negative video samples from the candidate video set based on the interaction behavior type, the candidate video that is selected to be played after being displayed in response to the search request can be determined as the video sample, that is, the candidate video is filtered first; and then the positive sample label or the negative sample label is added to the video sample based on the interaction behavior type. Since the candidate video that has positive interaction (that is, the positive video sample) under the search request because it best meets the search intent is much less than the candidate video that has no positive interaction (that is, the negative video sample) because it does not meet the search intent or meets the search intent to a low degree, the click rate value output by the trained model is extremely small, and a sufficient number of decimal places need to be retained to realize comparison, which is not convenient for application, and also increases the memory burden corresponding to the click rate and increases the sample quantity. The candidate video that is not selected to be played has a high probability of not meeting the search intent, and the quantity of the candidate video that is not selected to be played is much greater than that of the candidate video that is selected to be played. Based on this, by filtering out the candidate video that is not selected to be played first, the quantity of the negative video sample can be greatly reduced, the proportion of the positive video sample in the video sample is increased, which helps to reduce the training calculation amount, reduce the memory burden corresponding to the click rate, improve the training efficiency and the model application efficiency.

[0090] Optionally, when the acquisition unit 401 adds the positive sample label or the negative sample label to the video sample based on the interaction behavior type, the positive sample label can be added to the corresponding video sample when it is determined that the interaction behavior type is the positive interaction behavior type; and the negative sample label can be added to the corresponding video sample when it is determined that the interaction behavior type is not the positive interaction behavior type. That is, whether the video sample is the positive video sample is determined according to whether the interaction behavior type is the positive interaction behavior type, and the objective and specific positive interaction behavior type that can be detected can be used to represent that the abstract user search intent is met. Specifically, the positive interaction behavior can include, but is not limited to, a playing time longer than a time threshold (hereinafter referred to as long playing), liking, and following, that is, as long as there is any one or more of the positive interaction behaviors such as long playing, liking, and following in a video sample, the video sample is marked as the positive video sample; and if there is no positive interaction behavior after being selected to be played, that is, short playing (the playing time is less than or equal to the time threshold) and no liking and no following, the video sample is marked as the negative video sample.

[0091] The extraction unit 402 can extract a plurality of search words and content information corresponding to the video sample, and input the feature extraction network to extract search word embedding features corresponding to the plurality of search words and content embedding features corresponding to the content information. The extraction unit 402 extracts embedding features for standby, which can facilitate subsequent comparison of whether the search words and the content of the video sample are semantically close. It can be understood that the search words corresponding to a video sample refer to the current video sample that can be searched under the search words, so a video sample can correspond to multiple search words, and accordingly multiple search word embedding features can be obtained, while the content information corresponding to the video sample is unique, and accordingly one content embedding feature can be obtained. Here, unique refers to the case where multiple search words exist in parallel, and once the content to be selected is determined (for example, the text information and image information described below are selected), there is no multiple parallel content information for a video sample.

[0092] The content information is used to reflect the specific content of the video sample. In one example, the content information includes text information and image information. The text information can include hashtags, tags, cover texts, titles, etc. of the video sample, and can also include text converted from audio of the video sample by speech-to-text conversion, and can also include other extractable text, which is not limited in the present disclosure. In addition, since the user interacts based on the video image, the most intuitive understanding is the video image, and there is a lot of information loss from the video image to the text, so the image information is added here to supplement the lost information, so as to facilitate more fully understanding the video content. Of course, in other examples, the content information can only include text information or only include image information.

[0093] Corresponding to the example that the content information includes text information and image information, the content embedding features include text embedding features and image embedding features, the feature extraction network includes a text feature extraction network for extracting search word embedding features corresponding to the search words (the search words are essentially text) and text embedding features corresponding to the text information, for example, a BERT network can be used, and the feature extraction network further includes an image feature extraction network for extracting image embedding features corresponding to the image information, for example, a CNN can be used, and the specific image embedding features extracted are before the classification layer of the CNN. Figure 2 is a network structure diagram of the ResNet model in the related art, referring to Figure 2 When extracting the image embedding features, the output of the last layer fc4096 or the last layer avgpool in the figure can be extracted. Specifically, the feature extraction network can be a pre-trained network, wherein the CNN can be trained based on a public data set such as ImageNet.

