Page design evaluation method, topic scoring model training method and related equipment

Abstract

The embodiment of the invention provides a page design evaluation method, a theme scoring model training method and related equipment, and the method comprises the steps: obtaining a page screenshot corresponding to a to-be-evaluated shop page, the page screenshot comprising text information and an image, a text coordinate corresponding to the text information, and an image coordinate corresponding to the image; obtaining a text vector based on the text information, the text coordinate and the image coordinate; obtaining image features of the page screenshot based on the page screenshot and a preset image model; converting the image features by using a first feature conversion module to obtain an image vector; and based on the text vector, the image vector and a preset causal language model, obtaining a scoring result of the to-be-evaluated shop page. According to the method, the design efficiency of the website page of the shop can be improved, and the labor cost for designing the shop page can be reduced.

Description

Technical Field

[0001] The present application belongs to the field of interface design and relates to image processing technology, and in particular to a page design evaluation method, a training method for a topic scoring model and related equipment. Background Art

[0002] Cross-border e-commerce can be operated through independent websites, and users can purchase different products on different independent websites. The user interface (UI) design provided by stores running on independent websites is flexible and diverse, and is a key factor in measuring the personalized experience of stores. Since this interface design is too dependent on the personal experience and subjectivity of merchants or professional UI designers, and there is no unified evaluation standard, frequent design changes may occur, resulting in high labor costs and low interface design efficiency. Summary of the invention

[0003] The embodiments of the present application provide a page design evaluation method, a training method for a topic scoring model, a page design evaluation device, and a computer device, which can solve the technical problem that UI design relies on personal experience or subjective judgment.

[0004] A first aspect of an embodiment of the present application provides a page design evaluation method, comprising: obtaining a page screenshot corresponding to a store page to be evaluated, the page screenshot comprising text information and an image, text coordinates corresponding to the text information, and image coordinates corresponding to the image; obtaining a text vector based on the text information, the text coordinates, and the image coordinates; obtaining image features of the page screenshot based on the page screenshot and a preset image model; converting the image features using a first feature conversion module to obtain an image vector; and obtaining a scoring result for the store page to be evaluated based on the text vector, the image vector, and a preset causal language model.

[0005] In some optional embodiments of the present application, obtaining a text vector based on the text information, the text coordinates and the image coordinates includes: combining the text information, the text coordinates and the image coordinates to obtain sentence information; performing text processing and word segmentation processing on the sentence information to obtain multiple word units; performing vector conversion on each word unit to obtain a word vector corresponding to each word unit; and concatenating each word vector to obtain the text vector.

[0006] In some optional embodiments of the present application, the use of the first feature conversion module to convert the image features to obtain an image vector includes: processing the image features according to the first linear layer, dimensional conversion operation, and normalization operation of the first feature conversion module to obtain a first feature; processing the image features according to the second linear layer, the dimensional conversion operation, and the normalization operation of the first feature conversion module to obtain a second feature; processing the image features according to the third linear layer, the dimensional conversion operation, and the normalization operation of the first feature conversion module to obtain a third feature; and processing the image features using the fourth linear layer of the first feature conversion module. The first feature is processed by the first linear layer and the second linear layer of the first feature conversion module to obtain a seventh feature output by the sixth linear layer; the sum of the seventh feature and the sixth feature is normalized to obtain the image vector.

[0007] According to a second aspect of an embodiment of the present application, there is provided a training method for a topic rating model, the training method comprising: determining training data based on page screenshots of store pages corresponding to a plurality of stores; obtaining image set features based on the training data and an image model; obtaining a first prediction value based on the page screenshot, a first feature conversion module and a second feature conversion module; obtaining a second prediction value based on the page screenshot, the image model, the first feature conversion module and a causal language model; calculating a first loss value based on the image set features and the first prediction value; calculating a second loss value based on the second prediction value and an actual rating result; adjusting model parameters of the topic rating model based on the first loss value and the second loss value to obtain a trained topic rating model, the topic rating model comprising the image model, the first feature conversion module, the second feature conversion module and the causal language model.

[0008] In some optional embodiments of the present application, the training data is determined based on the page screenshots of store pages corresponding to multiple stores, including: extracting the sub-images in the page screenshots based on the image coordinates corresponding to the sub-images in the page screenshots to generate an image set; and determining the training data based on the image set and the number of sub-images in each page screenshot in the image set.

[0009] In some optional embodiments of the present application, the first prediction value is obtained based on the page screenshot, the first feature conversion module and the second feature conversion module, including: based on the page screenshot, using the image model to obtain the image features of the page screenshot; using the first feature conversion module to convert the image features to obtain an image vector; using the second feature conversion module to process the image vector to obtain the first prediction value.

[0010] In some optional embodiments of the present application, the second prediction value is obtained based on the page screenshot, the image model, the first feature conversion module and the causal language model, including: obtaining a text vector based on the text information of the page screenshot, the text coordinates corresponding to the text information and the image coordinates; extracting the features of the page screenshot using the image model to obtain the image features of the page screenshot; converting the image features using the first feature conversion module to obtain an image vector; and obtaining the second prediction value of the rating result for each store based on the text vector and the image vector and using the causal language model.

[0011] In some optional embodiments of the present application, obtaining the actual scoring result includes: determining a target store from the multiple stores; determining a traffic score based on the number of independent visitors to the store page of the target store and the transaction volume after visiting the store page; obtaining the daily transaction volume of each store, and determining the first store with the largest daily transaction volume and the second store with the smallest daily transaction volume from the multiple stores based on the daily transaction volume of each store; determining a transaction ratio based on the traffic score and the daily transaction volume of the target store; obtaining a transaction score based on the daily transaction volume of the target store, the daily transaction volume corresponding to the first store, the daily transaction volume corresponding to the second store and the transaction ratio; obtaining the actual scoring result based on the traffic score and the transaction score.

[0012] The third aspect of an embodiment of the present application provides a page design evaluation device, including: an acquisition module, used to acquire a page screenshot corresponding to a store page to be evaluated, the page screenshot including text information and an image, text coordinates corresponding to the text information, and image coordinates corresponding to the image; a text module, used to obtain a text vector based on the text information, the text coordinates and the image coordinates; an image module, used to obtain image features of the page screenshot based on the page screenshot and a preset image model; a feature module, used to convert the image features using a first feature conversion module to obtain an image vector; a scoring module, used to obtain a scoring result for the store page to be evaluated based on the text vector, the image vector and a preset causal language model.

[0013] In a fourth aspect of an embodiment of the present application, a training device for a topic scoring model is provided, comprising: a determination module, used to determine training data based on page screenshots of store pages corresponding to multiple stores; an image module, used to obtain image set features based on the training data and the image model; a prediction module, used to obtain a first prediction value based on the page screenshot, a first feature conversion module and a second feature conversion module; the prediction module is also used to obtain a second prediction value based on the page screenshot, the image model, the first feature conversion module and a causal language model; a calculation module, used to calculate a first loss value based on the image set features and the first prediction value; the calculation module is also used to calculate a second loss value based on the second prediction value and an actual scoring result; an evaluation module, used to adjust model parameters of the topic scoring model based on the first loss value and the second loss value to obtain a trained topic scoring model, wherein the topic scoring model includes the image model, the first feature conversion module, the second feature conversion module and the causal language model.

