Advertisement copy generation method and device, equipment and medium

Abstract

The application relates to an advertisement script generation method and device, equipment and medium in the computer technology field, which comprises the following steps: adopting a text generation model to generate a product text of an advertisement product, and predicting a generation probability distribution corresponding to each advertisement word element mapped to an encoding word table. Each generation probability distribution is weighted by a category probability distribution corresponding to the product category of the advertisement product to obtain a corresponding candidate probability distribution. According to each candidate probability distribution, a plurality of candidate word elements are selected from the encoding word table, and advertisement phrases starting with each candidate word element of a shop corresponding to the advertisement product are selected from a historical advertisement phrase library as candidate words to construct a corresponding candidate word set. For each advertisement word element, corresponding words are determined in the candidate word set or in the encoding word table according to the generation probability distribution / candidate probability distribution to construct an advertisement script. The application can generate high-quality advertisement scripts.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method for generating advertising copy and the corresponding apparatus, computer equipment, and computer-readable storage medium. Background Technology

[0002] E-commerce platforms typically have advertising pages where store users can place ads for the products they list in their stores to attract buyers on the e-commerce platform and drive traffic to promote sales.

[0003] When publishing advertisements, corresponding advertising copy is required. Professional copy usually leads to better promotional results. However, creating professional advertising copy is not easy. It requires a thorough understanding of the product's selling points and the target market, extracting the product's selling points to meet the needs of the target market, understanding certain writing standards, and possessing certain language expression skills. Therefore, to meet the advertising needs of e-commerce platforms, it is necessary to automatically generate advertising copy for the advertised products, assisting store users in reference, lowering the creation threshold, and making it easy to create professional advertising copy.

[0004] Currently, the automated generation of advertising copy for products mainly employs neural network models. These models extract textual information about the product, such as the product title and product details, along with corresponding semantic information, and then decode this semantic information to generate the advertising copy. While these models do achieve the goal of automated generation, they cannot guarantee that the generated advertising copy clearly conveys the product's marketing selling points, thus failing to meet the quality requirements for effective marketing promotion. Summary of the Invention

[0005] The primary objective of this application is to solve at least one of the aforementioned problems by providing a method for generating advertising copy and corresponding apparatus, computer equipment, and computer-readable storage medium.

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

[0007] A method for generating advertising copy, provided for one of the purposes of this application, includes the following steps:

[0008] The product text of the advertised product is obtained and input into the text generation model to predict the generation probability distribution of each advertising word element mapped to the corresponding encoded vocabulary. The product text includes the product description information of the product.

[0009] The preset category probability distribution for the product category of the advertised product is obtained, and each of the generated probability distributions is weighted to obtain the candidate probability distribution corresponding to each advertising term.

[0010] Based on each candidate probability distribution, select multiple candidate word elements corresponding to the preset conditions from the encoded word list, and select advertising phrases of the store corresponding to the advertised product that start with each of the candidate word elements from the preset historical advertising phrase library as candidate words, and construct a candidate word set corresponding to each candidate probability distribution.

[0011] For each advertising term, one term is randomly selected from its corresponding candidate term set as the corresponding term, or the term in the encoded term list is determined according to its generation probability distribution / candidate probability distribution as the corresponding term, and all corresponding terms are used to construct the advertising copy.

[0012] On the other hand, an advertising copy generation device provided to meet one of the purposes of this application includes a model prediction module, a probability distribution module, a word set construction module, and a copy construction module, wherein: the model prediction module is used to obtain the product text of the advertised product, input it into the text generation model, and predict the generation probability distribution corresponding to each advertising word element mapped to the encoded word list, the product text including the product description information of the product; the probability distribution module is used to obtain a category probability distribution preset for the product category of the advertised product, weight each generation probability distribution, and obtain a candidate probability distribution corresponding to each advertising word element; the word set construction module is used to select multiple candidate words elements corresponding to preset conditions from the encoded word list according to each candidate probability distribution, and select advertising phrases starting with each candidate word element from a preset historical advertising phrase library corresponding to the store of the advertised product as candidate words, and construct a candidate word set corresponding to each candidate probability distribution; the copy construction module is used to randomly select one from its corresponding candidate word set as a corresponding word for each advertising word element, or determine the word element in the encoded word list as a corresponding word according to its generation probability distribution / candidate probability distribution, and construct all corresponding words into advertising copy.

[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 copy 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 copy 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] First, this application uses a text generation model to predict the generation probability distribution of each advertising term in the advertising copy based on the product text. Then, it uses the category probability distribution corresponding to the product category to weight the generation probability distribution to obtain a candidate probability distribution. Based on the candidate probability distribution, it selects the corresponding candidate word set from the historical advertising phrase library. Finally, optionally, it determines the corresponding word for each advertising term in its candidate word set or in the model's vocabulary based on the generation probability distribution / candidate probability distribution. This introduces randomness into the determination of advertising terms in the advertising copy, enabling it to relatively randomly determine some advertising terms in the model's prediction results as high-quality advertising phrases in the historical advertising phrase library. This avoids heavy reliance on the text generation model's own vocabulary to determine advertising terms, enriches the expressive content of the generated advertising copy, and can further improve the quality of the generated advertising copy.

[0017] Secondly, based on the generation probability distribution of each advertising term obtained by the text generation model, this application uses the category probability distribution corresponding to the product category to weight the generation probability distribution. The category probability distribution can be, for example, reference data generated after extracting statistical features from the vocabulary used to express the selling points of the corresponding product category in high-quality advertising copy. In this way, the features corresponding to the word usage habits of the product category corresponding to the product text can be made explicit by using the category probability distribution, so that the weighted candidate probability distribution has a greater ability to represent the features corresponding to the word usage habits of the corresponding product category, and can be used to achieve the selection of high-quality advertising words.

[0018] Furthermore, because randomness is introduced during the copywriting stage, the generated copy, by appropriately using historical ad phrases, is more likely to produce elegant sentences. Moreover, since historical ad phrases can usually accurately express the product's selling points, the generated ad phrases are better able to express the product's selling points in the product text, which can improve the conversion rate of ads using the ad copy and increase advertising revenue. Attached Figure Description

[0019] 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:

[0020] Figure 1 This is a flowchart illustrating a typical embodiment of the advertising copy generation method of this application;

[0021] Figure 2 This is a schematic diagram illustrating the process of constructing a historical advertising phrase library in an embodiment of this application;

[0022] Figure 3 This is a schematic diagram of the encoding and decoding process of the text generation model in the embodiments of this application;

[0023] Figure 4This is a schematic diagram illustrating the process of obtaining the category probability distribution in an embodiment of this application;

[0024] Figure 5 This is a flowchart illustrating the process of determining the corresponding words for each advertising term in an embodiment of this application;

[0025] Figure 6 This is a flowchart illustrating the process of setting decision probabilities in one embodiment of this application;

[0026] Figure 7 This is a flowchart illustrating the process of setting decision probabilities in another embodiment of this application;

[0027] Figure 8 This is a schematic diagram of the process for obtaining a quality score for advertising copy in an embodiment of this application;

[0028] Figure 9 This is a schematic diagram of the advertising copy generation device of this application;

[0029] Figure 10 This is a schematic diagram of the structure of a computer device used in this application. Detailed Implementation

[0030] 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.

[0031] 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.

[0032] It will be understood by those skilled in the art 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.

[0033] One or more of the technical features of this application, unless explicitly specified herein, can be deployed on a server and accessed by a client remotely calling the online service interface provided by the server, or can be directly deployed and run on a client for access.

