CN117349669A

CN117349669A - Model training method, category detection method and computing device

Info

Publication number: CN117349669A
Application number: CN202311262808.2A
Authority: CN
Inventors: 马一凡
Original assignee: Hangzhou Alibaba Overseas Internet Industry Co ltd
Current assignee: Hangzhou Alibaba Overseas Internet Industry Co ltd
Priority date: 2023-09-27
Filing date: 2023-09-27
Publication date: 2024-01-05
Also published as: WO2025066434A1

Abstract

The embodiment of the application provides a model training method, a category detection method and computing equipment. Taking the object information of the issued object and the category path of the object as a training data set; selecting a plurality of training sample pairs of a current training batch from the training data set; the training sample pair comprises object information and corresponding category paths; inputting a plurality of training sample pairs into an object processing model, taking the training sample pairs as positive samples, taking object information in one positive sample and category paths in the other positive sample as negative samples, and training the object processing model; the object processing model is used for extracting object characteristics from object information of a target object and extracting category characteristics from a target category path of the target object; the object features and the category features are used to calculate the matching degree of the target object and the target category path. The technical scheme provided by the embodiment of the application realizes effective and accurate detection.

Description

Model training method, category detection method and computing device

Technical Field

The embodiment of the application relates to the technical field of computers, in particular to a model training method, a category detection method and computing equipment.

Background

In some online systems that provide objects for users to perform interactive actions, a hierarchical category structure is typically employed to organize and categorize the objects, where the category hierarchy typically includes a primary category and a secondary category … … leaf categories, where the objects are bound to the leaf categories, and the path from the primary category to the leaf category is a category path, with each object corresponding to a category path. When an online system publishes an object, an object provider typically selects an appropriate category path for the object to deliver. In practical application, due to the influence of factors such as cognitive errors of an object provider or update of category structures, the category path where the object is located may be wrong, that is, the object and the category path are misplaced, which may cause that the object appears in an irrelevant search result page, etc. under a search scene, so how to effectively detect whether the object is matched with the corresponding category path is a technical problem to be solved at present.

Disclosure of Invention

The embodiment of the application provides a model training method, a category detection method and computing equipment, which are used for solving the technical problem of how to effectively detect whether paths of objects and categories are matched in the prior art.

In a first aspect, an embodiment of the present application provides a model training method, including:

taking object information of the released object and a category path of the object as a training data set;

selecting a plurality of training sample pairs of a current training batch from the training data set; wherein the training sample pair comprises object information and corresponding category paths;

inputting the training sample pairs into an object processing model, taking the training sample pairs as positive samples, taking object information in one positive sample and category paths in the other positive sample as negative samples, and training the object processing model;

the object processing model is used for extracting object characteristics from object information of a target object and extracting category characteristics from a target category path of the target object; the object features and the category features are used for calculating the matching degree of the target object and the target category path.

Optionally, the inputting the plurality of training sample pairs into the object processing model, taking the training sample pairs as positive samples, wherein object information in one positive sample and category paths in another positive sample are negative samples, and training the object processing model includes:

Inputting a plurality of object information in the plurality of training sample pairs into the object processing model to obtain a plurality of first sample features;

inputting a plurality of category paths in the plurality of training sample pairs into the object processing model to obtain a plurality of second sample features;

performing matrix multiplication processing on a first matrix formed by the first sample features and a second matrix formed by the second sample features to obtain a third matrix; wherein the value of the ith row and the jth column in the third matrix represents the similarity between the ith object and the jth category path; wherein i=1, 2, 3 … … n, j=1, 2, 3 … … n; i is different from j; n is a positive integer to represent the number of pairs of samples of the plurality of training sample pairs;

and training the object processing model in a training mode of maximizing the value corresponding to the positive sample in the third matrix and minimizing the value corresponding to the negative sample.

Optionally, the method further comprises:

constructing a category correct sample pair and constructing a category misplaced sample pair; wherein the category correct sample pair comprises a sample object and a correct category path of the sample object; the category misplaced sample pair includes a sample object and an erroneous category path for the sample object;

Training a category identification model by using the correct category sample pair and the misplaced category sample pair respectively;

the category identification model is used for determining misplacement probability of the target object and the target category path based on the target object and the target category path; the misplacement probability is used for determining whether the target object and the target category path are misplaced or not in combination with the matching degree.

Optionally, the constructing the category correct sample pair includes one or more of the following:

constructing a correct category sample pair according to the object information of the manually maintained object and the corresponding category path;

according to the object information which is manually detected as the category correct object and the corresponding category path, constructing a category correct sample pair;

the method comprises the steps of,

constructing a correct category sample pair according to the object information of the object, the matching degree of which is determined by using the object processing model and meets the matching condition, and the corresponding category path;

the build category misplaced sample pairs include one or more of the following implementations:

according to the object information which is manually detected as the category error object and the corresponding category path, a category error sample pair is constructed;

The method comprises the steps of,

and constructing a category misplaced sample pair according to the object in the category correct sample pair and the similar category path corresponding to the category path in the category correct sample pair.

Optionally, the category identification model includes a first identification network and a second identification network;

the training category identification model comprises the following steps of:

selecting a plurality of training sample pairs including category correct sample pairs and/or category misplaced sample pairs;

inputting a plurality of object information in the training sample pairs into a first recognition network to obtain a plurality of first sample characteristics, and inputting a plurality of category paths into a second recognition model to obtain a plurality of second sample characteristics;

splicing the plurality of first sample features and the plurality of second sample features to obtain fusion features;

respectively carrying out dimension reduction treatment on the fusion characteristics to obtain output data;

calculating a second loss value by using the output data and the misplacement probability of different objects and paths of respective corresponding categories;

and updating the model parameters of the category identification model according to the second loss value.

Optionally, the method for generating the training data set includes the steps of:

Determining a published object and a category path of the object;

and adding the object information of the object and the path of the category corresponding to the object into a training data set under the condition that the object meets the screening condition for any object.

Optionally, the method further comprises:

and taking the sample object in the category correct sample pair and the category path as a training sample pair, and adding a training data set to retrain the object processing model.

Optionally, the screening condition includes one or more of the following implementations:

the exposure is greater than a first exposure threshold;

the method comprises the steps of,

the object score determined in conjunction with the historical interaction data of the object is greater than a second score threshold; wherein the historical interaction data comprises the number of actions of at least one interaction action type; the higher the number of acts, the higher the subject score.

Optionally, the training the object processing model in a training manner of maximizing the value corresponding to the positive sample in the third matrix and minimizing the value corresponding to the negative sample includes:

according to the value of the ith row and the jth column in the third matrix, calculating a first loss value of the current training batch by utilizing a contrast learning loss function;

And updating the object processing model according to the first loss value.

Optionally, the method for using the object information of the published object and the path of the category corresponding to the object as the training data set includes:

selecting a predetermined number of objects from under any category path of the on-line system;

adding the predetermined number of objects and the category path to a training dataset.

Optionally, the method further comprises:

and under the condition that the target category path is the correct category path, taking the target object and the target category path as training sample pairs to be added into the training data set so as to retrain the object processing model.

Optionally, the constructing a category misplaced sample pair according to the object in the category correct sample pair and the similar category path corresponding to the correct category path in the category correct sample pair includes:

determining a correct category path in the category correct sample pair;

determining a category path which is the same as the first category of the correct category path and has the word overlapping degree larger than the overlapping threshold value as a similar category path;

and forming a misplaced sample pair of the obtained category by the object information of the object in the correct sample pair of the category and the similar category path.