[0014] An embodiment of the present application also provides a computer device, including: a memory, and a processor, wherein the processor executes computer-readable instructions stored in the memory to implement the page design evaluation method and the topic scoring model training method.

[0015] The page design evaluation method provided in the embodiment of the present application takes a screenshot of the store page to be evaluated to obtain a page screenshot, thereby obtaining a text vector based on the text information, text coordinates and image coordinates of the page screenshot. The text coordinates and image coordinates can be used to determine the position of the text information and the image in the page screenshot, which can improve the accuracy of subsequent evaluation of the page screenshot to a certain extent. In addition, the image model is introduced to extract the image features of the page screenshot, and the image vector is obtained using the first feature conversion module. Finally, the text vector and the image vector are simultaneously input into the causal language model to obtain the scoring result of the store page to be evaluated, which solves the technical problem that the UI design depends on personal experience or subjective judgment, provides a design reference for users, and speeds up the efficiency of UI design. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0017] Figure 1 It is a schematic diagram of the application environment of the page design evaluation method provided in the embodiment of the present application.

[0018] Figure 2 It is a flow chart of a page design evaluation method provided in one embodiment of the present application.

[0019] Figure 3 It is a processing diagram of the first feature conversion module provided in one embodiment of the present application.

[0020] Figure 4 This is a flowchart of a page design evaluation method provided by another embodiment of the present application.

[0021] Figure 5 It is a flowchart of a method for training a topic scoring model provided in one embodiment of the present application.

[0022] Figure 6 It is a processing diagram of the second feature conversion module provided in one embodiment of the present application.

[0023] Figure 7 This is a flowchart of a method for training a topic scoring model provided in yet another embodiment of the present application.

[0024] Figure 8 It is a schematic diagram of a page design evaluation device provided in an embodiment of the present application.

[0025] Fig. 9 It is a schematic diagram of a training device for a topic scoring model provided in an embodiment of the present application.

[0026] Fig.10 It is a schematic diagram of the structure of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0027] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application is described in detail below with reference to the accompanying drawings and specific embodiments.

[0028] It should be noted that in this application, "at least one" means one or more, and "more than one" means two or more than two. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone, where A and B can be singular or plural. The terms "first", "second", "third", "fourth", etc. (if any) in the specification, claims and drawings of this application are used to distinguish similar objects, rather than to describe a specific order or sequence.

[0029] In the embodiments of the present application, words such as "exemplary" or "for example" are used to indicate examples, illustrations or descriptions. Any embodiment or design described as "exemplary" or "for example" in the embodiments of the present application should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary" or "for example" is intended to present related concepts in a concrete way. The following embodiments and features in the embodiments may be combined with each other without conflict.

[0030] Cross-border e-commerce can be operated through independent websites, and users can purchase different products on different independent websites. The user interface (UI) design provided by stores running on independent websites is flexible and diverse, and is a key factor in measuring the personalized experience of stores. Since this interface design is too dependent on the personal experience and subjectivity of merchants or professional UI designers, and there is no unified evaluation standard, frequent design changes may occur, resulting in high labor costs and low interface design efficiency.

[0031] Merchants can convert independent websites into mobile applications (Application, App), and users can install the App on their mobile phones and visit the store by clicking the App icon. When the APP runs in the foreground of the electronic device, the electronic device can display the application interface of the APP on the display screen, that is, the store page of the independent website store. Users can interact with the APP through the controls in the application interface of the APP. In addition, merchants can also keep the independent website in the form of a website, and users can visit the store by entering the URL through the mobile browser or clicking on the bookmark.

[0032] Among them, the application interface refers to the medium interface for interaction and information exchange between the application or operating system and the user, which can realize the conversion between the internal form of the information and the form that the user can receive. The application interface is a source code written in a specific computer language such as Java and Extensible Markup Language (XML). The interface source code is parsed and rendered on the electronic device, and finally presented as content that the user can recognize. The common form of the application interface is the graphical application interface (Graphic User Interface, GUI), which refers to an application interface related to computer operation displayed in a graphical manner. It can be a visual interface element such as text, icon, button, menu, tab, text box, dialog box, status bar, navigation bar, widget, etc. displayed on the display screen of the electronic device. In order to solve the technical problem that UI design depends on personal experience or subjective judgment, the embodiment of the present application provides a page design evaluation method, a training method for a theme scoring model, a page design evaluation device and a computer device, which can improve the design efficiency of the store page and reduce the labor cost of designing the store page, avoiding dependence on personal experience or subjective judgment. The following first introduces the application scenario.

[0033] Figure 1 Schematic diagram of the application environment of the page design evaluation method provided in the embodiment of the present application. Figure 1 As shown, the page design evaluation method and the training method of the topic scoring model described in the present application are applied to an e-commerce platform 10, which can be used to provide merchants' products and services to consumers. Although the present disclosure considers the use of devices, systems, and processes to purchase products and services, for the sake of simplicity, the description herein will refer to products. Throughout the present disclosure, all references to products should also be understood as references to products and / or services, including physical products, digital content, tickets, subscriptions, services to be provided, and the like.

[0034] "Merchant" and "consumer" may be understood as users that support an e-commerce environment, and all references to merchants and consumers throughout this disclosure should also be understood as references to merchant users, prospective users, service provider users, corporate or enterprise users, information technology users, computing entity users, and the like.

[0035] The e-commerce platform 10 may provide a centralized system for providing merchants with online resources and facilities for managing their businesses. The facilities described herein may be deployed in part or in whole by a machine that executes computer software, modules, program codes, and / or instructions on one or more processors that may be part of the e-commerce platform 10 or external to the e-commerce platform 10. Merchants may utilize the e-commerce platform 10 to manage commerce with consumers, such as by interacting with consumers via communications 120 of the e-commerce platform 10, or any combination thereof.

[0036] Online store 110 may represent a multi-tenant facility including multiple virtual storefronts. In an embodiment, a merchant may manage one or more storefronts in online store 110, such as through merchant device 20 (eg, computer, laptop, mobile computing device, etc.).

[0037] In some embodiments, the e-commerce platform 10 may be implemented by a processing facility including a processor and a memory, the processing facility storing a set of instructions that, when executed, cause the e-commerce platform 10 to perform e-commerce and supporting functions as described herein. The processing facility may be part of a server, a consumer end, a network infrastructure, a mobile computing platform, a cloud computing platform, a fixed computing platform, or other computing platform, and provides electronic connections and communications between and among electronic components of the e-commerce platform 10, merchant devices 20, payment gateways, application developers, channels, transportation providers, consumer devices 30, point-of-sale devices, and the like. The e-commerce platform 10 can be implemented as a cloud computing service, software as a service (SaaS), infrastructure as a service (IaaS), platform as a service (PaaS), desktop as a service (DaaS), management software as a service (MSaaS), mobile backend as a service (MBaaS), information technology management as a service (ITMaaS), etc. In some embodiments, elements of the e-commerce platform 10 can be implemented to operate on various platforms and operating systems, such as iOS, Android, on the Web, etc. (for example, the administrator 130 is implemented in multiple instances of a given online store for iOS, Android, and for the Web, each instance having similar functionality).

