CN113935401A

CN113935401A - Article information processing method, article information processing device, article information processing server and storage medium

Info

Publication number: CN113935401A
Application number: CN202111101577.8A
Authority: CN
Inventors: 刘柳; 胡懋地; 刘家骅; 张晓星
Original assignee: Beijing Sankuai Online Technology Co Ltd
Current assignee: Beijing Sankuai Online Technology Co Ltd
Priority date: 2021-09-18
Filing date: 2021-09-18
Publication date: 2022-01-14

Abstract

The application provides a method and a device for processing article information, a server and a storage medium, and belongs to the technical field of internet. The method comprises the following steps: acquiring attribute information and an article image of a target article; determining a first semantic feature corresponding to the attribute information and a second semantic feature corresponding to the article image; fusing the first semantic features and the second semantic features to obtain target semantic features of the target object; and determining standard article information of the target article based on the target semantic features. The first semantic features are semantic features corresponding to the attribute information of the article, and the second semantic features are semantic features corresponding to the image of the article, so that a target semantic feature obtained by fusing the first semantic features and the second semantic features is fused with a plurality of semantic features in different modes, the standard article information is determined through the target semantic features, and the accuracy of the determined standard article information can be improved.

Description

Article information processing method, article information processing device, article information processing server and storage medium

Technical Field

The present application relates to the field of internet technologies, and in particular, to a method and an apparatus for processing article information, a server, and a storage medium.

Background

With the development of internet technology, more and more users take the dishes out through an internet platform, and the users can input the names of the dishes, so that the server recommends a plurality of dishes matched with the names of the dishes based on the names of the dishes and the names of the dishes of a plurality of merchants. In order to improve the accuracy of the search, the server needs to determine a standard dish name corresponding to the dish name of the merchant, so as to recommend the dish based on the standard dish name, thereby improving the accuracy.

In the related art, the server determines semantic similarity between a dish name and a plurality of standard dish names, and determines a standard dish name corresponding to the dish name based on the semantic similarity.

However, due to the high personalization degree and incomplete information of the dish names of some merchants, the accuracy of determining the standard dish name corresponding to the dish name is low based on the semantic similarity between the dish name and the standard dish name.

Disclosure of Invention

The embodiment of the application provides a method and a device for processing article information, a server and a storage medium, which can improve the accuracy of information processing. The technical scheme is as follows:

according to an aspect of the embodiments of the present application, there is provided a method for processing article information, the method including:

acquiring attribute information and an article image of a target article;

determining a first semantic feature corresponding to the attribute information and a second semantic feature corresponding to the article image;

fusing the first semantic features and the second semantic features to obtain target semantic features of the target object;

and determining standard article information of the target article based on the target semantic features.

In one possible implementation, the attribute information includes at least two of an item name, a merchant name, and tag information; the process of determining the first semantic feature corresponding to the attribute information includes:

determining at least two semantic vectors of a first semantic vector, a second semantic vector and a third semantic vector, wherein the first semantic vector is a semantic vector corresponding to the item name, the second semantic vector is a semantic vector corresponding to the merchant name, and the third semantic vector is a semantic vector corresponding to the tag information;

determining the first semantic feature based on the at least two semantic vectors.

In another possible implementation manner, the determining the first semantic feature based on the at least two semantic vectors includes:

splicing the at least two semantic vectors to obtain a target semantic vector, or adding category information to each semantic vector, and splicing the at least two semantic vectors added with the category information to obtain the target semantic vector;

and determining the first semantic features corresponding to the target semantic vector.

In another possible implementation manner, the determining the first semantic feature corresponding to the target semantic vector includes:

and inputting the target semantic vector into a semantic vector model, and outputting the first semantic features, wherein the semantic vector model is used for converting the semantic vector into the semantic features.

In another possible implementation manner, the process of determining the second semantic feature corresponding to the item image includes:

determining at least one region of interest of the item image;

for each region of interest, determining a second semantic vector corresponding to the region of interest to obtain at least one second semantic vector;

determining the second semantic feature based on the at least one second semantic vector.

In another possible implementation manner, the fusing the first semantic feature and the second semantic feature to obtain the target semantic feature of the target item includes:

converting the first semantic feature and the second semantic feature into semantic features with the same dimension to obtain a first standard semantic feature and a second standard semantic feature;

determining a first attention feature corresponding to the first standard semantic feature and a second attention feature corresponding to the second standard semantic feature;

and fusing the first attention feature and the second attention feature to obtain the target semantic feature.

In another possible implementation manner, the determining a first attention feature corresponding to the first standard semantic feature and a second attention feature corresponding to the second standard semantic feature includes:

determining a first query vector, a first key vector and a first value vector corresponding to the first standard semantic feature and a second query vector, a second key vector and a second value vector corresponding to the second standard semantic feature; determining a first attention weight and a second attention weight based on the first query vector, the first key vector, the second query vector, and the second key vector; and performing weighting processing on the first standard semantic feature based on the first value vector and the first attention weight to obtain the first attention feature, and performing weighting processing on the second standard semantic feature based on the second value vector and the second attention weight to obtain the second attention feature.

In another possible implementation, the determining the first attention weight and the second attention weight includes:

determining the first attention weight based on a first similarity, a second similarity and the first value vector, wherein the first similarity is a similarity between the first query vector and the first key vector, and the second similarity is a similarity between the second query vector and the first key vector;

determining the second attention weight based on a third similarity, which is a similarity between the first query vector and the second key vector, a fourth similarity, which is a similarity between the second query vector and the second key vector, and the second value vector.

In another possible implementation, the first attention weight includes a first weight and a second weight;

said determining said first attention weight based on said first and second similarities and said first value vector, comprising:

normalizing the first similarity to obtain the first weight, and normalizing the second similarity to obtain the second weight;

determining the first attention weight based on a product between the first weight and the first vector of values and a product between the second weight and the first vector of values.

In another possible implementation, the second attention weight includes a third weight and a fourth weight;

the determining the second attention weight based on the third similarity, the fourth similarity, and the first value vector, comprising:

normalizing the third similarity to obtain a third weight, and normalizing the fourth similarity to obtain a fourth weight;

determining the first attention weight based on a product between the third weight and the second value vector and a product between the fourth weight and the second value vector.

In another possible implementation manner, the fusing the first attention feature and the second attention feature to obtain the target semantic feature includes:

fusing the first attention feature and the second attention feature to obtain an attention feature;

and performing pooling processing on the attention feature to obtain the target semantic feature.

