CN117312979A - Object classification method, classification model training method and electronic equipment - Google Patents

Abstract

The application provides an object classification method, a classification model training method, a device, electronic equipment and a computer readable storage medium; the object classification method comprises the following steps: acquiring preset category information and description information to be tested of an object to be tested; wherein the preset category information comprises preset categories of a plurality of category levels; performing hierarchical classification processing based on the description information to be detected and the preset category information through a classification model to obtain classification results of the object to be detected in a plurality of category levels; and determining target categories of the object to be tested in the category levels according to classification results of the object to be tested in the category levels. According to the method and the device, the object to be measured is classified layer by layer, the target category of each category level is determined, the conditions of a plurality of category levels can be comprehensively considered, and the classification precision is effectively improved.

Description

Object classification method, classification model training method and electronic device

Technical field

The present application relates to artificial intelligence technology, and in particular, to an object classification method, a classification model training method, a device, an electronic device, and a computer-readable storage medium.

Background technique

The category system is created to efficiently identify and apply specific objects. The category system includes multiple category levels, and each category level includes several categories. Category systems are common on e-commerce platforms. By determining the categories to which each product belongs, it can help the platform better maintain and manage the products on the site.

As for the category to which the product belongs, in the solution provided by the relevant technology, the text information of the product is usually matched with the last category level to determine the category to which the product belongs at the last category level. However, after research, the inventor found that because the category system itself is large and complex, the accuracy of classification based on the last category level is low, and it is easy to classify products into the wrong category.

Contents of the invention

Embodiments of the present application provide an object classification method, a classification model training method, a device, an electronic device, and a computer-readable storage medium, which can comprehensively consider the situation of multiple category levels, improve the training effect of the classification model, and thereby improve the classification accuracy.

The technical solution of the embodiment of this application is implemented as follows:

The embodiment of the present application provides an object classification method, including:

Obtain preset category information and test description information of the object to be tested; wherein the preset category information includes preset categories at multiple category levels;

The classification model performs level-by-level classification processing based on the description information to be tested and the preset category information to obtain classification results of the object to be tested at the multiple category levels;

According to the classification results of the object to be tested at the plurality of category levels, the target category of the object to be tested at the plurality of category levels is determined.

Through the above solution, the trained classification model is used to achieve multi-level classification of the object to be tested. Compared with the solution that only relies on the last category level for classification, each category level can be comprehensively considered, effectively improving the classification accuracy.

In the above scheme, it also includes:

For any category level, perform the following processing:

Perform screening processing on multiple preset categories of any one category level according to the classification results of the previous category level of any one category level;

Update the preset category information according to the filtered preset category.

Through the above solution, the preset category information is updated based on the existing classification results. In this way, it can comply with the rules of the category system, improve the accuracy and rationality of classification, and reduce the amount of calculation.

In the above solution, the classification model performs level-by-level classification processing based on the description information to be tested and the preset category information to obtain classification results of the object to be tested at the multiple category levels, include:

The following processing is performed by the classification model:

Extract description features to be tested from the description information to be tested, and extract preset category features from the preset categories included in the preset category information;

Calculate the probability that the object to be tested belongs to the preset category corresponding to the preset category feature according to the description feature to be tested and the preset category feature;

For any category level, the probability that the object to be tested belongs to multiple preset categories of the any category level is determined as the classification result of the object to be tested at the any category level. .

Through the above solution, the probability that the object to be tested belongs to the preset category can be predicted based on the learned feature association, ensuring the accuracy of the obtained probability.

The embodiment of the present application provides a classification model training method, including:

Obtain sample description information and sample category information of the sample object; the sample category information includes sample categories of the sample object at multiple category levels;

The classification model performs level-by-level classification processing based on the sample description information and the sample category information to obtain classification results at the multiple category levels;

Perform loss calculation according to the sample categories and classification results of the same category level to obtain hierarchical loss, and perform multi-level fusion processing on the hierarchical losses corresponding to the multiple category levels to obtain multi-level loss;

The classification model is trained according to the multi-level loss; wherein the trained classification model is used to predict the classification results of the object to be tested at the multiple category levels.

Through the above solution, comprehensive consideration of the hierarchical losses corresponding to multiple category levels can improve the training effect of the classification model, so that the trained classification model has the ability to classify at multiple levels, instead of just relying on the last category level. It can effectively improve classification accuracy.

In the above solution, the classification model performs level-by-level classification processing based on the sample description information and the sample category information to obtain classification results at the multiple category levels, including:

The following processing is performed by the classification model:

Extract sample description features from the sample description information, and extract sample category features from the sample category information;

For any category level, calculate the category probability distribution at the any category level based on the sample description features and the sample category features as a classification result at the any category level.

Through the above solution, an implementation method of level-by-level classification is provided, that is, extracting features from the input information to predict the category probability distribution of each category level, which can be adapted to the situation of each category level.

In the above solution, after extracting sample description features from the sample description information and extracting sample category features from the sample category information, the method further includes:

Perform linear projection processing on the sample description features and the sample category features;

The sample description features and the sample category features after linear projection are normalized.

Through the above solution, the sample description characteristics and the sample category characteristics are adjusted to be consistent in dimension by linear projection, and the sample description characteristics and the sample category characteristics are adjusted to be consistent in measurement by normalization method, which can Strengthen the comparability between features and enhance the learning effect.

In the above solution, the sample description information includes description information of multiple modalities; and extracting sample description features from the sample description information includes:

Extract modal description features from the description information of each modality;

Feature fusion processing is performed on modal description features corresponding to the plurality of modalities to obtain sample description features.

Through the above solution, comprehensive consideration of the description information of multiple modalities related to the sample object can improve the richness of the information, thereby improving the training effect of the classification model.

In the above solution, the number of sample objects includes multiple; the classification model performs level-by-level classification processing based on the sample description information and the sample category information to obtain classification results at the multiple category levels. ,include:

Combine multiple sample description information and multiple sample category information to obtain multiple information combinations; each information combination includes one sample description information and one sample category information;

The classification model performs level-by-level classification processing based on the information combination to obtain the classification results of the information combination at the multiple category levels;

The loss calculation based on the sample categories and classification results of the same category level results in hierarchical loss, including:

For any category level, perform the following processing:

According to the sample category of the target sample object corresponding to the information combination at the any category level, determine the category target label of the information combination at the any category level;

Determine the difference between the category label and the classification result of the information combination at any category level as the information combination loss;

Information combination fusion processing is performed on the information combination losses corresponding to the multiple information combinations to obtain the hierarchical loss of any category level.

Through the above solution, multiple information combinations are generated based on the idea of comparative learning, which can increase the richness of training data and thereby improve the effect of model training.

In the above scheme, it also includes:

When the sample description information and sample category information in the information combination correspond to the same sample object, determine the same sample object as the target sample object corresponding to the information combination;

When the sample description information and sample category information in the information combination correspond to different sample objects, the target sample object corresponding to the information combination is determined to be no object; wherein the no object is in the sample category of any category level There is no category.