[0038] In some embodiments, the online store 110 can serve the consumer device 30 through a web page provided by a server of the e-commerce platform 10. The server receives a request for the web page from a browser or other application installed on the consumer device 30, wherein the browser (or other application) is connected to the server via an IP address obtained by converting the domain name. In return, the server sends back the requested web page. The web page can be written in hypertext markup language (HTML), template language, JavaScript, etc., or any combination thereof, or the web page includes the above language. For example, HTML is a computer language that describes static information of a web page, such as the layout, format, and content of the web page. Website designers and developers can use template languages ​​to build web pages that combine static content (which is the same on multiple pages) and dynamic content (which changes from one page to the next). Template languages ​​can make it possible to reuse static elements that define the layout of a web page, while dynamically populating the page with data from an online store. Static elements can be written in HTML, and dynamic elements can be written in template languages. The template language elements in the file can act as placeholders so that the code in the file is compiled and sent to the consumer device 30, and then the template language is replaced by data from the online store 110, such as when the theme is installed. Templates and themes can take into account tags, objects, and filters. The consumer end device web browser (or other application) then displays the page accordingly.

[0039] In some embodiments, merchants can customize their online stores 110 using merchant-configurable domain names, customizable HTML themes, and the like. Merchants can customize the look and feel of their websites through a theme system, such as merchants can select and change the look and feel of their online stores 110 by changing their themes, while making the same basic products and business data be shown in the product hierarchy of the online store. Themes can be further customized through a theme editor, which is a design interface that enables users to flexibly customize the design of their websites. Themes can also be customized using theme-specific settings that can change aspects such as specific colors, fonts, and pre-built layout schemes. An online store can implement a content management system for website content. Merchants can write blog posts or static pages and publish them to their online stores 110, such as publishing through blogs, articles, and the like, and configuring navigation menus. Merchants can upload images (e.g., images of products), videos, content, data, and the like to the e-commerce platform 10, such as for the system to store (e.g., as data 140). In some embodiments, the e-commerce platform 10 may provide functionality for resizing images, associating images with products, adding text and associating text with images, adding images for new product variations, protecting images, and the like.

[0040] As described herein, the e-commerce platform 10 may provide transaction facilities for products to merchants, through multiple different channels including the online store 110, both over the phone and through the physical POS devices described herein. The e-commerce platform 10 may include business support services 150, administrators 130, etc. associated with running an online business. The business management engine 160 includes the basic or "core" functionality of the e-commerce platform 10, and as such, not all functionality supporting the online store 110, as described herein, may be suitable for inclusion.

[0041] The e-commerce platform 10 can help or assist merchants who are designing UI themes. The e-commerce platform 10, however, includes a theme scoring engine 170. The theme scoring engine 170 is an example of a computer-implemented system that obtains and analyzes images corresponding to user-designed UI themes and proactively provides suggestions to users to improve product image quality. In the example, when a merchant is designing a UI theme, the merchant device 20 sends the preliminarily designed store page to the theme scoring engine 170. The theme scoring engine 170 analyzes the store page to provide the merchant with a corresponding scoring result.

[0042] The topic scoring engine 170 includes a processor 171 and a memory 172. The processor 171 can be implemented by one or more processors executing instructions stored in the memory 172. Alternatively, some or all of the processors 171 can be implemented using dedicated circuits, such as an application-specific integrated circuit (ASIC), a graphics processing unit (GPU), or a programmed field programmable gate array (FPGA). The memory 172 stores a topic scoring model 1721, which can actually be a plurality of different topic scoring models, each of which is for a different type of topic. The memory 172 also stores a topic scoring model generator 1722, which facilitates the generation of the topic scoring model 1721. Once the topic scoring model 1721 is generated, the topic scoring model 1721 can be updated periodically, for example, when market trends change and / or when new data becomes available. The topic scoring model generator 1722 can also facilitate the update.

[0043] In the case where the topic scoring model 1721 is a machine learning algorithm or includes a machine learning algorithm, the topic scoring model generator 1722 includes a training algorithm to train the machine learning algorithm. The training algorithm may involve, for example, supervised learning, unsupervised learning, or reinforcement learning. The training algorithm may be stored as instructions executed by the processor 171.

[0044] The memory 172 stores image data 1724 used by the topic scoring model generator 1722 to train the machine learning algorithm. The image data 1724 is obtained from the store page of the existing online store 110 in the e-commerce platform 10. These existing store pages may include the homepage of the store, the homepage of the product, and the detail page.

[0045] In some cases, the image data 1724 is divided into a training data set and a test data set. The topic scoring model generator 1722 uses the training data to train the machine learning algorithm that implements the topic scoring model 1721. The test data set is used to test the accuracy of the machine learning algorithm after training.

[0046] The memory 172 also includes an image analyzer 1723. The image analyzer 1723 receives a store page and determines parameters of the store page, such as text information, text coordinates, and image coordinates. In some cases, the image analyzer 1723 is used to determine parameters of an existing store page for storage in the image data 1724. The image analyzer 1723 may also or alternatively be used to determine a store page received from a merchant and then input it into the topic scoring model 1721.

[0047] It is to be understood that the application scenarios illustrated in the embodiments of the present application do not constitute specific limitations. In other embodiments of the present application, the e-commerce platform 10 may also include more or fewer components, such as a network interface, a domain, a payment gateway, and the like.

[0048] Figure 2 is a flow chart of a page design evaluation method provided by an embodiment of the present application. Figure 2 As shown, the page design evaluation method provided in the embodiment of the present application is applied to an e-commerce platform (such as Figure 1 According to different requirements, the order of the steps in the flowchart can be changed, and some steps can be omitted. Figure 2 As shown, the following steps are included:

[0049] Step S201, obtaining a page screenshot corresponding to the store page to be evaluated.

[0050] In some embodiments of the present application, the store to be evaluated may be a store that needs to undergo page evaluation after UI theme design. The store may run on an independent station, which is a website that is independently built and operated. Merchants can log in to the website of the independent station or the corresponding application through an electronic device to enter the store. In the store, merchants can freely design the UI theme, product display, etc. of the store page. Among them, the store page may be the store homepage of the store, the display page of any product, and the details page of any product. The present application does not limit the store page, and it may be any page in a store running on an independent station.

[0051] In order to evaluate the UI theme design of the store page, a screenshot of the store page can be obtained by taking a screenshot of the store page. The method of obtaining the page screenshot includes but is not limited to calling the screenshot tool provided by the electronic device to capture the store page, calling a third-party screenshot tool to capture the store page, and using automated programming or scripts to capture the screenshot of the store page.

[0052] Step S202: obtaining a text vector based on the text information, the text coordinates and the image coordinates.

[0053] In some embodiments of the present application, the store page usually includes many interface elements, such as text, images, etc. Since the design of the store page has a certain logic, an image coordinate system can be established based on the page screenshot of the store page to obtain text coordinates and image coordinates, wherein the text coordinates represent the position of the text information in the page screenshot, and the text information may include a paragraph or a word. The image coordinates represent the position of the image in the page screenshot. In addition, there is no limit on the amount of text information and images in the page screenshot.

[0054] In some embodiments of the present application, after obtaining text information, text coordinates and image coordinates from the page screenshot, the text information, text coordinates and image coordinates are combined into a sentence information. The sentence information can be tokenized to obtain a text vector, wherein tokenization represents the process of segmenting a continuous text string into discrete words or symbols.

