An object label generation method, device and computer readable storage medium

Abstract

Embodiments of the present application disclose an object label generation method and device and a computer readable storage medium. After obtaining historical operation information of at least one target object in a preset time period, the embodiments of the present application filter target historical operation information corresponding to a target content of a preset content knowledge graph from the historical operation information. Then, based on the target historical operation information and the preset content knowledge graph, the historical object label of the target object is counted. Then, according to the target content and the target historical operation information, at least one predicted object label of the target object is determined from an object label set corresponding to the preset content knowledge graph. The historical object label and the predicted object label are fused to obtain the object label of the target object. The scheme can improve the accuracy of object label generation. The embodiments of the present application can be applied to various scenes such as cloud technology, artificial intelligence, intelligent transportation, and auxiliary driving.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of communication, in particular to an object label generation method, device and computer readable storage medium. BACKGROUND

[0002] In recent years, with the rapid development of Internet technology, the content in the Internet is also more and more rich. By browsing the content in the object, the interest information of the object to the content can be determined, and then the object label is generated. Through the object label, the image of the object can be concretized. The existing object label generation method is often obtained by manual editing or mining based on historical operation information.

[0003] In the research and practice process of the prior art, the inventors of the present application found that for the manual editing method, the coverage of the object is low, and by using the historical operation information for mining, the semantic of the mined object label is not uniform, the information is not rich, in addition, the source of the object label is single, therefore, the accuracy of the object label generation is low. SUMMARY

[0004] The embodiment of the present application provides an object label generation method, device and computer readable storage medium, which can improve the accuracy of object label generation.

[0005] An object label generation method comprises:

[0006] Obtain historical operation information of at least one target object in a preset time period;

[0007] Filter target historical operation information corresponding to a target content of a preset content knowledge graph from the historical operation information;

[0008] Based on the target historical operation information and the preset content knowledge graph, statistical historical object label of the target object is obtained;

[0009] According to the target content and the target historical operation information, at least one predicted object label of the target object is determined in an object label set corresponding to the preset content knowledge graph;

[0010] Fuse the historical object label and the predicted object label to obtain the object label of the target object.

[0011] Correspondingly, the embodiment of the present application provides an object label generation device, comprising:

[0012] An acquisition unit is configured to acquire historical operation information of at least one target object in a preset time period;

[0013] The screening unit is configured to screen, from the historical operation information, target historical operation information corresponding to a target content of a preset content knowledge graph;

[0014] The statistical unit is configured to count, based on the target historical operation information and the preset content knowledge graph, a historical object label of the target object;

[0015] The determination unit is configured to determine, from the target content and the target historical operation information, at least one predicted object label of the target object in an object label set corresponding to the preset content knowledge graph;

[0016] The fusion unit is configured to fuse the historical object label and the predicted object label to obtain an object label of the target object.

[0017] Optionally, in some embodiments, the determination unit can be specifically configured to perform feature extraction on text information of the target content to obtain a content feature of the target content; convert the content feature into an object feature of the target object based on the target historical operation information; and filter, from the object label set corresponding to the preset content knowledge graph, at least one object label corresponding to the object feature to obtain the predicted object label of the target object.

[0018] Optionally, in some embodiments, the determination unit can be specifically configured to filter, based on the target historical operation information, at least one content feature of a target content operated by the target object in a preset time interval from the content feature to obtain a basic content feature; perform maximum pooling processing on the basic content feature by using a trained object feature generation model to obtain a target pooled content feature; and perform nonlinear conversion on the target pooled content feature to obtain the object feature of the target object.

[0019] Optionally, in some embodiments, the determination unit can be specifically configured to obtain a sample content feature set corresponding to a historical content sample set operated by the target object, the sample content feature set including at least one sample content feature with a labeled classification result; determine at least one target sample content feature from the sample content feature set, and use sample content features in the sample content feature set except the target sample content feature as to-be-classified sample content features of to-be-classified sample content; and based on the labeled classification result, the target sample content feature, and the to-be-classified sample content feature, converge a preset object feature generation model to obtain a trained object feature generation model.

[0020] Optionally, in some embodiments, the determining unit can be specifically configured to: generate a sample object feature of the target object by using the preset object feature generation model on the target sample content feature; generate a classification sample content feature of the content to be classified by using the preset object feature generation model on the sample content feature to be classified; predict a classification result of the target object for the content to be classified based on the sample object feature and the classification sample content feature, to obtain a predicted classification result; and converge the preset object feature generation model according to the predicted classification result and the labeled classification result, to obtain a trained object feature generation model.

[0021] Optionally, in some embodiments, the determining unit can be specifically configured to: obtain an object label set corresponding to the preset content knowledge graph, and extract triple information in the preset content knowledge graph; extract a feature of an object label in the object label set by using the trained object label prediction model according to the triple information, to obtain an object label feature of at least one candidate object label corresponding to the object feature; and calculate a feature similarity between the object feature and the object label feature, to obtain a predicted object label of the target object.

[0022] Optionally, in some embodiments, the determining unit can be specifically configured to: filter out, by using the trained object label prediction model, at least one candidate object label corresponding to the object feature from the object label set, and extract a feature of the candidate object label, to obtain a basic object label feature of the candidate object label; extract a feature of the triple information by using the trained object label prediction model, to obtain a basic entity feature and an entity relationship feature, and fuse the basic entity feature and the entity relationship feature, to obtain an entity feature; and fuse the entity feature and the basic object label feature, to obtain an object label feature of the candidate object label.

[0023] Optionally, in some embodiments, the determining unit can be specifically configured to: obtain a target object label sample corresponding to a target object, and randomly negatively sample an object label negative sample from the object label set based on the target object label sample, to obtain an object label sample set; extract a triple information sample from the preset content knowledge graph, and construct a triple information negative sample based on the triple information sample, to obtain a triple information sample set; and converge the preset object label prediction model based on the object label sample set and the triple information sample set, to obtain a trained object label prediction model.

[0024] Optionally, in some embodiments, the determining unit can be specifically configured to determine object label loss information of the target object based on the object label sample set; determine triple loss information of the target object according to the triple information sample set; update network parameters of the preset object label prediction model based on the object label loss information and the triple loss information, and perform regularization processing on the updated network parameters to obtain the trained object label prediction model.

[0025] Optionally, in some embodiments, the determining unit can be specifically configured to calculate a feature similarity between the object feature and the object label feature, and take the feature similarity as a predicted interest degree of the target object for an object label corresponding to the object label feature; sort the candidate object labels based on the predicted interest degree; and filter out a predicted object label of the target object from the candidate object labels according to sorting information of the candidate object labels.

[0026] Optionally, in some embodiments, the statistical unit can be specifically configured to extract a graph relationship from the preset content knowledge graph, the graph relationship being used to indicate a relationship between the target content and a candidate historical object label; count an operation frequency of the target object for each target content in the target historical operation information; and determine a historical object label of the target object based on the operation frequency and the graph relationship.

[0027] Optionally, in some embodiments, the statistical unit can be specifically configured to identify a candidate historical object label corresponding to each target content in the graph relationship; sort the candidate historical object labels based on the operation frequency; filter out a historical object label of the target object from the candidate historical object labels according to sorting information of the candidate historical object labels, and determine a historical interest degree of each historical object label.

[0028] Optionally, in some embodiments, the fusion unit can be specifically configured to obtain a weighting coefficient corresponding to each object label type, weight the historical interest degree based on the weighting coefficient and the operation frequency to obtain a weighted historical interest degree of the historical object label, weight the predicted interest degree according to the weighting coefficient to obtain a weighted predicted interest degree, and fuse the weighted historical interest degree and the weighted predicted interest degree to obtain an object label of the target object.

[0029] Optionally, in some embodiments, the fusion unit can be specifically configured to fuse the weighted historical interest degree and the weighted predicted interest degree to obtain a fused interest degree, normalize the fused interest degree to obtain a target historical interest degree of the historical object label and a target predicted interest degree of the predicted object label, and filter out the object label of the target object from the historical object label and the predicted object label based on the target historical interest degree and the target predicted interest degree.

[0030] Optionally, in some embodiments, the fusion unit can be specifically configured to filter out a target historical object label from the historical object label according to the historical interest degree, filter out a target operation frequency corresponding to the target historical object label from the operation frequencies, determine an operation parameter of the target historical object based on the target operation frequency, fuse the operation parameter and the historical interest degree of the target historical object label to obtain a fused historical interest degree, and weight the fused historical interest degree based on the weighting coefficient to obtain the weighted historical interest degree of the historical object label.

[0031] Optionally, in some embodiments, the screening unit can be specifically configured to extract at least one content identifier from the preset content knowledge graph, identify a target content corresponding to the content identifier in the historical operation information, filter out initial historical operation information corresponding to the target content from the historical operation information, and clean the initial historical operation information to obtain target historical operation information corresponding to the target content.