[0094] Referring back to Figure 4The recall unit 403 can determine N search term embedding features that are close to the content embedding feature from the plurality of search term embedding features. As described above, one content embedding feature and a plurality of search term embedding features can be extracted from one video sample, and the embedding features are feature vectors that can reflect the semantics of the corresponding elements. For one video sample, by respectively calculating the distances between the plurality of search term embedding features extracted by the extraction unit 402 and the content embedding feature of the video sample, the semantic closeness of the corresponding search terms and the content information of the video sample can be understood, and N search terms with closer distances are recalled as candidate search terms, and N search terms that are closer to the semantics of the content of the video sample can be obtained, thereby achieving sufficient understanding and mining of the video content, and helping to provide rich search terms.

[0095] Optionally, the recall unit 403 can specifically determine N search term embedding features that are closest to the content embedding feature from the plurality of search term embedding features; or can determine N search term embedding features whose distances to the content embedding feature are less than a distance threshold from the plurality of search term embedding features. That is, when selecting N search term embedding features, the distance between the search term embedding feature and the content embedding feature needs to be determined first, and N search term embedding features with the smallest distance can be selected according to the distance size, in which case the distance between each search term embedding feature and the content embedding feature needs to be calculated, or N search term embedding features with a distance small enough (less than the distance threshold) can be selected according to the size relationship between the distance and the distance threshold, to achieve flexibility in specific operations.

[0096] Optionally, as described above, the content embedding features include text embedding features and image embedding features, and the recall unit 403 can specifically determine N search term embedding features close to at least one of the text embedding features and the image embedding features from the search term embedding features based on the at least one of the text embedding features and the image embedding features. That is, the search term embedding features can be recalled only by using the text embedding features (i.e., only the distance between the text embedding features and the search term embedding features is calculated), the search term embedding features can be recalled only by using the image embedding features (i.e., only the distance between the image embedding features and the search term embedding features is calculated), and the search term embedding features can be recalled by using both the above two embedding features (i.e., the distances between both the text embedding features and the image embedding features and the search term embedding features are calculated). For the third scheme, it can be further subdivided into two schemes. One is to consider the distances between the text embedding features and the image embedding features and the search term embedding features at the same time, for example, the statistical value of the two distances is calculated after the two distances are calculated, for example, the two distances are weighted and summed, specifically, the average of the distances is calculated, and then the N search term embedding features close to the statistical value are determined, so as to take into account the semantics of the text information and the image information. The other is to recall a plurality of search term embedding features by using the text embedding features and the image embedding features respectively, and then the N search term embedding features are obtained by combining the search term embedding features, which can appropriately take into account the semantics of the text information and the image information, and can simplify the calculation strategy and reduce the calculation load.

[0097] The estimation unit 404 can input the N search term embedding features and the content embedding features into the click rate estimation network to obtain the estimated click rate of each search term embedding feature. By inputting the search term embedding features and the content embedding features of the corresponding video sample into the click rate estimation network, the content information of the search term and the video sample can be considered at the same time, and then the estimated click rate is obtained. The positive video sample used for training is the video sample with which the user has positive interaction, so that the trained search term recommendation model can fully learn the relationship among the search term, the content of the video sample, and the interaction habit of the user who has interacted with the video sample, can further mine the search intention of the user on the basis of semantic comparison based on embedding features, can estimate the click rate with higher accuracy, and is helpful for recommending the search term that is more in line with the user's intention.

[0098] Specifically, after the recall unit 403 completes the recall of the search term embedding features, an inverted index can be constructed for the content embedding features, the search term embedding features, and the search term in the same way as the traditional retrieval method, so as to clearly determine the search term embedding features and the content embedding features corresponding to each search term, so as to avoid input errors and ensure the reliability of the model.

[0099] Still taking an example in which the content embedding features include text embedding features and image embedding features, the input of the click rate estimation network is specifically three types of features, i.e., search word embedding features, text embedding features, and image embedding features. The click rate estimation network can adopt a three-tower network, which is configured with full connection layers respectively corresponding to the above three types of features, to respectively receive the input of the above three types of features. The embedding features obtained from the three full connection layers can be spliced, and the spliced features are then input into a full connection layer, and then the estimated click rate is obtained through softmax.

[0100] The computing unit 405 can determine the value of the loss function of the search word recommendation model according to the estimated click rate and the sample label. The search word recommendation model is essentially to extract search word embedding features and content embedding features, recall search word embedding features according to the content embedding features, and finally estimate whether each search word corresponding to the recalled search word embedding feature will be clicked by a user (the model outputs in the form of an estimated click rate), that is, determine whether to recommend the search word corresponding to the recalled search word embedding feature to the user, so it belongs to a classification problem, and the loss function thereof can use a cross-entropy loss function.