In a second aspect, an embodiment of the present application provides a generic detection method, including:

acquiring object information of a target object and a target category path corresponding to the target object;

extracting target object characteristics from object information of the target object by using an object processing model; the object processing model takes a plurality of training sample pairs as positive samples, object information in one positive sample and category paths in the other positive sample as negative samples, and is obtained through training; the training sample pairs are selected from the training data set to obtain; the training data set is composed of object information of the issued object and category paths of the object;

extracting target category features from the target category path using the object processing model;

calculating the matching degree of the target object characteristics and the target category characteristics;

and detecting whether the target object and the category path are misplaced according to the matching degree.

Optionally, the method further comprises:

inputting the object information and the target category path into a category identification model, and calculating by using the category identification model to obtain misplacement probability; the category identification model is obtained by training a category misplaced sample pair by using the constructed and generated category correct sample pair;

Determining whether the target object and the category path are misplaced according to the matching degree comprises:

and determining whether the target object and the category path are misplaced according to the matching degree and the misplacement probability.

Optionally, the determining whether the target object and the category path are misplaced according to the matching degree and the misplacement probability includes:

determining that the target category path is an error category path of the target object under the condition that the matching degree is smaller than a first matching threshold value or the matching degree is smaller than a second matching threshold value and the misplacement probability is larger than a first probability threshold value; the second matching threshold is greater than the first matching threshold;

and determining that the target category path is the correct category path of the target object under the condition that the matching degree is larger than the second matching threshold or the misplacement probability is smaller than the first probability threshold.

Optionally, the method further comprises:

determining a plurality of objects under any category path based on search keyword hits;

respectively taking the plurality of objects as the target objects;

after detecting whether the target object and the category path are misplaced according to the matching degree, the method further comprises:

Selecting one or more objects of the correct category paths as recall objects of the search keywords according to the detection results of the objects;

or,

and selecting one or more objects of the correct category path according to the detection results of the objects, and selecting an object with the exposure larger than a second exposure threshold and/or the sales larger than a first sales threshold from the one or more objects of the incorrect category path as a recall object of the search keyword.

Optionally, the method further comprises:

setting an error mark for the target object according to the detection result of the target object when the target category path is an error category path;

in the case that any category path is hit based on a search keyword, an object with an error flag set under the any category path is not taken as a recall object of the search keyword.

Optionally, the method further comprises:

and providing the object with the exposure larger than the second exposure threshold or with the sales larger than the first sales threshold to related personnel so as to enable the related personnel to manually detect whether the object and the corresponding category path are misplaced.

Optionally, the obtaining the object information of the target object and the target category path corresponding to the target object includes:

Determining a target category path selected by an object provider and object information of the target object according to a category selection request of the object provider for the target object;

the method further comprises the steps of:

and sending misplacement prompt information to the object provider according to the detection result of the target object under the condition that the target category path is an error category path.

In a third aspect, embodiments of the present application provide a computing device comprising a processing component and a storage component;

the storage component stores one or more computer instructions; the one or more computer instructions are operable to be invoked by the processing component to implement the model training method as described in the first aspect above or the category detection method as described in the second aspect above.

In a fourth aspect, embodiments of the present application provide a computer readable storage medium storing a computer program, which when executed by a computer, implements a model training method as described in the first aspect or implements a category detection method as described in the second aspect.

In the embodiment of the application, a published object and a category path corresponding to the object are used as a training data set, a batch training mode is adopted, a plurality of training sample pairs are selected each time, each training sample pair comprises object information of the object and a category path corresponding to the object, the plurality of training sample pairs are input into an object processing model, the training sample pairs are used as positive samples, and a combination of the object information in one positive sample and the category path in another positive sample is used as a negative sample to train the object processing model; the object processing model is used for extracting object characteristics from object information of a target object and extracting category characteristics from a target category path of the target object; the object features and the category features are used for calculating the matching degree of the target object and the target category path. The model is trained by directly utilizing the object and category paths issued by the online system, so that a training data set is not required to be constructed, labels are not required to be marked, the training cost is reduced, the noise influence can be reduced by adopting a positive sample and negative sample comparison training mode, the model can learn the association between the object information and the category paths, the model processing accuracy can be ensured, and the purpose of effectively and accurately detecting the object and the category is finally realized. These and other aspects of the present application will be more readily apparent from the following description of the embodiments.

Drawings

In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief description will be given below of the drawings that are needed in the embodiments or the prior art descriptions, and it is obvious that the drawings in the following description are some embodiments of the present application, and that other drawings can be obtained according to these drawings without inventive effort for a person skilled in the art.

FIG. 1 illustrates a flow chart of one embodiment of a model training method provided herein;

FIG. 2a shows a training schematic of an object handling model in one practical application of an embodiment of the present application;

FIG. 2b shows a computational schematic of an object handling model in one practical application of an embodiment of the present application;

FIG. 3 is a schematic diagram showing training of a category identification model in one practical application of an embodiment of the present application;

FIG. 4 illustrates a flow chart of one embodiment of a generic visual detection method provided herein;

FIG. 5a shows a schematic view of a scene interaction in a practical application of the embodiment of the present application;

FIG. 5b is a schematic view of another scene interaction in a practical application according to an embodiment of the present application;

FIG. 6 is a schematic diagram of a category detection flow in a practical application according to an embodiment of the present application;

FIG. 7 is a schematic diagram of an embodiment of a model training apparatus provided herein;

FIG. 8 is a schematic diagram illustrating the construction of one embodiment of a generic visual inspection device provided herein;

FIG. 9 illustrates a schematic diagram of one embodiment of a computing device provided herein.

Detailed Description

In order to enable those skilled in the art to better understand the present application, the following description will make clear and complete descriptions of the technical solutions in the embodiments of the present application with reference to the accompanying drawings in the embodiments of the present application.

In some of the flows described in the specification and claims of this application and in the foregoing figures, a number of operations are included that occur in a particular order, but it should be understood that the operations may be performed in other than the order in which they occur or in parallel, that the order of operations such as 101, 102, etc. is merely for distinguishing between the various operations, and that the order of execution is not by itself represented by any order of execution. In addition, the flows may include more or fewer operations, and the operations may be performed sequentially or in parallel. It should be noted that, the descriptions of "first" and "second" herein are used to distinguish different messages, devices, modules, etc., and do not represent a sequence, and are not limited to the "first" and the "second" being different types.

The technical scheme of the embodiment of the application can be applied to an online system for providing the object for the user to execute the interactive behavior, for example, the object can refer to commodities, the online system is an online transaction platform, and the user can execute the interactive behavior such as object purchase, collection, additional purchase and the like. Of course, the object may also refer to content such as an article, news, etc., and the online system may be information providing information for viewing by a user, which is not limited in this application.

As described in the background art, how to effectively detect misplaced objects is a technical problem that needs to be solved at present. The inventor finds that in the process of implementing the application, detection of misplaced objects of categories can be considered to be implemented by using model capability, for example, a processing model for implementing text matching or text classification is implemented, however, such models are usually supervised models, for example, a supervised text matching model is adopted to calculate misplaced probability of the object and paths of the categories, the supervised models need to be obtained by training a large number of pre-labeled training samples, then the training samples are difficult to obtain, manual labeling workload is large, corresponding categories which cannot be covered in training data sets cannot be identified, and therefore, the models cannot be accurately detected.