[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 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.

[0036] 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.

[0037] 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.

[0038] The advertising copy 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. In this way, the method can be executed by 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.

[0039] Please see Figure 1 The advertising copy generation method of this application, in its typical embodiment, includes the following steps:

[0040] Step S1100: Obtain the product text of the advertised product, input it into the text generation model, and predict the generation probability distribution of each advertising word element mapped to the corresponding encoding vocabulary. The product text includes the product description information of the product.

[0041] The advertised products refer to goods advertised by merchants on e-commerce platforms. The product text of the advertised products mainly includes the product title. In other embodiments, the product text may further include any one or more types of data such as product details, product attribute data, and product category. The product text provides a description of the characteristics of the advertised products in text form. Therefore, semantic information corresponding to the selling points and characteristics of the advertised products can be extracted from the product text.

[0042] The text generation model is an Encoder-Decoder language model. The product text of the advertised item is used as input to the model. A pre-trained, convergent encoder is applied to encode the product text. This encoder can be implemented using an RNN (Recurrent Neural Network), but a self-attention layer-based encoder suitable for sequence tasks is recommended. The product text is segmented to obtain corresponding words. For each word in the product text, the corresponding encoding value is determined by querying an encoding vocabulary. Embedding vectors, i.e., word vectors, are constructed based on these encoding values. These word vectors constitute the initial feature sequence corresponding to the product text. The encoding vocabulary stores the mapping relationship between multiple words and their corresponding encoding data. Further, deep semantic features of the selling points of the advertised item are extracted from the initial feature sequence to obtain a text feature sequence. This text feature sequence is used as prefix information and input to the pre-trained, convergent decoder of the model for decoding. An autoregressive method is used to predict the probability distribution of each advertising word mapping to the corresponding word in the encoding vocabulary.

[0043] Step S1200: Obtain the category probability distribution preset for the product category of the advertised product, and weight each of the generated probability distributions to obtain the candidate probability distribution corresponding to each advertising word.

[0044] E-commerce platform stores all have a product category system in place to categorize and organize their vast array of goods. This category system can be multi-layered, containing multiple classification levels, each level encompassing multiple specific product categories. The e-commerce platform can provide a standardized template for constructing the category system, which merchants can then modify and finalize themselves.

[0045] Each product category has a corresponding preset category probability distribution, which represents the probability of each word in the encoded vocabulary for that product category. In one embodiment, the preset category probability distribution for each product category is implemented as follows:

[0046] It is understandable that e-commerce platforms are typically equipped with advertising systems. Merchants in various online stores can submit their advertising copy and configuration information to the advertising system, which will then display these to the public through the system's own advertising channels or third-party advertising channels. Furthermore, the advertising system can track the effectiveness of advertising campaigns, collecting relevant performance data and storing it in relation to the advertisement. This performance data includes, but is not limited to, click-through rate, conversion rate, collection rate, add-to-cart rate, purchase rate, and return on investment (ROI).

[0047] It is easy to understand that the performance data objectively reflects the effectiveness of advertising. The higher the performance data, the better the advertising performance and the higher the quality of the corresponding advertising copy. Therefore, by calling the data acquisition interface opened by the advertising system, the performance data corresponding to each advertisement can be obtained. Any one or more of the performance data that meets preset thresholds, such as a conversion rate higher than 0.8, a click-through rate higher than 0.8, or an add-to-cart rate of 0.8, are considered high-quality advertising copy. The preset threshold is set according to one or more of the performance data, and the specific value can be set as needed by those skilled in the art. It is known that advertising copy typically uses concise sentences or vocabulary to describe the selling points of a product or other advertising information. Therefore, by dividing the product categories to which each high-quality advertising copy belongs, high-quality advertising copy corresponding to each product category can be obtained.

[0048] Perform corresponding word segmentation on each high-quality advertising copy corresponding to each commodity category, remove stop words and punctuation marks to obtain a plurality of segmented words corresponding to each high-quality advertising copy, and construct a word segmentation set for each commodity category. For each commodity category, calculate the word frequency corresponding to each segmented word in its word segmentation set, wherein the calculation can be performed by dividing the number of occurrences of the segmented word in the word segmentation set by the total number of all segmented words in the word segmentation set, and the obtained result is the word frequency. In this way, the word frequency corresponding to each segmented word in the word segmentation set of each commodity category is obtained. Further, the word frequency corresponding to each segmented word in the word segmentation set of each commodity category is used as the probability of the token in the same coding vocabulary as the segmented words, so as to obtain a category probability distribution.

[0049] It can be easily understood from the above that the category probability distribution is reference data generated after condensing, extracting and statistically characterizing the vocabulary used by high-quality advertising copy to express the selling point features of commodities of corresponding categories.

[0050] For another implementation of the category probability distribution, reference may be made to the disclosure of some embodiments in the following part, which will not be described in this step for the time being.

[0051] Obtain the category probability distribution corresponding to the commodity category of the advertising commodity, and weight the generation probability distribution corresponding to each advertising token with the obtained category probability distribution. In one embodiment, an exemplary formula for the weighting is given as follows:

[0052] Pcan=Pgen*(1+w·Pbow)

[0053] wherein: Pgen is the generation probability distribution, Pbow is the category probability distribution, 1 is a regularization term for preventing the category probability distribution from being 0, 0<w≤1, and "*" represents an element-wise bitwise multiplication operation, that is, each probability in the generation probability distribution is multiplied bitwise by (1 + w multiplied by each corresponding probability in the category probability distribution). After the weighting, a candidate probability distribution corresponding to each advertising token is obtained.

[0054] Step S1300: selecting a plurality of candidate tokens corresponding to meeting preset conditions from the coding vocabulary according to each candidate probability distribution, selecting advertising phrases starting with each of the candidate tokens and corresponding to the store of the advertising commodity from a preset historical advertising phrase bank as candidate words, and constructing a candidate word set corresponding to each candidate probability distribution;

[0055] In one embodiment, a threshold can be set based on the candidate probability distribution corresponding to each advertising term, so as to select multiple candidate terms corresponding to each advertising term exceeding the threshold from the encoded vocabulary according to each candidate probability distribution. Specifically, the threshold is related to "w" set in the weighting process in step S1200. The larger "w" is, the larger the threshold is. When "w" is 1, the threshold can be set to a value between 0.8 and 1.5. Those skilled in the art can flexibly set the threshold according to the disclosure herein.

[0056] The historical advertising phrase library stores advertising phrases for each store under each product category. The specific implementation will be further revealed in subsequent embodiments, and this step will not be discussed here.

[0057] Furthermore, the store that published the advertisement for the advertised product is identified, as well as the product category corresponding to the advertised product. Based on the store and product category corresponding to the advertised product, the corresponding advertising phrase in the historical advertising phrase library is determined. The advertising phrases that begin with each of the aforementioned candidate word elements are obtained as candidate words, and a corresponding candidate word set is constructed based on the candidate words corresponding to each candidate probability distribution.

[0058] Step S1400: For each advertising word, randomly select one from its corresponding candidate word set as the corresponding word, or determine the word in the encoded word list as the corresponding word according to its generation probability distribution / candidate probability distribution, and construct all corresponding words into advertising copy.