Through the above scheme, when the sample description information and sample category information in the information combination correspond to the same sample object, the information combination is used as a positive sample; when the sample description information and sample category information in the information combination correspond to different sample objects, Combine this information as negative samples. In this way, the expansion of training data is achieved accurately and effectively.

An embodiment of the present application provides an object classification device, including:

The target acquisition module is used to obtain preset category information and test description information of the object to be tested; wherein the preset category information includes a preset category for each category level in multiple category levels;

A target processing module, configured to perform level-by-level classification processing based on the description information to be tested and the preset category information through a classification model to obtain classification results of the object to be tested at the multiple category levels;

A determining module, configured to determine the category of the object to be tested at the multiple category levels according to the classification results of the object to be tested at the multiple category levels.

An embodiment of the present application provides a classification model training device, including:

A sample acquisition module is used to obtain sample description information and sample category information of a sample object; the sample category information includes sample categories of the sample object at multiple category levels;

A sample processing module, configured to perform level-by-level classification processing based on the sample description information and the sample category information through a classification model to obtain classification results at the multiple category levels;

The loss calculation module is used to perform loss calculation according to the sample categories and classification results of the same category level to obtain the hierarchical loss, and perform multi-level fusion processing on the hierarchical losses corresponding to the multiple category levels to obtain the multi-level loss;

A training module configured to train the classification model according to the multi-level loss; wherein the trained classification model is used to predict the classification results of the object to be tested at the multiple category levels.

An embodiment of the present application provides an electronic device, including:

Memory, used to store executable instructions;

The processor is configured to implement the object classification method or classification model training method provided by the embodiments of the present application when executing executable instructions stored in the memory.

Embodiments of the present application provide a computer-readable storage medium that stores executable instructions for causing the processor to implement the object classification method or classification model training method provided by the embodiments of the present application when executed.

Embodiments of the present application provide a computer program product. The computer program product includes executable instructions for causing a processor to implement the object classification method or classification model training method provided by the embodiments of the present application when executed.

Description of the drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application. For those skilled in the art, other drawings can also be obtained based on these drawings without exerting creative efforts.

Figure 1 is a schematic structural diagram of a classification system provided by an embodiment of the present application;

Figure 2A is a schematic structural diagram of a server provided by an embodiment of the present application;

Figure 2B is another schematic structural diagram of a server provided by an embodiment of the present application;

Figure 3A is a schematic flow chart of the classification model training method provided by the embodiment of the present application;

Figure 3B is another schematic flow chart of the classification model training method provided by the embodiment of the present application;

Figure 3C is another schematic flow chart of the classification model training method provided by the embodiment of the present application;

Figure 4 is a schematic flow chart of the object classification method provided by the embodiment of the present application;

Figure 5 is a schematic diagram of product display on the e-commerce platform provided by the embodiment of the present application;

Figure 6 is a schematic diagram of constructing a feature combination matrix provided by an embodiment of the present application.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be described in further detail below in conjunction with the accompanying drawings. The described embodiments should not be regarded as limiting the present application. Those of ordinary skill in the art will not make any All other embodiments obtained under the premise of creative work belong to the scope of protection of this application.

In the following description, references to "some embodiments" describe a subset of all possible embodiments, but it is understood that "some embodiments" may be the same subset or a different subset of all possible embodiments, and Can be combined with each other without conflict. In the following description, the term "plurality" refers to at least two.

Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the technical field to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of the present application and are not intended to limit the present application.

Before further describing the embodiments of the present application in detail, the nouns and terms involved in the embodiments of the present application are explained. The nouns and terms involved in the embodiments of the present application are applicable to the following explanations.

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

Machine Learning (ML) is the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications cover all fields of artificial intelligence. Machine learning involves multiple disciplines such as probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specializes in studying how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge. The structure enables it to continuously improve its performance.

In this embodiment of the present application, the classification model may be an artificial neural network model. The artificial neural network is implemented based on machine learning theory and imitates the way biological neurons transmit signals to each other.

2) Forward Propagation: refers to the process of running the network layers of the neural network from left to right, until the last layer gets the output result. Back propagation refers to the process of updating the weight parameters and offsets in the network layer of the neural network to reduce this error when there is an error between the output result and the expected result. Backpropagation is the reverse order of forward propagation.

3) Category system: includes multiple category levels, each category level includes several preset categories. Taking the product category system as an example, the first category level can include categories such as "clothing" and "food"; for the "clothing" category, the next level (the second category level) can include Categories such as "tops" and "pants"; for the category "tops", the next level (the third category level) can include categories such as "T-shirts" and "down jackets".

4) Object: refers to the object that needs to be classified. The embodiment of the present application does not limit the type of object, which can be commodities, vehicles, humans, animals, etc.

5) Loss value (Loss): "Loss" means the penalty received by the model for failing to output the expected results. The loss function determines the performance of the model by comparing the output results of the model with the expected results, and then finds the optimization direction. Through the loss function The resulting value is the loss value. If the deviation between the output result and the expected result is very large, the loss value will be large; if the deviation is small, the loss value will be very low. The embodiment of the present application does not limit the type of the loss function. For example, it may be a cross-entropy loss function, a hinge loss function, a negative log-likelihood loss function, etc.

For the category classification of specific objects, in the solutions provided by related technologies, text information of the object is usually matched with each category in the last category level, and the category with the greatest text similarity is used as the category to which the object belongs. category. However, the inventor found through research that on the one hand, text information is relatively one-sided and the amount of information is insufficient. Pure text information can greatly confuse the judgment of categories; on the other hand, relying on the last category level for classification, The accuracy is low and it is easy to classify objects into wrong categories.

In view of this, embodiments of the present application provide an object classification method, classification model training method, device, electronic device and computer-readable storage medium, which can make the classification model have multi-level classification capabilities through machine learning, thereby improving object classification. Classification accuracy. The following describes exemplary applications of the electronic device provided by the embodiments of the present application. The electronic device provided by the embodiments of the present application can be implemented as various types of terminal devices or as a server.

Referring to Figure 1, Figure 1 is an architectural schematic diagram of the classification system 100 provided by the embodiment of the present application. The terminal device 400 is connected to the server 200 through the network 300, and the server 200 is connected to the database 500. The network 300 can be a wide area network or a local area network, or a A combination of the two.

In some embodiments, taking the electronic device as a server as an example, the classification model training method provided in the embodiments of the present application can be implemented by the server. For example, the server 200 obtains sample description information and sample category information of the sample object. The sample category information includes the sample categories to which the sample object belongs at multiple category levels. The server 200 may obtain the sample description information and sample category information of the sample object from the terminal device 400 or the database 500 . Then, the server 200 calls the classification model to perform level-by-level classification processing based on the sample description information and sample category information to obtain classification results at multiple category levels; the loss is calculated based on the sample categories and classification results at the same category level. Hierarchical loss: perform multi-level fusion processing on the hierarchical losses corresponding to multiple category levels to obtain multi-level loss; train the classification model based on the multi-level loss; among them, the trained classification model is used to predict the performance of the test object in multiple categories Level classification results.

