Object label identification method and device, computer device, readable storage medium and program product

Abstract

This application relates to an object tag recognition method, apparatus, computer device, computer-readable storage medium, and computer program product. The method includes: acquiring descriptive text of a target object; performing feature extraction processing on the descriptive data to obtain an embedded feature vector; performing tag classification processing on the embedded feature vector using different tag classification heads to obtain tag classification results for the embedded feature vector in multiple tag dimensions; searching for associated object tag information of the target object on social networks; and then filtering the tag classification results to obtain the object tags of the target object. In this embodiment, multi-task text classification is used to decouple the recognition of different identity tags corresponding to the target object, improving the efficiency of the identity tag recognition process. Furthermore, combining the social relationships of the target object for further recognition processing of its object tags effectively improves the accuracy of object tag recognition.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to an object tag identification method, apparatus, computer device, computer-readable storage medium, and computer program product. Background Technology

[0002] With the development of computer technology, object tagging technology has emerged. Object tagging technology links the attributes and behaviors of a target object with expected data transformations by assigning descriptive tags to the target object, serving as a virtual representative of the target object. It has wide applications in multiple fields. For example, in e-commerce, user object tags can help merchants accurately recommend products, improving user satisfaction and conversion rates; in advertising, user object tags can make ad targeting more precise, improving advertising effectiveness and return on investment (ROI).

[0003] Currently, it is possible to construct object-related behavior vectors by collecting and statistically analyzing various behavioral characteristics, and then identify the descriptive labels of target objects by classifying the behavior vectors using classification algorithms. However, this method depends on behavioral characteristics, resulting in insufficient recognition efficiency and accuracy. Summary of the Invention

[0004] Therefore, it is necessary to provide an object tag recognition method, apparatus, computer equipment, computer-readable storage medium, and computer program product that can improve the efficiency and accuracy of object tag recognition in response to the above-mentioned technical problems.

[0005] Firstly, this application provides an object label recognition method, including:

[0006] Get the description text of the target object;

[0007] The descriptive text is subjected to feature extraction processing to obtain the embedded feature vector of the descriptive text;

[0008] By performing label classification processing on the embedded feature vector using different label classification heads, the label classification results of the embedded feature vector in multiple label dimensions are obtained.

[0009] Find the associated object tag information of the target object on the social network;

[0010] Based on the associated object tag information, the tag classification results are filtered to obtain the object tags of the target object.

[0011] Secondly, this application also provides an object tag recognition device, comprising:

[0012] The data acquisition module is used to acquire the descriptive text of the target object;

[0013] The feature extraction module is used to perform feature extraction processing on the descriptive text to obtain the embedded feature vector of the descriptive text;

[0014] The multi-classification processing module is used to perform label classification processing on the embedded feature vector through different label classification heads to obtain the label classification results of the embedded feature vector in multiple label dimensions.

[0015] The association analysis module is used to find the associated object tag information of the target object on the social network;

[0016] The object tag recognition module is used to filter the tag classification results based on the associated object tag information to obtain the object tag of the target object.

[0017] In one embodiment, the association analysis module is specifically used to: obtain a relationship network graph of the target object in a social application; identify the associated objects of the target object in the social network based on the relationship network graph; and find the associated object tag information of the associated objects.

[0018] In one embodiment, the association analysis module is further configured to: identify the first-order neighboring entities of the target object in the social network based on the relationship network graph; and perform sampling processing on the first-order neighboring entities to obtain the associated objects of the target object in the social network.

[0019] In one embodiment, the association analysis module is further configured to: obtain object description text of the associated object; perform feature extraction processing on the object description text to obtain object embedding feature vector of the object description text; perform label classification processing on the object embedding feature vector through different label classification heads to obtain multiple object label classification results of the object embedding feature vector; and determine the associated object label information of the associated object based on the multiple object label classification results.

[0020] In one embodiment, the object tag recognition module is specifically used to: determine the social relationship tag distribution information of the target object based on the associated object tag information; compare the distribution probability values ​​of each associated object tag in the social relationship tag distribution information, and filter multiple target object tags from the associated object tags based on the comparison results; and compare the target object tags with the tag classification results to obtain the object tags of the target object.

[0021] In one embodiment, the object tag recognition module is further configured to: determine the identity tag of the target object based on the tag classification result; when the target object has multiple identity tags, determine a target identity tag that is consistent with the target object tag from the identity tags, and determine the object tag of the target object from the target identity tags based on the distribution probability of the target object tags; when the target object has a single identity tag, determine the single identity tag as the object tag of the target object; when the target object does not have an identity tag, take the tag with the highest distribution probability among the target object tags as the object tag of the target object.

[0022] In one embodiment, the apparatus further includes a classification head training module, configured to: acquire a text training dataset corresponding to different object labels; and train an initial classification head based on the text training dataset to obtain a label classification head corresponding to each object label.

[0023] In one embodiment, the classification head training module is specifically used for: obtaining a text processing pre-trained model; combining the text processing pre-trained model with the initial classification head to obtain a single-task training model; inputting the text training dataset into the single-task training model to obtain the model loss value corresponding to the sample data in the text training dataset; performing backpropagation processing based on the model loss value; freezing the parameters of the text processing pre-trained model during backpropagation and updating the parameters of the initial classification head to obtain the label classification head corresponding to each object label.

[0024] In one embodiment, the multi-classification processing module is specifically used to: read the pre-stored network configuration file corresponding to the target object; and perform label classification processing on the embedded feature vector based on the different label classification heads determined by the pre-stored network configuration file to obtain the label classification results of the embedded feature vector in multiple label dimensions.

[0025] In one embodiment, an exception handling module is further included, configured to: grant exemption permissions to the target object if the target object is determined to be a normal object based on the object tag of the target object; generate an exception risk identifier corresponding to the target object if the target object is determined to be an abnormal object based on the object tag of the target object; and generate a list of risk objects based on the object's permission information and identifier information.