[0101] The parameter adjusting unit 406 can adjust the parameters of the feature extraction network and the click rate estimation network according to the value of the loss function to obtain the trained search word recommendation model. Specifically, the model can be learned by using SGD back propagation gradient.

[0102] Optionally, still taking an example in which the content embedding features include text embedding features and image embedding features, the feature extraction network includes a text feature extraction network and an image feature extraction network, the image feature extraction network can include cascaded bottom convolution layers and semantic convolution layers, and a full connection layer connected with the semantic convolution layers, and the parameter adjusting unit 406 can specifically adjust the parameters of the semantic convolution layers, the full connection layer in the image feature extraction network, and the click rate estimation network according to the value of the loss function, and keep the parameters of the bottom convolution layers in the image feature extraction network and the text feature extraction network unchanged, to obtain the trained search word recommendation model. That is, during the training of the model, for the pre-trained image feature extraction network, the structure of the bottom convolution layers can be kept from being updated, and only the high-level semantic convolution layers and the full connection layer are updated, which is helpful to reduce the training load. For the text feature extraction network, it can be pre-trained and not updated during the training of the model.

[0103] Figure 5 FIG. 1 is a block diagram illustrating a search word recommendation device according to an example embodiment of the present disclosure.

[0104] Referring to Figure 5 , the search word recommendation device 500 includes an extraction unit 501, a recall unit 502, an estimation unit 503, and a recommendation unit 504.

[0105] The extraction unit 501 can extract a plurality of search words and content information corresponding to a reference video, and input the search word recommendation model into a feature extraction network to extract search word embedding features corresponding to the plurality of search words and content embedding features corresponding to the content information, wherein the search word recommendation model comprises a feature extraction network and a click rate estimation network. The reference video can be a video that the current user has interacted with most recently. This step extracts embedding features for later use, which facilitates subsequent comparison of whether the search words and the content of the reference video are semantically close.

[0106] The content information is used to reflect the specific content of the reference video. In an example, the content information includes text information and image information. The text information can include hashtags, tags, cover texts, titles, etc. of the reference video, and can also include text converted from audio of the reference video, and can also include other extractable text, which is not limited in the present disclosure. In addition, since the user interacts based on the video image, the most intuitive understanding is the video image, and there is a lot of information loss from the video image to the text, so the image information is added here to supplement the lost information, so as to facilitate a more comprehensive understanding of the video content. Of course, in other examples, the content information can only include text information or only include image information.

[0107] Corresponding to the example in which the content information includes text information and image information, the content embedding features include text embedding features and image embedding features, and the feature extraction network includes a text feature extraction network for extracting search word embedding features corresponding to search words (search words are essentially text) and text embedding features corresponding to text information, and an image feature extraction network for extracting image embedding features corresponding to image information.

[0108] Optionally, the search word recommendation model according to the example embodiment of the present disclosure can be trained using the training method described above, thus having all the beneficial technical effects of the training method, which will not be repeated here.

[0109] The recall unit 502 can determine N search word embedding features close to the content embedding features from the plurality of search word embedding features. One content embedding feature and a plurality of search word embedding features can be extracted from one reference video, and the embedding features are feature vectors that can reflect the semantics of the corresponding elements. For one reference video, by calculating the distance between the plurality of search word embedding features extracted by the extraction unit 501 and the content embedding features of the reference video, the semantic closeness of the corresponding search words and the content information of the reference video can be understood, and the N search words with closer distance are recalled as candidate search words, which can obtain N search words with closer semantic content to the reference video, thereby achieving a full understanding and mining of the video content, and helping to provide rich search words.

[0110] Optionally, the recalling unit 502 can specifically determine N search term embedding features closest to the content embedding feature from the plurality of search term embedding features, or can determine N search term embedding features with a distance less than a distance threshold from the content embedding feature from the plurality of search term embedding features. That is, when selecting N search term embedding features, the distance between the search term embedding feature and the content embedding feature needs to be determined first, and N search term embedding features with the smallest distance can be selected according to the distance size. At this time, the distance between each search term embedding feature and the content embedding feature needs to be calculated, or N search term embedding features with a distance small enough (less than the distance threshold) can be selected according to the size relationship between the distance and the distance threshold, to realize flexibility in specific operations.

