Text generation method, electronic equipment and computer readable storage medium

Abstract

The invention discloses a text generation method, electronic equipment and a computer readable storage medium, and relates to the field of artificial intelligence and text generation. The method comprises the steps that a structured description text of a target object and an example text corresponding to the structured description text are obtained, and the example text is used for describing an object of the same type as the target object; the structured description text and the example text are input into a first text generation model, a plurality of initial texts corresponding to the target object are generated through the first text generation model, and the attributes of the initial texts are the same as those of the example text; putting the plurality of initial texts into a preset scene, and collecting feedback information of the plurality of initial texts; and generating a target text of the target object by using a second text generation model based on the feedback information and the plurality of initial texts. The technical problem that the text generation effect is poor in the prior art is solved.

Description

Technical Field

[0001] The present application relates to the fields of artificial intelligence and text generation, and in particular, to a text generation method, an electronic device, and a computer-readable storage medium. Background Art

[0002] In recent years, large-scale language models have developed rapidly, and their application in intelligent text generation has significantly improved the efficiency and quality of text creation. However, the decoupling between text generation and actual delivery has become a major challenge. In the traditional process, the generation, testing, selection and final delivery of texts take a long time, and there is a lack of an effective feedback mechanism to guide the improvement of the model, resulting in the generated text may not accurately match user preferences, resulting in poor text generation results.

[0003] To address the above-mentioned problems, no effective solution has been proposed yet. Summary of the invention

[0004] The embodiments of the present application provide a text generation method, an electronic device, and a computer-readable storage medium to at least solve the technical problem of poor text generation effect in the related art.

[0005] According to one aspect of an embodiment of the present application, a text generation method is provided, comprising: obtaining a structured description text of a target object and an example text corresponding to the structured description text, wherein the example text is used to describe an object of the same type as the target object; inputting the structured description text and the example text into a first text generation model, and using the first text generation model to generate a plurality of initial texts corresponding to the target object, wherein the plurality of initial texts have the same attributes as the example text; placing the plurality of initial texts into a preset scene, and collecting feedback information of the plurality of initial texts, wherein the feedback information is used to indicate a user's preference for the plurality of initial texts in the preset scene; and using a second text generation model to generate a target text for the target object based on the feedback information and the plurality of initial texts, wherein the target text conforms to the feedback information.

[0006] According to one aspect of an embodiment of the present application, a text generation method is also provided, including: obtaining a structured description text of a target product and an example title text corresponding to the structured description text, wherein the example title text is used to describe a product of the same type as the target product; inputting the structured description text and the example title text into a first text generation model, and using the first text generation model to generate multiple initial title texts corresponding to the target product, wherein the multiple initial title texts have the same attributes as the example title text; placing the multiple initial title texts into a preset scene, and collecting feedback information of the multiple initial title texts, wherein the feedback information is used to indicate the degree of user preference for the multiple initial title texts in the preset scene; and using a second text generation model to generate a target title text for the target product based on the feedback information and the multiple initial title texts, wherein the target title text conforms to the feedback information.

[0007] According to one aspect of an embodiment of the present application, a text generation method is also provided, including: in response to an input instruction acting on an operation interface, displaying a structured description text of a target object and an example text corresponding to the structured description text on the operation interface, wherein the example text is used to describe an object of the same type as the target object; in response to a processing instruction acting on the operation interface, displaying a target text of the target object on the operation interface, wherein the target text is obtained by using a second text generation model based on feedback information and multiple initial texts, the feedback information is feedback information of multiple initial texts collected by placing multiple initial texts into a preset scene, the multiple initial texts are obtained by inputting the structured description text and the example text into a first text generation model, the multiple initial texts have the same attributes as the example text, and the target text conforms to the feedback information.

[0008] According to one aspect of an embodiment of the present application, a text generation method is also provided, including: obtaining a structured description text of a target object and an example text corresponding to the structured description text by calling a first interface, wherein the first interface includes a first parameter, and the parameter value of the first parameter includes the structured description text and the example text; obtaining the structured description text of the target object and the example text corresponding to the structured description text, wherein the example text is used to describe an object of the same type as the target object; inputting the structured description text and the example text into a first text generation model, and using the first text generation model to generate a plurality of initial texts corresponding to the target object, wherein the attributes of the plurality of initial texts are the same as those of the example text; placing the plurality of initial texts into a preset scene, and collecting feedback information of the plurality of initial texts, wherein the feedback information is used to indicate the degree of user preference for the plurality of initial texts in the preset scene; based on the feedback information and the plurality of initial texts, using a second text generation model to generate a target text of the target object, wherein the target text conforms to the feedback information; outputting the target text by calling a second interface, wherein the second interface includes a second parameter, and the parameter value of the second parameter includes the target text.

[0009] According to another aspect of the embodiments of the present application, a computing device is further provided, including: a memory storing an executable program; and a processor for running the program, wherein the method in each embodiment of the present application is executed when the program is running.

[0010] According to another aspect of the embodiments of the present application, an electronic device is also provided, including: a memory storing an executable program; a processor connected to the memory via a bus and used to run the program, wherein the method of each embodiment of the present application is executed when the program is running.

[0011] According to another aspect of an embodiment of the present application, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored executable program, wherein when the executable program is running, the device where the computer-readable storage medium is located is controlled to execute the methods in each embodiment of the present application.

[0012] According to another aspect of the embodiments of the present application, a computer program product is also provided, including a computer program, and when the computer program is executed by a processor, the method in each embodiment of the present application is implemented.

[0013] According to another aspect of an embodiment of the present application, a computer program product is also provided, including a non-volatile computer-readable storage medium, wherein the non-volatile computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method in each embodiment of the present application is implemented.

[0014] According to another aspect of the embodiments of the present application, a computer program is also provided, and when the computer program is executed by a processor, the methods in the various embodiments of the present application are implemented.

[0015] In an embodiment of the present application, a structured description text of a target object and an example text corresponding to the structured description text are obtained, wherein the example text is used to describe an object of the same type as the target object; the structured description text and the example text are input into a first text generation model, and a plurality of initial texts corresponding to the target object are generated using the first text generation model, wherein the plurality of initial texts have the same attributes as the example text; the plurality of initial texts are placed in a preset scene, and feedback information of the plurality of initial texts is collected, wherein the feedback information is used to indicate the degree of user preference for the plurality of initial texts in the preset scene; based on the feedback information and the plurality of initial texts, a target text of the target object is generated using a second text generation model, wherein the target text conforms to the feedback information, thereby achieving the purpose of improving the text generation effect. It is easy to notice that a plurality of initial texts having the same attributes as the sample text can be generated by using the first text generation model through the sample text describing the object of the same type as the target object. The plurality of initial texts can be placed in a preset scene, and the initial text features with a higher degree of user preference in the preset scene can be identified by analyzing the feedback information, and the second text generation model can be used to adjust its generation strategy so that the generated target text is more likely to have these features, while reducing or avoiding the initial text features with a lower degree of user preference from appearing in the target text, thereby improving the pertinence and user acceptance of text generation, and further solving the technical problem of poor text generation effect in related technologies.

[0016] It is easy to notice that the above general description and the following detailed description are only for the purpose of exemplifying and explaining the present application, and do not constitute a limitation of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0018] Figure 1 is a schematic diagram of an application scenario according to an embodiment of the present application;

[0019] Figure 2 is a flowchart of a text generation method according to an embodiment of the present application;

[0020] Figure 3 It is a structural flow chart of commodity information in an intelligent copywriting LLM according to an embodiment of the present application;

[0021] Figure 4is a flow chart of structured information output according to an embodiment of the present application;

[0022] Figure 5 is a flow chart of a large language model generated by a copywriting according to an embodiment of the present application;

[0023] Figure 6 is a schematic diagram of a process of constructing preference training data according to an embodiment of the present application;

[0024] Figure 7 is a schematic diagram of a model training process according to an embodiment of the present application;

[0025] Figure 8 is a flowchart of a text generation method according to an embodiment of the present application;

[0026] Fig. 9 is a flowchart of a text generation method according to an embodiment of the present application;

[0027] Fig.10 is a flowchart of a text generation method according to an embodiment of the present application;

[0028] Fig.11 is a schematic diagram of a text generation device according to an embodiment of the present application;

[0029] Fig.12 is a schematic diagram of a text generation device according to an embodiment of the present application;

[0030] Fig.13 is a schematic diagram of a text generation device according to an embodiment of the present application;

[0031] Fig.14 is a schematic diagram of a text generation device according to an embodiment of the present application;

[0032] Fig.15 is a structural block diagram of a computing device according to an embodiment of the present application;

[0033] Fig.16 It is a structural block diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0034] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present application.

[0035] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0036] The technical solution provided in this application is mainly implemented by a content generation model, where the content generation model refers to a deep learning model with large-scale model parameters. The content generation model can also be called a foundation model. The content generation model is pre-trained with large-scale unlabeled corpus to produce a pre-trained model with more than 100 million parameters. This model can adapt to a wide range of downstream tasks and has good generalization capabilities, such as large-scale language models (LLM), multi-modal pre-training models, etc.

[0037] It should be noted that when the content generation model is actually applied, the pre-trained model can be fine-tuned through a small number of samples so that the content generation model can be applied to different tasks. For example, the content generation model can be widely used in natural language processing (NLP), computer vision, speech processing and other fields, and can be specifically applied to computer vision tasks such as visual question answering (VQA), image description (IC), image generation, etc. It can also be widely used in natural language processing tasks such as text-based sentiment classification, text summary generation, and machine translation. Therefore, the main application scenarios of the content generation model include but are not limited to digital assistants, intelligent robots, search, online education, office software, e-commerce, intelligent design, etc. In the embodiment of the present application, data processing through a content generation model in a text generation scenario is used as an example for explanation.

[0038] First, some nouns or terms that appear in the description of the embodiments of the present application are subject to the following explanations:

[0039] Generative optimization: In the advertising creative delivery, first generate a variety of candidate advertising creative materials, then select the materials with good delivery effects through a carousel method, and eliminate the poor materials.

[0040] Preferred generation: By accumulating data on random delivery of intelligent materials over a long period of time through small batches of goods, combined with multi-dimensional information such as delivery scenario summaries, merchant information, and store crowds, preference alignment algorithms such as direct preference optimization (DPO) and proximal policy optimization (PPO) are used to achieve the preferred generation of intelligent creative materials (LLM), get rid of the "wide-net" generation and replacement mode, and significantly improve the return on investment (ROI) of production capacity.

[0041] Retrieval-Augmented Generation (RAG): refers to the process of adjusting the output of large language models so that they can reference knowledge bases other than the training data source before generating a response.

[0042] Preference alignment: Preference alignment technology aims to allow machines to continuously improve their behavior by receiving human feedback so as to better meet human expectations.

[0043] Direct Preference Optimization (DPO): A method for improving language models based on human preferences. Unlike Reinforcement Learning from Human Feedback (RLHF), DPO does not rely on explicit reward modeling or reinforcement learning. It targets the same goal as RLHF, but provides a simpler and more direct training method.