In some embodiments, taking the electronic device as a server as an example, the object classification method provided in the embodiments of this application can be implemented by the server. For example, the server 200 can obtain preset category information (for example, it can be stored locally in the server 200 or in the database 500 in advance), and obtain the test description information of the object to be tested from the terminal device 400 or the database 500 . Then, the server 200 calls the trained classification model to perform level-by-level classification processing based on the description information to be tested and the preset category information, to obtain the classification results of the object to be tested at multiple category levels, and based on the classification results of the object to be tested in multiple categories. The classification results at the category level determine the target category of the object to be tested at multiple category levels. The server 200 may store the obtained target categories of the object to be tested at multiple category levels in the database 500 or send them to the terminal device 400 .

In some embodiments, taking the electronic device as a terminal device as an example, the object classification method provided in the embodiments of this application can be implemented by the terminal device. For example, the server 200 may send the trained classification model to the terminal device 400, so that the terminal device 400 deploys the trained classification model locally. The terminal device 400 can obtain the preset category information and the test description information of the object to be tested, call the trained classification model to perform hierarchical classification processing based on the description information to be tested and the preset category information, and obtain the multi-level classification of the object to be tested. Classification results at multiple category levels, and based on the classification results of the object to be tested at multiple category levels, determine the target category of the object to be tested at multiple category levels.

In some embodiments, the terminal device 400 or the server 200 can implement the classification model training method or object classification method provided by the embodiments of the present application by running a computer program. For example, the computer program can be a native program or software module in the operating system; It can be a local (Native) application (APP, Application), that is, a program that needs to be installed in the operating system to run, such as an e-commerce platform application (client 410 in Figure 1); it can also be a small program, that is, A program that only needs to be downloaded to the browser environment to run; it can also be a small program that can be embedded in any APP, where the small program can be run or closed under the user's control. In summary, the computer program described above can be any form of application, module or plug-in.

In some embodiments, the server 200 may be an independent physical server, a server cluster or a distributed system composed of multiple physical servers, or may provide cloud services, cloud databases, cloud computing, cloud functions, cloud storage, Cloud servers for basic cloud computing services such as network services, cloud communications, middleware services, domain name services, security services, content delivery networks (Content Delivery Network, CDN), and big data and artificial intelligence platforms. Among them, cloud services can be classification models The training service or object classification service is for the terminal device 400 to call. The terminal device 400 can be a smart phone, a tablet computer, a notebook computer, a desktop computer, a smart TV, a smart watch, etc., but is not limited thereto. The terminal device and the server can be connected directly or indirectly through wired or wireless communication methods, which are not limited in the embodiments of this application.

Taking the electronic device provided by the embodiment of the present application as a server as an example, it can be understood that when the electronic device is a terminal device, modules such as a user interface, a presentation module, and an input processing module may also be included on the basis of Figure 2A. Referring to Figure 2A, Figure 2A is a schematic structural diagram of a server 200 provided by an embodiment of the present application. The server 200 shown in Figure 2A includes: at least one processor 210, a memory 250, and at least one network interface 220. The various components in server 200 are coupled together by bus system 240 . It can be understood that the bus system 240 is used to implement connection communication between these components. In addition to the data bus, the bus system 240 also includes a power bus, a control bus and a status signal bus. However, for the sake of clarity, the various buses are labeled bus system 240 in FIG. 2A.

The processor 210 may be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware. Components, etc., wherein the general processor can be a microprocessor or any conventional processor, etc.

Memory 250 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid state memory, hard disk drives, optical disk drives, etc. Memory 250 optionally includes one or more storage devices physically located remotely from processor 210 .

Memory 250 includes volatile memory or non-volatile memory, and may include both volatile and non-volatile memory. The non-volatile memory may be a read-only memory (ROM, Read Only Memory), and the volatile memory may be a random-access memory (RAM, Random Access Memory). The memory 250 described in the embodiments of this application is intended to include any suitable type of memory.

In some embodiments, the memory 250 is capable of storing data to support various operations, examples of which include programs, modules, and data structures, or subsets or supersets thereof, as exemplarily described below.

The operating system 251 includes system programs used to process various basic system services and perform hardware-related tasks, such as the framework layer, core library layer, driver layer, etc., which are used to implement various basic services and process hardware-based tasks;

Network communications module 252 for reaching other computing devices via one or more (wired or wireless) network interfaces 220. Exemplary network interfaces 220 include: Bluetooth, Wireless Compliance Certified (WiFi), and Universal Serial Bus ( USB, Universal Serial Bus), etc.;

In some embodiments, the classification model training device provided by the embodiment of the present application can be implemented in software. Figure 2A shows the classification model training device 2551 stored in the memory 250, which can be software in the form of programs, plug-ins, etc., It includes the following software modules: sample acquisition module 25511, sample processing module 25512, loss calculation module 25513 and training module 25514. These modules are logical, so they can be combined or further split according to the functions implemented. The functions of each module are explained below.

In some embodiments, the object classification device provided by the embodiments of the present application can also be implemented in software. Figure 2B shows the object classification device 2552 stored in the memory 250, which can be software in the form of programs, plug-ins, etc., including The following software modules: target acquisition module 25521, target processing module 25522, and determination module 25523. These modules are logical, so they can be arbitrarily combined or further split according to the functions implemented. The functions of each module are explained below. It is worth noting that, except for the object classification device 2552 shown in Figure 2B, the other structures can be the same as Figure 2A.

The classification model training method provided by the embodiment of the present application will be described with reference to the exemplary application and implementation of the electronic device provided by the embodiment of the present application.

Referring to Figure 3A, Figure 3A is a schematic flow chart of a classification model training method provided by an embodiment of the present application, which will be described in conjunction with the steps shown in Figure 3A.

In step 101, sample description information and sample category information of the sample object are obtained; the sample category information includes sample categories of the sample object at multiple category levels.

In the embodiment of this application, the category system has been established in advance. The category system includes multiple category levels. Each category level includes several preset categories. There are correlations between categories at different category levels. , that is, there is an association relationship between a category in a certain category level and at least one category in the next category level. On the basis of the established category system, specific objects need to be classified into the correct categories. Therefore, the relevant data of the sample objects can be used as training data to train the classification model, so that the classification model has the ability to classify accurately, where , sample objects refer to objects that have been labeled with categories.

The relevant data of the sample object includes sample description information and sample category information. The sample description information is used to describe the sample object. The sample description information can be single-modal information, such as text information; it can also be multi-modal information. Such as including text information, picture information, etc. The sample category information includes the sample category to which the sample object belongs at each category level. The sample category refers to the category that has been marked and determined to be correct.

Taking the e-commerce scenario as an example, the e-commerce category system can include three category levels. The first category level can include "clothing", "food", and "electrical appliances" categories; for the "clothing" category, The next level (the second category level) can include the categories of "tops" and "pants", that is, the "clothes" category has an associated relationship with the categories of "tops" and "pants". The same is true below; for "tops" and "pants" categories, For the "tops" category, the next level (the third category level) can include "T-shirts" and "down jackets" categories. In this scenario, the sample object can be a product that has been manually labeled with a good category. The sample description information can include pictures and texts of the product (such as product introduction), etc. The sample category information includes the sample object in the e-commerce category system. The sample category to which each category level belongs. For example, the sample category information can be "clothes-tops-T-shirts", the sample category "clothes" corresponds to the first category level, and the sample category "tops" corresponds to the second There are three category levels, and the sample category "T-shirt" corresponds to the third category level.