In order to realize effective and accurate detection of misplaced objects, the inventor has researched a series of technical schemes of the application, in the embodiment of the application, the published objects and the corresponding class paths of the objects are directly used as training data sets, a training batch mode is adopted, a plurality of training sample pairs are selected each time, each training sample pair consists of object information of the objects and the corresponding class paths, a plurality of training sample pairs are input into an object processing model, each training sample pair is used as a positive sample, and the combination of the object information in one positive sample and the class paths in the other positive sample is used as a negative sample, so that the object processing model is trained; the object processing model can be used for extracting object characteristics from object information of the target object and extracting category characteristics from a target category path of the target object; the object features and the category features are used for calculating the matching degree of the target object and the target category path. According to the embodiment of the application, the model training is directly carried out by using the issued object and category path, so that a training data set is not required to be constructed, a label is not required to be marked manually, and the training cost is reduced; the published object and category paths still contain a large number of accurate objects although noise exists, and model training is performed according to the noise, so that the model has good generalization capability, hidden association information between the object and category paths can be learned, the model processing accuracy can be ensured, and the model can be free from noise influence through a comparison training mode of positive samples and negative samples, and effective detection is finally realized. The following description of the embodiments of the present application will be made clearly and fully with reference to the accompanying drawings, in which it is evident that the embodiments described are only some, but not all, of the embodiments of the present application. All other embodiments, which can be made by those skilled in the art based on the embodiments herein without making any inventive effort, are intended to be within the scope of the present application.

Fig. 1 is a flowchart of an embodiment of a model training method provided in the embodiments of the present application, where the technical solution of the present embodiment may be executed by a processing end, and the processing end may be a server end in an online system, or may be other nodes independent of the server end of the online system.

In practical applications, the online system is generally composed of a first user terminal, a second user terminal and a server terminal, where the first user terminal and the second user terminal respectively establish a connection with the server terminal through a network. The network provides a medium for communication links between the first user terminal and the second user terminal and the service terminal respectively. The network may include various connection types, such as wired, wireless communication links, or fiber optic cables, among others.

The first user side may be an object-oriented provider operated by the object provider to publish objects, etc., and the second user side may be a consumer-oriented user for the consumer to conduct interactive actions such as object searching, browsing, etc.

The first user terminal and the second user terminal can interact with the service terminal through the network to receive or send messages and the like. For example, the first user side can sense category selection operation, release operation and the like of the object provider, and send a corresponding request to the server side for the server side to perform corresponding processing; the second user side can sense the interactive behavior executed by the consumer user and send a corresponding interactive request to the server side, and the server side can process the interactive request and feed back a processing result to the second user side.

The first client or the second client may be a browser, an APP (Application), or a web Application such as H5 (HyperText Markup Language5, 5 th edition of hypertext markup language) Application, or a light Application (also called applet, a lightweight Application) or cloud Application, etc., and the first client or the second client may be deployed in an electronic device, and needs to run depending on the device or some APPs in the device. The electronic device may for example have a display screen and support information browsing etc. as may be a personal mobile terminal such as a mobile phone, tablet, personal computer, desktop computer, smart phone, smart watch etc.

The processing end or the service end may include a server that provides various services, for example, a server for background training that provides support for a model used on the user end 101, and for example, a server that processes interaction information sent by the user end.

It should be noted that, the processing end or the server end may be implemented as a distributed server cluster formed by a plurality of servers, or may be implemented as a single server. The server may also be a server of a distributed system or a server that incorporates a blockchain. The server may also be a cloud server that provides cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (Content Delivery Network, CDN), and basic cloud computing services such as big data and artificial intelligence platforms, or an intelligent cloud computing server or intelligent cloud host with artificial intelligence technology.

It should be noted that, in the embodiments of the present application, the use of user data may be involved, and in practical applications, user specific personal data may be used in the schemes described herein within the scope allowed by applicable legal regulations in the country where the applicable legal regulations are met (for example, the user explicitly agrees to the user to actually notify the user, etc.).

It should be noted that, in the embodiment of the present application, the user described in the present application is generally referred to as a "virtual user" in a network virtual environment, and the real user may register a user account in a server through a registration manner to obtain a user identity in the network environment.

The model training method provided by the embodiment shown in fig. 1 may include the following steps:

101: and taking the object information of the published object and the category path of the object as a training data set.

The object provider may request that the online system publish an object, i.e., specifically, an object published in the online system, to upload the object in the online system. The objects published by the online system may refer to the full amount of online system on-shelf objects, i.e., all current objects in the online system that are available for users to perform interactive actions.

In order to facilitate the user to understand the object, the online system will generally provide an object detail page of the object to introduce the object related information, where the object detail page may include basic information such as an object title, a detail content, an object picture, etc., so that the object information may include the object title, the detail content, and/or the object picture, for example. In practical applications, since the object titles are generally refined and generated in combination with the characteristics of the object, the object information may at least include the object titles, and one object title may be expressed in text form, for example, "men shoes 2023, fashion and Korean style full-size basketball shoes, men running shoes, and the like.

The category path of the object may refer to a category path selected for the object when the object provider issues the object, and in order to facilitate the object provider to select the object, optionally, multiple category paths matched with the object may be determined according to related information of the object requested to issue by the object provider, and the multiple category paths may be recommended to the object provider, so that the object provider may select the category path to which the object belongs for the object.

The category path is composed of a primary category, a secondary category … … to a leaf category, for example, one category path may be expressed as: fashion- & gt men- & gt shoes- & gt sports shoes, "sports shoes" are leaf categories for binding an object, "fashion" is a first class category, and the second class category "men" and the third class category "shoes" are required to pass from "fashion" to "sports shoes".

In addition, in order to improve the quality of the training data set, the data may be cleaned, and optionally, taking the object information of the published object and the path of the corresponding category of the object as the training data set may include: determining a category path of an object published by the online system and the object; and adding object information of the object and a path of the category corresponding to the object into the training data set under the condition that the object meets the screening condition for any object.

And adding the object information category corresponding to the object meeting the screening condition into the training data set.

The screening conditions may include, for example, one or more of the following implementations:

the exposure is greater than a first exposure threshold;

the method comprises the steps of,

The exposure of the object may refer to, for example, the number of times the object is browsed by the user.

The at least one interaction behavior type may include browsing, collecting, adding to shopping cart or purchasing, and the like, and the object may be scored by counting the number of behaviors of the at least one interaction behavior type, where the object score represents the object quality to a certain extent, and the higher the object score, the better the object quality. The scoring method is various, and may be implemented by a scoring rule or a model, which is not limited in the present application.

Optionally, since the number of objects under the leaf sub-category appears as long tail distribution, the number of objects is often larger, and optionally, taking the object information of the published object and the path of the corresponding category of the object as the training data set may be: selecting a predetermined number of objects from under any category path of the on-line system; a predetermined number of objects and category paths are added to the training dataset.

That is, only a predetermined number of objects are selected as training data under each category path, alternatively, a predetermined number of objects may be randomly selected, and the predetermined number is 5 ten thousand in practical application.

102: a plurality of training sample pairs for a current training batch are selected from the training dataset.

The training sample pair is composed of object information and corresponding category paths.

In the training data set, object information of each object and a corresponding category path form a group of training sample pairs.

Model training may be performed in a batch-wise fashion, and multiple training sample pairs may be selected from the training dataset for the current training batch. The number of the sample pairs of the plurality of training sample pairs can be set in combination with the actual running situation, and generally, the larger the number of the sample pairs is, the more accurate the model is, and the relatively higher the calculated amount is.

103: and inputting a plurality of training sample pairs into the object processing model, taking the training samples as positive samples, taking object information in one positive sample and category paths in the other positive sample as negative samples, and training the object processing model.

After a plurality of training sample pairs are selected, model training can be performed by using the plurality of training sample pairs, in the embodiment of the application, an unsupervised training mode is adopted, each training sample pair is used as a positive sample, and object information in any one positive sample and category paths in another positive sample form a negative sample to train an object processing model.

For example, assume that the data set formed by the training sample pairs is [ (t) ₁ ，c ₁ )(t ₂ ，c ₂ )……(t _n ，c _n )]T represents object information, c represents a category path; the ith object information and the ith category path form a positive sample, and the ith object information and the rest category paths except the ith category path can form a negative sample. Where i=1, 2, 3 … … n, n is a positive integer, to represent the number of pairs of training pairs.