[0044] At present, the integrated project of intelligent text generation and delivery has achieved a steady increase in the revenue per thousand impressions (RPM) of display ads in the application of main title improvement. The iteration of traditional intelligent text (including other creative materials) depends on the delivery effect, but it takes at least one week from production, release, cold start carousel, online selection to final delivery. In addition, the "wide-net" generation optimization leads to problems such as slow iteration speed and low production capacity. LLM technology is developing rapidly, and the content generation model can easily generate some high-quality copywriting. However, these copywriting may not necessarily get positive effects in actual delivery. The reason is that the generation and delivery of the copywriting are decoupled, and there is no feedback mechanism from the delivery side to the generation model, so positive effects may not be achieved.

[0045] This application abandons the traditional generative optimization and proposes an improved generation paradigm. The innovations include: 1. Proposing a structured LLM for product information to achieve controllable generation of selling point word count, etc.; 2. Proposing a RAG-based text generation LLM to improve the diversity of intelligent texts; 3. Proposing an improved generation LLM, based on actual delivery effects, using a preference alignment algorithm to generate intelligent texts with relatively good delivery effects.

[0046] According to an embodiment of the present application, a text generation method is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0047] Considering that the model parameters of the content generation model are huge and the computing resources of the mobile terminal are limited, the above method provided in the embodiment of the present application can be applied to Figure 1 The application scenarios shown are not limited to these. Figure 1 is a schematic diagram of an application scenario according to an embodiment of the present application. Figure 1 In the application scenario shown, the content generation model is deployed in the server 10, and the server 10 can be connected to one or more client devices 20 through a local area network connection, a wide area network connection, an Internet connection, or other types of data networks. The client device 20 here may include but is not limited to: a smart phone, a tablet computer, a laptop computer, a PDA, a personal computer, a smart home device, a car device, etc. The client device 20 can interact with the user through a graphical user interface to implement the call of the content generation model, thereby implementing the method provided in the embodiment of the present application.

[0048] In an embodiment of the present application, a system composed of a client device and a server can perform the following steps: the client device generates a structured description text and a sample text corresponding to the structured description text. The server inputs the structured description text and the sample text into a first text generation model, and generates multiple initial texts corresponding to a target object using the first text generation model; the multiple initial texts are placed in a preset scene, and feedback information of the multiple initial texts is collected; based on the feedback information and the multiple initial texts, a target text of the target object is generated using a second text generation model.

[0049] It should be noted that, with the rapid development of high-performance computing units, in other application scenarios, the above method provided in the embodiment of the present application can also be applied to the model all-in-one machine. In an optional embodiment, a plurality of models are built into the model all-in-one machine, and the user can choose to adjust a model as needed to obtain the user's own model, so that the high-performance computing unit built into the model all-in-one machine can directly call the adjusted model to execute the above method provided in the embodiment of the present application. In another optional embodiment, a trained model is built into the content generation model all-in-one machine, so that the high-performance computing unit built into the model all-in-one machine can directly call the model to execute the above method provided in the embodiment of the present application.

[0050] Furthermore, when users need to train their own models, they can also upload their own data sets through the client, which are sent by the client to the server, so that the server can adjust the pre-trained model with the data set to obtain the user's own model and then deploy it to the production environment. In order to facilitate users' needs for model adjustment, the server can provide complete adjustment tools, development frameworks and processes, and can support multiple adjustment strategies, so that the adjusted model can better adapt to applications in different fields and achieve a high degree of customization.

[0051] Under the above operating environment, this application provides Figure 2 The text generation method shown. Figure 2 is a flowchart of a text generation method according to an embodiment of the present application. Figure 2 As shown, the method may include the following steps:

[0052] Step S202, obtaining a structured description text of the target object and a sample text corresponding to the structured description text;

[0053] The sample text is used to describe an object of the same type as the target object.

[0054] The above-mentioned target objects refer to the goods or services for which advertising copy needs to be created, and there is no limitation here.

[0055] The above-mentioned structured description text can be a kind of refined and formatted information, which presents the key attributes of the product (such as brand, style, and function) in a standard form, making it easier for the algorithm to understand and process it, and reducing errors and illusions in the generation process.

[0056] The sample texts mentioned above refer to advertising copy samples that have performed well in the past. These copy samples demonstrate the effectiveness of attributes and expressions for similar products, and provide references for the generation of new copy.

[0057] In an optional embodiment, the original information of the target product is first collected, including the product title, pictures, videos, attributes, etc., and then a specific algorithm is used to process this information, extract key points, and form a structured description text. In addition, the system can call a sample text library consisting of high-quality copywriting that matches the type of target product. This step can ensure the accuracy and consistency of the input information, providing high-quality raw materials for subsequent intelligent copywriting generation. At the same time, by introducing sample text, the system can learn effective copywriting attributes and expressions, thereby improving the starting level of copywriting generation.

[0058] Through structured processing of information and guidance from sample texts, the diversity and quality of copywriting can be significantly improved, reducing the problem of hallucinations in the model generation process, ensuring that the generated copywriting more accurately reflects the characteristics of the product and is more attractive to target users in terms of attributes.

[0059] Step S204, inputting the structured description text and the sample text into a first text generation model, and using the first text generation model to generate a plurality of initial texts corresponding to the target object;

[0060] Among them, the properties of multiple initial texts are the same as those of the sample texts.

[0061] The first text generation model mentioned above can be an intelligent copywriting LLM, which is a trained large-scale language model that can understand the input product structured description text and example text, and generate diverse but consistent copywriting.

[0062] The above-mentioned structured description text may refer to the conversion of the original information of the product (such as title, picture, video, attributes, etc.) into a descriptive text in a standard format through multimodal information processing technology, so as to facilitate model understanding and generation.

[0063] The above-mentioned sample text can be a sample of copywriting with excellent results in history, which is used to generate copywriting with similar attributes using the model.

[0064] The above-mentioned multiple initial texts can be a series of preliminary copywriting generated by the intelligent copywriting creation LLM based on the structured description text and the sample text. While the attributes of these copywritings are consistent with the sample text, the content is adjusted for the target product.

[0065] In an optional embodiment, the system inputs the structured description text of the target object and the sample text that matches its type into the intelligent copywriting creation LLM. By understanding these input texts, the model uses its internal generation mechanism and retrieval-augmented generation technology (RAG) to generate multiple initial texts with the same attributes as the sample text. Specifically, RAG technology allows the model to retrieve and refer to high-quality text instances related to the target object when generating copywriting, ensuring that the output copywriting not only covers the core concerns of the object, but also remains consistent with the sample text in expression and attributes, thereby improving the diversity and quality of the text while reducing the illusion problems that may be encountered during the generation process. Through the generation capabilities of the intelligent copywriting creation LLM and the retrieval function of the RAG technology, the diversity and quality of the text are significantly improved, while maintaining consistency with the attributes of the sample text.

[0066] Step S206, placing the multiple initial texts into a preset scene, and collecting feedback information of the multiple initial texts;

[0067] The feedback information is used to indicate the user's preference for multiple initial texts in a preset scenario.

[0068] The above-mentioned preset scenario can be a specific display environment of the target object. For example, it can refer to a specific display scenario of an advertisement, but it is not limited thereto. The display scenario can refer to when a user visits an e-commerce platform for the first time or searches for a product, the system guesses and displays product advertisements that may attract the user's attention based on their past behavior, preferences, and search keywords. The display scenario can also refer to the system displaying advertisements for related or complementary products to the purchased product to the user after the user completes the purchase, in order to promote secondary purchases or improve user satisfaction. The display scenario can also be an advertising display scenario for non-core traffic or non-peak hours, such as non-prominent locations in recommendation flows, and advertising displays for low-activity user groups.

[0069] The above feedback information refers to the user's response data to the delivered text, including but not limited to indicators such as click-through rate, conversion rate, and user dwell time. These data can reflect how the text performs in the preset scenario and the user's preference for the text content. By analyzing the feedback information, it is possible to determine which texts can more effectively attract users, thereby improving the effectiveness of advertising.

[0070] The above-mentioned user preference refers to the user's preference for the initial text, which can usually be indirectly measured by indicators such as click-through rate in feedback information. A high click-through rate often means that the user is interested in the text content, so the user preference for the text in the preset scenario is high. Other indicators in the feedback information, such as conversion rate, can also further reveal whether the text can effectively guide users to complete purchases and other behaviors, thereby fully understanding user preferences.

[0071] In an optional embodiment, multiple initial texts (ad titles or descriptions) can be placed in different preset scenarios. These scenarios can be multiple advertising scenarios, and the user traffic and behavior patterns of each scenario are different. The system will record the specific performance of each text in different scenarios, such as the number of clicks, the number of impressions, the number of conversions, etc. Based on these data, the click-through rate and conversion rate of each text can be calculated.

[0072] By delivering multiple initial texts in preset scenarios and collecting feedback information, the system can identify the text types and expressions that users prefer. This not only increases the attractiveness of the ads, but also improves the overall advertising effectiveness, including higher user engagement, click-through rate, and conversion rate. In addition, through continuous delivery and feedback loops, the system can continuously improve the text generation model so that the generated text is more in line with user preferences, thereby achieving continuous improvement in advertising effectiveness.

[0073] Step S208: Generate a target text of the target object using a second text generation model based on the feedback information and the multiple initial texts.

[0074] Among them, the target text conforms to the feedback information.

[0075] The second text generation model mentioned above in this solution specifically refers to the intelligent copywriting optimization generation LLM, which is a trained model that can generate better text based on historical delivery results and user preferences. The role of the model at this stage is to integrate feedback information, adjust its generation strategy, ensure that the output text can more accurately align with user preferences, and improve market effects.

[0076] The target texts mentioned above are the texts used for delivery. They are the result of improvements based on the initial texts and user feedback. These texts not only contain the key information of the product, but also are more in line with the preferences of the target audience in terms of expression and attributes, which is expected to increase click-through rate and conversion rate.

[0077] In an optional embodiment, the system first analyzes the feedback information collected from the preset scenario to identify the text features with better performance, including attributes, information structure and expression methods. These features are then used as utilization signals and input into the second text generation model, which will adjust its generation strategy based on these feedback signals to generate better target text. Optionally, the model may be trained through the DPO algorithm to improve its strategy and ensure that the output text is more likely to be clicked and converted by users.

[0078] By improving the second text generation model based on user feedback information, the delivery effect of the generated text can be significantly improved. The target text is more in line with user preferences in terms of attributes and expressions, which increases click-through rate and conversion rate and attracts users. This closed-loop improvement mechanism closely links text generation with delivery effect, realizes continuous iteration and improvement of text generation, and ensures that the delivered text is a better version that has been screened and improved.

[0079] The above steps describe the core process of the integrated solution for intelligent copy generation and delivery, from product information processing to improving delivery text.

[0080] In an optional embodiment, the system extracts structured description text of the target object (such as a product), which is a systematic summary of the product's features, attributes, and selling points for model understanding and generation. At the same time, the system also collects sample texts, which represent descriptions of objects of the same type as the target object. Usually, the sample text has a positive effect on delivery and can guide the generation of texts with consistent attributes. After obtaining the structured description text and sample text, the system inputs this information into the first text generation model. The model generates multiple initial texts related to the target object based on the attributes of the sample text. These texts are consistent with the sample text in attributes, and are combined with specific information of the target object to form a diverse text sample.

