Advertisement text generation method and device, equipment and medium

Abstract

This application relates to a method, apparatus, device, and medium for generating advertising text in the field of computer technology. The method includes: acquiring the title of an advertised product and its reference advertising text; using a text generation model's encoder-decoder network to decode the title time-step by time to obtain decoded feature information, and determining its generation probability distribution mapped to a first reference vocabulary referenced by the encoder-decoder network; using a text generation model's encoder to obtain encoded feature information of the comprehensive text composed of the product title and its reference advertising text, and determining its key feature information relative to the decoded feature information as a pointer probability distribution; calculating and determining a matching weight based on the encoded feature information and the decoded feature information, and smoothly weighting the generation probability distribution and the pointer probability distribution with the matching weight to obtain an optimized probability distribution; and determining the lexical units of the target advertising text based on their word selection in the first reference vocabulary or the comprehensive text. This application can generate high-quality advertising text.

Description

Technical Field

[0001] This application relates to the field of e-commerce technology, and in particular to a method for generating advertising text and the corresponding apparatus, computer equipment, and computer-readable storage medium. 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] Traditional advertising text generation technology typically takes information about the product to be advertised as input and generates corresponding advertising text based on a pre-trained and converged encoder-decoder network. However, in the inference and testing phases of advertising text generation, the encoder-decoder network is a black box, and there is very little human intervention to effectively improve the generation effect. Generally, we can only start with the advertising texts used by similar products in their historical advertising campaigns. However, how to fully leverage the advertising creativity of similar product advertising texts to generate high-quality advertising texts remains an urgent problem to be solved. Summary of the Invention

[0004] The primary objective of this application is to solve at least one of the aforementioned problems by providing a method for generating advertising text and corresponding apparatus, computer equipment, and computer-readable storage medium.

[0005] To achieve the various objectives of this application, the following technical solution is adopted:

[0006] An advertising text generation method provided for one of the purposes of this application includes the following steps:

[0007] Obtain the product title of the advertised product and its reference advertising text;

[0008] The decoding feature information of the product title is obtained by decoding the text generation model through an encoding and decoding network step by step, and the generation probability distribution of the decoding feature information mapped to the first reference vocabulary referenced by the encoding and decoding network is determined.

[0009] An encoder using a text generation model obtains the encoded feature information of the comprehensive text consisting of the product title and its reference advertising text, and determines the key feature information of the encoded feature information relative to the decoded feature information as a pointer probability distribution;

[0010] The matching weight is calculated and determined based on the encoded feature information and the decoded feature information. The generated probability distribution and the pointer probability distribution are then smoothly weighted with the matching weight to obtain an optimized probability distribution. The optimized probability distribution includes an indication of the probability of taking a word corresponding to the first reference vocabulary or the comprehensive text.

[0011] The lexical units of the target advertising text are determined by extracting words from the first reference vocabulary or the comprehensive text based on the optimized probability distribution.

[0012] On the other hand, an advertising text generation apparatus provided to suit one of the purposes of this application includes a data acquisition module, a first probability distribution module, a second probability distribution module, a third probability distribution module, and a lexical determination module, wherein: the data acquisition module is used to acquire the product title of the advertised product and its reference advertising text; the first probability distribution module is used to decode the product title step-by-step using an encoder-decoder network of a text generation model to obtain the decoding feature information of the product title, and determine the generation probability distribution of the decoding feature information mapped to the first reference vocabulary referenced by the encoder-decoder network; the second probability distribution module is used to obtain the product title and its reference advertising text using an encoder of a text generation model. The system uses the encoded feature information of the comprehensive text composed of the reference advertising text to determine the key feature information of the encoded feature information relative to the decoded feature information, as a pointer probability distribution; a third probability distribution module is used to calculate and determine the matching weight based on the encoded feature information and the decoded feature information, and to smoothly weight the generated probability distribution and the pointer probability distribution with the matching weight to obtain an optimized probability distribution, wherein the optimized probability distribution includes the probability of taking words corresponding to words from the first reference vocabulary or the comprehensive text; a lexical determination module is used to determine the lexicals of the target advertising text by taking words from the first reference vocabulary or the comprehensive text according to the optimized probability distribution.

[0013] 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 generation method described in this application.

[0014] 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 generation method, which, when invoked by a computer, performs the steps included in the method.

[0015] The technical solution of this application has many advantages, including but not limited to the following aspects:

[0016] This application employs a text generation model. On one hand, it determines the generation probability distribution based on the product title of the advertised product. On the other hand, it determines the pointer probability distribution based on the comprehensive text composed of the product title and its reference advertising text. Then, by using matching weights, it smoothly weights the generation probability distribution and the pointer probability distribution to obtain an optimized probability distribution. Based on the optimized probability distribution, it extracts words from the reference vocabulary or the comprehensive text to determine the lexical units of the target advertising text, thus generating the target advertising text. It is evident that this model can combine a thorough understanding of the advertised product's own information with ample reference to the advertising creativity of the reference advertising text, thereby generating high-quality advertising text for the advertised product. Attached Figure Description

[0017] 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:

[0018] Figure 1 This is a flowchart illustrating a typical embodiment of the advertising text generation method of this application;

[0019] Figure 2 This is a flowchart illustrating the process of filtering reference advertising text that matches the product title of the advertised product in an embodiment of this application;

[0020] Figure 3 This is a flowchart illustrating the process of determining the generation probability distribution of each time step corresponding to the construction of the target advertising text in an embodiment of this application.

[0021] Figure 4 This is a flowchart illustrating the process of constructing the pointer probability distribution for each time step corresponding to the target advertisement text in an embodiment of this application.

[0022] Figure 5 This is a flowchart illustrating the process of constructing the optimized probability distribution of the target advertising text at each time step in an embodiment of this application.

[0023] Figure 6 This is a schematic diagram illustrating the training process of the text generation model in an embodiment of this application;

[0024] Figure 7 This is a schematic block diagram of the advertising text generation device of this application;

[0025] Figure 8 This is a schematic diagram of the structure of a computer device used in this application. Detailed Implementation

[0026] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0027] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this application means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any units and all combinations of one or more associated listed items.