[0032] In addition, an electronic device is also provided in the embodiments of the present application, which comprises a processor and a memory, the memory stores an application program, and the processor is configured to run the application program in the memory to implement the object label generation method provided in the embodiments of the present application.

[0033] In addition, a computer readable storage medium is also provided in the embodiments of the present application, which stores a computer program, and the computer program is adapted to be loaded by a processor to execute the steps in any one of the object label generation methods provided in the embodiments of the present application.

[0034] The embodiment of the application acquires historical operation information of at least one target object in a preset time period, screens target historical operation information corresponding to target content of a preset content knowledge graph from the historical operation information, then, based on the target historical operation information and the preset content knowledge graph, counts historical object labels of the target object, then, according to the target content and the target historical operation information, determines at least one predicted object label of the target object in an object label set corresponding to the preset content knowledge graph, and fuses the historical object labels and the predicted object label to obtain object labels of the target object. Since the scheme mines historical operation information of the target object and also uses associated information of the preset content knowledge graph, the scheme can improve semantic inconsistency of the object labels, and can mine the historical object labels and predict potential object labels of the object, so that accuracy of generation of the object labels can be improved. BRIEF DESCRIPTION OF DRAWINGS

[0035] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative effort.

[0036] Figure 1 is a scene diagram of the object label generation method provided by the embodiment of the present application;

[0037] Figure 2 is a flow diagram of the object label generation method provided by the embodiment of the present application;

[0038] Figure 3 is a diagram of a subgraph in a video knowledge graph provided by the embodiment of the present application;

[0039] Figure 4 is a flow diagram of counting the historical object labels provided by the embodiment of the present application;

[0040] Figure 5 is a diagram of training a preset object feature generation model provided by the embodiment of the present application;

[0041] Figure 6 is a flow diagram of predicting the predicted object labels of the target object provided by the embodiment of the present application;

[0042] Figure 7 is a flow diagram of generating the object labels of the target object provided by the embodiment of the present application;

[0043] Figure 8 is another flow diagram of the object label generation method provided by the embodiment of the present application;

[0044] Figure 9 is a structural schematic diagram of an object label generation device provided by an embodiment of the present application.

[0045] Figure 10 is a structural schematic diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0046] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person skilled in the art without creative work fall within the protection scope of the present application.

[0047] The embodiments of the present application provide an object label generation method, device and computer readable storage medium. The object label generation device can be integrated in an electronic device, which can be a server or a terminal or the like.

[0048] The server can be a stand-alone physical server, a server cluster or a distributed system composed of multiple physical servers, or a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery network (CDN), and basic cloud computing services such as big data and artificial intelligence platforms. The terminal can include but is not limited to a mobile phone, a computer, a smart voice interaction device, a smart home appliance, a vehicle-mounted terminal, an aircraft, and the like, but is not limited thereto. The terminal and the server can be directly or indirectly connected through wired or wireless communication, which is not limited in the present application. The embodiments of the present application can be applied to various scenarios, including but not limited to cloud technology, artificial intelligence, intelligent transportation, and assisted driving. The object label generated in the embodiments of the present application can be applied to application scenarios such as advertisement delivery, video personalized recommendation, and product operation. The implicit information (features) of the object and the object label generated in the process of generating the object label can also be used in application scenarios such as object learning, content recommendation, and content recall model learning processes.

[0049] For example, referring to Figure 1For example, the object label generation apparatus is integrated in an electronic device. After the electronic device obtains historical operation information of at least one target object in a preset time period, the electronic device filters target historical operation information of a target content corresponding to a preset content knowledge graph from the historical operation information. Then, based on the target historical operation information and the preset content knowledge graph, the electronic device counts historical object labels of the target object. Then, according to the target content and the target historical operation information, the electronic device determines at least one predicted object label of the target object in an object label set corresponding to the preset content knowledge graph. The historical object labels and the predicted object labels are fused to obtain object labels of the target object, thereby improving the accuracy of object label generation.

[0050] The object label can be understood as abstracting the attributes and behaviors of an object into a label, and is a kind of expression mode for visualizing an object. In the present scheme, the object label focuses on virtualizing the interests of an object, thereby obtaining an interest label of the object, and visualizing a user. Therefore, the object label can also be an interest label of the object.

[0051] The following will be described in detail. It should be noted that the order of the following embodiments is not limited as the preferred order of the embodiments.

[0052] As shown in Figure 2 The object label generation method is executed by an electronic device, and the specific process is as follows:

[0053] 101. Obtain historical operation information of at least one target object in a preset time period.

[0054] The historical operation information can be understood as information about the operation of a target object on content or business in a content platform or a business platform. The target object can be a user of the content platform or the business platform.

[0055] There are various ways to obtain the historical operation information of the target object, which can be as follows:

[0056] For example, the historical operation log of at least one target object can be obtained, and the operation information in the preset time period can be read from the historical operation log to obtain the historical operation information of the target object. Alternatively, the historical operation log of at least one target object can be obtained, and the target historical operation log in the preset time period can be filtered from the historical operation log. The operation information of the target object on the content can be read from the target historical operation log to obtain the historical operation information of the target object.

[0057] 102. Filter target historical operation information of a target content corresponding to a preset knowledge graph from the historical operation information.

[0058] The preset content knowledge graph can be understood as a pre-set knowledge graph related to the content. The knowledge graph is a knowledge base containing entities and relationships, and is commonly represented by triplets such as (head entity, relationship, tail entity). The knowledge graph has the advantages of containing semantic relationships of related entities, diverse relationships, and good interpretability, and has been widely used in search, recommendation, and question answering in recent years. Taking the preset knowledge graph as the knowledge graph corresponding to the video, Figure 3 The screenshot is a subgraph of the video knowledge graph, from Figure 3 As can be seen from the knowledge graph, the knowledge graph contains a variety of information, such as video entities, abstract entities, character entities, label entities, various relationships, and entity-relationship triplets.

[0059] Among them, the way to filter out the target historical operation information corresponding to the target content of the preset knowledge graph in the historical operation information can be various, and can be as follows:

[0060] For example, at least one content identifier can be extracted from the preset content knowledge graph, and the target content corresponding to the content identifier can be identified in the historical operation information. The initial historical operation information corresponding to the target content is filtered out in the historical operation information, and the initial historical operation information is cleaned to obtain the target historical operation information corresponding to the target content.

[0061] Among them, the target content can be the content associated in the content knowledge graph, and the form of the target content can be various, such as video, audio, text or other content. The way to clean the initial historical operation information corresponding to the target content can be various, such as identifying the operation duration of the target content in the initial historical operation information, obtaining the content duration of the target content, and calculating the duration ratio of the operation duration and the content duration. The operation information of the target content whose duration ratio does not exceed the preset ratio threshold is deleted in the initial historical operation information to obtain the target historical operation information corresponding to the target content.

[0062] Among them, the preset ratio threshold can be set according to actual application, such as 0.5 or other threshold. Taking the content as a video and the initial operation information as video watching information as an example, the preset ratio threshold can be used to remove noise information in the video watching information, thereby extracting effective video watching information.

[0063] 103, based on the target historical operation information and the preset content knowledge graph, the historical object label of the target object is counted.

[0064] Among them, the historical object label can be understood as an object label determined based on the target historical operation information of the target object, mainly indicating the object label of the target object in the historical stage.

[0065] The manner of counting the historical object label of the target object based on the target historical operation information and the preset content knowledge graph can be various, and can be as follows:

[0066] For example, the graph relationship is extracted in the preset content knowledge graph, the operation frequency of the target object on each target content is counted in the target historical operation information, and the historical object label of the target object is determined based on the operation frequency and the graph relationship.

[0067] The graph relationship is used to indicate the relationship between the target content and the candidate historical object label, and the candidate historical object label corresponding to the target content can be determined based on the graph relationship.

[0068] The manner of counting the operation frequency of the target object on each target content in the target historical operation information can be various, such as recognizing the operation information of the target object on the target content in the target historical operation information, and counting the number of the operation information, so that the operation frequency of the target object on the target content can be obtained.

[0069] After counting the operation frequency of the target content, the historical object label of the target object can be determined based on the operation frequency and the graph relationship. The manner of determining the historical object label can be various, such as recognizing the candidate historical object label corresponding to each target content in the graph relationship, sorting the candidate historical object label based on the operation frequency, filtering the historical object label of the target object from the candidate historical object label according to the sorting information of the candidate historical object label, and determining the historical interest degree of each historical object label.