[0081] Next, the system delivers these initial texts to preset scenarios to actually test the user's preference for the texts. The selection of preset scenarios is based on the analysis of traffic and target user groups, and is intended to collect feedback information on its effectiveness. Through delivery, the system collects feedback information on multiple initial texts, which reflects the user's response to the texts in the preset scenarios, such as click-through rate, conversion rate, etc., thus providing direct data on the user's preference.

[0082] Finally, based on the collected feedback information and the initial text, the system uses the second text generation model to generate the target text. The target text is an improved description that is more in line with user preferences and is expected to achieve better market results in the preset scenario. Through this closed-loop process, the system can continuously iterate and improve the text generation strategy to ensure that the target text delivered each time is a market-proven version that can attract user attention and promote conversion.

[0083] The above steps realize a closed loop from intelligent processing of product information to improvement of text delivery effect, which improves the quality of text generation and market responsiveness. By introducing structured description text, the illusion problem of generated text is reduced and the accuracy of information is improved. The use of sample text guides the consistency of text attributes and increases the attractiveness of text. The collection and analysis of feedback information provides a data basis for subsequent text improvement, ensuring that the generated target text is closer to user preferences and improving click-through rate and conversion rate.

[0084] Through the above steps, a structured description text of the target object and a sample text corresponding to the structured description text are obtained, wherein the sample text is used to describe an object of the same type as the target object; the structured description text and the sample text are input into a first text generation model, and a plurality of initial texts corresponding to the target object are generated by using the first text generation model, wherein the attributes of the plurality of initial texts are the same as those of the sample text; the plurality of initial texts are placed in a preset scene, and feedback information of the plurality of initial texts is collected, wherein the feedback information is used to indicate the user's preference for the plurality of initial texts in the preset scene; based on the feedback information and the plurality of initial texts, a target text of the target object is generated by using a second text generation model, wherein the target text meets the feedback information. information, thereby achieving the purpose of improving the text generation effect; it is easy to notice that, through the sample text describing the same type of object as the target object, a plurality of initial texts with the same attributes as the sample text can be generated by using the first text generation model, and the plurality of initial texts can be put into the preset scene. By analyzing the feedback information, the initial text features with a high degree of user preference in the preset scene are identified, and the second text generation model is used to adjust its generation strategy, so that the generated target text is more likely to have these features, while reducing or avoiding the initial text features with a low degree of user preference from appearing in the target text, thereby improving the pertinence of text generation and user acceptance, and further solving the technical problem of poor text generation effect in the related art.

[0085] In the above embodiment of the present application, based on feedback information and multiple initial texts, a target text of a target object is generated using a second text generation model, including: based on the feedback information, selecting a first text and a second text from multiple initial texts, wherein the user preference degree of the first text is greater than the user preference degree of initial texts other than the first text, and the user preference degree of the second text is less than the user preference degree of initial texts other than the second text; based on the first text and the second text, generating the target text using the second text generation model.

[0086] The above-mentioned feedback information refers to the user's response feedback to multiple initial texts in specific preset scenarios, which is used to quantify the user's preference for the text, such as user click-through rate, dwell time, conversion rate and other data indicators.

[0087] The first text and the second text mentioned above are representative samples selected from multiple initial texts, among which the user preference for the first text is relatively high, while the user preference for the second text is relatively low. They are used to guide the training of subsequent models to generate target texts that are more in line with user preferences.

[0088] The above-mentioned second text generation model can generate LLM for intelligent copywriting optimization, a trained model that can generate better target text based on existing feedback information and text samples to improve user participation and response effect in preset scenarios.

[0089] The target text mentioned above refers to the descriptive text generated by the second text generation model (LLM), which is guided and improved by preference information, and is used to more efficiently display it to users in preset scenarios, in order to obtain higher user preference and market response.

[0090] In an optional embodiment, the system generates a series of diverse initial texts through intelligent copywriting LLM based on structured descriptions and sample texts. Then, these texts are randomly placed in preset scenarios to collect user feedback on different texts, including click-through rate, conversion rate, etc., to quantify user preference. After collecting the feedback information, the system will filter out the first text (i.e., the text with a higher user preference) and the second text (the text with a lower user preference) from multiple initial texts. This selection process is essentially the system ranking the initial texts according to the feedback data, and using preference differences to guide subsequent model training.

[0091] Based on the preference differences between the first text and the second text, the system uses the second text generation model to learn and adjust to generate a target text that is closer to the user's preferences. This guidance process can be achieved using a preference alignment algorithm (such as DPO). By comparing and analyzing the user's preferences for different texts, the model can learn how to generate text that is more attractive to the target user.

[0092] For example, suppose there is a newly released children's sports shoe. The system first generates multiple initial text descriptions based on the product information description and sample text, including initial text 1, initial text 2, and initial text 3. These texts are then placed in different scenes of display advertising, and the system collects user feedback on each description. Assume that initial text 2 has a relatively high click-through rate and conversion rate, while the click-through rate and conversion rate of initial text 3 are relatively low. At this time, initial text 2 becomes the first text, and initial text 3 becomes the second text.

[0093] Based on the preference differences between the first and second texts, the system uses the second text generation model to learn how to adjust the generation strategy to generate a target text that is more attractive to users. For example, the system may use the model to pay more attention to the description of brand identity, seasonality, comfort, and sports attributes because these elements are emphasized in the first text and verified by higher user preferences.

[0094] The above steps can improve the market responsiveness of texts through a closed-loop intelligent generation mechanism. In practical applications, this process can accelerate the iteration speed of intelligently generated texts, reduce resource waste in the generation process, and ensure that the generated texts are more accurately aligned with user preferences, thereby improving delivery effects.

[0095] In the above embodiment of the present application, based on the first text and the second text, a target text is generated using a second text generation model, including: obtaining a scene description text of a preset scene and an associated description text of a target object; inputting the scene description text, the associated description text, the first text and the second text into the second text generation model, and generating the target text using the second text generation model.

[0096] The above scenario description text is the specific environment or conditions when the text is delivered. The user traffic characteristics and behavior patterns in different scenarios are different.

[0097] The above-mentioned associated description text can be additional descriptive information related to the target object. Exemplarily, the associated description text can be a product information description, such as the detailed attributes, selling points and price of the product, etc., but it is not limited to this. The associated description text can also be other types of description text associated with the target object.

[0098] In an optional embodiment, the system needs to first determine the preset scene in which the target object is located, and collect the scene description and associated description text related to the scene and the product. The scene description provides the context for the delivery, while the associated description text covers the characteristics of the merchant and the consumer, as well as the information of the product itself. After collecting the necessary information, the system provides the scene description text, the associated description text, the first text with a higher user preference and the second text with a lower user preference as input to the second text generation model. This model analyzes these inputs, especially under the guidance of the scene description and the first text, and learns how to generate target text that is more attractive to the target user group, while avoiding the poorly performing features in the second text.

[0099] Furthermore, based on the above input, the second text generation model generates target texts that are more in line with the context of the preset scenario and the information in the description of the target object. At the same time, due to the learning and adjustment of the model, the expression method and information presentation are more in line with the needs to adapt to user preferences. This process realizes a closed-loop improvement from product information to market verification, and then to model training and text generation.

[0100] By combining scene description and related description text, as well as user feedback, the second text generation model is used to generate target text that is more in line with market and user preferences, thereby improving the attractiveness and conversion efficiency of the copy.

[0101] In the above-mentioned embodiment of the present application, a first text and a second text are selected from a plurality of initial texts based on feedback information, including: in response to the feedback information being the amount of user interaction in a preset scenario, sorting the plurality of initial texts based on the amount of user interaction to obtain a sorting result; matching the plurality of initial texts based on the sorting result to obtain a plurality of text pairs, wherein the degree of text difference in different text pairs is different; screening the plurality of text pairs based on the degree of text difference to obtain a target text pair, wherein the degree of text difference of the target text pair is greater than the degree of text difference of other text pairs in the plurality of text pairs except the target text pair, and the target text pair includes the first text and the second text.

[0102] The above ranking result is a relative ranking of the initial texts in terms of user preference obtained by analyzing and ranking the user interaction amounts.

[0103] The above text pairs are two texts selected as a group from the sorted initial texts for subsequent text difference and preference comparison analysis.

[0104] The above-mentioned text difference degree refers to the difference between the first text and the second text in the text pair in terms of expression, information content or attributes, and is used to evaluate the diversity of the text generated by the model.

[0105] The above-mentioned target text pair is a pair of texts selected from multiple text pairs, which has a relatively large degree of text difference and includes texts with relatively high and relatively low user preference, and is used for training and adjusting the second text generation model.

[0106] In an optional embodiment, the system sorts multiple initial texts based on the amount of user interaction (such as click-through rate, conversion rate, etc.) in the feedback information to obtain a sorting result, which reflects the user's preference for different texts. The goal of sorting is to identify the texts that the user prefers and dislikes. Next, the system selects paired texts (i.e., text pairs) from the sorted multiple initial texts for comparison, and the degree of difference between these text pairs varies. The purpose of constructing text pairs is to subsequently analyze the impact of text differences on user preferences. On the basis of establishing multiple text pairs, the system needs to screen out the target text pair by analyzing the degree of text difference of each text pair. The target text pair here refers to a pair with a relatively large degree of text difference among all text pairs. The target text pair contains a first text with a relatively high degree of user preference and a second text with a relatively low degree of preference, which is used for subsequent training and adjustment of the guidance model.

[0107] The determination of the target text pair mentioned above helps the learning and adjustment of the second text generation model. The model will generate a target text that is more in line with the user's preferences based on the information of the first text (higher preference) and the second text (lower preference) in the target text pair. For example, if text A emphasizes brand and sports characteristics, and text B focuses on price and promotion information, the model may learn that in advertising scenarios, users prefer descriptions that include brand and sports characteristics, and thus pay more attention to these elements when generating target texts.

[0108] The above steps significantly improve the quality and diversity of text generation through a closed-loop feedback mechanism and difference analysis. Specifically, by sorting the initial text based on the amount of user interaction in the feedback information and screening out target text pairs with relatively large text differences, the system can more accurately identify and understand user preferences, and guide the intelligent copywriting optimization generation LLM to generate more attractive description texts for users. This mechanism not only speeds up the text generation and optimization process and reduces unnecessary resource consumption, but also ensures that the generated text can more accurately align with user preferences and improve key indicators such as click-through rate and conversion rate.

[0109] In the above embodiment of the present application, the method also includes: obtaining multiple sample texts of the sample object; placing the multiple sample texts into multiple sample scenarios, and collecting sample feedback information of the multiple sample texts, wherein the sample feedback information is used to indicate the user preference for the multiple sample texts in the multiple sample scenarios; selecting a first sample text and a second sample text from the multiple sample texts based on the sample feedback information, wherein the user preference for the first sample text is greater than the user preference for sample texts other than the first sample text, and the user preference for the second sample text is less than the user preference for sample texts other than the second sample text; based on the first sample text and the second sample text, generating a target sample text using an initial text generation model; adjusting the model parameters of the initial text generation model based on the first sample text, the second sample text and the target sample text to obtain a second text generation model.

[0110] The sample object mentioned above may refer to a commodity or service instance for which a description text needs to be generated.

[0111] The above-mentioned multiple sample texts may be a series of text sets for describing sample objects, generated by the initial text generation model based on the structured information of the goods or services.

[0112] The above sample scenarios refer to preset environments or conditions for delivering and testing sample texts.