In another possible implementation manner, the determining standard item information of the target item based on the target semantic feature includes:

determining similarity between the target item and a plurality of standard item information based on the target semantic features and semantic features of the plurality of standard item information;

determining standard item information of the target item from the plurality of standard item information based on similarities between the target item and the plurality of standard item information.

According to another aspect of the embodiments of the present application, there is provided an article information processing apparatus, including:

the acquisition module is used for acquiring attribute information and an article image of a target article;

the first determining module is used for determining a first semantic feature corresponding to the attribute information and a second semantic feature corresponding to the article image;

the fusion module is used for fusing the first semantic feature and the second semantic feature to obtain a target semantic feature of the target article;

and the second determination module is used for determining standard article information of the target article based on the target semantic features.

In one possible implementation, the attribute information includes at least two of an item name, a merchant name, and tag information; a first determination module comprising:

a first determining unit, configured to determine at least two semantic vectors of a first semantic vector, a second semantic vector, and a third semantic vector, where the first semantic vector is a semantic vector corresponding to the item name, the second semantic vector is a semantic vector corresponding to the merchant name, and the third semantic vector is a semantic vector corresponding to the tag information;

a second determining unit for determining the first semantic feature based on the at least two semantic vectors.

In another possible implementation manner, the second determining unit is configured to splice the at least two semantic vectors to obtain a target semantic vector, or add category information to each semantic vector and splice the at least two semantic vectors to which the category information is added to obtain the target semantic vector; and determining the first semantic features corresponding to the target semantic vector.

In another possible implementation manner, the second determining unit is configured to input the target semantic vector into a semantic vector model, and output the first semantic feature, where the semantic vector model is configured to convert a semantic vector into a semantic feature.

In another possible implementation manner, the first determining module further includes:

a third determination unit for determining at least one region of interest of the item image;

the fourth determining unit is used for determining a second semantic vector corresponding to each interested area to obtain at least one second semantic vector;

a fifth determining unit for determining the second semantic features based on the at least one second semantic vector.

In another possible implementation manner, the fusion module includes:

the conversion unit is used for converting the first semantic feature and the second semantic feature into semantic features with the same dimension to obtain a first standard semantic feature and a second standard semantic feature;

a sixth determining unit, configured to determine a first attention feature corresponding to the first standard semantic feature and a second attention feature corresponding to the second standard semantic feature;

and the fusion unit is used for fusing the first attention feature and the second attention feature to obtain the target semantic feature.

In another possible implementation manner, the sixth determining unit is configured to determine a first query vector, a first key vector, and a first value vector corresponding to the first standard semantic feature, and a second query vector, a second key vector, and a second value vector corresponding to the second standard semantic feature; determining a first attention weight and a second attention weight based on the first query vector, the first key vector, the second query vector, and the second key vector; and performing weighting processing on the first standard semantic feature based on the first value vector and the first attention weight to obtain the first attention feature, and performing weighting processing on the second standard semantic feature based on the second value vector and the second attention weight to obtain the second attention feature.

In another possible implementation manner, the sixth determining unit is configured to determine the first attention weight based on a first similarity, a second similarity and the first value vector, where the first similarity is a similarity between the first query vector and the first key vector, and the second similarity is a similarity between the second query vector and the first key vector; determining the second attention weight based on a third similarity, which is a similarity between the first query vector and the second key vector, a fourth similarity, which is a similarity between the second query vector and the second key vector, and the second value vector.

In another possible implementation, the first attention weight includes a first weight and a second weight; the sixth determining unit is configured to perform normalization processing on the first similarity to obtain the first weight, and perform normalization processing on the second similarity to obtain the second weight; determining the first attention weight based on a product between the first weight and the first vector of values and a product between the second weight and the first vector of values.

In another possible implementation, the second attention weight includes a third weight and a fourth weight; the sixth determining unit is configured to perform normalization processing on the third similarity to obtain the third weight, and perform normalization processing on the fourth similarity to obtain the fourth weight; determining the second attention weight based on a product between the third weight and the second value vector and a product between the fourth weight and the second value vector.

In another possible implementation manner, the fusion unit is configured to fuse the first attention feature and the second attention feature to obtain an attention feature; and performing pooling processing on the attention feature to obtain the target semantic feature.

In another possible implementation manner, the second determining module is configured to determine, based on the target semantic feature and semantic features of a plurality of standard item information, a similarity between the target item and the plurality of standard item information; determining standard item information of the target item from the plurality of standard item information based on similarities between the target item and the plurality of standard item information.

According to another aspect of embodiments of the present application, there is provided a server, including: the system comprises a processor and a memory, wherein at least one program code is stored in the memory, and the at least one program code is loaded by the processor and executed to realize the operation in the article information processing method according to any one of the possible implementation modes.

According to another aspect of embodiments of the present application, there is provided a computer-readable storage medium having stored therein at least one program code, which is loaded by a processor and has an operation performed in a processing method to implement the item information.

According to another aspect of embodiments of the present application, there is provided a computer program product or a computer program comprising computer program code stored in a computer readable storage medium. The processor of the computer device reads the computer program code from the computer-readable storage medium, and the processor executes the computer program code, so that the computer device performs the operations performed by the determination method for processing item information in any one of the above-described possible implementations.

The technical scheme provided by the embodiment of the application has the beneficial effects that at least:

the embodiment of the application provides a method for processing article information, wherein a first semantic feature is a semantic feature corresponding to attribute information of an article, and a second semantic feature is a semantic feature corresponding to an article image, so that a target semantic feature obtained by fusing the first semantic feature and the second semantic feature is fused with a plurality of semantic features in different modalities, standard article information is determined through the target semantic feature, and the accuracy of the determined standard article information can be improved.

Drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the description of the embodiments are briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present application, and it is obvious for those skilled in the art to obtain other drawings based on these drawings without creative efforts.

FIG. 1 is a schematic illustration of an implementation environment provided by an embodiment of the present application;

fig. 2 is a flowchart of a method for processing item information according to an embodiment of the present disclosure;

fig. 3 is a flowchart of a method for processing item information according to an embodiment of the present application;

FIG. 4 is a schematic diagram of a method for determining a first semantic feature and a second semantic feature provided by an embodiment of the present application;

FIG. 5 is a schematic diagram of a method for determining a first semantic feature according to an embodiment of the present disclosure;

FIG. 6 is a schematic diagram of a method for determining a target semantic feature through a self-attention mechanism according to an embodiment of the present disclosure;

fig. 7 is a schematic diagram of a method for processing item information according to an embodiment of the present disclosure;

fig. 8 is a block diagram of an article information processing apparatus according to an embodiment of the present application;

fig. 9 is a block diagram of an article information processing apparatus according to an embodiment of the present application;

fig. 10 is a schematic structural diagram of a server according to an embodiment of the present application.