[0111] Optionally, as described above, the content embedding feature includes a text embedding feature and an image embedding feature, and the recalling unit 502 can specifically determine N search term embedding features close to at least one of the text embedding feature and the image embedding feature from the search term embedding features based on the at least one of the text embedding feature and the image embedding feature. That is, the search term embedding feature can be recalled only by using the text embedding feature (i.e., only the distance between the text embedding feature and the search term embedding feature is calculated), the search term embedding feature can be recalled only by using the image embedding feature (i.e., only the distance between the image embedding feature and the search term embedding feature is calculated), or the search term embedding feature can be recalled by using both of the above embedding features (i.e., the distances between both the text embedding feature and the image embedding feature and the search term embedding feature are calculated). For the third scheme, it can be further subdivided into two schemes. One is to consider the distances between the text embedding feature and the image embedding feature and the search term embedding feature at the same time, for example, to calculate two distances respectively and then to obtain a statistical value of the two distances, for example, to perform weighted summation on the two distances, specifically, to obtain the average distance, and then to determine N search term embedding features closer according to the statistical value, so as to take into account the semantics of the text information and the image information. The other is to respectively recall a plurality of search term embedding features by using the text embedding feature and the image embedding feature respectively, and then to obtain N search term embedding features by aggregating the search term embedding features, which can appropriately take into account the semantics of the text information and the image information, and can also simplify the calculation strategy and reduce the calculation load.

[0112] The estimation unit 503 can input the N search word embedding features and the content embedding features into a click rate estimation network to obtain an estimated click rate of each search word embedding feature. By inputting the search word embedding features and the content embedding features of the reference video into the click rate estimation network, the content information of the search word and the reference video can be considered at the same time, and then the estimated click rate is obtained. The positive video samples used for training the search word recommendation model are video samples with positive interaction of the user, so that the trained search word recommendation model can fully learn the relationship among the search word, the content of the reference video, and the interaction habit of the user who has interacted with the reference video, can further mine the search intention of the user on the basis of semantic comparison based on embedding features, estimate a click rate with higher accuracy, and help recommend search words that are more in line with the user's intention.

[0113] Specifically, after the recall unit 502 completes the recall of the search word embedding features, an inverted index can be constructed for the content embedding features, the search word embedding features, and the search words in a manner similar to traditional retrieval methods, so as to clearly determine the search word embedding features and the content embedding features corresponding to each search word, so as to avoid input errors and ensure model reliability.

[0114] Still taking an example in which the content embedding features include text embedding features and image embedding features, the input of the click rate estimation network is specifically the search word embedding features, the text embedding features, and the image embedding features. The click rate estimation network can adopt a three-tower network and be configured with full connection layers corresponding to the above three types of features to respectively receive the input of the above three types of features. The embedding features obtained from the three full connection layers can be spliced, and the spliced features are then input into a full connection layer, and then the estimated click rate is obtained through softmax.

[0115] The recommendation unit 504 can select M search word embedding features from the N search word embedding features according to the estimated click rate, and take the search words corresponding to the M search word embedding features as the recommended search words corresponding to the reference video. Wherein, M≤N. By selecting the M search word embedding features in combination with the estimated click rate, and correspondingly obtaining M recommended search words, the recommendation range can be further narrowed from the N search word embedding features recalled, which helps to recommend search words that are rich and in line with the user's intention.

[0116] Optionally, the search word recommendation apparatus according to an example embodiment of the present disclosure can further include a determination unit and a sending unit. The determination unit can determine an interactive user having an interactive association relationship with the reference video. The sending unit can send the recommended search words to a terminal corresponding to the interactive user, so that the terminal displays the recommended search words. By sending the recommended search words obtained by the recommendation unit 504 to the terminal corresponding to the interactive user, the interactive user can use the recommended search words to initiate a search, and then search for a video that is more in line with the search intention.

[0117] Figure 6 is a block diagram of an electronic device according to an exemplary embodiment of the present disclosure.

[0118] Referring to Figure 6 , the electronic device 600 includes at least one memory 601 having a set of computer executable instructions stored therein and at least one processor 602, which when executing the set of computer executable instructions, performs a method of training a search term recommendation model or a method of recommending a search term according to an exemplary embodiment of the present disclosure.

[0119] As an example, the electronic device 600 can be a PC computer, a tablet device, a personal digital assistant, a smart phone, or other devices capable of executing the above-mentioned set of instructions. Here, the electronic device 600 is not necessarily a single electronic device, but can be a collection of any devices or circuits capable of executing the above-mentioned instructions (or set of instructions) individually or jointly. The electronic device 600 can also be a part of an integrated control system or a system manager, or can be configured as a portable electronic device that interfaces with a local or remote (e.g., via wireless transmission).