[0026] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0027] Get the description text of the target object;

[0028] The descriptive text is subjected to feature extraction processing to obtain the embedded feature vector of the descriptive text;

[0029] By performing label classification processing on the embedded feature vector using different label classification heads, the label classification results of the embedded feature vector in multiple label dimensions are obtained.

[0030] Find the associated object tag information of the target object on the social network;

[0031] Based on the associated object tag information, the tag classification results are filtered to obtain the object tags of the target object.

[0032] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0033] Get the description text of the target object;

[0034] The descriptive text is subjected to feature extraction processing to obtain the embedded feature vector of the descriptive text;

[0035] By performing label classification processing on the embedded feature vector using different label classification heads, the label classification results of the embedded feature vector in multiple label dimensions are obtained.

[0036] Find the associated object tag information of the target object on the social network;

[0037] Based on the associated object tag information, the tag classification results are filtered to obtain the object tags of the target object.

[0038] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0039] Get the description text of the target object;

[0040] The descriptive text is subjected to feature extraction processing to obtain the embedded feature vector of the descriptive text;

[0041] By performing label classification processing on the embedded feature vector using different label classification heads, the label classification results of the embedded feature vector in multiple label dimensions are obtained.

[0042] Find the associated object tag information of the target object on the social network;

[0043] Based on the associated object tag information, the tag classification results are filtered to obtain the object tags of the target object.

[0044] The aforementioned object tag recognition method, apparatus, computer equipment, computer-readable storage medium, and computer program product first obtain the descriptive text of the target object to acquire the basic data for the tag recognition process. Then, feature extraction processing is performed on the descriptive text to obtain embedded feature vectors. Different tag classification heads are used to classify the embedded feature vectors, resulting in tag classification results for each embedded feature vector across multiple tag dimensions. By directly classifying the embedded feature vectors using multiple tag classification heads instead of multiple binary tag classification models, the processing efficiency of the tag classification process is improved. Finally, the associated object tag information of the target object on social networks is searched; and the tag classification results are filtered based on the associated object tag information to obtain the object tags of the target object. This further filters the tag classification results by combining the social relationships of the target object, ensuring the accuracy of the tag recognition process. This application achieves decoupling of the recognition of different identity tags corresponding to the target object through multi-task text classification, improving the efficiency of the identity tag recognition process. Furthermore, by combining the social relationships of the target object for further recognition processing of its object tags, the accuracy of object tag recognition is effectively improved. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is an application environment diagram of the object label recognition method in one embodiment;

[0047] Figure 2 This is a flowchart illustrating an object label recognition method in one embodiment;

[0048] Figure 3 This is a flowchart of the multi-task reasoning steps in one embodiment;

[0049] Figure 4 This is a schematic diagram of the tag distribution information processing steps in one embodiment;

[0050] Figure 5 This is a schematic diagram of a single-task training step in one embodiment;

[0051] Figure 6This is a schematic diagram illustrating the step of configuring the tag classification header through a pre-stored network configuration file in one embodiment;

[0052] Figure 7 This is a flowchart of the steps from single-task training to multi-task inference in one embodiment;

[0053] Figure 8 This is a structural block diagram of an object label recognition device in one embodiment;

[0054] Figure 9 This is an internal structural diagram of a computer device in one embodiment;

[0055] Figure 10 This is a diagram of the internal structure of a computer device in another embodiment. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0057] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0058] The object tag recognition method provided in this application embodiment can be applied to, for example, Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or placed on a cloud or other network server. When a user on terminal 102 wants to identify the identity tag of a specified target object, they can submit the target object's description text to server 104. Server 104 then performs the relevant processing for object tag identification. Server 104 obtains the target object's description text; performs feature extraction processing on the description text to obtain the embedded feature vector; performs tag classification processing on the embedded feature vector using different tag classification heads to obtain the tag classification results of the embedded feature vector in multiple tag dimensions; searches for the target object's associated object tag information on social networks; and filters the tag classification results based on the associated object tag information to obtain the target object's object tag. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, projection devices, etc. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted displays. Head-mounted displays can be virtual reality (VR) devices, augmented reality (AR) devices, smart glasses, etc. Server 104 can be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server providing cloud computing services.

[0059] In one exemplary embodiment, such as Figure 2 As shown, an object label recognition method is provided, which can be applied to... Figure 1 Taking server 104 as an example, the explanation includes the following steps 201 to 209. Wherein:

[0060] Step 201: Obtain the description text of the target object.

[0061] The target object is the object to be identified by the object tag recognition method in this application. The object tag of the target object can be identified through various related data. The target object can be specifically represented in the form of a subject, which refers to various core elements, including but not limited to: application accounts, application groups, and video account accounts. The descriptive text refers to the text data obtained by summarizing and transforming the various basic data of the target object. The descriptive text can include static descriptive data describing the static attributes of the target object, and dynamic descriptive data describing the specific behaviors of the target object. After obtaining the original descriptive data of the target object, this original descriptive data can be anonymized using methods such as data replacement, and then the anonymized data can be used as the descriptive text.

[0062] For example, when a user on terminal 102 wants to identify the object tags of a target object for analysis, they can send a request to server 104 via terminal 102, carrying relevant information about the target object in the request, such as descriptive text of the target object or the target object's identity identifier in the system. Server 104, upon receiving the request, can perform object tag identification processing, specifically based on the descriptive text carried in the request. Alternatively, it can search for relevant descriptive data of the target object in the system based on the object identifier carried in the request, then generate the descriptive text of the target object, and then perform object tag identification processing. In a specific embodiment, this application is specifically used in the field of system risk control, and the object tags for the target object can specifically include abnormal objects and normal objects, etc. In this case, basic information filled in by the object when registering in the system, as well as the object's operation record information in the system, can be collected, and then the collected information can be anonymized to generate descriptive text for tag identification. In another embodiment, this application is applied to the field of system content recommendation, where the historical content browsed by the target object through the system can be collected, and then, after data anonymization processing, descriptive text corresponding to this historical browsing data can be generated for object tag identification.