Detailed Description

To make the objects, technical solutions and advantages of the present application more clear, embodiments of the present application will be described in further detail below with reference to the accompanying drawings.

Fig. 1 is a schematic diagram of an implementation environment provided by an embodiment of the present application. Referring to fig. 1, the implementation environment includes a terminal 101 and a server 102.

The terminal 101 and the server 102 are connected via a wireless or wired network. Moreover, a client that the server 102 provides services may be installed on the terminal 101, and a user corresponding to the terminal 101 may implement functions such as data transmission and message interaction with the server 102 through the client. The client may be a client installed on the terminal 101 and including an internet access function. For example, the client may be a browser, a social application, a gaming application, or a take-away application, among others.

The terminal 101 may be a computer, a mobile phone, a tablet computer or other electronic devices. The server 102 may be a server, a server cluster composed of several servers, or a cloud computing service center.

The article information processing method provided in the embodiment of the present application can be applied to a plurality of scenarios:

for example, an application in a search scenario for user point takeaway: when a merchant uploads the dish information of a dish to a server, the server determines the standard dish name of the dish through the processing method of the item information provided in the embodiment of the application, and associates the dish information of the dish with the standard dish name. And subsequently, when the user inputs the name of the dish, the server acquires the name of the standard dish matched with the name of the dish based on the name of the dish and the stored name of the standard dish, and outputs the information of the dish related to the name of the standard dish.

For another example, the application is in a recommendation scenario for user point takeaway: when a merchant uploads the dish information of a dish to a server, the server determines the standard dish name of the dish through the processing method of the item information provided in the embodiment of the application, and associates the dish information of the dish with the standard dish name. The server determines the name of a dish liked by the user based on the historical browsing record or the historical order information of the user, acquires a standard dish name matched with the name of the dish based on the name of the dish and the stored standard dish name, and recommends the dish information associated with the standard dish name to the user. For example, the server determines that the dish name of the dish liked by the user is 'spicy and hot mixed noodles' based on the historical browsing record of the user, acquires the standard dish name matched with the dish name as 'mixed noodles', and recommends 'mixed noodles' to the user.

For another example, the method is applied to an automatic association scene of the dish name uploaded by the merchant and the description information of the dish: and the server stores the corresponding relation between the standard dish name and the description information. The description information comprises information such as food materials, tastes and methods of dishes. When a merchant uploads the dish information of a dish to a server, the server determines the standard dish name of the dish through the processing method of the item information provided in the embodiment of the application; and the server determines the description information matched with the dish name based on the corresponding relation between the stored standard dish name and the description information, and associates the dish name uploaded by the merchant with the description information of the dish.

For another example, the method is applied to an operation analysis scene of dish supply and sale: when a merchant uploads the dish information of a dish to a server, the server determines the standard dish name of the dish through the processing method of the item information provided in the embodiment of the application, and associates the dish information, sales information and the standard dish name of the dish. And subsequently, when the operation analysis is carried out on the dish, the server determines the dish supply amount and the dish sales amount corresponding to the standard dish name based on the incidence relation among the dish information, the sales amount information and the standard dish name of the dish, and carries out the operation analysis on the dish according to the dish supply amount and the dish sales amount.

Fig. 2 is a flowchart of a method for processing item information according to an embodiment of the present application. Referring to fig. 2, the item information processing method includes the steps of:

201. and acquiring attribute information and an article image of the target article.

202. And determining a first semantic feature corresponding to the attribute information and a second semantic feature corresponding to the article image.

203. And fusing the first semantic features and the second semantic features to obtain the target semantic features of the target object.

204. And determining standard article information of the target article based on the target semantic features.

In one possible implementation, the attribute information includes at least two of an item name, a merchant name, and tag information; the process for determining the first semantic features corresponding to the attribute information comprises the following steps:

determining at least two semantic vectors of a first semantic vector, a second semantic vector and a third semantic vector, wherein the first semantic vector is a semantic vector corresponding to an article name, the second semantic vector is a semantic vector corresponding to a merchant name, and the third semantic vector is a semantic vector corresponding to tag information;

based on the at least two semantic vectors, a first semantic feature is determined.

In another possible implementation, determining the first semantic feature based on at least two semantic vectors includes:

and determining a first semantic feature corresponding to the target semantic vector.

In another possible implementation manner, determining a first semantic feature corresponding to the target semantic vector includes:

and inputting the target semantic vector into a semantic vector model, and outputting a first semantic feature, wherein the semantic vector model is used for converting the semantic vector into the semantic feature.

In another possible implementation manner, the process of determining the corresponding second semantic feature of the item image includes:

determining at least one region of interest of an image of the item;

based on the at least one second semantic vector, a second semantic feature is determined.

In another possible implementation manner, the fusing the first semantic feature and the second semantic feature to obtain the target semantic feature of the target item, includes:

converting the first semantic features and the second semantic features into semantic features with the same dimension to obtain first standard semantic features and second standard semantic features;

and fusing the first attention feature and the second attention feature to obtain a target semantic feature.

In another possible implementation manner, determining a first attention feature corresponding to the first standard semantic feature and a second attention feature corresponding to the second standard semantic feature includes:

determining a first query vector, a first key vector and a first value vector corresponding to the first standard semantic features and a second query vector, a second key vector and a second value vector corresponding to the second standard semantic features;

determining a first attention weight and a second attention weight based on the first query vector, the first key vector, the second query vector, and the second key vector;

and performing weighting processing on the first standard semantic features based on the first value vector and the first attention weight to obtain first attention features, and performing weighting processing on the second standard semantic features based on the second value vector and the second attention weight to obtain second attention features.

In another possible implementation, determining the first attention weight and the second attention weight includes:

determining a first attention weight based on a first similarity and a second similarity, wherein the first similarity is the similarity between the first query vector and the first key vector, and the second similarity is the similarity between the second query vector and the first key vector;

the second attention weight is determined based on a third similarity, which is a similarity between the first query vector and the second key vector, and a fourth similarity, which is a similarity between the second query vector and the second key vector.

determining a first attention weight based on the first similarity and the second similarity, comprising:

and normalizing the first similarity to obtain a first weight, and normalizing the second similarity to obtain a second weight.

determining a second attention weight based on the third similarity and the fourth similarity, comprising:

and normalizing the third similarity to obtain a third weight, and normalizing the fourth similarity to obtain a fourth weight.

and performing pooling processing on the attention feature to obtain a target semantic feature.