Specifically, model training can be performed by adopting a contrast learning (Contrastive Learning) mode, so that the matching degree corresponding to the positive sample is maximized and the matching degree corresponding to the negative sample is minimized as a training target, and the object is trained to process the model.

The object processing model can be used for extracting object characteristics from object information of the target object and extracting category characteristics from a target category path of the target object; the object features and the category features are used for calculating the matching degree of the target object and the target category path.

The target object is any object to be detected. And combining the matching degree of the target object and the target category path, namely determining whether the target category path is a misplaced category distance, so that the effective and accurate detection of misplaced objects can be realized.

In the embodiment, the model training is directly performed by using the object and category path issued by the online system, so that a training data set is not required to be constructed, a label is not required to be marked manually, and the training cost is reduced; the on-line system can release the object and the category path, and the on-line system can release the category path of the object and the category path of the object.

In some embodiments, the inputting a plurality of training sample pairs into the object processing model, taking each training sample pair as a positive sample, wherein object information in one positive sample and a category path in another positive sample are negative samples, and the training object processing model includes:

inputting a plurality of object information in a plurality of training sample pairs into an object processing model to obtain a plurality of first sample features; inputting a plurality of category paths in a plurality of training sample pairs into an object processing model to obtain a plurality of second sample features; performing matrix multiplication processing on a first matrix formed by a plurality of first sample features and a second matrix formed by a plurality of second sample features to obtain a third matrix; wherein the value of the ith row and the jth column in the third matrix represents the similarity between the ith object and the jth category path; wherein i=1, 2, 3 … … n, j=1, 2, 3 … … n; i is different from j; n is a positive integer representing the number of pairs of training samples; and training the object processing model in a training mode of maximizing the value corresponding to the positive sample in the third matrix and minimizing the value corresponding to the negative sample.

Wherein each training sample pair can be represented as (t _i ，c _i )，t _i Representing the ith object information, c _i Representing an i-th category path; the data set formed by the training sample pairs is [ (t) ₁ ，c ₁ )(t ₂ ，c ₂ )……(t _n ，c _n )]The method comprises the steps of carrying out a first treatment on the surface of the Wherein n pieces of object information in n training sample pairs are input into the object processing model, n first sample features can be obtained, and the n first sample features can form a first matrix et= [ T ] ₁ ，T ₂ ，……，T ₂ ]Likewise, n category paths in n training sample pairs may be input into the object processing model to obtain n second sample features, which may formSecond matrix ec= [ C ₁ ，C ₂ ，……，C ₂ ]. Multiplying the first matrix by the second matrix, i.e. ET EC, yields a third matrix of n, which can be expressed as:

the value T of the diagonal position in the third matrix ₁ C ₁ ，T ₂ C ₂ ，…，T _n C _n Respectively corresponding to the positive samples, and the numerical values of the rest positions are corresponding to the negative samples.

The value of the ith row and jth column in the third matrix may represent the matching degree of the object information and the category path, and may be denoted as P (c _j |t _j )。

And training the object processing model by a training mode of maximizing the value corresponding to the positive sample in the third matrix and minimizing the value corresponding to the negative sample, namely adopting a contrast learning mode.

The matching degree of the object in the positive sample and the category path should be a value 1, and the matching degree of the object in the negative sample and the category path should be a value 0, so that the value corresponding to the positive sample in the third matrix is infinitely close to the value 1, and the value corresponding to the negative sample is infinitely close to 0, so as to train the object processing model. The positive sample corresponds to the i-th row and i-th column, and the negative sample corresponds to the i-th row and non-i-th column.

In practical applications, training of the object processing model may be implemented in combination with the loss function, so in some embodiments, training the object processing model in a training manner that maximizes a value corresponding to a positive sample in the third matrix and minimizes a value corresponding to a negative sample may include:

according to the value of the ith row and the jth column in the third matrix, calculating a first loss value of the current training batch by utilizing a contrast learning loss function; and updating the object processing model according to the first loss value.

Alternatively, the contrast learning loss function may be implemented, for example, as follows, although the present application is not limited thereto:

Where τ represents the temperature coefficient, which is a scalar quantity.

The first loss value represents a difference value between a model prediction result and a real result, a gradient of a model parameter can be calculated according to the first loss value, and the object processing model is updated based on the gradient by adopting a counter-propagation mode. In practical application, n can be 512, the temperature coefficient can be set to 1, and the learning rate of the model can be 3e-5. The learning rate is one of the important super-parameters of the training model, and represents the step size of the gradient moving to the loss function in each iteration. In addition, the object information and the category path may be represented in text form, the object information and the category path may be processed by word segmentation, and then a text sequence is formed by the word segmentation result and then input into a model, in order to ensure computing performance, the sequence length may be not greater than a predetermined length, for example, the predetermined length may be 128, and after the object information or the category path is processed by word segmentation, if the number of obtained word segmentation elements is greater than 128, only the first 128 word segmentation elements may be taken to form a corresponding text sequence.

The object processing model may be implemented by using BERT (Bidirectional Encoder Representations from Transformers, a bi-directional semantic coding characterization model constructed based on a transformer), and the object processing model may include a first processing model and a second processing model, where the first processing model and the second processing model may be, for example, a BERT-Small (Small configuration BERT) model, a Projector (mapper) component, and an L2 Norm (L2 normalization, to change a set of numbers into between 0 and 1) component, the BERT-Small model is a pre-training model, may be used to extract feature vectors, the Projector component refers to a linear mapping layer, may be used as a fully connected network, to increase the number of feature parameters, to increase the model expression capability, and the L2 Norm component is used to normalize the feature vectors, to increase the stability of the model. Wherein, the BERT-Small model may adopt a 4-layer transducer model, the attention header number is 8, the hidden layer dimension is 512, and the projector component may be a linear mapping layer with input and output dimensions of 512.

For ease of understanding, in the training diagram of the object processing model shown in fig. 2a, multiple object information in the current training batch may be simultaneously input into the first processing model 201, and multiple category paths may be simultaneously input into the second processing model 202. The first process model 201 and the second process model 202 have the same model structure and share parameters.

In the computational schematic diagram of the object processing model shown in fig. 2b, the first processing model 201 is processed to obtain a plurality of first sample features T ₁ ，T ₂ ，……，T ₂ Forming a first matrix ET, and processing the second processing model 202 to obtain a plurality of second sample features C ₁ ，C ₂ ，……，C ₂ A second matrix EC is formed. The first matrix ET and the second matrix EC obtain a third matrix through a matrix multiplication process.

And then, combining the numerical values in the third matrix, calculating a first loss value by using a loss function, namely updating the parameters of the first processing model or the second processing model, and achieving the purpose of training the object processing model.

As an alternative, the first processing model and the second processing model may be the same model, the model structures are the same and parameters are the same, the object information and the category path may share the same processing model, the structures are consistent, and the parameters are shared.

In addition, as another alternative, the first processing model and the second processing model may be two different models, and the models have the same structure, but different parameters, and the parameters of the first processing model and the second processing model may be respectively adjusted when the object processing model is trained.

In some embodiments, to further ensure model accuracy, the method may further include:

and under the condition that the target category path is the correct category path, taking the target object and the target category path as training sample pairs to be added into a training data set so as to retrain the object processing model.

The target category path is a correct category path, and the training sample pair is a correct category sample pair, so that the reliability of the training sample pair can be improved, and the model is more beneficial to learning the association between the object information and the category path.