[0063] Step 203: Perform feature extraction processing on the descriptive text to obtain the embedded feature vector of the descriptive text.

[0064] Feature extraction refers to extracting representative information from raw data to enable better training and learning of the model. Feature extraction plays a crucial role in machine learning and data analysis, significantly improving model performance through dimensionality reduction, increased accuracy, reduced overfitting, enhanced interpretability, and accelerated processing. In this application, the scheme involves processing descriptive text using a relevant machine learning model to obtain corresponding feature vector data. Embedding vectors are a method of mapping objects in a high-dimensional space to a low-dimensional space. In natural language processing and machine learning, embedding vectors are commonly used to represent words, sentences, or other text data. In this application, the descriptive text is converted into a fixed-length embedding feature vector, thereby mapping objects with similar semantics to similar vector spaces. In this way, computers can effectively perform similarity comparisons, thereby improving the performance of tasks such as search and classification.

[0065] For example, this application specifically uses machine learning to achieve corresponding tag recognition processing for the acquired descriptive text. To process unstructured descriptive text more efficiently, feature extraction can be used. The descriptive text is input into a pre-trained text data processing model, which extracts features from the descriptive text to obtain embedded feature vectors. The length of the embedded feature vectors is fixed for different descriptive texts. These embedded feature vectors can effectively identify various tags of the target object. In a specific embodiment, a pre-trained text model can be used to extract text features. The machine learning model is repeatedly trained using a large amount of collected text data to obtain a pre-trained text model that can effectively extract features from the text.

[0066] Step 205: Perform label classification processing on the embedded feature vector using different label classification heads to obtain the label classification results of the embedded feature vector in multiple label dimensions.

[0067] In this context, the label classification head refers to the head used for binary classification. The classification head plays a crucial role in object detection networks, identifying the category of the target. In object detection tasks, the classification head is responsible for outputting the category information of the target. For example, the classification head can output the category probability of each candidate box through convolutional layers, thereby determining which category the target belongs to. For instance, for a label classification head for label A, it can determine the probability of the embedded feature vector corresponding to the label A based on the embedded feature vector, and then perform classification based on a preset label threshold. For example, if the label threshold is set to a probability higher than 0.4, indicating that the embedded feature vector can be classified into the category with label A, then during the label classification process, if the probability predicted by the classification head is 0.5, which is higher than the label threshold, it means that the embedded feature vector is classified into the category with label A. If the probability predicted by the classification head is 0.3, which is lower than the label threshold, it means that the embedded feature vector cannot be classified into the category with label A. In different label classification heads, each label classification head corresponds to a label dimension, and the label classification result of the embedded feature vector in that label dimension can be determined through a single label classification head.

[0068] For example, this application can specifically perform target object-related label classification processing through multi-task label recognition. During label classification, basic data is collected in advance for different label types to train label classification heads. During training, the label classification heads can be connected to a general feature extraction backbone network to train label classification heads corresponding to different label dimensions. In the specific object label recognition process, the appropriate label classification head can be selected for the target object to perform label recognition processing. By inputting the embedded feature vectors into different label classification heads, the label classification heads perform label classification processing, obtaining the label classification results for each label classification head, that is, the label classification results of the embedded feature vectors in multiple label dimensions. In one embodiment, multiple different label classification heads can be pre-trained, and then the label classification head is selected for label classification processing according to the specific object label recognition scenario. For example, if five label classification heads A, B, C, D, and E are pre-trained, in a risk recognition scenario, only label classification heads A and B need to be selected for recognition. In this case, label classification heads A and B can be added to the end of the feature extraction network to form a complete model. In another content recommendation scenario, three tag classification heads (B, C, and D) are needed for identification. In this case, these three tag classification heads need to be added to the feature extraction network to form the model. In one embodiment, the model network suitable for tag processing can refer to... Figure 3As shown, after obtaining the descriptive text as the basic inference data, it is input into the model. First, feature extraction is performed through the backbone network to obtain the embedded feature vector. Specifically, a pre-trained text processing model suitable for text processing can be selected as the backbone network for feature extraction. Then, through the full task setting module, the label classification heads selected for the current target object are determined to be three classification heads: Label Classification Head 1, Label Classification Head 2, and Label Classification Head 3. At this point, the three classification heads can be connected to the lower end of the model, and the embedded features are input into the three classification heads respectively for classification and recognition processing, and the corresponding label classification results are output.

[0069] Step 207: Find the associated object tag information of the target object on the social network.

[0070] Social Network Service (SNS) refers to the networking of interpersonal relationships achieved through internet platforms. It includes various social networking software such as blogs, wikis, tags, and RSS, which construct a virtual social network service platform. In the scheme of this application, it refers to a social relationship network related to a target object identified through a social application. For example, the target object's account on the social application can be considered a subject, while other accounts that are friends with that account can be considered first-order adjacent subjects, and friends of friends are second-order adjacent subjects. All subjects together constitute the social network. Associated objects refer to other objects in the social network that are associated with the current target object. Associated object tag information is the specific information of the tags of other objects. Associated object tag information can be extracted from the descriptive text corresponding to the associated objects through a tag classification process similar to that described above.

[0071] For example, since there can be multiple tag classification results, multiple different tags can be added to the target object based on these results. To more accurately identify the object tag that best matches the target object's own attributes, verification processing can be performed by combining the tag classification results with social networks. Furthermore, the applicant also considered that tag identification for objects with rich content is relatively easy to achieve, while predicting the identity tags of newly registered accounts and alternate accounts is very difficult. However, for social platforms, the behavior of new user registration or creating alternate accounts is continuous, making this situation common and unavoidable. Therefore, how to perform tag identification during cold starts is a problem that needs to be solved. At this point, tag classification results combined with the target object's social network can be used to achieve object tag identification in cold start situations. When combining social networks for analysis, firstly, for the current target object, the associated object tag information of its related objects on social networks is searched. Specifically, for the current target object, its related objects on social networks can be searched first, and then for each related object, the corresponding tag classification result can be extracted from the descriptive text of the related object through the above-mentioned tag classification process, thus obtaining the related object tag information.

