Advertisement text optimization method and device, equipment, medium and product

Abstract

The application relates to an advertisement text optimization method and device, equipment, medium and product. The method comprises the following steps: obtaining a product picture and a product description text of a target product; matching a plurality of creative advertisements in an advertisement creative library according to the product picture to obtain creative texts; performing word frequency statistics based on the creative texts to determine the distribution probability of each word segmentation, and mapping the distribution probability to the reference probability distribution of a reference word table referenced by a corresponding advertisement text generation model; applying the advertisement text generation model to generate a text feature sequence corresponding to the advertisement text based on the product description text, wherein the generation probability distribution of a target word element in the reference word table is obtained at each time step, and the target word element is determined after the generation probability distribution is corrected based on the reference probability distribution. The application can determine the creative text by using the product picture, and can guide the advertisement text generation model to generate the advertisement text by using the word frequency features of the creative text, so that the quality of the generated advertisement text is better.

Description

Technical Field

[0001] This application relates to advertising text generation technology, and more particularly to an advertising text optimization method and corresponding apparatus, computer equipment, computer-readable storage medium, and computer program product. Background Technology

[0002] In e-commerce, advertising is a crucial way to attract potential customers and drive purchases. One of the key elements presented to users during advertising is the advertising copy. In practice, creating effective advertising copy requires a thorough understanding of the product and target market, as well as familiarity with advertising copywriting guidelines. Based on this understanding, one must extract the product's selling points, employ appropriate expression techniques, and craft copy that resonates with the market and maximizes commercial success. This clearly demonstrates the significant time and effort required for advertising copy creation. Furthermore, writing advertising copy demands a high level of expertise, which merchants often lack, further complicating the process.

[0003] To address these issues, deep learning-based neural network models are often used to automatically generate relevant advertising text for merchants and users. Currently, advertising copy generation technology typically involves inputting product-related information, including the product title, details page text, and product images. This is done using encoder-decoder models, autoregressive language models, or multimodal Transformer structures. The model is then trained on a large-scale text corpus, and the converged model is used to generate the copy. Widely used models include GPT3, GPT2, T5, BART, and CTRL. During the copy inference or testing phase, the model acts as a black box, predicting and generating the copy, with minimal human intervention.

[0004] When creating advertising copy, businesses often focus on referencing previous advertisements for similar products, as well as those from other businesses, especially competitors. Therefore, strengthening the influence of advertising creativity related to similar products in copywriting—that is, making fuller use of advertising creativity from similar products—remains a topic that urgently needs further research. Summary of the Invention

[0005] The purpose of this application is to solve the above-mentioned problems by providing an advertising text optimization method and corresponding apparatus, computer equipment, computer-readable storage medium, and computer program product.

[0006] To suit the various purposes of this application, the following technical solution is adopted:

[0007] In one aspect, to suit one of the purposes of this application, an advertising text optimization method is provided, comprising:

[0008] Obtain the product image and product description text of the target product;

[0009] Based on the product image of the target product, multiple creative advertisements are matched in the advertising creative library to obtain the creative text in the creative advertisements, wherein the product image of each creative advertisement is semantically matched with the product image of the target product;

[0010] Based on the word frequency statistics of all matched creative texts, the distribution probability of each word segment is determined and mapped to the reference probability distribution of the reference vocabulary referenced by the corresponding advertising text generation model.

[0011] An advertising text generation model that uses time-step decoding is applied to generate a text feature sequence corresponding to the advertising text based on the product description text. In this model, at each time step, the generation probability distribution of the target word corresponding to the target position in the text feature sequence is obtained in the reference vocabulary. The target word is then determined after the generation probability distribution is corrected by the reference probability distribution.

[0012] On the other hand, to suit one of the purposes of this application, an advertising text optimization device is provided, comprising:

[0013] The data acquisition module is configured to acquire product images and product description text for the target product;

[0014] The creative matching module is configured to match multiple creative ads in the advertising creative library based on the product image of the target product, and obtain the creative text in the creative ads, wherein the product image of each creative ad is semantically matched with the product image of the target product;

[0015] The probability mapping module is set to perform word frequency statistics on all matched creative texts, determine the distribution probability of each word segment, and map it to the reference probability distribution of the reference vocabulary referenced by the corresponding advertising text generation model;

[0016] The text generation module is configured to use a time-step decoding advertising text generation model to generate a text feature sequence corresponding to the advertising text based on the product description text. In each time step, the generation probability distribution of the target word corresponding to the target position in the text feature sequence is obtained in the reference vocabulary. The target word is determined after correcting the generation probability distribution with the reference probability distribution.

[0017] In another aspect, a computer device provided for one of the purposes of this application includes a central processing unit and a memory, the central processing unit being used to invoke and run a computer program stored in the memory to perform the steps of the advertising text optimization method described in this application.

[0018] In another aspect, a computer-readable storage medium is provided to suit another purpose of this application, which stores, in the form of computer-readable instructions, a computer program implemented according to the described advertising text optimization method, which, when invoked by a computer, performs the steps included in the method.

[0019] In another aspect, a computer program product provided for another purpose of this application includes a computer program / instructions that, when executed by a processor, implement the steps of the advertising text optimization method described in any embodiment of this application.

[0020] Compared with existing technologies, this application has several technological advantages, including but not limited to:

[0021] First, this application matches creative advertisements with product images of the target product and obtains the creative text therein, which is used to generate a reference probability distribution. Then, the reference probability distribution is used to correct the generation probability distribution of the advertisement text generation model at each time step based on the product description text of the target product. In this way, the historical creative text is used to guide the advertisement text generation model to predict high-quality advertisement text.

[0022] Secondly, when determining the creative text in the creative advertisement used to generate the reference probability distribution, this application is based on matching the semantic vectors of the product images of the target product and the product images of the creative advertisement. It takes advantage of the characteristics of product images, such as high image coverage, relatively consistent data distribution, and the ability to represent similar products, to accurately and comprehensively match creative advertisements of similar products of the target product, thereby obtaining creative text that is closer to the actual needs, which serves as important reference information in guiding the model to generate advertisement text.

[0023] Furthermore, in the process of guiding the advertising text generation model to generate advertising text, this application uses the reference probability distribution to correct the generation probability distribution of the current time step for each time step, and then determines the target word corresponding to the current time step based on the generation probability distribution. The next time step will continue to reason based on the target word determined in the current time step, and sequentially advance the prediction of each target word in the advertising text, realizing the reasoning and generation process of advertising text related to the preceding text, making the generated advertising text more fluent in expression, and more in line with the requirements of advertising characteristics in meaning because it references historical creative text. Attached Figure Description

[0024] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0025] Figure 1 This is a schematic diagram of the network architecture of an exemplary advertising text generation model in this application;

[0026] Figure 2 This is a flowchart illustrating one embodiment of the advertising text optimization method of this application.

[0027] Figure 3 This is a schematic diagram illustrating the process of filtering creative text based on product images in an embodiment of this application.

