Product picture label generation method and device, equipment, medium and product

Abstract

The application discloses a commodity picture label generation method and device, equipment, medium and product, and the method comprises the steps of calling a picture description text generation model to generate a commodity description text corresponding to a commodity picture; calling a commodity word extraction model to extract a plurality of key commodity words contained in the commodity description text; calling a search word acquisition algorithm to acquire associated search words corresponding to each key commodity word, and storing each key commodity word and the associated search words as candidate picture labels in a picture label list; determining a plurality of candidate picture labels in the picture label list which are ranked in the front and whose commodity static information and commodity dynamic information satisfy preset conditions, and taking each candidate picture label as a picture label of the commodity picture. The application can provide a picture label automatic generation service for a network e-commerce platform, generate picture labels for the commodity pictures of each commodity in the platform, and save the time cost of manually setting the picture labels.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of network e-commerce, in particular to a method for generating a picture label of a commodity, and further relates to a device, equipment, nonvolatile storage medium and computer program product corresponding to the method. BACKGROUND

[0002] In the existing network platform, a corresponding picture label (picture ALT label) is usually set for a picture. The picture label refers to a keyword prompt given to the picture. By setting the picture label for the picture, the user in the platform can search for the corresponding picture through the keyword. In the network e-commerce platform, the seller user in the platform uploads the corresponding commodity picture of the commodity sold online to introduce the commodity. The picture label associated with the sold commodity is set for the commodity picture. The commodity picture with the picture label related to the commodity has accessibility in search, so that the commodity of the commodity picture is more easily searched by the buyer user through the keyword when the buyer user searches for the commodity to be purchased. However, the picture label of the existing commodity picture is generally set by the seller user. In addition to consuming a large amount of human cost of the seller user, the picture label set by the seller user for the commodity picture may not be concise and express the characteristic attributes of the corresponding commodity, so that the set picture label cannot effectively improve the probability of the commodity being searched by the buyer user, or the probability of the commodity being searched is affected because the picture label set for the commodity picture is too long.

[0003] In view of the problems existing in setting the picture label for the picture, the present applicant makes corresponding exploration for solving the problems. SUMMARY

[0004] The present application aims to meet the needs of users and provide a method for generating a picture label of a commodity, and further relates to a device, equipment, nonvolatile storage medium and computer program product corresponding to the method.

[0005] To achieve the purpose of the present application, the following technical solutions are adopted:

[0006] A picture description text generation model is called to generate a commodity description text corresponding to the commodity picture;

[0007] A commodity word extraction model is called to extract a plurality of key commodity words contained in the commodity description text;

[0008] The associated search words corresponding to each of the key commodity words are obtained, and each of the key commodity words and the associated search words are stored in a picture label list as alternative picture labels, and each of the alternative picture labels in the picture label list is sorted according to the search volume;

[0009] Determine multiple candidate picture tags in the picture tag list that are ranked high and whose static information and dynamic information of the goods meet preset conditions, and use each of the candidate picture tags as a picture tag of the product picture.

[0010] In further embodiments, before the step of calling the picture description text generation model to generate the product description text corresponding to the product picture, the method comprises

[0011] Obtaining product key information of product objects corresponding to multiple product pictures, the product key information comprising product title text and product attribute text;

[0012] Calling a product description text generation model to generate training product description texts for each of the product key information;

[0013] Using each of the product pictures and its corresponding training product description text as a training sample to train the picture description text generation model until it converges.

[0014] In further embodiments, in the step of calling the picture description text generation model to generate the product description text corresponding to the product picture, the method comprises:

[0015] Extracting an image semantic feature vector of the product picture, inputting the image semantic feature vector into an input encoding layer, and obtaining image semantic feature encoding data output by the input encoding layer;

[0016] Inputting the image semantic feature encoding data into an output decoding layer, obtaining image semantic feature decoding data output by the output decoding layer, and performing normalization processing on the image semantic feature decoding data to obtain the product description text.

[0017] In further embodiments, in the step of calling the picture description text generation model to generate the product description text corresponding to the product picture, the method comprises:

[0018] Extracting an image semantic feature vector of the product picture, inputting the image semantic feature vector into an input encoding layer, and obtaining image semantic feature encoding data output by the input encoding layer;

[0019] Inputting the image semantic feature encoding data into an output decoding layer, using a diversity decoding strategy, and obtaining multiple product description texts corresponding to multiple image semantic feature decoding data output by the output decoding layer;

[0020] Obtaining a product title corresponding to the product picture, and calling a product title matching rule to determine a product description text that matches the product title from the multiple product description texts.

[0021] In a further embodiment, in the step of calling a commodity word extraction model to extract a plurality of key commodity words contained in the commodity description text, the step comprises:

[0022] extracting a text semantic feature vector of the commodity description text;

[0023] According to the text semantic feature vector, determine the words or phrases in the commodity description text that hit the preset commodity attribute type, and the commodity attribute type includes commodity category, commodity brand, commodity color or commodity size;

[0024] Filter out the words or phrases in the commodity description text that exceed the preset score as key commodity words.

[0025] In a further embodiment, obtain the associated search words corresponding to each key commodity word, store each key commodity word and the associated search word as an alternative picture tag in a picture tag list, and in the step of sorting each alternative picture tag in the picture tag list according to search volume, the step comprises:

[0026] Combine a plurality of key commodity words corresponding to a commodity picture to generate a plurality of key commodity word combinations, and obtain one or more associated search words corresponding to each key commodity word combination;

[0027] Generate a picture tag list for the commodity picture, and store each key commodity word and each key commodity word as an alternative picture tag in the picture tag list;

[0028] Obtain the search volume of each alternative picture tag in the picture tag list and perform reverse order sorting.

[0029] In a further embodiment, in the step of determining a plurality of alternative picture tags in the picture tag list that are ranked in the front and whose commodity static information and commodity dynamic information satisfy a preset condition, and storing each alternative picture tag as a picture tag of the commodity picture, the step comprises:

[0030] Identify the semantic feature similarity between each alternative picture tag in the picture tag list and the commodity picture as commodity static information;

[0031] Obtain the commodity competition degree and commodity conversion rate of each alternative picture tag in the picture tag list as commodity dynamic information;

[0032] Filter out the compliant alternative picture tags in each alternative picture tag, whose semantic similarity of commodity static information exceeds a preset similarity threshold, and whose commodity competition degree in commodity dynamic information satisfies a low competition rule and whose commodity conversion rate exceeds a preset minimum conversion rate threshold;

[0033] A plurality of target candidate picture tags meeting the ranking-in-front rule are determined from the picture tag list, and each of the target candidate picture tags is used as a picture tag of the product picture.

[0034] A product picture tag generation device is provided for the purpose of the present application, comprising:

[0035] A description text generation module is configured to invoke a product picture description text generation model to generate a product description text corresponding to a product picture.

[0036] A product word extraction module is configured to invoke a product word extraction model to extract a plurality of key product words contained in the product description text.

[0037] A tag list generation module is configured to obtain an associated search word corresponding to each of the key product words, and store each of the key product words and the associated search word as a candidate picture tag in a picture tag list, wherein each of the candidate picture tags in the picture tag list is ranked according to search volume.

[0038] A picture tag screening module is configured to screen a plurality of candidate picture tags that meet a preset condition in terms of product static information and product dynamic information from the picture tag list, and use each of the candidate picture tags as a picture tag of the product picture.

[0039] In a further embodiment, the description text generation module comprises:

[0040] An image feature extraction submodule is configured to extract an image semantic feature vector of a product picture, input the image semantic feature vector into an input encoding layer, and obtain image semantic feature encoding data output by the input encoding layer.

[0041] A unique description text generation module is configured to input the image semantic feature encoding data into an output decoding layer, obtain image semantic feature decoding data output by the output decoding layer, and perform normalization processing on the image semantic feature decoding data to obtain a product description text.