[0070] The historical interest degree can be the degree of interest of the target object in the target content corresponding to the historical object label. The manner of determining the historical interest degree of each historical object label can be various, such as extracting the interest feature of the target object on the historical object label from the sorting information of the historical object label, and determining the historical interest degree of the historical object label based on the interest feature.

[0071] For example, the target content is a video, and the manner of counting the historical object label of the target object can be as follows: Figure 4As shown, effective video ID information can be extracted according to a preset content knowledge graph (KG), and the object video information of the target object in the historical operation log in the database within a period of time is extracted, and the effective video ID in the knowledge graph is extracted. The effective video information of the target object is calculated. At present, it is considered that the video viewing time length ratio is more than 0.5 as an effective viewing behavior. According to the effective viewing behavior, the frequency of all historical object labels is counted in the graph relationship, and then the topN candidate historical portrait labels are cut off according to the frequency as the historical object labels of the target object.

[0072] 104、According to the target content and the target historical operation information, at least one predicted object label of the target object is determined in the object label set corresponding to the preset content knowledge graph.

[0073] Among them, the predicted object label can be understood as the potential object label of the target object, which represents the potential interest label of the target object.

[0074] Among them, there are many ways to determine at least one predicted object label of the target object in the object label set corresponding to the preset content knowledge graph, which can be as follows:

[0075] For example, the text information of the target content can be extracted to obtain the content features of the target content, and the content features can be converted into object features of the target object based on the target historical operation information. At least one object label corresponding to the object features is filtered out in the object label set corresponding to the preset content knowledge graph, and the predicted object label of the target object is obtained. Specifically, it can be as follows:

[0076] S1, the text information of the target content is extracted to obtain the content features of the target content.

[0077] Among them, the text information can be the text information contained in the target content, or the text information of the introduction or attribute information of the target content. The type of text information can be various, for example, taking the target content as a video as an example, the text information of the target content can include video title, video introduction or video subtitle text information.

[0078] Among them, there are many ways to extract the text information of the target content, which can be as follows:

[0079] For example, the text information can be extracted from the target content, and the text type of the text information can be identified. Based on the text type, the target text information is filtered out in the text information, and the semantic feature extraction of the target text information is performed to obtain the content features of the target content.

[0080] The text type is used to indicate the type of text information, and the text type can be various, such as a title, abstract information (introduction information), subtitles, and the like. Based on the text type, the target text information can be filtered out from the text information in various ways. For example, taking a video as an example, based on the text type, the video title and the abstract information of the video are filtered out from the text information, and the video title and the abstract information are taken as the target text information.

[0081] After the target text information is filtered out, semantic feature extraction can be performed on the target text information. The semantic feature extraction can be performed in various ways, such as using a pre-trained semantic feature extraction model to extract semantic features from the video title and the abstract information, and taking the semantic features as content features of the target content. The dimension of the content features can be 768 or other dimensions.

[0082] S2, based on the target historical operation information, converting the content features into object features of the target object.

[0083] The object features can be feature information representing the target object.

[0084] The content features can be converted into object features of the target object in various ways, as follows:

[0085] For example, based on the target historical operation information, at least one target content feature of the target object operated in a preset time interval can be filtered out from the content features to obtain a basic content feature. A trained object feature generation model is used to perform maximum pooling processing on the basic content feature to obtain a target pooled content feature. The target pooled content feature is nonlinearly converted to obtain the object features of the target object.

[0086] The maximum pooling processing of the basic content feature by the trained object feature generation model can be performed in various ways. For example, the trained object feature generation model can be used to perform pooling processing on the basic content feature. At this time, the pooled basic content feature with the same feature dimension as the basic content feature is outputted, and the basic content feature with the largest feature value in the pooled basic content feature is filtered out as the initial pooled content feature corresponding to the target object. The initial pooled content feature is activated by the activation layer (ReLU layer) of the trained object feature generation model, and then the initial pooled content feature after the activation is compressed to 128 dimensions, thereby obtaining the target pooled content feature.

[0087] Wherein, the trained object feature generation model can be set according to actual application, in addition, it should be pointed out that, the trained object feature generation model can be pre-set by maintenance personnel, or can be trained by the object label generation device itself, that is, before the step "performing maximum pooling processing on the basic content feature by using the trained object feature generation model", the object label generation method can further include:

[0088] Obtaining a sample content feature set corresponding to a historical content sample set operated by the target object, the sample content feature set including at least one sample content feature with an annotated classification result, determining at least one target sample content feature in the sample content feature set, and taking the sample content features in the sample content feature set except the target sample content feature as the sample content features of the sample content to be classified, converging the preset object feature generation model based on the annotated classification result, the target sample content feature and the sample content feature to be classified, and obtaining the trained object feature generation model.

[0089] Wherein, the annotated classification result can be understood as the classification result of whether the target object is interested in the sample content corresponding to the target sample content feature. There are many ways to converge the preset object feature generation model based on the annotated classification result, the target sample content feature and the sample content feature to be classified, such as converting the target sample content feature into a sample object feature of the target object by using the preset object feature generation model, converting the sample content feature to be classified into a classification sample content feature of the classification content to be classified by using the preset object feature generation model, predicting the classification result of the target object for the classification content to be classified based on the sample object feature and the classification sample content feature, obtaining the predicted classification result, and converging the preset feature generation model according to the predicted classification result and the annotated classification result, obtaining the trained object feature generation model.

[0090] Wherein, taking a video as an example, the process of training the preset object feature generation model can be as shown in Figure 5 The sample content feature of the target video watched by the target object in a period of time is converted into a sample object feature of the target object, then the classification sample content feature of the video sample to be classified is also converted into a sample content feature to be classified, the converted feature is the sample content feature to be classified of the video sample to be classified, the classification calculation of whether the target object is interested in watching the video to be classified is calculated through the classification network (softmax), so as to obtain the predicted classification result, the predicted classification result is compared with the annotated classification result, so as to obtain the loss information of the target object, and the object feature generation model is converged based on the loss information, and the trained object feature generation model is obtained.

[0091] S3, filtering out at least one object label corresponding to the object feature from the object label set corresponding to the preset content knowledge graph to obtain a predicted object label of the target object.

[0092] The predicted object label can be understood as a potential object label of the target object.

[0093] The filtering of the at least one object label corresponding to the object feature from the object label set can be performed in various ways, and can be performed as follows:

[0094] For example, the object label set corresponding to the preset content knowledge graph is obtained, and triple information is extracted from the preset content knowledge graph. According to the triple information, the object label set is subjected to feature extraction by using the trained object label prediction model to obtain object label features of at least one candidate object label corresponding to the object feature, the feature similarity between the object feature and the object label feature is calculated, and the predicted object label of the target object is obtained.

[0095] The triple information can be understood as information composed of a head node, a relationship node and a tail node associated with the content in the preset content knowledge graph. The head node and the tail node can be entities, and the relationship node can be a relationship between entities. According to the triple information, the object label set can be subjected to feature extraction by using the trained object label prediction model in various ways, such as filtering out at least one candidate object label corresponding to the object feature from the object label set by using the trained object label prediction model, and extracting features of the candidate object label to obtain basic object label features of the candidate object label. The trained object label prediction model is used to extract features of the triple information to obtain basic entity features and entity relationship features, and the basic entity features and the entity relationship features are fused to obtain entity features. The entity features and the basic object label features are fused to obtain object label features of the candidate object label.

[0096] After obtaining the object label features of the candidate object label, the feature similarity between the object feature and the object label feature can be calculated. The feature similarity can be calculated in various ways, such as calculating the feature similarity between the object feature and the object label feature, and taking the feature similarity as a predicted interest degree of the object label corresponding to the object label feature of the target object to the object label. Based on the predicted interest degree, the candidate object label is sorted, and the predicted object label of the target object is filtered out from the candidate object label according to the sorting information of the candidate object label.

[0097] The calculating the feature similarity can be understood as calculating the correlation between the candidate object label feature and the object feature, so as to obtain the predicted interest degree of the object label corresponding to the object label feature of the target object. Then, the candidate object label is sorted based on the predicted interest degree. The manner of screening the predicted object label of the target object from the candidate object label based on the sorting information of the candidate object label can be various. For example, at least one candidate object label whose sorting position exceeds a preset position threshold can be screened from the candidate object label, so as to obtain the predicted object label of the target object.

[0098] The trained object label prediction model can be set according to actual application. In addition, it should be noted that the trained object label prediction model can be pre-set by maintenance personnel or trained by the object label generation device itself. That is, before the step of "screening at least one candidate object label corresponding to the object feature from the object label set by using the trained object label prediction model", the object label generation method can further include:

[0099] The target object label sample corresponding to the target object is obtained, and the object label negative sample is randomly negatively sampled from the object label set based on the target object label sample, so as to obtain the object label sample set. The triple information sample is extracted from the preset content knowledge graph, and the triple information negative sample is constructed based on the triple information sample, so as to obtain the triple information sample set. The preset object label prediction model is converged based on the object sample label set and the triple information sample set, so as to obtain the trained object label prediction model.