[0028] Those skilled in the art will understand that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0029] Those skilled in the art will understand that the terms "client," "terminal," and "terminal device" as used herein include both devices that receive wireless signals, devices that only possess wireless signal receiver capabilities without transmission capabilities, and devices with receiving and transmitting hardware, devices that have receiving and transmitting hardware capable of bidirectional communication over a bidirectional communication link. Such devices may include: cellular or other communication devices such as personal computers or tablets, having single-line displays, multi-line displays, or cellular or other communication devices without multi-line displays; PCS (Personal Communications Service) that can combine voice, data processing, fax, and / or data communication capabilities; PDAs (Personal Digital Assistants) that may include radio frequency receivers, pagers, internet / intranet access, web browsers, notebooks, calendars, and / or GPS (Global Positioning System) receivers; and conventional laptops and / or handheld computers or other devices that have and / or include radio frequency receivers. As used herein, "client," "terminal," and "terminal device" can be portable, transportable, installed in a means of transportation (air, sea, and / or land), or suitable and / or configured to operate locally and / or in a distributed manner, operating in any other location on Earth and / or in space. "Client," "terminal," and "terminal device" as used herein can also be a communication terminal, an internet access terminal, or a music / video playback terminal, such as a PDA, a MID (Mobile Internet Device), and / or a mobile phone with music / video playback capabilities, or a smart TV, set-top box, etc.

[0030] The hardware referred to by the names "server," "client," and "service node" in this application is essentially an electronic device with the equivalent capabilities of a personal computer. It is a hardware device with the necessary components revealed by the von Neumann architecture, such as a central processing unit (including an arithmetic logic unit and a control unit), memory, input devices, and output devices. The computer program is stored in its memory, and the central processing unit loads the program stored in the secondary storage into the main memory to run it, execute the instructions in the program, and interact with the input and output devices to complete specific functions.

[0031] It should be noted that the concept of "server" used in this application can also be extended to the case of server clusters. Based on the network deployment principles understood by those skilled in the art, the servers should be logically divided. Physically, these servers can be independent of each other but accessible through interfaces, or they can be integrated into a single physical computer or a computer cluster. Those skilled in the art should understand this flexibility and should not use it to constrain the implementation of the network deployment method in this application.

[0032] Unless otherwise expressly specified, one or more technical features of this application may be deployed on a server and accessed by a client through remote invocation of the online service interface provided by the server, or they may be directly deployed and run on a client for access.

[0033] 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.

[0034] Unless otherwise specified, all data involved in this application may be stored remotely on a server or on a local terminal device, as long as it is suitable for use by the technical solution of this application.

[0035] Those skilled in the art will understand that although the various methods in this application are described based on the same concept and thus present commonality among them, they can be performed independently unless otherwise specified. Similarly, the various embodiments disclosed in this application are all based on the same inventive concept; therefore, concepts expressed in the same way, as well as concepts that are appropriately changed for convenience but are expressed differently, should be understood equivalently.

[0036] 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.

[0037] The advertising text generation method of this application can be programmed into a computer program product and deployed on a client or server for execution. For example, in an exemplary application scenario of this application, it can be deployed on the server of an e-commerce platform, thereby allowing human-computer interaction with the process of the computer program product through a graphical user interface by accessing the interface opened after the computer program product is running.

[0038] Please see Figure 1 The advertising text generation method of this application, in its typical embodiment, includes the following steps:

[0039] Step S1100: Obtain the product title of the advertised product and its reference advertising text;

[0040] When an online store user needs to generate advertising text for a product to be advertised, in one embodiment, an ad generation page can be used to specify the product corresponding to the advertising text. Typically, the advertising product can be a product from the online store's product 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 is stored in the product database, and the user can specify the advertising product through the ad generation page.

[0041] When users of an online store need to specify the advertised product, they can usually determine it by specifying the product's unique identifier, i.e., the product ID. Based on the product ID, the product information of the advertised product can be retrieved from the product database.

[0042] The product information includes, but is not limited to, different types of data such as images and text. Images in the product information include those uploaded by users of online stores on e-commerce platforms when listing products for display purposes. These images showcase the product from an overall and multi-faceted perspective, including the main product image and product detail images. The main product image is typically used as the primary image for displaying the product, as it usually shows the complete picture of the product. In this application, it is recommended to use the main product image of the advertised product as the product image used for subsequent processing. The product title in the product information is the text entered by users of online stores on e-commerce platforms when listing products as a title.

[0043] It can collect creative ads that users of the current online store have historically placed on the advertising system provided by the e-commerce platform. In addition, it can collect creative ads that have historically placed from third-party advertising systems. Each creative ad contains a product image and its creative text. The creative text is a descriptive text used to display the selling points of the corresponding product. The product image can be the image that best showcases the corresponding product, such as the main product image.

[0044] The product images from historical creative advertisements delivered from a third-party advertising system can be pre-encoded using an image encoder to obtain corresponding image encoding vectors. Similarly, the creative text of the creative advertisements can be pre-encoded using a text encoder to obtain corresponding text encoding vectors. The image encoding vectors and text encoding vectors of the corresponding creative advertisements are then stored together for later retrieval. The image encoder can be any of the following models: a pre-trained convergent ViT (Vision Transformer) model, a deep convolutional model such as EfficientNet, DenseNet, or ResNet. The text encoder can be any of the following models: a pre-trained convergent model capable of vectorizing text, such as BERT, RoBERTa, XLM-RoBERTa, or MPNet.

[0045] In one embodiment, similarly, the image encoder is used to encode the product image of the advertised product selected by the user in the online store to obtain the corresponding image encoding vector, and the text encoder is used to encode the product title of the advertised product to obtain the corresponding text encoding vector.

[0046] Further, the image encoding vector of the advertised product image is calculated, and the vector distance between it and the image encoding vectors of product images in various creative advertisements historically placed on the e-commerce platform's advertising system and third-party advertising systems is used as the first similarity. The vector distance between the text encoding vector of the advertised product title and the text encoding vector of the creative text in each creative advertisement is calculated as the second similarity. For each creative advertisement, its first and second similarities are multiplied by their respective preset weights and then added together to calculate the corresponding first comprehensive similarity. Based on this, creative advertisements with a first comprehensive similarity exceeding the highest one (excluding historically placed creative advertisements of the advertised product) are selected, and their creative text is obtained as reference advertisement text matching the product title of the advertised product. The preset weights of the first and second similarities can be flexibly set by those skilled in the art, as long as the sum of their preset weights is 1.

[0047] In a further embodiment, based on calculating the image similarity and text similarity of each creative advertisement historically placed on the advertising system provided by the e-commerce platform and the third-party advertising system, similarly, the creative text in the historical creative advertisements of the advertised product is encoded using the text encoder to obtain the corresponding text encoding vector. The vector distance between the text encoding vector of the creative text of the advertised product and the text encoding vector of the creative text in each creative advertisement is calculated as the third similarity. For each creative advertisement, its first similarity, second similarity, and third similarity are multiplied by their respective preset weights and then added together to calculate the corresponding second comprehensive similarity. Based on this, creative advertisements other than those historically placed creative advertisements of the advertised product with a second comprehensive similarity exceeding the highest are selected from the creative advertisements, and their creative text is obtained as reference advertising text that matches the product title of the advertised product. The preset weights of the first similarity, second similarity, and third similarity can be flexibly set by those skilled in the art, as long as the sum of the preset weights of the three is 1.