In step 102, the classification model performs level-by-level classification processing based on the sample description information and the sample category information to obtain classification results at multiple category levels.

Here, the sample description information and sample category information are used as input data of the classification model, and forward propagation processing is performed in the classification model. That is, the classification model performs level-by-level classification processing based on the sample description information and sample category information. The obtained The output results are classification results at each category level. The classification model may be an artificial neural network model, and forward propagation processing refers to processing the input data of the network layers in the order from front to back in the classification model.

In step 103, loss calculation is performed based on the sample categories and classification results of the same category level to obtain a hierarchical loss, and multi-level fusion processing is performed on the hierarchical losses corresponding to multiple category levels to obtain a multi-level loss.

For each category level, based on the classification results of the category level and the sample categories at that category level in the sample category information, the hierarchical loss of the category level is calculated. The hierarchical loss represents the classification model. The error between the output results at the category level and the expected results. The embodiments of this application do not limit the loss function used in loss calculation, and may be various loss functions used to solve classification problems, such as cross-entropy loss functions, etc.

In this way, the hierarchical losses corresponding to multiple category levels can be obtained. Then, these hierarchical losses are subjected to multi-level fusion processing to obtain multi-level losses. The multi-level loss represents the total error of the classification model at all category levels. The embodiment of the present application does not limit the method of multi-level fusion processing. For example, it can be weighted summation processing, and the weight of each category level can be set according to the actual application scenario.

In step 104, a classification model is trained based on multi-level losses; wherein the trained classification model is used to predict the classification results of the object to be tested at multiple category levels.

For example, on the basis of obtaining multi-level losses, the classification model is back-propagated based on the multi-level losses, and the weight parameters and offsets of the classification model are updated during the back-propagation process, thus achieving the purpose of training the classification model. In the process of backpropagation, the gradient descent (Gradient Descent) algorithm can be used to find the weight parameters and offsets that minimize the multi-level loss. The trained classification model has the ability to accurately perform level-by-level classification. Therefore, the trained classification model can be used to predict the classification results of the object to be tested at each category level.

In some embodiments, when training the classification model according to the multi-level loss, it also includes: stopping training the classification model when a preset stopping condition is met. Here, the stopping condition can be set in advance. For example, the stopping condition can be that the number of training rounds of the classification model reaches the training round threshold, or it can be that the performance index of the classification model reaches the performance index threshold. The performance index here can be accuracy or F1- score, etc. Through the above method, the training process can be precisely controlled, and the waste of computing resources can be reduced while ensuring the performance of the classification model.

As shown in Figure 3A, the embodiment of the present application comprehensively considers the hierarchical loss of each category level, so that the trained classification model has the ability to accurately perform level-by-level classification, which can improve the classification accuracy and avoid relying solely on the last category level. The problem caused by high classification error rate.

In some embodiments, see Figure 3B. Figure 3B is a schematic flowchart of the classification model training method provided by the embodiment of the present application. Step 102 shown in Figure 3A can be implemented through steps 201 to 202. Each step will be explained in combination. .

In step 201, sample description features are extracted from sample description information, and sample category features are extracted from sample category information.

Here, feature extraction is performed on the sample description information to obtain sample description features, and feature extraction is performed on the sample category information to obtain sample category features. The embodiments of the present application do not limit the method of feature extraction. For example, the type of information that requires feature extraction can be implemented through a specific model corresponding to the information type (the specific model here is part of the classification model). For example, for text (for example, For text in the sample description information or sample category information embodied in text form), feature extraction can be achieved through a text feature extraction model, such as the TextTransformer model; for pictures (such as pictures in the sample description information) In other words, feature extraction can be achieved through image feature extraction models, such as convolutional neural networks (Convolutional Neural Networks, CNN) or Vision Transformer models. It is worth noting that for sample description information and sample category information, the types of models used to implement feature extraction may be the same or partially the same, but the weight parameters and offsets in the models are different.

In some embodiments, the sample description information includes description information of multiple modalities; the above-mentioned extraction of sample description features from the sample description information can be achieved in this way: extracting the modality description from the description information of each modality Features; perform feature fusion processing on the modal description features corresponding to multiple modalities to obtain sample description features.

The sample description information may include description information in a single modality, such as only picture information or text information; it may also include description information in multiple modalities, such as picture information, text information, and sound information at the same time. Descriptive information can describe the sample object more comprehensively.

When the sample description information includes description information of multiple modalities, the modal description features can be extracted from the description information of each modality first, and then feature fusion processing is performed on all modal description features to obtain the sample description features. The method of feature fusion processing is not limited, for example, it can be direct summation or weighted summation. Through the above-mentioned multi-modal fusion, the comprehensiveness of the obtained sample description features can be improved, thereby improving the effect of model training and fully learning the data patterns between various modalities and categories.

In step 202, for any category level, the category probability distribution at any category level is calculated according to the sample description characteristics and the sample category characteristics as a classification result at any category level.

Here, for each category level, the category probability distribution is calculated based on the sample description characteristics and sample category characteristics, that is, the probability corresponding to each preset category in the category level is calculated. In this way, any category level obtained The dimension of the category probability distribution (the dimension of the category probability distribution represented in the form of a vector) is the same as the number of preset categories at any category level. It is worth noting that for different category levels, the process of obtaining the category probability distribution is independent of each other; it can be obtained in sequence from the first category level to the last category level in the category system. Category probability distribution at each category level.

In some embodiments, after extracting the sample description features from the sample description information and extracting the sample category features from the sample category information, it also includes: performing linear projection processing on the sample description features and sample category features; The sample description features and sample category features are normalized.

Here, linear projection processing can be performed on the sample description features and sample category features. The purpose of linear projection is to project the sample description features and sample category features into the same dimension to facilitate subsequent calculations. On the basis of linear projection, the sample description features and sample category features after linear projection can be normalized. Normalization refers to turning the dimensionless values in the features into dimensionless values, that is, Become a scalar, which can enhance the comparability between different features and improve the effect of model training. The embodiment of the present application does not limit the normalization method. For example, it may be L1 normalization or L2 normalization. Through the above method, the sample description features and sample category features can be made more suitable for calculation and comparison, and the model training effect can be effectively enhanced.

As shown in Figure 3B, by extracting features in the information and predicting the category probability distribution of each category level based on the extracted features, classification can be achieved while adapting to the actual situation of each category level.

In some embodiments, see Figure 3C. Figure 3C is a schematic flowchart of the classification model training method provided by the embodiment of the present application. Step 102 shown in Figure 3A can be implemented through steps 301 to 302. Each step will be explained in combination. .

In step 301, multiple sample description information and multiple sample category information are combined to obtain multiple information combinations; each information combination includes one sample description information and one sample category information.