[0113] The above sample feedback information is feedback data collected from users by placing sample texts in sample scenarios, which is used to measure the user's preference for different texts, including key indicators such as click-through rate and conversion rate. The first sample text and the second sample text are selected from multiple sample texts, representing sample texts with relatively high and relatively low user preferences, respectively, and are used to guide and adjust model parameters.

[0114] The target sample text is generated by the initial text generation model and adjusted based on the preference information of the first sample text and the second sample text.

[0115] The above-mentioned model parameter adjustment is to modify the parameters of the initial text generation model according to the comparative analysis of the first sample text, the second sample text and the target sample text, so as to generate text that better meets the user's preferences and scenario requirements.

[0116] In an optional embodiment, the system generates multiple sample texts based on the structured information of the sample object through the initial text generation model, and these texts describe different aspects of the sample object. Subsequently, the sample texts are placed in multiple sample scenarios to collect user feedback information in different scenarios and evaluate the user's preference for the sample texts.

[0117] The system analyzes the collected sample feedback information (such as click-through rate, conversion rate, etc.), sorts the sample texts based on user preference, and selects the first sample text (with a higher user preference) and the second sample text (with a lower user preference). This selection process is essentially a quantitative analysis of user feedback, as well as the sorting and classification of text effects. Based on the preference information of the first sample text and the second sample text, the system uses the initial text generation model to generate the target sample text. This process involves adjusting the model generation strategy in order to generate text that is more attractive to users and better meets the needs of the scene.

[0118] Finally, the system adjusts the parameters of the initial text generation model based on the comparative analysis of the first sample text, the second sample text, and the target sample text. This adjustment is usually performed using a preference alignment algorithm, with the goal of enabling the model to generate text that better meets user preferences and scenario requirements, thereby obtaining a second text generation model.

[0119] The above steps continuously iterate and improve the text generation model through feedback from actual delivery effects to generate higher quality description text. Specifically, by testing and analyzing user feedback information of multiple sample texts in multiple sample scenarios, the system can identify which description methods are more popular with users and which perform poorly. Based on these analysis results, the system uses the model to generate target sample text that is closer to user preferences, and continuously improves the model's generation strategy through model parameter adjustment to ensure that the generated text is more accurately aligned with user preferences and improve key indicators such as click-through rate and conversion rate in the scenario.

[0120] In the above embodiment of the present application, the model parameters of the initial text generation model are adjusted based on the first sample text, the second sample text and the target sample text to obtain the second text generation model, including: constructing a first loss function based on the first sample text and the target sample text, wherein the first loss function is used to increase the similarity between the target sample text and the first sample text; constructing a second loss function based on the second sample text and the target sample text, wherein the second loss function is used to reduce the similarity between the target sample text and the second sample text; adjusting the model parameters based on the first loss function and the second loss function to obtain the second text generation model.

[0121] The first sample text mentioned above refers to a text sample that is identified as having a relatively high user preference through actual delivery and user feedback information.

[0122] The above-mentioned target sample text is generated by the initial text generation model and adjusted based on the preference information of the first sample text and the second sample text, aiming to be a text that is closer to user preferences and scenario requirements.

[0123] The second sample text mentioned above is opposite to the first sample text. The second sample text is a text sample with a relatively low user preference, which is used for comparison and adjustment of model parameters.

[0124] The first loss function mentioned above is used to increase the similarity between the target sample text and the first sample text (text with high user preference) to take advantage of the model's tendency to generate text that is more favored by users.

[0125] The second loss function is used to reduce the similarity between the target sample text and the second sample text (text with low user preference) so as to use the model to avoid generating text that the user does not like.

[0126] In an optional embodiment, the system constructs a first loss function based on the first sample text and the target sample text, aiming to increase the similarity between the two, and at the same time, constructs a second loss function based on the second sample text and the target sample text, aiming to reduce the similarity between the target sample text and the second sample text (text with low user preference). By constructing these two types of loss functions, the system can guide the adjustment of model parameters from both positive and negative aspects.

[0127] After constructing the loss function, the system adjusts the model parameters of the initial text generation model according to the guidance of the first loss function and the second loss function. This adjustment process usually involves the training and iteration of the machine learning algorithm. By reducing the similarity with low-preference texts and increasing the similarity with high-preference texts, the model learns how to adjust the generation strategy to output texts that are more in line with user preferences.

[0128] By continuously adjusting the parameters of the initial text generation model, we eventually obtain a second text generation model trained with a preference alignment algorithm. This model can more accurately generate text with high user preference when describing products, reduce the cost of generating and eliminating inefficient text, and improve conversion rates.

[0129] The above steps improve the quality and market responsiveness of intelligently generated copywriting through the closed-loop feedback mechanism and the guidance of the loss function, reduce the cost of generating and eliminating inefficient copywriting, and improve the conversion rate. Specifically, by increasing the similarity with the first sample text (high user preference text) and reducing the similarity with the second sample text (low user preference text), the system can use the initial text generation model to generate more accurate and user-attracting copywriting, which not only reduces the waste of resources in generating a large number of inefficient copywriting, but also improves the click-through rate, conversion rate and user engagement in the scene, thereby achieving higher advertising effectiveness and market responsiveness.

[0130] In the above embodiment of the present application, based on the first sample text and the second sample text, the target sample text is generated by using the initial text generation model, including: obtaining sample scene description texts of multiple sample scenes and sample association description texts of sample objects; inputting the sample scene description texts, the sample association description texts, the first sample text and the second sample text into the initial text generation model, and generating the target sample text by using the initial text generation model.

[0131] The above-mentioned sample scenario description text provides the environment or condition information of the sample preset scenario, such as user traffic, purchase intention, etc., so as to use the model to generate copy that is more in line with the characteristics of the scenario.

[0132] The above-mentioned sample-related description text is information text closely related to the sample object. For example, the sample-related description text may include detailed attributes, selling points, brand information, price, etc., as well as information such as merchant reputation, consumer purchasing power and preferences, to enhance the personalization and accuracy of the model-generated copy.

[0133] In an optional embodiment, the system first collects sample scene description texts of multiple sample scenes and sample associated description texts of sample objects. This information covers scene characteristics and product characteristics, as well as background information of merchants and consumers, and constitutes a comprehensive information basis for guiding the model to generate copy. The above-collected sample scene description texts, sample associated description texts, first sample texts (with a relatively high degree of user preference) and second sample texts (with a relatively low degree of user preference) are input into the initial text generation model. The system uses this information to generate new target sample texts using the model. The input of multi-dimensional information can ensure that the copy generated by the model is more personalized and fits the scene requirements. At the same time, by comparing the first sample text and the second sample text, the model can learn the expression method and information presentation that are more popular with users.

[0134] By combining multi-dimensional information input, the initial text generation model is guided to generate more accurate and effective copy. Specifically, the system collects and analyzes user feedback on multiple sample texts in different sample scenarios, and determines the first sample text with a higher user preference and the second sample text with a lower user preference. Subsequently, based on the scene description, product information, merchant and consumer background information, the system uses the model to generate target sample text that is closer to user preferences and scene requirements. This process is actually a closed-loop improvement based on user feedback information, which enables the generated copy to more accurately attract the target user group, improve click-through rate, conversion rate and user engagement, and thus achieve higher advertising effectiveness and market responsiveness.

[0135] By combining multi-dimensional information input, the personalization and scene adaptability of the copy are significantly enhanced, and the user's willingness to click and buy is improved. Specifically, by inputting information such as sample scene description text and sample associated description text into the initial text generation model, the system can use the model to generate copy that is more in line with user needs and scene characteristics, which not only improves the attractiveness of the copy, but also ensures that the copy can more accurately convey the characteristics of the product and stimulate the user's purchasing interest. This mechanism not only reduces the waste of resources in the generative optimization process through closed-loop feedback and the integration of multi-dimensional information, but also continuously iterates and improves the text generation strategy according to the dynamic changes of user preferences and scene needs, ensuring the market effect and personalization level of the copy.

[0136] In the above embodiment of the present application, obtaining a structured description text of a target object and an example text corresponding to the structured description text includes: obtaining multimodal data of the target object, wherein the multimodal data is used to describe the target object through different modalities; converting the multimodal data into an initial description text of the target object; structuring the initial description text to obtain a structured description text; and retrieving an example text from a database based on the type of the target object, wherein the database is used to store example texts of objects of different types.

[0137] The above-mentioned multimodal data refers to data about the target object collected through multiple information modes (such as text, images, videos, audio, etc.). Structured description text: The text information converted and processed from multimodal data contains clear and organized key information about the target object, such as product attributes, brands, functions, etc.

[0138] The above-mentioned initial description text is text information extracted directly from multimodal data without structured processing. It is usually messy or lengthy and requires further processing before it can be used for model input.

[0139] The above database stores a collection of sample texts of different types of objects. These sample texts can be verified and effective copywriting or descriptions, which are used to guide and inspire the model to generate higher quality texts.

[0140] The above sample texts are representative text samples of specific types of objects, which are used to enhance the model generation capabilities. By learning the expression and information structure of these sample texts, the model can generate copy that is more in line with scene requirements and user preferences.

[0141] In an optional embodiment, the system first obtains multimodal data of the target object, which may include multiple information sources such as product titles, attribute descriptions, pictures, videos, etc., constituting a comprehensive description of the target object. Subsequently, the system converts the collected multimodal data into initial description text. This process may involve technologies such as optical character recognition, image annotation, and video annotation to ensure the integrity and accuracy of the text information. The obtained initial description text is processed through key information summary and structured LLM model to extract and organize the key information of the product, such as brand, appearance, ingredient materials, efficacy and function, etc., to obtain a structured description text. This step is to convert lengthy and messy text information into a clear, organized structured form to facilitate the model to understand and generate more accurate copywriting. Finally, based on the type of target object, the system retrieves relevant sample texts from the database. These sample texts provide a reference to help the model learn and imitate effective copywriting attributes and information presentation methods to generate higher quality and more user-attractive text.

[0142] The above steps significantly enhance the model's input information and generation capabilities through the fusion processing and structured transformation of multimodal data, reduce the LLM illusion problem, and improve the accuracy and quality of copy generation. Specifically, by collecting multimodal data of products and converting and structuring it into clear, organized structured description text, the system can provide the model with more accurate and comprehensive product information, thereby generating copy that is more in line with product characteristics and user preferences. In addition, by retrieving example texts of specific types of objects in the database, the model can learn a wider and more diverse range of copy attributes, improve the diversity and attractiveness of the generated copy, and demonstrate the innovation and practicality of the technical solution in the field of e-commerce advertising copy generation.

[0143] Through structured information input and efficient retrieval mechanism, the problem of hallucination in the LLM model when generating copy is significantly reduced, and the accuracy and quality of the copy are improved. Specifically, structured processing can ensure that the model receives clear and orderly product information, thereby avoiding the appearance of statements that are inconsistent with the actual product during the generation process, and improving the accuracy and credibility of the copy. At the same time, the mechanism of retrieving sample text provides a rich learning resource for the model, helping the model to generate more attractive and diverse copy, further improving users' willingness to click and buy, and demonstrating the practical value and market potential of the technical solution in the field of e-commerce advertising copy generation.