[0120] In the electronic device 600, the processor 602 can include a central processor (CPU), a graphics processor (GPU), a programmable logic device, a dedicated processor system, a microcontroller, or a microprocessor. As an example and not a limitation, the processor can also include an analog processor, a digital processor, a microprocessor, a multi-core processor, a processor array, a network processor, etc.

[0121] The processor 602 can execute instructions or codes stored in the memory 601, where the memory 601 can also store data. The instructions and data can also be sent and received over a network via a network interface device, which can employ any known transmission protocol.

[0122] The memory 601 can be integrated with the processor 602, for example, arranging RAM or flash memory within an integrated circuit microprocessor, etc. In addition, the memory 601 can include a separate device, such as an external disk drive, a storage array, or other storage devices usable by any database system. The memory 601 and the processor 602 can be operatively coupled or can communicate with each other, for example, through an I / O port, a network connection, etc., so that the processor 602 can read files stored in the memory.

[0123] In addition, the electronic device 600 can also include a video display (such as a liquid crystal display) and a user interaction interface (such as a keyboard, a mouse, a touch input device, etc.). All components of the electronic device 600 can be connected to each other via a bus and / or a network.

[0124] According to an exemplary embodiment of the disclosure, there can also be provided a computer-readable storage medium that, when instructions in the computer-readable storage medium are executed by at least one processor, causes the at least one processor to perform the training method of a search term recommendation model or the search term recommendation method according to an exemplary embodiment of the disclosure. Examples of the computer-readable storage medium here include read-only memory (ROM), random-access programmable read-only memory (PROM), electrically erasable programmable read-only memory (EEPROM), random-access memory (RAM), dynamic random-access memory (DRAM), static random-access memory (SRAM), flash memory, non-volatile memory, CD-ROM, CD-R, CD+R, CD-RW, CD+RW, DVD-ROM, DVD-R, DVD+R, DVD-RW, DVD+RW, DVD-RAM, BD-ROM, BD-R, BD-R LTH, BD-RE, Blu-ray or optical disc memory, a hard disk drive (HDD), a solid state drive (SSD), a card-type memory such as a multimedia card, a secure digital (SD) card, or an extreme digital (XD) card, a magnetic tape, a floppy disk, a magneto-optical data storage device, an optical data storage device, a hard disk, a solid state disk, and any other device configured to store a computer program and any associated data, data files, and data structures in a non-transitory manner and provide the computer program and any associated data, data files, and data structures to a processor or computer so that the processor or computer can execute the computer program. The computer program in the above-described computer-readable storage medium can be executed in an environment deployed in a computer device such as a client, a host, an agent device, a server, etc., and, in addition, in one example, the computer program and any associated data, data files, and data structures are distributed over a networked computer system so that the computer program and any associated data, data files, and data structures are stored, accessed, and executed in a distributed manner by one or more processors or computers.

[0125] According to an exemplary embodiment of the disclosure, there can also be provided a computer program product including computer instructions that, when executed by at least one processor, cause the at least one processor to perform the training method of a search term recommendation model or the search term recommendation method according to an exemplary embodiment of the disclosure.

[0126] According to the training method, the search term recommendation method and the device of the search term recommendation model according to the exemplary embodiments of the present disclosure, the search term recommendation model comprises a feature extraction network and a click rate estimation network. The feature extraction network is used to extract search term embedding features and content embedding features of a reference video, so as to facilitate semantic comparison between a plurality of search terms corresponding to the reference video and the content of the reference video, and then recall search terms with similar content semantics to the reference video, so as to realize sufficient understanding and mining of the content of the reference video, and help to provide rich search terms. The click rate estimation network is used to process search term embedding features and content embedding features corresponding to the recalled search terms, and estimate the click rate of the recalled search terms, that is, estimate the degree of closeness of the search terms to the user search intent, so as to facilitate sorting of the recalled search terms according to the click rate. Since the search term recommendation model is trained by using the historical interaction behavior data of the current user, the search intent of the user can be further mined on the basis of semantic comparison based on embedding features, and it is helpful to recommend search terms that are more close to the user intent.

[0127] Other embodiments of the present disclosure will be apparent to those skilled in the art from consideration of the specification and practice of the features disclosed herein. This application is intended to cover any variations, uses or adaptations of the present disclosure following, in general, the principles of the present disclosure and including such departures from the present disclosure that come within known

[0128] It should be understood that the present disclosure is not limited to the precise structures as herein described and illustrated in the drawings, and that various modifications and changes can be made without departing from the scope thereof. The scope of the present disclosure is limited only by the claims that follow.