In another possible implementation manner, determining standard item information of the target item based on the target semantic features includes:

determining similarity between the target item and the plurality of standard item information based on the semantic features of the target item and the semantic features of the plurality of standard item information;

the standard article information of the target article is determined from the plurality of standard article information based on the similarity between the target article and the plurality of standard article information.

Fig. 3 is a flowchart of a method for processing item information according to an embodiment of the present application, and is executed by a server. Referring to fig. 3, the method for processing the article information includes the steps of:

301. the server acquires attribute information and an article image of the target article.

Optionally, the target item is a commodity that is shelved by a merchant through an internet platform. The commodity may be a dietetic commodity, for example, shredded pork with a fish flavour or the like; or may be a clothing-like article such as jeans. The attribute information includes at least one of an item name, a merchant name, and tag information. In one possible implementation, the attribute information includes at least an item name. In another possible implementation, the attribute information further includes at least one of a merchant name and tag information. The item image is an image including the target item.

In one possible implementation manner, the server receives attribute information and an article image of the article uploaded by the terminal in real time. Correspondingly, the method comprises the following steps: the merchant uploads the attribute information and the article image of the article to the server based on the terminal, and the server receives the attribute information and the article image of the article uploaded by the terminal to obtain the attribute information and the article image of the target article.

In the embodiment of the application, the server receives the attribute information and the article image of one article every time the terminal uploads the attribute information and the article image of one article, so that the server can timely receive the attribute information and the article image of the article uploaded by a merchant, and the timeliness of acquiring the attribute information and the article image of the target article is improved.

It should be noted that, each time the server acquires the attribute information and the item image of a target item, the server may directly perform step 302 to process the attribute information and the item image of the target item. In another possible implementation manner, the server performs batch processing on the attribute information and the object images of the plurality of target objects acquired within the first preset time period. Correspondingly, the method comprises the following specific steps: and in response to the fact that the receiving time length reaches a first preset time length, the server processes the received attribute information and the received object image of each target object through the object information processing method. In the embodiment of the present application, the value of the first preset duration is not specifically limited, and may be set and modified as needed. For example, the first preset time period is 5min, 10min, 60min, and the like.

In the embodiment of the application, the server periodically processes the attribute information and the article image of the article uploaded by the terminal, so that the quantity of article information processed each time is increased, and the processing efficiency of the article information is improved.

In a possible implementation manner, the number of the article information uploaded by the merchant in the daytime is large, the number of the article information uploaded by the merchant in the nighttime is small, and the server can adjust the processing method according to the time range of the current time. Correspondingly, the method comprises the following specific steps: the server determines the time range of the current time, and if the time range is daytime, the server directly executes step 302 to process the attribute information and the object image of the target object every time the server acquires the attribute information and the object image of one target object; and if the time range is night, responding to that the receiving time length reaches a first preset time length, and processing the received attribute information and the received object image of each target object by the server through the object information processing method. In the embodiment of the present application, the time ranges of day and night are not particularly limited. For example, the time ranges of the day are: 9 in the morning: 00 to 9 pm: 00; the time range at night was: afternoon 9: 00 to 9 am: 00.

in the embodiment of the application, the server flexibly adjusts the method for processing the article information according to the current time, directly processes the article information when the quantity of the received article information is large, and processes the article information in batches when the quantity of the received article information is small, so that the processing efficiency of the article information is improved.

302. The server determines a first semantic feature corresponding to the attribute information.

In one possible implementation, the attribute information includes at least two of an item name, a merchant name, and tag information. Correspondingly, the method comprises the following steps: the server determines at least two semantic vectors of a first semantic vector, a second semantic vector and a third semantic vector, wherein the first semantic vector is a semantic vector corresponding to an article name, the second semantic vector is a semantic vector corresponding to a merchant name, and the third semantic vector is a semantic vector corresponding to tag information; based on the at least two semantic vectors, a first semantic feature is determined. Optionally, the tag information includes a category of the item. For example, the article is "leek egg patty", and the label information of the article is "signboard patty".

In the embodiment of the application, because the first semantic features are determined by a plurality of semantic vectors, and the plurality of semantic vectors correspond to the attribute information of different information sources such as article names, merchant names, label information and the like, when the first semantic features are determined, the attribute information of the different information sources can be verified mutually, so that the accuracy of the determined first semantic features is improved.

In a possible implementation manner, the server determines the first semantic feature based on the at least two semantic vectors by: the server splices at least two semantic vectors to obtain a target semantic vector, and determines a first semantic feature corresponding to the target semantic vector. In the embodiment of the application, the server splices the semantic vectors to obtain the target semantic vector, so that the efficiency of determining the target semantic vector is improved.

In one possible implementation, the server determines the first semantic feature by a semantic vector model. Correspondingly, the step of the server determining the first semantic feature corresponding to the target semantic vector is as follows: the server inputs the target semantic vector into the semantic vector model and outputs a first semantic feature. The semantic vector model is used for converting the semantic vector into semantic features.

In the embodiment of the application, the server determines the first semantic features corresponding to the target semantic vector through the semantic vector model, and the semantic vector model is a model trained in advance, so that the accuracy of determining the first semantic features is improved.

In another possible implementation, the attribute information includes at least an item name. Correspondingly, the step of the server determining the first semantic features corresponding to the attribute information is as follows: the server determines a first semantic vector and determines a first semantic feature based on the first semantic vector; the first semantic vector is a semantic vector corresponding to the name of the article.

In the embodiment of the application, the name of the article is the main attribute information of the article, so that when the first semantic feature is determined, the interference of the secondary attribute information can be avoided, and the accuracy of the determined first semantic feature is improved.

In another possible implementation, the attribute information further includes at least one of a merchant name and tag information; correspondingly, the step of the process that the server determines the first semantic features corresponding to the attribute information is as follows: the server determines a first semantic vector and at least one of a second semantic vector and a third semantic vector, wherein the first semantic vector is a semantic vector corresponding to an article name, the second semantic vector is a semantic vector corresponding to a merchant name, and the third semantic vector is a semantic vector corresponding to tag information; the first semantic feature is determined based on the first semantic vector and at least one of the second semantic vector and the third semantic vector.

In the embodiment of the application, because the first semantic features are determined by a plurality of semantic vectors, and the plurality of semantic vectors correspond to attribute information of different information sources such as article names, merchant names, label information and the like, when the first semantic features are determined, the attribute information of the different information sources can be verified mutually, and the plurality of semantic vectors include the semantic vector corresponding to the article name, and the article name is the main attribute information of the article, so that the main attribute information is considered when the first semantic features are determined, and the attribute information of the different information sources can be verified mutually, thereby further improving the accuracy of the determined first semantic features.