In the embodiment of this application, the sample description information and sample category information of the sample object can be regarded as positive samples for model training. However, if the model training is only based on positive samples, it is easy for the classification model to fall into an overfitting state, which greatly affects the model training. Reduce the generalization ability of the classification model. Therefore, negative samples can be constructed and model training based on the idea of contrastive learning.

For example, when the number of sample objects includes multiple samples, multiple sample description information and multiple sample category information can be obtained. Furthermore, the multiple sample description information and multiple sample category information can be combined to obtain multiple samples. Information combinations, wherein each information combination includes a sample description information and a sample category information. The combination processing method can be exhaustive combination. For example, if there are N pieces of sample description information and N pieces of sample category information, N×N information combinations can be obtained.

To facilitate understanding, examples are given. The sample object includes sample object A and sample object B. Sample object A corresponds to sample description information A1 and sample category information A2. Sample object B corresponds to sample description information B1 and sample category information B2. After combination processing, it can Four information combinations are obtained, namely A1-A2, A1-B2, B1-A2, and B1-B2. There are two situations for the obtained information combination. The first situation is that the sample description information and sample category information in the information combination correspond to the same sample object (such as A1-A2, B1-B2). Information that meets this situation The combination is a positive sample; the second case is that the sample description information and sample category information in the information combination correspond to different sample objects (such as A1-B2, B1-A2), and the information combination that meets this situation is a negative sample.

In step 302, a classification model is used to perform level-by-level classification processing based on the information combination to obtain classification results of the information combination at multiple category levels.

For each information combination, the classification model is called to perform level-by-level classification processing based on the sample description information and sample category information in the information combination, that is, forward propagation processing is performed to obtain the classification results of the information combination at each category level.

In Figure 3C, step 103 shown in Figure 3A can be implemented through steps 303 to 306, which will be described in combination with each step.

In step 303, the category target label of the information combination at any category level is determined based on the sample category of the target sample object corresponding to the information combination at any category level.

Here, after determining the classification results at each category level corresponding to each information combination, the level loss is calculated at each category level based on the classification results. For ease of understanding, the following article takes the first category level as an example. illustrate.

First, for each information combination, determine the category target label of the information combination at the first category level based on the sample category of the target sample object corresponding to the information combination at the first category level. The category target label is Desired result.

The embodiment of this application does not limit the form of the class target label. For example, it can be in the form of a One-Hot vector. A One-Hot vector is a vector with only one element with a value of 1 and other elements with a value of 0. On this basis, in the process of determining the hierarchical loss of the first category level, the dimension of the obtained category target label is equal to the number (type) of the preset categories in the first category level. Each element of corresponds to a preset category in the first category level, and the element with a value of 1 corresponds to the sample category of the target sample object in the first category level. Correspondingly, the classification result of the first category level can also be in the form of a vector (such as a category probability distribution represented in vector form), and has the same dimensions as the category label of the first category level. The classification result The sum of the values of the multiple elements included is 1, and the value of each element represents the probability of belonging to the corresponding preset category.

For example, there are three preset categories in the first category level, namely "clothing", "food", and "electrical appliances". In the process of determining the level loss of the first category level, the determined categories The dimension of the target tag is 3, in the form [Element 1, Element 2, Element 3], where Element 1 corresponds to the category "Clothes", Element 2 corresponds to the category "Food", and Element 3 corresponds to the category "Electrical Appliances". Of course , the one-to-one correspondence between elements and categories can be adjusted and is not limited to the above examples. Using this example, if the classification result of a certain information combination at the first category level is [0.1, 0.7, 0.2], it means that the probability that the information combination belongs to "clothing" is 10%, and the probability of belonging to "food" is 70%, the probability of belonging to "electrical appliances" is 20%.

In some embodiments, when the sample description information and sample category information in the information combination correspond to the same sample object, the same sample object is determined as the target sample object corresponding to the information combination; when the sample description information and sample category information in the information combination When the target information corresponds to different sample objects, the target sample object corresponding to the information combination is determined as no object; among them, the sample category with no object at any category level is no category.

When the sample description information and sample category information in the information combination correspond to the same sample object, it is proved that the information combination is a positive sample, then the same sample object is determined as the target sample object corresponding to the information combination, and based on the target sample object Sample category to determine the category target label. Taking the above example again, if the sample category of the determined target sample object at the first category level is "clothing", then it is determined that the category target label of the corresponding information combination at the first category level is [1, 0 ,0].

When the sample description information and sample category information in the information combination correspond to different sample objects, it is proved that the information combination is a negative sample, and the target sample object corresponding to the information combination is determined to be no object. For no objects, the category label can be fixed, for example, a vector with all values 0. Taking the above example again, then the sample category of the first category level without objects is no category, and then determine The category target label of the corresponding information combination at the first category level is [0, 0, 0].

Through the above method, positive samples and negative samples can be accurately divided, training data can be expanded accurately and effectively, and accurate class target labels can also be obtained.

In step 304, the difference between the category label and the classification result of the information combination at any category level is determined as the information combination loss.

For each information combination, the difference between the category label (expected result) of the information combination at the first category level and the classification result (output result) of the information combination at the first category level is determined as the information For the combination loss, for example, the category label of the information combination at the first category level and the classification result of the information combination at the first category level can be substituted into the loss function to obtain the information combination loss.

In step 305, information combination fusion processing is performed on the information combination losses corresponding to multiple information combinations to obtain the hierarchical loss of any category level.

Here, after obtaining the information combination loss of each information combination at the first category level, these information combination losses are subjected to information combination fusion processing to obtain the hierarchical loss at the first category level. The embodiment of the present application does not limit the method of information combination and fusion processing. For example, it may be direct summation or weighted summation.

The above is only an example of the process of determining the hierarchical loss of the first category level. The process of determining the hierarchical loss of the remaining category levels can be performed by reference.

In step 306, multi-level fusion processing is performed on hierarchical losses corresponding to multiple category levels to obtain multi-level losses.

As shown in Figure 3C, the embodiment of the present application generates multiple information combinations based on the idea of contrastive learning and determines the corresponding class target labels, which can improve the richness of training data and enable the classification model to be fully trained based on positive samples and negative samples. It can effectively improve the generalization ability of the trained classification model and avoid overfitting.

In some embodiments, see FIG. 4 , which is a schematic flowchart of an object classification method provided by an embodiment of the present application, which will be described with reference to each of the steps shown.

In step 401, preset category information and test description information of the object to be tested are obtained; wherein the preset category information includes preset categories of multiple category levels.

Here, the classification model can be used to predict the classification results of the object to be tested at each category level. First, the preset category information and the test description information of the object to be tested can be obtained, where the classification model can be trained by the classification model training method provided by the embodiment of the present application. The preset category information can be determined based on the category system. For example, it can include all preset categories at each category level in the category system. Of course, categories can also be selected according to the needs in actual application scenarios. Some but not all of the default categories in the system can be added to the default category information. For example, if some default categories do not need to be classified, then these default categories do not need to be added to the default category information.

It is worth noting that the description information to be tested and the sample description information are of the same type. For example, they both include only picture information, or they both include picture information and text information.