[0028] Figure 4 This is a schematic diagram illustrating the process of obtaining a reference probability distribution based on creative text in an embodiment of this application.

[0029] Figure 5 This is a flowchart illustrating the reasoning process of the advertising text generation model in this application embodiment.

[0030] Figure 6 This is a schematic diagram of the process of generating a probability distribution by correcting a reference probability distribution in an embodiment of this application.

[0031] Figure 7 This is a flowchart illustrating the autoregressive process of the decoder in the advertising text generation model of this application embodiment.

[0032] Figure 8 This is a schematic block diagram of the advertising text optimization device of this application;

[0033] Figure 9 This is a schematic diagram of the structure of a computer device used in this application. Detailed Implementation

[0034] Unless otherwise specified, the neural network models referenced or potentially referenced in this application may be deployed on a remote server and invoked remotely on the client, or deployed on a client with the capability to invoke directly. In some embodiments, when running on the client, the corresponding intelligence may be acquired through transfer learning in order to reduce the requirements on the client's hardware resources and avoid excessive consumption of the client's hardware resources.

[0035] Unless otherwise expressly stated, the various embodiments disclosed in this application can be combined in a cross-cutting manner to flexibly construct new embodiments, as long as such combination does not depart from the inventive spirit of this application and can meet the needs of the prior art or solve a certain deficiency in the prior art. Those skilled in the art should be aware of such modifications.

[0036] Please see Figure 1In the network architecture of the advertising text generation model exemplarily provided in this application, the advertising text generation model includes an encoder, a decoder, and a vocabulary classifier. The encoder receives the input of the encoded vector of the product description text, extracts the deep semantic information therein, obtains the encoded vector sequence, and inputs it into the decoder. The decoder performs autoregressive decoding using the encoded vector sequence as prefix information. Through the vocabulary classifier, it predicts the generation probability distribution corresponding to each prefix information step by step, and then updates the prefix information with the encoded features of the words determined according to the generation probability distribution to continue predicting the generation probability distribution of the next time step, thereby realizing autoregressive decoding. The vocabulary classifier can map the hidden vectors decoded by the decoder to the reference vocabulary referenced by the advertising text generation model, thereby obtaining a probability distribution mapped to each word in the reference vocabulary as the generation probability distribution.

[0037] In some embodiments disclosed later in this application, the generation probability distribution generated at each time step can be corrected by means of a reference probability distribution generated based on historical creative text, thereby optimizing the generation probability distribution at each time step. This optimizes the word selection at each time step by referencing word frequency features in historical creative text, making the generated advertising text more consistent with the writing style of high-quality historical creative advertisements.

[0038] The encoder described is suitable for vector representation and multi-layer semantic space transformation learning of input text, and can be selected from a variety of known models, including but not limited to RNN, BiLSTM, BiGRU, RoBERTa, ALBert, ERNIE, BERT-WWM, etc.

[0039] The decoder described herein is adapted to learn and adjust the state information of the hidden layer at the current time step by utilizing the preceding encoded vector sequence and the state information of the last hidden layer of the encoder, through multiple layers of self-attention and cross-self-attention mechanisms. Selectable autoregressive pre-trained models include, but are not limited to, GPT2, XLNet, DistilGPT2, etc., and can also employ classic models such as RNN, LSTM, and GRU.

[0040] In one embodiment, a Transformer model, which includes both an encoder and a decoder, may be used as an option for the advertising text generation model.

[0041] The vocabulary classifier can be constructed using the Softmax function, which is connected after the last hidden layer of the decoder to predict the word corresponding to the current time step and calculate the generation probability distribution of the corresponding hidden layer vector mapped to each word in the entire reference vocabulary.

[0042] The above exemplary advertising text generation model, after being pre-trained to a convergent state, can be used to implement the technical solutions of this application, including the methods, apparatus, devices, media, products, etc. of this application. The specific implementation methods of each technical solution will be disclosed below.

[0043] The advertising text optimization method of this application can be programmed into a computer program product and deployed on a client or server for implementation. For example, in the exemplary application scenario of this application, it can be deployed and implemented on the server of an e-commerce platform.

[0044] Please see Figure 2 In one embodiment of the advertising text optimization method of this application, the following steps are included:

[0045] Step S1100: Obtain the product image and product description text of the target product;

[0046] When an advertiser wants to generate advertising text for a target product, in one embodiment, an ad generation page can be used, where the target product corresponding to the advertising text to be generated can be specified. Typically, the advertiser is a merchant user of an online store or its agent user, and the target product can be a product from the online store's product information database, or it can be a customized product provided by the user with specific product information. For ease of understanding, it is assumed that all product-related information exists in the product information database, and the user can specify the target product on the ad generation page.

[0047] When a user identifies a target product, they can typically determine it by specifying the product's unique characteristic information, namely the product ID. Based on this product ID, the product information of the target product can be retrieved from the product information database.

[0048] The product information includes, but is not limited to, different types of data such as images and text. Images in the product information include various images showcasing the product from an overall perspective and from different angles. Typically, a primary image is selected as the default display image, which usually displays the full picture of the product. In this application, it is recommended to use the default display image of the target product as the product image used for subsequent processing. Text in the product information includes, but is not limited to, any one or any combination of product titles, detail text, attribute data, category tags, etc. In this application, it is recommended to use the product title as the product description text of the target product used for subsequent processing.

[0049] Therefore, it is easy to understand that when a user specifies a target product, the product image and description text of that target product can be obtained from the product information database.

[0050] Step S1200: Match multiple creative advertisements in the advertising creative library based on the product image of the target product, and obtain the creative text in the creative advertisement, wherein the product image of each creative advertisement and the product image of the target product constitute semantic matching;

[0051] This application provides an advertising creative library containing multiple creative advertisements. These creative advertisements can be advertisements historically placed in the advertising system. Each creative advertisement contains a product image and its corresponding creative text, which is descriptive text showcasing the product's selling points. Typically, within the same creative advertisement, there is a certain correlation between the creative text and the product image. For example, when the creative text contains the phrase "seven-color gradient skin tone, the princess in the mobile phone," the corresponding product image might be a picture of a mobile phone with a seven-color gradient back cover.

[0052] In one embodiment, the advertising creative library of this application includes two sub-libraries: a private advertising creative library and a public advertising creative library. The private advertising creative library stores creative advertisements corresponding to the products of the current user's online store, while the public advertising creative library stores creative advertisements published by other users' online stores. Accordingly, the private advertising creative library is preset to be accessible only to the current merchant user, while the public advertising creative library is preset to be accessible to all merchant users of the e-commerce platform. By mapping all creative advertisements in the private and public advertising creative libraries to index tags determined by a unified algorithm, any creative advertisement can be quickly identified using the index tags. The index tags can be natural number sequences or the results obtained by encoding the corresponding creative advertisement using algorithms such as MD5, SHA1, and Base64.

[0053] In the advertising creative library of this application, the product images in each creative advertisement are pre-extracted with corresponding encoding vectors using a preset image encoder and stored accordingly for retrieval in this step.