[0072] Step 209: Filter the tag classification results based on the associated object tag information to obtain the object tags of the target object.

[0073] For example, when identifying object tags, the filtering can be performed by combining associated object tag information and tag classification results. This is because when a target object possesses a certain tag potential, other adjacent objects on its social network may also exhibit the same tag potential. For instance, for an anomalous object, other adjacent objects on its social network may also be anomalous objects. Therefore, the tag classification results can be filtered based on associated object tag information, thereby selecting the object tags that best match the target object's attributes from the tag classification results. In another embodiment, when there are multiple tag classification results, the tag classification results can also be sorted based on associated object tag information, thereby assigning multiple object tags to the target object and determining the order between the various object tags.

[0074] The aforementioned object label recognition method first obtains the descriptive text of the target object to acquire the basic data for the label recognition process. Then, it performs feature extraction on the descriptive text to obtain embedded feature vectors. Different label classification heads are used to classify the embedded feature vectors, resulting in label classification results for each embedded feature vector across multiple label dimensions. By directly classifying the embedded feature vectors using multiple label classification heads instead of multiple binary classification models, the processing efficiency of the label classification process is improved. Finally, the method searches for the associated object label information of the target object on social networks and filters the label classification results based on this information to obtain the object labels of the target object. This further filters the label classification results by incorporating the target object's social relationships, ensuring the accuracy of the label recognition process. This application achieves decoupling of the recognition of different identity labels corresponding to the target object through multi-task text classification, improving the efficiency of the identity label recognition process. Furthermore, by combining the target object's social relationships for further object label recognition processing, the accuracy of object label recognition is effectively improved.

[0075] In one exemplary embodiment, step 207 includes: obtaining a relationship network graph of the target object in a social application; identifying associated objects of the target object in the social network based on the relationship network graph; and finding the associated object tag information of the associated objects.

[0076] Among them, the relationship network diagram is a concretization of social relationships. By obtaining the relationship network diagram of the target object in the social application, the intangible social relationships can be transformed into a tangible relationship network diagram, thereby determining the social relationships between different objects through the relationship network diagram.

[0077] For example, since the relationships between entities in a social application can reflect the social relationships of different objects, this application can specifically perform tag recognition processing based on the relationships between the target object and the entities in the social application. In the processing, firstly, a relationship network graph of the target object in the social application can be obtained. Specifically, for the target object, its friend list in the social application can be searched, and then each friend is treated as an object to build a social network graph for the target object. To further expand the social network graph, for each friend, the friends of those friends can be searched and added as new objects to the social network graph. Then, based on the relationship network graph, the associated objects of the target object in the social network can be identified. Specifically, depending on the application scenario, first-order or first-order and second-order entities can be selected as associated objects for calculation. Then, the tag information of each associated object is searched. In one embodiment, this application is applicable to extracting the corresponding object tags for each target object in the social application. At this time, a relationship network graph containing all objects can be established for multiple objects in the social application, and then the descriptive text of each object is obtained, and then the tag classification result of each object is extracted. In the object tag recognition process, for each object, its own tag classification result can be re-authenticated by using the tag classification results of related objects in the relationship network graph, thereby improving the accuracy of tag recognition. In this embodiment, the related objects of the target object are identified by using the relationship network graph of the target object's social applications, and then the tag information related to the related objects is determined. This can effectively improve the accuracy of related object identification, thus ensuring the tag recognition effect.

[0078] Furthermore, identifying the associated objects of a target object in a social network based on a relationship network graph includes: identifying the first-order neighboring entities of the target object in the social network based on the relationship network graph; and sampling these first-order neighboring entities to obtain the associated objects of the target object in the social network. First-order neighboring entities refer to objects in the relationship network graph that have a direct relationship with the target object, such as a friend relationship between two entities, or, in enterprise-type social applications, two entities belonging to the same department. In this application's solution, the first-order neighboring entities of the target object can be identified through direct relationships in the relationship network graph. Since the number of first-order neighboring entities may vary for different objects, to standardize the calculation and simplify the process, sampling can be performed on the first-order neighboring entities to determine the associated objects of the target object in the social network. The sampling method can be the TOPN algorithm, selecting N first-order neighboring entities as associated objects for each target object. In this embodiment, by identifying first-order neighboring entities and sampling to identify the associated objects of the target object, the accuracy and efficiency of associated object identification can be effectively guaranteed.

[0079] Further, finding the associated object label information of related objects includes: obtaining the object description text of the related objects; performing feature extraction processing on the object description text to obtain the object embedding feature vector of the object description text; performing label classification processing on the object embedding feature vector through different label classification heads to obtain multiple object label classification results of the object embedding feature vector; and determining the associated object label information of the related objects based on the multiple object label classification results. For example, this application can use the same machine learning model to perform label recognition processing on both the target object and related objects. When extracting label information for related objects, the object description text of the related objects can be obtained first, and then the same text feature extraction model can be used to perform feature extraction processing on the object description text to obtain the object embedding feature vector of the object description text. Then, the object embedding feature vector can be processed for label classification using different label classification heads. It is worth noting that the label classification head selected here should be consistent with the label classification head selected for extracting the label classification results of the target object. After the label classification is completed, it is necessary to determine the associated object label information of the related objects based on the multiple object label classification results. Here, all the labels existing in the related objects can be determined based on the binary classification results of the label classification heads, and these can be used as the associated object label information of the related objects. In this embodiment, by performing feature extraction and classification on the subject description tags, the accuracy of tag extraction for associated objects can be effectively guaranteed, while ensuring the consistency of the tag recognition process.

[0080] In an exemplary embodiment, step 209 includes: determining the social relationship tag distribution information of the target object based on the associated object tag information; comparing the distribution probability values ​​of each associated object tag in the social relationship tag distribution information, and filtering multiple target object tags from the associated object tags based on the comparison results; and comparing the target object tags with the tag classification results to obtain the object tags of the target object.