In step 402, the classification model performs level-by-level classification processing based on the description information to be tested and the preset category information to obtain classification results of the object to be tested at multiple category levels.

Here, the classification model is called to perform level-by-level classification processing based on the description information to be tested and the preset category information to obtain the classification results of the objects to be tested at each category level. The hierarchical classification here may refer to classification in order from the first category level to the last category level.

In some embodiments, the above-mentioned classification process based on the description information to be tested and the preset category information through the classification model can be implemented in this way to obtain the classification results of the object to be tested at multiple category levels: The following processing is performed through the classification model: extracting the description features to be tested from the description information to be tested, and extracting the preset category features from the preset categories included in the preset category information; based on the description features to be tested and the preset category features , calculate the probability that the object to be tested belongs to the preset category corresponding to the preset category characteristics; for any category level, determine the probability that the object to be tested belongs to multiple preset categories at any category level as the probability to be tested. Classification results of test objects at any category level.

Here, for each preset category in the preset category information, the probability that the object to be tested belongs to each preset category is separately predicted. First, the description features to be tested are extracted from the description information to be tested, and the preset category features are extracted from the preset categories included in the preset category information, where there is a relationship between the preset categories and the preset category features. One-to-one relationship. It is worth noting that the preset category features here represent features of the preset category, and the preset category features themselves are not preset.

Then, based on the description characteristics to be tested and the preset category characteristics, the probability that the object to be tested belongs to the preset category corresponding to the preset category characteristics is calculated. If the category probability distribution is calculated based on the sample description features and sample category features when training the classification model, here, the category probability distribution can also be calculated based on the description features to be tested and the preset category features, and the category can be found. The probability corresponding to the preset category (here refers to the preset category corresponding to the preset category characteristics) in the category probability distribution.

For example, in the preset category information, there are three preset categories in the first category level, namely "clothing", "food", and "electrical appliances". For the "clothing" category, first extract the "clothing" category features, and then calculate the category probability distribution based on the description features to be measured and the "clothing" category features. Since the "clothing" category is located at the first category level, the dimension of the calculated category probability distribution is equal to the number of preset categories of the preset category information at the first category level, that is, the dimension is 3 , each element in the category probability distribution corresponds to a preset category of the preset category information at the first category level. If the elements in the category probability distribution correspond to "clothing", "food", and "electrical appliances" from front to back, and the category probability distribution is specifically [0.6, 0.3, 0.1], then it can be concluded that the object to be tested belongs to "clothing" The probability of the category is 60%. As for the values of other elements in the category probability distribution, there is no need to pay attention to the process of calculating the probability that the object to be tested belongs to the "clothing" category.

In this way, for any category level, the probability that the object to be tested belongs to each preset category of the category level can be obtained, and these probabilities are the classification results of the object to be tested at that category level. Through the above method, the probability that the object to be measured belongs to the corresponding preset category can be accurately calculated based on the description features to be measured and the preset category features, and accurate predictions can be achieved for each preset category.

In some embodiments, when the classification model is used to perform level-by-level classification processing based on the description information to be tested and the preset category information, it also includes: performing the following processing for any one category level: according to any one category level The classification results of the previous category level are filtered for multiple preset categories at any category level; the preset category information is updated according to the filtered preset categories.

In the process of classifying level by level in the order from the first category level to the last category level, for any category level except the first category level, you can The classification results of the previous category level are used to filter multiple preset categories at any category level.

For ease of understanding, the second category level is used as an example. For the second category level, the classification result of the first category level can be used to determine the category to which the object to be tested belongs at the first category level. Category (named target category for ease of differentiation), then you can filter out the default categories in the second category level that are associated with the target category of the first category level, and based on The default category information filtered out in the second category level updates the default category information, so that the second category level in the default category information only includes the filtered default category.

For example, in the default category information, the first category level includes "clothing" and "food" categories, and the "clothing" category is associated with the "tops" and "pants" categories at the second category level. , the "Food" category is associated with the "Vegetables" and "Fruits" categories at the second category level. If the target category of the first category level is "Clothes", then filter out the second category level. "Top" and "Pants" categories that are associated with the "Clothing" category, and update the default category information so that the default category information only includes "Top" and "Pants" at the second category level "Category.

Through the above method, layer-by-layer reasoning can be performed based on the category association relationships between different category levels, thereby reducing the amount of calculation and at the same time correlating target categories at different category levels.

In step 403, target categories of the object to be tested at multiple category levels are determined based on the classification results of the object to be tested at multiple category levels.

Here, for each category level, the target category of the object to be tested at the category level can be determined based on the classification result of the object to be tested at the category level. For example, for any category level, when the classification result of the object to be tested at any category level refers to the probability that the object to be tested belongs to each preset category of any category level, the value can be The preset category corresponding to the highest probability is used as the target category.

It is worth noting that in some embodiments, step 403 can be executed during the execution of step 402. For example, after obtaining the classification result of the object to be tested at a certain category level, it is determined that the object to be tested is classified at the category level. target category, thereby determining the classification result of the object to be tested at the next category level based on the inference principle.

As shown in Figure 4, the embodiment of the present application can accurately obtain the target category of the object to be tested at each category level through level-by-level classification processing and comprehensively consider the situation at each category level. According to the category system, the target category of the object to be tested can be obtained. Objects are accurately classified.

Below, an exemplary application of the embodiment of the present application in an actual application scenario will be described. To facilitate understanding, a commodity classification scenario is used as an example.

Step 1: Construct training data for the classification model.

Here, the data input to the classification model includes three types, namely picture information, text information and category information. Among them, picture information and text information are both description information of the product. The picture information can be the appearance of the product, and the text information can include the name of the product. , product introduction text, etc. Figure 5 is an example of a display diagram of a product in the e-commerce platform provided by the embodiment of the present application, including product appearance 51, product name 52, product introduction text 53 and category information 54, where the product appearance 51 includes six on the left Small image and a larger image on the right.

For sample products (products with labeled categories), the relevant picture information, text information, and category information can be identified from the display image of the sample product, or can be obtained directly from the database of the e-commerce platform.

In order to improve the model training effect, multiple sample products can be accurately identified, and the sample image information of each sample product (referring to the image information of the sample product, the same below), sample text information and sample category information can be obtained as the basis for the classification model. training data.

Step 2: Use the classification model to perform hierarchical classification based on the training data.

For ease of understanding, description will be made in conjunction with FIG. 6 .

Step 1) For each sample product, extract sample category features from the sample category information through the first text feature extraction model, extract sample image features from the sample image information through the image feature extraction model, and extract sample image features from the sample image information through the second text feature extraction model. Sample text features are extracted from the sample text information. Sample picture features, sample text features, and sample category features are recorded as I_f, T_f, and C_f respectively.

Among them, the first text feature extraction model, the picture feature extraction model, and the second text feature extraction model are all part of the classification model. The embodiments of this application do not limit the types of the text feature extraction model (the first text feature extraction model or the second text feature extraction model) and the picture feature extraction model. For example, the text feature extraction model can be a TextTransformer model, and the picture feature extraction model can be It is a CNN model or a Vision Transformer model.