[0042] In a preferred embodiment, the description text generation module further comprises:

[0043] An image feature extraction submodule is configured to extract an image semantic feature vector of a product picture, input the image semantic feature vector into an input encoding layer, and obtain image semantic feature encoding data output by the input encoding layer.

[0044] A diverse description text generation submodule is configured to input the image semantic feature encoding data into an output decoding layer, use a diversity decoding strategy, and obtain a plurality of product description texts corresponding to a plurality of image semantic feature decoding data output by the output decoding layer.

[0045] The text matching submodule is configured to obtain a product title corresponding to the product picture, call a product title matching rule, and determine a product description text that matches the product title from a plurality of product description texts.

[0046] In a further embodiment, the product word extraction module includes:

[0047] The text segmentation processing submodule is configured to perform text segmentation processing on the product description text to obtain a text segmentation sequence.

[0048] The semantic feature extraction submodule is configured to extract semantic features of each word in the text segmentation sequence to obtain a text semantic feature sequence.

[0049] The product attribute type determination submodule is configured to determine an entity type to which each word belongs based on the semantic features of each word in the text semantic feature sequence, and filter out words whose entity type is a preset product attribute type as key product words. The product attribute type includes a product category, a product brand, a product color, or a product size.

[0050] In a further embodiment, the label list generation module includes:

[0051] The associated search word acquisition submodule is configured to combine a plurality of key product words corresponding to the product picture to generate a plurality of key product word combinations, and obtain one or more associated search words corresponding to each key product word combination.

[0052] The picture label list generation submodule is configured to generate a picture label list for the product picture, and store each key product word and each key product word as a candidate picture label in the picture label list.

[0053] The picture label sorting submodule is configured to obtain a search volume of each candidate picture label in the picture label list and perform reverse sequence sorting.

[0054] In a further embodiment, the picture label screening module includes:

[0055] The product static information generation submodule is configured to identify a semantic feature similarity between each candidate picture label in the picture label list and the product picture as product static information.

[0056] The product dynamic information generation submodule is configured to obtain a product competition degree and a product conversion rate of each candidate picture label in the picture label list as product dynamic information.

[0057] The compliance picture tag screening submodule is configured to screen out a compliance candidate picture tag from the candidate picture tags, wherein the compliance candidate picture tag has a semantic similarity of static information of the commodity exceeding a preset similarity threshold, a competition degree of the commodity in dynamic information of the commodity satisfying a low-competition rule, and a conversion rate of the commodity exceeding a preset minimum conversion rate threshold.

[0058] The target picture tag determination submodule is configured to determine a plurality of target candidate picture tags from the compliance candidate picture tags in the picture tag list, wherein the target candidate picture tags satisfy an ordering-early rule, and each of the target candidate picture tags is used as a picture tag of the commodity picture.

[0059] To solve the above technical problems, the embodiments of the present application further provide a commodity picture tag generation device, which comprises a memory and a processor, and the memory stores computer readable instructions. When the computer readable instructions are executed by the processor, the processor executes the steps of the commodity picture tag generation method.

[0060] To solve the above technical problems, the embodiments of the present application further provide a storage medium storing computer readable instructions. When the computer readable instructions are executed by one or more processors, the one or more processors execute the steps of the commodity picture tag generation method.

[0061] To solve the above technical problems, the embodiments of the present application further provide a computer program product, which comprises a computer program and computer instructions. When the computer program and computer instructions are executed by a processor, the processor executes the steps of the commodity picture tag generation method.

[0062] Compared with the prior art, the advantages of the present application are as follows:

[0063] The application can provide a picture tag automatic generation service for automatically generating picture tags of product pictures for a network e-commerce platform to save the labor cost consumed by a buyer user in the network e-commerce platform for editing and setting picture tags for product pictures. After obtaining a product picture of a product object, a product description text describing the product object displayed by the image of the product picture is generated, and key product words describing the product features of the product object contained in the product description text are extracted. Based on the key product words, associated search words associated with the product features of the product object in a search engine or a product search service are obtained, and then each key product word and the associated search word are stored in a picture tag list acting on the product picture as a candidate picture tag. The picture tag list is sorted in reverse order according to the search volume of each candidate picture tag, that is, the candidate picture tag with higher search volume is sorted higher in the picture tag list. Then, the product picture similarity, product competition degree, product conversion rate and search volume are used as screening dimensions to screen the candidate picture tags that meet the conditions of each screening dimension from the picture tag list as the picture tags of the product picture. It can be seen that in addition to automatically generating the picture tags of the product picture for the buyer user, the generated picture tags also conform to the product object displayed by the image of the product picture, and have the characteristics of low competition degree, high conversion rate and high search volume. The generated picture tags can effectively improve the probability of the product object being searched by the buyer user. BRIEF DESCRIPTION OF DRAWINGS

[0064] The above and / or additional aspects and advantages of the present application will become apparent and more readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings, in which:

[0065] Figure 1 A typical network deployment architecture diagram related to the technical solutions of the present application;

[0066] Figure 2 A flowchart of a typical embodiment of the product picture tag generation method of the present application;

[0067] Figure 3 A flowchart of a specific implementation of the present application for training a picture description text generation model;

[0068] Figure 4 A flowchart of a specific implementation of the present application for calling a picture description text generation model to generate a product description text for a product picture;

[0069] Figure 5 A flowchart of a specific implementation of the present application for calling a picture description text generation model of a sampling diversity generation strategy to generate multiple product description texts for a product picture for product title matching screening;

[0070] Figure 6 A flowchart of a specific embodiment of the present application for calling a commodity word extraction model to extract key commodity words contained in a commodity description text;

[0071] Figure 7 A flowchart of a specific embodiment of the present application for obtaining associated search words of the key commodity words, and storing the key commodity words and the associated search words as alternative picture tags into a picture tag list for search volume sorting;

[0072] Figure 8 A flowchart of a specific embodiment of the present application for screening out alternative picture tags with a high ranking and commodity static information and commodity dynamic information satisfying preset conditions from the picture tag list, and taking the alternative picture tags as picture tags of the commodity picture;

[0073] Figure 9 A principle block diagram of a typical embodiment of the commodity picture tag generation device of the present application;

[0074] Figure 10 A basic structure block diagram of the commodity picture tag generation device of one embodiment of the present application. DETAILED DESCRIPTION

[0075] Embodiments of the present application are described in detail below with reference to the accompanying drawings. The embodiments described below are examples for explaining the present application and should not be interpreted as limiting the present application.

[0076] It should be understood by those skilled in the art that the singular forms “a,” “an,” and “the” used in this application include plural references unless specifically stated otherwise. It should be further understood that the use of the term “include” in the specification of the present application means that the stated features, integers, steps, operations, elements, and / or components are present, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there can be intermediate elements. In addition, “connected” or “coupled” used herein can include wireless connection or wireless coupling. The phrase “and / or” used herein includes all or any single unit and all combinations of the associated listed items.

[0077] As those skilled in the art will appreciate, unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. It will be further appreciated that terms, such as those defined in commonly used dictionaries, should be given their ordinary and customary meaning, unless explicitly defined otherwise herein and unless the context clearly dictates otherwise.

[0078] As those skilled in the art will appreciate, unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. It will be further appreciated that terms, such as those defined in commonly used dictionaries, should be given their ordinary and customary meaning, unless explicitly defined otherwise herein and unless the context clearly dictates otherwise. As those skilled in the art will appreciate, unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. It will be further appreciated that terms, such as those defined in commonly used dictionaries, should be given their ordinary and customary meaning, unless explicitly defined otherwise herein and unless the context clearly dictates otherwise. As those skilled in the art will appreciate, unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. It will be further appreciated that terms, such as those defined in commonly used dictionaries, should be given their ordinary and customary meaning, unless explicitly defined otherwise herein and unless the context clearly dictates otherwise.