[0081] The social relationship tag distribution information refers to the distribution of tags among all associated objects of a target object. Specifically, after identifying the associated objects of the target object, the tags present in each associated object are determined. For example, associated object A has tags 1, 2, and 4; associated object 2 has tags 1, 2, and 3, and so on. Finally, all associated objects are counted, and it is determined that among all associated objects corresponding to the target object, there are 5 instances of tag 1, 4 instances of tag 2, 2 instances of tag 3, and 4 instances of tag 4. This is the social relationship tag distribution information of the target object. The probability distribution value of each associated object tag refers to the probability of each associated object tag appearing. For example, if there are 5 associated objects, and a total of 4 associated object tags are counted, then the probability distribution value is 80%.

[0082] For example, this application can specifically summarize the tag information of associated objects by analyzing the tag distribution of associated objects, and then compare the tag information of associated objects with the tag classification results of target objects to determine the tags of target objects. First, the social relationship tag distribution information of target objects is determined based on the tag information of associated objects. This process can specifically involve statistically analyzing the tag information of all objects, determining which social relationship tags associated objects possess, and how many of each social relationship tag there are. Specifically, the social relationship tag distribution information can be displayed in the form of a histogram. In a specific embodiment, such as... Figure 4 As shown, after extracting information from the social network, the specific label distribution of labels 1, 2, and 3 is displayed in the form of a histogram. Then, the probability values ​​of the distribution of each associated object label in the social relationship label distribution information are compared, and the target object label is determined based on the comparison results. That is, by comparing the probability values ​​of associated object labels, multiple social relationship labels that appear most frequently in the associated objects can be identified. For example, in the case of multiple social relationship labels, the top three labels with the highest probability values ​​can be identified as target object labels through comparison. Then, the target object labels are compared with the label classification results to determine the object label of the target object from the label classification results. In this embodiment, label filtering is performed based on social relationship label distribution information, and then the object label is determined by comparing the filtering results with the label classification results, which can effectively improve the efficiency and accuracy of object label identification.

[0083] In an exemplary embodiment, by comparing the target object label and the label classification result, the object label of the target object is obtained, including:

[0084] The identity tags of the target object are determined based on the tag classification results.

[0085] When a target object has multiple identity tags, the target identity tag that matches the target object tag is determined from the identity tags, and the object tag of the target object is determined from the target identity tags based on the distribution probability of the target object tag.

[0086] If the target object has a unique identity tag, the unique identity tag will be determined as the object tag of the target object.

[0087] If the target object does not have an identity tag, the tag with the highest probability distribution among the target object's tags will be used as the target object's object tag.

[0088] Among them, identity tags are the tags that the target object may possess, as determined by the tag classification head. These tags may not be accurate and need to be further filtered and verified in conjunction with social relationships to determine the final object tags.

[0089] For example, this application can specifically combine the identity tags of the target object determined by classification and the target object tags determined based on social relationships to perform object tag identification and authentication. The specific comparison process can be discussed in different scenarios. When the target object is determined to have multiple identity tags based on the tag classification results, the identity tags and target object tags need to be compared to determine if there are any identical tags. If only one identical tag exists, that tag is the target object's object tag. If multiple identical tags exist, their distribution probabilities in the social relationship tag distribution information can be compared, and the tag with the highest probability can be selected as the object tag. Specifically, if no identity tag and target object tag match, corresponding identification error information can be reported, which may indicate a tag classification error during the classification process. If the target object has only one single identity tag, that single identity tag can be directly identified as the target object's object tag. If the target object has no identity tag, the tag with the highest distribution probability among the target object tags can be used as the target object's object tag, i.e., the target object's object tag is directly obtained through the target object's social relationships. This is beneficial for collecting tags of objects without specific information within the system, i.e., achieving identity tag identification in cold start situations. In this embodiment, corresponding object label recognition processing is performed for different label classification results, thereby improving the object label recognition effect under different conditions. At the same time, object labels can be directly identified even in cold start conditions, ensuring the applicability of the label recognition method of this application.

[0090] In an exemplary embodiment, the method further includes: obtaining a text training dataset corresponding to different object labels; and training an initial classification head based on the text training dataset to obtain a label classification head corresponding to each object label.

[0091] For example, for each object label, its text recognition model consists of a pre-trained text model with shared parameters for multiple tasks and a classification head. The pre-trained text model with shared parameters is used to acquire the embedded feature representation of the input data, while the classification head aims at a binary classification task, that is, to determine whether the target object has an object label based on the current input data. Therefore, during the training phase, the acquired embedded feature representations are used to train different classification heads for different object labels. Thus, text training datasets corresponding to different object labels can be constructed during the training phase. For example, for the object label "student," some object description data corresponding to the "student" label can be acquired to construct some text training datasets corresponding to the "student" label. Then, the constructed text training datasets are used to train the classification head, enabling the trained label classification head to perform the corresponding label classification and recognition processing. In this embodiment, for different object labels, the training of the corresponding label classification heads is completed by constructing text training datasets for each type of object label. This allows for the expansion of recognizable label types directly by adding new label classification heads when adding new label types, thus avoiding a large amount of feature engineering and keyword library construction and maintenance. At the same time, the decoupling of different object label recognition can prevent the impact of existing object labels when adding new object labels, effectively improving model training efficiency.

[0092] In an exemplary embodiment, training an initial classification head based on a text training dataset to obtain a label classification head corresponding to each object label includes: obtaining a text processing pre-trained model; combining the text processing pre-trained model with the initial classification head to obtain a single-task training model; inputting the text training dataset into the single-task training model to obtain the model loss value corresponding to the sample data in the text training dataset; performing backpropagation based on the model loss value; freezing the parameters of the text processing pre-trained model during backpropagation and updating the parameters of the initial classification head to obtain a label classification head corresponding to each object label.