[0054] Similarly, for the product image of the target product, the corresponding encoding vector is extracted using the image encoder. Then, the data distance between the encoding vector of the target product image and the encoding vectors of the product images of each creative advertisement in the advertising creative library is calculated. The similarity is determined based on this data distance. Then, a preset threshold is used to filter the similarity of each creative advertisement, selecting those with similarity higher than the preset threshold to form a target creative advertisement set. The creative advertisements in the target creative advertisement set are considered to be semantically matched with the product image of the target product. In an alternative embodiment, instead of using a preset threshold to filter similarity, the creative advertisements can be sorted according to their similarity, and then a preset number of creative advertisements can be selected as the target creative advertisements to form a corresponding target creative advertisement set.

[0055] Therefore, based on the product images of the target product, semantic matching of images is usually performed. Target creative advertisements that semantically match the product images of the target product can be retrieved from the advertising creative library. There is a semantic similarity matching relationship between the product images in these target creative advertisements and the product images of the target product.

[0056] At this point, all the creative texts from the matched creative ads have been extracted and constructed into a creative text set. The creative texts in this set are the target creative texts obtained by performing image semantic matching from the ad creative library based on the product images of the target products.

[0057] When calculating the data distance, any of the large-scale vector retrieval engines such as Faiss, ElasticSearch, and Milvus can be used, or any existing algorithm such as cosine similarity, dot product, Manhattan distance, or Euclidean distance can be used.

[0058] It should be noted that in this application, matching creative advertisements based on the encoded vectors of images has a significant positive effect on the subsequent determination of reference probability distribution based on the creative text in the matched creative advertisements to guide the advertising text generation model to predict target words. This is mainly reflected in the high image coverage, relatively consistent data distribution, and the ability of images to represent similar products.

[0059] Step S1300: Perform word frequency statistics based on all matched creative texts, determine the distribution probability of each word segment, and map it to the reference probability distribution of the reference vocabulary referenced by the corresponding advertising text generation model;

[0060] All the matched creative texts constitute the creative text set. In order to generate a reference probability distribution, data mining needs to be performed based on the creative text set.

[0061] In one embodiment, each creative text in the matched creative text set is segmented into words to determine the segmentation sequence corresponding to each creative text. The segmentation sequence includes multiple words and is actually equivalent to an ordered set of words contained in the corresponding creative text. The segmentation sequence can be regarded as a corresponding segmentation set.

[0062] In another embodiment, for each creative text in the creative text set, any method such as named entity recognition, syntactic structure analysis, TextRank algorithm, YAKE algorithm, etc., can be applied to first extract key sentences, and then perform word segmentation based on the key sentences to determine the corresponding word segmentation sequence, which serves as the word segmentation set corresponding to the creative text.

[0063] When segmenting the creative text, the same segmenter used by the advertising text generation model during encoding input can be employed. Since the segments of the advertising text generation model are segmented according to the segments in the reference vocabulary referenced by the advertising text generation model, it can be ensured that the result of segmenting the creative text corresponds to the segments understood by the advertising text generation model during reasoning. This can enhance the information reference value of the creative text for the reasoning process of the advertising text generation model.

[0064] After determining the word segmentation sequence corresponding to each matched creative text, word frequency statistics can be performed on each word in the entire word segmentation sequence obtained from the entire creative text set. The probability of each word appearing in the entire creative text set can be calculated, thereby obtaining the distribution probability of each word in the entire creative text set. It is easy to understand that the probability of each word appearing indicates the degree of widespread use of that word in all creative texts. It can be regarded as an indicator of the popularity of using that word to write creative texts. Such an indication is obviously of reference value for guiding the advertising text generation model to select words for advertising texts.

[0065] To facilitate the conversion of the probability distribution of all word segments in the entire creative text set into data that can be directly referenced by the advertising text generation model, a reference probability distribution can be created for each word element in the reference vocabulary referenced by the advertising text generation model. This reference probability distribution provides a corresponding probability for each word element in the reference vocabulary. Therefore, it is equivalent to determining the reference information value of each word element in the reference vocabulary through the reference probability distribution, which can be used to guide the word selection of the advertising text generation model at each time step.

[0066] Step S1400: Apply the advertising text generation model that decodes step by step over time to generate a text feature sequence corresponding to the advertising text based on the product description text. In this step, the generation probability distribution of the target word corresponding to the target position in the text feature sequence is obtained in the reference word list at each time step. The target word is determined after the generation probability distribution is corrected by the reference probability distribution.

[0067] To adapt to the input format requirements of the advertising text generation model, the product description text of the target product can be determined as the input of the advertising text generation model. Word embedding is performed on the product description text according to the reference vocabulary to determine the encoding features of each word in the product description text in the reference vocabulary, thereby obtaining the corresponding encoding vector.

[0068] Then, under the encoder of the advertising text generation model, deep semantic information is extracted from the encoded vector to obtain a corresponding encoded vector sequence. This encoded vector sequence is used as prefix information and input into the decoder of the advertising text generation model for decoding. The decoder, based on its own autoregressive mechanism, decodes step by step according to the encoded vector sequence. During the decoding process, the results obtained from the hidden layer are mapped to the vocabulary classifier through a fully connected layer to obtain the corresponding generation probability distribution. This generation probability distribution represents the classification probability of each word in the reference vocabulary. After selecting the target word corresponding to the current time step according to the classification probability in the generation probability distribution, the encoded features of the target word in the reference vocabulary are appended to the prefix information, and the prediction of the next generation probability distribution continues until the end symbol is encountered. Thus, the entire process gradually determines each target word in the advertising text and obtains the corresponding text feature sequence, which contains tokens of multiple words. Based on these tokens, the specific text in the reference vocabulary can be determined, thereby converting the text feature sequence into advertising text.

[0069] It should be noted that, in order to use the reference probability distribution to guide word selection at each time step, the generation probability distribution at each time step can be superimposed on the reference probability distribution. The probability of the corresponding target position, i.e. the corresponding word element, in the generation probability distribution can be corrected using each probability in the reference probability distribution, thereby updating the generation probability distribution. Then, based on the generation probability distribution, the target word element corresponding to the current time step can be determined, thus realizing the word selection operation at the current time step guided by the reference probability distribution.

[0070] When using the reference probability distribution to correct the generation probability distribution at each time step, the two can be flexibly superimposed. For example, the reference probability distribution can be directly used as the weight of the word corresponding to the generation probability distribution for all corrections, or the classification probabilities of some words in the generation probability distribution can be individually corrected based on the reference probability distribution. Those skilled in the art can implement this flexibly based on the principles disclosed in this application.