Claims

1. A training method for a search term recommendation model, characterized in that, The search term recommendation model comprises a feature extraction network and a click rate estimation network, and the training method comprises: obtaining video samples and sample labels thereof; extracting a plurality of search terms and content information corresponding to the video samples, and inputting the feature extraction network to extract a plurality of search term embedding features corresponding to the plurality of search terms and content embedding features corresponding to the content information; determining N search term embedding features close to the content embedding features from the plurality of search term embedding features; inputting the N search term embedding features and the content embedding features into the click rate estimation network to obtain an estimated click rate of each search term embedding feature; determining a value of a loss function of the search term recommendation model according to the estimated click rate and the sample labels; adjusting parameters of the feature extraction network and the click rate estimation network according to the value of the loss function to obtain a trained search term recommendation model; wherein the obtaining of the video samples and the sample labels thereof comprises: obtaining a candidate video set; the candidate video set comprises a plurality of candidate videos, and the candidate videos are videos displayed in response to search requests; determining the interaction behavior type of the candidate videos under search requests; based on the interaction behavior type, filtering positive video samples and negative video samples from the candidate video set.

2. The training method of claim 1, wherein, The filtering of the positive video samples and the negative video samples from the candidate video set based on the interaction behavior type comprises: determining a candidate video selected for playing after being displayed in response to a search request as the video sample; based on the interaction behavior type, adding a positive sample label or a negative sample label to the video sample.

3. The training method of claim 2, wherein, The adding of the positive sample label or the negative sample label to the video sample based on the interaction behavior type comprises: determining that the interaction behavior type is a positive interaction behavior type, and adding a positive sample label to the corresponding video sample; determining that the interaction behavior type is not the positive interaction behavior type, and adding a negative sample label to the corresponding video sample.

4. The training method of claim 3, wherein, The positive interaction behavior type comprises at least one of a playing time longer than a time threshold, a like, and a follow.

5. The training method of claim 1, wherein, The content information comprises text information and image information, and the feature extraction network comprises a text feature extraction network and an image feature extraction network, wherein inputting the plurality of search terms and the content information into the feature extraction network to extract the plurality of search term embedding features corresponding to the plurality of search terms and the content embedding features corresponding to the content information comprises: inputting the plurality of search terms and the text information into the text feature extraction network to obtain the plurality of search term embedding features corresponding to the plurality of search terms and text embedding features corresponding to the text information; inputting the image information into the image feature extraction network to obtain image embedding features corresponding to the image information.

6. The training method of claim 5, wherein, The determination of N search term embedding features close to the content embedding features from the plurality of search term embedding features comprises: determining at least one search term embedding feature close to the text embedding feature from the plurality of search term embedding features, and / or determining at least one search term embedding feature close to the image embedding feature from the plurality of search term embedding features, to obtain the N search term embedding features.

7. The training method of claim 1, wherein The determining the N search term embedding features close to the content embedding feature from the plurality of search term embedding features comprises: determining the N search term embedding features closest to the content embedding feature from the plurality of search term embedding features; or determining the N search term embedding features with a distance less than a distance threshold to the content embedding feature from the plurality of search term embedding features.

8. The training method of claim 5, wherein, The image feature extraction network comprises cascaded bottom convolution layers and semantic convolution layers, and a full connection layer connected to the semantic convolution layers, and the adjusting the parameters of the feature extraction network and the click rate estimation network according to the value of the loss function to obtain the trained search term recommendation model comprises: adjusting the parameters of the semantic convolution layers, the full connection layer and the click rate estimation network in the image feature extraction network according to the value of the loss function, and keeping the parameters of the bottom convolution layers of the image feature extraction network and the text feature extraction network unchanged, to obtain the trained search term recommendation model.

9. A search term recommendation method characterized by comprising: comprises: extracting a plurality of search terms and content information corresponding to a reference video, and inputting the search terms and the content information into a feature extraction network of a search term recommendation model to extract a plurality of search term embedding features corresponding to the plurality of search terms and a content embedding feature corresponding to the content information, wherein the search term recommendation model comprises the feature extraction network and a click rate estimation network, and the search term recommendation model is trained by using the training method in any one of claims 1 to 8; determining N search term embedding features close to the content embedding feature from the plurality of search term embedding features; inputting the N search term embedding features and the content embedding feature into the click rate estimation network to obtain an estimated click rate of each search term embedding feature; selecting M search term embedding features from the N search term embedding features according to the estimated click rates, and taking search terms corresponding to the M search term embedding features as recommended search terms corresponding to the reference video.