[0048] When calculating the aforementioned vector distance, it can be implemented using any of the large-scale vector retrieval engines such as Faiss, ElasticSearch, and Milvus, or any existing algorithm such as cosine similarity, dot product, Manhattan distance, or Euclidean distance can be used.

[0049] Therefore, it is easy to understand that when a user of an online store specifies an advertised product, the product title and image of that product can be obtained from the product database. Further referring to the implementation of the above embodiment, a reference advertising text matching the product title can be determined. Based on this, the product title and its reference advertising text can be obtained.

[0050] Step S1200: Using a text generation model's encoding and decoding network, decode the product title time-step by time to obtain the decoding feature information, and determine the generation probability distribution of the decoding feature information mapped to the first reference vocabulary referenced by the encoding and decoding network;

[0051] The encoding / decoding network of the text generation model can adopt the Transformer model, which includes both an encoder and a decoder. Word embedding is performed on the product title based on the first reference vocabulary referenced by the text generation model's encoding / decoding network to determine the encoding features of each word in the product title within the first reference vocabulary, thereby obtaining the corresponding first encoding vector. Under the action of the encoder of the encoding-decoding network, deep semantic information is extracted based on the first encoded vector to obtain the corresponding first encoded vector sequence. Further, the decoder performs decoding step by step based on its own autoregressive mechanism. For the decoding of each time step, the decoding is performed according to the first encoded vector sequence and the decoding results of each time step before the current time step. Specifically, the first encoded vector sequence is used as prefix information and input to the decoder of the encoding-decoding network for decoding to obtain the decoding feature sequence of the last hidden layer of the decoder as decoding feature information. The decoding feature information is mapped to the vocabulary classifier through a fully connected layer to obtain the corresponding generation probability distribution. After selecting the corresponding word of the current time step as the decoding result according to the generation probability distribution using a preset decoding strategy, the encoding feature of the word in the first reference vocabulary is appended to the prefix information and input to the encoder for decoding of the next time step. The prediction of the generation probability of the next time step continues until the end symbol is encountered.

[0052] The generated probability distribution represents the classification probability of each word in the first reference vocabulary. The decoding strategy may include, but is not limited to, Top_N algorithm, Top_K algorithm, beam search algorithm, and greedy search algorithm; those skilled in the art can choose one to implement as needed. 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 probability distribution of the corresponding hidden layer vector mapped to each word in the entire first reference vocabulary.

[0053] Based on the above, it can be understood that in the process of decoding the product title step by step, the decoding feature information corresponding to the last hidden layer of the decoder can be obtained for each time step, and the generation probability distribution of the decoding feature information mapped to the first reference vocabulary can be determined.

[0054] Step S1300: The encoder of the text generation model obtains the encoded feature information of the comprehensive text composed of the product title and its reference advertising text, and determines the key feature information of the encoded feature information relative to the decoded feature information as a pointer probability distribution.

[0055] The product title and its reference advertising text are concatenated to form a composite text, for example: "Product Title [SEP] Reference Advertising Text". Word embedding is performed on the composite text using a second reference vocabulary referenced by the encoder of the text generation model. This determines the encoding features of each word in the composite text within the second reference vocabulary, obtaining the corresponding second encoding vector. Under the action of this encoder, deep semantic information is extracted based on the second encoding vector, obtaining the second encoding vector sequence of the last hidden layer of the encoding layer as encoding feature information. It can be understood that each element in the encoding feature information is the encoded representation of each word in the corresponding composite text.

[0056] The encoder of the text generation model is an encoder suitable for processing long text encoding, such as Longformer, BigBird, Reformer, etc. Those skilled in the art can choose one of them as needed.

[0057] Furthermore, the decoding feature information and the encoding feature information corresponding to each time step in the decoding process of the text generation model's encoding and decoding network are input into the attention layer to perform feature interaction based on the additive attention mechanism, so as to obtain the key feature information of the comprehensive text corresponding to each time step as a pointer probability distribution.

[0058] Step S1400: Calculate and determine the matching weight based on the encoded feature information and the decoded feature information, and smoothly weight the generated probability distribution and the pointer probability distribution with the matching weight to obtain an optimized probability distribution. The optimized probability distribution includes the probability of taking a word corresponding to the first reference vocabulary or the comprehensive text.

[0059] Based on obtaining the key feature information of the comprehensive text corresponding to each time step, the key feature information is used to weight each element in the encoded feature information, that is, the encoded representation of each word in the last hidden layer of the encoder in the corresponding comprehensive text. The weighted elements are summed to obtain the context vector of the comprehensive text.

[0060] Furthermore, a matching weight is calculated and determined based on the decoded feature information and the context vector. This matching weight is specifically represented as a probability with a value range of [0,1]. This probability determines the probability of generating a word from the first reference vocabulary, rather than the probability of copying a single word from the comprehensive text. Therefore, by subtracting the matching weight from 1, the probability of copying a single word from the comprehensive text, rather than generating a word from the first reference vocabulary, can be determined. In this way, the generation probability distribution can be weighted with the matching weight, and the pointer probability distribution can be weighted with 1 minus the matching weight. The weighted distributions are then summed to obtain an optimized probability distribution.

[0061] Step S1500: Determine the lexical units of the target advertising text by extracting words from the first reference word list or the comprehensive text according to the optimized probability distribution.

[0062] Based on the disclosure of steps S1100-1400, it is not difficult to understand that during the time-step decoding process of the encoding and decoding network of the text generation model, the corresponding generation probability distribution, pointer probability distribution, and matching weight can be determined for each time step, and then the corresponding optimized probability distribution can be obtained.

[0063] It can be seen that, under the effect of the matching weight in the optimized probability distribution, for each word predicted by the text generation model, the closer the matching weight is to 1, the stronger the attention of the generation probability distribution to the prediction, that is, the greater the probability of taking a word from the first reference vocabulary to determine the predicted word; the closer the matching weight is to 0, the stronger the attention of the pointer probability distribution to the prediction, that is, the greater the probability of copying a word from the comprehensive text to determine the predicted word.

[0064] Based on the foregoing, according to the optimized probability distribution corresponding to each time step, a preset decoding strategy is adopted to extract words from the first reference vocabulary or the comprehensive text to determine the corresponding word units as word units of the target advertising text, and construct the target advertising text. The decoding strategy may include, but is not limited to: Top_N algorithm, Top_K algorithm, cluster search algorithm, greedy search algorithm, and those skilled in the art can select one to implement as needed.

[0065] In one embodiment, a greedy search algorithm is used to find a local optimum to a certain extent. Specifically, for each time step, the word with the highest probability is determined according to its optimization probability distribution. Then, the word determined for each time step is used as the word of the target advertisement text to construct the target advertisement text.