[0071] When selecting words at each time step, different word selection algorithms are adapted. For the same product description text, multiple target words can be identified at each time step to form multiple outputs. Accordingly, each target word is appended to the text feature sequence already obtained at the current time step to differentiate into multiple text feature sequences at the current time step. Then, the encoded features of each target word are concatenated with the prefix information of the current time step, allowing for further inference of the target words for the next time step for each differentiated text feature sequence. By analogy, it is easy to understand that multiple text feature sequences can ultimately be obtained, thus yielding multiple advertising texts. The word selection algorithm, based on given parameters, operates on the generation distribution probability corresponding to each time step, optimizing and determining one or more target words from each classification probability in the generation distribution probability. The word selection algorithm includes, but is not limited to, various algorithms such as Top_K algorithm, cluster search algorithm, and greedy search algorithm.

[0072] Therefore, it is easy to understand that, based on the product description text of the target product and the reference probability distribution obtained from the creative text set determined based on the product image of the target product, one or more advertising texts can be generated through the advertising text generation model of this application.

[0073] As can be seen from the above embodiments, this application contains multiple technical advantages, including but not limited to:

[0074] First, this application matches creative advertisements with product images of the target product and obtains the creative text therein, which is used to generate a reference probability distribution. Then, the reference probability distribution is used to correct the generation probability distribution of the advertisement text generation model at each time step based on the product description text of the target product. In this way, the historical creative text is used to guide the advertisement text generation model to predict high-quality advertisement text.

[0075] Secondly, when determining the creative text in the creative advertisement used to generate the reference probability distribution, this application is based on matching the semantic vectors of the product images of the target product and the product images of the creative advertisement. It takes advantage of the characteristics of product images, such as high image coverage, relatively consistent data distribution, and the ability to represent similar products, to accurately and comprehensively match creative advertisements of similar products of the target product, thereby obtaining creative text that is closer to the actual needs, which serves as important reference information in guiding the model to generate advertisement text.

[0076] Furthermore, in the process of guiding the advertising text generation model to generate advertising text, this application uses the reference probability distribution to correct the generation probability distribution of the current time step for each time step, and then determines the target word corresponding to the current time step based on the generation probability distribution. The next time step will continue to reason based on the target word determined in the current time step, and sequentially advance the prediction of each target word in the advertising text, realizing the reasoning and generation process of advertising text related to the preceding text, making the generated advertising text more fluent in expression, and more in line with the requirements of advertising characteristics in meaning because it references historical creative text.

[0077] Based on any embodiment of this application, please refer to Figure 3 Based on the product image of the target product, multiple creative advertisements are matched in the advertising creative library to obtain the creative text of the creative advertisements, including:

[0078] Step S1210: Use an image encoder to extract the deep semantic information of the product image of the target product and obtain the corresponding encoding vector;

[0079] This application provides an image encoder, which is pre-trained to a convergent state, for extracting deep semantic information from the product images of this application. The image encoder can be implemented using a deep learning-based neural network model, such as a Vision Transformer (ViT) model, EfficientNet model, DenseNet model, ResNet model, etc., which are pre-trained to a convergent state.

[0080] For the product image of the target product, after it is formatted to fit the scale required by the image encoder, it is input into the image encoder. The image encoder performs feature representation on it and extracts its deep semantic information, thereby obtaining its corresponding encoding vector.

[0081] Step S1220: Calculate the similarity between the encoding vector of the product image of the target product and the encoding vector of the product images in each creative advertisement in the advertising creative library;

[0082] Similar to the product images of the target product, the product images carried by each creative advertisement in the advertising creative library of this application also have their deep semantic information extracted in advance using the image encoder, and the encoding vectors corresponding to their deep semantic information are associated and stored in the corresponding creative advertisement for retrieval.

[0083] Accordingly, any vector matching method can be used to calculate the data distance between the encoding vector of the target product and the encoding vector of the product image of each creative advertisement in the advertising creative library. As mentioned above, any tool such as Faiss, ElasticSearch, and Milvus can be used to perform vector matching calculation based on data distance. Alternatively, any data distance algorithm such as cosine similarity, dot product, Manhattan distance, and Euclidean distance can be used to perform vector matching calculation. Regardless of the method used, the vector similarity between the product image of the target product and the product images of each creative advertisement in the advertising creative library can be determined.

[0084] Step S1230: Select some creative advertisements from the advertising creative library based on the similarity, and use them as target creative advertisements that match the product images of the target product;

[0085] After the above process, each creative ad in the ad creative library obtains the similarity score corresponding to the product image. Therefore, various methods can be used to select some creative ads from them as target creative ads that match the product image of the target product.

[0086] In one embodiment, a preset threshold can be used to compare the similarity of each creative advertisement with the preset threshold, and the creative advertisement with a similarity higher than the preset threshold can be used as the target creative advertisement.

[0087] In another embodiment, a preset number of creative advertisements can be used to sort each creative advertisement according to the similarity, and then select the preset number of creative advertisements that rank highest as the target creative advertisement.

[0088] It is easy to understand that the encoding vector of the product image carried by the target creative advertisement is exactly matched with the encoding vector of the product image of the target product.

[0089] Step S1240: Extract the creative text corresponding to the product images in the target creative advertisement.

[0090] Once all target creative ads have been identified from the ad creative library, the creative text of each target creative ad can be extracted to form a creative text set, which will be used to construct a reference probability distribution later.

[0091] Based on the above embodiments, it is easy to understand that determining creative text matching the target product from the advertising creative library based on vector matching of product images has significant positive implications, which are reflected in the following aspects:

[0092] Firstly, the image coverage is high: as the main information carrier of a product, product images are an important part of creative advertising. Basically, any advertisement with creative text will usually contain corresponding product images, so the corresponding creative text can be obtained by matching product images.

[0093] Secondly, the data distribution is relatively consistent: the product images of the target product to be promoted and the product images in the creative advertisement are essentially display images used to showcase the advantages and features of the product, and have a relatively consistent data distribution. In contrast, the product title and creative text are two information formats with significantly different styles and distributions. Therefore, prioritizing the use of product images for creative advertisement searches has a more accurate recall effect than using text for searches.

[0094] Third, images can represent similar products: the product images of the target product and the product images in the creative advertisement can effectively represent whether they belong to the same category or are similar products through the shape of the products. Therefore, when matching based on the semantics of each other's images, more accurate matching results can be obtained.

[0095] Based on any embodiment of this application, please refer to Figure 4 Based on the word frequency statistics of all matched creative texts, the probability distribution of each word segment is determined and mapped to the reference probability distribution of the reference vocabulary used by the corresponding advertising text generation model, including:

[0096] Step S1310: Segment all the matched creative texts into words to obtain the word segmentation sequence corresponding to each creative text;

[0097] In this application, each creative text matched based on the product image of the target product can be segmented using the word segmenter of the advertising text generation model that encodes the product description text, thereby obtaining a word segmentation sequence corresponding to each creative text. This word segmentation sequence contains the individual words corresponding to the creative text. Furthermore, since the word segmenter used is the one corresponding to the input of the advertising text generation model, it necessarily references the reference vocabulary for word segmentation. Therefore, the words in the word segmentation sequence are also words existing in the reference vocabulary as lexical units. Thus, each creative text can obtain its corresponding word segmentation sequence.