Step 2) Perform feature fusion processing on I_f and T_f of the sample product. For example, a summation process can be performed to obtain the sample description feature I_f+T_f.

Step 3) Perform linear projection processing on I_f+T_f and C_f of the sample product to project to the same dimension. Then perform L2 normalization processing on the linearly projected I_f+T_f and C_f to obtain I (corresponding to I_f+T_f) and T (corresponding to C_f) respectively. Linear projection processing and L2 normalization processing are not shown in FIG. 6 .

Among them, the L2 normalization formula is as follows:

x＝[x ₁ ,x ₂ ,…,x _n ]

y＝[y ₁ ,y ₂ ,…,y _n ]

In the above formula, x represents the original feature vector, y represents the feature vector after L2 normalization, and n represents the dimension of the feature vector.

Step 4) Exhaustively combine multiple I and T to obtain multiple feature combinations. Each feature combination includes an I and a T, as shown in Figure 6, forming a feature combination matrix, where I1 represents The I and T1 of the first sample product represent the T of the first sample product, and the rest can be deduced by analogy; I1.T1 represents the feature combination composed of I1 and T1, and the rest can be deduced by analogy.

For each feature combination, the category probability distribution Logits at each category level are calculated based on the two features within it. The category probability distribution can be reflected in the form of a vector.

Step 3: Calculate multi-level loss.

For each category level, the level loss is calculated separately, and then all level losses are merged into multi-level losses. The same loss function can be used at each category level, such as the cross-entropy loss function. The formula is as follows:

In the above formula, L represents the loss value, y ⁽ⁱ⁾ represents the output result corresponding to the i-th sample product, represents the expected result corresponding to the i-th sample product, and N represents the number of sample products.

For ease of understanding, the process of obtaining the hierarchical loss of the first category level is given as an example. In the feature combination matrix, if two features in the feature combination correspond to the same sample product, then the feature combination is a positive sample (that is, the bold feature combination in Figure 6). Taking I1.T1 as an example, then I1.T1 corresponds to The category label Labels is a One-Hot vector. The element with a value of 1 in the One-Hot vector corresponds to the first sample product (that is, the same sample product corresponding to the two features in the feature combination) at the first category level. sample category; if the two features in the feature combination do not correspond to the same sample product, the feature combination is a negative sample, and the corresponding category label Labels are all 0 vectors. Logits are the output results, and Labels are the expected results. On this basis, traverse each feature combination in the feature combination matrix row by row (that is, in the order of I1.T1 to I1.TN, I2.T1 to I2.TN...IN.T1 to IN.TN), and you will traverse to The Logits and Labels corresponding to the feature combination are substituted into the above loss function, and then the obtained loss values are summed to obtain loss_i_1; by column (that is, according to I1.T1 to IN.T1, I1.T2 to IN.T2...I1. The order from TN to IN.TN) traverses each feature combination in the feature combination matrix, substitutes the Logits and Labels corresponding to the traversed feature combinations into the above loss function, and then sums the obtained loss values to obtain loss_t_1. Finally, calculate the hierarchical loss loss_1=(loss_i_1+loss_t_1)/2 of the first category level. This process is essentially a process of finding the loss value of each feature combination and summing it up.

Similarly, the hierarchical losses of subsequent category levels are calculated in sequence. Assuming that there are 5 categories in total, the multi-level loss Loss=x1*loss_1+x2*loss_2+x3*loss_3+x4*loss_4+x5*loss_5

Among them, loss_2 represents the hierarchical loss of the second category level, and the rest can be deduced by analogy; x1, x2, x3, x4, and x5 are custom parameters that can be determined according to the training situation. For example, they can be set to 3, 2, and 2 respectively. ,1,1.

Step 4: Train the classification model based on multi-level losses until the stopping condition is met.

Step 5: Use the trained classification model to predict the target categories to which the object to be tested belongs at multiple category levels based on inference principles. Among them, the inference principle refers to making predictions within the range of multiple preset categories associated with the target category in the next category level on the basis of determining the target category of a certain category level.

Through the embodiments of this application, at least the following technical effects can be achieved: 1) Give full play to the advantages of multi-modality (pictures and text), integrate the information of pictures and text, and improve the effect of model training compared with traditional text matching solutions; Classification accuracy; 2) Constructing a multi-level classification loss function can achieve level-by-level classification. Compared with the scheme of classifying directly at the last category level, the accuracy is higher, such as clothes-tops-T-shirts. Hierarchical recognition will be more accurate than directly identifying T-shirts.

The following continues to describe an exemplary structure in which the classification model training device 2551 provided by the embodiment of the present application is implemented as a software module. In some embodiments, as shown in FIG. 2A , the software module stored in the classification model training device 2551 of the memory 250 can Includes: sample acquisition module 25511, used to obtain sample description information and sample category information of sample objects; sample category information includes sample categories of sample objects at multiple category levels; sample processing module 25512, used to pass the classification model Perform level-by-level classification processing based on sample description information and sample category information to obtain classification results at multiple category levels; the loss calculation module 25513 is used to calculate the loss based on sample categories and classification results at the same category level. Hierarchical loss: perform multi-level fusion processing on hierarchical losses corresponding to multiple category levels to obtain multi-level loss; training module 25514 is used to train the classification model based on multi-level loss; among them, the trained classification model is used to predict the test Classification results of objects at multiple category levels.

In some embodiments, the sample processing module 25512 is also used to perform the following processing through the classification model: extract sample description features from sample description information, extract sample category features from sample category information; for any category level, according to The sample description features and sample category features calculate the category probability distribution at any category level as the classification result at any category level.

In some embodiments, the sample processing module 25512 is also used to: perform linear projection processing on the sample description features and sample category features; and perform normalization processing on the linearly projected sample description features and sample category features.

In some embodiments, the sample description information includes description information of multiple modalities; the sample processing module 25512 is also used to: extract modal description features from the description information of each modality; Perform feature fusion processing on state description features to obtain sample description features.

In some embodiments, the number of sample objects includes multiple; the sample processing module 25512 is also used to: combine multiple sample description information and multiple sample category information to obtain multiple information combinations; each information combination includes A sample description information and a sample category information; the classification model performs level-by-level classification processing based on the information combination to obtain the classification results of the information combination at multiple category levels; the loss calculation module 25513 is also used to target any category level, perform the following processing: according to the sample category of the target sample object corresponding to the information combination at any category level, determine the category target label of the information combination at any category level; determine the category of the information combination at any category level The difference between the target label and the classification result is used as the information combination loss; the information combination losses corresponding to multiple information combinations are processed by information combination fusion to obtain the hierarchical loss at any category level.

In some embodiments, the loss calculation module 25513 is also used to: when the sample description information and sample category information in the information combination correspond to the same sample object, determine the same sample object as the target sample object corresponding to the information combination; when the information combination When the sample description information and sample category information in the combination correspond to different sample objects, the target sample object corresponding to the information combination is determined as no object; among them, the sample category with no object at any category level is no category.