[0079] The hardware referred to by the names "server", "client", "worker node" and the like in the present application is essentially an electronic device with the equivalent capability of a personal computer, which is a hardware device with a central processing unit (including an arithmetic unit and a controller), a memory, an input device, and an output device, and other necessary components disclosed by the von Neumann principle. The computer program is stored in the memory, the central processing unit calls the program stored in the external memory into the memory for running, executes the instructions in the program, interacts with the input and output devices, and thereby completes the specific functions.

[0080] It should be noted that the concept of "server" in the present application can also be extended to the case of server cluster. According to the network deployment principle understood by those skilled in the art, the servers should be logically divided, and in physical space, these servers can be independent of each other but can be called through an interface, or can be integrated into a physical computer or a computer cluster. Those skilled in the art should understand this variation and should not be restricted by the implementation of the network deployment of the present application.

[0081] Please refer to Figure 1 The hardware basis required for the implementation of the technical solutions related to the present application can be deployed according to the architecture shown in the figure. The server 80 referred to in the present application is deployed in the cloud and serves as an online server, which can be responsible for further connecting related data servers and other servers that provide related support, etc., to form a logically associated service cluster to provide services for related terminal devices such as the smart phone 81 and the personal computer 82 shown in the figure or a third-party server (not shown). The smart phone and the personal computer can access the Internet through a publicly known network access method and establish a data communication link with the server 80 in the cloud to run the terminal application program related to the services provided by the server.

[0082] For the server, the application program is usually built as a service process, and the corresponding program interface is opened for the application program running on various terminal devices to make remote calls. The related technical solutions suitable for running on the server in the present application can be implemented in the server in this way.

[0083] The application program refers to the application program running on the server or the terminal device. This application program implements the related technical solutions of the present application in a programming manner, and the program code can be saved in a non-volatile storage medium that can be recognized by a computer in the form of computer executable instructions, and can be called into the memory by the central processing unit for running. The related device of the present application is constructed through the running of the application program on the computer.

[0084] For the server, the application program is usually built as a service process, opens the corresponding program interface, and is remotely called by the application program running on various terminal devices. The related technical solutions suitable for running on the server in the present application can be implemented in the server in this way.

[0085] Those skilled in the art should know that the various methods of the present application, although based on the same concept and common to each other, are independently executable unless otherwise specified. Similarly, for each embodiment disclosed in the present application, it is based on the same inventive concept, so the same concept and the concept appropriately transformed although the concept is different, should be understood as the same.

[0086] Please refer to Figure 2 A commodity picture tag generation method of the present application includes the following steps in a typical embodiment:

[0087] Step S11, calling a picture description text generation model to generate a commodity description text corresponding to the commodity picture:

[0088] The commodity picture refers to the picture uploaded by the seller user in the network e-commerce platform for the commodity object for sale, for example, the commodity cover picture of the commodity object in the network e-commerce platform, etc. After the seller user uploads the commodity picture for the commodity object, the picture tag automatic generation service built by the present method can automatically generate the corresponding picture tag for the uploaded commodity picture. The picture tag refers to the ALT tag, which is used to describe the image content displayed in the commodity picture in the form of text. For example, a mobile phone with a description of A brand can be set with the picture tags of "A brand" and "mobile phone". Setting the picture tags of the commodity picture in accordance with the description of the commodity object can improve the accessibility of the commodity picture and its search ranking in the search, so that the commodity object in the network e-commerce platform is easy to be searched by the buyer user through the keyword, so as to improve the exposure rate and sales of the commodity object described by the commodity picture.

[0089] After obtaining the product picture uploaded by the seller user for the product object, the picture description text generation model is called to generate the product description text describing the product object shown in the product picture. The picture description text generation model is a neural network model constructed based on the neural network idea, has an image feature extraction layer for extracting an image semantic feature vector of the product picture, and is generally constructed based on a deep residual network (ResNet). After the image semantic feature vector is extracted, the image semantic feature vector is input into an input encoding layer for encoding, and the image semantic feature encoding data encoded by the input encoding layer is input into an output decoding layer to output image semantic feature decoding data. The image semantic feature decoding data is normalized by a Softmax algorithm to generate a text corresponding to the image semantic feature decoding data. The text is taken as the product description text corresponding to the product picture. The input encoding layer and the output decoding layer are constructed based on the Encoder and Decoder structure of the Transformer. As can be seen, the picture description text generation model is a neural network model constructed by the image feature extraction layer based on the deep residual network (ResNet), the Encoder and Decoder based on the Transformer, and the encoding and decoding layer based on the Softmax structure.

[0090] In an embodiment, the output encoding layer in the picture description text generation model can output image semantic feature decoding data corresponding to the image semantic feature decoding data, use a diversified generation strategy for the image semantic feature decoding data, generate a plurality of product description texts, and then obtain a product title corresponding to the product picture. The product title matching rule is called to determine the product description text that matches the product title from the plurality of product description texts. The diversified generation strategy can be a temperature control strategy (Temperature Scaling), a top-k sampling strategy (Top-k Sampling), a random sampling strategy (Random Sampling), or the like, which can be implemented by a person skilled in the art as needed.

[0091] The training sample used by the picture description text generation model is generally composed of a commodity picture and a training commodity description text. The commodity picture is input into the picture description text generation model that is not trained to convergence. Then, the commodity description text generated by the picture description text generation model and the training commodity description text are used to calculate the loss by using a loss function. When the loss meets the condition, the picture description text generation model is trained to completion. The training commodity description text corresponding to the commodity picture in the training sample is generally generated by a commodity description text generation model that is trained to convergence. The commodity key information of the commodity object in the commodity picture is used as the input of the commodity description text generation model, such as the commodity title text, the commodity attribute text (commodity category, commodity brand, commodity material, commodity size, or commodity color), and the like, to generate a text that can describe the commodity object as the training commodity description text.

[0092] In step S12, a commodity word extraction model is called to extract a plurality of key commodity words contained in the commodity description text.

[0093] After the commodity description text of the commodity picture is generated by calling the picture description text generation model that is trained to convergence, the commodity word extraction model that is trained to convergence is called to extract a plurality of key commodity words in the commodity description text.

[0094] The commodity word extraction model is generally a neural network model constructed based on the named entity recognition (NER) idea. After the commodity description text is processed by word segmentation, the corresponding text segmentation sequence is obtained. Then, the text segmentation sequence is input into the commodity word extraction model. The commodity word extraction model extracts the semantic features of each word in the text segmentation sequence by using a semantic feature extraction layer composed of a BiLSTM model, a Bert model, or a Bert model and a BiLSTM model, to generate a corresponding text semantic feature sequence. Then, the text semantic feature sequence is input into an entity type recognition layer constructed by CRF to determine the entity type of each word in the text semantic feature sequence based on the semantic features of each word in the text semantic feature sequence. The target words with the preset commodity attribute type are selected from each word to be the key commodity words.

[0095] The preset commodity attribute types in the entity types generally include commodity category types, commodity brand types, commodity color types, commodity size types, or commodity material types, etc. For example, when the commodity description text is "A brand blue small size silk shirt", "A brand" hits the commodity brand type, "blue" hits the commodity color type, "small size" hits the commodity size type, "silk" hits the commodity material type, and "shirt" hits the commodity category type. It can be seen that by training the commodity word extraction model to convergence, the entity type corresponding to each word or each word group in the commodity description text is determined, and then the words or word groups with the entity type of the preset commodity attribute type are screened out as the key commodity words in the commodity description text. That is, the words or word groups related to the commodity attribute in the commodity description text are taken as the key commodity words, the key commodity words are taken as the reference words for searching the associated search words of the commodity picture, and the key commodity words are taken as the alternative picture labels of the commodity picture.