[0059] It is understood that each advertising term corresponds to a candidate probability distribution with its own set of candidate words. Therefore, for each advertising term, one can be randomly selected from its corresponding set of candidate words as the corresponding word. The candidate words selected for the latter advertising term will differ from those selected for the former. Alternatively, the term with the highest predicted probability from the corresponding model encoding in the lexicon, determined based on its generation probability distribution or candidate probability distribution, can be selected as the corresponding word. For these two options, corresponding decision probabilities can be set to determine which option to execute. Based on this, an advertising copy is constructed using the corresponding word for each advertising term. Furthermore, this step can be performed multiple times to obtain multiple advertising copy.

[0060] The server can push multiple advertising copy corresponding to the advertised product to the page where the advertising copy for the advertised product is created, so as to provide creative references for the store's merchants and lower the creative threshold.

[0061] 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:

[0062] First, this application uses a text generation model to predict the generation probability distribution of each advertising term in the advertising copy based on the product text. Then, it uses the category probability distribution corresponding to the product category to weight the generation probability distribution to obtain a candidate probability distribution. Based on the candidate probability distribution, it selects the corresponding candidate word set from the historical advertising phrase library. Finally, optionally, it determines the corresponding word for each advertising term in its candidate word set or in the model's vocabulary based on the generation probability distribution / candidate probability distribution. This introduces randomness into the determination of advertising terms in the advertising copy, enabling it to relatively randomly determine some advertising terms in the model's prediction results as high-quality advertising phrases in the historical advertising phrase library. This avoids heavy reliance on the text generation model's own vocabulary to determine advertising terms, enriches the expressive content of the generated advertising copy, and can further improve the quality of the generated advertising copy.

[0063] Secondly, based on the generation probability distribution of each advertising term obtained by the text generation model, this application uses the category probability distribution corresponding to the product category to weight the generation probability distribution. The category probability distribution can be, for example, reference data generated after extracting statistical features from the vocabulary used to express the selling points of the corresponding product category in high-quality advertising copy. In this way, the features corresponding to the word usage habits of the product category corresponding to the product text can be made explicit by using the category probability distribution, so that the weighted candidate probability distribution has a greater ability to represent the features corresponding to the word usage habits of the corresponding product category, and can be used to achieve the selection of high-quality advertising words.

[0064] Furthermore, because randomness is introduced during the copywriting stage, the generated copy, by appropriately using historical ad phrases, is more likely to produce elegant sentences. Moreover, since historical ad phrases can usually accurately express the product's selling points, the generated ad phrases are better able to express the product's selling points in the product text, which can improve the conversion rate of ads using the ad copy and increase advertising revenue.

[0065] Please see Figure 2 In a further embodiment, before step S1100, obtaining the product text of the advertised product, the following steps are also included:

[0066] Step S1000: Obtain the advertising copy of the ads that have been placed in the advertising system. Each advertising copy is placed by a store in the e-commerce platform and is used to promote the products corresponding to its description.

[0067] E-commerce platforms are typically equipped with advertising systems. By opening up corresponding advertising publishing interfaces to merchants in various online stores, they can obtain the advertising copy and configuration information corresponding to the advertisements placed by merchants in any store into the advertising system, and submit them to the advertising system's own advertising publishing channels or third-party advertising publishing channels for public display.

[0068] In this application, the advertising copy mainly includes advertising text, which is expressed in natural language and can contain text in any language. Each store can be deployed on an independent site of an e-commerce platform. Each store can list a large number of products, and users can enter the store to access the transaction page of any product to place an order and make a payment, thereby realizing e-commerce transactions. Each store can place advertisements corresponding to any product in its store into the advertising system, providing the corresponding advertising copy during the advertising process, and submitting it to the advertising system through the advertising publishing interface to complete the publication. The advertising copy can describe the selling points and other advertising information of the corresponding product using text in the appropriate language. The selling points of the product can include any information such as the product's name, brand, features, and attributes.

[0069] Once the corresponding advertising copy for a product is submitted and the corresponding advertisement is successfully launched, the advertising system will calculate various performance data for the advertisement based on the number of user visits to the corresponding advertisement display page or display position after the advertisement is launched. These data include, but are not limited to, click-through rate, conversion rate, collection rate, add-to-cart rate, purchase rate, return on investment, etc.

[0070] Click-through rate (CTR) is a commonly used term in internet advertising. It refers to the click-through rate of online ads (image ads / text ads / keyword ads / ranking ads / video ads, etc.), which is the actual number of clicks on the ad (strictly speaking, it can be the number of times the ad is clicked and a redirect is triggered to the specified page) divided by the number of times the ad is displayed (showcontent).

[0071] Conversion Rate (CVR) refers to the percentage of internet users who click on an online advertisement and enter the promoting website and subsequently convert, typically reflecting the direct benefit of the advertisement. It was initially proposed in the "China Online Marketing (Advertising) Effectiveness Evaluation Criteria" at the China Internet Association's Online Marketing Working Committee Member Conference held on June 18, 2009. The statistical period is usually hourly, daily, weekly, or monthly, and can be set as needed. The statistical objects include various advertising formats such as flash ads, image ads, text link ads, advertorials, email ads, video ads, and multimedia ads. CVR = (Conversions / Clicks) * 100%.

[0072] The collection rate is the ratio of the total number of products displayed on the product page after a web ad is clicked and the user is redirected to the collection page to the total number of users who click on the ad, i.e., the number of visitors.

[0073] The add-to-cart rate is similar to the collection rate, referring to the ratio of the total number of items added to the shopping cart by users who clicked on the online advertisement and were redirected to the product page, to the total number of users who clicked on the ad, i.e., the number of visitors.

[0074] The purchase rate, similar to the add-to-cart rate, refers to the ratio of the total number of products displayed on the product page after a click on an online advertisement leads to the total number of users who click on the ad, i.e., the number of visitors.

[0075] The Return on Advertising Spend (ROAS) is a marketing metric that measures the effectiveness of online advertising. ROAS = Total Advertising Revenue / Advertising Cost.

[0076] It's easy to understand that for previously placed ads, the data acquisition interface provided by the advertising system can be used to obtain not only the ad copy but also the corresponding performance data for each ad.

[0077] Step S1010: Construct an advertising keyword library corresponding to each product category in the product category system. Each advertising keyword library stores advertising phrases extracted from the advertising copy of the corresponding product category. Each advertising phrase contains two or more word elements.

[0078] E-commerce platform stores all have a product category system in place to categorize and organize their vast array of goods. This category system can be multi-layered, containing multiple classification levels, each level encompassing multiple specific product categories. The e-commerce platform can provide a standardized template for constructing the category system, which merchants can then modify and finalize themselves.

[0079] The advertising copy obtained from the advertising system contains numerous concise and precise advertising phrases that accurately describe the selling points of the product. Therefore, one or more advertising phrases can be extracted from each advertising copy according to certain rules to construct the advertising terminology library of this application, and the extracted advertising phrases are stored. Each advertising phrase is stored in the advertising terminology library corresponding to the product category of the product in the advertising copy it belongs to. The advertising phrase can be a phrase containing two or more word elements in the advertising copy, and the word elements can be two words, three words, or four words.

[0080] In one embodiment, part-of-speech (POS) structure information can be provided in advance. This POS structure information is used to define the lexical construction rules for the advertising phrases to be extracted from the advertising copy. Therefore, it can be represented by including multiple phrase constructions. For example, the phrase constructions are represented in the following form:

[0081] Noun & Noun

[0082] Adjectives & Nouns

[0083] Verbs & Nouns

[0084] It can be seen that each phrase is constructed to represent a combination of word elements that are the same (noun & noun) or different parts of speech (adjective & noun, verb & noun) as an advertising phrase.