In some embodiments, as shown in FIG. 2B , the software module stored in the object classification device 2552 of the memory 250 may include: a target acquisition module 25521 for acquiring preset category information and test description information of the object to be tested. ; Among them, the preset category information includes preset categories for each category level in multiple category levels; the target processing module 25522 is used to perform layer-by-layer based on the description information to be tested and the preset category information through the classification model level classification processing to obtain the classification results of the object to be tested at multiple category levels; the determination module 25523 is used to determine the classification results of the object to be tested at multiple category levels based on the classification results of the object to be tested at multiple category levels. Head.

In some embodiments, the target processing module 25522 is also configured to perform the following processing for any category level: based on the classification result of the previous category level of any category level, perform multiple pre-processing for any category level. Set categories for filtering processing; update the default category information based on the filtered default categories.

In some embodiments, the target processing module 25522 is also configured to perform the following processing through the classification model: extract the description features to be tested from the description information to be tested, and extract the preset categories from the preset categories included in the preset category information. Characteristics; based on the description characteristics to be tested and the preset category characteristics, calculate the probability that the object to be tested belongs to the preset category characteristics corresponding to the preset category; for any category level, the object to be tested belongs to any category level. The probabilities of multiple preset categories are determined as the classification results of the object to be tested at any category level.

Embodiments of the present application provide a computer program product or computer program. The computer program product or computer program includes executable instructions, and the executable instructions are stored in a computer-readable storage medium. The processor of the electronic device reads the executable instructions from the computer-readable storage medium, and the processor executes the executable instructions, so that the electronic device executes the classification model training method or object classification method described in the embodiments of the present application.

Embodiments of the present application provide a computer-readable storage medium storing executable instructions. The executable instructions are stored therein. When the executable instructions are executed by a processor, they will cause the processor to execute the method provided by the embodiments of the present application, such as , the classification model training method shown in Figure 3A, Figure 3B and Figure 3C, or the object classification method shown in Figure 4.

In some embodiments, the computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface memory, optical disk, or CD-ROM; it may also include one or any combination of the above memories. Various equipment.

In some embodiments, executable instructions may take the form of a program, software, software module, script, or code, written in any form of programming language, including compiled or interpreted languages, or declarative or procedural languages, and their May be deployed in any form, including deployed as a stand-alone program or deployed as a module, component, subroutine, or other unit suitable for use in a computing environment.

As an example, executable instructions may, but do not necessarily correspond to, files in a file system and may be stored as part of a file holding other programs or data, for example, in a Hyper Text Markup Language (HTML) document. in one or more scripts, stored in a single file specific to the program in question, or in multiple collaborative files (e.g., files storing one or more modules, subroutines, or portions of code).

As examples, executable instructions may be deployed to execute on one computing device, or on multiple computing devices located at one location, or alternatively, on multiple computing devices distributed across multiple locations and interconnected by a communications network execute on.

The above are only examples of the present application and are not used to limit the protection scope of the present application. Any modifications, equivalent substitutions and improvements made within the spirit and scope of this application are included in the protection scope of this application.

Claims (10)

1. An object classification method, comprising:

acquiring preset category information and description information to be tested of an object to be tested; wherein the preset category information comprises preset categories of a plurality of category levels;

performing hierarchical classification processing based on the description information to be detected and the preset category information through a classification model to obtain classification results of the object to be detected in the category levels;

and determining target categories of the object to be tested in the category levels according to classification results of the object to be tested in the category levels.

2. The method according to claim 1, wherein the method further comprises:

For any one category level, the following processing is performed:

screening a plurality of preset categories of any category hierarchy according to the classification result of the previous category hierarchy of the any category hierarchy;

and updating the preset category information according to the screened preset categories.

3. The method according to claim 1, wherein the step-by-step classification processing is performed by a classification model based on the description information to be measured and the preset category information to obtain classification results of the object to be measured at the plurality of category levels, including:

the following processing is performed by the classification model:

extracting description features to be detected from the description information to be detected, and extracting preset category features from preset categories included in the preset category information;

according to the description characteristic to be detected and the preset category characteristic, calculating the probability that the object to be detected belongs to the preset category corresponding to the preset category characteristic;

and aiming at any category level, determining the probability that the object to be detected respectively belongs to a plurality of preset categories of the any category level as a classification result of the object to be detected in the any category level.

4. A method of training a classification model, comprising:

acquiring sample description information and sample category information of a sample object; the sample category information comprises sample categories of the sample object at a plurality of category levels;

performing hierarchical classification processing based on the sample description information and the sample category information through a classification model to obtain classification results at the category levels;

performing loss calculation according to sample categories and classification results of the same category level to obtain level loss, and performing multi-level fusion processing on the level loss corresponding to each category level to obtain multi-level loss;

training the classification model according to the multi-level loss; the trained classification model is used for predicting classification results of the object to be tested in the category levels.

5. The method of claim 4, wherein the performing, by a classification model, a hierarchical classification process based on the sample description information and the sample category information to obtain classification results at the plurality of category levels comprises:

the following processing is performed by the classification model:

extracting sample description features from the sample description information and extracting sample category features from the sample category information;

And calculating category probability distribution at any category level according to the sample description characteristics and the sample category characteristics aiming at any category level, and taking the category probability distribution as a classification result at any category level.

6. The method of claim 5, wherein the extracting sample description features from the sample description information, and wherein after extracting sample category features from the sample category information, the method further comprises:

performing linear projection processing on the sample description characteristic and the sample category characteristic;

and carrying out normalization processing on the sample description characteristic and the sample category characteristic after linear projection.

7. The method of claim 5, wherein the sample description information includes description information of a plurality of modalities; the extracting the sample description feature from the sample description information comprises the following steps:

extracting a mode description characteristic from the description information of each mode;

and carrying out feature fusion processing on the mode description features respectively corresponding to the multiple modes to obtain sample description features.

8. The method of claim 4, wherein the number of sample objects comprises a plurality; the step-by-step classification processing is performed by a classification model based on the sample description information and the sample category information to obtain classification results at the plurality of category levels, including:

Combining the plurality of sample description information and the plurality of sample category information to obtain a plurality of information combinations; each information combination comprises a sample description information and a sample category information;

performing hierarchical classification processing based on information combination through the classification model to obtain classification results of the information combination in the category levels;

the step of calculating the loss according to the sample category and the classification result of the same category level to obtain the level loss comprises the following steps:

for any one category level, the following processing is performed:

determining a category label of the information combination at any category level according to the sample category of the target sample object corresponding to the information combination at any category level;

determining the difference between the category label of any category level and the classification result of the information combination as the information combination loss;

and carrying out information combination fusion processing on the information combination losses corresponding to the plurality of information combinations respectively to obtain the hierarchy loss of any category hierarchy.

9. An electronic device, comprising:

a memory for storing executable instructions;

a processor for implementing the object classification method of any one of claims 1 to 3 or the classification model training method of any one of claims 4 to 8 when executing executable instructions stored in the memory.

10. A computer readable storage medium storing executable instructions for implementing the object classification method of any one of claims 1 to 3 or the classification model training method of any one of claims 4 to 8 when executed by a processor.