[0055] The sentence information is tokenized to obtain a text vector, including: processing the sentence information text, removing irrelevant characters, special symbols, etc. After the text processing, the sentence information is segmented, and the sentence information can be divided based on a preset segmentation strategy to obtain multiple word units, where the segmentation strategy includes word segmentation, character segmentation, etc. After obtaining the word unit, each word unit is vectorized and converted into a vector representation of a fixed dimension to obtain the word vector corresponding to each word unit. Since the word vector can capture the semantic relationship between word units, each word vector can be spliced ​​based on the semantic relationship to obtain a text vector.

[0056] Step S203: obtaining image features of the page screenshot based on the page screenshot and a preset image model.

[0057] In some embodiments of the present application, the image model may be any image processing model that has been trained to meet the accuracy requirements, and the image model may be used as an image feature extractor. The image model may include but is not limited to the vit_base_patch14_reg4_dinov2 model and the ResNext101 model.

[0058] Among them, the vit_base_patch14_reg4_dinov2 model is a visual model that combines ViT and DINOv2. It uses a 14x14 patch size for image segmentation and combines DINOv2's self-supervised learning method to enhance feature extraction capabilities. ViT (Vision Transformer) is an image processing model based on the Transformer architecture. It divides the image into small blocks and inputs them into a fully connected network to obtain an embedding vector, which is then processed by the Transformer encoder. DINOv2 is a large visual model that aims to learn robust visual features without supervision. DINOv2 mainly extracts useful image features from a large amount of unlabeled data through self-supervised learning, thereby improving the generalization ability and performance of the model. The ResNext101 model is a deep residual network model that combines the advantages of ResNet and Xception, and improves the performance and efficiency of the model by introducing grouped convolution and multi-branch structures. ResNet is a deep learning model that mainly implements residual learning by introducing skip connections, allowing the network to learn the residual between input and output instead of directly learning the mapping from input to output. Xception is a deep convolutional neural network architecture that consists of 36 convolutional layers that form the basis of network feature extraction. Each layer is followed by a logistic regression layer, and a fully connected layer may be randomly inserted before the logistic regression layer.

[0059] In some embodiments of the present application, a page screenshot is input into an image model for feature extraction to obtain image features of the page screenshot, including: resizing, normalizing, and other preprocessing of the page screenshot to obtain a preprocessed image. The preprocessed image is segmented using the ViT model to obtain a plurality of image blocks. Each image block is input into a linear layer, in which each image block is converted into an embedding vector of a fixed length, and a position code is added to the embedding vector corresponding to each image block, which is used to describe the relative position relationship between different image blocks. The embedding vector carrying the position code is input into a Transformer encoder for processing. The relationship between the embedding vectors carrying the position code is captured by the Transformer encoder's multi-head self-attention mechanism and feed-forward neural networks to obtain the embedding vector output by the Transformer encoder, and the embedding vector output by the Transformer encoder is used as the image feature of the page screenshot.

[0060] Step S204: using a first feature conversion module to convert the image features to obtain an image vector.

[0061] In some embodiments of the present application, the first feature conversion module may be an adapter module for converting image features into image vectors, wherein the structure of the adapter module includes multiple linear layers, for example, a first linear layer, a second linear layer, a third linear layer, and a fourth linear layer.

[0062] For ease of understanding, the following Figure 3 The structure of the adapter module shown in Figure 1 describes the process of converting image features into image vectors.

[0063] The image features are input into the first linear layer of the first feature conversion module, such as Figure 3 As shown, the number of features of the image feature is L, and the dimension of each feature is H. The weight matrix Wq is used in the first linear layer to map the dimension of the image feature from H to 16m, where 16m means that the image features after linear transformation are divided into 16 groups, each with m features, and m can be set to 256, so 16m is 4096. Then, the intermediate result output by the first linear layer is subjected to a dimension conversion operation (such as a reshape operation) to readjust the dimension, and after the dimension is adjusted, a normalization (LayerNorm) operation is performed to obtain the first feature (size is 16×L×m). Reshape is a function that can readjust the number of rows, columns, and dimensions of a matrix.

[0064] The image features are input into the second linear layer of the first feature conversion module, such as Figure 3 As shown, the dimension of the image feature is mapped from H dimension to 16m in the second linear layer through the weight matrix Wk. Then, the intermediate result output by the second linear layer is subjected to a dimension conversion operation to re-adjust the dimension, and a normalization operation is performed after the dimension adjustment to obtain the second feature (with a size of 16×L×m).

[0065] The image features are input into the third linear layer of the first feature conversion module, such as Figure 3 As shown, the dimension of the image feature is mapped from H dimension to 16m in the third linear layer through the weight matrix Wv. Then, the intermediate result output by the third linear layer is subjected to a dimension conversion operation to re-adjust the dimension, and a normalization operation is performed after the dimension adjustment to obtain the third feature (with a size of 16×L×m).

[0066] The image features are input into the fourth linear layer of the first feature conversion module, such as Figure 3 As shown, in the fourth linear layer, the dimension of the image feature is mapped from H dimension to 16m, thereby obtaining the fourth feature.

[0067] like Figure 3 As shown, the first feature is processed by using the activation function elu(x) plus the constant 1 to obtain the first activation feature. The second feature is processed by using the activation function elu(x) plus the constant 1 to obtain the second activation feature.

[0068] like Figure 3 As shown, the third feature, the first activation feature and the second activation feature are similarly calculated using the attention score calculation formula (LA formula) to obtain the fifth feature. Among them, the LA formula is used to determine the similarity between each query (Qa) and all keys (Ka), so as to calculate the weight of the value (V) corresponding to each query. In an embodiment of the present application, Qa is the first activation feature, Ka is the second activation feature, and V is the third feature. Qa, Ka and V are all matrices. The LA formula is expressed as follows:

[0069]

[0070] Where, the i-th LA formula is used to query the j-th key and value.

[0071] A dimension conversion operation is performed on the fifth feature. The fifth feature after the dimension conversion operation is added to the fourth feature and normalized to obtain a sixth feature.

[0072] like Figure 3As shown, the sixth feature is input into the fifth linear layer of the first feature conversion module, the dimension of the sixth feature is mapped from 16m to 64m (64m is 16384), and then input into the sixth linear layer, the dimension is mapped from 64m to 16m. The seventh feature output by the sixth linear layer is obtained.

[0073] like Figure 3 As shown, the seventh feature and the sixth feature are normalized to obtain an image vector with a size of L×16m.

[0074] Step S205: obtaining a scoring result for the store page to be evaluated based on the text vector, the image vector and a preset causal language model.

[0075] In some embodiments of the present application, the causal language model may be a causal language model similar to the GPT structure, for example, the Vicuna-7b-v1.5-16k model. The text vector and the image vector are merged into a multimodal vector, and the multimodal vector is used as the input of the causal language model to obtain the scoring result of the store page to be evaluated.