[0100] The manner of constructing the triple information negative sample based on the triple information sample can be various. For example, the head node, tail node and relation node in each triple information sample are extracted, a tail node not connected with the head node is randomly selected from the tail node as a target tail node, the head node relation node and the target tail node are taken as the triple information negative sample, and the triple information sample and the triple information negative sample are taken as the triple information sample set.

[0101] After the triple information sample set is constructed, the preset object label prediction model can be converged based on the object label sample set and the triple information sample set. The manner of convergence can be various. For example, the object label loss information of the target object is determined based on the object label sample set, the triple loss information of the target object is determined according to the triple information sample set, the network parameters of the preset object label prediction model are updated based on the object label loss information and the triple loss information, and the updated network parameters are regularized to obtain the trained object label prediction model.

[0102] The object label loss information is primarily used to learn object label features by calculating the loss information of the target object's object label preference during training of the preset object label prediction model. The triplet loss information can be learned from the knowledge graph using the TransR fusion network in the preset object label prediction model. During training, it is mainly used to calculate the loss information of the preset knowledge graph network to learn entity embeddings and relation embeddings. In addition to the above two loss information components, the overall optimization of the preset object label prediction model can also include regularization terms for all network parameters to prevent overfitting. Furthermore, to make the optimization objective of the preset object label prediction model approach the global optimum more quickly, a method of dynamically decaying the learning rate based on the detection metric status can be used. After training the preset object label prediction model, the process of predicting the predicted object label corresponding to the target object based on the target object's object label can be as follows: Figure 6 As shown, at least one candidate object label is selected from the set of object labels corresponding to the preset content knowledge graph. Feature extraction is performed on the candidate object labels to obtain basic object label features (portrait embedding). Feature extraction is also performed on the triple information in the preset content knowledge graph to obtain basic entity features (entity embedding) and relation features (relation embedding). The basic entity features and relation features are fused to obtain entity features. The entity features are then fused with the basic object label features to obtain the object label features for each candidate object label. Finally, the correlation between object feature labels and object features is calculated, and the results are sorted based on the calculated correlation. Thus, at least one predicted object label for the target object can be obtained.

[0103] 105. Merge the historical object tags and the predicted object tags to obtain the object tags of the target object.

[0104] For example, the weighting coefficients corresponding to each object tag type can be obtained, and the historical interest can be weighted based on the weighting coefficients and operation frequency to obtain the weighted historical interest of the historical object tags. The predicted interest can be weighted according to the weighting coefficients to obtain the weighted predicted interest. The weighted historical interest and the weighted predicted interest are then fused to obtain the object tag of the target object.

[0105] The weighting coefficient can be understood as a weight corresponding to each object label type. There are various ways to weight the historical interest degree based on the weighting coefficient and the operation frequency. For example, the target historical object label can be filtered out from the historical object label based on the historical interest degree, the target operation frequency corresponding to the target historical object label can be filtered out from the operation frequency, and the operation parameter of the target historical object label can be determined based on the target operation frequency. The operation parameter is fused with the historical interest degree of the target historical object label to obtain a fused historical interest degree. The fused historical interest degree is weighted based on the weighting coefficient to obtain the weighted historical interest degree of the historical object label.

[0106] The historical object label can be filtered out from the historical object label based on the historical interest degree.

[0107] The operation parameter of the target historical object label can be determined based on the target operation frequency. For example, a preset basic operation parameter can be obtained, and the preset basic operation parameter is fused with the target operation frequency to obtain the operation parameter of the historical object label.

[0108] The fused historical interest degree can be weighted based on the weighting coefficient. For example, the weighting coefficient can be directly multiplied by the fused historical interest degree to obtain the weighted historical interest degree of the historical object label. For details, refer to formula (1):

[0109]

[0110] score his，u，t is the weighted interest degree of the historical object label, is the weighting coefficient (weight). The object label can be divided into video entities, IPs, series, characters, 1-3 level labels, etc. Different types of object labels have different weights. r t is the target operation frequency, and a is a preset basic operation parameter, which is set according to actual application. For example, a can be 0.9 or other numerical values. score base，u is the historical interest degree of the target historical object label.

[0111] The prediction interest degree can be weighted based on the weighting coefficient. For example, the target weighting coefficient corresponding to the prediction object label can be filtered out from the weighting coefficient, and the prediction interest degree is weighted based on the target weighting coefficient to obtain the weighted prediction interest degree of the prediction object label. For details, refer to formula (2):

[0112]

[0113] wherein score pre，u，t is the weighted predicted interest degree of the predicted object label, is the target weighting coefficient, score preM，u，t is the predicted interest degree of the predicted object label.

[0114] After the historical interest degree and the predicted interest degree are weighted respectively, the weighted historical interest degree and the weighted predicted interest degree can be fused. The fusion manner can be various, such as fusing the weighted historical interest degree and the weighted predicted interest degree to obtain a fused interest degree, normalizing the fused interest degree to obtain the target historical interest degree of the historical object label and the target predicted interest degree of the predicted object label. The specific formula is shown in formula (3):

[0115] Score fusion，u，t = f Norm (score his，u，t + score pre，u，t ) (3)

[0116] wherein score fusion，u，t is the target historical interest degree of the historical object label and the target predicted interest degree of the predicted object label, f Norm is a parameter of the normalization processing, score his，u，t is the weighted historical interest degree of the historical object label, and score pre，u，t is the weighted predicted interest degree of the predicted object label.

[0117] Based on the target historical interest degree and the target predicted interest degree, the object label of the target object can be screened from the historical object label and the predicted object label in various manners. For example, the object label with the maximum interest degree value can be screened from the historical object label and the predicted object label as the object label of the target object. Alternatively, the historical object label and the predicted object label can be sorted based on the target historical interest degree and the target predicted interest degree, and the object label of the target object can be screened from the historical object label and the predicted object label based on the sorting result.

[0118] In the present scheme, the object label generation process of the target object can be as follows: Figure 7As shown, mainly includes three parts, the first part is based on the knowledge graph of historical object label mining, such as shown in figure 71, the second part is based on object label prediction model of potential object label prediction, such as shown in figure 72, the third part is the multi-element object label (historical object label and predicted object label) fusion. In the process of historical object label mining, mainly based on the historical operation information of the target object is parsed for the target content corresponding to the preset content knowledge graph target historical operation information, and then, through the historical object label mining strategy is mined. In the process of predicting potential object label (predicted object label), mainly through three models to predict the predicted object label, mainly includes content feature generation model generates the content feature of the target content, object feature generation model converts the content feature into the object feature of the target object, object label prediction model based on the preset knowledge graph and object feature generates the object label feature of the candidate object label of the target object, and then, through the predicted object label calculation strategy to predict the predicted object label. Finally, the predicted object label and the historical object label are fused, so as to determine the object label of the target object.

[0119] Among them, it needs to be explained here that the main role of the scheme to introduce the knowledge graph to generate the object label of the target object can be that first, with the help of the correlation between entities, the deep interest information of the object is learned, and the information of the relationship of each similar object is learned, which improves the problem of cold start of the user; Second, the label graph in the knowledge graph is processed through semantic segmentation, alignment and rewriting, so it is more unified in semantics, and therefore the label interest portrait learned through the knowledge graph improves the problem of inconsistent semantics of the previous interest label; Third, the entity graph of the knowledge graph has the ability to generate series of abstract entities such as IP, which enriches the object label set of the object interest portrait.

[0120] Among them, the top 10 fusion object labels calculated by the scheme for a certain online object can be shown in table 1.

[0121] Table 1

[0122]

[0123]

[0124] Among them, HS is the object label mined based on the historical operation information, and PR is the object label predicted by the predicted object label.

[0125] In addition, the object label generation process of the present solution is evaluated by using the DCG (a search engine quality index) in the form of a percentage. The result of the historical object label mined by using the historical operation information on line is P20=55 points, and the result of the object label generated by using the present solution is P20=63 points. Compared with the ranking result mined on line, the relative improvement is 14.5%.

[0126] As can be seen from the above, after obtaining the historical operation information of at least one target object in a preset time period, the target historical operation information of the target content corresponding to the preset content knowledge graph is screened out in the historical operation information, and then the historical object label of the target object is counted based on the target historical operation information and the preset content knowledge graph. Then, according to the target content and the target historical operation information, at least one predicted object label of the target object is determined in the object label set corresponding to the preset content knowledge graph. The historical object label and the predicted object label are fused to obtain the object label of the target object. Since the present solution mines the historical operation information of the target object as a prerequisite and also uses the associated information of the preset content knowledge graph, the problem of inconsistent semantics of the object label can be improved, and the potential object label of the object can be predicted while the historical object label is mined. Therefore, the accuracy of the object label generation can be improved.