[0085] For each ad copy, in order to obtain the ad phrase, a preset word segmentation method can be applied first, such as using the N-Gram algorithm or Jieba word segmenter to segment it. At the same time, a preset part-of-speech extractor or other preset neural network models for part-of-speech tagging can be used to tag each word segmentation to obtain the part-of-speech corresponding to each word segmentation.

[0086] Furthermore, based on the part of speech corresponding to each word in each ad copy, a combination of word segments that conforms to the specifications of the phrase construction can be matched as an ad phrase.

[0087] Step S1020: Statistically determine the recommendation score of each advertising phrase under the constraints of product category and store conditions by associating the advertising copy containing the advertising phrase. The recommendation score is determined by summarizing the category dimension score, store dimension score, and advertising dimension score of the advertising phrase.

[0088] The category-based scoring of the advertising phrase can be achieved through the following process:

[0089] First, we count the frequency of each ad phrase in the ad copy of products within the same product category. Taking each product category as an independent unit, we count the number of times each ad phrase w in the ad word library corresponding to each product category j appears in all the ad copy of ads already run for that product category; this is its frequency. w_j Then, count the number of ad copy for all ads that have been placed for the product category. j Therefore, we can obtain the ratio of each ad phrase appearing in all ad copy that has been placed. w_j ,Right now:

[0090] Ratio w_j =freqency w_j / count j

[0091] Furthermore, the occurrence ratio of all advertising phrases in the advertising keyword library for each category can be normalized to achieve numerical standardization, adjusting the statistical units of each occurrence ratio to the numerical space of [0,1]. In one embodiment, the softmax function is applied for normalization to transform the occurrence ratio of each advertising phrase under each category. An example formula is shown below:

[0092]

[0093] Where k represents the category to which the advertising phrase belongs, and j represents any one of the categories.

[0094] After conversion, each ad phrase under each category can obtain its corresponding category-specific score (ScoreCatego). w .

[0095] The store-level rating of the aforementioned ad phrase can be achieved through the following process:

[0096] On a store-by-store basis, for each product category, the frequency of each ad phrase in the ad copy of all products advertised in that store is calculated. This allows us to determine the frequency of each ad phrase used by a store within the same product category across multiple ad copy campaigns for that store. w_j_s .

[0097] For each ad phrase, stores whose word frequency exceeds a preset threshold are identified as "used stores." The total number of stores in the same product category that have advertised for each product category, and the total number of "used stores," are also determined. The preset threshold can be an empirical threshold or a measured threshold, and can be set as needed by those skilled in the art. Specifically, for each store, the word frequency of each ad phrase used is compared with the preset threshold. When the word frequency exceeds the preset threshold, the store is identified as a "used store" that frequently uses the ad phrase. Conversely, if the word frequency does not exceed the preset threshold, the store is identified as a "non-used store" that uses the ad phrase infrequently.

[0098] For each ad phrase under each product category, the corresponding stores that have used the ad can be determined according to the above principles, thus determining the total number of stores that have used the ad. used_j Furthermore, for all stores that have advertised products within the same product category j, they can be identified as stores within the same product category that have advertised products in that category, thus allowing us to determine the total number of stores within the same product category. all_i .

[0099] For each store, given a fixed total number of stores using the same ad phrase within a product category, a higher total number of stores using that ad phrase indicates lower freshness, and vice versa. This means the ad phrase is more effective at differentiating the store from others. Therefore, the ratio of the total number of stores using the ad phrase to the total number of stores using it within the same product category determines its widespread use. Furthermore, the following formula can be applied to determine the freshness (ScoreStore) of each ad phrase for each store and each product category. w :

[0100]

[0101] The 1 is used to avoid regularization terms with a denominator of zero, and the word frequency of the advertising phrase. w_j_s This can be seen as adjusting the weight. It is not difficult to understand that the higher the word frequency, the fresher the advertising phrase is, indicating that the store not only uses the advertising phrase frequently, but it is also likely a common word that distinguishes the store from other stores.

[0102] To facilitate the calculation of recommendation scores, a min-max normalization process is applied to normalize the freshness of all ad phrases for each store by category. This yields the store-level score for each ad phrase within each category. For ease of understanding, ScoreStore will be used as the reference hereafter. w This indicates the store-specific rating.

[0103] The advertising dimension scoring of the aforementioned advertising phrase can be achieved through the following process:

[0104] By calling the data acquisition interface provided by the advertising system, the system can retrieve the corresponding ad copy for each ad phrase within the same product category, along with the corresponding performance data. The performance data of the ad copy for each ad phrase within the same product category is averaged to obtain the average performance data for each ad phrase within its respective product category. To facilitate the calculation of recommendation scores, a min-max normalization process is applied to normalize the average performance data of each ad phrase by product category, thus obtaining the ScoreCTR for each ad phrase within each product category. w .

[0105] The recommendation rating of the advertising phrase can be achieved through the following process:

[0106] Using product categories and stores as constraints, a recommendation score is calculated for each product category within each store. This recommendation score is a weighted sum of the ad phrase's category-level score, store-level score, and ad-level score within the product category. Example formula:

[0107] Score w =c1*ScoreCate w +c2*ScoreStore w +c3*ScoreCTR w

[0108] Among them, Score w The recommended ratings for advertising phrases in different stores are provided. C1, C2, and C3 are the preset weights corresponding to the category dimension rating, store dimension rating, and advertising dimension rating of the advertising phrase, respectively, which can be preset as needed by those skilled in the art.

[0109] Step S1030: Based on the recommendation scores of each advertising phrase in the advertising terminology library under the constraints of product category and store, obtain the advertising phrases with higher recommendation scores for different stores in the advertising terminology library corresponding to each product category to construct a historical advertising phrase library.

[0110] A higher recommendation score can be preset as a threshold. Then, from the advertising keyword library corresponding to each product category, select some advertising phrases corresponding to different stores with recommendation scores exceeding the preset threshold. It is easy to understand that these advertising phrases have higher information contribution value for their respective product categories and stores. These advertising phrases are associated with their respective product categories and stores and stored in the historical advertising phrase library.

[0111] In this embodiment, a standardized processing procedure is employed. Based on the advertising phrases extracted from the advertising copy obtained from the advertising system, the information contribution value of each advertising phrase is quantified not only from the perspective of product category but also from the perspective of each store. Furthermore, by combining the information contribution value of each advertising phrase in the advertising dimension, a recommendation score for each advertising phrase is determined under the constraints of store and product category. This further enables the batch determination of the historical advertising phrase library required by the massive number of online stores on the e-commerce platform. The processing procedure is highly efficient, and the generated advertising phrases, under the guidance of their recommendation scores, can accurately provide the information contribution value of the advertising phrases in the advertising copy, providing reliable basic data for the generation of advertising copy in this application.

[0112] Please see Figure 3In a more detailed embodiment, step S1100, the step of predicting the generation probability distribution corresponding to each advertising term in the encoded vocabulary, includes the following steps:

[0113] Step S1110: Encode the product text of the advertisement product using the encoder of the text generation model to obtain the text feature sequence;

[0114] The product text can be any one or more of the following: product title, product details text, and product category. When multiple texts are involved, they can be concatenated to obtain the product text.

[0115] The text generation model is an Encoder-Decoder language model. The product text of the advertised item is used as input to the model. A pre-trained, convergent encoder is applied to encode the product text. This encoder can be implemented using an RNN (Recurrent Neural Network), but a self-attention layer-based encoder suitable for sequence tasks is recommended. The product text is segmented to obtain corresponding words. For each word in the product text, the corresponding encoding value is determined by querying an encoding vocabulary. Embedding vectors, i.e., word vectors, are constructed based on these encoding values. These word vectors constitute the initial feature sequence corresponding to the product text. The encoding vocabulary stores the mapping relationship data between multiple words and their corresponding encoding data. Further, deep semantic features of the selling points of the advertised item are extracted from the initial feature sequence to obtain a text feature sequence.