[0076] In the above embodiment, a screenshot of the store page to be evaluated is taken to obtain a page screenshot, thereby obtaining a text vector based on the text information, text coordinates and image coordinates of the page screenshot. The text coordinates and image coordinates can be used to determine the position of the text information and the image in the page screenshot, which can improve the accuracy of subsequent evaluation of the page screenshot to a certain extent. In addition, the image model is introduced to extract the image features of the page screenshot, and the image vector is obtained using the first feature conversion module. Finally, the text vector and the image vector are simultaneously input into the causal language model to obtain the scoring result of the store page to be evaluated, which solves the technical problem that the UI design depends on personal experience or subjective judgment, provides a design reference for users, and speeds up the efficiency of UI design.

[0077] To better understand the Figure 2 The process shown below is combined with Figure 4 The process framework diagram shown is further explained.

[0078] like Figure 4 As shown, the text information, text coordinates and image coordinates are tokenized to obtain a text vector. The page screenshot is processed using the image model to obtain image features. The image features are processed by the first feature conversion module to obtain an image vector. The text vector and the image vector are merged, and the merged text vector and image vector are input into the causal language model to obtain a scoring result.

[0079] Figure 5is a flowchart of a method for training a topic scoring model provided in an embodiment of the present application. Figure 5 As shown, the topic scoring model includes an image model, a first feature conversion module, a second feature conversion module and a causal language model, wherein the second feature conversion module is as shown in Figure 5 After the topic scoring model training is completed, the second feature conversion module may not participate in the training process shown in FIG. Figure 2 The training method of the topic scoring model is applied to a computer device (such as Figure 1 According to different requirements, the order of the steps in the flowchart can be changed, and some steps can be omitted. Figure 5 As shown, the following steps are included:

[0080] Step S501, obtaining screenshots of store pages corresponding to multiple stores as training data.

[0081] In some embodiments of the present application, multiple stores may be stores running on an independent station. In order to collect rich training data, screenshots of store pages corresponding to different types of stores may be obtained. The store page may be the store homepage of the store, the display page of any product, and the details page of any product. The present application does not limit the store page, which may be any page in a store running on an independent station. The method of obtaining a page screenshot includes but is not limited to calling a screenshot tool provided by an electronic device to capture a store page, calling a third-party screenshot tool to capture a store page, and using automated programming or scripts to capture a screenshot of a store page.

[0082] In some embodiments of the present application, each page screenshot includes multiple sub-images. An image coordinate system is established for the page screenshot, and the image coordinates of all sub-images in the page screenshot are determined. Each sub-image in the page screenshot is intercepted based on the image coordinates, and an image set is constructed based on each sub-image. In order to ensure the quality of the training data and improve the training accuracy of the topic scoring model, the image set can be preprocessed, including:

[0083] If the number of sub-images in the image set is equal to a preset number (eg, 20 images), the image set is used as training data.

[0084] If the number of sub-images in the image set is greater than the preset number, the image set is screened. Based on the image coordinates corresponding to each sub-image, the midpoint coordinates of the page screenshot are determined. Sub-images are screened starting from the midpoint coordinates, and based on the image coordinates of each sub-image, multiple neighbor sub-images that have an adjacent relationship with the midpoint coordinates and whose image size is greater than or equal to the preset size are selected. When the number of multiple neighbor sub-images is equal to the preset number, the multiple neighbor sub-images are used as training data. When the number of multiple neighbor sub-images is less than the preset number, multiple neighbor sub-images that have an adjacent relationship with the midpoint coordinates and whose image size is less than the preset size are selected based on the image coordinates of each sub-image, and when the number of multiple neighbor sub-images is equal to the preset number, the multiple neighbor sub-images are used as training data.

[0085] If the number of sub-images in the image set is less than a preset number, a corresponding blank image is supplemented according to the number of sub-images to be supplemented, and the pixel value of the blank image can be 0 or 255.

[0086] Step S502: obtaining image set features based on the training data and the image model.

[0087] In the embodiments of the present application, the image model may include but is not limited to the vit_base_patch14_reg4_dinov2 model, the ResNext101 model, and the like. Figure 2 The description of the vit_base_patch14_reg4_dinov2 model and the ResNext101 model in step S203 of the illustrated embodiment will not be repeated here.

[0088] Input the training data into the image model to obtain the image set features, including: resizing and normalizing each sub-image in the training data to obtain the pre-processed image. Use the ViT model to segment the pre-processed image to obtain multiple image blocks. Input each image block into the linear layer (embedding layer), convert each image block into an embedding vector (embedding vectors) of fixed length in the linear layer, add a position code to the embedding vector corresponding to each image block, and the position code is used to describe the relative position relationship between different image blocks. Input the embedding vector carrying the position code into the Transformer encoder for processing. The relationship between the embedding vectors carrying the position code is captured by the multi-head self-attention mechanism (multi-head self-attention) and feed-forward neural networks of the Transformer encoder to obtain the embedding vector output by the Transformer encoder, and the embedding vector output by the Transformer encoder is used as the image feature of each sub-image, and the image set feature is obtained based on the image feature of each sub-image.

[0089] Step S503: obtaining a first prediction value based on the page screenshot, the first feature conversion module and the second feature conversion module.

[0090] In some embodiments of the present application, the processing process and structure of the first conversion module can refer to Figure 3 As shown, the description is not repeated here. The second feature conversion module can be an inv-adapter module. The image features of the page screenshot are extracted using the image model. The image features are processed using the first feature conversion module, and the following is executed cyclically: Figure 3 The processing process shown is repeated until the number of iterations reaches a preset number (eg, 3 times) to obtain an image vector. The image vector is input into the second feature conversion module for processing to obtain a first prediction value.

[0091] Among them, the structure of the inv-adapter module is similar to that of the adapter module. The structure of the inv-adapter module has one more linear layer than that of the adapter module (such as Figure 6 For ease of understanding, let's combine Figure 6 The process of inputting the image vector into the second feature conversion module for processing to obtain the first prediction value is described.

[0092] The image features are input into the first linear layer of the second feature conversion module, such as Figure 6 As shown, the number of features of the image feature is L, and the dimension of each feature is H. The dimension of the image feature is mapped from H to 32m in the first linear layer through the weight matrix Wq, where m can be set to 256, then 32m is 8192. Then, the intermediate result output by the first linear layer is subjected to a dimension conversion operation (such as a reshape operation) to readjust the dimension, and after the dimension is adjusted, a normalization (LayerNorm) operation is performed to obtain the first feature (size is 32×L×m). Reshape is a function that can readjust the number of rows, columns, and dimensions of a matrix.

[0093] The image features are input into the second linear layer of the second feature conversion module, such as Figure 6 As shown, the dimension of the image feature is mapped from H dimension to 32m in the second linear layer through the weight matrix Wk. Then, the intermediate result output by the second linear layer is subjected to a dimension conversion operation to re-adjust the dimension, and a normalization operation is performed after the dimension adjustment to obtain the second feature (with a size of 32×L×m).

[0094] The image features are input into the third linear layer of the second feature conversion module, such as Figure 6As shown, the dimension of the image feature is mapped from H dimension to 32m in the third linear layer through the weight matrix Wv. Then, the intermediate result output by the third linear layer is subjected to a dimension conversion operation to re-adjust the dimension, and a normalization operation is performed after the dimension adjustment to obtain the third feature (with a size of 32×L×m).

[0095] The image features are input into the fourth linear layer of the second feature conversion module, such as Figure 6 As shown, in the fourth linear layer, the dimension of the image feature is mapped from H dimension to 32m, thereby obtaining the fourth feature.