[0098] Step S1320: Perform word frequency statistics based on each word segmentation sequence to determine the probability of occurrence of each word in the entire word segmentation sequence;

[0099] Based on the word segmentation sequence of the creative text of each target creative advertisement, word frequency statistics can be performed. In this embodiment, the bag-of-words (BOW) model can be used to directly establish the probability distribution of each word in the entire word segmentation sequence, obtaining a probability distribution table that contains the occurrence probability of each word appearing in the entire word segmentation sequence. This occurrence probability is determined based on the word frequency of its corresponding word.

[0100] Step S1330: Map the probability of each word segment in the entire word set to the probability of each corresponding word segment in the reference probability distribution, which represents the probability of each word element in the reference word list.

[0101] To convert the probability distribution of word segmentation determined based on the creative text set into the reference probability distribution of this application, in one embodiment, an empty probability vector can be created corresponding to the reference lexicon. This probability vector contains multiple elements, each corresponding to a word element in the reference lexicon, and the probability of each element is initialized to 0. Then, using the probability distribution table obtained through word frequency statistics in the previous step, the occurrence probability of each word segment is converted into the probability of the corresponding word segment element in the probability vector. Thus, the construction of the probability vector is completed, and this probability vector is the reference probability distribution that this application aims to determine.

[0102] In another embodiment, the initialization is also based on the aforementioned probability vector. Then, a preset threshold is used. When the probability of a certain word in the probability distribution table is higher than a pre-given preset threshold, the probability corresponding to the position of the corresponding word in the probability vector is modified. This further optimizes the reference information provided by the probability distribution table. Only when the word frequency statistical features meet the preset conditions are the corresponding words introduced into the reference probability distribution, which plays a guiding role in the subsequent time-step word selection of the advertising text generation model.

[0103] Based on the above embodiments, it is easy to understand that the reference probability distribution of this application is determined based on the word frequency statistical features of the creative text of the target creative advertisement that matches the product image of the target product. Since the word frequency statistical features represent the popularity of the corresponding word segment in historical advertising copy, quantifying them into a reference probability distribution to guide the advertising text generation model in reasoning about the target word units used in the advertising text can improve the quality of the advertising text generated by the advertising text generation model, making the generated advertising text more effective in achieving good advertising results.

[0104] Based on any embodiment of this application, please refer to Figure 5 An advertising text generation model using time-step decoding is applied to generate a text feature sequence corresponding to the advertising text based on the product description text, including:

[0105] Step S1410: Encode the product description text according to the reference vocabulary used by the advertising text generation model to obtain the encoding vector;

[0106] To meet the input requirements of the advertising text generation model, the product description text needs to be encoded. For this purpose, a pre-defined word segmenter can be used. This segmenter uses a reference vocabulary as a guide to segment the product description file, obtaining corresponding word sequences. Then, based on the reference vocabulary, each segmented word is converted into the encoding features of the corresponding word element in the reference vocabulary. These encoding features are then constructed into an encoding vector, which is input into the advertising text generation model for inference. In some embodiments, the positional encoding of each word element can be further combined with the encoding features of the word elements to enrich the reference information of the encoding vector.

[0107] Step S1420: The encoder of the advertising text generation model extracts the deep semantic information of the encoded vector to obtain the corresponding encoded vector sequence;

[0108] As mentioned earlier, the encoder in the advertising text generation model is mainly based on RNN. Therefore, it is suitable for extracting deep semantic information by referring to the contextual information of the input encoded vector, realizing a deep semantic representation of the input encoded vector, and obtaining an encoded vector sequence. A self-attention mechanism can also be applied to the encoder to make key features more prominent, making the obtained encoded vector sequence more helpful for subsequent decoding.

[0109] Step S1430: Input the decoder of the advertising text generation model with the encoded vector sequence as prefix information, and autoregressively decode the generation probability distribution of each time step, wherein each generation probability distribution stores the probability of each word in the reference vocabulary.

[0110] The decoder in the advertising text generation model takes the prefix information formed by the encoded vector sequence as its initial input. Based on the self-attention mechanism, it performs serial decoding, decoding step by step. The decoded results are then classified and mapped by a vocabulary classifier to obtain the generation probability distribution of each word in the reference vocabulary. Based on the classification probability in the generation probability distribution, the word corresponding to the current time step can be selected. The encoded features of this word in the reference vocabulary are then appended to the prefix information to achieve autoregression. Then, the decoding continues to the next time step based on the updated prefix information until the end symbol is encountered.

[0111] Step S1440: In the autoregressive decoding process, for each time step, the generation probability distribution generated at that time step is corrected by weighting the reference probability distribution. The target word corresponding to that time step is determined according to the corrected generation probability distribution. The encoding features of the target word are appended to the prefix information corresponding to that time step to generate the generation probability distribution of the next time step.

[0112] Specifically, during the autoregressive decoding process of the decoder, at each time step, the reference vocabulary can be queried based on the generation probability distribution obtained by the vocabulary classifier to determine one or more target words.

[0113] Taking the identification of a single target word as an example, this target word is usually the word with the highest probability in the generation probability distribution. The encoded features of this target word are appended to the prefix information corresponding to the current time step to obtain updated prefix information. This updated prefix information is then passed to the decoder for the next time step to generate the generation probability distribution for the next time step. This process continues until a terminator is encountered, completing the prediction of all generation probability distributions and thus the prediction of each target word in the advertising text. The target words obtained step by step in each time step can be arranged into a corresponding text feature sequence according to the chronological order of the time steps. Each text feature sequence can be converted into a corresponding advertising text by determining the corresponding word in the reference vocabulary based on its encoded features.

[0114] It should be further explained that this application modifies the generation probability distribution corresponding to each time step by referring to the reference probability distribution constructed in this application. In this embodiment, since both the generation probability distribution and the reference probability distribution are mapped to the reference vocabulary, they have the same dimension. Accordingly, the generation probability distribution of the current time step can be modified by weighting the reference probability distribution. Then, a word selection algorithm is applied to the modified generation probability distribution to determine one or more target words corresponding to the current time step.

[0115] The word selection algorithm described above can be used to select one or more target words corresponding to the current time step based on the generation probability distribution of the current time step. After determining multiple target words, each target word is appended to the prefix information of the current time step, which can differentiate into multiple updated prefix information. Subsequently, the generation probability distribution corresponding to the next time step can be inferred based on each updated prefix information. Furthermore, in each time step, the generation probability distribution of the time step can be corrected by combining the reference probability distribution and then combined with the word selection algorithm to determine the multiple target words to be selected in that time step. This achieves gradual differentiation across multiple time steps. Each time step differentiates the text feature sequence obtained in the previous time step into multiple text feature sequences, ultimately obtaining the text feature sequences corresponding to multiple advertising texts, and thus determining multiple advertising texts.