[0066] As can be seen from the typical embodiments of this application, the technical solution of this application has many advantages, including but not limited to the following aspects:

[0067] This application employs a text generation model. On one hand, it determines the generation probability distribution based on the product title of the advertised product. On the other hand, it determines the pointer probability distribution based on the comprehensive text composed of the product title and its reference advertising text. Then, by using matching weights, it smoothly weights the generation probability distribution and the pointer probability distribution to obtain an optimized probability distribution. Based on the optimized probability distribution, it extracts words from the reference vocabulary or the comprehensive text to determine the lexical units of the target advertising text, thus generating the target advertising text. It is evident that this model can combine a thorough understanding of the advertised product's own information with ample reference to the advertising creativity of the reference advertising text, thereby generating high-quality advertising text for the advertised product.

[0068] Please see Figure 2 In a further embodiment, step S1100, the step of obtaining the product title of the advertised product and its reference advertising text, includes the following steps:

[0069] Step S1110: Obtain the product title and product image of the advertised product;

[0070] When users of an online store need to generate advertising text for a specific product, they can usually determine this by specifying the product's unique identifier, i.e., the product ID. Based on this product ID, the product title and product image of the advertised product can be retrieved from the online store's product database.

[0071] This application provides an advertising creative library, which comprises two sub-libraries: a private advertising creative library and a public advertising creative library. The private advertising creative library stores creative advertisements corresponding to products from a user's online store that have been published on the e-commerce platform's advertising system. The public advertising creative library stores creative advertisements that have been placed on third-party advertising systems. Accordingly, the private advertising creative library is preset to be accessible only to users of the current online store, while the public advertising creative library is preset to be accessible to all users within the e-commerce platform. Each creative advertisement includes a product image and its corresponding creative text. The creative text is a descriptive text showcasing the product's selling points, and the product image can be the most representative image of the product, such as the main product image. This implementation effectively isolates the data of creative advertisements from different online stores, ensuring the privacy of each online store's creative advertisements.

[0072] Furthermore, by mapping each creative ad in the private and public creative ad libraries to a corresponding index tag determined by a unified algorithm, the corresponding creative ad can be quickly identified based on the index tag for subsequent retrieval. The index tag can be a natural number or the result of encoding the corresponding creative ad using hash functions such as MD5 and SHA1, or algorithms such as Base64.

[0073] In the advertising creative library of this application, the product image in each creative advertisement is pre-extracted with a corresponding image encoding vector using a preset image encoder, and the creative text in each creative advertisement is pre-extracted with a corresponding text encoding vector using a preset text encoder. The image encoding vector and text encoding vector corresponding to each creative advertisement are then stored along with the index tag of the corresponding creative advertisement for later retrieval. The image encoder can be any model such as a pre-trained convergent ViT (Vision Transformer) model, a deep convolutional model such as EfficientNet, DenseNet, or ResNet. The text encoder can be any model such as a pre-trained convergent model capable of vectorizing text, such as BERT, RoBERTa, XLM-RoBERTa, or MPNet.

[0074] Step S1120: Use a text encoder to extract the deep semantic information of the product title of the advertised product and obtain the corresponding text encoding vector; use an image encoder to extract the deep semantic information of the product image of the advertised product and obtain the corresponding image encoding vector.

[0075] Similarly, for the product images of the advertised products, the image encoder is also used to extract deep semantic information based on image convolution to obtain the corresponding vectorized representation, i.e., the image encoding vector. For the product titles of the advertised products, the text encoder is also used to extract deep semantic information based on bidirectional feature encoding to obtain the corresponding vectorized representation, i.e., the text encoding vector.

[0076] Step S1130: Calculate the first similarity between the image encoding vector of the product image of the advertised product and the image encoding vector of the product image in each creative advertisement in the advertising creative library, and calculate the second similarity between the text encoding vector of the product title of the advertised product and the text encoding vector of the reference advertisement text in each creative advertisement in the advertising creative library.

[0077] Generally, the similarity between two objects can be represented by the vector distance between their respective vectors. Based on this, the vector distance between the image encoding vector of the advertised product's image and the image encoding vectors of the product images of various creative advertisements in the advertising creative library is used as the first similarity. Similarly, the vector distance between the text encoding vector of the advertised product's title and the text encoding vectors of the product titles of various creative advertisements in the advertising creative library is used as the second similarity.

[0078] When calculating the aforementioned vector distance, it can be implemented using any of the large-scale vector retrieval engines such as Faiss, ElasticSearch, and Milvus, or any existing algorithm such as cosine similarity, dot product, Manhattan distance, or Euclidean distance can be used.

[0079] Step S1140: Select reference advertising texts that match the product title from the advertising creative library based on the first similarity and the second similarity.

[0080] Weights can be preset for the first and second similarities, and the sum of their weights is 1. For example, if image similarity better represents the similarity between the advertised product and the creative advertisement than text similarity, the weight of the first similarity can be set to be greater than that of the second similarity, such as 0.6 for the first similarity and 0.4 for the second similarity. Conversely, the weight of the first similarity can be set to be less than that of the second similarity. Furthermore, if image similarity and text similarity equally represent the similarity between the advertised product and the creative advertisement, the weights of the first and second similarities can be set to be equal, both at 0.5. Referring to the disclosure herein, those skilled in the art can set the corresponding weights for the first and second similarities as needed.

[0081] The first similarity and the second similarity are multiplied by their respective weights and then added together to calculate the first comprehensive similarity. Based on this, the first comprehensive similarity corresponding to each creative advertisement in the advertising creative library can be obtained. Then, creative advertisements with a first comprehensive similarity exceeding the highest one, excluding the creative advertisements that were historically placed for the advertised product, are selected, and their creative texts are obtained as reference advertising texts that match the product title of the advertised product.

[0082] In this embodiment, by determining the first similarity between the product image of the advertised product and the product images in each creative advertisement, and the second similarity between the product title of the advertised product and the creative text in each creative advertisement, and then combining the similarity of the two dimensions to obtain a comprehensive similarity, the confidence level of the similarity between the advertised product and each creative advertisement is ensured. This allows for the accurate identification of creative advertisements that match the product title of the advertised product from each creative advertisement based on their corresponding comprehensive similarity. The creative text in these advertisements is then used as reference advertisement text, laying the foundation for the subsequent generation of high-quality target advertisement text.

[0083] Please see Figure 3 In a further embodiment, step S1200, using a text generation model's codec network to decode the product title time-step by time to obtain the decoding feature information, and determining the generation probability distribution of the decoding feature information mapped to the first reference vocabulary referenced by the codec network, includes the following steps:

[0084] Step S1210: Encode the product title according to the first reference vocabulary referenced by the encoding and decoding network of the text generation model to obtain the first encoding vector;

[0085] To adapt to the input requirements of the encoding / decoding network of the text generation model, the product title needs to be encoded. For this purpose, a first reference vocabulary referenced by the encoding / decoding network can be used to convert each word segment in the product title into the encoding features of the corresponding word units in the first reference vocabulary. These encoding features are then constructed into corresponding first encoding vectors, which are input into the encoding / decoding network for inference. In some embodiments, the positional encoding of each word unit can be further combined with the encoding features of the word units to enrich the reference information of the encoding vector.