[0116] Step S1120: Input the text feature sequence as prefix information into the decoder of the text generation model for decoding, and predict the generation probability distribution of each advertising word element mapped to the corresponding encoded vocabulary.

[0117] In one embodiment, GPT-2 (Generative Pre-Training) can be used as the decoder of the text generation model. Other decoders, such as XLNet and DistilGPT2, are also suitable for constructing the decoder. Structurally, GPT-2 is built using a transformer decoder module and employs an autoregressive mechanism to generate individual lexical units. This autoregressive mechanism involves appending each newly generated lexical unit to the previously generated sequence, which then becomes the new input for the next step of the model, ultimately generating a complete sequence of lexical units.

[0118] The text feature sequence is used as prefix information and input into the pre-trained and converged GPT-2 for decoding. An autoregressive method is used to predict the probability distribution of each advertising word corresponding to each word in the encoding vocabulary.

[0119] In this embodiment, the encoder of the text generation model encodes the product text of the advertised product, extracts text features representing the selling points of the advertised product, and obtains a vectorized text feature sequence. This sequence is then used as the decoder of the prefix information application model for decoding. Based on the text features, the corresponding generation probability distribution of each advertising term mapped to the encoded vocabulary is accurately predicted. This allows the subsequent construction of corresponding words for the advertising copy based on the determined advertising terms according to the generation probability distribution to accurately describe the selling points of the advertised product.

[0120] Please see Figure 4 In a further embodiment, step S1200, obtaining the preset category probability distribution for the product category of the advertised product, includes the following steps:

[0121] Step S1210: Use a preset word segmenter to segment the advertising phrases of each product category stored in the preset historical advertising phrase library to obtain the word segments corresponding to each advertising phrase;

[0122] The historical advertising phrase library stores advertising phrases for each store under each product category. The specific implementation will be further revealed in subsequent embodiments, and this step will not be discussed here.

[0123] The word segmenter can be a Byte-level BPE (Byte-level Byte Pair Encoding). Alternatively, word segmenters that can be used include: WordPiece, SentencePiece, Byte-Pair Encoding, N-Gram algorithm, HanLP, Jieba word segmenter, etc., and those skilled in the art can choose any one as needed.

[0124] Byte-level BPE is used to segment the advertising phrases of each product category stored in the preset historical advertising phrase library to obtain the corresponding word units for each advertising phrase.

[0125] Step S1220: Construct a bag of words for each product category using the bag-of-words model, which contains the word segments corresponding to the advertising phrases of the corresponding product category and their number in the bag of words;

[0126] The bag-of-words model can be a BOW (Bag of Words). Those skilled in the art should understand that when constructing the bag of words for a text, the BOW model does not focus on the grammatical structure of the text or the order of the words, but rather on the frequency of each word. Accordingly, the BOW model is used to count the frequency of each segmented word in each product category, and then, based on the segmented words corresponding to the advertising phrases included in each product category and their frequencies, a bag of words for each product category is constructed. An exemplary example of this bag of words is provided below.

[0127] {(token_1,count_1), (token_2,count_2)……(token_n,count_n)}

[0128] Where: token is the word segmentation unit, and count is the number of times the word segmentation unit appears.

[0129] Step S1230: Determine the word frequency corresponding to each word segment in the bag of words for each product category, and use the word frequency corresponding to each word segment in each product category as the probability of the word segment in the same encoded word list as the word segment to obtain the category probability distribution.

[0130] Furthermore, maximum value normalization can be used to determine the word frequency corresponding to each word segment in the bag-of-words for each product category. An example formula for maximum value normalization is as follows:

[0131] frequency_i=count_i / max(count_j)

[0132] Where i and j are positive integers from 1 to n. count_i is the number of occurrences of each word segment in the bag-of-words for each product category, max(count_j) is the maximum number of occurrences of each word segment in the bag-of-words for each product category, and frequency_i is the word frequency of each word segment in the bag-of-words for each product category.

[0133] The frequency of each word segment corresponding to each word segment of each product category is taken as the probability of the word segment in the same encoding word list as the word segment. Correspondingly, the probability of other word segments in the encoding word list is set to 0, thus obtaining the probability of each word segment in the encoding word list under each product category, i.e., the category probability distribution.

[0134] In this embodiment, advertising phrases for each product category stored in the historical advertising phrase library are segmented to obtain corresponding segmented word units. A bag-of-words structure is then constructed for each product category's segmented word units. The frequency of each segmented word unit in the bag-of-words structure is determined, and its probability is used as the probability of a word unit in the same encoded vocabulary as the segmented word unit. This yields the category probability distribution for each product category. Essentially, the distribution features of word units corresponding to each product category can be extracted from historical advertising data and converted into the distribution features of word units in the corresponding encoded vocabulary, i.e., the category probability distribution. Subsequently, this category probability distribution can be used to weight the model's predicted generation probability distribution to explicitly reveal the product category characteristics of the advertised product, thus guiding the model's output.

[0135] Please see Figure 5 In a further embodiment, step S1400, which involves randomly selecting one word from the corresponding candidate word set for each advertising word element, or determining the word element in the encoded word list as the corresponding word based on its generation probability distribution / candidate probability distribution, includes the following steps:

[0136] Step S1410: For each advertising keyword, obtain its corresponding decision probability, where the decision probabilities of some advertising keywords are different;

[0137] For each advertising term, a corresponding decision probability can be set, which is used to determine the selection object of the advertising term as the candidate word set corresponding to the advertising term, or to generate a probability distribution / candidate probability distribution.

[0138] In one embodiment, for the first N advertising keywords, the decision probability is set to 0, where N is greater than or equal to 1. For subsequent advertising keywords, the decision probability is set to decrease from a preset value higher than 0. The decision probability can be set as needed by those skilled in the art as disclosed herein.

[0139] In another embodiment, for the first advertising keyword, its decision probability is set to a preset value higher than 0. For subsequent advertising keywords, the decision probability of each advertising keyword is set inversely proportional to the number of candidate words that have been determined as the corresponding words. The decision probability can be set as needed by those skilled in the art as disclosed herein.

[0140] Step S1420: Based on the decision probability, determine that the word selection object of the advertising word element is the candidate word set or generation probability distribution / candidate probability distribution corresponding to the advertising word element;

[0141] It can be understood that the target word for the advertising keyword is the candidate word set corresponding to that keyword, or the generation probability distribution / candidate probability distribution. The decision probability refers to the probability that the target word is the candidate word set, and the probability that the target word is the generation probability distribution / candidate probability distribution is (1 - decision probability). For example, the decision probability is 86%, which is the probability that the target word for the advertising keyword is the candidate word set corresponding to that keyword. Additionally, there is a 14% probability that the target word is the generation probability distribution / candidate probability distribution. When the 86% hit rate is achieved, the target word is the candidate word set corresponding to the advertising keyword; conversely, when the 14% hit rate is achieved, the target word is the generation probability distribution / candidate probability distribution. Therefore, the target word for the advertising keyword can be determined based on the hit rate of the decision probability.