[0116] As can be understood from the above embodiments, with the help of the advertising text generation model of this application, one or more advertising texts can be predicted based on the same product description text, facilitating the selection of the best within a set. Because the advertising text generation model has serialization processing capabilities, it can predict the corresponding text feature sequence with reference to the context. Furthermore, during the serialization process, the generation probability distribution at each time step is superimposed with the reference probability distribution prepared in this application to optimize the target word units. Therefore, its semantic understanding capability is strong, enabling the generation of fluent and more suitable advertising texts for promotion. Since the prediction of these advertising texts references the information provided by historical creative texts, it is expected that better advertising results can be achieved smoothly.

[0117] Based on any embodiment of this application, please refer to Figure 6 The generation probability distribution generated at this time step is weighted and corrected using the reference probability distribution, including:

[0118] Step S2100: The reference probability distribution is weighted using preset weights to obtain a weighted reference probability distribution;

[0119] In one embodiment, before adjusting the generation probability distribution for each time step, the reference probability distribution is first weighted using preset weights. This provides a means to adjust the impact of the reference probability distribution on the advertising text generation model, making the process of guiding the advertising text generation model to produce advertising text based on the reference probability distribution more controllable. After weighting the reference probability distribution using preset weights, the weighted reference probability distribution can be obtained.

[0120] Step S2200: Calculate the element-wise product of the weighted reference probability distribution and the generation probability distribution at the current time step to obtain the weighted optimized weight distribution;

[0121] Then, the element-wise product (Hadamard product) of the weighted reference probability distribution and the generated probability distribution is calculated, and element-wise weighting is performed. Thus, based on the generated probability distribution, the weighted reference probability distribution is introduced, and the amplitude of each element in the generated probability distribution is extracted through calculation to form a weighted distribution.

[0122] Step S2300: Calculate the sum of the elements of the generation probability distribution and the weight distribution to obtain the generation probability distribution updated for the current time step.

[0123] Finally, based on the generation probability distribution at the current time step, the sum of its elements and the weight distribution obtained in the previous step is calculated to adjust and update the generation probability distribution at the current time step. The updated generation probability distribution can then be used to determine the target word corresponding to the current time step.

[0124] Based on the above process, it can be abstracted and simplified into the following exemplary formula:

[0125] P′ gen =P gen *(1+w·P bow )

[0126] Among them, P gen Let P be the generation probability distribution at the current time step, where w is a preset weight that can take values ​​within the half-open interval (0,1]. bow As a reference probability distribution, P′ gen This represents the updated generation probability distribution, where * denotes the Hadamard product operation.

[0127] In this embodiment, when applying the updated generation probability distribution obtained in this embodiment and using a preset threshold, i.e., a probability threshold, to select target words at the current time step, the probability threshold can be set in relation to the preset weight, for example, proportional to the preset weight, so that the larger the preset weight, the larger the probability threshold. Specifically, when the preset weight is 1, the probability weight can take any value between 0.8 and 1.5. By using the probability threshold to determine target words with probabilities higher than the updated generation probability distribution, and determining multiple outputs at the current time step, multiple text feature sequences can be differentiated, ultimately resulting in multiple advertising texts.

[0128] As can be seen from the above embodiments, when using the reference probability distribution to correct the generation probability distribution of each time step, the weight can be preset as a control method to extract the weight distribution of key information in the generation probability distribution of the current time step, and then adjust the generation probability distribution accordingly. This effectively combines the prior knowledge obtained from the statistical word frequency of historical creative texts into the generation probability distribution of each time step, determines the target word unit of each time step, guides the word selection and differentiation process of the advertising text generation model, and makes the final output advertising text more in line with the quality requirements of advertising text and more likely to meet the good advertising expectations.

[0129] Based on any embodiment of this application, please refer to Figure 7 The generation probability distribution generated at the current time step is weighted and corrected using the reference probability distribution. The target word corresponding to the current time step is determined based on the corrected generation probability distribution. The encoded features of the target word are appended to the prefix information corresponding to the current time step to generate the generation probability distribution for the next time step, including:

[0130] Step S1441: Starting from the generation probability distribution obtained in the first time step, after correcting the generation probability distribution using a reference probability distribution, determine multiple target words with probabilities higher than a preset threshold based on the corrected generation probability distribution, and construct each text feature sequence with the encoding features of each target word as the guide.

[0131] At the first time step, the encoded vector sequence output by the encoder is used as initial prefix information input to the decoder. The decoder decodes the vector and maps it through a vocabulary classifier to obtain the corresponding generation probability distribution. This generation probability distribution is then corrected using the reference probability distribution to obtain an updated generation probability distribution. Since the generation probability distribution stores the probability of each word in the reference vocabulary mapped to the current time step in vector form, a first quantity can be set, corresponding to the updated generation probability distribution at the current time step. Multiple words with the highest probabilities corresponding to this first quantity are selected as target words. Each target word is used as a precursor to construct multiple corresponding text feature sequences.

[0132] Step S1442: Each text feature sequence is individually appended to the prefix information. The generation probability distribution of the next time step is predicted based on the prefix information. After correcting the generation probability distribution using a reference probability distribution, multiple target words with probabilities higher than a preset threshold are determined according to the corrected generation probability distribution. The encoding features of each target word are appended to the corresponding text feature sequence to differentiate multiple updated text feature sequences.

[0133] For one or more target words generated at the current time step, their encoded features can be appended to the prefix information used to generate the generation probability distribution at the current time step, thus differentiating multiple prefix information and updating the prefix information. This allows for the generation probability distribution of the next time step to be generated based on the updated prefix information. The updated generation probability distribution is obtained by superimposing the reference probability distribution on the generation probability distribution of the current time step, and then word selection continues based on the updated generation probability distribution of the current time step, following the logic of the first time step.

[0134] As time steps progress, it is easy to understand that when a target word is appended to the prefix information corresponding to the current time step, it is actually equivalent to appending a text feature sequence corresponding to the current time step to the prefix information corresponding to the first time step.

[0135] A text feature sequence from a previous time step, after each target word in the current time step is generated, can be followed by each target word in the current time step, thus differentiating into multiple text feature sequences. It's easy to understand that as time steps progress, by concatenating the corresponding text feature sequences with multiple target words determined according to the first quantity, multiple text feature sequences can be continuously differentiated. When the last time step is reached, a large number of text feature sequences can be obtained. Since each text feature sequence can correspond to a specific advertising text, a large number of advertising texts can actually be predicted based on a single product description text.

[0136] Step S1443: Continue iterating the above process for the next time step until the preset conditions are met and the process terminates, obtaining the final determined updated text feature sequence for conversion into the corresponding advertising text.

[0137] For each subsequent time step, the process of step S1442 can be iterated until a preset condition is met, at which point the iteration terminates. The preset condition can refer to the end symbol in the encountered prefix information. When the iteration terminates, it means that the entire inference process ends, and the final obtained text feature sequence can be used as the text feature sequence corresponding to the advertising text actually predicted by the advertising text generation model, and can be used to convert the corresponding advertising text.