In some embodiments, to further improve the accuracy of category detection, the method may further include:

constructing a category correct sample pair and constructing a category misplaced sample pair; wherein, the correct category sample pair comprises a sample object and a correct category path of the sample object; misplacing a sample pair including a sample object and an incorrect category path for the sample object;

Training a category identification model by using the category correct sample pair and the category misplaced sample pair respectively;

the category identification model is used for determining misplacement probability of the target object and the target category path based on the target object and the target category path; the misplacement probability is used for determining whether the target object and the target category path are misplaced or not according to the matching degree.

The category identification model is a supervised model, and the training label corresponding to the category correct sample pair can be the misplacement probability of the sample object and the correct category path; the training label of the category misplaced sample pair can be the misplacement probability of the sample object and the error category path, and the misplacement probability of the sample object and the correct category path can be represented by 0; the misplacement probability of the sample object and the error category path can be represented by 1, and the misplacement probability calculated by the category identification model can be a value between 0 and 1.

Wherein, respectively utilize the correct sample pair of category and the misplaced sample pair of category, training category recognition model can be: taking a sample object of a correct category sample pair and a correct category path of the sample object as model input data, taking misplacement probability of the sample object and the correct category path as a training label, and training a category identification model; and taking the sample object of the class misplaced sample pair and the error class path of the sample object as model input data, taking the misplaced probability of the sample object and the error class path as a training label, and training a class identification model.

In some embodiments, the category identification model may include a first identification network and a second identification network, where training the category identification model may include:

selecting a plurality of training sample pairs including category correct sample pairs and/or category misplaced sample pairs; inputting a plurality of object information in a plurality of training sample pairs into a first recognition network to obtain a plurality of first sample characteristics, and inputting a plurality of category paths into a second recognition network to obtain a plurality of second sample characteristics; splicing the first sample features and the second sample features to obtain fusion features; respectively performing dimension reduction treatment on the fusion characteristics to obtain output data; calculating a second loss value by using the output data and the misplacement probability of the object and the correct category path; updating model parameters of the category identification model according to the second loss value;

training of the category identification model may also be performed in batches, each training batch may be selected from a plurality of training sample pairs, which may include category correct sample pairs and/or category misplaced sample pairs.

The merging feature may be a vector combination or a vector product, and the like, where the merging feature is obtained by stitching a plurality of first sample features and a plurality of second sample features, and assuming that each first sample feature is a d-dimensional vector and each second sample feature is a d-dimensional vector, the plurality of first sample features form an m-x-d-dimensional vector, the plurality of second sample features form an m-x-d-dimensional vector, m is a pair number of samples of a plurality of training sample pairs, and the plurality of first sample features and the plurality of second sample features may be combined to obtain an m-x 2 d-dimensional vector, for example.

The dimension reduction processing is performed on the fusion features, for example, output data in m×1 dimensions can be obtained. Based on the output data and the misplacement probabilities of the different objects and the respective corresponding paths, a second loss value may be calculated, such that model parameters of the category identification model may be updated according to the second loss value.

The second loss value may be calculated, for example, using a cross entropy loss function, although the application is not limited thereto.

Wherein, the misplacement probability of the object and the correct category path may be 0, and the misplacement probability of the object and the wrong category path may be 1.

The first identification network and the second identification network may be implemented by adopting the same model structure, and may share or not share model parameters, that is, the first identification network and the second identification network may have the same model parameters and the same model structure, and when performing model training, the model parameters of the first identification network and the second identification network may be uniformly adjusted, or may also have the same model result and different model parameters, and when performing model training, the model parameters of the first identification network and the second identification network may be respectively adjusted.

Optionally, the category identification network may further include a connection layer configured to splice the plurality of first sample features and the plurality of second sample features to obtain a fusion feature; and the system also comprises an output layer for respectively carrying out dimension reduction processing on the fusion characteristics to obtain output data.

The first identification network and the second identification network may have the same model structure as the first process model and the second process model, and may be composed of, for example, a BERT-Small model, a Projector component, and an L2 Norm component.

For ease of understanding, as shown in fig. 3, which shows a training diagram of a category recognition model, object information may be input to a first recognition network 301, a category path may be input to a second recognition network 302, and the model results of the first recognition network 301 and the second recognition network 302 are the same, and parameters are different.

The first sample features obtained by the first recognition network 301 and the second sample features obtained by the second recognition network 302 may be spliced by the connection layer 303 to obtain a fusion feature, the fusion feature may be subjected to dimension reduction by the output layer 304 to obtain output data, a second loss value may be calculated by using a loss function based on the output data, and model parameters of the first recognition network, the second recognition network, the connection layer, the output and the like may be updated based on the first loss value, so as to achieve the purpose of training a category recognition model.

And calculating the gradient of the model parameter according to the second loss value, and updating the category identification model based on the gradient by adopting a counter-propagation mode. In practical application, m can be 300, and the learning rate of the model can be 3e-5. In addition, the object information and the category path may be represented in text form, the object information and the category path may be processed by word segmentation, and then a text sequence is formed by the word segmentation result and then input into a model, in order to ensure computing performance, the sequence length may be not greater than a predetermined length, for example, the predetermined length may be 128, and after the object information or the category path is processed by word segmentation, if the number of obtained word segmentation elements is greater than 128, only the first 128 word segmentation elements may be taken to form a corresponding text sequence.

In some embodiments, the above-described construction category correct sample pairs may include one or more of the following implementations:

constructing a correct category sample pair according to the manually maintained object and the corresponding category path in the online system;

according to the object which is manually detected to be the correct category and the corresponding category path, constructing a correct category sample pair;

the method comprises the steps of,

for example, in the case of an online system that is an online trading platform, in some promotional option activities, the operation and maintenance personnel will specifically maintain some items, and the path of the category corresponding to the maintained items is correct, so that the manually maintained object and the path of the category corresponding thereto can be used as a correct category sample pair. For example, in the search scenario, the search result may be manually evaluated to determine whether the object and the category path match, and thus, the object and the corresponding category path that are manually detected as being of the correct category may be used as the correct category pair.

In addition, the object processing model obtained through the training can be utilized to determine the matching degree of the object information of any object issued by the on-line system and the corresponding category path, so that the object information and the category intersection, the matching degree of which meets the matching condition, can be selected as the correct category sample pair.

The matching condition may be, for example, that the degree of matching is greater than a specified threshold value, or the like.

In some embodiments, the above-described build category misplaced sample pairs may include one or more of the following implementations:

according to the object information of the object which is manually detected to be the category error and the corresponding category path, constructing a category misplaced sample pair;

the method comprises the steps of,

That is, in one implementation, the object information of the object manually detected as the category error and the corresponding category path thereof may be used as the category misplaced sample pair.

In addition, in order to further improve the accuracy of the model, as a further implementation manner, a class misplaced sample pair can be constructed according to a class correct sample pair, and since one object only corresponds to one class correct sample pair, by searching a similar class path corresponding to the class path in the class correct sample pair, the similar class and the object information in the class correct sample pair can be used as the class misplaced sample pair.

Of course, as another alternative, the object information of the object in the category correct sample pair may be formed into the category misplaced sample pair with any category path other than the correct category path.

In some embodiments, constructing the category misplaced sample pair based on the objects in the category correct sample pair and similar category paths corresponding to the correct category paths in the category correct sample pair may include:

determining a correct category path in the category correct sample pair;

determining a similar category path satisfying a similar condition to the correct category path;

the obtained category misplaced sample pair is formed by the object information of the object in the category correct sample pair and the similar category path.

The similar condition may be, for example, that the first class is the same and the word overlap is greater than the overlap threshold, and thus, a class path that is the same as the first class of the correct class path and that the word overlap is greater than the overlap threshold may be referred to as a similar class path.