[0086] Step S1220: The encoder of the encoding and decoding network is used to extract the deep semantic information of the first encoded vector to obtain the corresponding first encoded vector sequence;

[0087] The encoder of the encoder-decoder network is suitable for learning vector representation and multi-layer semantic space transformation 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.

[0088] The encoder of the aforementioned encoder-decoder network extracts deep semantic information by referencing the contextual information of the input first encoded vector, thereby achieving a deep semantic representation of the input first encoded vector and obtaining a first encoded vector sequence. A self-attention mechanism can also be applied to the encoder to make the features that require special attention explicit, making the obtained first encoded vector sequence more helpful for subsequent decoding.

[0089] Step S1230: Using the decoder in the encoding / decoding network, the first encoding vector sequence is used as input to obtain the corresponding decoding feature sequence as decoding feature information. The decoding feature information is then autoregressively decoded step by step to obtain the generation probability distribution at each time step, wherein each generation probability distribution includes the probability of each word in the first reference vocabulary.

[0090] The decoder of the encoder-decoder network is adapted to learn and adjust the state information of the hidden layer at the current time step by utilizing the first encoded vector sequence preceding the current time step 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 classic models such as RNN, LSTM, and GRU can also be used.

[0091] The decoder in the aforementioned codec network uses the prefix information formed by the first encoded vector sequence as initial input. Based on the self-attention mechanism, it performs serialized decoding, decoding step by step, to obtain the decoding feature sequence of the last hidden layer of the decoder as decoding feature information. The decoding feature information is then mapped through a vocabulary classifier to obtain the probability distribution of each word mapped to the first reference vocabulary, i.e., the generated probability distribution. Based on the generated probability distribution, the word corresponding to the current time step can be selected. The encoding feature of this word in the first reference vocabulary is then appended to the prefix information to achieve autoregression. Then, the decoding continues at the next time step based on the updated prefix information until the end symbol is encountered.

[0092] Based on the above, it can be understood that in the process of decoding the product title step by step, the decoding feature information corresponding to the last hidden layer of the decoder can be obtained for each time step, and the generation probability distribution of the decoding feature information mapped to the first reference vocabulary can be determined.

[0093] In this embodiment, the encoding and decoding network of the text generation model refers to the contextual information of the product title of the advertised product. During the time-step autoregressive decoding process, the generation probability distribution corresponding to each time step is predicted, so that the accuracy of the generation probability distribution can be guaranteed based on a full understanding of the semantics of the product title.

[0094] Please see Figure 4 In a further embodiment, step S1300, which involves using an encoder of a text generation model to obtain the encoded feature information of the comprehensive text consisting of the product title and its reference advertising text, and determining the key feature information of the encoded feature information relative to the decoded feature information as a pointer probability distribution, includes the following steps:

[0095] Step S1310: Combine the product title with its reference advertising text to form a comprehensive text;

[0096] The encoder, employing a text generation model, can recognize separators such as [SEP]. These separators are then concatenated with the product title and its reference advertising text to form a composite text, exemplified by "Product Title [SEP] Reference Advertising Text". This approach facilitates the encoder's focus on closely related terms between the product title and the reference advertising text within the composite text, and also allows for deeper integration of the product title and reference text during the encoding stage.

[0097] Step S1320: Encode the synthesized text according to the second reference vocabulary referenced by the encoder of the text generation model to obtain the second encoding vector;

[0098] To adapt to the encoder input requirements of the text generation model, the synthesized text needs to be encoded. For this purpose, a second reference lexicon referenced by the encoder can be used to convert each word segment in the synthesized text into the encoding features of the corresponding word units in the second reference lexicon. These encoding features are then constructed into corresponding second encoding vectors, which are input into the encoder for inference. In some embodiments, the positional encoding of each word unit can be further combined with the encoding features of the word units to enrich the reference information of the encoding vectors.

[0099] Step S1330: Use the encoder to extract the deep semantic information of the second encoding vector and obtain the corresponding second encoding vector sequence as encoding feature information;

[0100] The encoder of the text generation model is an encoder suitable for processing long text encoding, such as Longformer, BigBird, Reformer, etc. Those skilled in the art can choose one of them as needed.

[0101] The encoder of the aforementioned encoder-decoder network extracts deep semantic information by referencing the contextual information of the input second encoded vector, thereby achieving a deep semantic representation of the input second encoded vector. The resulting sequence of second encoded vectors from the last hidden layer of the encoder is used as encoded feature information. Each element in this encoded feature information represents the encoded representation of each word in the corresponding synthesized text. A self-attention mechanism can also be applied to the encoder to make the features requiring special attention explicit, resulting in a more accurate feature representation of the obtained second encoded vector sequence.

[0102] Step S1340: Input the decoding feature information and the encoding feature information corresponding to each time step in the decoding process of the text generation model's encoding and decoding network into the attention layer for feature interaction, and obtain the corresponding key feature information as pointer probability distribution.

[0103] For each time step in the decoding process of the text generation model's encoder-decoder network, the decoding feature information of the last hidden layer of the corresponding decoder and the encoding feature information of the last hidden layer of the encoder are input into the attention layer for feature interaction based on an additive attention mechanism. This yields key feature information of the comprehensive text corresponding to each time step, serving as a pointer probability distribution. Specifically, the key feature information is calculated using additive attention based on the decoding and encoding feature information. An illustrative formula is shown below:

[0104]

[0105]

[0106] in: The key feature information is the pointer probability distribution. , These are all parameters that can be learned. This refers to the decoding feature information of the last hidden layer of the decoder in the encoder-decoder network of the text generation model. This represents the encoded feature information of the last hidden layer of the encoder in the text generation model at the current time step. and These are all vectors with the same number of terms as the total text. The softmax function is used to... Normalized to a vector that sums to 1 , Each element in the array has a value range of [0,1]. The closer the value is to 1, the higher the probability that the corresponding word in the comprehensive text will be used as the output of the current time step.

[0107] In this embodiment, by performing feature interaction between the decoding feature information of the last hidden layer of the decoder of the text generation model's encoder-decoder network and the encoding feature information of the last hidden layer of the encoder of the text generation model at each time step during the decoding process of the encoder-decoder network of the text generation model, based on the additive attention mechanism, key feature information can be effectively summarized in the comprehensive text with high linear complexity. In addition, the computational complexity is lower than that of the general dot product attention mechanism, and the execution efficiency is high.