[0096] The key commodity words related to the commodity attribute in the commodity description text are extracted by calling the commodity word extraction model after generating the commodity description text for the commodity picture by the picture description text generation model. It can be seen that by using the picture description text generation model and the commodity word extraction model, the key commodity words representing the key features of the commodity object shown in the commodity picture are generated as the alternative picture labels, that is, the key features of the commodity object shown in the commodity picture are identified, and the key features are taken as the alternative picture labels acting on the commodity picture. In order to further improve the available picture labels of the commodity picture, the search word obtaining algorithm will be used later to obtain the associated search words associated with the key commodity words representing the key features of the commodity object shown in the commodity picture, to expand the words or word groups used to describe the key features of the commodity object shown in the commodity picture, and the associated search words will be taken as the alternative picture labels acting on the commodity picture to increase the number of alternative picture labels of the commodity picture.

[0097] In step S13, the associated search words corresponding to each key commodity word are obtained, and each key commodity word and the associated search word are stored in the picture label list as an alternative picture label. Each alternative picture label in the picture label list is sorted by search volume:

[0098] After extracting the key commodity words from the commodity description text generated for the commodity picture, the associated search words associated with the key commodity words are obtained, and the search volume of the key commodity words and the associated search words is obtained.

[0099] The associated search terms refer to search terms related to the key commodity terms, which are used together with the key commodity terms in the search in the third-party search engine (such as Google search engine or Baidu search engine) or the commodity search service of the network e-commerce platform. For example, when the key commodity terms are "trousers" and "large size", the search terms commonly used together with the key commodity terms for search can be "men", "work pants" or "jeans", and the search terms can be used as the associated search terms of the key commodity terms. In addition, the search terms that are the same as the search object of the key commodity terms can also be used as the associated search terms. For example, when the key commodity terms are "bicycle", the search terms that are the same as the search object of the key commodity terms can be "bicycle" or "bicycle", and the search terms can be used as the associated search terms of the key commodity terms.

[0100] Regarding the way of obtaining the associated search terms corresponding to the key commodity terms, the source of the associated search terms is generally related. When the associated search terms are obtained from the third-party search engine, the search keyword optimization interface provided by the third-party search engine can be used. For example, when the third-party search engine is Google search engine, the Google AdWords Keyword Planner provided by the Google platform can be used to obtain the associated search terms of the key commodity terms in the Google search engine. When the associated search terms are obtained from the network e-commerce platform, the associated search terms related to the key commodity terms of the network e-commerce platform can be obtained by using the crawler tool, or the search terms associated with the key commodity terms in the commodity search service of the network e-commerce platform can be obtained as the associated search terms.

[0101] In order to improve the number of obtained associated search terms and optimize the relevance of the obtained associated search terms and the key commodity, the key commodity terms can be combined and then the associated search terms are obtained. That is, a plurality of key commodity terms are combined into a long text, and the associated search terms corresponding to the long text are obtained. For example, the key commodity terms corresponding to the commodity picture are "A brand", "male", "trousers", and "black". The key commodity terms "trousers" and "black" can be combined into key commodity combination A, and the key commodity terms "A brand" and "black" can be combined into key commodity combination B. The associated search terms obtained for the key commodity combination A can be "dark color clothing", and the associated search terms obtained for the key commodity combination B can be "A brand dark color clothing".

[0102] The search volume refers to the search times of the key commodity word or the associated search word as a search word in a third-party search engine or a network e-commerce platform. Of course, it can also be the search times in a period of time. The search volume is generally related to the third-party search engine or the network e-commerce platform from which the associated search word is obtained. For example, when the associated search word is obtained from a third-party search engine, the search volume of the key commodity word and the associated search word will be the total search times or the search times in a period of time of both in the third-party search engine. If the associated search word is obtained from the commodity search service of a network e-commerce platform, the search volume of the key commodity word and the associated search word will be the total search times or the search times in a period of time of both in the commodity search service. Of course, the search volume of the key commodity word and the associated search word can not be limited by the source of the associated search word. One of other third-party search engines or other network e-commerce platforms can be selected as the search volume of the key commodity word and the associated search word for statistics. The search volume of the key commodity word and the associated search word obtained from multiple third-party search engines or network e-commerce platforms can also be calculated by weight to determine the comprehensive search volume of the key commodity word and the associated search word in different third-party search engines or network e-commerce platforms.

[0103] After obtaining the associated search word corresponding to the key commodity word, the key commodity word and the associated search word can be stored in the picture tag list as candidate picture tags. The picture tag list acts on the commodity picture to which the key commodity word belongs, and based on the search volume of each of the key commodity word and the associated search word, the key commodity word and the associated search word stored in the picture tag list are sorted in reverse order from large to small.

[0104] Step S14, determining a plurality of candidate picture tags in the picture tag list that are sorted in front and whose commodity static information and commodity dynamic information satisfy the preset condition, and taking each of the candidate picture tags as a picture tag of the commodity picture:

[0105] After storing the key commodity word and the associated search word as candidate picture tags in the picture tag list and sorting in reverse order based on the search volume, a plurality of candidate pictures that are sorted in front and whose commodity static information and commodity dynamic information satisfy the preset condition are selected from the picture tag list as picture tags of the commodity picture.

[0106] The commodity static information refers to the semantic feature similarity between the text of the alternative picture label and the commodity picture, that is, the similarity between the text of the alternative picture label and the image displayed by the commodity picture. Generally, the semantic feature vector of the text of the alternative picture label is extracted, and the semantic feature vector of the commodity picture is extracted, and the semantic feature vectors of both are mapped to the same semantic space, and the distance between the two semantic feature vectors is calculated as the semantic feature similarity between the text of the alternative picture label and the commodity picture. The semantic feature similarity calculation can be implemented using any one of the large-scale vector retrieval engines such as Faiss, ElasticSearch, Milvus, etc. It can also use any ready-made algorithm such as cosine similarity, inner product, Manhattan distance, Euclidean distance, etc. to calculate. Correspondingly, the preset condition of commodity static information is generally to judge whether the feature similarity between the alternative picture label and the commodity picture is greater than or equal to the preset similarity threshold. If yes, the semantic between the text of the alternative picture label and the commodity picture is similar, and the commodity static information of the alternative picture label meets the preset condition.

[0107] The commodity dynamic information includes the commodity competition degree of the alternative picture label. The commodity competition degree is determined by the competition intensity, competition cost and competition risk when the alternative picture label is used as the picture label of the commodity picture. Generally, the competition intensity is determined by the sales amount of the e-commerce store using the alternative picture label as the picture label, the online time of the store and the sales amount of the commodity. The competition cost is determined by the payment required to use the alternative picture label as the picture label to the search engine or the network e-commerce platform. The competition risk degree is determined by the competition cost and the conversion rate of using the alternative picture label as the picture label. The commodity competition degree is generally divided into three levels: high, medium and low. The higher the commodity competition degree, the more intense the competition intensity, the more the competition cost, and the higher the competition risk. Therefore, the preset condition for the commodity competition degree is generally to judge whether the commodity competition degree is low. If yes, it indicates that the commodity competition degree of the commodity object is low when the alternative picture label is used as the picture label, which meets the preset condition.

[0108] The commodity dynamic information includes not only the competition degree of the commodity using the alternative picture label as the picture label of the commodity, but also the conversion rate of the commodity using the alternative picture label as the picture label of the commodity. The conversion rate is generally determined by counting the click rate and the purchase rate of the commodity object after the commodity object using the alternative picture label as the picture label of the commodity is searched. The conversion rate can reflect the purchase intention of the buyer when the commodity object using the alternative picture label as the picture label of the commodity is searched. Therefore, the preset condition for the conversion rate is generally to determine whether the conversion rate is greater than or equal to a preset minimum conversion rate threshold. If yes, it indicates that the conversion rate of the commodity object using the alternative picture label as the picture label of the commodity is high, and the alternative picture label meets the preset condition.