[0142] Step S1430: When the word selection object is a candidate word set, randomly select one of the candidate words as the corresponding word;

[0143] When the word selection target of the advertising word element is the candidate word set of the advertising word element, the selected candidate words corresponding to all the advertising word elements before the advertising word element are determined, and then a candidate word other than those selected candidate words in the candidate word set corresponding to the advertising word element is randomly determined as the corresponding word.

[0144] Step S1440: When the word selection object is a generation probability distribution / candidate probability distribution, determine the word elements in the encoding vocabulary as the corresponding words according to its generation probability distribution / candidate probability distribution.

[0145] When the word selection target for the advertising term is a generation probability distribution / candidate probability distribution, the term with the highest probability can be determined as the corresponding word based on the generation probability distribution of the advertising term, i.e., the probability of it being distributed among various terms in the corresponding encoded word list. Alternatively, the term with the highest probability can be determined as the corresponding word based on the candidate probability distribution of the advertising term, i.e., the probability of it being distributed among various terms in the corresponding encoded word list.

[0146] In this embodiment, the decision probability corresponding to each advertising term is used to flexibly determine the word selection object corresponding to the advertising term and select the corresponding word corresponding to the advertising term. This makes the advertising copy constructed from the corresponding words of each advertising term appear to be a natural embedding of candidate words, rather than a rigid and fixed embedding that looks like a template and affects the viewing experience of the advertising copy.

[0147] Please see Figure 6 In a preferred embodiment, step S1410, obtaining the corresponding decision probability for each advertising keyword, includes the following steps:

[0148] Step S1411: For the first N advertising keywords, set their decision probability to 0, where N is greater than or equal to 1;

[0149] A dynamic adjustment method can be set for the decision probability. First, for the first N advertising words predicted by the text generation model, such as 8 or 10, the decision probability corresponding to each advertising word can be set to 0. That is, the word selection objects of these first N advertising words can only be the generation probability distribution, ensuring that they will not be the corresponding candidate word set.

[0150] Step S1412: For subsequent advertising keywords, set their decision probability to decrease from a preset value higher than 0.

[0151] Secondly, for subsequent advertising keywords, the decision probability corresponding to the first advertising keyword immediately following the previous N advertising keywords can be set to a relatively large preset value higher than 0. Then, the decision probability corresponding to each subsequent advertising keyword decreases by a certain difference until a preset minimum threshold is reached. At this point, the decision probability corresponding to each advertising keyword is equal to the preset minimum threshold, which can be a value higher than or equal to 0. For example, the decision probability corresponding to the first advertising keyword is 60%, and the decision probability corresponding to each subsequent advertising keyword decreases by a difference of 2%, with a preset minimum threshold of 10%. That is, the corresponding decision probabilities are 58%, 56%, ... 10%, 10%, 10%. The specific values ​​corresponding to the minimum threshold, decision probability, and decreasing difference can be set as needed by those skilled in the art.

[0152] Please see Figure 7 Alternatively, it may include the following steps:

[0153] Step S1411': For the first advertising keyword, set its decision probability to a preset value that is higher than 0;

[0154] A dynamic adjustment method can be set for the decision probability. First, for the first advertising keyword, its decision probability can be set to a large preset value higher than 0, such as 80%. The specific value can be set as needed by those skilled in the art.

[0155] Step S1412': For subsequent advertising keywords, set the decision probability corresponding to each advertising keyword inversely proportional to the number of candidate words that have been determined as the corresponding words.

[0156] Secondly, for subsequent ad keywords, the decision probability of each ad keyword can be set inversely proportional to the number of candidate words identified as corresponding words relative to each ad keyword, including previous ad keywords and the current ad keyword. That is, the more candidate words there are, the lower the corresponding decision probability, down to 0. Accordingly, the decision probability can be set to decrease by a certain difference based on the number of candidate words until it reaches 0. For example, with a preset minimum threshold of 10%, when the number of candidate words identified as corresponding words is 0, the decision probability of each ad keyword is 80%; when the number of candidate words identified as corresponding words is 3, the decision probability of each ad keyword is 40%; and when the number of candidate words identified as corresponding words is 6, the decision probability of each ad keyword is 0. The specific values ​​corresponding to the number of candidate words and the decreasing difference can be set as needed by those skilled in the art.

[0157] This embodiment reveals a method for dynamically adjusting the decision probability of advertising keywords, so that the product copy constructed from the corresponding words of each advertising keyword has a sufficient number of embedded corresponding words as candidate words, but not too many.

[0158] Please see Figure 8 In the extended embodiment, after step S1400, which involves constructing all corresponding words into advertising copy, the following steps are also included:

[0159] Step S1500: Determine the descriptive similarity and promotional value between the product text and the advertising copy of the advertised product;

[0160] In one embodiment, multiple categories representing different levels of similarity can be set for determining descriptive similarity. These categories can be mapped to different categories based on the degree of descriptive similarity, making it convenient to quickly filter out some advertising copy with low descriptive similarity through category filtering.

[0161] The determination of descriptive similarity can be implemented using a neural network model, calculated by extracting feature vectors from the product text, advertising copy, or a combination thereof of the advertised product. The preferred neural network model is a recurrent neural network (RNN), such as LSTM (Long Short-Term Memory Recurrent Neural Network), BiLSTM (Bidirectional Long Short-Term Memory Recurrent Neural Network), Transformer, BERT, RoBERTa, ALBert, ERNIE, BERT-WWM, etc. Based on the principles disclosed in this application, the neural network model can employ one of the optimization algorithms such as Adam or AdamW to establish a multi-task classification model. It can be pre-trained to convergence using a sufficient number of corresponding training samples, thereby enabling it to learn the ability to determine descriptive similarity based on the given product text, advertising copy, or a combination thereof of the advertised product.

[0162] Therefore, in one embodiment, a dual-tower model can be built using two isomorphic basic neural network models. The two basic neural network models extract feature vectors from the product text and advertising copy of the advertised product, respectively, and then concatenate them to obtain a comprehensive feature vector. This comprehensive feature vector is then further classified and mapped to obtain the classification probability of mapping to a preset category as the descriptive similarity. In another embodiment, a single basic neural network model can be used to extract features from the combined text of the product text and advertising copy of the advertised product to obtain a comprehensive feature vector. This comprehensive feature vector is then classified and mapped to obtain the classification probability of mapping to a preset category as the descriptive similarity.

[0163] The promotional value between the product text and the advertising copy is mainly used to characterize whether the advertising copy is suitable for promotion. The higher the promotional value, the higher the information contribution value of the corresponding advertising copy for the promotion of the advertised product. The lower the promotional value, the lower the information contribution value of the advertising copy for the promotion of the advertised product.

[0164] The determination of the promotional value can also be implemented using a neural network model. This is achieved by extracting feature vectors from the product text, advertising copy, or a combination thereof of the advertised product. The preferred neural network model is a recurrent neural network (RNN), such as LSTM (Long Short-Term Memory Recurrent Neural Network), BiLSTM (Bidirectional Long Short-Term Memory Recurrent Neural Network), Transformer, BERT, RoBERTa, ALBert, ERNIE, BERT-WWM, etc. Based on the principles disclosed in this application, the neural network model can use one of the optimization algorithms such as Adam or AdamW to establish a multi-task classification model. It can be pre-trained to convergence using a sufficient number of corresponding training samples, thereby enabling it to learn the ability to determine the promotional value based on the given product text, advertising copy, or a combination thereof of the advertised product.