[0108] Please see Figure 5 In a further embodiment, step S1400, which involves calculating and determining a matching weight based on the encoded feature information and the decoded feature information, and then smoothly weighting the generated probability distribution and the pointer probability distribution with the matching weight to obtain an optimized probability distribution, includes the following steps:

[0109] Step S1410: Weight the encoded feature information with the key feature information to obtain the context vector of the comprehensive text;

[0110] Example formulas are shown below:

[0111]

[0112] in: In the encoding feature information of the last hidden layer of the encoder of the text generation model, the corresponding encoded representation of the i-th word in the comprehensive text is given. This is to synthesize the key feature information corresponding to the i-th word in the text.

[0113] Step S1420: Calculate and determine the matching weight based on the decoded feature information and the context vector, and then sum the weights of the generation probability distribution and the pointer probability distribution to obtain the optimized probability distribution.

[0114] An example formula for calculating the matching weight is shown below:

[0115]

[0116] Where p_gen is the matching weight, and w and bais are learnable parameters.

[0117] An example of the exemplary formula for the optimized probability distribution is as follows:

[0118]

[0119] in, For the predicted word units, for The probability, To generate a probability distribution, For the integrated text and The probability distribution of pointers corresponding to all identical lexical terms.

[0120] In this embodiment, by smoothly weighting the generation probability distribution and the pointer probability distribution with matching weights, the attention of both to the final optimized probability distribution is rationally planned. This allows the determination of the target ad text's lexical units based on the optimized probability distribution to either copy lexical units from the comprehensive text based on the pointer probability distribution or determine lexical units from the first reference vocabulary based on the generation probability distribution. This effectively solves the problem of the text generation model generating lexical units in the comprehensive text that are outside the first reference vocabulary, while also making full use of the reference ad text from historical ad placements.

[0121] Please see Figure 6 In a further embodiment, the training process of the text generation model includes the following steps:

[0122] Step S2100: Obtain a single training sample from the prepared training set. Each training sample in the training set contains the product title and its reference advertising text. Each training sample is labeled with the advertising text used by the product in the past when advertising was placed as a supervision label.

[0123] Obtain sufficient product titles from online stores that have historically run advertisements on e-commerce platforms. Further, referring to steps S1110-1140, determine the reference advertising text matching each product title. Thus, construct a training set using each product title and its reference text as a separate training sample. Additionally, obtain the advertising text used by each training sample's products during historical advertising campaigns and label it accordingly as the supervision label for each training sample.

[0124] Step S2200: Using the text generation model's encoding and decoding network, decode the product title of the training sample step by step to obtain the decoding feature information, and determine the generation probability distribution of the decoding feature information mapped to the first reference vocabulary referenced by the encoding and decoding network;

[0125] This can be implemented with reference to step S1200, which will not be elaborated here. During the process of decoding the product titles of the training samples step by step, the decoding feature information corresponding to the last hidden layer of the decoder in the encoder-decoder network can be obtained at each time step, and the generation probability distribution of this decoding feature information mapped to the first reference vocabulary can be determined.

[0126] Step S2300: The encoder of the text generation model obtains the encoding feature information of the comprehensive text composed of the product title of the training sample and its reference advertising text, and determines the key feature information of the encoding feature information relative to the decoding feature information as a pointer probability distribution.

[0127] This can be implemented by referring to step S1300, which will not be described in detail here.

[0128] Step S2400: Calculate and determine the matching weight based on the encoded feature information and the decoded feature information, and smoothly weight the generated probability distribution and the pointer probability distribution with the matching weight to obtain an optimized probability distribution. Determine the word units of the target advertisement text by taking words from the first reference word list or the comprehensive text according to the optimized probability distribution, and construct the target advertisement text. The optimized probability distribution includes the probability corresponding to taking words from the first reference word list or the comprehensive text.

[0129] This can be achieved by referring to steps S1400-1500, which will not be elaborated upon here. The goal is to obtain the target advertising text corresponding to the training samples.

[0130] Step S2500: Determine the loss value of the target advertisement text based on the supervised labels of the training samples. If the loss value does not reach a preset threshold, update the weights of the text generation model and continue to call other training samples to perform iterative training until the model converges.

[0131] The preset cross-entropy loss function is invoked. This function can be flexibly set by those skilled in the art based on prior knowledge or experimental experience. The cross-entropy loss value of the target advertisement text is calculated based on the supervision labels of the training samples. When the loss value reaches a preset threshold, it indicates that the text generation model has been trained to a convergent state, and the model training can be terminated. When the loss value does not reach the preset threshold, it indicates that the model has not converged. Therefore, the model is updated with gradients based on the loss value. Usually, the weight parameters of each part of the model are corrected through backpropagation to make the model closer to convergence. Then, the next sample data in the training set is called to iteratively train the model until the model is trained to a convergent state.

[0132] In this embodiment, by supervising the convergence of the text generation model, the text generation model learns to accurately determine the corresponding optimized probability distribution based on the product title and its reference advertising text, and thus constructs high-quality target advertising text by selecting words according to the optimized probability distribution.

[0133] Please see Figure 7 This application provides an advertising text generation apparatus to meet one of its objectives. It is a functional embodiment of the advertising text generation method of this application. The apparatus includes a data acquisition module 1100, a first probability distribution module 1200, a second probability distribution module 1300, a third probability distribution module 1400, and a lexical determination module 1500. Specifically: the data acquisition module 1100 is used to acquire the product title of the advertised product and its reference advertising text; the first probability distribution module 1200 is used to decode the product title step-by-step using a text generation model's encoding / decoding network to obtain decoding feature information, and to determine the generation probability distribution of the decoding feature information mapped to the first reference vocabulary referenced by the encoding / decoding network; the second probability distribution module 1300... An encoder using a text generation model obtains encoded feature information of the comprehensive text consisting of the product title and its reference advertising text, and determines key feature information of the encoded feature information relative to the decoded feature information as a pointer probability distribution; a third probability distribution module 1400 is used to calculate and determine matching weights based on the encoded feature information and the decoded feature information, and smoothly weights the generation probability distribution and the pointer probability distribution with the matching weights to obtain an optimized probability distribution, the optimized probability distribution including the probability of taking words corresponding to words from the first reference vocabulary or the comprehensive text; a lexical determination module 1500 is used to determine the lexicals of the target advertising text by taking words from the first reference vocabulary or the comprehensive text according to the optimized probability distribution.

[0134] In a further embodiment, the data acquisition module 1100 includes: a data acquisition submodule, used to acquire the product title and product image of the advertised product; an image encoding submodule, used to extract deep semantic information of the product title of the advertised product using a text encoder to obtain a corresponding text encoding vector, and to extract deep semantic information of the product image of the advertised product using an image encoder to obtain a corresponding image encoding vector; a similarity calculation submodule, used to calculate a first similarity between the image encoding vector of the product image of the advertised product and the image encoding vector of the product image in each creative advertisement in the advertising creative library, and to calculate a second similarity between the text encoding vector of the product title of the advertised product and the text encoding vector of the reference advertisement text in each creative advertisement in the advertising creative library; and a filtering and matching submodule, used to filter reference advertisement text that matches the product title from the advertising creative library based on the first similarity and the second similarity.