Wherein the word overlap can be calculated from the ratio of the number of identical words to the total number of words. Different category paths can be represented by text sequences of specified lengths, and the ratio of the same number of words to the total number of words corresponding to the specified lengths can be used as the word overlap degree by counting the same number of words.

In some embodiments, the method may further comprise:

and taking the sample object and the category path in the category correct sample pair as a training sample pair, and adding a training data set to retrain the object processing model.

Fig. 4 is a flowchart of an embodiment of a method for detecting a category provided in an embodiment of the present application, where the technical solution of the present application is introduced from a model application perspective, and the technical solution of the present embodiment may be executed by a processing end, or executed by a server end of an online system, etc., and the method may include the following steps:

401: object information of a target object and a target category path corresponding to the target object are obtained.

The target object may be any object issued by the online system, or may be any object requested to be issued by the object provider.

402: and extracting target object characteristics from the object information of the target object by using the object processing model.

403: and extracting the target category characteristics from the target category path by using the object processing model.

The object processing model takes a plurality of training sample pairs as positive samples, object information in one positive sample and category paths in the other positive sample as negative samples, and is obtained through training; a plurality of training sample pairs are selected from the training data set to obtain; the training data set is composed of object information of objects issued by the online system and category paths of the objects.

The specific generation manner of the object processing model may be described in detail in the embodiment shown in fig. 1, and will not be described herein.

The object handling model may extract target object features and target category features, respectively.

404: and calculating the matching degree of the target object characteristics and the target category characteristics.

The matching degree can be obtained by multiplying the target object characteristics and the target category characteristics, and in the case that the characteristics are represented by vectors, namely, the characteristics are obtained by calculating a vector dot product, the matching degree is consistent with the calculation mode when the object processing model is subjected to model training.

405: and detecting whether the target object and the target category path are misplaced according to the matching degree.

For example, if the matching degree is smaller than a certain threshold, the target object and the target category path may be considered to be misplaced, and if the matching degree is larger than the certain threshold, the target object and the target category path may be considered to be correct.

In addition, to further improve detection accuracy, in some embodiments, the method may further include:

and inputting the object information and the target category path into a category identification model, and calculating by using the category identification model to obtain the misplacement probability.

The category identification model is obtained by training a category misplaced sample pair by using the constructed and generated category correct sample pair; the specific generation manner of the category identification model may be described in detail in the foregoing corresponding embodiments, and will not be described herein.

Determining whether the paths of the target object and the category are wrong according to the matching degree comprises the following steps:

I.e., whether the target object and category paths are misplaced or not can be determined by combining the matching degree and the misplacement probability.

As an alternative, it may be that the target category path is determined to be an error category path of the target object in the case that the matching degree is smaller than the first matching threshold, or the matching degree is smaller than the second matching threshold and the misplacement probability is larger than the first probability threshold; the second matching threshold is greater than the first matching threshold;

and under the condition that the matching degree is larger than the second matching threshold or the misplacement probability is smaller than the first probability threshold, determining the target category path as the correct category path of the target object.

For example, in practical applications, the first matching threshold may be 0.2, the second matching threshold may be 0.6, the first probability threshold may be 0.85, etc.

The technical solution of the embodiment of the present application may be applied to a search scenario, and the search action may be executed based on a search keyword provided by a user through a second user side, so as to perform object recall, and in some embodiments, the method may further include:

determining a plurality of objects under any category path based on search keyword hits; taking a plurality of objects as target objects respectively;

as an alternative, after detecting whether the target object and the category path are misplaced according to the matching degree, the method may further include:

and selecting one or more objects of the correct category path as recall objects of the search keywords according to detection results of the plurality of objects.

That is, by adopting the technical scheme of the embodiment of the application, a plurality of objects in any category path hit by the search keyword can be screened, and one or more objects in the correct category path can be used as recall objects.

After the recall object is obtained, sorting processing and the like can be performed on the recall object, and finally object prompt information of the recall object is fed back to the second user side as a search result according to the sorting result for the user to check.

As another alternative, after detecting whether the target object and the category path are misplaced according to the matching degree, the method may further include:

According to the detection results of the objects, one or more objects of the correct category path are selected, and from the one or more objects of the wrong category path, an object with the exposure amount larger than a second exposure threshold or with the sales amount larger than a first sales threshold is selected as a recall object of the search keyword.

That is, in addition to recalling one or more objects of the correct category path, one or more objects of the incorrect category path may be filtered, and high exposure and/or high sales objects may be recalled.

Furthermore, in some embodiments, the method may further comprise:

And determining whether the corresponding category path is a correct category path or an incorrect category path through manual detection of the high-exposure and/or high-sales objects.

According to the manual detection result, a category correct sample pair can be generated for the object of the correct category path, and a category misplaced sample pair can be generated for the object of the wrong category path, so that the category recognition model can be retrained, and fine adjustment and update can be realized for the category recognition model.

Furthermore, in some embodiments, the method may further comprise:

setting an error mark for the target object according to the detection result of the target object under the condition that the target category path is an error category path;

in the case where a path of any category is hit based on the search keyword, an object for which an error flag is set under the path of any category is not taken as a recall object of the search keyword.

That is, the object determined to be misplaced in the category path may be added to the blacklist and may no longer be hit by the search keyword.

In the scene interaction schematic diagram shown in fig. 5a, a user may send a search request to a server side 502 through a second user side 501, where the search request carries a search keyword, the server side 502 may determine any category path hit by the search keyword, recall only an object under the any category path, where an error flag is not set, and generate a search result based on the recall result and feed back the search result to the second user side 501; of course, the server 502 may perform real-time detection processing on a plurality of objects in any category path to determine the object in the correct category path and the object in the wrong category path, and determine the recalled object according to the above.

In addition, the technical solution of the embodiment of the present application may also be applied to an object release scenario, where the method is used to prompt whether the object provider has a wrong path selection, so in some embodiments, the obtaining the object information of the target object and the target category path corresponding to the target object may include:

determining a target category path selected by the object provider and object information of the target object according to a category selection request of the object provider for the target object;

the method may further comprise: and according to the detection result of the target object, sending misplacement prompt information to the object provider under the condition that the target category path is an error category path.

Alternatively, the misplacement prompt information may be sent to the first user side, so that the first user side may generate the misplacement prompt information.

The category selection request is generated and sent by the first user end in response to the object provider category selection operation.

In the scenario interaction schematic diagram shown in fig. 5b, the object provider may send, to the server 502 through the first user 503, a category selection request for a target object to be published, the server 502 may determine a target category path selected by the object provider and object information of the target object according to the category selection request, may perform category detection on the target category path, and in case that it is determined that the target category path is an error category path, feedback a misplacement prompt message to the first user 503, so that the first user 503 outputs misplacement prompt information, where the misplacement prompt message may prompt the object provider whether to modify the selected target category path, and so on. And in the case that the target category path is the correct category path, the target object can be issued according to the correct category path.

In an actual application of the embodiment of the present application, the online system may refer to an online transaction platform, the object may refer to a commodity provided by the online transaction platform, and the technical scheme of the present application is described below by taking the commodity provided by the online transaction platform as an example with reference to a schematic view of a category detection flow shown in fig. 6.

Firstly, commodity information and corresponding category paths of the issued commodity are used as training data sets according to the commodity issued by an online transaction platform, and each commodity information and the corresponding category path form a training sample pair; then, a plurality of training sample pairs of the current training batch can be selected from the training data set, then, the plurality of training sample pairs can be input into the object processing model, the training sample pairs are used as positive samples, commodity information in one positive sample and category paths in the other positive sample are used as negative samples, and a comparison learning mode can be adopted to train the object processing model. Therefore, semi-supervision data of the class can be effectively utilized, the model has good generalization capability, and the requirement on the manual annotation data volume is reduced.