[0096] like Figure 6 As shown, the first feature is processed by using the activation function elu(x) plus the constant 1 to obtain the first activation feature. The second feature is processed by using the activation function elu(x) plus the constant 1 to obtain the second activation feature.

[0097] like Figure 6 As shown in FIG. 1 , the third feature, the first activation feature, and the second activation feature are similarly calculated using the attention score calculation formula (LA formula), thereby obtaining the fifth feature. Figure 2 Step S204 in the illustrated embodiment.

[0098] A dimension conversion operation is performed on the fifth feature and the feature is input into the fifth linear layer. In the fifth linear layer, the dimension of the fifth feature on which the dimension conversion operation is performed is mapped from 32m to 60m, thereby obtaining the sixth feature output by the fifth linear layer.

[0099] The sixth feature is added to the fourth feature and a normalization operation is performed to obtain the seventh feature.

[0100] like Figure 6 As shown, the seventh feature is input into the sixth linear layer of the second feature conversion module, the dimension of the seventh feature is mapped from 60m to 64m, and then input into the seventh linear layer, the dimension is mapped from 64m to 60m. The eighth feature output by the seventh linear layer is obtained.

[0101] like Figure 6 As shown, the eighth feature and the seventh feature are normalized to obtain a first prediction value with a size of L×60m.

[0102] In the above embodiment, the inv-adapter module is introduced for training, which can ensure that the image vector and the subsequently obtained text vector are aligned in the same space.

[0103] Step S504: obtaining a second prediction value based on the page screenshot, the image model, the first feature conversion module, and the causal language model.

[0104] In some embodiments of the present application, the text information in the page screenshot, the text coordinates corresponding to the text information, and the image coordinates are extracted, and the text information, the text coordinates, and the image coordinates are combined into a sentence information. The sentence information can be tokenized to obtain a text vector. The image model is used to extract the features of the page screenshot to obtain the image features of the page screenshot. The image features are converted using the first feature conversion module to obtain an image vector. Based on the text vector and the image vector, a causal language model is used to obtain a second prediction value for the rating result of each store. The specific description of step S504 can be found in Figure 2 Steps S202 to S205 in the illustrated embodiment are not described again herein.

[0105] Step S505: Calculate a first loss value based on the image set features and the first prediction value.

[0106] In some embodiments of the present application, the image set feature is used as a sample label, and the difference between the sample label and the first prediction value is calculated to obtain a first loss value.

[0107] Step S506: Calculate a second loss value based on the second predicted value and the actual scoring result.

[0108] In some embodiments of the present application, obtaining the actual scoring result includes: determining a target store from multiple stores, the target store can be any one of the multiple stores, and the present application does not limit this. Counting the number of independent visitors to the store page of the target store, that is, the number of different users who visit the store page. Counting the transaction volume after visiting the store page. Determine the traffic score based on the number of independent visitors and the transaction volume after visiting the store page. Expressed as follows:

[0109]

[0110] In the formula, c4 represents the traffic score, uv represents the number of independent visitors, and cns represents the transaction volume after visiting the store page.

[0111] In some embodiments of the present application, the daily transaction volume of each of the multiple stores is counted, and the first store with the largest daily transaction volume and the second store with the smallest daily transaction volume are determined from the multiple stores based on the daily transaction volume of each store. The transaction ratio is determined based on the traffic score and the daily transaction volume of the target store.

[0112] Based on the daily transaction volume of the target store, the daily transaction volume of the first store, the daily transaction volume of the second store, and the transaction ratio, the transaction score is obtained, which is expressed by the following formula:

[0113]

[0114] gns=1 / (1+e -gn );

[0115] In the formula, net_amount represents the daily transaction volume of the target store, min_net_amount represents the daily transaction volume corresponding to the second store, sp represents the transaction ratio, max_net_amount represents the daily transaction volume corresponding to the first store, and gns represents the transaction score.

[0116] After the transaction score is determined, the actual score result can be obtained according to the traffic score and the transaction score. The actual score result is used as the sample label, and the second loss value is calculated based on the second predicted value and the actual score result.

[0117] Step S507: adjust the model parameters of the topic scoring model based on the first loss value and the second loss value to obtain a trained topic scoring model.

[0118] In some embodiments of the present application, the first loss value and the second loss value are weighted and summed to obtain a total loss value. If the total loss value is less than a preset loss threshold, it is determined that the training of the topic scoring model is completed. If the total loss value is greater than or equal to the preset loss threshold, the model parameters of the topic scoring model can be adjusted through back propagation until the total loss value is less than the preset loss threshold.

[0119] In the above embodiment, the introduction of the inv-adapter module in the training process of the topic scoring model can improve the performance and generalization ability of the topic scoring model, and help improve the training efficiency and convergence speed of the topic scoring model.

[0120] To better understand the Figure 5 The process shown below is combined with Figure 7 The process framework diagram shown is further explained, where: Figure 7 and Figure 4 Some of the processing procedures are the same.

[0121] like Figure 7 As shown, the double arrows indicate that during training, there will be a process of forward reasoning and back-propagation of gradients, in which the back-propagation gradients will automatically update the corresponding model parameters. Specifically, the page screenshot is processed to obtain training data, and the training data is input into the image model to obtain image set features. The page screenshot is input into the image model to obtain image features. The image features are input into the first feature conversion module to obtain an image vector, and the image vector is input into the second feature conversion module to obtain a first prediction value. The first loss value is calculated based on the image features and the first prediction value. Figure 7The flow direction indicated by the bold line shown in the figure introduces the second feature conversion module, which can ensure that the image vector and the text vector are aligned in the same space and that the first prediction value has the original image features of the page screenshot.

[0122] Tokenize the image coordinates, text information, and text coordinates in the page screenshot to obtain a text vector. Merge the image vector and the text vector and input them into the causal language model to obtain a second prediction value. Calculate the second loss value based on the second prediction value and the actual scoring result.

[0123] The model parameters of the topic scoring model are calculated and adjusted based on the first loss value and the second loss value to obtain a trained topic scoring model.

[0124] See also Figure 8 , is a schematic diagram of a page design evaluation device provided in an embodiment of the present application. It is a functional embodiment of the page design evaluation method of the present application, and the page design evaluation device includes an acquisition module 810, a text module 820, an image module 830, a feature module 840 and a scoring module 850, wherein: the acquisition module 810 is used to acquire a page screenshot corresponding to the store page to be evaluated, and the page screenshot includes text information and an image, text coordinates corresponding to the text information, and image coordinates corresponding to the image; the text module 820 is used to obtain a text vector based on the text information, the text coordinates and the image coordinates; the image module 830 is used to obtain the image features of the page screenshot based on the page screenshot and a preset image model; the feature module 840 is used to convert the image features using a first feature conversion module to obtain an image vector; the scoring module 850 is used to obtain a scoring result of the store page to be evaluated based on the text vector, the image vector and a preset causal language model.

[0125] Based on any embodiment of the present application, the text module 820 is also used to: combine the text information, the text coordinates and the image coordinates to obtain sentence information; perform text processing and word segmentation processing on the sentence information to obtain multiple word units; perform vector conversion on each word unit to obtain the word vector corresponding to each word unit; and concatenate each word vector to obtain the text vector.