[0135] In a further embodiment, the first probability distribution module 1200 includes: a first text encoding submodule, configured to encode the product title according to a first reference vocabulary referenced by the encoding / decoding network of the text generation model to obtain a first encoding vector; a first semantic extraction submodule, configured to extract deep semantic information of the first encoding vector using the encoder of the encoding / decoding network to obtain a corresponding first encoding vector sequence; and an autoregressive decoding submodule, configured to use the decoder in the encoding / decoding network with the first encoding vector sequence as input to obtain a corresponding decoding feature sequence as decoding feature information, and autoregressively decode the decoding feature information step by step to obtain a generation probability distribution for each time step, wherein each generation probability distribution includes the probability of each word in the first reference vocabulary.

[0136] In a further embodiment, the second probability distribution module 1300 includes: a text splicing submodule, used to splice the product title and its reference advertising text to form a comprehensive text; a second text encoding submodule, used to encode the comprehensive text according to the second reference vocabulary referenced by the encoder of the text generation model to obtain a second encoding vector; a second semantic extraction submodule, used to extract the deep semantic information of the second encoding vector using the encoder to obtain the corresponding second encoding vector sequence as encoding feature information; and a feature interaction submodule, used to input the decoding feature information and the encoding feature information corresponding to each time step in the decoding process of the encoder-decoder network of the text generation model to the attention layer for feature interaction to obtain the corresponding key feature information as a pointer probability distribution.

[0137] In a further embodiment, the third probability distribution module 1400 includes: a weighted summation submodule, used to perform weighted summation on the encoded feature information with the key feature information to obtain the context vector of the comprehensive text; and a probability distribution optimization submodule, used to calculate and determine the matching weight based on the decoded feature information and the context vector, and to sum the generated probability distribution and the pointer probability distribution with the matching weight respectively to obtain an optimized probability distribution.

[0138] In a further embodiment, the lexical determination module 1500 includes: a sampling word extraction unit, used to extract words from the first reference word list or the comprehensive text according to an optimized probability distribution and a preset sampling strategy to determine the lexical units of the target advertising text.

[0139] In a further embodiment, the training process of the text generation model includes: a sample acquisition module, used to acquire a single training sample in a prepared training set, wherein each training sample in the training set contains a product title and its reference advertising text, and each training sample is correspondingly labeled with the advertising text used by the product in historical advertising as a supervision label; a first probability distribution determination module, used to use the encoder-decoder network of the text generation model to decode the product title of the training sample step by step to obtain the decoding feature information, and determine the generation probability distribution of the decoding feature information mapped to the first reference vocabulary referenced by the encoder-decoder network; and a second probability distribution determination module, used to use the encoder of the text generation model to obtain the encoding feature information of the comprehensive text composed of the product title and its reference advertising text of the training sample, and determine the encoding feature information. The key feature information relative to the decoded feature information is used as a pointer probability distribution; the advertising text construction module is used to calculate and determine the matching weight based on the encoded feature information and the decoded feature information, and smoothly weight the generation probability distribution and the pointer probability distribution with the matching weight to obtain an optimized probability distribution. Based on the optimized probability distribution, words are taken from the first reference vocabulary or the comprehensive text to determine the lexical units of the target advertising text, and the target advertising text is constructed. The optimized probability distribution includes the probability corresponding to the words taken from the first reference vocabulary or the comprehensive text; the iterative training module is used to determine the loss value of the target advertising text based on the supervision label of the training samples, update the weights of the text generation model when the loss value does not reach a preset threshold, and continue to call other training samples to perform iterative training until the model converges.

[0140] To address the aforementioned technical problems, embodiments of this application also provide computer equipment. For example... Figure 8The diagram shows the internal structure of a computer device. 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 the computer-readable instructions are executed by the processor, they enable the processor to implement an advertising text generation method. The processor of the computer device provides computing and control capabilities, supporting the operation of the entire computer device. The memory of the computer device may store computer-readable instructions, which, when executed by the processor, enable the processor to execute the advertising text generation 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 8 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.

[0141] In this embodiment, the processor is used to execute... Figure 7 The specific functions of each module and its sub-modules are defined within the device. The memory stores the program code and various types of 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 generation device of this application. The server can call the server's program code and data to execute the functions of all sub-modules.

[0142] 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 generation method of any embodiment of this application.

[0143] 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.

[0144] In summary, this application is able to generate high-quality advertising texts for advertising products by combining a thorough understanding of the information of the advertised product itself with ample reference to the advertising creative ideas in the advertising text.

[0145] Those skilled in the art will understand that the steps, measures, and solutions in the various operations, methods, and processes discussed in this application can be alternated, modified, combined, or deleted. Furthermore, other steps, measures, and solutions in the various operations, methods, and processes discussed in this application can also be alternated, modified, rearranged, decomposed, combined, or deleted. Furthermore, steps, measures, and solutions in the prior art that are similar to those disclosed in this application can also be alternated, modified, rearranged, decomposed, combined, or deleted.

[0146] The above description is only a partial embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for generating advertising text, characterized in that, Includes the following steps: Obtain the product title of the advertised product and its reference advertising text; The decoding feature information of the product title is obtained by decoding the text generation model through an encoding and decoding network step by step, and the generation probability distribution of the decoding feature information mapped to the first reference vocabulary referenced by the encoding and decoding network is determined. An encoder using a text generation model obtains encoded feature information of the composite text consisting of the product title and its reference advertising text. Key feature information of this encoded feature information relative to the decoded feature information is determined as a pointer probability distribution. This process includes: concatenating the product title and its reference advertising text to form the composite text; encoding the composite text according to a second reference vocabulary referenced by the encoder of the text generation model to obtain a second encoded vector; extracting deep semantic information from the second encoded vector using the encoder to obtain a corresponding second encoded vector sequence as encoded feature information; and inputting the decoded feature information and the encoded feature information corresponding to each time step in the decoding process of the text generation model's encoder-decoder network into an attention layer for feature interaction to obtain corresponding key feature information as a pointer probability distribution. The matching weight is calculated and determined based on the encoded feature information and the decoded feature information. The generated probability distribution and the pointer probability distribution are then smoothly weighted with the matching weight to obtain an optimized probability distribution. The optimized probability distribution includes an indication of the probability of taking a word corresponding to the first reference vocabulary or the comprehensive text. Based on the optimized probability distribution, a preset sampling strategy is used to extract words from the first reference vocabulary or the comprehensive text to determine the lexical units of the target advertising text.