[0144] This application includes three core algorithm modules, namely, a large language model for intelligent copywriting creation, a large language model for intelligent copywriting optimization generation, and project efficiency improvement and customer expression. For the large language model for intelligent copywriting creation, the open source LLM can be adjusted through structured input of product information, imitation of intelligent copywriting attributes based on RAG, and building a data flywheel by collecting high-quality copywriting from users / merchants, so as to improve the generation quality and diversity of intelligent copywriting LLM and reduce the illusion problem. For the large language model for intelligent copywriting optimization generation, the data of random delivery of intelligent copywriting can be accumulated through small batches of goods for a long time, combined with multi-dimensional information such as delivery scenario summary, merchant information, and store crowd, and preference alignment algorithms such as DPO / PPO can be used to realize intelligent copywriting optimization generation LLM, get rid of the "wide-net" generation and replacement mode, and significantly improve production capacity ROI. For project efficiency improvement and customer expression, the advertiser title is used as the entry point to achieve project efficiency improvement in different scenarios, and complete the tools for copywriting optimization generation capabilities, so as to improve the customer usage and adoption rate of copywriting generation and recommendation tools.

[0145] Figure 3 This is a structured flow chart of product information in an intelligent copywriting LLM according to an embodiment of the present application, mainly to improve and control the generation of copywriting, to ensure that the copywriting can accurately reflect the characteristics of the product and meet the specific needs of advertising. The figure contains multiple key structures and information flows, aiming to illustrate the conversion process from raw product information to structured and controllably generated intelligent copywriting.

[0146] Figure 3 It contains three core algorithm modules: intelligent copywriting creation language model, intelligent copywriting optimization generation language model, project efficiency improvement and customer expression. In the intelligent copywriting creation language model, an excellent title sample library can be obtained according to manual design, production and operation manual mining, and data flywheel. The excellent title sample library and product structured information can be input into the intelligent copywriting creation language model to obtain intelligent creative titles. Various creative titles, such as merchant product titles (for example, main titles, natural traffic titles, etc.) can be published as intelligent materials. In the intelligent copywriting optimization generation language model, the intelligent material is input into the random carousel experiment bucket, the scene statistical effect is analyzed, and the click-through rate sorting result (Click-Through Rate Sorting, referred to as CTR) is obtained. Information such as products, merchants, users, delivery scenarios, and click-through rate sorting results can be input into the intelligent copywriting optimization generation language model, and the intelligent copywriting optimization generation language model is used to process it to obtain high-quality copywriting for products in specific delivery scenarios. In project efficiency improvement and customer expression, we can obtain mainstream large-scale delivery and recommended imitation samples of new language attributes based on the high-quality copy of the product in the specific delivery scenario.

[0147] The data flywheel mentioned above refers to a continuous cycle of data collection, analysis and re-application, which aims to continuously improve the performance of models or algorithms. The random carousel experiment bucket mentioned above is a common A / B testing strategy in the field of online advertising or recommendation systems, which is used to evaluate the effectiveness of different creative or recommended content. A small portion of traffic or display opportunities can be randomly assigned to a group of experimental smart copywriting to observe and measure their performance in the actual environment.

[0148] For the above-mentioned intelligent copywriting creation large language model, when the input is only the main title of the product and lacks complete product information, there are problems such as insufficient diversity of generated results and serious hallucination problems. When the input is the main title of the product and text information extracted from visual materials, the product information is lengthy and messy, and there is no standard format classification. There are problems such as lengthy input information, overly complex model tasks, inaccurate acquisition of product core information, and serious hallucination problems. This application can summarize and summarize various messy and lengthy original product information based on a multimodal model, and output clean and semantically structured information.

[0149] Figure 4 is a flow chart of structured information output according to an embodiment of the present application, such as Figure 4 As shown, it includes input information, algorithm module and key information structured output. Among them, the input information can be divided into four categories. The first category is text such as titles and attributes. The second category is detail page pictures. The text in the picture can be converted into editable and searchable text through optical character recognition (OCR). The third category is product pictures. The text describing the picture content can be automatically generated through the picture caption (Caption). The fourth category is product videos. The text describing the video content can be automatically generated through the video caption. Among them, the algorithm module is implemented through key information summary and structured large language model. Among them, the key information structured output can be a structured output of the key information of the product.

[0150] Without fine-tuning the large language model, the problem of hallucination when the model generates titles can be effectively alleviated by referencing the structured product information. For example, the original input information causes the model to generate an incorrect brand name. After introducing structured information, the brand name and other information can be correctly extracted.

[0151] The "hallucination problem" when generating titles with large language models (LLMs) usually refers to the fact that when the model generates text, it may generate content that is inconsistent with the input information, such as incorrect brand names, non-existent product features, or information that is irrelevant to the product. This problem may be caused by factors such as bias in the model training data, inaccurate understanding of the model's context, and randomness in the generation process.

[0152] By introducing the structured product information, the hallucination problem can be effectively alleviated and the brand name can be correctly extracted. The input format can be clarified. Structured information provides a clear and consistent input format, which enables the model to more accurately locate and understand the various attributes of the product, including the brand name, when processing data. Structured data is usually organized by category, such as "product name", "brand", "effect", etc., which helps the model learn the association and distinction between different attributes. It can reduce interference information. The original input information may contain a large amount of irrelevant or disorganized data, which may interfere with the generation process of the model. Through structuring, irrelevant information can be filtered out, and only the product features that are critical to generating titles, such as brand, appearance, material, etc., can be retained. This helps the model focus on important information and reduce generation errors. It can enhance signal strength. When the signal received by the model is clearer and more targeted, it can more accurately capture the specific attributes of the product, such as brand. In structured data, brand names are usually clearly marked, which provides clear instructions to the model to give priority to such information when generating titles, thereby avoiding the generation of incorrect brand names.

[0153] With structured information, higher quality training data sets can be constructed. These data sets can better reflect the real attributes and characteristics of the product, allowing the model to better understand and remember brand information during the learning process, so that it can more accurately call and use this information when generating titles. Structured information can help the model better understand the context of the product, that is, the product's position in the market, the target consumer group, the product's characteristics, etc. This understanding helps the model generate titles that are highly consistent with the product's attributes, including the correct brand name.

[0154] The introduction of structured information not only reduces the complexity of model processing, but also improves the accuracy and relevance of the content generated by the model. It is an effective way to alleviate the problem of hallucination and improve the accuracy of brand names. In specific applications, structured information clearly defines the various dimensions of the product, allowing the model to directly reference this information when generating copywriting, avoiding the fuzzy extraction of key information such as brand names from unstructured text, thereby greatly reducing the possibility of generation errors.

[0155] Based on the large language model of retrieval-enhanced copywriting generation, efficient generation of various styles of title copywriting is the key to optimal generation. The way of generating specific styles of title copywriting through prompt words is relatively inefficient when adding new copywriting types. When adding new title types, it is necessary to redesign the prompt words and train the model. This application can abstractly upgrade the generation of prompt word control to a general imitation model. The model can be trained only once. In the future, it is only necessary to accumulate different title copywriting examples. By retrieving example copywriting of similar products through RAG, new titles of corresponding attribute types can be produced without training the model again, which greatly improves efficiency.

[0156] Figure 5 is a flowchart of generating a large language model according to an embodiment of the present application, such as Figure 5 As shown in the figure, the product structured information is used as input and sent to the intelligent copywriting large language model to provide a detailed description and key features of the product. The preferred example copy is also used as input to provide the model with examples of copywriting of various attributes and types. The model will use the retrieval enhancement method to retrieve example copywriting similar to the current product in order to imitate the attributes and expressions therein. The intelligent copywriting large language model generates new title copywriting with consistent attributes by learning the product structured information and imitating the samples in the preferred example copywriting library. The generated title copy is used for actual advertising, and the attractiveness and effectiveness of the copywriting are evaluated through real market feedback to obtain the preferred example copywriting.

[0157] The above-mentioned product structured information is a structured expression of the detailed attributes of the product, including key information such as brand, appearance, material, and function, organized in a unified format to make the information clear and easy to process. The intelligent copywriting large language model is trained to understand and imitate the writing attributes and structure of the input examples to generate new copywriting content. Based on the retrieval-enhanced method, information in the external knowledge base can be retrieved and referenced during the generation process to improve the accuracy and diversity of the generated copy.

[0158] The above-mentioned preferred sample copy library is a database that collects various high-quality copy, including examples of title copy with different attributes and types. These examples can be successful advertising copy in the past, high-quality titles customized by merchants, or excellent copy from other sources, which are used to utilize and inspire the generation of new copy. Delivery refers to applying the generated copy title to actual advertising delivery to test its effect and improve its effect.

[0159] Figure 5The whole process from product information to generating and improving title copywriting is outlined. The core of this process is to link product structured information and the preferred sample copywriting library through RAG technology, so that the intelligent copywriting large language model can generate more accurate copywriting that meets specific attributes. Specifically, the model first receives structured product information, which includes the detailed attributes of the product, and then retrieves copywriting examples that match the current product attributes in the preferred sample copywriting library through RAG technology. The model imitates the attributes and structure of these examples to generate new title copywriting, which not only contains the key information of the product, but also has attractive writing attributes similar to the examples. The generated title copywriting will then be used for actual advertising, and its effect will be verified through market feedback. If the effect is good, these copywriting may be added to the preferred sample copywriting library to provide more high-quality reference styles for future copywriting generation. The whole process forms a cycle, continuously accumulating and improving the generation ability of the model, realizing the diversity generation and efficiency improvement of title copywriting without repeatedly training or adjusting the model.

[0160] After a small-scale carousel experiment and accumulating effect data, the intelligent copywriting large language model is used to align the effect of the generated model based on CTR ranking.

[0161] The training data involved in this application can be composed of three parts, namely, instructions, preferred samples (Chosen) and eliminated samples (Rejected). The Instruction data contains specific guidance information, usually including detailed information of the product (such as product title, structured information, merchant information, etc.) and specific description of the delivery scenario. In the Chinese context, it can be understood as "instructions", which provides the model with contextual information on how to generate copy. Chosen refers to copy samples that have been screened out from historical delivery and have relatively good results. These copies usually have a high CTR (click-through rate) or conversion rate, so they are selected as positive learning materials for the model to help the model understand which copies are more likely to attract the attention of the target audience in a specific scenario. In contrast to the Chosen samples, the Rejected data contains copy samples that have poor delivery effects under the same product and scenario conditions. Although these samples seem negative, they are equally important because they can help the model learn to avoid generating similar inefficient copy.

[0162] The construction of preference training data aims to enable the intelligent copywriting optimization generation LLM to learn how to generate more attractive and effective copywriting in different products and delivery scenarios through the three types of information: Instruction, Chosen, and Rejected. This training method can improve the copywriting generation quality of the model by comparing positive and negative samples, so that it can not only accurately reflect product information, but also take into account the preferences of the target audience, thereby achieving better advertising effects in actual applications.