[0093] For example, a text processing pre-trained model refers to a model obtained through large-scale pre-training for extracting text features. This model can be used at various stages to extract feature information from the input text data. Combining the text processing pre-trained model with the initial classification head yields a single-task training model suitable for single-label classification. The next step is to train this single-task training model. In a specific embodiment, the structure of the single-task training model can be referred to... Figure 5As shown. Each single-task training model consists of three modules: the backbone network of the pre-trained text processing model, single-task setting information, and a trainable initial classifier head. Iterative training of the model can be completed by inputting training data into the model. During model training, a corresponding single-task training model can be constructed for each initial classifier head, and then the label classifier head can be trained efficiently in parallel. In each round of training, the text training dataset can be input into the single-task training model, and then the model parameters in the initial classifier head can be updated through backpropagation. First, for the input training dataset, the model loss value corresponding to the sample data can be obtained through the loss parameters in the classifier head. The model loss value satisfies the following formula:

[0094]

[0095] in Indicates the number of samples. and They represent the first The true label (0 or 1) and predicted probability (between 0 and 1) of each sample are defined. In the reverse process, the gradient of the loss function with respect to each parameter is calculated using the above formula, and the model parameters are updated using the gradient descent algorithm. Furthermore, during parameter updates, the parameters of the pre-trained text processing model are frozen, and only the parameters of the initial classification head are updated. In this embodiment, the application trains a model separately for each object label, thereby decoupling the recognition of different object labels, ensuring that subsequent additions of new object labels or iterations of existing identity labels do not affect each other. The training model is constructed by combining the pre-trained text processing model with the initial classification head, completing the training of the classification head and ensuring the efficiency of the model training process.

[0096] In one embodiment, step 207 includes: reading the pre-stored network configuration file corresponding to the target object; performing label classification processing on the embedded feature vector based on the different label classification heads determined by the pre-stored network configuration file, and obtaining the label classification results of the embedded feature vector in multiple label dimensions.

[0097] Among them, the pre-stored network configuration file refers to the file data used to configure the model network structure. It contains the type information of the selected classification heads. The label classification head of the current target object can be dynamically configured by reading the pre-stored network configuration file.

[0098] For example, this application can specifically implement the configuration processing of the label classification head through a pre-stored network configuration file. Developers can pre-configure pre-stored network configuration files for different scenarios. During the label recognition process, after the descriptive text of the target object is imported into the system, the system can read the pre-stored network configuration file corresponding to the target object based on the scenario to which the target object belongs, and then search for multiple different label classification heads in the scenario to perform label classification processing on the embedded feature vector, thereby obtaining the label classification results of the embedded feature vector in multiple label dimensions. Figure 6 As shown, in one embodiment, this application is applied to object tag recognition in a content recommendation scenario. In this case, based on the scenario to which the target object belongs, a pre-stored network configuration file A for the content recommendation scenario is read, along with tag classification heads a, b, and c for the content recommendation scenario. These three tag classification heads represent the recommendation success rates of different types of content. These three tag classification heads are then loaded into the model, and the imported feature vectors are processed using these three tag classification heads to obtain the corresponding tag classification results, thereby determining whether the corresponding type of content can be recommended to the target object. In another abnormal object recognition scenario, a pre-stored network configuration file B for the object anomaly recognition scenario is read, along with tag classification heads d, e, and f for the object anomaly recognition scenario. Different tag classification heads represent the anomaly recognition results of the target object under different detection methods. In this case, only these three tag classification heads need to be found and loaded into the model. The imported feature vectors are then processed using these three tag classification heads to obtain the user's recognition results under different anomaly detection methods. In this embodiment, different classification heads are dynamically loaded through configuration files to predict identity labels. When the amount of inference data is large, this approach can effectively reduce the pressure on model inference. At the same time, the dynamic loading of classification heads can also well support the addition and removal of object labels, thereby ensuring the efficiency and accuracy of the object label recognition process.

[0099] In an exemplary embodiment, the method further includes: granting exemption permissions to the target object if the target object is determined to be a normal object based on the object tag of the target object; generating an abnormal risk identifier corresponding to the target object if the target object is determined to be an abnormal object based on the object tag of the target object; and finally generating a list of risk objects based on the object's permission information and identifier information.

[0100] For example, this application can be used in the process of object anomaly identification. By extracting the descriptive text of the target object, it identifies whether the target object is abnormal. Simultaneously, it determines the cause of the anomaly through the identified anomaly tags. Each object tag can correspond to an anomaly cause. When the target object's object tag is normal and does not correspond to any anomaly cause, the target object can be determined as normal, and exemption permissions can be granted to prevent normal objects from being treated as anomalies and falling within the scope of anomalies. When the target object's object tag is abnormal, and the object tag indicates the anomaly type, the target object can be determined as an anomaly, and a corresponding anomaly risk label can be added to the target object based on the anomaly type indicated by the object tag. Finally, a risk object list is generated based on the permission information and label information of each object for subsequent implementation of measures such as account suspension. In this embodiment, object tags are used to identify the type of the target object, thereby enabling appropriate processing based on the object type, effectively improving the efficiency and accuracy of risk handling and enhancing system security performance.

[0101] This application also provides an application scenario, which is illustrated by taking the above-mentioned object label recognition method as an example. The object label recognition method specifically includes:

[0102] When users want to implement content recommendation functions on a content platform, they can use the object tag recognition method of this application to train an object tag recognition model, and use the object tag recognition model to identify the object tags of platform users, determine information such as user interests, and then recommend content that can arouse their interest based on the object tags.