[0127] According to the method described in the above embodiment, the following will be further described by way of example.

[0128] In the present embodiment, the method is executed by an electronic device, the electronic device is a server, the target content is a target video, and the historical operation information is historical viewing information. The present embodiment will be described by way of example.

[0129] As shown in Figure 8 An object label generation method, and the specific process is as follows:

[0130] 201. The server obtains the historical viewing information of at least one target object in a preset time period.

[0131] For example, the server can obtain the historical operation log of at least one target object, read the operation information in the preset time period in the historical operation log, and thus obtain the historical viewing information of the target object, or the server can also obtain the historical operation log of at least one target object, screen out the target historical operation log in the preset time period in the historical operation log, read the operation information of the target object for the video in the target historical operation log, and thus obtain the historical viewing information of the target object.

[0132] 202. The server screens out the target historical viewing information of the target video corresponding to the preset knowledge graph in the historical viewing information.

[0133] For example, the server extracts at least one video identifier in the preset video knowledge graph, identifies a target video corresponding to the video identifier in the historical viewing information, and screens and obtains initial historical viewing information corresponding to the target video in the historical viewing information. The server identifies an operation duration for the target video in the initial historical viewing information, obtains a video duration of the target video, calculates a duration ratio of the operation duration and the video duration, deletes operation information of the target video with the duration ratio less than 0.5 in the initial historical viewing information, and obtains target historical viewing information corresponding to the target video.

[0134] 203. The server counts historical object labels of the target object based on the target historical viewing information and the preset video knowledge graph.

[0135] For example, the server extracts a graph relationship in the preset video knowledge graph, identifies operation information of the target object on the target video in the target historical viewing information, and counts a number of times of the operation information, so as to obtain an operation frequency of the target object on the target video. The server identifies a candidate historical object label corresponding to each target video in the graph relationship, sorts the candidate historical object labels based on the operation frequency, screens a historical object label of the target object from the candidate historical object labels according to the sorting information of the candidate historical object labels, extracts a feature of the historical object label from the sorting information of the historical object label, and determines a historical interest degree of the historical object label based on the feature of the historical object label.

[0136] 204. The server extracts a feature of text information of the target video to obtain a video feature of the target video.

[0137] For example, the server can extract text information from the target video, identify a text type of the text information, screen video title and abstract information of the target video from the text information based on the text type, and take the video title and the abstract information as target text information. The server extracts a semantic feature from the video title and the abstract information by using a pre-trained semantic feature extraction model, obtains the semantic feature, and takes the semantic feature as the video feature of the target video. The dimension of the video feature can be 768.

[0138] 205. The server converts the video feature into an object feature of the target object based on the target historical viewing information.

[0139] For example, the server can filter out, based on the target historical viewing information, video features of at least one target video in which the target object operates in a preset time interval in the video features, to obtain basic video features. The server can perform pooling processing on the basic video features by using the trained object feature generation model, and at this time, output the pooled basic video features with the same feature dimension as the basic video features, and filter out, in the pooled basic video features, a basic video feature with the largest feature value as an initial pooled video feature corresponding to the target object. The server can perform activation processing on the initial pooled video feature by using an activation layer (ReLU layer) of the trained object feature generation model, and then compress the initial pooled video feature after the activation processing to 128 dimensions, to obtain a target pooled video feature. The server can perform nonlinear conversion on the target pooled video feature, to obtain an object feature of the target object.

[0140] Optionally, before performing the maximum pooling processing on the basic video features by using the trained object feature generation model, the server can also train a preset object feature generation model, to obtain a trained object feature generation model. The training process can be various, for example, the server can convert sample video features of target videos watched by the target object in a period of time into sample object features of the target object, and then perform feature conversion on classification sample video features of a video sample to be classified, to obtain classified sample video features of the video sample to be classified. The server can calculate a classification result of whether the target object is interested in watching the video to be classified by using a classification network (softmax), to obtain a predicted classification result. The server can compare the predicted classification result with a labeled classification result, to obtain loss information of the target object. Based on the loss information, the server can converge the object feature generation model, to obtain the trained object feature generation model.

[0141] 206. The server filters out, in an object label set corresponding to the preset video knowledge graph, at least one object label corresponding to the object feature, to obtain a predicted object label of the target object.

[0142] For example, the server can obtain an object label set corresponding to a preset video knowledge graph, and extract triple information in the preset video knowledge graph. The server can filter out, in the object label set, at least one candidate object label corresponding to the object feature by using a trained object label prediction model, and perform feature extraction on the candidate object label, to obtain basic object label features of the candidate object label. The server can perform feature extraction on the triple information by using the trained object label prediction model, to obtain entity features and entity relationship features, and fuse the entity features and the entity relationship features, to obtain entity features. The server can fuse the entity features and the basic object label features, to obtain object label features of the candidate object label.

[0143] The server calculates a feature similarity between the object feature and the object feature label, takes the feature similarity as a predicted interest degree of an object label corresponding to an object label feature pair of the target object, sorts the candidate object labels based on the predicted interest degree, and filters out a predicted object label of the target object from the candidate object labels according to the sorting information of the candidate object labels.

[0144] Optionally, before the server filters out at least one candidate object label corresponding to the object feature from the object label set by using the trained object label prediction model, the server can also train the preset object label training model to obtain the trained object label prediction model. The training process can be various, for example, a target object label sample corresponding to a target object is obtained, and an object label negative sample is randomly negatively sampled from the object label set based on the target object label sample to obtain an object label sample set. A triple information sample is extracted from the preset video knowledge graph, and a head node, a tail node and a relation node in each triple information are extracted from the triple information sample. A tail node that is not connected to the head node is randomly filtered out from the tail node as a target tail node, and the head node and the target tail node are taken as a triple information negative sample. The triple information sample and the triple information negative sample are taken as a triple information sample set.

[0145] The server determines object label loss information of the target object based on the object label sample set, determines triple loss information of the target object based on the triple information sample set, updates network parameters of the preset object label prediction model based on the object label loss information and the triple loss information, and performs regularization processing on the updated network parameters to obtain the trained object label prediction model.

[0146] 207、The server fuses the historical object label and the predicted object label to obtain an object label of the target object.

[0147] For example, the server can obtain a weighting coefficient corresponding to each object label type, filter out a historical object label with the largest historical interest degree from the historical object label, and thus obtain a target historical object label. The server fuses the preset basic operation parameter and the target operation frequency to obtain an operation parameter of the historical object label. The server fuses the operation parameter and the historical interest degree of the target historical object label to obtain a fused historical interest degree. The server multiplies the weighting coefficient and the fused historical interest degree to obtain a weighted historical interest degree of the historical object label. For details, refer to formula (1).

[0148] The server filters out a target weighting coefficient corresponding to the predicted object label from the weighting coefficient, weights the predicted interest degree based on the target weighting coefficient, and thus obtains a weighted predicted interest degree of the predicted object label. For details, refer to formula (2).

[0149] The server fuses the weighted historical interest degree and the weighted predicted interest degree to obtain a fused interest degree, and performs normalization processing on the fused interest degree to obtain a target historical interest degree of a historical object label and a target predicted interest degree of a predicted object label, which can be shown in formula (3). The object label with the largest interest degree value is selected from the historical object label and the predicted object label as the object label of the target object, or the historical object label and the predicted object label can be sorted based on the target historical interest degree and the target predicted interest degree, and the object label of the target object is selected from the historical object label and the predicted object label based on the sorting result.

[0150] As can be seen from the above, the server of the embodiment obtains the historical viewing information of at least one target object in a preset time period, screens the target historical viewing information of the target video corresponding to the preset video knowledge graph from the historical viewing information, then, based on the target historical viewing information and the preset video knowledge graph, the historical object label of the target object is counted, then, according to the target video and the target historical viewing information, at least one predicted object label of the target object is determined from the object label set corresponding to the preset video knowledge graph, and the historical object label and the predicted object label are fused to obtain the object label of the target object; since the scheme mines the historical viewing information of the target object and also uses the associated information of the preset video knowledge graph, it can improve the problem of inconsistent semantics of the object label, and can mine the historical object label and predict the potential object label of the object at the same time, so as to improve the accuracy of the object label generation.

[0151] In order to better implement the above method, the embodiment of the application also provides an object label generation device.

[0152] For example, as shown in Figure 9 The object label generation device can include an acquisition unit 301, a screening unit 302, a counting unit 303, a determination unit 304, and a fusion unit 305, as follows:

[0153] (1) The acquisition unit 301;

[0154] The acquisition unit 301 is configured to acquire historical operation information of at least one target object in a preset time period.