[0165] Therefore, in one embodiment, a dual-tower model can be built using two isomorphic basic neural network models. The two basic neural network models extract feature vectors from the product text and advertising copy of the advertised product, respectively, and then concatenate them to obtain a comprehensive feature vector. This comprehensive feature vector is then further classified and mapped to obtain the classification probability of being mapped to a preset category, which is used as the promotional value. In another embodiment, a single basic neural network model can be used to extract features from the combined text of the product text and advertising copy of the advertised product to obtain a comprehensive feature vector. This comprehensive feature vector is then classified and mapped to obtain the classification probability of being mapped to a preset category, which is used as the promotional value.

[0166] Step S1600: Determine the textual perplexity of the statements in the advertising copy;

[0167] Text perplexity can be used to evaluate the fluency of a sentence. Therefore, text perplexity can be used to quantify whether advertising copy meets grammatical requirements, uses appropriate word choice, and is semantically coherent. In one embodiment, when an advertising copy contains multiple independent sentences, the fact that they are separate sentences can be disregarded, and the individual sentences can be directly concatenated together as a single sentence for quantifying the overall text perplexity of the advertising copy. In another embodiment, the text perplexity of each sentence in the advertising copy can be determined independently, and then the average of the text perplexity of all sentences can be used as the overall text perplexity of the advertising copy.

[0168] In one embodiment, to determine the text perplexity of a given statement, a pre-trained causal language model (CLM) can be used. A causal language model, as opposed to a masked language model, is similar to the decoder in a Transformer mechanism. It uses a diagonal masking matrix, ensuring that each token only sees information about tokens preceding it, not those following it. The training objective is to predict the next token based on the preceding tokens. While the probability of word combinations is typically calculated based on probability distributions, the causal language model predicts the current token based on all preceding token information, making it directly applicable to text generation tasks. It can be understood that while an encoder-decoder model uses the complete Transformer structure, the causal language model only uses the decoder structure of the Transformer (omitting the encoder-decoder attention mechanism, as there is no encoder structure).

[0169] Therefore, the embedding vectors obtained by embedding the word units of the advertising copy can be sequentially input into a causal training model for step-by-step decoding to obtain the predicted probability corresponding to each word unit. The result of multiplying the predicted probabilities of these word units is used as the text perplexity of the statement, which can reflect the legality and rationality of the advertising copy to a certain extent. It is easy to understand that the lower the text perplexity of the advertising copy, the more grammatically correct and expressive it is.

[0170] Step S1700: Combine description similarity, promotional value, and text confusion to obtain a quality score for the advertising copy.

[0171] After obtaining the description similarity, promotional value, and text perplexity of the advertising copy, these data can be fused to obtain a comprehensive result, which serves as the quality score for the advertising copy. In one embodiment, the quality score can be obtained by weighted summation using preset weights, as exemplified by the following formula:

[0172]

[0173] Among them, ω1, ω2, and ω3 are preset weights, which can be preset as needed by those skilled in the art. Similar, Value, and Perplexity have been pre-normalized to the confidence interval of [0,1]. Therefore, the final quality score is further normalized to the numerical interval of [0,1] by taking the mean of the coefficient 1 / 3, so as to make the quality score more intuitive.

[0174] It is understood that the execution of step S1400 has a certain degree of randomness. Therefore, this step can be executed multiple times to obtain multiple advertising copy corresponding to the advertised product. These multiple advertising copy can be used to obtain corresponding quality scores according to the technology implemented in this embodiment, thereby selecting one or more advertising copy with higher quality scores. The server can then push the advertising copy with higher quality scores to the page where the advertising copy corresponding to the advertised product was created, providing creative references to merchants and lowering the creation threshold.

[0175] In this embodiment, the quality of advertising copy is quantitatively evaluated from three dimensions: description similarity, promotional value, and textual confusion, thereby meeting the quality evaluation requirements for marketing and promotion of advertising copy.

[0176] Please see Figure 9This application provides an advertising copy generation device to meet one of its objectives. It is a functional embodiment of the advertising copy generation method of this application. The device includes a model prediction module 1100, a probability distribution module 1200, a word set construction module 1300, and a copy construction module 1400. Specifically: the model prediction module 1100 is used to acquire the product text of the advertised product, input it into a text generation model, and predict the generation probability distribution corresponding to each advertising word element mapped to the encoded word list. The product text includes product description information. The probability distribution module 1200 is used to acquire the category probability distribution preset for the product category of the advertised product and the generation probability distributions. The algorithm performs weighted summation to obtain the candidate probability distribution corresponding to each advertising term; the term set construction module 1300 is used to select multiple candidate terms that meet preset conditions from the encoded term list according to each candidate probability distribution, and select advertising phrases that start with each candidate term from the preset historical advertising phrase library corresponding to the store of the advertised product as candidate words, and construct a candidate term set corresponding to each candidate probability distribution; the copywriting construction module 1400 is used to randomly select one from its corresponding candidate term set as a corresponding word for each advertising term, or determine the term in the encoded term list as a corresponding word according to its generation probability distribution / candidate probability distribution, and construct all corresponding words as advertising copy.

[0177] In a further embodiment, before the model prediction module 1100, the system further includes: a copywriting acquisition module, used to acquire advertising copy of advertisements already placed in the advertising system, each advertising copy being placed by a store in the e-commerce platform to promote products corresponding to its description; a keyword library construction module, used to construct advertising keyword libraries corresponding to each product category in the product category system, each advertising keyword library storing advertising phrases extracted from the advertising copy of products in its corresponding product category, each advertising phrase containing two or more word units; a rating determination module, used to associate the advertising copy containing the advertising phrases to statistically determine the recommendation rating of each advertising phrase under the constraints of product category and store, the recommendation rating being determined based on the sum of the category dimension rating, store dimension rating, and advertising dimension rating of the advertising phrase; and a database construction module, used to construct a historical advertising phrase library by acquiring a portion of advertising phrases with higher recommendation ratings for different stores in the advertising keyword library corresponding to each product category based on the recommendation ratings of each advertising phrase in the advertising keyword library under the constraints of product category and store.

[0178] In a further embodiment, the model prediction module 1100 includes: an encoding submodule, used to apply the encoder of the text generation model to encode the product text of the advertising product to obtain a text feature sequence; and a decoding submodule, used to input the text feature sequence as prefix information into the decoder of the text generation model for decoding, and predict the generation probability distribution corresponding to each advertising word in the encoded vocabulary.

[0179] In a further embodiment, the probability distribution module 1200 includes: a text segmentation submodule, used to segment advertising phrases of various product categories stored in a preset historical advertising phrase library using a preset segmenter to obtain segmented word units corresponding to each advertising phrase; a bag-of-words construction submodule, used to construct a bag-of-words for each product category using a bag-of-words model, which includes segmented word units corresponding to the advertising phrases of the corresponding product category and their quantity in the bag-of-words; and a category probability distribution submodule, used to determine the word frequency corresponding to each segmented word unit in the bag-of-words for each product category, and use the word frequency corresponding to each segmented word unit of each product category as the probability of word units in the same encoded vocabulary as the segmented word units to obtain the category probability distribution.

[0180] In a further embodiment, the copywriting construction module 1400 includes: a decision probability acquisition submodule, used to acquire the corresponding decision probability for each advertising keyword, wherein the decision probabilities of some advertising keywords are different; a keyword selection object judgment submodule, used to determine, based on the decision probability, that the keyword selection object of the advertising keyword is the candidate word set or generation probability distribution / candidate probability distribution corresponding to the advertising keyword; a first corresponding word submodule, used to randomly determine one of the candidate words as the corresponding word when the keyword selection object is a candidate word set; and a second corresponding word submodule, used to determine the word in the encoded word list as the corresponding word based on its generation probability distribution / candidate probability distribution when the keyword selection object is a generation probability distribution / candidate probability distribution.