[0126] Based on any embodiment of the present application, the feature module 840 is further used to: process the image feature according to the first linear layer, dimensional conversion operation and normalization operation of the first feature conversion module to obtain a first feature; process the image feature according to the second linear layer, the dimensional conversion operation and the normalization operation of the first feature conversion module to obtain a second feature; process the image feature according to the third linear layer, the dimensional conversion operation and the normalization operation of the first feature conversion module to obtain a third feature; and, process the image feature using the fourth linear layer of the first feature conversion module to obtain a fourth feature; process the first feature and the second feature using an activation function to obtain a first activation feature corresponding to the first feature and a second activation feature corresponding to the second feature; calculate the similarity of the third feature, the first activation feature and the second activation feature based on the attention score calculation formula to obtain a fifth feature; obtain a sixth feature based on the fifth feature and the fourth feature; input the sixth feature into the fifth linear layer and the sixth linear layer of the first feature conversion module in sequence to obtain a seventh feature output by the sixth linear layer; normalize the sum of the seventh feature and the sixth feature to obtain the image vector.

[0127] See also Fig. 9 , is a schematic diagram of a training device for a topic scoring model provided in an embodiment of the present application. It is a functional embodiment of a training method for a topic scoring model of the present application. The training device for the topic scoring model includes a determination module 910, an image module 920, a prediction module 930, a calculation module 940, and an evaluation module 950, wherein: the determination module 910 is used to determine training data based on page screenshots of store pages corresponding to multiple stores; the image module 920 is used to obtain image set features based on the training data and the image model; the prediction module 930 is used to obtain a first prediction value based on the page screenshot, the first feature conversion module, and the second feature conversion module; the prediction module 930 is also used to obtain a first prediction value based on the page screenshot, the first feature conversion module, and the second feature conversion module. screenshot, the image model, the first feature conversion module and the causal language model to obtain a second prediction value; a calculation module 940 is used to calculate a first loss value based on the image set features and the first prediction value; the calculation module 940 is also used to calculate a second loss value based on the second prediction value and the actual scoring result; an evaluation module 950 is used to adjust the model parameters of the topic scoring model based on the first loss value and the second loss value to obtain a trained topic scoring model, wherein the topic scoring model includes the image model, the first feature conversion module, the second feature conversion module and the causal language model.

[0128] Based on any embodiment of the present application, the determination module 910 is further used to: extract the sub-image in the page screenshot based on the image coordinates corresponding to the sub-image in the page screenshot to generate an image set; determine the training data based on the image set and the number of sub-images in each page screenshot in the image set.

[0129] Based on any embodiment of the present application, the prediction module 930 is also used to: based on the page screenshot, use the image model to obtain the image features of the page screenshot; use the first feature conversion module to convert the image features to obtain an image vector; use the second feature conversion module to process the image vector to obtain a first prediction value.

[0130] Based on any embodiment of the present application, the prediction module 930 is also used to: obtain a text vector based on the text information of the page screenshot, the text coordinates corresponding to the text information, and the image coordinates; use the image model to extract the features of the page screenshot to obtain the image features of the page screenshot; use the first feature conversion module to convert the image features to obtain an image vector; based on the text vector and the image vector, use the causal language model to obtain the second prediction value of the rating result for each store.

[0131] Based on any embodiment of the present application, the actual scoring result is obtained, including: determining a target store from the multiple stores; determining a traffic score based on the number of independent visitors to the store page of the target store and the transaction volume after visiting the store page; obtaining the daily transaction volume of each store, and determining the first store with the largest daily transaction volume and the second store with the smallest daily transaction volume from the multiple stores based on the daily transaction volume of each store; determining a transaction ratio based on the traffic score and the daily transaction volume of the target store; obtaining a transaction score based on the daily transaction volume of the target store, the daily transaction volume corresponding to the first store, the daily transaction volume corresponding to the second store, and the transaction ratio; obtaining the actual scoring result based on the traffic score and the transaction score.

[0132] Another embodiment of the present application further provides a computer device. Fig.10 is a schematic diagram of the structure of a computer device provided in an embodiment of the present application, such as Fig.10 As shown, in one embodiment of the present application, the computer device 11 may be a tablet computer, an augmented reality (AR) / virtual reality (VR) device, a laptop computer, a netbook, a server cluster, a large server, etc. The present embodiment of the application does not impose any restrictions on the specific type of the computer device 11. The computer device 11 is used to carry Figure 1The e-commerce platform 10 is shown.

[0133] like Fig.10 As shown, the computer device 11 may include, but is not limited to, a communication module 1101, a memory 1102, a processor 1103, an input / output (I / O) interface 1104, and a bus 1105. The processor 1103 is coupled to the communication module 1101, the memory 1102, and the I / O interface 1104 through the bus 1105.

[0134] Those skilled in the art will appreciate that the schematic diagram is merely an example of the computer device 11 and does not constitute a limitation on the computer device 11. The computer device 11 may include more or fewer components than shown in the diagram, or a combination of certain components, or different components. For example, the computer device 11 may also include a network access device, etc.

[0135] The communication module 1101 may include a wired communication module and / or a wireless communication module. The wired communication module may provide one or more wired communication solutions such as Universal Serial Bus (USB), Controller Area Network (CAN), etc. The wireless communication module may provide one or more wireless communication solutions such as Wireless Fidelity (Wi-Fi), Bluetooth (BT), mobile communication network, Frequency Modulation (FM), near field communication (NFC), infrared (IR), etc.

[0136] The memory 1102 can be used to store computer-readable instructions and / or modules. The processor 1103 implements various functions of the computer device 11 by running or executing the computer-readable instructions and / or modules stored in the memory 1102 and calling the data stored in the memory 1102. The memory 1102 can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, an application required for at least one function (such as a sound playback function, an image playback function, etc.), etc.; the data storage area can store data created according to the use of the computer device 11, etc. The memory 1102 can include non-volatile and volatile memories, such as: a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other storage devices.

[0137] The memory 1102 may be an external memory and / or an internal memory of the computer device 11. Furthermore, the memory 1102 may be a memory in a physical form, such as a memory stick, a TF card (Trans-flash Card), and the like.

[0138] The processor 1103 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc. The processor 1103 is the computing core and control center of the computer device 11, and uses various interfaces and lines to connect various parts of the entire computer device 11, and execute the operating system of the computer device 11 and various installed application programs, program codes, etc.

[0139] Exemplarily, the computer-readable instructions may be divided into one or more modules / sub-modules / units, one or more modules / sub-modules / units are stored in the memory 1102, and executed by the processor 1103 to complete the present application. One or more modules / sub-modules / units may be a series of computer-readable instruction segments capable of completing a specific function, and the computer-readable instruction segments are used to describe the execution process of the computer-readable instructions in the computer device 11.

[0140] If the module / unit integrated in the computer device 11 is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present application implements all or part of the processes in the above-mentioned embodiment method, and can also instruct the relevant hardware to complete it through computer-readable instructions. The computer-readable instructions can be stored in a computer-readable storage medium, and when the computer-readable instructions are executed by the processor, the steps of the above-mentioned method embodiments can be implemented.

[0141] The computer-readable instructions include computer-readable instruction codes, which may be in source code form, object code form, executable files, or some intermediate form, etc. The computer-readable medium may include: any entity or device capable of carrying the computer-readable instruction codes, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM).