[0138] Based on the above exemplary autoregressive process, it can be understood that in the decoding process of the advertising text generation model, this application can continuously differentiate the text feature sequences corresponding to each time step by selecting multiple target words at each time step, thereby achieving the optimization of the text feature sequences, realizing the quality control of the advertising text predicted by the advertising text generation model, and ensuring that the advertising text generation model generates a batch of advertising texts with better overall quality.

[0139] Please see Figure 8 To meet one of the purposes of this application, an advertising text optimization device is provided, which is a functional embodiment of the advertising text optimization method of this application. The device includes a data acquisition module 1100, a creative matching module 1200, a probability mapping module 1300, and a text generation module 1400. The data acquisition module 1100 is configured to acquire product images and product description text of a target product. The creative matching module 1200 is configured to match multiple creative advertisements in an advertising creative library based on the product images of the target product, and obtain the creative text in the creative advertisements. Each creative advertisement's product image and the product image of the target product constitute... Semantic matching; the probability mapping module 1300 is configured to perform word frequency statistics based on all matched creative texts, determine the distribution probability of each word segment, and map it to the reference probability distribution of the reference vocabulary referenced by the corresponding advertising text generation model; the text generation module 1400 is configured to apply a time-step decoding advertising text generation model to generate a text feature sequence corresponding to the advertising text based on the product description text, wherein, at each time step, the generation probability distribution of the target word corresponding to the target position in the text feature sequence is obtained in the reference vocabulary, and the target word is determined after correcting the generation probability distribution with the reference probability distribution.

[0140] Based on any embodiment of this application, the creative matching module 1200 includes: an image encoding unit, configured to extract deep semantic information of the product image of the target product using an image encoder to obtain a corresponding encoding vector; a similarity calculation unit, configured to calculate the similarity between the encoding vector of the product image of the target product and the encoding vector of the product images in various creative advertisements in the advertising creative library; an advertisement filtering unit, configured to filter some creative advertisements from the advertising creative library according to the similarity, as target creative advertisements matching the product image of the target product; and a text extraction unit, configured to extract creative text corresponding to the product image in the target creative advertisement.

[0141] Based on any embodiment of this application, the probability mapping module 1300 includes: a word segmentation processing unit, configured to segment all matched creative texts to obtain word segmentation sequences corresponding to each creative text; a word frequency statistics unit, configured to perform word frequency statistics based on each word segmentation sequence to determine the probability of occurrence of each word segmentation in the entire word segmentation sequence; and a conversion mapping unit, configured to map the probability of occurrence of each word segmentation in the entire word segmentation set to the probability of each corresponding word segmentation in the reference probability distribution corresponding to the probability of each word element in the reference word list.

[0142] Based on any embodiment of this application, the text generation module 1400 includes: a vector embedding unit, configured to encode the product description text according to a reference vocabulary referenced by the advertising text generation model to obtain an encoded vector; an encoding processing unit, configured to extract deep semantic information of the encoded vector using the encoder of the advertising text generation model to obtain a corresponding encoded vector sequence; a decoding processing unit, configured to input the encoded vector sequence as prefix information into the decoder of the advertising text generation model, and autoregressively decode the generation probability distribution of each time step step by time, wherein each generation probability distribution stores the probability of each word mapped to the reference vocabulary; and a regression processing unit, configured to, during the autoregressive decoding process, for each time step, weight and correct the generation probability distribution generated at that time step using the reference probability distribution, determine the target word corresponding to that time step according to the corrected generation probability distribution, and append the encoded features of the target word to the prefix information corresponding to that time step to generate the generation probability distribution of the next time step.

[0143] Based on any embodiment of this application, the regression processing unit includes: a weighting processing subunit, configured to weight the reference probability distribution using preset weights to obtain a weighted reference probability distribution; a product calculation subunit, configured to calculate the element-wise product of the weighted reference probability distribution and the generation probability distribution at the current time step to obtain a weighted optimized weight distribution; and a summation update subunit, configured to calculate the element-wise sum of the generation probability distribution and the weight distribution to obtain the updated generation probability distribution at the current time step.

[0144] Based on any embodiment of this application, the regression processing unit includes: an initial processing subunit, configured to start from the generation probability distribution obtained at the first time step, and after correcting the generation probability distribution using a reference probability distribution, determine multiple target words with probabilities higher than a preset threshold according to the corrected generation probability distribution, and construct each text feature sequence with the encoding features of each target word as a guide; an iterative processing subunit, configured to append each text feature sequence separately to the prefix information, predict the generation probability distribution at the next time step according to the prefix information, and after correcting the generation probability distribution using a reference probability distribution, determine multiple target words with probabilities higher than a preset threshold according to the corrected generation probability distribution, and append the encoding features of each target word to the corresponding text feature sequence to differentiate into multiple updated text feature sequences; and an iterative control subunit, configured to continue iterating the above process at the next time step until a preset condition is met and the process terminates, obtaining the finally determined updated text feature sequence for conversion into corresponding advertising text.

[0145] To address the aforementioned technical problems, embodiments of this application also provide computer equipment. For example... Figure 9 As shown, the computer device includes a processor, a computer-readable storage medium, a memory, and a network interface connected via a system bus. The computer-readable storage medium stores an operating system, a database, and computer-readable instructions. The database may store a sequence of control information. When executed by the processor, the computer-readable instructions enable the processor to implement a product search category identification method. The processor provides computing and control capabilities to support the operation of the entire computer device. The memory stores computer-readable instructions, which, when executed by the processor, enable the processor to execute the advertising text optimization method of this application. The network interface of the computer device is used for communication with a terminal. Those skilled in the art will understand that… Figure 9 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 computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0146] In this embodiment, the processor is used to execute... Figure 8 The system contains the specific functions of each module and its sub-modules, and the memory stores the program code and various data required to execute these modules or sub-modules. The network interface is used for data transmission between the user terminal and the server. In this embodiment, the memory stores the program code and data required to execute all modules / sub-modules in the advertising text optimization device of this application, and the server can call the server's program code and data to execute the functions of all sub-modules.

[0147] This application also provides a storage medium storing computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the steps of the advertising text optimization method of any embodiment of this application.

[0148] This application also provides a computer program product, including a computer program / instructions that, when executed by one or more processors, implement the steps of the method described in any embodiment of this application.

[0149] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. This computer program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. The aforementioned storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0150] In summary, this application utilizes product images to determine creative text, uses the word frequency features of the creative text to generate a reference probability distribution to adjust the generation probability distribution obtained at each time step during the decoding process of the advertising text generation model, and then predicts the target words at the corresponding time step based on the generation probability distribution. These target words constitute the advertising text. Thus, the advertising text generation model can be guided step by step to generate advertising text, and high-quality advertising text can be generated by referring to historical creative text.