[0181] In a preferred embodiment, the decision probability acquisition submodule includes: a first initial setting unit, configured to set the decision probability of the first N advertising keywords to 0, where N is greater than or equal to 1; and a first subsequent setting unit, configured to set the decision probability of subsequent advertising keywords to decrease from a preset value higher than 0.

[0182] Alternatively, it may include: a second initial setting unit, used to set the decision probability of the first advertising keyword to a preset value higher than 0; and a second subsequent setting unit, used to set the decision probability of each advertising keyword inversely proportional to the number of candidate words that have been determined as the corresponding words in advance.

[0183] In an extended embodiment, after the copywriting construction module 1400, the system further includes: a first calculation module for determining the descriptive similarity and promotional value between the product text of the advertised product and the advertising copy; a second calculation module for determining the textual confusion of the statements in the advertising copy; and a comprehensive calculation module for integrating the descriptive similarity, promotional value, and textual confusion to obtain a quality score for the advertising copy.

[0184] To address the aforementioned technical problems, embodiments of this application also provide computer equipment. For example... Figure 10 The 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 copy 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 copy 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 10 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the 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.

[0185] In this embodiment, the processor is used to execute... Figure 9 The specific functions of each module and its sub-modules are defined within the device. 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 copy generation device of this application. The server can call the server's program code and data to execute the functions of all sub-modules.

[0186] 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 copy generation method of any embodiment of this application.

[0187] 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.

[0188] In summary, this application, by interfering with the encoding process of the text generation model, guides the model to generate high-quality advertising copy. In addition, it introduces randomness into the determination of advertising terms in the advertising copy, ensuring that the advertising copy contains an appropriate amount of high-quality advertising phrases, thereby further improving the quality of the advertising copy.

[0189] 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.

[0190] 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 copy, characterized in that, The steps include the following: The product text of the advertised product is obtained and input into the text generation model to predict the generation probability distribution of each advertising word element mapped to the corresponding encoded vocabulary. The product text includes the product description information of the product. The preset category probability distribution for the product category of the advertised product is obtained, and each of the generated probability distributions is weighted to obtain the candidate probability distribution corresponding to each advertising term. Based on each candidate probability distribution, select multiple candidate word elements corresponding to the preset conditions from the encoded word list, and select advertising phrases of the store corresponding to the advertised product that start with each of the candidate word elements from the preset historical advertising phrase library as candidate words, and construct a candidate word set corresponding to each candidate probability distribution. For each advertising term, one term is randomly selected from its corresponding candidate term set as the corresponding term, or the term in the encoded term list is determined according to its generation probability distribution / candidate probability distribution as the corresponding term, and all corresponding terms are used to construct the advertising copy.

2. The advertising copy generation method according to claim 1, characterized in that, Before obtaining the product text of the advertised product, the following steps are also included: Get the ad copy of the ads that have been placed in the advertising system. Each ad copy is placed by a store in the e-commerce platform and is used to promote the products that correspond to its description. Construct an advertising keyword library corresponding to each product category in the product category system. Each advertising keyword library stores advertising phrases extracted from the advertising copy of the corresponding product category. Each advertising phrase contains more than two word elements. The recommendation score for each ad phrase is determined by statistically analyzing the ad copy containing the ad phrase, under the constraints of product category and store. The recommendation score is determined by summarizing the category dimension score, store dimension score, and ad dimension score of the ad phrase. Based on the recommendation scores of each advertising phrase in the advertising keyword library under the constraints of product category and store, a historical advertising phrase library is constructed by obtaining the advertising phrases with higher recommendation scores for different stores in the advertising keyword library corresponding to each product category.

3. The advertising copy generation method according to claim 1, characterized in that, The step of predicting the generation probability distribution of each advertising terminology mapped to the corresponding encoded vocabulary includes the following steps: An encoder using a text generation model encodes the product text of an advertisement to obtain a text feature sequence; The text feature sequence is used as prefix information and input into the decoder of the text generation model to predict the generation probability distribution of each advertising term mapped to the corresponding encoded vocabulary.

4. The advertising copy generation method according to claim 1, characterized in that, The step of obtaining the preset category probability distribution for the product category of the advertised product includes the following steps: A preset word segmenter is used to segment the advertising phrases of each product category stored in the preset historical advertising phrase library to obtain the corresponding word segments of each advertising phrase; The bag-of-words model is used to construct the bag of words for each product category, which contains the word segments corresponding to the advertising phrases of the corresponding product category and their number in the bag of words; Determine the word frequency corresponding to each word segmentation unit in the bag-of-words for each product category, and use the word frequency corresponding to each word segmentation unit in each product category as the probability of the word segmentation unit in the same encoded vocabulary as the word segmentation unit to obtain the category probability distribution.

5. The advertising copy generation method according to claim 1, characterized in that, The step of randomly selecting a word from its corresponding candidate word set as the corresponding word for each advertising word element, or determining the word element in the encoded vocabulary as the corresponding word based on its generation probability distribution / candidate probability distribution, includes the following steps: For each ad keyword, obtain its corresponding decision probability, where the decision probabilities of some ad keywords are different; Based on the decision probability, the word selection target of the advertising term is determined to be the candidate word set or generation probability distribution / candidate probability distribution corresponding to the advertising term; When the word selection object is a candidate word set, one of the candidate words is randomly selected as the corresponding word; When the word selection object is a generation probability distribution / candidate probability distribution, the word elements in the encoding vocabulary are determined as the corresponding words according to its generation probability distribution / candidate probability distribution.

6. The advertising copy generation method according to claim 1, characterized in that, The steps for obtaining the corresponding decision probability for each advertising keyword include the following: For the first N advertising keywords, set their decision probability to 0, where N is greater than or equal to 1; For subsequent advertising keywords, their decision probability is set to decrease from a preset value higher than 0. Alternatively, it may include the following steps: For the first ad keyword, set its decision probability to a preset value that is higher than 0; For subsequent ad keywords, the decision probability of each ad keyword is set inversely proportional to the number of candidate keywords that have been identified as the corresponding words.

7. The advertising copy generation method according to claim 1, characterized in that, After constructing all corresponding words into advertising copy, the following steps are also included: Determine the descriptive similarity and promotional value between the product text and the advertising copy of the advertised product; Determine the textual perplexity of statements in the advertising copy; The quality score of advertising copy is obtained by combining description similarity, promotional value, and textual confusion.

8. An advertising copy generation device, characterized in that, include: The model prediction module is used to obtain the product text of the advertised product, input it into the text generation model, and predict the generation probability distribution of each advertising word unit mapped to the corresponding encoding vocabulary. The product text includes the product description information of the product. The probability distribution module is used to obtain a preset category probability distribution for the product category of the advertised product, and then weight each of the generated probability distributions to obtain a candidate probability distribution for each advertised word. The word set construction module is used to select multiple candidate word elements that meet preset conditions from the encoded word list according to each candidate probability distribution, and select advertising phrases that start with each of the candidate word elements from the preset historical advertising phrase library corresponding to the store of the advertised product as candidate words, and construct a candidate word set corresponding to each candidate probability distribution. The copy construction module is used to randomly select one word from its corresponding candidate word set as the corresponding word for each advertising word, or to determine the word in the encoded word list as the corresponding word based on its generation probability distribution / candidate probability distribution, and to construct advertising copy from all corresponding words.

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 7.

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

Images