Alternatively, referring to fig. 4, the semantic vector model is a BERT (Bidirectional Encoder Representation from transformations, pre-trained language Representation) model. The name of the article is pink girl, the name of the merchant is XX tea, and the label information is fruit tea. The first semantic vector is a semantic vector corresponding to pink girl, the second semantic vector is a semantic vector corresponding to XX tea, and the third semantic vector is a semantic vector corresponding to fruit tea; splicing the first semantic vector, the second semantic vector and the third semantic vector to obtain a target semantic vector; and inputting the target semantic vector into a BERT model to obtain a first semantic feature.

In the embodiment of the application, the server determines the first semantic features corresponding to the target semantic vectors through the BERT model, and the BERT model is a pre-training model with the most advanced performance, so that the accuracy of determining the first semantic features is further improved.

In another possible implementation, in order to distinguish different classes of semantic vectors, the server needs to add class information to at least two semantic vectors. Correspondingly, the server determines the first semantic features based on the at least two semantic vectors by the steps of: the server adds category information to each semantic vector, and splices at least two semantic vectors added with the category information to obtain a target semantic vector; and determining a first semantic feature corresponding to the target semantic vector.

Alternatively, referring to fig. 5, the type information includes [ CLS ]]Identification bits and numbers. The first semantic vector is represented by semantic vector i, the second semantic vector is represented by semantic vector j, and the third semantic vector is represented by semantic vector k. Add [ CLS ] to each semantic vector]Identification bits and numbers; wherein the number of the semantic vector i comprises an ID₁To ID_i(ii) a The number of semantic vector j includes an ID₁To ID_j(ii) a The number of semantic vectors k comprises an ID₁To ID_k(ii) a Will add [ CLS]Splicing the identification bits and the three numbered semantic vectors to obtain a target semantic vector; and inputting the target semantic vector into a BERT model to obtain a first semantic feature. Optionally, [ CLS]The identification bits are used for distinguishing different target semantic vectors, and the numbers are used for distinguishing a first semantic vector, a second semantic vector and a third semantic vector.

In the embodiment of the application, the server adds the category information to each semantic vector, so that the spliced target semantic vector also contains the category information, and when the semantic features are determined through the semantic vector model, the first semantic features corresponding to the target semantic vectors can be determined according to the category information, so that the efficiency of determining the first semantic features is improved.

In another possible implementation, the server first concatenates the semantic vectors and then adds the category information. Correspondingly, the process of obtaining the target semantic vector by the server is as follows: the server splices at least two semantic vectors to obtain spliced semantic vectors; and adding category information to the spliced semantic vector to obtain a target semantic vector.

In the embodiment of the application, the server directly adds the category information to the spliced semantic vectors to obtain the target semantic vectors, so that when the semantic features are determined through the semantic vector model, the first semantic features corresponding to the target semantic vectors can be determined according to the category information, and the efficiency of determining the first semantic features is improved.

In a possible implementation manner, in order to ensure that the coding lengths of the target semantic vectors corresponding to different target articles are consistent, the server determines that the coding length of the first semantic vector is a first preset length, the coding length of the second semantic vector is a second preset length, and the coding length of the third semantic vector is a third preset length. Correspondingly, the steps of the server determining the first semantic vector, the second semantic vector and the third semantic vector are as follows: the method comprises the steps that a server determines an initial semantic vector corresponding to an article name, and carries out bit supplementing or intercepting on the initial semantic vector to obtain a first semantic vector with a first preset length; determining an initial semantic vector corresponding to the name of the merchant, and performing bit complementing or intercepting on the initial semantic vector to obtain a second semantic vector with a second preset length; and determining an initial semantic vector corresponding to the tag information, and performing bit complementing or intercepting on the initial semantic vector to obtain a third semantic vector with a third preset length.

In the embodiment of the present application, the numerical values of the first preset length, the second preset length, and the third preset length are not specifically limited, and may be set and modified as needed. Optionally, the server obtains a first preset length, a second preset length, and a third preset length by performing pre-training on an MLM (Masked Language Model) based on a plurality of item data in the internet platform.

In the embodiment of the application, because the coding length of each semantic vector is fixed, the coding length of the target semantic vector obtained by splicing the plurality of semantic vectors is fixed, and further, before the target semantic vector is input into the semantic vector model, the first semantic feature corresponding to the target semantic vector can be determined without adjusting the input length of the target semantic vector, so that the efficiency of determining the first semantic feature is improved.

303. The server determines a second semantic feature corresponding to the item image.

In one possible implementation, the server determines the semantic features corresponding to the item image through the region of interest in the item image. Correspondingly, the method comprises the following steps: the server determines at least one region of interest of the item image; for each region of interest, determining a second semantic vector corresponding to the region of interest to obtain at least one second semantic vector; based on the at least one second semantic vector, a second semantic feature is determined.

In the embodiment of the application, the interested area in the article image is an area with high correlation with the article, and the server determines the semantic features corresponding to the article image according to the interested area in the article image, so that the correlation between the semantic features and the article is improved, and the accuracy of the determined second semantic features is improved.

Optionally, with continued reference to fig. 4, the server performs feature extraction on the image information through the fast R-CNN model to obtain at least one ROI (Region of Interest) corresponding to the image information. In the embodiment of the application, the Faster R-CNN model is an image model widely applied to the image field, and the efficiency and the accuracy of feature extraction can be improved through the slave image model.

Optionally, the step of determining, by the server, the second semantic vector corresponding to the region of interest includes: and the server performs pooling processing on each region of interest to obtain at least one second semantic vector with the same dimension. The pooling process may be a maximum pooling process or an average pooling process.

In a possible implementation manner, when the number of the second semantic vectors with the same dimension is one, the server directly determines that the second semantic vector is the second semantic feature. And when the number of the second semantic vectors with the same dimensionality is multiple, the server adds the multiple second semantic vectors to obtain a second semantic feature.

In the embodiment of the application, the data amount needing to be processed is reduced by performing pooling processing on the region of interest, so that the efficiency of determining the second semantic vector is improved.

In another possible implementation manner, the server determines the semantic vector corresponding to the whole article image, and then determines the second semantic feature according to the region of interest. Correspondingly, the step of determining the second semantic features corresponding to the article image by the server is as follows: the server determines an integral semantic vector corresponding to the article image; determining at least one region of interest of an image of the item; for each interested region, determining a second semantic vector corresponding to the interested region from the whole semantic vectors to obtain at least one second semantic vector; based on the at least one second semantic vector, a second semantic feature is determined.