In addition, a correct category sample pair and a misplaced category sample pair can be constructed, and the construction modes of the correct category sample pair and the misplaced category sample pair can be detailed in the above, wherein the method can comprise the step of constructing the misplaced category sample pair of the Hard (error) type by a category path similarity algorithm according to the objects in the correct category sample pair and category misplaced category sample pairs formed by similar category paths corresponding to category paths in the correct category sample pair.

As shown in fig. 6, information about the commodity requested to be issued by the merchant may be stored in the commodity information base 60, and any one commodity in the commodity information base 60 may be a target commodity to be detected. Based on the commodity title and the target category path of the target commodity, the object processing model 61 and the category identification model 62 obtained through the training described above may be respectively processed, so as to obtain the corresponding matching degree simfire and misplacement probability mismatscore. In combination with the matching degree and the misplacement probability, a misplacement detection operation 601 may be performed, where the misplacement detection condition may be, for example:

((MiscatScore>0.8)&(SimScore＜0.6))||(SimScore<0.2)。

and combining the detection results of the misplaced detection conditions, and determining the commodity of the path of the correct category and the commodity of the path of the wrong category. The commodity of the correct category path may be recalled, and the commodity judging module 602 may be further executed to judge the exposure pv and/or the sales order for the commodity of the incorrect category path, where the commodity judging conditions may be, for example:

(pv<500)&(order<2)。

that is, for low exposure, such as merchandise with pv less than 500 and low sales, such as merchandise with order less than 2, the item may be recalled without being recalled during the object recall phase 604 of the search scene, while for high exposure, such as merchandise with pv greater than 500 or high sales, such as merchandise with order greater than 2, the item may be recalled and sent to the relevant personnel for human detection 603, and a correct category sample pair or a misplaced category sample pair may be constructed based on the human detection results to continue training the category identification model. Of course, correct pairs of samples for categories may also be used to continue training the object handling model.

In the embodiment of the application, based on a comparison learning mode, an object processing model can be constructed, and the model is pre-trained by using semi-supervision data of category issued by merchants, so that the model learns associated information and knowledge implied by commodity titles and category paths of commodities. In addition, a category identification model can be constructed, and the model is trained by manufacturing a category misplaced sample pair and a category correct sample pair, so that the identification of misplaced commodities by the model is realized; the technical scheme of the embodiment of the application can be applied to commodity searching scenes, so that searching correlation can be improved, misplaced sample pairs and correct category sample pairs of categories can be deposited, and the purpose of continuously fine-tuning and optimizing the model is achieved.

Fig. 7 is a schematic structural diagram of an embodiment of a model training apparatus according to an embodiment of the present application, where the apparatus may include:

a data acquisition module 701, configured to use object information of a published object and a category path of the object as a training data set;

a data selection module 702, configured to select a plurality of training sample pairs of a current training batch from the training data set; the training sample pair comprises object information and corresponding category paths;

A first training module 703, configured to input a plurality of training sample pairs into the object processing model, and train the object processing model by taking the training sample pairs as positive samples, and taking object information in one positive sample and category paths in another positive sample as negative samples;

In some embodiments, the first training module is specifically configured to input a plurality of object information in a plurality of training sample pairs into the object processing model to obtain a plurality of first sample features; inputting a plurality of category paths in a plurality of training sample pairs into an object processing model to obtain a plurality of second sample features; performing matrix multiplication processing on a first matrix formed by a plurality of first sample features and a second matrix formed by a plurality of second sample features to obtain a third matrix; wherein the value of the ith row and the jth column in the third matrix represents the similarity between the ith object and the jth category path; wherein i=1, 2, 3 … … n, j=1, 2, 3 … … n; i is different from j; n is a positive integer representing the number of pairs of training samples; and training the object processing model in a training mode of maximizing the value corresponding to the positive sample in the third matrix and minimizing the value corresponding to the negative sample.

In some embodiments, the apparatus may further comprise:

the sample construction module is used for constructing correct category sample pairs and misplaced category sample pairs; wherein, the correct category sample pair comprises a sample object and a correct category path of the sample object; misplacing a sample pair including a sample object and an incorrect category path for the sample object;

the second training module is used for training a category identification model by using the correct category sample pair and the misplaced category sample pair respectively;

In some embodiments, the sample construction module forms the pair of classically correct samples in particular according to one or more of the following implementations:

constructing a correct category sample pair according to the object information of the manual maintenance object and the corresponding category path;

the method comprises the steps of,

In some embodiments, the sample construction module forms the class misplaced sample pairs in particular according to one or more of the following implementations:

the method comprises the steps of,

In some embodiments, the category identification model includes a first identification network and a second identification network;

the second training module is specifically: selecting a plurality of training sample pairs including category correct sample pairs and/or category misplaced sample pairs; inputting a plurality of object information in a plurality of training sample pairs into a first recognition network to obtain a plurality of first sample characteristics, and inputting a plurality of category paths into a second recognition model to obtain a plurality of second sample characteristics; splicing the first sample features and the second sample features to obtain fusion features; respectively performing dimension reduction treatment on the fusion characteristics to obtain output data; calculating a second loss value by using the output data and the misplacement probability of different objects and paths of respective corresponding categories; and updating the model parameters of the category identification model according to the second loss value.

In some embodiments, the data acquisition module specifically determines the published object and the category path in which the object is located; and adding object information of the object and a path of the category corresponding to the object into the training data set under the condition that the object meets the screening condition for any object.

In some embodiments, the screening conditions include one or more of the following implementations:

the exposure is greater than a first exposure threshold;

the method comprises the steps of,

In some embodiments, the training the object processing model by the first training module in a training manner that maximizes the value corresponding to the positive sample in the third matrix and minimizes the value corresponding to the negative sample includes: according to the value of the ith row and the jth column in the third matrix, calculating a first loss value of the current training batch by utilizing a contrast learning loss function; and updating the object processing model according to the first loss value.

In some embodiments, the data acquisition module is further configured to add the sample object in the category correct sample pair and the category path as a sample pair to the training data set to retrain the object handling model.

In some embodiments, the data acquisition module is specifically configured to select a predetermined number of objects from under any category path of the online system; a predetermined number of objects and category paths are added to the training dataset.

In some embodiments, the data acquisition module is further configured to add the target object and the target category path as a training sample pair to the training data set to retrain the object handling model if the target category path is a correct category path.

In some embodiments, the sample construction module constructs a category misplaced sample pair according to the object in the category correct sample pair and a similar category path corresponding to the correct category path in the category correct sample pair, including:

determining a correct category path in the category correct sample pair; determining a category path which is the same as the first category of the correct category path and has the word overlapping degree larger than the overlapping threshold value as a similar category path; the obtained category misplaced sample pair is formed by object information of objects in the category correct sample pair and similar category paths.

The model training apparatus shown in fig. 7 may perform the model training method described in the embodiment shown in fig. 1, and its implementation principle and technical effects will not be described again. The specific manner in which the respective modules and units of the model training apparatus in the above embodiment perform operations has been described in detail in the embodiments related to the method, and will not be described in detail here.

Fig. 8 is a schematic structural diagram of an embodiment of a generic visual inspection device provided in an embodiment of the present application, where the device may include:

an information obtaining module 801, configured to obtain object information of a target object and a target category path corresponding to the target object;

a first extraction module 802, configured to extract a target object feature from object information of a target object using an object processing model; the object processing model takes a plurality of training sample pairs as positive samples, object information in one positive sample and category paths in the other positive sample are taken as negative samples, and training is carried out to obtain the object processing model; a plurality of training sample pairs are selected from the training data set to obtain; the training data set is composed of object information of the object issued by the online system and category paths of the object;

a second extraction module 803 for extracting target category features from the target category path using the object processing model;

a first calculating module 804, configured to calculate a matching degree between the target object feature and the target category feature;

and a detection module 805, configured to detect whether the target object and the category path are misplaced according to the matching degree.