10. The search term recommendation method of claim 9, wherein, The content information comprises text information and image information, and the feature extraction network comprises a text feature extraction network and an image feature extraction network, wherein inputting the plurality of search terms and the content information into the feature extraction network to extract the plurality of search term embedding features corresponding to the plurality of search terms and the content embedding feature corresponding to the content information comprises: inputting the plurality of search terms and the text information into the text feature extraction network to obtain the plurality of search term embedding features corresponding to the plurality of search terms and a text embedding feature corresponding to the text information; inputting the image information into the image feature extraction network to obtain an image embedding feature corresponding to the image information.

11. The search term recommendation method of claim 10, wherein, The determining the N search term embedding features close to the content embedding feature from the plurality of search term embedding features comprises: determining at least one search term embedding feature close to the text embedding feature from the plurality of search term embedding features, and / or determining at least one search term embedding feature close to the image embedding feature from the plurality of search term embedding features, to obtain the N search term embedding features.

12. The search term recommendation method as claimed in claim 9, characterized by, The determining the N search term embedding features close to the content embedding feature from the plurality of search term embedding features comprises: determining the N search term embedding features closest to the content embedding feature from the plurality of search term embedding features; or determining the N search term embedding features with a distance less than a distance threshold to the content embedding feature from the plurality of search term embedding features.

13. The search term recommendation method of claim 9, wherein, The selecting the M search term embedding features from the N search term embedding features according to the estimated click rates comprises: selecting the M search term embedding features corresponding to the maximum estimated click rates from the N search term embedding features; or selecting the M search term embedding features corresponding to the estimated click rates greater than a click rate threshold from the N search term embedding features.

14. The search term recommendation method of claim 9, wherein, The method further comprises: determining an interactive user having an interactive association relationship with the reference video; sending the recommended search term to a terminal corresponding to the interactive user, so that the terminal displays the recommended search term.

15. A training device for a search term recommendation model, characterized in that, The search term recommendation model comprises a feature extraction network and a click rate estimation network, and the training apparatus comprises: an acquisition unit configured to acquire a video sample and a sample label thereof; an extraction unit configured to extract a plurality of search terms and content information corresponding to the video sample, and input the feature extraction network to extract a plurality of search term embedding features corresponding to the plurality of search terms and a content embedding feature corresponding to the content information; a recall unit configured to determine N search term embedding features close to the content embedding feature from the plurality of search term embedding features; an estimation unit configured to input the N search term embedding features and the content embedding feature into the click rate estimation network to obtain an estimated click rate of each search term embedding feature; a calculation unit configured to determine a value of a loss function of the search term recommendation model according to the estimated click rate and the sample label; a parameter adjustment unit configured to obtain a trained search term recommendation model by adjusting parameters of the feature extraction network and the click rate estimation network according to the value of the loss function; The acquisition unit is further configured to: acquire a candidate video set; the candidate video set comprises a plurality of candidate videos, and the candidate videos are videos displayed in response to a search request; determine an interactive behavior type of the candidate videos under a search request; filter out positive video samples and negative video samples from the candidate video set based on the interactive behavior type.

16. The exercise device of claim 15, wherein, The acquisition unit is further configured to: determine a candidate video selected for playing after being displayed in response to a search request as the video sample; add a positive sample label or a negative sample label to the video sample based on the interaction behavior type.

17. The exercise device of claim 16, wherein, The acquisition unit is further configured to: determine that the interaction behavior type is a positive interaction behavior type, and add a positive sample label to the corresponding video sample; determine that the interaction behavior type is not the positive interaction behavior type, and add a negative sample label to the corresponding video sample.

18. The exercise device of claim 17, wherein, The positive interaction behavior type includes at least one of a play duration greater than a duration threshold, a like, and a follow.

19. The exercise device of claim 15, wherein, The content information includes text information and image information, and the feature extraction network includes a text feature extraction network and an image feature extraction network. The extraction unit is further configured to: input the plurality of search terms and the text information into the text feature extraction network to obtain a plurality of search term embedding features corresponding to the plurality of search terms and a text embedding feature corresponding to the text information; input the image information into the image feature extraction network to obtain an image embedding feature corresponding to the image information.