[0155] For example, the acquisition unit 301 can be specifically configured to acquire the historical operation log of the at least one target object, read the operation information of the preset time period from the historical operation log, and obtain the historical operation information of the target object, or can also acquire the historical operation log of the at least one target object, filter the target historical operation log of the preset time period from the historical operation log, read the operation information of the target object for the content from the target historical operation log, and obtain the historical operation information of the target object.

[0156] (2) the screening unit 302;

[0157] The screening unit 302 is configured to filter the target historical operation information of the target content corresponding to the preset content knowledge graph from the historical operation information.

[0158] For example, the screening unit 302 can be specifically configured to extract at least one content identifier from the preset content knowledge graph, identify the target content corresponding to the content identifier from the historical operation information, filter the initial historical operation information corresponding to the target content from the historical operation information, clean the initial historical operation information, and obtain the target historical operation information corresponding to the target content.

[0159] (3) the statistical unit 303;

[0160] The statistical unit 303 is configured to count the historical object label of the target object based on the target historical operation information and the preset content knowledge graph.

[0161] For example, the statistical unit 303 can be specifically configured to extract a graph relationship from the preset content knowledge graph, count the operation frequency of the target object for each target content from the target historical operation information, and determine the historical object label of the target object based on the operation frequency and the graph relationship.

[0162] (4) the determination unit 304;

[0163] The determination unit 304 is configured to determine at least one predicted object label of the target object from the object label set corresponding to the preset content knowledge graph according to the target content and the target historical operation information.

[0164] For example, the determination unit 304 can be specifically configured to perform feature extraction on the text information of the target content to obtain the content feature of the target content, convert the content feature into the object feature of the target object based on the target historical operation information, filter at least one object label corresponding to the object feature from the object label set corresponding to the preset content knowledge graph, and obtain the predicted object label of the target object.

[0165] (5) the fusion unit 305;

[0166] The fusion unit 305 is configured to fuse the historical object label and the predicted object label to obtain the object label of the target object.

[0167] For example, the fusion unit 305 can be specifically configured to obtain a weighting coefficient corresponding to each object label type, weight the historical interest degree based on the weighting coefficient and the operation frequency to obtain a weighted historical interest degree of the historical object label, weight the predicted interest degree based on the weighting coefficient to obtain a weighted predicted interest degree, and fuse the weighted historical interest degree and the weighted predicted interest degree to obtain the object label of the target object.

[0168] In specific implementation, the above units can be implemented as independent entities, or can be combined as the same or several entities, and the specific implementation of the above units can be referred to the method embodiments above, which will not be described herein.

[0169] As known from the above, after the obtaining unit 301 obtains the historical operation information of at least one target object in a preset time period, the screening unit 302 screens the target historical operation information of the target content corresponding to the preset content knowledge graph from the historical operation information, then the statistical unit 303 statistically obtains the historical object label of the target object based on the target historical operation information and the preset content knowledge graph, then the determination unit 304 determines at least one predicted object label of the target object from the object label set corresponding to the preset content knowledge graph according to the target content and the target historical operation information, and the fusion unit 305 fuses the historical object label and the predicted object label to obtain the object label of the target object. Since the scheme mines the historical operation information of the target object and further uses the associated information of the preset content knowledge graph, the semantic inconsistency of the object label can be improved, the potential object label of the target object can be predicted while the historical object label is mined, and thus the accuracy of the object label generation can be improved.

[0170] The embodiment of the present application further provides an electronic device, as shown in the figure, which shows a structural schematic diagram of the electronic device related to the embodiment of the present application, and specifically: Figure 10

[0171] The electronic device can include a processor 401 with one or more processing cores, a memory 402 with one or more computer readable storage media, a power supply 403, an input unit 404, and the like. Those skilled in the art can understand that the electronic device structure shown in the figure does not constitute a limitation on the electronic device, and can include more or fewer components than the figure, or combine certain components, or different component arrangements. Among them: Figure 10

[0172] ​​The processor 401 is the control center of the electronic device, connects each part of the entire electronic device by various interfaces and lines, and performs various functions of the electronic device and processes data by running or executing software programs and / or modules stored in the memory 402 and calling data stored in the memory 402. Optionally, the processor 401 can include one or more processing cores; preferably, the processor 401 can integrate an application processor and a modem processor, wherein the application processor mainly processes the operating system, user interface, and application program, etc., and the modem processor mainly processes wireless communication. It can be understood that the above-mentioned modem processor can also not be integrated into the processor 401.

[0173] The memory 402 can be used to store software programs and modules, and the processor 401 executes various functions and data processing by running the software programs and modules stored in the memory 402. The memory 402 can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application program required by a function (such as a sound playing function, an image playing function, etc.), etc.; and the data storage area can store data created according to the use of the electronic device, etc. In addition, the memory 402 can include a high-speed random access memory, and can also include a non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other volatile solid-state memory device. Accordingly, the memory 402 can also include a memory controller to provide access for the processor 401 to the memory 402.

[0174] The electronic device also includes a power supply 403 for supplying power to each component, and preferably, the power supply 403 can be logically connected to the processor 401 through a power management system, so as to realize functions such as management of charging, discharging, and power consumption management through the power management system. The power supply 403 can also include one or more than one direct current or alternating current power supply, a recharging system, a power failure detection circuit, a power converter or inverter, a power state indicator, etc. Any component.

[0175] The electronic device can also include an input unit 404, which can be used to receive input digital or character information, and generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function control.

[0176] Although not shown, the electronic device can also include a display unit, etc., which will not be described here. Specifically, in the present embodiment, the processor 401 in the electronic device will load the executable file corresponding to the process of one or more than one application program into the memory 402 according to the following computer program, and run the application program stored in the memory 402 by the processor 401, so as to realize various functions, as follows:

[0177] The historical operation information of the at least one target object in a preset time period is acquired, target historical operation information corresponding to target content of a preset knowledge graph is screened out from the historical operation information, the historical object label of the target object is counted based on the target historical operation information and the preset knowledge graph, at least one predicted object label of the target object is determined from an object label set corresponding to the preset knowledge graph according to the target content and the target historical operation information, the historical object label and the predicted object label are fused, and the object label of the target object is obtained.

[0178] For example, the electronic device can acquire a historical operation log of at least one target object, read operation information in a preset time period from the historical operation log, and thus obtain the historical operation information of the target object, or can also acquire a historical operation log of at least one target object, screen out a target historical operation log in a preset time period from the historical operation log, and read operation information of the target object for content from the target historical operation log, and thus obtain the historical operation information of the target object. At least one content identifier is extracted from a preset content knowledge graph, and target content corresponding to the content identifier is identified from the historical operation information, initial historical operation information corresponding to the target content is screened out from the historical operation information, the initial historical operation information is cleaned, and target historical operation information corresponding to the target content is obtained. Graph relationships are extracted from the preset content knowledge graph, and operation frequencies of the target object for each target content are counted from the target historical operation information. Based on the operation frequencies and the graph relationships, the historical object label of the target object is determined. The text information of the target content is subjected to feature extraction, and content features of the target content are obtained. Based on the target historical operation information, the content features are converted into object features of the target object, at least one object label corresponding to the object features is screened out from an object label set corresponding to the preset content knowledge graph, and predicted object labels of the target object are obtained. A weighting coefficient corresponding to each object label type is acquired, and the historical interest degree is weighted based on the weighting coefficient and the operation frequency, to obtain a weighted historical interest degree of the historical object label. According to the weighting coefficient, the predicted interest degree is weighted to obtain a weighted predicted interest degree. The weighted historical interest degree and the weighted predicted interest degree are fused to obtain the object label of the target object.

[0179] The specific implementation of each operation can refer to the foregoing embodiments, which will not be repeated here.

[0180] From the above, the embodiment of the present application obtains the historical operation information of at least one target object in a preset time period, screens the target historical operation information for the target content corresponding to the preset content knowledge graph in the historical operation information, then, based on the target historical operation information and the preset content knowledge graph, the historical object label of the target object is counted, then, according to the target content and the target historical operation information, at least one predicted object label of the target object is determined in the object label set corresponding to the preset content knowledge graph, the historical object label and the predicted object label are fused to obtain the object label of the target object; Since the scheme is based on the premise of mining the historical operation information of the target object, and also with the aid of the associated information of the preset content knowledge graph, the problem of inconsistent semantics of the object label can be improved, and the potential object label of the object can be predicted while the historical object label is mined, so the accuracy of the object label generation can be improved.

[0181] Those skilled in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by a computer program, or by a computer program controlling related hardware, which can be stored in a computer readable storage medium and loaded and executed by a processor.