[0109] The compliant alternative picture labels are the alternative picture labels that are semantically related to the picture label of the commodity. When the compliant alternative picture labels are used as the picture label of the commodity, the competition degree of the commodity object to which the commodity picture belongs is low and the purchase intention of the buyer is strong. In addition, the top-ranked compliant alternative picture labels in the picture label list, i.e., the compliant alternative picture labels with high search volume, are selected as the picture label of the commodity. For example, the compliant alternative picture labels ranked before a preset rank in the picture label list are selected as the picture label of the commodity. For example, the compliant alternative picture labels ranked before the 50th rank in the picture label list are selected as the picture label of the commodity. Alternatively, a preset number of top-ranked compliant alternative picture labels are selected. For example, when the current compliant alternative picture labels are the compliant alternative picture label A ranked 5th, the compliant alternative picture label B ranked 10th, the compliant alternative picture label C ranked 25th, the compliant alternative picture label D ranked 49th, and the compliant alternative picture label E ranked 67th, three top-ranked compliant alternative picture labels, i.e., the compliant alternative picture label A, the compliant alternative picture label B, and the compliant alternative picture label C, are selected as the picture label of the commodity.

[0110] The above typical embodiments and their variant embodiments fully disclose the implementation of the commodity picture label generation method of the present application. However, various variant embodiments of the method can be derived by transforming and expanding some technical means, as summarized below:

[0111] In one embodiment, before the step of invoking the picture description text generation model to generate the commodity description text corresponding to the commodity picture, the method comprises the steps of: Figure 3 , calling the picture description text generation model to generate the commodity description text corresponding to the commodity picture,

[0112] Step S07, obtaining the product key information of the product object corresponding to the plurality of product pictures, the product key information including product title text and product attribute text:

[0113] The product key information belongs to the product title text and the product attribute text of the product picture to which the product picture belongs. Generally, the product title text refers to the product title text edited and uploaded by the seller user for the product object. The product attribute text generally includes product category text representing the product category of the product object, product brand text representing the product brand of the product object, product material text representing the product material of the product object, product size text representing the product size of the product object, and product color text representing the product color of the product object, and other texts for introducing the product attributes of the product object.

[0114] Step S08, calling a product description text generation model to generate training product description text for each product key information:

[0115] The product description text generation model generally refers to a neural network-based model that extracts the text semantic features of the product title text and each product attribute text included in the product construction information, and then extracts the keywords in the product title text and each product attribute text based on the text semantic feature information. The text semantic features of the keywords are combined into training product description text for training the picture description text generation model. The network for keyword extraction and keyword combination can be constructed based on the CNN algorithm or the FCN algorithm, and the network for text semantic feature extraction can be constructed based on the Bert model.

[0116] In addition, the ChatGpt model can also be used as a product description text generation model to generate text based on the product title text and each product attribute text included in the product construction information as training product description text. By inputting the obtained product title text and each product attribute text into the ChatGpt model, and driving the ChatGpt model to generate text using the product title text and each product attribute text, a short text generated by using the product title text and each product attribute text to make sentences is obtained as training product description text.

[0117] Step S09, using each of the product pictures and its corresponding training product description text as training samples to train the picture description text generation model to convergence:

[0118] The training commodity description text corresponding to each commodity picture is generated, and after the commodity pictures and the training commodity description texts are used as training samples, each commodity picture is input into the picture description text generation model which is not trained to convergence, and then the training commodity description text generated by the picture description text generation model is used for loss calculation by using a loss function until the loss meets the condition, and then the picture description text generation model is trained.

[0119] In the embodiment, the training commodity description text for model training is generated based on the commodity title text and the commodity attribute text of the commodity object to which the commodity picture belongs by using the commodity description text generation model, and then the picture description text generation model is trained by using the commodity picture and the training commodity model text as training samples, so that a large number of training samples can be generated for the picture description text generation model to improve the training speed of the picture description text generation model to convergence.

[0120] In one embodiment, referring to Figure 4 The step of generating the commodity description text corresponding to the commodity picture by calling the picture description text generation model includes the following steps.

[0121] In step S111, the image semantic feature vector of the commodity picture is extracted, and the image semantic feature vector is input into the input encoding layer to obtain the image semantic feature encoding data output by the input encoding layer.

[0122] The picture description text generation model is a neural network model constructed based on the neural network idea, and has an image feature extraction layer for extracting the image semantic feature vector of the commodity picture. The image feature extraction layer is generally constructed based on a deep residual network (ResNet).

[0123] After the image semantic feature of the commodity picture is extracted by the picture feature extraction layer, the image semantic feature vector is input into the input encoding layer for encoding to obtain the image semantic feature encoding data generated by the input encoding layer after the image semantic feature training is encoded. The input encoding layer is generally constructed based on the Encoder structure of the Transformer.

[0124] In step S112, the image semantic feature encoding data is input into the output decoding layer to obtain the image semantic feature decoding data output by the output decoding layer, and the image semantic feature decoding data is normalized to obtain the commodity description text.

[0125] The image semantic feature encoding data generated by the input encoding layer is input into the output decoding layer, and the image semantic feature encoding data encoded by the input encoding layer is decoded by the output decoding layer to generate image semantic feature decoding data. The image semantic feature decoding data is normalized by using the Softmax algorithm to generate the text corresponding to the image semantic feature decoding data. The text is used as the commodity description text corresponding to the commodity picture. The input encoding layer is generally constructed based on the Decoder (decoding) structure of the Transformer.

[0126] In this embodiment, the picture description text generation model constructed based on the encoding and decoding structure of the Transformer is used to generate the commodity description text describing the commodity object displayed in the image of the commodity picture, so that the picture label generated based on the commodity description text can more accurately describe the commodity object displayed in the commodity picture.

[0127] In one embodiment, referring to Figure 5 , the step of calling the picture description text generation model to generate the commodity description text corresponding to the commodity picture includes:

[0128] Step S111', the image semantic feature vector of the commodity picture is extracted, and the image semantic feature vector is input into the input encoding layer to obtain the image semantic feature encoding data output by the input encoding layer:

[0129] This step refers to the related embodiments in step S111, and this step will not be described here.

[0130] Step S112', the image semantic feature encoding data is input into the output decoding layer, and a diversity decoding strategy is used to obtain a plurality of commodity description texts corresponding to a plurality of image semantic feature decoding data output by the output decoding layer:

[0131] After obtaining the image semantic feature encoding data generated by the input encoding layer training and encoding the image semantic feature, the image semantic feature encoding data will be input into the output decoding layer using the diversity decoding strategy for decoding. The diversity generation strategy can be a temperature control strategy (Temperature Scaling), a top-k sampling strategy (Top-k Sampling) or a random sampling strategy (Random Sampling), so that the decoding layer sampling diversity generation strategy decodes the image semantic feature encoding data to generate a plurality of image semantic feature decoding data, and generates the commodity description texts corresponding to each of the image semantic feature decoding data, so that the output decoding layer can generate a plurality of commodity description texts based on one image semantic feature encoding data.

[0132] Step S113', obtaining the product title corresponding to the product picture, calling a product title matching rule, and determining a product description text that is most matched with the product title of the product object to which the product picture belongs from a plurality of product description texts:

[0133] After obtaining the plurality of product description texts output by the output decoding layer, the product title matching rule is called to filter the product description text that is most matched with the product title of the product object to which the product picture belongs from the product description texts.

[0134] The product title matching rule can be a keyword matching based on a regular expression or a text matching model constructed based on a neural network. The text matching model performs semantic similarity calculation between the text semantic features of each product description text and the product title, so as to take the product description text with the highest semantic similarity with the product title as the product description text of the product picture.

[0135] In this embodiment, the picture description text generation model generates a plurality of product description texts for the product picture through the diversity generation strategy, and filters the product description text that is most matched with the product title of the product picture from the generated plurality of product description texts, so that the product description text not only takes the product picture as the generation object, but also further filters the product title corresponding to the product picture, so as to improve the accuracy of the product description text in describing the product object displayed by the product picture.