[0163] Figure 6 is a schematic diagram of a process of constructing preference training data according to an embodiment of the present application, such as Figure 6 As shown in the figure, the construction of preference training data is a closed-loop and dynamically improved process, which aims to enable the intelligent copywriting optimization generation large language model to generate higher quality and more suitable copywriting for the delivery scenario through comparative learning. The model first receives a description containing product information and delivery scenarios, such as delivery scenario description, merchant level description, store population stratification (consumer purchasing power description), and product structured information. It can be trained through machine learning training methods (Direct Preference Optimization, referred to as DPO). It can filter out high-confidence sample pairs from small-scale delivery data based on multiple conditions (page views, click-through rate, conversion rate). The model will analyze the preferred samples and eliminated samples. Through comparative learning, the model identifies which copywriting is more popular and which copywriting is not effective. This process allows the model to "learn" to distinguish the characteristics of good and bad copywriting, such as language attributes, information presentation methods, emotional tonality, etc. Based on the results of comparative learning, the model improves its generation strategy to improve the click-through rate and conversion rate of the output copy. When generating new copy, the model will refer to the product and scenario information provided in the instruction, and draw on the advantages of the preferred samples, avoid the defects of the eliminated samples, and generate copy that is more in line with the preferences of the target audience. The generated copy will be actually put into use, and its effect will be evaluated through market feedback. The effect data will be used to update the sample library of the preferred samples and the eliminated samples, thus forming a data flywheel and continuously improving the model generation strategy.

[0164] Through this process, the intelligent copywriting optimization-based content generation model can be continuously iterated to improve its copywriting generation capabilities for specific products and scenarios, realize the commercialization of the copywriting optimization-based generation paradigm, significantly improve the copywriting delivery effect, and reduce the inefficiency of generative optimization. This training method enhances the model's "preference alignment" capability by comparing positive and negative samples, allowing it to more accurately reflect the needs and preferences of the target audience when generating copy.

[0165] For the training of large language models for intelligent copywriting optimization generation, we can first obtain a fusion model that can generate diverse titles. This fusion model is used as a reference model. At the same time, the strategy model is initialized using the parameters of the fusion model, and the output preference of the strategy model is improved through the direct preference loss (DPO Loss), making the probability of outputting positive samples higher.

[0166] Figure 7 is a schematic diagram of a model training process according to an embodiment of the present application, such as Figure 7 As shown in the figure, the function of the fusion model is to integrate various product information and scenario requirements to generate diversified title copy. The diversity generation capability of the fusion model is crucial for subsequent preference alignment. The eliminated samples and the preferred samples can be input into the trained fusion model (reference model) and the policy model. The trained fusion model is used as a benchmark for improving the policy model and is used to compare the output of the policy model. The policy model initializes parameters from the fusion model (i.e., reference model) and is improved through direct preference loss to learn to generate copy titles that users prefer.

[0167] Through the above process, the copy generated by the model can be effectively improved to make it closer to the user's real preferences, thus achieving better results in actual advertising. This method directly integrates user feedback into the model training process, avoiding the long-term iteration cycle and inefficient "wide-net" generation of traditional generative optimization, significantly improving the quality of copy generation and the efficiency of advertising.

[0168] Compared with the original product titles, the titles generated by the DPO preference model can put some important information of the product in the front, such as core selling points, product endorsements and brands, and the expression is also more attractive to users.

[0169] Through the construction of data flywheel, the introduction of product structured information to reduce the illusion of large language model, the intelligent copywriting large language model with controllable generation of selling point word count and other multi-dimensional, and the optimization-based generation of large language model for delivery effect and other core improvements, the integration of intelligent copywriting generation and delivery is achieved. This application proposes a structured large language model of product information to achieve controllable generation of selling point word count, etc.; proposes RAG-based copywriting generation LLM to improve the diversity of intelligent copywriting; proposes optimization-based generation LLM, based on actual delivery effect, using preference alignment algorithm to generate intelligent copywriting with relatively better delivery effect.

[0170] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of relevant countries and regions, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0171] It should be noted that, for the aforementioned method embodiments, for the sake of simplicity, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the present application is not limited by the described order of actions, because according to the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present application.

[0172] Through the description of the above implementation methods, those skilled in the art can clearly understand that the method according to the above embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a number of instructions for a terminal device (which can be a mobile phone, computer, server, or network device, etc.) to execute the methods of each embodiment of the present application.

[0173] Figure 8 is a flowchart of a text generation method according to an embodiment of the present application, such as Figure 8 As shown, the method may include the following steps:

[0174] Step S802, obtaining a structured description text of the target product and an example title text corresponding to the structured description text;

[0175] The example title text is used to describe the same type of product as the target product.

[0176] The system collects and processes detailed information about the target product and converts it into structured description text. At the same time, the system retrieves sample title texts that match the target product type from the historical database. These sample titles represent known copywriting attributes and content structures that attract user attention.

[0177] Step S804, inputting the structured description text and the sample title text into a first text generation model, and using the first text generation model to generate a plurality of initial title texts corresponding to the target product;

[0178] Among them, the properties of multiple initial title texts are the same as those of the sample title texts.

[0179] The system inputs the structured description text and the sample title text into a pre-trained first text generation model, and uses the model to generate a series of initial title texts with similar attributes to the sample title texts. These texts contain key information about the target products while trying to imitate the attractiveness and expression of the sample titles.

[0180] Step S806, placing multiple initial title texts into a preset scene, and collecting feedback information of the multiple initial title texts;

[0181] The feedback information is used to indicate the user's preference for multiple initial title texts in a preset scenario.

[0182] The initial title text generated above will be placed in the preset scenario, and the system will collect user feedback on these texts, including click-through rate, dwell time, etc., to quantify the user's preference for different attributes and content.

[0183] Step S808: Based on the feedback information and the multiple initial title texts, a target title text of the target product is generated using a second text generation model.

[0184] Among them, the target title text conforms to the feedback information.

[0185] Based on the feedback information of the initial title text, the system uses the second text generation model to generate the final target title text. This improvement process will ensure that the target title text is closer to user preferences, can effectively attract user clicks and attention in the preset scenario, and improve the market responsiveness of the text.

[0186] By providing accurate product information through structured description text, sample title text guides text generation attributes, and user feedback information is used to evaluate and improve the copywriting effect, the final generated title text can more accurately align user preferences and scenario requirements, improve conversion rate and user engagement. Through this process, the system can generate higher quality and more market-friendly product title copywriting, which not only improves the attractiveness of the copywriting, but also reduces the cost of generating and optimizing inefficient copywriting, demonstrating the innovation and commercial value of the technical solution in the field of e-commerce copywriting generation. Through the closed-loop improvement mechanism and the integration of multi-dimensional information, the system significantly improves the quality and efficiency of intelligently generated text, providing merchants with more efficient and accurate text generation strategies.

[0187] Through the above steps, a structured description text of the target product and an example title text corresponding to the structured description text are obtained, wherein the example title text is used to describe a product of the same type as the target product; the structured description text and the example title text are input into a first text generation model, and a plurality of initial title texts corresponding to the target product are generated using the first text generation model, wherein the attributes of the plurality of initial title texts are the same as those of the example title text; the plurality of initial title texts are placed in a preset scene, and feedback information of the plurality of initial title texts is collected, wherein the feedback information is used to indicate the degree of user preference for the plurality of initial title texts in the preset scene; based on the feedback information and the plurality of initial title texts, a target title text of the target product is generated using a second text generation model, wherein , the target title text conforms to the feedback information, and the purpose of improving the text generation effect is achieved; it is easy to notice that, through the sample text describing the same type of object as the target object, a plurality of initial texts with the same attributes as the sample text can be generated by the first text generation model, and the plurality of initial texts can be put into the preset scene. By analyzing the feedback information, the initial text features with higher user preference in the preset scene are identified, and the second text generation model is used to adjust its generation strategy, so that the generated target text is more likely to have these features, and at the same time, the initial text features with lower user preference are reduced or avoided from appearing in the target text, thereby improving the pertinence and user acceptance of text generation, and thus solving the technical problem of poor text generation effect in the related art.

[0188] Fig. 9 is a flowchart of a text generation method according to an embodiment of the present application, such as Fig. 9 As shown, the method may include the following steps:

[0189] Step S902, in response to an input instruction acting on the operation interface, displaying a structured description text of the target object and an example text corresponding to the structured description text on the operation interface;

[0190] The sample text is used to describe an object of the same type as the target object.

[0191] When the user enters an instruction or selects a target object on the operation interface, the system responds and displays the structured description text of the target object and the sample text of the same type of object. The structured description text provides clear information about the target object, and the sample text provides a reference for the attributes and content for generating high-quality text.

[0192] Step S904, in response to the processing instruction acting on the operation interface, displaying the target text of the target object on the operation interface;

[0193] Among them, the target text is obtained by using the second text generation model based on the feedback information and multiple initial texts, the feedback information is the feedback information of multiple initial texts collected by putting multiple initial texts into a preset scene, the multiple initial texts are obtained by inputting structured description texts and sample texts into the first text generation model using the first text generation model, the multiple initial texts have the same attributes as the sample text, and the target text conforms to the feedback information.

[0194] In response to further instructions from the user on the operation interface, the system displays the target text that has been improved based on the user's feedback. This target text is generated through the following process: First, the structured description text and the sample text are input into the first text generation model, and a series of initial texts with the same properties as the sample texts are generated; then, these initial texts are put into the preset scene, and the system collects and analyzes the feedback information to determine the user's preference for the text; finally, based on the feedback information, the second text generation model generates a target text that is closer to the user's preference, and this text is displayed on the operation interface for the user to view or use.

[0195] Through the above steps, in response to the input instruction on the operation interface, the structured description text of the target object and the sample text corresponding to the structured description text are displayed on the operation interface, wherein the sample text is used to describe an object of the same type as the target object; in response to the processing instruction on the operation interface, the target text of the target object is displayed on the operation interface, wherein the target text is obtained by using the second text generation model based on the feedback information and the multiple initial texts, the feedback information is the feedback information of the multiple initial texts collected by putting the multiple initial texts into the preset scene, the multiple initial texts are obtained by inputting the structured description text and the sample text into the first text generation model, the multiple initial texts have the same attributes as the sample text, and the target The text conforms to the feedback information, and the purpose of improving the text generation effect is achieved; it is easy to notice that a plurality of initial texts with the same attributes as the sample text can be generated by using the first text generation model through the sample text that describes the same type of object as the target object, and the plurality of initial texts can be placed in the preset scene. By analyzing the feedback information, the initial text features with a higher degree of user preference in the preset scene are identified, and the second text generation model is used to adjust its generation strategy, so that the generated target text is more likely to have these features, while reducing or avoiding the initial text features with a lower degree of user preference from appearing in the target text, thereby improving the pertinence of text generation and user acceptance, and thus solving the technical problem of poor text generation effect in the related art.

[0196] Fig.10 is a flowchart of a text generation method according to an embodiment of the present application, such as Fig.10 As shown, the method may include the following steps:

[0197] Step S1002, acquiring a structured description text of a target object and a sample text corresponding to the structured description text by calling a first interface;

[0198] The first interface includes a first parameter, and a parameter value of the first parameter includes a structured description text and an example text, and the example text is used to describe an object of the same type as the target object.

[0199] The above-mentioned first interface can be an interface for data interaction between the cloud server and the client. The structured description text and the example text corresponding to the structured description text can be passed into the interface function as the first parameter of the interface function to achieve the purpose of uploading the structured description text and the example text corresponding to the structured description text to the cloud server.

[0200] Step S1004, inputting the structured description text and the sample text into a first text generation model, and using the first text generation model to generate a plurality of initial texts corresponding to the target object;

[0201] Among them, the properties of multiple initial texts are the same as those of the sample texts.