[0182] Therefore, the embodiment of the present application provides a computer readable storage medium, which stores a computer program capable of being loaded by a processor to execute the steps in any object label generation method provided by the embodiment of the present application. For example, the computer program can execute the following steps:

[0183] Obtain the historical operation information of at least one target object in a preset time period, screen the target historical operation information for the target content corresponding to the preset knowledge graph in the historical operation information, based on the target historical operation information and the preset knowledge graph, count the historical object label of the target object, according to the target content and the target historical operation information, determine at least one predicted object label of the target object in the object label set corresponding to the preset knowledge graph, and fuse the historical object label and the predicted object label to obtain the object label of the target object.

[0184] For example, the historical operation log of the at least one target object is acquired, operation information of a preset time period is read out from the historical operation log, and thus the historical operation information of the target object is obtained, or the historical operation log of the at least one target object is acquired, the target historical operation log of the preset time period is screened out from the historical operation log, and the operation information of the target object for the content is read out from the target historical operation log, and thus the historical operation information of the target object is obtained. The at least one content identifier is extracted from the preset content knowledge graph, and the target content corresponding to the content identifier is identified in the historical operation information, the initial historical operation information corresponding to the target content is screened out from the historical operation information, the initial historical operation information is cleaned, and thus the target historical operation information corresponding to the target content is obtained. The graph relationship is extracted from the preset content knowledge graph, the operation frequency of the target object for each target content is counted in the target historical operation information, and the historical object label of the target object is determined based on the operation frequency and the graph relationship. The text information of the target content is subjected to feature extraction, and thus the content feature of the target content is obtained, the content feature is converted into the object feature of the target object based on the target historical operation information, at least one object label corresponding to the object feature is screened out from the object label set corresponding to the preset content knowledge graph, and thus the predicted object label of the target object is obtained. The weighting coefficient corresponding to each object label type is acquired, the historical interest degree is weighted based on the weighting coefficient and the operation frequency, and thus the weighted historical interest degree of the historical object label is obtained, the predicted interest degree is weighted according to the weighting coefficient, and thus the weighted predicted interest degree is obtained, and the weighted historical interest degree and the weighted predicted interest degree are fused, so as to obtain the object label of the target object.

[0185] The specific implementation of each operation can refer to the foregoing embodiments, and will not be described herein again.

[0186] The computer readable storage medium can include a read only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0187] Due to the computer program stored in the computer readable storage medium, the steps of any object label generation method provided in the embodiments of the present application can be executed, and thus the beneficial effects of any object label generation method provided in the embodiments of the present application can be achieved. Details can refer to the foregoing embodiments, and will not be described herein again.

[0188] According to an aspect of the present application, a computer program product is provided, which includes a computer program stored in a computer readable storage medium. A processor of an electronic device reads the computer program from the computer readable storage medium, and the processor executes the computer program, so that the computer device executes the method provided in various optional implementations of the object label generation aspect or the object label determination aspect.

[0189] The object label generation method, device and computer readable storage medium provided by the embodiments of the present application are described in detail above, and the principles and implementation manners of the present application are described by applying specific examples. The above embodiment description is only used to help understand the method of the present application and its core idea. Meanwhile, for those skilled in the art, according to the idea of the present application, the specific implementation manner and application range will be changed, and the above description should not be understood as a limitation of the present application.

Claims

1. An object label generation method characterized by, The method comprises the following steps: acquiring historical operation information of at least one target object in a preset time period; screening target historical operation information corresponding to a target content of a preset content knowledge graph from the historical operation information; based on the target historical operation information and the preset content knowledge graph, counting the historical object label of the target object, comprising: extracting a graph relationship in the preset content knowledge graph, the graph relationship being used to indicate the relationship between the target content and a candidate historical object label; counting the operation frequency of the target object for each target content in the target historical operation information; based on the operation frequency and the graph relationship, determining the historical object label of the target object; determining at least one predicted object label of the target object in the object label set corresponding to the preset content knowledge graph according to the target content and the target historical operation information, comprising: performing feature extraction on the text information of the target content to obtain content features of the target content; converting the content features into object features of the target object based on the target historical operation information; acquiring an object label set corresponding to the preset content knowledge graph, and extracting triple information in the preset content knowledge graph; according to the triple information, performing feature extraction on the object labels in the object label set by using a trained object label prediction model to obtain object label features of at least one candidate object label; calculating the feature similarity between the object features and the object label features to obtain the predicted object label of the target object; fusing the historical object label and the predicted object label to obtain the object label of the target object.

2. The object label generation method according to claim 1, characterized by, The method further comprises the following steps: based on the target historical operation information, screening at least one content feature of the target content operated by the target object in a preset time interval from the content features to obtain basic content features; performing maximum pooling processing on the basic content features by using a trained object feature generation model to obtain target pooled content features; performing nonlinear conversion on the target pooled content features to obtain the object features of the target object.

3. The object label generation method according to claim 2, characterized by, The method further comprises the following steps before performing maximum pooling processing on the basic content features by using a trained object feature generation model to obtain target pooled content features: acquiring a sample content feature set corresponding to a historical content sample set operated by the target object, the sample content feature set comprising at least one sample content feature with an annotated classification result; determining at least one target sample content feature in the sample content feature set, and regarding the sample content features in the sample content feature set except the target sample content features as classification sample content features of classification sample content; based on the annotated classification result, the target sample content feature and the classification sample content feature, converging a preset object feature generation model to obtain a trained object feature generation model.

4. The object label generation method according to claim 3, characterized by, The preset object feature generation model is converged based on the annotation classification result, the target sample content feature and the to-be-classified sample content feature, and a trained object feature generation model is obtained, including: The preset object feature generation model is used for feature conversion on the target sample content feature, and a sample object feature of the target object is obtained; The preset object feature generation model is used for feature conversion on the to-be-classified sample content feature, and a classification sample content feature of the to-be-classified sample content is obtained; Based on the sample object feature and the classification sample content feature, a classification result of the target object for the to-be-classified sample content is predicted, and a predicted classification result is obtained; According to the predicted classification result and the annotation classification result, the preset object feature generation model is converged, and a trained object feature generation model is obtained.

5. The object label generation method according to claim 1, characterized by, According to the triple information, the object label feature corresponding to at least one candidate object label of the object feature is obtained by using the trained object label prediction model to perform feature extraction on the object label in the object label set, including: The trained object label prediction model is used to filter out at least one candidate object label corresponding to the object feature in the object label set, and the feature of the candidate object label is extracted to obtain the basic object label feature of the candidate object label; The trained object label prediction model is used to perform feature extraction on the triple information to obtain the basic entity feature and the entity relationship feature, and the basic entity feature and the entity relationship feature are fused to obtain the entity feature; The entity feature and the basic object label feature are fused to obtain the object label feature of the candidate object label.

6. The object label generation method according to claim 5, characterized by, Before the trained object label prediction model is used to filter out at least one candidate object label corresponding to the object feature in the object label set, it further includes: A target object label sample corresponding to a target object is obtained, and based on the target object label sample, a negative sample of an object label is randomly sampled in the object label set to obtain an object label sample set; Triple information samples are extracted in a preset content knowledge graph, and triple information negative samples are constructed based on the triple information samples to obtain a triple information sample set; The preset object label prediction model is converged based on the object label sample set and the triple information sample set to obtain a trained object label prediction model.

7. The object label generation method according to claim 6, characterized in that, The preset object label prediction model is converged based on the object label sample set and the triple information sample set to obtain a trained object label prediction model, including: Based on the object label sample set, object label loss information of the target object is determined; According to the triple information sample set, triple loss information of the target object is determined; Based on the object label loss information and the triple loss information, the network parameters of the preset object label prediction model are updated, and the updated network parameters are regularized to obtain the trained object label prediction model.

8. The object label generation method of claim 1, wherein, The calculating the feature similarity between the object feature and the object label feature to obtain the predicted object label of the target object comprises: calculating the feature similarity between the object feature and the object label feature, and taking the feature similarity as a predicted interest degree of the target object to the object label corresponding to the object label feature; ranking the candidate object labels based on the predicted interest degree; screening the predicted object label of the target object from the candidate object labels according to the ranking information of the candidate object labels.

9. The object label generation method according to claim 8, characterized by, The determining the historical object label of the target object based on the operation frequency and the graph relationship comprises: identifying candidate historical object labels corresponding to each target content in the graph relationship; ranking the candidate historical object labels based on the operation frequency; screening the historical object label of the target object from the candidate historical object labels according to the ranking information of the candidate historical object labels, and determining a historical interest degree of each historical object label.