In the embodiment of the application, the region of interest is a region with high correlation with the article, so that the second semantic vector corresponding to the region of interest is determined from the whole semantic vector based on the region of interest, the correlation with the article is high, and further the correlation between the semantic features determined based on the second semantic vector and the article is improved, so that the accuracy of the determined second semantic features is improved.

304. And the server fuses the first semantic features and the second semantic features to obtain the target semantic features of the target object.

In one possible implementation, the server fuses the two semantic features through a self-attention mechanism. Accordingly, the step of the server determining the target semantic features of the target item includes (1) to (3):

(1) and the server converts the first semantic feature and the second semantic feature into semantic features with the same dimensionality to obtain a first standard semantic feature and a second standard semantic feature.

Optionally, the server maps the first semantic feature and the second semantic feature to a vector space of the same dimension through MLP (multi layer Perceptron), so as to obtain the first standard semantic feature and the second standard semantic feature of the same dimension.

(2) The server determines a first attention feature corresponding to the first standard semantic feature and a second attention feature corresponding to the second standard semantic feature.

In one possible implementation, the server determines the first attention feature and the second attention feature according to an attention model. Accordingly, this step includes the following steps s21-s 23:

s 21: the server determines a first query vector, a first key vector and a first value vector corresponding to the first standard semantic features and a second query vector, a second key vector and a second value vector corresponding to the second standard semantic features.

Optionally, trained first query weight, first key weight and first value weight and second query weight, second key weight and second value weight are set in the attention model. Correspondingly, the method comprises the following steps: the server determines a first query weight, a first key weight and a first value weight corresponding to the first standard semantic feature according to the attention model, determines a first query vector based on the product of the first standard semantic feature and the first query weight, determines a first key vector based on the product of the first standard semantic feature and the first key weight, and determines a first value vector based on the product of the first standard semantic feature and the first value weight; and determining a second query weight, a second key weight and a second value weight corresponding to the second standard semantic feature according to the attention model, determining a second query vector based on the product of the second standard semantic feature and the second query weight, determining a second key vector based on the product of the second standard semantic feature and the second key weight, and determining a second value vector based on the product of the second standard semantic feature and the second value weight.

For example, referring to fig. 6, the attention model is a self-attention model. The first standard semantic feature uses x_tRepresenting, second criterion semantic characteristics by x_rMeaning that the first query weight is W^QMeaning that the first key weight is W^KIndicating that the first value is weighted by W^VAnd (4) showing. Q for the first query vector_tIs shown, then q_t＝W^Qx_t(ii) a K for the first key vector_tRepresents, then k_t＝W^Kx_t(ii) a Vector of first values with v_tDenotes that v is_t＝W^Vx_t. Q for the second query vector_rIs shown, then q_r＝W^Qx_r(ii) a K for the second key vector_rRepresents, then k_r＝W^Kx_r(ii) a Second value vector v_rDenotes that v is_r＝W^Vx_r。

In the embodiment of the application, the server determines the first attention feature and the second attention feature through the attention model, and the attention model can combine semantic information of the context, so that the accuracy of the determined first attention feature and the second attention feature is improved.

s 22: the server determines a first attention weight and a second attention weight based on the first query vector, the first key vector, the second query vector, and the second key vector.

In one possible implementation, the server determines the first attention weight and the second attention weight according to a similarity between the query vector and the key vector. Correspondingly, the method comprises the following steps: the server determines a first attention weight based on a first similarity and a second similarity, wherein the first similarity is the similarity between the first query vector and the first key vector, and the second similarity is the similarity between the second query vector and the first key vector; the second attention weight is determined based on a third similarity, which is a similarity between the first query vector and the second key vector, and a fourth similarity, which is a similarity between the second query vector and the second key vector.

In the embodiment of the application, the server determines the first attention weight and the second attention weight according to the similarity between the query vector and the key vector, and the query vector and the key vector comprise text features and image features, so that the semantic features of multiple modalities are fused, and the accuracy of the determined first attention weight and the second attention weight is improved.

In one possible implementation, the first attention weight includes a first weight and a second weight. Correspondingly, the server determines the first attention weight by the steps of: the server normalizes the first similarity to obtain a first weight, and normalizes the second similarity to obtain a second weight. Optionally, the server performs normalization processing on the first similarity and the second similarity through a Softmax function. For example, with continued reference to FIG. 6, the first weight is used

Representing, the second weight being

Indicating that the first similarity is alpha_ttExpressed by a, the second similarity is_rtAnd (4) showing. Accordingly, the method can be used for solving the problems that,

optionally, the step of determining, by the server, the first similarity between the first query vector and the first key vector is: the server determines a first similarity according to the first query vector and the first key vector through a first formula;

the formula I is as follows:

wherein alpha is_ttDenotes a first degree of similarity, k_tRepresenting a first key vector, q_tA first query vector is represented that represents a first query vector,

representing the distance between the first query vector and the first key vector.

Optionally, the step of the server determining the second similarity between the second query vector and the first key vector comprises: the server determines a second similarity according to the second query vector and the first key vector through a second formula;

the formula II is as follows:

wherein alpha is_rtRepresenting a second degree of similarity, k_tRepresenting a first key vector, q_rA second query vector is represented that represents a second query vector,

representing the distance between the second query vector and the first key vector.

In one possible implementation, the second attention weight includes a third weight and a fourth weight. Correspondingly, the server determines the second attention weight by the steps of: and the server normalizes the third similarity to obtain a third weight, and normalizes the fourth similarity to obtain a fourth weight. Optionally, the server performs normalization processing on the third similarity and the fourth similarity through a Softmax function. For example, with continued reference to FIG. 6, the third weight is used

Representing, fourth weight

Indicating that the third similarity is alpha_trThe fourth similarity is represented by alpha_rrAnd (4) showing. Accordingly, the method can be used for solving the problems that,

optionally, the step of determining, by the server, a third similarity between the first query vector and the second key vector is: the server determines a third similarity according to the first query vector and the second key vector through a third formula;

the formula III is as follows:

wherein alpha is_trDenotes the third similarity, k_rRepresenting a second key vector, q_tA first query vector is represented that represents a first query vector,

representing the distance between the first query vector and the second key vector.

Optionally, the step of determining, by the server, a fourth similarity between the second query vector and the second key vector is: the server determines a fourth similarity according to the second query vector and the second key vector by the following formula IV;

the formula four is as follows:

wherein alpha is_rrRepresents a fourth degree of similarity, k_rRepresenting a second key vector, q_rA second query vector is represented that represents a second query vector,

representing the distance between the second query vector and the second key vector.