[0136] In one embodiment, referring to Figure 6 The step of calling the product word extraction model to extract a plurality of key product words contained in the product description text includes:

[0137] Step S121, performing word segmentation processing on the product description text to obtain a text segmentation sequence:

[0138] After obtaining the product description text, the product description text is generally subjected to word segmentation processing to generate a text segmentation sequence corresponding to the product description text. For example, when the product description text is a Chinese text, the Jieba word segmentation algorithm can be used to perform word segmentation processing on the product description text. When the product description text is a non-Chinese text, the Subword word segmentation algorithm (BPE, WordPiece, ULM) or the like can be used to perform word segmentation processing on the product description text. Of course, a named entity recognition model constructed based on a neural network can also be used to perform word segmentation processing on the product description text of any language.

[0139] Step S122, extracting semantic features of each word in the text segmentation sequence to obtain a text semantic feature sequence:

[0140] After obtaining the text segmentation sequence corresponding to the commodity description text, semantic feature extraction is performed on each text segmentation in the text segmentation sequence to obtain a text semantic feature sequence composed of semantic features of each text segmentation. The semantic feature extraction layer for performing semantic feature extraction can be constructed by a BiLSTM model, a Bert model, or a Bert model and a BiLSTM model.

[0141] In step S123, based on the semantic features of each text segmentation in the text semantic feature sequence, the entity type to which each text segmentation belongs is determined, and text segmentations with an entity type of a preset commodity attribute type are screened out as key commodity words. The commodity attribute type includes a commodity category, a commodity brand, a commodity color, or a commodity size.

[0142] After extracting the semantic features of each text segmentation in the text segmentation sequence to generate a corresponding text semantic feature sequence, the text semantic feature sequence is input into an entity type recognition layer constructed by a CRF. Based on the semantic features of each text segmentation in the text semantic feature sequence, the entity type to which each text segmentation belongs is determined, and target text segmentations with an entity type of a preset commodity attribute type are screened out from each text segmentation. These target text segmentations are used as key commodity words.

[0143] In this embodiment, a commodity word extraction model constructed based on the idea of named entity recognition is used to determine the entity type corresponding to each word or each word group in the commodity description text, and then words or word groups with an entity type of a preset commodity attribute type are screened out as key commodity words in the commodity description text. That is, words or word groups related to commodity attributes in the commodity description text are used as key commodity words. These key commodity words can be used as alternative picture labels for the commodity picture, improving the commodity description accuracy of the picture labels of the commodity picture.

[0144] In one embodiment, referring to Figure 7 , the associated search words corresponding to each key commodity word are obtained, and each key commodity word and the associated search word are stored in the picture label list as alternative picture labels. In the step of sorting each alternative picture label in the picture label list by search volume, the following steps are included.

[0145] In step S131, a plurality of key commodity words corresponding to a commodity picture are combined to generate a plurality of key commodity word combinations, and one or more associated search words corresponding to each key commodity word combination are obtained.

[0146] The key commodity words are combined and then associated search words are obtained to combine multiple key commodity words into long texts, and associated search words corresponding to the long texts are obtained. For example, the key commodity words corresponding to a commodity picture are "A brand", "male", "pants", and "black". The key commodity words "pants" and "black" can be combined into key commodity word combination A, and the key commodity words "A brand" and "black" can be combined into key commodity combination B. The associated search words obtained for the key commodity word combination A can be "dark color clothing", and the associated search words obtained for the key commodity word combination B can be "A brand dark color clothing", so as to increase the number of obtained associated search words and optimize the relevance of the obtained associated search words to the key commodities.

[0147] In step S132, a picture label list acting on the commodity picture is generated, and each key commodity word and each key commodity word are stored in the picture label list as alternative picture labels.

[0148] Regarding the manner of obtaining associated search words corresponding to key commodity words, generally, the source of the associated words is related. When the associated search words obtained from a third-party search engine, the search keyword optimization interface provided by the third-party search engine can be used to obtain the associated search words acting on the key commodity words in the third-party search engine. For example, when the third-party search engine is a Google search engine, the search keyword optimization interface provided by the Google AdWords Keyword Planner of the Google platform can be used to obtain the associated search words acting on the key commodity words in the Google search engine. When the associated search words are obtained from a network e-commerce platform, the associated search words related to the key commodity words on the network e-commerce platform can be obtained by using a crawler tool, or the search words associated with the key commodity words in the commodity search service of the network e-commerce platform can be obtained as the associated search words.

[0149] The key commodity words and the associated search words can be stored in the picture label list as alternative picture labels.

[0150] In step S133, the search volumes of each alternative picture label in the picture label list are obtained and sorted in reverse order.

[0151] The search volume refers to the search times of the key commodity words or the associated search words as search words in a third-party search engine or a network e-commerce platform. Of course, it can also be the search times in a period of time. After determining the search volumes of each alternative picture label in the picture label list, the storage order of the alternative picture labels in the picture label list is sorted in reverse order based on the search volumes of the alternative picture labels from large to small.

[0152] In this embodiment, by obtaining the associated search words of the key commodity words, the number of alternative picture labels of the commodity picture is expanded, so that more diverse and search-related words can be selected as the picture labels of the commodity picture, and the search volume of the alternative picture labels is obtained as a reference parameter for subsequent picture label screening, so that the selected picture labels for the commodity picture can improve the probability of the commodity object of the commodity picture being searched by the buyer user and improve the exposure of the commodity object.

[0153] In one embodiment, referring to Figure 8 , the step of determining a plurality of alternative picture labels in the picture label list that are ranked in the front and whose commodity static information and commodity dynamic information satisfy a preset condition, and taking each of the alternative picture labels as a picture label of the commodity picture, includes:

[0154] In step S141, the semantic feature similarity between each alternative picture label in the picture label list and the commodity picture is identified as commodity static information:

[0155] The semantic feature similarity refers to the semantic similarity between the text of the alternative picture label and the image displayed by the commodity picture. Generally, the semantic feature vector of the text of the alternative picture label is extracted, and the semantic feature vector of the commodity picture is extracted, and the semantic feature vectors of both sides are mapped to the same semantic space, and the distance between the two semantic feature vectors is calculated as the semantic feature similarity between the text of the alternative picture label and the commodity picture. The semantic feature similarity calculation can use any one of Faiss, ElasticSearch, Milvus, etc. Large-scale vector retrieval engine, or any ready-made algorithm such as cosine similarity, inner product, Manhattan distance, Euclidean distance, etc. can be used to calculate.

[0156] In step S142, the commodity competition degree and commodity conversion rate of each alternative picture label in the picture label list are obtained as commodity dynamic information:

[0157] The competition degree of the commodity is determined by the competition intensity, competition cost and competition risk when the alternative picture label is used as the picture label of the commodity. Generally, the competition intensity is determined by the sales volume of the e-commerce store, the online time of the store and the sales volume of the commodity. The competition cost is determined by the payment required to use the alternative picture label as the picture label to the search engine or the network e-commerce platform. The competition risk degree is determined by the competition cost and the conversion rate of using the alternative picture label as the picture label. The competition degree of the commodity is generally divided into three levels: high, medium and low. The higher the competition degree of the commodity, the more intense the competition intensity, the more the competition cost, and the higher the competition risk.

[0158] The commodity dynamic information includes the commodity conversion rate of the alternative picture label in addition to the commodity competition degree of the alternative picture label. The commodity conversion rate is generally determined by counting the click rate and purchase rate of the commodity object after the commodity object is searched using the alternative picture label as the commodity picture, and can reflect the purchase willingness of the buyer user when the commodity object is searched using the alternative picture label as the picture label of the commodity picture.

[0159] In step S143, the compliant alternative picture label is screened out from each of the alternative picture labels, in which the semantic similarity of the commodity static information exceeds the preset similarity threshold, and the commodity competition degree in the commodity dynamic information satisfies the low competition rule and the commodity conversion rate exceeds the preset minimum conversion rate threshold:

[0160] The preset condition of the commodity static information is generally to determine whether the feature similarity between the alternative picture label and the commodity picture is greater than or equal to the preset similarity threshold. If yes, the text of the alternative picture label and the commodity picture are semantically similar, and the commodity static information of the alternative picture label satisfies the preset condition.