[0103] First, the object label recognition model needs to be trained. Examples of the model training and inference process can be found in [reference needed]. Figure 7 As shown, this process first requires collecting basic information and historical browsing data of platform users as foundational data. After collecting basic data from various platform users and anonymizing the data, the anonymized text data is used as training data. Specifically, based on historical browsing information, the types of content a user is interested in can be determined, object tags can be added to platform users, and the corresponding text data of the platform users can be used as a training sample for that object tag. By organizing data for each platform user, a corresponding training set can be constructed for each object tag. Then, the training processing of the tag classification header under each object tag type needs to be completed, such as... Figure 7As shown, for each label classification head, training can be achieved by constructing a single-task training module. A pre-trained text model is selected as the backbone network for feature extraction, and then the initial label classification heads are concatenated into the pre-trained text model to construct a single-task training model. By importing the training set data for each object label type into the single-task training model, the parameters of the classification heads in the single-task training model are updated using backpropagation, thereby training the label classification heads for each object label type.

[0104] After training the tag classification heads, the model can be deployed to the platform environment to perform tag recognition for platform users. When a user visits the platform and seeks content recommendations, the platform will generate de-identified descriptive text based on the user's basic information. This de-identified text will be imported into the model for inference processing. This data will first undergo feature extraction processing using the same pre-trained text model as in the training phase to obtain corresponding embedded feature vectors. These embedded feature vectors will then be imported into multiple different tag classification heads to determine whether the platform user possesses the corresponding tags.

[0105] Since platform users may identify multiple tags through tag classification heads, it is necessary to filter these tags and select the highest priority tag as the object tag. To identify the object tag, a relationship network graph of the current platform users can be obtained. Specifically, a relationship network graph can be built based on the friend relationships of platform users to determine the first-order neighbors of the target object in the relationship network graph. Then, sampling processing is performed on the first-order neighbors to obtain the target object's associated objects in the social network. These associated objects are other platform users, and tag priority can be identified based on the analysis of these platform users. First, the object description text of the associated objects is obtained from the platform; feature extraction processing is performed on the object description text to obtain the object embedding feature vector; tag classification processing is performed on the object embedding feature vector using different tag classification heads to obtain multiple object tag classification results; the social relationship tag distribution information of the target object is determined based on the associated object tag information; the distribution probability values ​​of each associated object tag in the social relationship tag distribution information are compared, and the target object tag is determined based on the comparison results; the target object tag is obtained by comparing the target object tag with the tag classification results. For the specific comparison process, the target object's identity tag can be determined first based on the tag classification results. Then, if the target object has multiple identity tags, the target identity tag that matches the target object's tag is selected from the identity tags. Based on the distribution probability of the target object's tags, the target object's object tag is determined from the target identity tags. If the target object has a single identity tag, that single identity tag is determined as the target object's object tag. If the target object has no identity tag, the tag with the highest distribution probability among the target object tags is taken as the target object's object tag. After determining the object tag, content under the object tag category can be used as the primary recommended content, and recommended content from other identity tags can be used as secondary recommended content to generate a recommendation page corresponding to the current platform user, completing the content recommendation process. Furthermore, the classification results output by the tag classification head can be manually inspected to identify incorrect classification cases. These incorrect cases are fed back into the training data as negative samples to participate in model training, completing the optimization training of the tag classification head and ensuring subsequent recognition performance.

[0106] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0107] Based on the same inventive concept, this application also provides an object tag recognition device for implementing the object tag recognition method described above. The solution provided by this device is similar to the implementation described in the above method; therefore, the specific limitations in one or more object tag recognition device embodiments provided below can be found in the limitations of the object tag recognition method described above, and will not be repeated here.

[0108] In one exemplary embodiment, such as Figure 8 As shown, an object label recognition device is provided, comprising:

[0109] The data acquisition module 801 is used to acquire the description text of the target object.

[0110] The feature extraction module 803 is used to perform feature extraction processing on the descriptive text to obtain the embedded feature vector of the descriptive text.

[0111] The multi-classification processing module 805 is used to perform label classification processing on the embedded feature vector through different label classification heads to obtain the label classification results of the embedded feature vector in multiple label dimensions.

[0112] The association analysis module 807 is used to find the associated object tag information of the target object on the social network.

[0113] The object label recognition module 809 is used to filter the label classification results based on the associated object label information to obtain the object label of the target object.

[0114] In one embodiment, the association analysis module 807 is specifically used to: obtain a relationship network diagram of the target object in a social application; identify the associated objects of the target object in the social network based on the relationship network diagram; and find the associated object tag information of the associated objects.

[0115] In one embodiment, the association analysis module 807 is further configured to: identify the first-order neighboring entities of the target object in the social network based on the relationship network graph; and perform sampling processing on the first-order neighboring entities to obtain the associated objects of the target object in the social network.

[0116] In one embodiment, the association analysis module 807 is further configured to: obtain the object description text of the associated object; perform feature extraction processing on the object description text to obtain the object embedding feature vector of the object description text; perform label classification processing on the object embedding feature vector through different label classification heads to obtain multiple object label classification results of the object embedding feature vector; and determine the associated object label information of the associated object based on the multiple object label classification results.

[0117] In one embodiment, the object tag recognition module 809 is specifically used to: determine the social relationship tag distribution information of the target object based on the associated object tag information; compare the distribution probability values ​​of each associated object tag in the social relationship tag distribution information, and filter multiple target object tags from the associated object tags based on the comparison results; and compare the target object tags with the tag classification results to obtain the object tags of the target object.

[0118] In one embodiment, the object label recognition module 809 is specifically used to: determine the identity label of the target object based on the label classification result; when the target object has multiple identity labels, determine the target identity label that is consistent with the target object label from the identity labels, and determine the object label of the target object from the target identity labels based on the distribution probability of the target object labels; when the target object has a single identity label, determine the single identity label as the object label of the target object; when the target object does not have an identity label, take the label with the highest distribution probability among the target object labels as the object label of the target object.

[0119] In one embodiment, a classification head training module is further included, which is used to: obtain a text training dataset corresponding to different object labels; and train an initial classification head based on the text training dataset to obtain a label classification head corresponding to each object label.

[0120] In one embodiment, the classification head training module is specifically used to: obtain a text processing pre-trained model, combine the text processing pre-trained model with the initial classification head to obtain a single-task training model; input the text training dataset into the single-task training model to obtain the model loss value corresponding to the sample data in the text training dataset; perform backpropagation processing based on the model loss value, freeze the parameters of the text processing pre-trained model during backpropagation, and update the parameters of the initial classification head to obtain the label classification head corresponding to each object label.