[0202] Step S1006, placing the multiple initial texts into a preset scene, and collecting feedback information of the multiple initial texts;

[0203] The feedback information is used to indicate the user's preference for multiple initial texts in a preset scenario.

[0204] Step S1008, generating a target text of the target object using a second text generation model based on the feedback information and the multiple initial texts;

[0205] Among them, the target text conforms to the feedback information.

[0206] Step S1010: output the target text by calling the second interface.

[0207] The second interface includes a second parameter, and a parameter value of the second parameter includes a target text.

[0208] The above-mentioned second interface can be an interface for data interaction between the cloud server and the client. The cloud server can pass the target text into the interface function as the second parameter of the interface function to achieve the purpose of sending the target text to the client.

[0209] Through the above steps, the structured description text of the target object and the sample text corresponding to the structured description text are obtained by calling the first interface, wherein the first interface includes a first parameter, and the parameter value of the first parameter includes the structured description text and the sample text, and the sample text is used to describe an object of the same type as the target object; the structured description text and the sample text are input into the first text generation model, and the first text generation model is used to generate a plurality of initial texts corresponding to the target object, wherein the attributes of the plurality of initial texts are the same as those of the sample text; the plurality of initial texts are placed in a preset scene, and feedback information of the plurality of initial texts is collected, wherein the feedback information is used to indicate the user's preference for the plurality of initial texts in the preset scene; based on the feedback information and the plurality of initial texts, a target text of the target object is generated using a second text generation model, wherein the target text conforms to the feedback information; The target text is output by calling the second interface, wherein the second interface includes a second parameter, and the parameter value of the second parameter includes the target text information, thereby achieving the purpose of improving the text generation effect; it is easy to notice that a plurality of initial texts with the same attributes as the sample text can be generated by using the first text generation model through the sample text that describes the object of the same type as the target object, and the plurality of initial texts can be delivered to the preset scene, and the initial text features with a higher degree of user preference in the preset scene can be identified by analyzing the feedback information, and the second text generation model can be used to adjust its generation strategy, so that the generated target text is more likely to have these features, while reducing or avoiding the initial text features with a lower degree of user preference from appearing in the target text, thereby improving the pertinence and user acceptance of text generation, and thus solving the technical problem of poor text generation effect in the related art.

[0210] According to an embodiment of the present application, a text generation device for implementing the above-mentioned text generation method is also provided. Fig.11 is a schematic diagram of a text generation device according to an embodiment of the present application, such as Fig.11 As shown, the device 1100 includes: an acquisition module 1102 , a first generation module 1104 , a collection module 1106 , and a second generation module 1108 .

[0211] Among them, the acquisition module is used to obtain the structured description text of the target object and the example text corresponding to the structured description text, wherein the example text is used to describe the object of the same type as the target object; the first generation module is used to input the structured description text and the example text into the first text generation model, and use the first text generation model to generate multiple initial texts corresponding to the target object, wherein the multiple initial texts have the same attributes as the example text; the collection module is used to put the multiple initial texts into a preset scene and collect feedback information of the multiple initial texts, wherein the feedback information is used to indicate the user's preference for the multiple initial texts in the preset scene; the second generation module is used to generate the target text of the target object based on the feedback information and the multiple initial texts using the second text generation model, wherein the target text conforms to the feedback information.

[0212] It should be noted that the acquisition module 1102, the first generation module 1104, the acquisition module 1106, and the second generation module 1108 correspond to steps S202 to S208 in the above embodiment, and the four modules and the corresponding steps implement the same examples and application scenarios, but are not limited to the contents disclosed in the above embodiment. It should be noted that the above modules or units can be hardware components or software components stored in a memory and processed by one or more processors, and the above modules can also be run in the server 10 provided in the above embodiment as part of the device.

[0213] In the above embodiment of the present application, the second generation module is also used to select a first text and a second text from multiple initial texts based on feedback information, wherein the user preference level of the first text is greater than the user preference level of the initial texts other than the first text, and the user preference level of the second text is less than the user preference level of the initial texts other than the second text; based on the first text and the second text, a target text is generated using a second text generation model.

[0214] In the above embodiment of the present application, the second generation module is also used to obtain the scene description text of the preset scene and the associated description text of the target object; input the scene description text, the associated description text, the first text and the second text into the second text generation model, and generate the target text using the second text generation model.

[0215] In the above embodiment of the present application, the second generation module is also used to respond to feedback information of the amount of user interaction in a preset scenario, sort multiple initial texts based on the amount of user interaction to obtain a sorting result; match the multiple initial texts based on the sorting result to obtain multiple text pairs, wherein the degree of text difference in different text pairs is different; screen the multiple text pairs based on the degree of text difference to obtain a target text pair, wherein the degree of text difference of the target text pair is greater than the degree of text difference of other text pairs in the multiple text pairs except the target text pair, and the target text pair includes a first text and a second text.

[0216] In the above embodiment of the present application, the device is also used to obtain multiple sample texts of the sample object; put the multiple sample texts into multiple sample scenes, and collect sample feedback information of the multiple sample texts, wherein the sample feedback information is used to represent the user preference for the multiple sample texts in the multiple sample scenes; select a first sample text and a second sample text from the multiple sample texts based on the sample feedback information, wherein the user preference for the first sample text is greater than the user preference for sample texts other than the first sample text, and the user preference for the second sample text is less than the user preference for sample texts other than the second sample text; based on the first sample text and the second sample text, generate a target sample text using an initial text generation model; adjust the model parameters of the initial text generation model based on the first sample text, the second sample text and the target sample text to obtain a second text generation model.

[0217] In the above embodiment of the present application, the device is also used to construct a first loss function based on the first sample text and the target sample text, wherein the first loss function is used to increase the similarity between the target sample text and the first sample text; construct a second loss function based on the second sample text and the target sample text, wherein the second loss function is used to reduce the similarity between the target sample text and the second sample text; adjust the model parameters based on the first loss function and the second loss function to obtain a second text generation model.

[0218] In the above embodiments of the present application, the device is also used to obtain sample scene description texts of multiple sample scenes and sample associated description texts of sample objects; input the sample scene description texts, sample associated description texts, first sample texts and second sample texts into the initial text generation model, and generate target sample texts using the initial text generation model.

[0219] In the above embodiments of the present application, the acquisition module is also used to acquire multimodal data of the target object, wherein the multimodal data is used to describe the target object through different modalities; convert the multimodal data into an initial description text of the target object; structure the initial description text to obtain a structured description text; and retrieve sample text from a database based on the type of the target object, wherein the database is used to store sample texts of different types of objects.

[0220] According to an embodiment of the present application, a text generation device for implementing the above-mentioned text generation method is also provided. Fig.12 is a schematic diagram of a text generation device according to an embodiment of the present application, such as Fig.12 As shown, the device 1200 includes: an acquisition module 1202 , a first generation module 1204 , a collection module 1206 , and a second generation module 1208 .

[0221] Among them, the acquisition module is used to obtain the structured description text of the target product and the example title text corresponding to the structured description text, wherein the example title text is used to describe the product of the same type as the target product; the first generation module is used to input the structured description text and the example title text into the first text generation model, and use the first text generation model to generate multiple initial title texts corresponding to the target product, wherein the multiple initial title texts have the same attributes as the example title text; the collection module is used to put the multiple initial title texts into a preset scene, and collect feedback information of the multiple initial title texts, wherein the feedback information is used to indicate the user's preference for the multiple initial title texts in the preset scene; the second generation module is used to generate the target title text of the target product based on the feedback information and the multiple initial title texts using the second text generation model, wherein the target title text conforms to the feedback information.

[0222] It should be noted that the acquisition module 1202, the first generation module 1204, the acquisition module 1206, and the second generation module 1208 correspond to steps S802 to S808 in the above embodiment, and the four modules and the corresponding steps implement the same examples and application scenarios, but are not limited to the contents disclosed in the above embodiment. It should be noted that the above modules or units can be hardware components or software components stored in a memory and processed by one or more processors, and the above modules can also be run in the server 10 provided in the above embodiment as part of the device.

[0223] According to an embodiment of the present application, a text generation device for implementing the above-mentioned text generation method is also provided. Fig.13 is a schematic diagram of a text generation device according to an embodiment of the present application, such as Fig.13 As shown, the device 1300 includes: a first display module 1302 and a second display module 1304 .

[0224] Among them, the first display module is used to respond to the input instructions acting on the operation interface, and display the structured description text of the target object and the example text corresponding to the structured description text on the operation interface, wherein the example text is used to describe the object of the same type as the target object; the second display module is used to respond to the processing instructions acting on the operation interface, and display the target text of the target object on the operation interface, wherein the target text is obtained by using the second text generation model based on feedback information and multiple initial texts, the feedback information is feedback information of multiple initial texts collected by putting multiple initial texts into a preset scene, the multiple initial texts are obtained by inputting the structured description text and the example text into the first text generation model, the multiple initial texts have the same attributes as the example text, and the target text conforms to the feedback information.

[0225] It should be noted that the first display module 1302 and the second display module 1304 correspond to steps S902 to S904 in the above embodiment, and the two modules and the corresponding steps implement the same examples and application scenarios, but are not limited to the contents disclosed in the above embodiment. It should be noted that the above modules or units can be hardware components or software components stored in a memory and processed by one or more processors, and the above modules can also be run as part of the device in the server 10 provided in the above embodiment.

[0226] According to an embodiment of the present application, a text generation device for implementing the above-mentioned text generation method is also provided. Fig.14 is a schematic diagram of a text generation device according to an embodiment of the present application, such as Fig.14 As shown, the device 1400 includes: an acquisition module 1402 , a first generation module 1404 , a collection module 1406 , a second generation module 1408 , and an output module 1410 .

[0227] Among them, the acquisition module is used to obtain the structured description text of the target object and the example text corresponding to the structured description text by calling the first interface, wherein the first interface includes a first parameter, the parameter value of the first parameter includes the structured description text and the example text, and the example text is used to describe an object of the same type as the target object; the first generation module is used to input the structured description text and the example text into the first text generation model, and use the first text generation model to generate multiple initial texts corresponding to the target object, wherein the multiple initial texts have the same attributes as the example text; the collection module is used to put the multiple initial texts into a preset scene, and collect feedback information of the multiple initial texts, wherein the feedback information is used to indicate the user's preference for the multiple initial texts in the preset scene; the second generation module is used to generate the target text of the target object based on the feedback information and the multiple initial texts using the second text generation model, wherein the target text conforms to the feedback information; the output module is used to output the target text by calling the second interface, wherein the second interface includes a second parameter, and the parameter value of the second parameter includes the target text.

[0228] It should be noted that the acquisition module 1402, the first generation module 1404, the acquisition module 1406, the second generation module 1408, and the output module 1410 correspond to steps S1002 to S1010 in the above embodiment, and the five modules and the corresponding steps implement the same examples and application scenarios, but are not limited to the contents disclosed in the above embodiment. It should be noted that the above modules or units can be hardware components or software components stored in a memory and processed by one or more processors, and the above modules can also be run in the server 10 provided in the above embodiment as part of the device.

[0229] It should be noted that the preferred implementation scheme involved in the above embodiments of the present application is the same as the scheme provided in the above embodiments, as well as the application scenario and implementation process, but is not limited to the scheme provided in the above embodiments.