10. The object label generation method according to claim 9, characterized in that, The fusing the historical object label and the predicted object label to obtain the object label of the target object comprises: obtaining a weighting coefficient corresponding to each object label type, and weighting the historical interest degree based on the weighting coefficient and the operation frequency to obtain a weighted historical interest degree of the historical object label; weighting the predicted interest degree based on the weighting coefficient to obtain a weighted predicted interest degree; fusing the weighted historical interest degree and the weighted predicted interest degree to obtain the object label of the target object.

11. The object label generation method according to claim 10, characterized in that, The fusing the weighted historical interest degree and the weighted predicted interest degree to obtain the object label of the target object comprises: fusing the weighted historical interest degree and the weighted predicted interest degree to obtain a fused interest degree; normalizing the fused interest degree to obtain a target historical interest degree of the historical object label and a target predicted interest degree of the predicted object label; screening the object label of the target object from the historical object label and the predicted object label based on the target historical interest degree and the target predicted interest degree.

12. The object label generating method according to claim 10, wherein The weighting the historical interest degree based on the weighting coefficient and the operation frequency to obtain the weighted historical interest degree of the historical object label comprises: screening a target historical object label from the historical object label according to the historical interest degree; screening a target operation frequency corresponding to the target historical object label from the operation frequency, and determining an operation parameter of the target historical object based on the target operation frequency; fusing the operation parameter and the historical interest degree of the target historical object label to obtain a fused historical interest degree, and weighting the fused historical interest degree based on the weighting coefficient to obtain the weighted historical interest degree of the historical object label.

13. The object label generating method according to any one of claims 1 to 8, characterized by, The screening target historical operation information corresponding to a target content of a preset content knowledge graph from the historical operation information comprises: extract at least one content identifier from the preset content knowledge graph, and identify the target content corresponding to the content identifier in the historical operation information; screen out initial historical operation information corresponding to the target content from the historical operation information; clean the initial historical operation information to obtain target historical operation information corresponding to the target content.

14. An object label generating apparatus characterized by comprising: Comprise: An acquisition unit is configured to acquire historical operation information of at least one target object in a preset time period; A screening unit is configured to screen out target historical operation information corresponding to target content of a preset content knowledge graph from the historical operation information; A statistical unit is configured to statistically obtain historical object labels of the target object based on the target historical operation information and the preset content knowledge graph, including: extracting graph relationships from the preset content knowledge graph, the graph relationships being used to indicate relationships between the target content and candidate historical object labels; statistically obtaining operation frequencies of the target object for each target content in the target historical operation information; and determining historical object labels of the target object based on the operation frequencies and the graph relationships; A determination unit is configured to determine at least one predicted object label of the target object in an object label set corresponding to the preset content knowledge graph according to the target content and the target historical operation information, including: performing feature extraction on text information of the target content to obtain content features of the target content; converting the content features into object features of the target object based on the target historical operation information; obtaining an object label set corresponding to the preset content knowledge graph, and extracting triple information from the preset content knowledge graph; performing feature extraction on object labels in the object label set by using a trained object label prediction model according to the triple information, to obtain object label features of at least one candidate object label corresponding to the object features; and calculating feature similarity between the object features and the object label features, to obtain predicted object labels of the target object; A fusion unit is configured to fuse the historical object labels and the predicted object labels to obtain object labels of the target object.

15. The object label generation apparatus according to claim 14, wherein The determination unit is specifically configured to screen out, based on the target historical operation information, content features of at least one target content operated by the target object in a preset time interval from the content features, to obtain basic content features; perform maximum pooling processing on the basic content features by using a trained object feature generation model, to obtain target pooled content features; perform nonlinear conversion on the target pooled content features, to obtain the object features of the target object.

16. The object label generation apparatus according to claim 15, wherein The determination unit is specifically configured to acquire sample content feature sets corresponding to a historical content sample set operated by the target object, the sample content feature sets including at least one sample content feature with an annotated classification result. determine at least one target sample content feature in the sample content feature set, and determine sample content features in the sample content feature set other than the target sample content feature as to-be-classified sample content features of to-be-classified sample content; converge the preset object feature generation model based on the labeled classification result, the target sample content feature, and the to-be-classified sample content feature, to obtain a trained object feature generation model. 17.The object label generation apparatus of claim 16, characterized in that, the determination unit is specifically configured to: perform feature conversion on the target sample content feature by using the preset object feature generation model, to obtain a sample object feature of the target object; and perform feature conversion on the to-be-classified sample content feature by using the preset object feature generation model, to obtain a classification sample content feature of the to-be-classified sample content; predict a classification result of the target object for the to-be-classified sample content based on the sample object feature and the classification sample content feature, to obtain a predicted classification result; and converge the preset object feature generation model based on the predicted classification result and the labeled classification result, to obtain a trained object feature generation model. 18.The object label generation apparatus of claim 14, characterized in that, the determination unit is specifically configured to: filter at least one candidate object label corresponding to the object feature from the object label set by using the trained object label prediction model, and perform feature extraction on the candidate object label, to obtain a basic object label feature of the candidate object label; perform feature extraction on the triple information by using the trained object label prediction model, to obtain a basic entity feature and an entity relationship feature, and fuse the basic entity feature and the entity relationship feature, to obtain an entity feature; and fuse the entity feature and the basic object label feature, to obtain an object label feature of the candidate object label. 19.The object label generation apparatus of claim 18, characterized in that, the determination unit is specifically configured to: obtain a target object label sample corresponding to a target object, and randomly negatively sample an object label negative sample from the object label set based on the target object label sample, to obtain an object label sample set; extract a triple information sample from a preset content knowledge graph, and construct a triple information negative sample based on the triple information sample, to obtain a triple information sample set; and converge a preset object label prediction model based on the object label sample set and the triple information sample set, to obtain a trained object label prediction model. 20.The object label generation apparatus of claim 19, characterized in that, The determining unit is specifically configured to determine object label loss information of the target object based on the set of object label samples, determine triple loss information of the target object according to the set of triple information samples, update network parameters of the preset object label prediction model based on the object label loss information and the triple loss information, and perform regularization processing on the updated network parameters to obtain the trained object label prediction model. 21.The object label generation apparatus of claim 14, wherein, The determining unit is specifically configured to calculate a feature similarity between the object feature and the object label feature, and take the feature similarity as a predicted interest degree of the target object to the object label corresponding to the object label feature; sort the candidate object labels based on the predicted interest degree; and filter out a predicted object label of the target object from the candidate object labels according to sorting information of the candidate object labels. 22.The object label generation apparatus of claim 21, wherein, The statistical unit is specifically configured to identify a candidate historical object label corresponding to each target content in the graph relationship; sort the candidate historical object labels based on the operation frequency; filter out a historical object label of the target object from the candidate historical object labels according to sorting information of the candidate historical object labels, and determine a historical interest degree of each historical object label. 23.The object label generation apparatus of claim 22, wherein, The fusion unit is specifically configured to obtain a weighting coefficient corresponding to each object label type, weight the historical interest degrees based on the weighting coefficient and the operation frequency to obtain weighted historical interest degrees of the historical object labels, weight the predicted interest degrees based on the weighting coefficient to obtain weighted predicted interest degrees, and fuse the weighted historical interest degrees and the weighted predicted interest degrees to obtain the object label of the target object. 24.The object label generation apparatus of claim 23, wherein, The fusion unit is specifically configured to fuse the weighted historical interest degrees and the weighted predicted interest degrees to obtain fused interest degrees, perform normalization processing on the fused interest degrees to obtain a target historical interest degree of the historical object label and a target predicted interest degree of the predicted object label; and filter out the object label of the target object from the historical object label and the predicted object label based on the target historical interest degree and the target predicted interest degree. 25.The object label generation apparatus of claim 23, wherein, ​ The fusion unit is specifically configured to filter out a target historical object label from the historical object labels according to the historical interest degrees, filter out a target operation frequency corresponding to the target historical object label from the operation frequencies, determine an operation parameter of the target historical object based on the target operation frequency, fuse the operation parameter and the historical interest degree of the target historical object label, obtain a fused historical interest degree, and weight the fused historical interest degree based on the weighting coefficient to obtain a weighted historical interest degree of the historical object label. 26.The object label generation apparatus of any one of claims 14-21, wherein, The screening unit is specifically configured to extract at least one content identifier from the preset content knowledge graph, identify a target content corresponding to the content identifier in the historical operation information, screen out initial historical operation information corresponding to the target content from the historical operation information, and clean the initial historical operation information to obtain target historical operation information corresponding to the target content.

27. An electronic device, comprising: The computer readable storage medium stores a computer program, and the computer program is adapted to be loaded by the processor to execute the steps in the object label generation method of any one of claims 1-13.

28. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the steps in the object label generation method of any one of claims 1-13.

29. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program is adapted to be loaded by the processor to execute the steps in the object label generation method of any one of claims 1-13.

Images