2. The advertising text generation method according to claim 1, characterized in that, The steps to obtain the product title and its reference advertising text for the advertised product include the following: Obtain the product title and product image of the advertised product; A text encoder is used to extract the deep semantic information of the product title of the advertised product to obtain the corresponding text encoding vector. An image encoder is used to extract the deep semantic information of the product image of the advertised product to obtain the corresponding image encoding vector. Calculate the first similarity between the image encoding vector of the product image of the advertised product and the image encoding vector of the product image in each creative advertisement in the advertising creative library, and calculate the second similarity between the text encoding vector of the product title of the advertised product and the text encoding vector of the reference advertisement text in each creative advertisement in the advertising creative library. Reference advertising text that matches the product title is selected from the advertising creative library based on the first similarity and the second similarity.

3. The advertising text generation method according to claim 1, characterized in that, The decoding feature information of the product title is obtained by decoding step-by-step using an encoding / decoding network of a text generation model. The generation probability distribution of the decoding feature information mapped to the first reference vocabulary referenced by the encoding / decoding network is determined, including the following steps: The product title is encoded according to the first reference vocabulary referenced by the encoding and decoding network of the text generation model to obtain the first encoding vector; The encoder of the aforementioned codec network extracts the deep semantic information of the first encoded vector to obtain the corresponding first encoded vector sequence; The decoder in the encoding / decoding network takes the first encoded vector sequence as input to obtain the corresponding decoded feature sequence as decoded feature information. The decoded feature information is then autoregressively decoded step by step to obtain the generation probability distribution at each time step, wherein each generation probability distribution includes the probability of each word in the first reference vocabulary.

4. The advertising text generation method according to claim 1, characterized in that, The step of calculating and determining the matching weight based on the encoded feature information and the decoded feature information, and then smoothly weighting the generated probability distribution and the pointer probability distribution with the matching weight to obtain the optimized probability distribution includes the following steps: The context vector of the comprehensive text is obtained by weighted summation of the encoded feature information using the key feature information. The matching weights are calculated based on the decoded feature information and the context vector. The generated probability distribution and the pointer probability distribution are weighted by the matching weights and then summed to obtain the optimized probability distribution.

5. The advertising text generation method according to any one of claims 1 to 4, characterized in that, The training process of the text generation model includes the following steps: Obtain a single training sample from the prepared training set. Each training sample in the training set contains the product title and its reference advertising text. Each training sample is labeled with the advertising text used by the product in the past when it was advertised as a supervision label. The text generation model uses an encoding / decoding network to decode the product titles of the training samples step by step to obtain the decoding feature information, and determines the generation probability distribution of the decoding feature information mapped to the first reference vocabulary referenced by the encoding / decoding network; The encoder of the text generation model obtains the encoded feature information of the comprehensive text composed of the product title and the reference advertising text of the training sample, and determines the key feature information of the encoded feature information relative to the decoded feature information as a pointer probability distribution; The matching weight is calculated and determined based on the encoded feature information and the decoded feature information. The generated probability distribution and the pointer probability distribution are smoothly weighted by the matching weight to obtain an optimized probability distribution. The target advertising text is determined by taking words from the first reference word list or the comprehensive text according to the optimized probability distribution, and the target advertising text is constructed. The optimized probability distribution includes the probability corresponding to taking words from the first reference word list or the comprehensive text. The loss value of the target advertisement text is determined based on the supervised labels of the training samples. If the loss value does not reach a preset threshold, the weights of the text generation model are updated, and other training samples are called to carry out iterative training until the model converges.

6. An advertising text generation device, characterized in that, include: The data acquisition module is used to acquire the product title of the advertised product and its reference advertising text; The first probability distribution module is used to decode the product title step by step using an encoding and decoding network of a text generation model to obtain the decoding feature information, and to determine the generation probability distribution of the decoding feature information mapped to the first reference vocabulary referenced by the encoding and decoding network. The second probability distribution module is used to obtain the encoded feature information of the comprehensive text composed of the product title and its reference advertising text using the encoder of the text generation model, and to determine the key feature information of the encoded feature information relative to the decoded feature information as a pointer probability distribution. This includes: concatenating the product title and its reference advertising text to form the comprehensive text; encoding the comprehensive text according to the second reference vocabulary referenced by the encoder of the text generation model to obtain a second encoded vector; extracting the deep semantic information of the second encoded vector using the encoder to obtain a corresponding second encoded vector sequence as encoded feature information; and inputting the decoded feature information and the encoded feature information corresponding to each time step in the decoding process of the text generation model's encoder-decoder network into the attention layer for feature interaction to obtain the corresponding key feature information as a pointer probability distribution. The third probability distribution module is used to calculate and determine the matching weight based on the encoded feature information and the decoded feature information, and to smoothly weight the generated probability distribution and the pointer probability distribution with the matching weight to obtain an optimized probability distribution. The optimized probability distribution includes the probability of taking a word corresponding to the first reference vocabulary or the comprehensive text. The lexical determination module is used to determine the lexical units of the target advertising text by extracting words from the first reference vocabulary or the comprehensive text according to the optimized probability distribution and a preset sampling strategy.

7. The advertising text generation device according to claim 6, characterized in that, The data acquisition module includes: a data acquisition submodule for acquiring the product title and product image of the advertised product; an image encoding submodule for extracting deep semantic information of the product title of the advertised product using a text encoder to obtain a corresponding text encoding vector, and extracting deep semantic information of the product image of the advertised product using an image encoder to obtain a corresponding image encoding vector; a similarity calculation submodule for calculating a first similarity between the image encoding vector of the product image of the advertised product and the image encoding vector of the product image in each creative advertisement in the advertising creative library, and calculating a second similarity between the text encoding vector of the product title of the advertised product and the text encoding vector of the reference advertisement text in each creative advertisement in the advertising creative library; and a filtering and matching submodule for filtering reference advertisement text that matches the product title from the advertising creative library based on the first similarity and the second similarity.

8. The advertising text generation device according to claim 6, characterized in that, The first probability distribution module includes: a first text encoding submodule, used to encode the product title according to a first reference vocabulary referenced by the encoding / decoding network of the text generation model to obtain a first encoding vector; a first semantic extraction submodule, used to extract deep semantic information of the first encoding vector using the encoder of the encoding / decoding network to obtain a corresponding first encoding vector sequence; and an autoregressive decoding submodule, used to use the decoder in the encoding / decoding network with the first encoding vector sequence as input to obtain a corresponding decoding feature sequence as decoding feature information, and autoregressively decode the decoding feature information step by step to obtain the generation probability distribution at each time step, wherein each generation probability distribution includes the probability of each word in the first reference vocabulary.

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 5.

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 5, which, when invoked by a computer, executes the steps included in the corresponding method.

Images