In the embodiment of the application, the server performs normalization processing on the plurality of similarity degrees, and converts the plurality of similarity degrees into the same numerical value range, so that the accuracy of the determined first standard semantic feature and the second standard semantic feature is improved.

s 23: the server carries out weighting processing on the first standard semantic features based on the first value vector and the first attention weight to obtain first attention features, and carries out weighting processing on the second standard semantic features based on the second value vector and the second attention weight to obtain second attention features.

In one possible implementation, the method includes the following steps: the server carries out weighting processing on the first standard semantic features based on the product between the first weight and the first value vector and the product between the second weight and the first value vector to obtain first attention features; and performing weighting processing on the second standard semantic features based on the product between the third weight and the second value vector and the product between the fourth weight and the second value vector to obtain second attention features.

Optionally, with continued reference to FIG. 6, the first attention vector is expressed in z_tIndicating that the second attention vector is z_rRepresenting, the first value vector by v_tRepresenting, second value vector by v_rAnd (4) showing. The server carries out weighting processing on the first standard semantic features to obtain a first attention vector

The server carries out weighting processing on the second standard semantic features to obtain a second attention vector

In the embodiment of the application, the server performs weighting processing on the standard semantic features according to the value vector and the attention weight to obtain the attention features, and both the value vector and the attention weight comprise text features and image features, so that the semantic features of multiple modalities are fused, and the accuracy of the determined first attention vector and the second attention vector is improved.

(3) And the server fuses the first attention feature and the second attention feature to obtain a target semantic feature.

In one possible implementation, with continued reference to fig. 6, the present steps are: the server fuses the first attention feature and the second attention feature to obtain an attention feature; and performing pooling processing on the attention feature to obtain a target semantic feature.

In the embodiment of the application, the server fuses the characteristics of two modes of the text and the image through an attention mechanism, so that the supplement and interaction of multi-mode information are realized, and the integrity and the robustness of the obtained characteristic semantic information are ensured.

305. And the server determines standard commodity information of the target object based on the target semantic information.

In one possible implementation, the method includes the following steps: the server determines the similarity between the target item and the plurality of standard item information based on the semantic features of the target item and the semantic features of the plurality of standard item information; the standard article information of the target article is determined from the plurality of standard article information based on the similarity between the target article and the plurality of standard article information.

In the embodiment of the application, the text information sources are enriched by introducing a plurality of attribute information such as article names, merchant names and label names, the cross-mode information sources are added by introducing article images, the text information sources and the article image sources are fused on the presentation layer, the defect that the traditional information sources are single is overcome, and the accuracy of the determined standard article information is improved. For example, the multi-source information processing method in the implementation of the present application is compared with the conventional single-source information processing method by using the same sample data. The accuracy of the conventional treatment method was 97.24%, whereas the accuracy of the treatment method in the practice of the present application was 99.81%.

In one possible implementation, the plurality of standard item information is information in a standard item information base; the server determines semantic features of the information of the plurality of standard articles through a BRET model; and determining the similarity between the target object and the information of the plurality of standard objects through cosine similarity. Alternatively, referring to fig. 7, for each standard item information, the server is implemented by a double tower model, which includes three modules of an input layer, a presentation layer, and a matching layer. The server inputs attribute information, an article image and standard article information of a target article on an input layer; determining a multi-modal semantic feature corresponding to the target article and a text semantic feature corresponding to the standard article information through a BRET model and a Fast R-CNN model in the presentation layer; and then determining the similarity between the target object and the standard object information through cosine similarity.

Fig. 8 is a schematic structural diagram of an article information processing apparatus according to an embodiment of the present application. Referring to fig. 8, the apparatus includes:

an obtaining module 801, configured to obtain attribute information and an article image of a target article;

a first determining module 802, configured to determine a first semantic feature corresponding to the attribute information and a second semantic feature corresponding to the article image;

a fusion module 803, configured to fuse the first semantic feature and the second semantic feature to obtain a target semantic feature of the target item;

a second determining module 804, configured to determine standard item information of the target item based on the target semantic feature.

In one possible implementation, referring to fig. 9, the attribute information includes at least two of an item name, a merchant name, and tag information; a first determining module 802, comprising:

a first determining unit 8021, configured to determine at least two semantic vectors of a first semantic vector, a second semantic vector, and a third semantic vector, where the first semantic vector is a semantic vector corresponding to an article name, the second semantic vector is a semantic vector corresponding to a merchant name, and the third semantic vector is a semantic vector corresponding to tag information;

a second determining unit 8022, configured to determine the first semantic feature based on the at least two semantic vectors.

In another possible implementation manner, the second determining unit 8022 is configured to splice at least two semantic vectors to obtain a target semantic vector, or add category information to each semantic vector, and splice at least two semantic vectors to which the category information is added to obtain the target semantic vector; and determining a first semantic feature corresponding to the target semantic vector.

In another possible implementation manner, the second determining unit 8022 is configured to input the target semantic vector into a semantic vector model, and output the first semantic feature, where the semantic vector model is configured to convert the semantic vector into a semantic feature.

In another possible implementation manner, with continued reference to fig. 9, the first determining module 802 further includes:

a third determining unit 8023 for determining at least one region of interest of the article image;

a fourth determining unit 8024, configured to determine, for each region of interest, a second semantic vector corresponding to the region of interest to obtain at least one second semantic vector;

a fifth determining unit 8025, configured to determine the second semantic feature based on the at least one second semantic vector.

In another possible implementation, with continued reference to fig. 9, the fusion module 803 includes:

a conversion unit 8031, configured to convert the first semantic feature and the second semantic feature into semantic features with the same dimension, so as to obtain a first standard semantic feature and a second standard semantic feature;

a sixth determining unit 8032, configured to determine a first attention feature corresponding to the first standard semantic feature and a second attention feature corresponding to the second standard semantic feature;

the fusion unit 8033 is configured to fuse the first attention feature and the second attention feature to obtain a target semantic feature.

In another possible implementation manner, the sixth determining unit 8032 is configured to determine a first query vector, a first key vector, and a first value vector corresponding to the first standard semantic feature, and a second query vector, a second key vector, and a second value vector corresponding to the second standard semantic feature; determining a first attention weight and a second attention weight based on the first query vector, the first key vector, the second query vector, and the second key vector; and performing weighting processing on the first standard semantic features based on the first value vector and the first attention weight to obtain first attention features, and performing weighting processing on the second standard semantic features based on the second value vector and the second attention weight to obtain second attention features.