[0121] In one embodiment, the multi-classification processing module 805 is specifically used to: read the pre-stored network configuration file corresponding to the target object; and perform label classification processing on the embedded feature vector based on the different label classification heads determined by the pre-stored network configuration file to obtain the label classification results of the embedded feature vector in multiple label dimensions.

[0122] In one embodiment, the system further includes an exception handling module, configured to: grant exemption permissions to the target object if the target object is determined to be a normal object based on the object tag of the target object; and generate an exception risk identifier corresponding to the target object if the target object is determined to be an abnormal object based on the object tag of the target object, and generate a list of risk objects based on the object's permission information and identifier information.

[0123] Each module in the aforementioned object tag recognition device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0124] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 9 As shown, this computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores data related to object tag identification. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When executed by the processor, the computer program implements an object tag identification method.

[0125] In one exemplary embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 10As shown, the computer device includes a processor, memory, input / output interfaces, a communication interface, a display unit, and an input device. The processor, memory, and input / output interfaces are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interfaces. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The input / output interfaces are used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. When executed by the processor, the computer program implements an object tag identification method. The display unit is used to form a visually visible image and can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0126] Those skilled in the art will understand that Figure 10 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0127] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0128] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0129] In one embodiment, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, causing the computer device to perform the steps in the above method embodiments.

[0130] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0131] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0132] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. An object label recognition method characterized by, The method includes: Get the description text of the target object; The descriptive text is subjected to feature extraction processing to obtain the embedded feature vector of the descriptive text; By performing label classification processing on the embedded feature vector using different label classification heads, the label classification results of the embedded feature vector in multiple label dimensions are obtained. Find the associated object tag information of the target object on the social network; Based on the associated object tag information, the tag classification results are filtered to obtain the object tags of the target object.

2. The method of claim 1, wherein, The process of finding the associated object tag information of the target object on the social network includes: Obtain the relationship network diagram of the target object in the social application; Identify the associated objects of the target object in the social network based on the relationship network graph; Find the associated object tag information of the associated object.

3. The method of claim 2, wherein, The method of identifying the target object's associated objects in the social network based on the relationship network graph includes: Identify the first-order neighbor entities of the target object in the social network based on the relationship network graph; The first-order adjacent entities are sampled to obtain the associated objects of the target object in the social network.

4. The method of claim 2, wherein, The associated object tag information for finding the associated object includes: Obtain the object description text of the associated object; The object description text is subjected to feature extraction processing to obtain the object embedding feature vector of the object description text; By performing label classification processing on the object embedding feature vector using different label classification heads, multiple object label classification results of the object embedding feature vector are obtained; The associated object tag information of the associated object is determined based on the classification results of the multiple object tags.

5. The method of claim 1, wherein, The process of filtering the tag classification results based on the associated object tag information to obtain the object tags of the target object includes: Based on the associated object tag information, determine the social relationship tag distribution information of the target object; Compare the distribution probability values ​​of each associated object tag in the social relationship tag distribution information, and filter multiple target object tags from the associated object tags based on the comparison results; By comparing the target object label with the label classification result, the object label of the target object is obtained.

6. The method of claim 5, wherein, The comparison between the target object label and the label classification result yields the following object labels for the target object: The identity label of the target object is determined based on the label classification results; When the target object has multiple identity tags, a target identity tag that matches the target object tag is determined from the identity tags, and the object tag of the target object is determined from the target identity tag based on the distribution probability of the target object tag; If the target object has a unique identity tag, the unique identity tag is determined as the object tag of the target object; If the target object does not have an identity tag, the tag with the highest probability distribution among the target object tags shall be used as the object tag of the target object.

7. The method of claim 1, wherein, The method further includes: Obtain the text training dataset corresponding to different object labels; The initial classification head is trained based on the text training dataset to obtain the label classification head corresponding to each object label.

8. The method of claim 7, wherein, The step of training the initial classification head based on the text training dataset to obtain the label classification head corresponding to each object label includes: Obtain a pre-trained text processing model, and combine the pre-trained text processing model with the initial classification head to obtain a single-task training model; Input the text training dataset into the single-task training model to obtain the model loss value corresponding to the sample data in the text training dataset; Backpropagation is performed based on the model loss value. During backpropagation, the parameters of the pre-trained text processing model are frozen, and the parameters of the initial classification head are updated to obtain the label classification head corresponding to each object label.

9. The method according to any one of claims 1 to 8, characterized in that, The step of performing label classification processing on the embedded feature vector using different label classification heads to obtain the label classification results of the embedded feature vector in multiple label dimensions includes: Read the pre-stored network configuration file corresponding to the target object; Based on the different label classification heads determined by the pre-stored network configuration file, the embedded feature vector is subjected to label classification processing to obtain the label classification results of the embedded feature vector in multiple label dimensions.

10. The method according to any one of claims 1 to 8, characterized in that, The method further includes: If the target object is determined to be a normal object based on its object tag, an exemption permission is granted to the target object. If the target object is determined to be an abnormal object based on its object tag, an abnormal risk identifier corresponding to the target object is generated. A list of risky objects is generated based on the object's permission and identification information.

11. An object label recognition apparatus characterized by comprising: The device includes: The data acquisition module is used to acquire the descriptive text of the target object; The feature extraction module is used to perform feature extraction processing on the descriptive text to obtain the embedded feature vector of the descriptive text; The multi-classification processing module is used to perform label classification processing on the embedded feature vector through different label classification heads to obtain the label classification results of the embedded feature vector in multiple label dimensions. The association analysis module is used to find the associated object tag information of the target object on the social network; The object tag recognition module is used to filter the tag classification results based on the associated object tag information to obtain the object tag of the target object.

12. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 10.

13. A computer readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 10.

14. A computer program product comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 10.

Images