[0230] An embodiment of the present application may provide a computing device. Fig.15 is a structural block diagram of a computing device according to an embodiment of the present application. Fig.15 As shown, the computing device 100 may include: one or more ( Fig.15 (only one is shown) processor 102, memory 104, storage controller, and peripheral interfaces.

[0231] The above-mentioned computing device can be understood as an integrated intelligent terminal, including but not limited to a server, a desktop computer, a PC (Personal Computer), a model all-in-one machine, etc., and the computing device can be pre-installed with the model described in the above-mentioned embodiments of the present application.

[0232] Specifically, the computing device can pre-set multiple types of models, including but not limited to models in the fields of natural language processing, visual processing, speech processing, code processing, multimodal task processing, etc., so as to provide a variety of model choices. In different product forms, the computing device can support one or more model usage methods, including but not limited to model training, model calling, model fine-tuning, model deployment, model reasoning and application, etc. In some product forms, the computing device also supports model management, including but not limited to multi-type model management (supporting the management of multiple types of models such as discriminants and generative models), model version control (supporting the control of different model versions), model evaluation (based on model evaluation tools, evaluating the performance and effect of the model), etc. In other product forms, the computing device can also create applications based on models, provide API calling capabilities, and can call models to the created applications through API interfaces, while providing application management tools to achieve management and monitoring of applications.

[0233] Furthermore, the computing device can also include data management (supporting the creation and management of model tuning data sets), training center (providing rich training resources to help users learn and master AI technology), and basic management and control capabilities (providing enterprise-level basic management and control capabilities to ensure the security and efficient operation of the system). Through the above functions, a comprehensive, integrated AI development, training, deployment and application device is provided.

[0234] Among them, the memory can be used to store software programs and modules, such as program instructions / modules corresponding to the methods and devices in the embodiments of the present application. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, that is, implementing the methods in the above embodiments. The memory may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include a memory remotely arranged relative to the processor, and these remote memories may be connected to the terminal A via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0235] The processor may call the executable program stored in the memory through the transmission device to execute the method described in any one of the above embodiments.

[0236] An embodiment of the present application may provide an electronic device. Fig.16 is a structural block diagram of an electronic device according to an embodiment of the present application. Fig.16As shown, the electronic device may include: an input / output device 112 ; a memory 114 and a processor 116 , wherein the processor 116 is connected to the input / output device 112 and the memory 114 via a bus 118 .

[0237] Among them, the memory can be used to store software programs and modules, such as program instructions / modules corresponding to the methods and devices in the embodiments of the present application. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, that is, implementing the methods in the above embodiments. The memory may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include a memory remotely arranged relative to the processor, and these remote memories may be connected to the terminal A via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0238] The processor may call the executable program stored in the memory through the transmission device to execute the method described in any one of the above embodiments.

[0239] Those skilled in the art will understand that Fig.16 The structure shown is for illustration only, and the computing device may also be a smart phone (such as an Android phone, an iOS phone, etc.), a tablet computer, a PDA, a mobile Internet device (MID), a PAD, and other terminal devices. The figure does not limit the structure of the above computing device. For example, the computing device 100 may also include more or fewer components (such as a network interface, a display device, etc.) than those shown in the figure, or have a configuration different from that shown in the figure.

[0240] A person of ordinary skill in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing the hardware related to the terminal device through a program, and the program can be stored in a computer-readable storage medium, and the storage medium may include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0241] The embodiment of the present application further provides a computer-readable storage medium. Optionally, in this embodiment, the computer-readable storage medium can be used to store the program code executed by the method provided in the above embodiment.

[0242] Optionally, in this embodiment, the above storage medium may be located in a computing device.

[0243] Optionally, in this embodiment, the computer-readable storage medium is configured to store an executable program, and when the executable program is running, the device where the computer-readable storage medium is located is controlled to execute the method described in any one of the above embodiments.

[0244] The embodiment of the present application further provides a computer program product. Optionally, in this embodiment, the computer program product may include a computer program, and the computer program implements the method provided in the embodiment when executed by a processor.

[0245] The embodiments of the present application also provide a computer program product. Optionally, the computer program product may include a non-volatile computer-readable storage medium, which may be used to store a computer program, and when the computer program is executed by a processor, the method provided in the embodiments is implemented.

[0246] The embodiment of the present application further provides a computer program. Optionally, in this embodiment, when the computer program is executed by a processor, the method provided in the above embodiment is implemented.

[0247] In the above embodiments of the present application, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0248] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0249] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0250] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0251] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, server or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, disk or optical disk, etc., which can store program code.

[0252] The above is only a preferred implementation of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A text generation method, characterized in that: include: Acquire a structured description text of a target object and an example text corresponding to the structured description text, wherein the example text is used to describe an object of the same type as the target object; Inputting the structured description text and the sample text into a first text generation model, and using the first text generation model to generate a plurality of initial texts corresponding to the target object, wherein the plurality of initial texts have the same attributes as the sample text; Putting the multiple initial texts into a preset scene, collecting feedback information of the multiple initial texts, wherein the feedback information is used to indicate the user's preference for the multiple initial texts in the preset scene; Based on the feedback information and the multiple initial texts, a target text of the target object is generated using a second text generation model, wherein the target text conforms to the feedback information.

2. The method according to claim 1, characterized in that Based on the feedback information and the multiple initial texts, generating a target text of the target object using a second text generation model includes: Based on the feedback information, selecting a first text and a second text from the multiple initial texts, wherein the user preference degree of the first text is greater than the user preference degree of the initial texts other than the first text, and the user preference degree of the second text is less than the user preference degree of the initial texts other than the second text; Based on the first text and the second text, the target text is generated using the second text generation model.

3. The method according to claim 2, characterized in that Based on the first text and the second text, generating the target text by using the second text generation model includes: Obtaining a scene description text of the preset scene and an associated description text of the target object; The scene description text, the associated description text, the first text and the second text are input into the second text generation model, and the target text is generated by using the second text generation model.

4. The method according to claim 2, characterized in that: Selecting a first text and a second text from the plurality of initial texts based on the feedback information includes: In response to the feedback information being the amount of user interaction in the preset scenario, sorting the multiple initial texts based on the amount of user interaction to obtain a sorting result; Matching the multiple initial texts based on the sorting results to obtain multiple text pairs, wherein the text differences in different text pairs are different; The multiple text pairs are screened based on the text difference degree to obtain a target text pair, wherein the text difference degree of the target text pair is greater than the text difference degrees of other text pairs in the multiple text pairs except the target text pair, and the target text pair includes the first text and the second text.

5. The method according to claim 1, characterized in that The method further comprises: Get multiple sample texts of the sample object; Putting the plurality of sample texts into a plurality of sample scenarios, and collecting sample feedback information of the plurality of sample texts, wherein the sample feedback information is used to indicate the user preference degree for the plurality of sample texts in the plurality of sample scenarios; Selecting a first sample text and a second sample text from the plurality of sample texts based on the sample feedback information, wherein the user preference degree of the first sample text is greater than the user preference degree of sample texts other than the first sample text, and the user preference degree of the second sample text is less than the user preference degree of sample texts other than the second sample text; Based on the first sample text and the second sample text, generating a target sample text by using an initial text generation model; The model parameters of the initial text generation model are adjusted based on the first sample text, the second sample text and the target sample text to obtain the second text generation model.

6. The method according to claim 5, characterized in that The method further comprises: adjusting the model parameters of the initial text generation model based on the first sample text, the second sample text and the target sample text to obtain the second text generation model, including: Constructing a first loss function based on the first sample text and the target sample text, wherein the first loss function is used to increase the similarity between the target sample text and the first sample text; Constructing a second loss function based on the second sample text and the target sample text, wherein the second loss function is used to reduce the similarity between the target sample text and the second sample text; The model parameters are adjusted based on the first loss function and the second loss function to obtain the second text generation model.

7. The method according to claim 5, characterized in that Based on the first sample text and the second sample text, generating a target sample text by using an initial text generation model includes: Acquire sample scene description texts of the plurality of sample scenes and sample associated description texts of the sample objects; The sample scene description text, the sample association description text, the first sample text and the second sample text are input into the initial text generation model, and the target sample text is generated by using the initial text generation model.

8. The method according to claim 1, characterized in that Obtaining a structured description text of a target object and a sample text corresponding to the structured description text, including: Acquire multimodal data of the target object, wherein the multimodal data is used to describe the target object in different modalities; Converting the multimodal data into an initial description text of the target object; Performing structural processing on the initial description text to obtain the structured description text; Based on the type of the target object, the sample text is retrieved from a database, wherein the database is used to store sample texts of objects of different types.

9. A text generation method, characterized in that: include: Acquire a structured description text of a target product and an example title text corresponding to the structured description text, wherein the example title text is used to describe a product of the same type as the target product; Inputting the structured description text and the example title text into a first text generation model, and using the first text generation model to generate a plurality of initial title texts corresponding to the target product, wherein the plurality of initial title texts have the same attributes as the example title text; Putting the multiple initial title texts into a preset scene, collecting feedback information of the multiple initial title texts, wherein the feedback information is used to indicate the user's preference for the multiple initial title texts in the preset scene; Based on the feedback information and the multiple initial title texts, a target title text of the target product is generated using a second text generation model, wherein the target title text conforms to the feedback information.

10. A text generation method, characterized in that: include: In response to an input instruction acting on an operation interface, displaying a structured description text of a target object and an example text corresponding to the structured description text on the operation interface, wherein the example text is used to describe an object of the same type as the target object; In response to a processing instruction acting on the operation interface, a target text of the target object is displayed on the operation interface, wherein the target text is obtained by using a second text generation model based on feedback information and multiple initial texts, the feedback information is feedback information of the multiple initial texts collected by placing the multiple initial texts into a preset scene, the multiple initial texts are obtained by inputting the structured description text and the sample text into a first text generation model and using the first text generation model, the multiple initial texts have the same attributes as the sample text, and the target text conforms to the feedback information.

11. A text generation method, characterized in that: include: Acquire a structured description text of a target object and a sample text corresponding to the structured description text by calling a first interface, wherein the first interface includes a first parameter, a parameter value of the first parameter includes the structured description text and the sample text, and the sample text is used to describe an object of the same type as the target object; Inputting the structured description text and the sample text into a first text generation model, and using the first text generation model to generate a plurality of initial texts corresponding to the target object, wherein the plurality of initial texts have the same attributes as the sample text; Putting the multiple initial texts into a preset scene, collecting feedback information of the multiple initial texts, wherein the feedback information is used to indicate the user's preference for the multiple initial texts in the preset scene; Based on the feedback information and the multiple initial texts, generating a target text of the target object using a second text generation model, wherein the target text conforms to the feedback information; The target text is output by calling a second interface, wherein the second interface includes a second parameter, and a parameter value of the second parameter includes the target text.

12. An electronic device, characterized in that: include: A memory storing an executable program; A processor, connected to the memory via a bus, and configured to run the program, wherein the program executes the method described in any one of claims 1 to 11 when running.

13. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored executable program, wherein when the executable program is executed, the device where the computer-readable storage medium is located is controlled to execute the method according to any one of claims 1 to 11.

14. A computer program product, characterized in that The invention comprises a computer program which, when executed by a processor, implements the method according to any one of claims 1 to 11.

Images