[0161] The preset condition of the commodity competition degree is generally to determine whether the commodity competition degree is a low-level commodity competition degree. If yes, it indicates that the commodity competition degree of the commodity object is low when the alternative picture label is used as the picture label, and the preset condition is satisfied.

[0162] The preset condition set for the commodity conversion rate is generally to determine whether the commodity conversion rate is greater than or equal to the preset minimum conversion rate threshold. If yes, it indicates that the commodity conversion rate of the commodity object of the commodity picture is high when the alternative picture label is used as the picture label of the commodity picture, and the alternative picture label satisfies the preset condition.

[0163] The compliant alternative picture label is screened out from the alternative picture label stored in the picture label list, in which the feature similarity between the alternative picture label and the commodity picture is greater than or equal to the preset similarity threshold, the commodity competition degree is a low-level commodity competition degree, and the commodity conversion rate is greater than or equal to the minimum conversion rate threshold. The compliant alternative picture label is semantically related to the commodity picture. When the compliant alternative picture label is used as the picture label of the commodity picture, the competition degree of the commodity object to which the commodity picture belongs is low and the purchase willingness of the buyer user is strong.

[0164] In step S144, a plurality of target alternative picture labels are determined from the compliant alternative picture labels of the picture label list, in which the ranking satisfies the ranking front rule, and each of the target alternative picture labels is used as the picture label of the commodity picture:

[0165] After the screening of the various compliance candidate picture tags from the picture tag list, the compliance candidate picture tags that are ranked high in the picture tag list, i.e. the compliance candidate picture tags with high search volume, are screened out from the compliance candidate picture tags. Generally, the compliance candidate picture tags ranked before a preset ranking in the picture tag list are selected as the picture tags of the product picture, or a preset number of compliance candidate picture tags ranked high in the picture tag list are selected.

[0166] In this embodiment, by using the product picture similarity, the product competition degree, the product conversion rate and the search volume as the screening parameters, the candidate picture tags that meet the conditions of the various screening parameters are screened out from the picture tag list as the picture tags of the product picture, so that the picture tags screened for the product picture can meet the product object displayed by the product picture and have the characteristics of low competition degree, high conversion rate and high search volume.

[0167] Further, each step in the method disclosed in the above embodiments can be functioned to construct a product picture tag generation device of the present application. According to this idea, please refer to Figure 9 In one typical embodiment, the device comprises: a description text generation module 11 configured to call a picture description text generation model to generate a product description text corresponding to a product picture; a product word extraction module 12 configured to call a product word extraction model to extract a plurality of key product words contained in the product description text; a tag list generation module 13 configured to obtain associated search words corresponding to each of the key product words, and store each of the key product words and the associated search words as candidate picture tags in a picture tag list, wherein each of the candidate picture tags in the picture tag list is ranked according to search volume; and a picture tag screening module 14 configured to screen a plurality of candidate picture tags ranked high in the picture tag list and meeting preset conditions of product static information and product dynamic information, and use each of the candidate picture tags as a picture tag of the product picture.

[0168] In one embodiment, the description text generation module 11 comprises: an image feature extraction sub-module configured to extract an image semantic feature vector of a product picture, input the image semantic feature vector into an input encoding layer, and obtain image semantic feature encoding data output by the input encoding layer; and a unique description text generation module configured to input the image semantic feature encoding data into an output decoding layer, obtain image semantic feature decoding data output by the output decoding layer, and perform normalization processing on the image semantic feature decoding data to obtain a product description text.

[0169] In another embodiment, the description text generation module 11 further comprises: an image feature extraction submodule for extracting an image semantic feature vector of the product picture, inputting the image semantic feature vector into an input encoding layer, and obtaining image semantic feature encoding data output by the input encoding layer; a diverse description text generation submodule for inputting the image semantic feature encoding data into an output decoding layer, using a diversity decoding strategy, and obtaining a plurality of product description texts corresponding to a plurality of image semantic feature decoding data output by the output decoding layer; and a description text matching submodule for obtaining a product title corresponding to the product picture, calling a product title matching rule, and determining a product description text that is more matched with the product title from the plurality of product description texts.

[0170] In an embodiment, the product word extraction module 12 comprises: a text segmentation processing submodule for performing text segmentation processing on the product description text to obtain a text segmentation sequence; a semantic feature extraction submodule for extracting semantic features of each segmentation word in the text segmentation sequence to obtain a text semantic feature sequence; and a product attribute type determination submodule for determining an entity type to which each segmentation word belongs based on the semantic features of each segmentation word in the text semantic feature sequence, screening out segmentation words whose entity types are preset product attribute types as key product words, and the product attribute types include a product category, a product brand, a product color, or a product size.

[0171] In an embodiment, the label list generation module 13 comprises: an associated search word acquisition submodule for combining a plurality of key product words corresponding to a product picture to generate a plurality of key product word combinations, and obtaining one or more associated search words corresponding to each key product word combination; a picture label list generation submodule for generating a picture label list acting on the product picture, and storing each key product word and each key product word as an alternative picture label in the picture label list; and a picture label sorting submodule for obtaining a search volume of each alternative picture label in the picture label list and performing reverse order sorting.

[0172] In one embodiment, the image tag filtering module 14 includes: a product static information generation submodule, used to identify the semantic feature similarity between each candidate image tag in the image tag list and the product image as product static information; a product dynamic information generation submodule, used to obtain the product competition degree and product conversion rate of each candidate image tag in the image tag list as product dynamic information; a compliant image tag filtering submodule, used to filter out compliant candidate image tags whose semantic similarity of product static information exceeds a preset similarity threshold, and whose product competition degree in product dynamic information meets the low competition rule and whose product conversion rate exceeds a preset minimum conversion rate threshold; and a target image tag determination submodule, used to determine multiple target candidate image tags in the image tag list whose sorting meets the sorting priority rule, and use each target candidate image tag as the image tag of the product image.

[0173] To address the aforementioned technical problems, this application also provides a product image label generation device for running a computer program implemented according to the product image label generation method. Please refer to the following for details. Figure 10 , Figure 10 This is a basic structural block diagram of the product image label generation device in this embodiment.

[0174] like Figure 10 The diagram shows the internal structure of a product image label generation device. This device includes a processor, non-volatile storage medium, memory, and a network interface connected via a system bus. The non-volatile storage medium stores an operating system, a database, and computer-readable instructions. The database stores control information sequences. When the processor executes the computer-readable instructions, it enables the processor to implement a product image label generation method. The processor provides computing and control capabilities, supporting the operation of the entire device. The memory stores computer-readable instructions, which, when executed by the processor, enable the processor to execute a product image label generation method. The network interface is used for communication with a terminal. Those skilled in the art will understand that… Figure 10 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the product image label generation device to which the present application is applied. A specific product image label generation device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0175] The processor in this embodiment is used to execute the specific functions of each module / submodule in the product picture label generation device of the present application, and the memory stores the program codes and various data required for executing the above modules. The network interface is used for data transmission between the user terminal or the server. The memory in this embodiment stores the program codes and data required for executing all modules / submodules in the product picture label generation device, and the server can call the program codes and data of the server to execute the functions of all submodules.

[0176] The present application also provides a non-volatile storage medium, wherein the product picture label generation method is written into a computer program and stored in the storage medium in the form of computer readable instructions. When the computer readable instructions are executed by one or more processors, it means that the program is running in the computer, thereby causing the one or more processors to execute the steps of the product picture label generation method of any of the above embodiments.

[0177] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by a computer program instructing related hardware. The computer program can be stored in a computer readable storage medium. When the program is executed, it can include the processes of the above-mentioned embodiments. The storage medium can be a non-volatile storage medium such as a magnetic disc, an optical disc, a read-only memory (ROM), or a random access memory (RAM).