In some embodiments, the apparatus may further comprise:

the second calculation module is used for inputting the object information and the target category path into a category identification model, and calculating by using the category identification model to obtain misplacement probability; the category identification model is obtained by training the category misplaced sample pairs by using the constructed and generated category correct sample pairs;

The detection module is specifically configured to determine whether the target object and the category path are misplaced according to the matching degree and the misplacement probability.

In some embodiments, the detection module is specifically configured to determine that the target category path is an error category path of the target object when the matching degree is less than the first matching threshold, or the matching degree is less than the second matching threshold and the misplacement probability is greater than the first probability threshold; the second matching threshold is greater than the first matching threshold;

In some embodiments, the apparatus may further comprise:

an object determining module for determining a plurality of objects under an arbitrary category path based on search keyword hits; taking a plurality of objects as target objects respectively;

the object recall module is used for selecting one or more objects of the correct category paths as recall objects of the search keywords according to detection results of the objects;

or,

selecting one or more objects of the correct category path according to the detection results of the plurality of objects, and selecting an object with the exposure larger than a second exposure threshold and/or the sales larger than a first sales threshold from the one or more objects of the incorrect category path as a recall object of the search keyword.

In some embodiments, the apparatus may further comprise:

the marking module is used for setting an error mark for the target object according to the detection result of the target object under the condition that the target category path is an error category path; in the case where a path of any category is hit based on the search keyword, an object for which an error flag is set under the path of any category is not taken as a recall object of the search keyword.

In some embodiments, the apparatus may further comprise:

the providing module is used for providing the object with the exposure larger than the second exposure threshold or the sales larger than the first sales threshold to related personnel so as to enable the related personnel to manually detect whether the object and the corresponding category path are misplaced or not.

In some embodiments, the information obtaining module may specifically determine, according to a category selection request of the object provider for the target object, a target category path selected by the object provider and object information of the target object;

the apparatus may further include:

and the prompt module is used for sending misplacement prompt information to the object provider according to the detection result of the target object under the condition that the target category path is an error category path.

The category detection device shown in fig. 8 may perform the category detection method shown in the embodiment shown in fig. 4, and its implementation principle and technical effects are not repeated. The specific manner in which the respective modules and units of the above-described embodiment of the category detection device perform operations has been described in detail in the embodiments related to the method, and will not be described in detail here.

Embodiments of the present application also provide a computing device, as shown in fig. 9, which may include a storage component 901 and a processing component 902;

the storage component 901 stores one or more computer instructions for execution by the processing component 902 to implement the model training method of the embodiment shown in fig. 1 or the category detection method shown in fig. 4.

Of course, the computing device may necessarily include other components as well, such as input/output interfaces, display components, communication components, and the like.

The input/output interface provides an interface between the processing component and a peripheral interface module, which may be an output device, an input device, etc. The communication component is configured to facilitate wired or wireless communication between the computing device and other devices, and the like.

Wherein the processing component may include one or more processors to execute computer instructions to perform all or part of the steps of the methods described above. Of course, the processing component may also be implemented as one or more Application Specific Integrated Circuits (ASICs), digital Signal Processors (DSPs), digital Signal Processing Devices (DSPDs), programmable Logic Devices (PLDs), field Programmable Gate Arrays (FPGAs), controllers, microcontrollers, microprocessors or other electronic elements for executing the methods described above.

The storage component is configured to store various types of data to support operations at the terminal. The memory component may be implemented by any type or combination of volatile or nonvolatile memory devices such as Static Random Access Memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic or optical disk.

The display component may be an Electroluminescent (EL) element, a liquid crystal display or a micro display having a similar structure, or a retina-directly displayable or similar laser scanning type display.

It should be noted that, the above computing device may be an elastic computing host provided for a physical device or a cloud computing platform, or the like. It may be implemented as a distributed cluster of multiple servers or terminal devices, or as a single server or single terminal device.

The embodiment of the application further provides a computer readable storage medium, and a computer program is stored, and when the computer program is executed by a computer, the model training method of the embodiment shown in fig. 1 or the category detection method of the embodiment shown in fig. 4 can be implemented. The computer-readable medium may be contained in the electronic device described in the above embodiment; or may exist alone without being incorporated into the electronic device.

The embodiment of the application further provides a computer program product, which comprises a computer program loaded on a computer readable storage medium, and the computer program can implement the model training method of the embodiment shown in fig. 1 or the category detection method of the embodiment shown in fig. 2 when the computer program is executed by a computer. In such embodiments, the computer program may be downloaded and installed from a network, and/or installed from a removable medium. The computer program, when executed by a processor, performs the various functions defined in the system of the present application.

It will be clear to those skilled in the art that, for convenience and brevity of description, specific working procedures of the above-described systems, apparatuses and units may refer to corresponding procedures in the foregoing method embodiments, which are not repeated herein.

The apparatus embodiments described above are merely illustrative, wherein the elements illustrated as separate elements may or may not be physically separate, and the elements shown as elements may or may not be physical elements, may be located in one place, or may be distributed over a plurality of network elements. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art will understand and implement the present invention without undue burden.

From the above description of the embodiments, it will be apparent to those skilled in the art that the embodiments may be implemented by means of software plus necessary general hardware platforms, or of course may be implemented by means of hardware. Based on this understanding, the foregoing technical solution may be embodied essentially or in a part contributing to the prior art in the form of a software product, which may be stored in a computer readable storage medium, such as ROM/RAM, a magnetic disk, an optical disk, etc., including several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute the method described in the respective embodiments or some parts of the embodiments.

Finally, it should be noted that: the above embodiments are only for illustrating the technical solution of the present application, and are not limiting thereof; although the present application has been described in detail with reference to the foregoing embodiments, it should be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit and scope of the corresponding technical solutions.

Claims

1. A method of model training, comprising:

2. The method of claim 1, wherein inputting the plurality of training sample pairs into an object processing model with the training sample pairs as positive samples, object information in one positive sample being negative with category paths in another positive sample, training the object processing model comprising:

3. The method as recited in claim 1, further comprising:

4. A method according to claim 3, wherein said constructing pairs of class-correct samples comprises one or more of the following implementations:

The method comprises the steps of,

the method comprises the steps of,

5. The method of claim 3, wherein the category identification model includes a first identification network and a second identification network;

the training category identification model comprises the following steps of:

6. The method of claim 1, wherein the taking object information of the published object and the object corresponding category path as a training data set comprises:

determining a published object and a category path of the object;

7. The method as recited in claim 4, further comprising:

8. A method for detecting a species of eye, comprising:

9. The method as recited in claim 8, further comprising:

10. The method of claim 9, wherein the determining whether the target object is misplaced with the category path based on the degree of matching and the probability of misplacement comprises:

11. The method as recited in claim 10, further comprising:

respectively taking the plurality of objects as the target objects;

Selecting one or more objects of the correct category paths as recall objects of the search keywords according to the detection results of the objects; or,

12. The method as recited in claim 8, further comprising:

13. A computing device comprising a processing component and a storage component;

the storage component stores one or more computer instructions; the one or more computer instructions are for execution by the processing component to implement the model training method of any one of claims 1-7 or the category detection method of any one of claims 8-12.

14. A computer-readable storage medium, characterized in that a computer program is stored, which, when being executed by a computer, implements the model training method according to any one of claims 1 to 7 or implements the category detection method according to any one of claims 8 to 12.