20. The exercise device of claim 19, wherein, The recall unit is further configured to: determine at least one search term embedding feature close to the text embedding feature from the plurality of search term embedding features, and / or determine at least one search term embedding feature close to the image embedding feature from the plurality of search term embedding features, to obtain the N search term embedding features.

21. The exercise device of claim 15, wherein, The recall unit is further configured to: determine N search term embedding features closest to the content embedding feature from the plurality of search term embedding features; or determine N search term embedding features with a distance less than a distance threshold to the content embedding feature from the plurality of search term embedding features.

22. The exercise device of claim 19, wherein, The image feature extraction network includes cascaded bottom convolution layers and semantic convolution layers, and a fully connected layer connected to the semantic convolution layers. The parameter adjustment unit is further configured to: adjust parameters of the semantic convolution layers, the fully connected layer, and the click rate estimation network in the image feature extraction network according to the value of the loss function, and keep the bottom structure of the image feature extraction network and the parameters of the text feature extraction network unchanged, to obtain a trained search term recommendation model.

23. A search term recommendation apparatus characterized by comprising: comprise: an extraction unit configured to extract a plurality of search terms and content information corresponding to a reference video, and input the plurality of search terms and the content information into a feature extraction network of a search term recommendation model to extract a plurality of search term embedding features corresponding to the plurality of search terms and a content embedding feature corresponding to the content information, wherein the search term recommendation model comprises the feature extraction network and a click rate estimation network, and the search term recommendation model is trained using a training method as claimed in any one of claims 1 to 8; a recall unit configured to determine N search term embedding features close to the content embedding feature from the plurality of search term embedding features; an estimation unit configured to input the N search term embedding features and the content embedding feature into the click rate estimation network to obtain an estimated click rate of each search term embedding feature; and an output unit configured to output the N search term embedding features with the estimated click rate of each search term embedding feature. The recommendation unit is configured to: select M search word embedding features from the N search word embedding features according to the estimated click rate, and take search words corresponding to the M search word embedding features as recommended search words corresponding to the reference video.

24. The search term recommendation apparatus of claim 23, wherein The content information includes text information and image information, and the feature extraction network includes a text feature extraction network and an image feature extraction network. The extraction unit is further configured to: input the plurality of search words and the text information into the text feature extraction network to obtain the plurality of search word embedding features corresponding to the plurality of search words and a text embedding feature corresponding to the text information; input the image information into the image feature extraction network to obtain an image embedding feature corresponding to the image information.

25. The search term recommendation device of claim 24, wherein, The recall unit is further configured to: determine at least one search word embedding feature close to the text embedding feature from the plurality of search word embedding features, and / or determine at least one search word embedding feature close to the image embedding feature from the plurality of search word embedding features, to obtain the N search word embedding features.

26. The search term recommendation device of claim 23, wherein, The recall unit is further configured to: determine N search word embedding features closest to the content embedding feature from the plurality of search word embedding features; or determine N search word embedding features with a distance less than a distance threshold to the content embedding feature from the plurality of search word embedding features.

27. The search term recommendation device of claim 23, wherein, The recommendation unit is further configured to: select M search word embedding features with the largest estimated click rate from the N search word embedding features; or select M search word embedding features with an estimated click rate greater than a click rate threshold from the N search word embedding features. The search word recommendation device further includes:

28. The search term recommendation device of claim 23, wherein, a determination unit configured to determine an interactive user having an interactive association relationship with the reference video; a sending unit configured to send the recommended search words to a terminal corresponding to the interactive user, so that the terminal displays the recommended search words. comprise:

29. An electronic device, comprising: at least one processor; at least one memory storing computer executable instructions, wherein the computer executable instructions, when executed by the at least one processor, cause the at least one processor to perform the training method of the search word recommendation model as claimed in any one of claims 1 to 8 or the search word recommendation method as claimed in any one of claims 9 to 14. The computer readable storage medium stores computer executable instructions, and when the instructions in the computer readable storage medium are executed by at least one processor, the at least one processor is caused to perform the training method of the search word recommendation model as claimed in any one of claims 1 to 8 or the search word recommendation method as claimed in any one of claims 9 to 14.

30. A computer-readable storage medium, characterized in that, The computer instructions are executed by at least one processor to implement the training method of the search word recommendation model as claimed in any one of claims 1 to 8 or the search word recommendation method as claimed in any one of claims 9 to 14.

31. A computer program product comprising computer instructions, characterized in that, ​

Images