[0178] In summary, the present application can provide a picture label automatic generation service for a network e-commerce platform to generate picture labels for the product pictures of each product in the platform, thereby saving the time cost of manually setting picture labels.

[0179] It should be understood that although each step in the flowchart of the accompanying drawings is displayed in sequence according to the direction of the arrow, these steps are not necessarily executed in sequence according to the direction of the arrow. Unless otherwise stated herein, the execution of these steps is not strictly limited in sequence, and they can be executed in other orders. Moreover, at least part of the steps in the flowchart of the accompanying drawings can include multiple sub-steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution sequence is not necessarily sequential, but can be executed alternately or alternately with at least part of other steps or sub-steps or stages of other steps.

[0180] Those skilled in the art can understand that the steps, measures and schemes in the various operations, methods and processes discussed in the present application can be alternated, changed, combined or deleted. Further, other steps, measures and schemes in the various operations, methods and processes discussed in the present application can also be alternated, changed, rearranged, decomposed, combined or deleted. Further, the steps, measures and schemes in the various operations, methods and processes in the prior art can also be alternated, changed, rearranged, decomposed, combined or deleted.

[0181] The above only describes some embodiments of the present application. It should be pointed out that those skilled in the art can make several improvements and refinements without departing from the principles of the present application, and these improvements and refinements should also be considered as the protection scope of the present application.

Claims

1. A product picture label generation method characterized by, The method comprises the following steps: calling a picture description text generation model to generate a product description text corresponding to a product picture; calling a product word extraction model to extract a plurality of key product words contained in the product description text; obtaining associated search words corresponding to each of the key product words, and storing each of the key product words and the associated search words as candidate picture labels in a picture label list, wherein each of the candidate picture labels in the picture label list is sorted according to search volume; determining a plurality of candidate picture labels in the picture label list that are ranked in the front and whose product static information and product dynamic information satisfy a preset condition, and taking each of the candidate picture labels as a picture label of the product picture, comprising: recognizing a semantic feature similarity between each candidate picture label in the picture label list and the product picture as product static information; obtaining a product competition degree and a product conversion rate of each candidate picture label in the picture label list as product dynamic information; screening compliant candidate picture labels from each of the candidate picture labels, wherein the semantic similarity of the product static information of each of the candidate picture labels exceeds a preset similarity threshold, the product competition degree in the product dynamic information satisfies a low-competition rule, and the product conversion rate exceeds a preset minimum conversion rate threshold; determining a plurality of target candidate picture labels from each of the compliant candidate picture labels in the picture label list that satisfy a ranking-in-the-front rule, and taking each of the target candidate picture labels as a picture label of the product picture.

2. The product picture label generation method according to claim 1, characterized by, Before the step of calling the picture description text generation model to generate the product description text corresponding to the product picture, the method comprises: obtaining product key information of a product object corresponding to a plurality of product pictures, wherein the product key information comprises a product title text and a product attribute text; calling the product description text generation model to generate a training product description text for each of the product key information; training the picture description text generation model as a training sample together with the training product description text corresponding thereto until the picture description text generation model converges.

3. The product picture tag generation method according to claim 1, characterized by, In the step of calling the picture description text generation model to generate the product description text corresponding to the product picture, the method comprises: extracting an image semantic feature vector of the product picture, inputting the image semantic feature vector into an input encoding layer to obtain image semantic feature encoding data output by the input encoding layer; inputting the image semantic feature encoding data into an output decoding layer to obtain image semantic feature decoding data output by the output decoding layer, and performing normalization processing on the image semantic feature decoding data to obtain the product description text.

4. The product picture tag generation method according to claim 1, characterized by, In the step of calling the picture description text generation model to generate the product description text corresponding to the product picture, the method comprises: extracting an image semantic feature vector of the product picture, inputting the image semantic feature vector into an input encoding layer to obtain image semantic feature encoding data output by the input encoding layer; inputting the image semantic feature encoding data into an output decoding layer, using a diversity decoding strategy, and obtaining a plurality of product description texts corresponding to a plurality of image semantic feature decoding data output by the output decoding layer; Obtaining the corresponding commodity title of the commodity picture, calling the commodity title matching rule, and determining the commodity description text that matches the commodity title from multiple commodity description texts.

5. The product picture tag generation method according to claim 1, characterized by, In the step of calling the commodity word extraction model and extracting multiple key commodity words contained in the commodity description text, the following steps are included: Performing word segmentation processing on the commodity description text to obtain a text segmentation sequence: Extracting the semantic features of each word in the text segmentation sequence to obtain a text semantic feature sequence: Based on the semantic features of each word in the text semantic feature sequence, determining the entity type to which each word belongs, and screening the words whose entity type is a preset commodity attribute type as key commodity words, wherein the commodity attribute type includes commodity category, commodity brand, commodity color, or commodity size.

6. The product picture label generation method according to any one of claims 1 to 5, characterized by, In the step of obtaining the associated search words corresponding to each key commodity word, storing each key commodity word and the associated search word as a candidate picture label in a picture label list, and sorting each candidate picture label in the picture label list according to search volume, the following steps are included: Combining multiple key commodity words corresponding to the commodity picture to generate multiple key commodity word combinations, and obtaining one or more associated search words corresponding to each key commodity word combination; Generating a picture label list for the commodity picture, and storing each key commodity word and each key commodity word as a candidate picture label in the picture label list; Obtaining the search volume of each candidate picture label in the picture label list and performing reverse order sorting.

7. A product picture label generation apparatus characterized by comprising: The method includes the following steps: A description text generation module is configured to call a picture description text generation model to generate a commodity description text corresponding to a commodity picture. A commodity word extraction module is configured to call a commodity word extraction model to extract multiple key commodity words contained in the commodity description text. A label list generation module is configured to obtain the associated search words corresponding to each key commodity word, store each key commodity word and the associated search word as a candidate picture label in a picture label list, and sort each candidate picture label in the picture label list according to search volume. A picture label screening module is configured to screen multiple candidate picture labels in the picture label list that are ranked in the front and whose commodity static information and commodity dynamic information satisfy a preset condition, and store each candidate picture label as a picture label of the commodity picture, including the following steps: A commodity static information generation submodule is configured to identify the semantic feature similarity between each candidate picture label in the picture label list and the commodity picture as commodity static information. A commodity dynamic information generation submodule is configured to obtain the commodity competition degree and commodity conversion rate of each candidate picture label in the picture label list as commodity dynamic information. A compliant picture label screening submodule is configured to screen compliant candidate picture labels from each candidate picture label, wherein the semantic similarity of the commodity static information of the compliant candidate picture label exceeds a preset similarity threshold, the commodity competition degree in the commodity dynamic information satisfies a low competition rule, and the commodity conversion rate exceeds a preset minimum conversion rate threshold. The target picture label determination submodule is configured to determine a plurality of target candidate picture labels from the picture label list, and each of the target candidate picture labels is used as a picture label of the product picture.

8. The merchandise picture tag generation apparatus of claim 7, wherein The label list generation module comprises: The associated search term acquisition submodule is configured to combine a plurality of key product terms corresponding to the product picture, generate a plurality of key product term combinations, and acquire one or more associated search terms corresponding to each of the key product term combinations; The picture label list generation submodule is configured to generate a picture label list for the product picture, and store each of the key product terms and each of the key product terms as a candidate picture label in the picture label list; The picture label sorting submodule is configured to sort each of the candidate picture labels in the picture label list in a reverse order according to a search volume of each of the candidate picture labels.

9. A product picture label generation apparatus comprising a central processing unit and a memory, characterized by The central processing unit is configured to call and run a computer program stored in the memory to perform the steps of the method according to any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that, The computer program is stored in the form of computer readable instructions and is implemented according to the method of any one of claims 1 to 6. When the computer program is called and run by a computer, the steps included in the corresponding method are performed.

Images