Claims

1. A method for optimizing advertising text, characterized in that, include: Obtain the product image and product description text of the target product; Based on the product image of the target product, multiple creative advertisements are matched in the advertising creative library to obtain the creative text in the creative advertisements, wherein the product image of each creative advertisement is semantically matched with the product image of the target product; Based on the word frequency statistics of all matched creative texts, the distribution probability of each word segment is determined and mapped to the reference probability distribution of the reference vocabulary referenced by the corresponding advertising text generation model. An advertising text generation model that uses time-step decoding is applied to generate a text feature sequence corresponding to the advertising text based on the product description text. In each time step, the generation probability distribution of the target word corresponding to the target position in the text feature sequence is obtained in the reference vocabulary. The target word is determined after the generation probability distribution is corrected by the reference probability distribution. The step of correcting the generation probability distribution generated at the current time step by weighting the reference probability distribution includes: weighting the reference probability distribution with preset weights to obtain a weighted reference probability distribution; calculating the element-wise product of the weighted reference probability distribution and the generation probability distribution at the current time step to obtain a weighted optimized weight distribution; and obtaining the element-wise sum of the generation probability distribution and the weight distribution to obtain the generation probability distribution updated for the current time step.

2. The advertising text optimization method according to claim 1, characterized in that, Based on the product image of the target product, multiple creative advertisements are matched from the advertising creative library to obtain the creative text of the creative advertisements, including: An image encoder is used to extract deep semantic information from the product image of the target product to obtain the corresponding encoding vector; Calculate the similarity between the encoding vector of the product image of the target product and the encoding vector of the product images in each creative advertisement in the advertising creative library; Based on the similarity, a subset of creative ads are selected from the ad creative library and used as target creative ads to match the product images of the target product. Extract the creative text corresponding to the product images in the target creative advertisement.

3. The advertising text optimization method according to claim 1, characterized in that, Based on the word frequency statistics of all matched creative texts, the probability distribution of each word segment is determined and mapped to the reference probability distribution of the reference vocabulary used by the corresponding advertising text generation model, including: All matched creative texts are segmented into words to obtain the segmentation sequence corresponding to each creative text. Based on the word frequency statistics of each word segmentation sequence, the probability of occurrence of each word in the entire word segmentation sequence is determined; The reference probability distribution, which represents the probability of each word element in the reference vocabulary, maps the occurrence probability of each word segment in the entire word set to the probability of each corresponding word element in the reference probability distribution.

4. The advertising text optimization method according to claim 1, characterized in that, An advertising text generation model using time-step decoding is applied to generate a text feature sequence corresponding to the advertising text based on the product description text, including: The product description text is encoded according to the reference vocabulary used by the advertising text generation model to obtain an encoding vector; The encoder of the advertising text generation model extracts the deep semantic information of the encoded vector to obtain the corresponding encoded vector sequence; The decoder of the advertising text generation model is input with the encoded vector sequence as prefix information, and the generation probability distribution of each time step is decoded by autoregression step by step, wherein each generation probability distribution stores the probability of each word in the reference vocabulary. In the autoregressive decoding process, for each time step, the generation probability distribution generated at that time step is weighted and corrected using the reference probability distribution. The target word corresponding to that time step is determined based on the corrected generation probability distribution. The encoded features of the target word are appended to the prefix information corresponding to that time step to generate the generation probability distribution for the next time step.

5. The advertising text optimization method according to claim 1, characterized in that, The generation probability distribution generated at the current time step is weighted and corrected using the reference probability distribution. The target word corresponding to the current time step is determined based on the corrected generation probability distribution. The encoded features of the target word are appended to the prefix information corresponding to the current time step to generate the generation probability distribution for the next time step, including: Starting from the generation probability distribution obtained at the first time step, after correcting the generation probability distribution using a reference probability distribution, multiple target words with probabilities higher than a preset threshold are determined based on the corrected generation probability distribution, and text feature sequences are constructed with the encoding features of each target word as the guide. Each text feature sequence is individually appended to the prefix information. The generation probability distribution of the next time step is predicted based on the prefix information. After correcting the generation probability distribution using a reference probability distribution, multiple target words with probabilities higher than a preset threshold are determined according to the corrected generation probability distribution. The encoding features of each target word are appended to the corresponding text feature sequence to differentiate multiple updated text feature sequences. Continue iterating the above process at the next time step until the preset conditions are met and the process terminates, obtaining the final determined updated text feature sequence for conversion into the corresponding advertising text.

6. The advertising text optimization method according to any one of claims 1 to 5, characterized in that, The advertising creative library includes a private advertising creative library and a public advertising creative library. The private advertising creative library is preset to be accessible only to the current merchant user, while the public advertising creative library is preset to be accessible to all merchant users on the e-commerce platform.

7. An advertising text optimization device, characterized in that, include: The data acquisition module is configured to acquire product images and product description text for the target product; The creative matching module is configured to match multiple creative ads in the advertising creative library based on the product image of the target product, and obtain the creative text in the creative ads, wherein the product image of each creative ad is semantically matched with the product image of the target product; The probability mapping module is set to perform word frequency statistics on all matched creative texts, determine the distribution probability of each word segment, and map it to the reference probability distribution of the reference vocabulary referenced by the corresponding advertising text generation model; The text generation module is configured to use a time-step decoding advertising text generation model to generate a text feature sequence corresponding to the advertising text based on the product description text. In each time step, the generation probability distribution of the target word corresponding to the target position in the text feature sequence is obtained in the reference vocabulary. The target word is determined after the generation probability distribution is corrected by the reference probability distribution. The step of correcting the generation probability distribution generated at the current time step by weighting the reference probability distribution includes: weighting the reference probability distribution with preset weights to obtain a weighted reference probability distribution; calculating the element-wise product of the weighted reference probability distribution and the generation probability distribution at the current time step to obtain a weighted optimized weight distribution; and obtaining the element-wise sum of the generation probability distribution and the weight distribution to obtain the generation probability distribution updated for the current time step.

8. The advertising text optimization device according to claim 7, characterized in that, The regression processing unit includes: an initial processing subunit, configured to start from the generation probability distribution obtained at the first time step, correct the generation probability distribution using a reference probability distribution, determine multiple target words with probabilities higher than a preset threshold based on the corrected generation probability distribution, and construct text feature sequences with the encoding features of each target word as the guide; an iterative processing subunit, configured to append prefix information to each text feature sequence individually, predict the generation probability distribution at the next time step based on the prefix information, correct the generation probability distribution using a reference probability distribution, determine multiple target words with probabilities higher than a preset threshold based on the corrected generation probability distribution, and append the encoding features of each target word to the corresponding text feature sequence to differentiate into multiple updated text feature sequences; and an iterative control subunit, configured to continue iterating the above process at the next time step until a preset condition is met and the process terminates, obtaining the finally determined updated text feature sequence for conversion into the corresponding advertising text.

9. A computer device comprising a central processing unit and a memory, characterized in that, The central processing unit is used to invoke and run a computer program stored in the memory to perform the steps of the method as described in any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that, It stores, in the form of computer-readable instructions, a computer program implemented according to any one of claims 1 to 6, which, when invoked by a computer, executes the steps included in the corresponding method.

Images