In another possible implementation manner, the sixth determining unit 8032 is configured to determine the first attention weight based on a first similarity, a second similarity and the first value vector, where the first similarity is a similarity between the first query vector and the first key vector, and the second similarity is a similarity between the second query vector and the first key vector; determining a second attention weight based on a third similarity, which is a similarity between the first query vector and the second key vector, a fourth similarity, which is a similarity between the second query vector and the second key vector, and the second value vector.

In another possible implementation, the first attention weight includes a first weight and a second weight; a sixth determining unit 8032, configured to perform normalization processing on the first similarity to obtain a first weight, and perform normalization processing on the second similarity to obtain a second weight; the first attention weight is determined based on a product between the first weight and the first value vector and a product between the second weight and the first value vector.

In another possible implementation, the second attention weight includes a third weight and a fourth weight; a sixth determining unit 8032, configured to perform normalization processing on the third similarity to obtain a third weight, and perform normalization processing on the fourth similarity to obtain a fourth weight; the second attention weight is determined based on a product between the third weight and the second value vector and a product between the fourth weight and the second value vector.

In another possible implementation manner, the fusion unit 8033 is configured to fuse the first attention feature and the second attention feature to obtain an attention feature; and performing pooling processing on the attention feature to obtain a target semantic feature.

In another possible implementation manner, the second determining module 804 is configured to determine similarity between the target item and the plurality of standard item information based on the target semantic feature and semantic features of the plurality of standard item information; the standard article information of the target article is determined from the plurality of standard article information based on the similarity between the target article and the plurality of standard article information.

The embodiment of the application provides a processing device of article information, because the first semantic features are semantic features corresponding to attribute information of an article, and the second semantic features are semantic features corresponding to an article image, a target semantic feature obtained by fusing the first semantic features and the second semantic features is fused with a plurality of semantic features of different modalities, standard article information is determined through the target semantic features, and the accuracy of the determined standard article information can be improved.

It should be noted that: in the article information processing apparatus provided in the above embodiment, when the article information is processed, only the division of each functional module is illustrated, and in practical applications, the functions may be distributed by different functional modules as needed, that is, the internal structure of the server may be divided into different functional modules to complete all or part of the functions described above. In addition, the article information processing apparatus and the article information processing method provided in the above embodiments belong to the same concept, and specific implementation processes thereof are described in the method embodiments and are not described herein again.

Fig. 10 is a schematic structural diagram of a server according to an embodiment of the present application, where the server 1000 may generate a relatively large difference due to different configurations or performances, and may include one or more processors (CPUs) 1001 and one or more memories 1002, where the memory 702 stores at least one program code, and the at least one program code is loaded and executed by the processors 1001 to implement the method for processing the item information provided by the above-mentioned method embodiments. Of course, the server may also have components such as a wired or wireless network interface, a keyboard, and an input/output interface, so as to perform input/output, and the server may also include other components for implementing the functions of the device, which are not described herein again.

The embodiment of the present application also provides a computer-readable storage medium, in which at least one program code is stored, and the at least one program code is loaded by a processor and has operations in the item information processing method for implementing the above-mentioned embodiment.

The embodiments of the present application also provide a computer program product, and when instructions in the computer program product are executed by a processor of a server, the server is enabled to execute the method for processing the article information in the above-mentioned method embodiments.

It will be understood by those skilled in the art that all or part of the steps of implementing the above embodiments may be implemented by hardware, or may be implemented by a program instructing relevant hardware, where the program may be stored in a storage medium, and the storage medium may be a read-only memory, a magnetic disk, an optical disk, or the like.

The above description is only an alternative embodiment of the present application and should not be construed as limiting the present application, and any modification, equivalent replacement, or improvement made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A method for processing item information, the method comprising:

acquiring attribute information and an article image of a target article;

2. The method of claim 1, wherein the attribute information includes at least two of an item name, a merchant name, and tag information; the process of determining the first semantic feature corresponding to the attribute information includes:

3. The method of claim 2, wherein determining the first semantic feature based on the at least two semantic vectors comprises:

4. The method of claim 3, wherein the determining the first semantic feature corresponding to the target semantic vector comprises:

5. The method according to claim 1, wherein the step of determining the corresponding second semantic feature of the item image comprises:

determining at least one region of interest of the item image;

6. The method according to claim 1, wherein the fusing the first semantic feature and the second semantic feature to obtain the target semantic feature of the target item comprises:

7. The method of claim 6, wherein the determining a first attention feature corresponding to the first standard semantic feature and a second attention feature corresponding to the second standard semantic feature comprises:

determining a first query vector, a first key vector and a first value vector corresponding to the first standard semantic feature and a second query vector, a second key vector and a second value vector corresponding to the second standard semantic feature;

and performing weighting processing on the first standard semantic feature based on the first value vector and the first attention weight to obtain the first attention feature, and performing weighting processing on the second standard semantic feature based on the second value vector and the second attention weight to obtain the second attention feature.

8. The method of claim 7, wherein determining a first attention weight and a second attention weight based on the first query vector, the first key vector, the second query vector, and the second key vector comprises:

determining the first attention weight based on a first similarity and a second similarity, wherein the first similarity is a similarity between the first query vector and the first key vector, and the second similarity is a similarity between the second query vector and the first key vector;

determining the second attention weight based on a third similarity and a fourth similarity, the third similarity being a similarity between the first query vector and the second key vector, the fourth similarity being a similarity between the second query vector and the second key vector.

9. The method of claim 8, wherein the first attention weight comprises a first weight and a second weight;

the determining the first attention weight based on the first similarity and the second similarity comprises:

and normalizing the first similarity to obtain the first weight, and normalizing the second similarity to obtain the second weight.

10. The method of claim 8, wherein the second attention weight comprises a third weight and a fourth weight;

the determining the second attention weight based on the third similarity and the fourth similarity comprises:

and normalizing the third similarity to obtain the third weight, and normalizing the fourth similarity to obtain the fourth weight.

11. The method of claim 6, wherein the fusing the first attention feature and the second attention feature to obtain the target semantic feature comprises:

12. The method according to any one of claims 1-11, wherein the determining standard item information of the target item based on the target semantic features comprises:

13. An apparatus for processing article information, the apparatus comprising:

14. A server, characterized in that the server comprises:

a processor and a memory, the memory having stored therein at least one instruction, the at least one instruction being loaded and executed by the processor to implement operations in the method of processing item information of any one of claims 1 to 12.

15. A computer-readable storage medium having stored therein at least one instruction, which is loaded and executed by a processor to implement the operations performed in the method for processing the article information according to any one of claims 1 to 12.