[0142] The memory 1102 in the computer device 11 stores computer-readable instructions. The specific implementation method of the processor 1103 for the above-mentioned computer-readable instructions can be referred to Figure 2 , Figure 4 , Figure 6 as well as Figure 8 The description of the relevant steps in the corresponding embodiments will not be repeated here.

[0143] The I / O interface 1104 is used to provide a channel for user input or output. For example, the I / O interface 1104 can be used to connect various input and output devices, such as a mouse, keyboard, touch device, display screen, etc., so that the user can enter information or visualize information.

[0144] The bus 1105 is at least used to provide a channel for mutual communication among the communication module 1101 , the memory 1102 , the processor 1103 , and the I / O interface 1104 in the computer device 11 .

[0145] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of modules is only a logical function division, and there may be other division methods in actual implementation.

[0146] The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0147] In addition, each functional module in each embodiment of the present application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The above integrated unit can be implemented in the form of hardware or in the form of hardware plus software functional modules.

[0148] Therefore, no matter from which point of view, the embodiments should be regarded as illustrative and non-restrictive, and the scope of the present application is limited by the appended claims rather than the above description, so it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present application. Any attached figure mark in the claims should not be regarded as limiting the claims involved.

[0149] In addition, it is clear that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or devices can also be implemented by one unit or device through software or hardware. The words first, second, etc. are used to indicate names, and do not indicate any particular order.

[0150] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present application and are not intended to limit it. Although the present application has been described in detail with reference to the preferred embodiments, a person of ordinary skill in the art should understand that the technical solution of the present application may be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present application.

Claims

1. A page design evaluation method, characterized in that: include: Obtain a page screenshot corresponding to the store page to be evaluated, wherein the page screenshot includes text information and an image, text coordinates corresponding to the text information, and image coordinates corresponding to the image; Obtaining a text vector based on the text information, the text coordinates and the image coordinates; Based on the page screenshot and a preset image model, obtaining image features of the page screenshot; Using a first feature conversion module to convert the image features to obtain an image vector; Based on the text vector, the image vector and a preset causal language model, a scoring result for the store page to be evaluated is obtained.

2. The page design evaluation method according to claim 1, characterized in that: The obtaining of a text vector based on the text information, the text coordinates and the image coordinates includes: Combining the text information, the text coordinates and the image coordinates to obtain sentence information; Performing text processing and word segmentation processing on the sentence information to obtain multiple word units; Performing vector conversion on each word unit to obtain the word vector corresponding to each word unit; Each word vector is concatenated to obtain the text vector.

3. The page design evaluation method according to claim 1, characterized in that: The step of converting the image features using the first feature conversion module to obtain an image vector includes: The image features are processed according to the first linear layer, the dimension conversion operation, and the normalization operation of the first feature conversion module to obtain a first feature; the image features are processed according to the second linear layer, the dimension conversion operation, and the normalization operation of the first feature conversion module to obtain a second feature; the image features are processed according to the third linear layer, the dimension conversion operation, and the normalization operation of the first feature conversion module to obtain a third feature; and the image features are processed using the fourth linear layer of the first feature conversion module to obtain a fourth feature; Processing the first feature and the second feature using an activation function to obtain a first activation feature corresponding to the first feature and a second activation feature corresponding to the second feature; Performing similarity calculation on the third feature, the first activation feature, and the second activation feature based on an attention score calculation formula to obtain a fifth feature; Based on the fifth feature and the fourth feature, a sixth feature is obtained; Inputting the sixth feature into the fifth linear layer and the sixth linear layer of the first feature conversion module in sequence to obtain a seventh feature output by the sixth linear layer; The sum of the seventh feature and the sixth feature is normalized to obtain the image vector.

4. A method for training a topic scoring model, characterized in that: The training method comprises: Determine training data based on screenshots of store pages corresponding to multiple stores; Based on the training data and the image model, obtaining image set features; Based on the page screenshot, the first feature conversion module and the second feature conversion module, a first prediction value is obtained; Obtaining a second prediction value based on the page screenshot, the image model, the first feature conversion module, and the causal language model; Calculating a first loss value based on the image set feature and the first prediction value; Calculating a second loss value based on the second predicted value and the actual scoring result; Based on the first loss value and the second loss value, the model parameters of the topic scoring model are adjusted to obtain a trained topic scoring model, wherein the topic scoring model includes the image model, the first feature conversion module, the second feature conversion module and the causal language model.

5. The training method of the topic scoring model according to claim 4, characterized in that: The determining of training data based on page screenshots of store pages corresponding to the plurality of stores includes: Extracting the sub-images in the page screenshot based on the image coordinates corresponding to the sub-images in the page screenshot to generate an image set; The training data is determined based on the image set and the number of sub-images of each page screenshot in the image set.

6. The training method of the topic scoring model according to claim 4, characterized in that: The obtaining of a first prediction value based on the page screenshot, the first feature conversion module and the second feature conversion module includes: Based on the page screenshot, obtaining image features of the page screenshot using the image model; Using a first feature conversion module to convert the image features to obtain an image vector; The image vector is processed by using the second feature conversion module to obtain a first prediction value.

7. The method for training a topic scoring model according to claim 4, characterized in that: The obtaining a second prediction value based on the page screenshot, the image model, the first feature conversion module, and the causal language model includes: Obtaining a text vector based on the text information of the page screenshot, the text coordinates corresponding to the text information, and the image coordinates; Extracting features of the page screenshot using the image model to obtain image features of the page screenshot; Using a first feature conversion module to convert the image features to obtain an image vector; Based on the text vector and the image vector, the second prediction value of the rating result for each store is obtained by using the causal language model.

8. The method for training a topic scoring model according to claim 4, characterized in that: The method includes obtaining the actual scoring result, including: determining a target store from the plurality of stores; Determine a traffic score based on the number of unique visitors to the store page of the target store and the transaction volume after visiting the store page; Obtaining the daily transaction volume of each store, and determining a first store with the largest daily transaction volume and a second store with the smallest daily transaction volume from the multiple stores according to the daily transaction volume of each store; Determining a transaction ratio based on the traffic score and the transaction volume of the target store on that day; Obtaining a transaction score based on the daily transaction volume of the target store, the daily transaction volume corresponding to the first store, the daily transaction volume corresponding to the second store, and the transaction ratio; Based on the traffic score and the transaction score, the actual scoring result is obtained.

9. A page design evaluation device, characterized in that: include: An acquisition module, used to acquire a page screenshot corresponding to the store page to be evaluated, wherein the page screenshot includes text information and an image, text coordinates corresponding to the text information, and image coordinates corresponding to the image; A text module, used for obtaining a text vector based on the text information, the text coordinates and the image coordinates; An image module, used for obtaining image features of the page screenshot based on the page screenshot and a preset image model; A feature module, used for converting the image feature by using the first feature conversion module to obtain an image vector; The scoring module is used to obtain a scoring result for the store page to be evaluated based on the text vector, the image vector and a preset causal language model.

10. A computer device, characterized in that: include: Memory; and processor; The processor executes computer-readable instructions stored in the memory to implement the page design evaluation method described in any one of claims 1 to 3 and the training method of the topic scoring model described in any one of claims 4 to 8.

Images